Semantic understanding of large-scale aerial scenes represents a critical challenge in 3D computer vision, hindered by the prohibitive cost of dense annotation. This paper introduces EvoPropGS, a novel approach for the semantic segmentation of 3D Gaussian Splatting models that requires only minimal supervision. Our core insight is to leverage the inherent structural repetitions within aerial environments to propagate semantic information from a sparse set of annotations across the entire 3D scene. Our approach constructs a prompt library by pairing SAM-generated mask candidates with DINOv2 feature embeddings from annotated views. For unannotated regions, we generate pseudo-labels by matching region proposals with these featured prompts via cosine similarity. We then formulate optimal prompt selection as a discrete optimization problem solved via evolutionary search, guided by our novel fitness function that evaluates both 3D consistency and 2D semantic coherence. Extensive experiments demonstrate that EvoPropGS achieves accurate segmentation with only 2 percent annotated pixels.

Open-Vocabulary Object Detection (OVOD) shows promise in remote sensing (RS), but due to its unique value, there are challenges such as the predominance of background regions, sparse labels, limited semantic information, and difficulties in semi-supervised training. To tackle these challenges, we propose the Semi-Supervised Open-Vocabulary Aerial Object Detection with Dual-Perception Prior Denoising (SOAR), which explicitly models the background embeddings of each scene to indirectly construct foreground priors, thereby capitalizing on the abundant background information present in RS imagery. We further introduce a query enhancement module that integrates language and foreground prior information to enhance the effectiveness of query selection and feature augmentation. During the decoding stage of semi-supervised training, we perform denoising and reconstruction of the foreground priors to generate pseudo-labels that support the training process. Additionally, we address the sparsity of label information through expansion and aggregation techniques, further improving model performance. Experimental evaluations reveal that, in the open-vocabulary object detection task on the DIOR dataset, our method achieves a mean Average Precision (mAP) of 68.5% and Harmonic Mean (HM) of 55.9%, outperforming the previous state-of-the-art model’s mAP of 61.6% and HM of 53.6%. Our approach offers a novel solution to the open-vocabulary challenge in aerial object detection.

In recent years, the rapid progress of deep learning has driven notable advancements in satellite video tracking, a critical task for applications such as environmental monitoring, disaster management, and defense. Despite these strides, existing approaches remain constrained by their inability to handle dynamic challenges, such as target appearance variations, complex motion patterns, and occlusions. Traditional methods often suffer from static template matching or overly complex update mechanisms, compromising their robustness and practicality in real-world scenarios. To address these limitations, we propose a paradigm shift in satellite video tracking by integrating historical trajectory knowledge with visual features. This fusion enhances the tracker's perceptual understanding of targets over time, enabling more adaptive and resilient tracking. By aligning spatial, temporal, and cross-modal information, our approach effectively bridges the gap between fragmented observations and coherent tracking performance, even under challenging conditions like small target detection and cluttered backgrounds. Extensive experiments conducted on multiple satellite video tracking benchmarks demonstrate the superiority of our method, with HTTrack achieving success rates of 51.5% on SV248S, 52.9% on SatSOT, and 32.6% on VISO, significantly outperforming state-of-the-art trackers and marking a step forward in achieving robust, accurate, and scalable satellite video tracking.

Remote sensing (RS) image scene classification has wide applications in the field of RS. Although existing methods have achieved remarkable performance, there are still limitations in feature extraction and lightweight design. Current multi-branch models, although performing well, have large parameter counts and high computational costs, making them difficult to deploy on resource-constrained edge devices such as unmanned aerial vehicles (UAVs). On the other hand, lightweight models like StarNet, having less parameter, but rely on element-wise multiplication to generate features and lack the capture of explicit long-range spatial feature, resulting in insufficient classification accuracy. To address these issues, this letter proposes a lightweight mamba-based hybrid network, namely LMHMamba, whose core is an innovative lightweight multi-feature hybrid mamba (LMHM) module. This module combines the advantage of StarNet in implicitly generating high-dimensional nonlinear features, introduces a lightweight state space module to enhance spatial feature learning capabilities, and then uses local and global attention modules to emphasize local and global features. This enables effective multi-dimensional feature fusion while maintaining low parameter. We validate the performance of LMHMamba model on three remote sensing scene classification datasets and compare it with mainstream lightweight models and the latest methods. Experimental results show that LMHMamba achieves advanced levels in both classification accuracy and computational efficiency, significantly outperforming existing lightweight models, providing an efficient solution for edge deployment. Code is available at https://github.com/yizhilanmaodhh/LMHMamba.

Remote sensing image change detection (CD), as a pivotal technology for monitoring Earth’s surface dynamics, plays a crucial role in urban planning, resource management, and disaster assessment. Despite the success of deep learning-based methods, they still suffer from two significant limitations. Firstly, the inadequate exploitation of frequency-domain information restricts their ability to capture subtle structural and edge changes. Secondly, the suboptimal interaction strategies that predominantly rely on attention mechanisms or direct feature exchange that fails to fully model complex semantic differences between bi-temporal images. To address these challenges, we propose a Wavelet Attention-based Group Interaction Network (WAGINet), which leverages wavelet attention for joint frequency-spatial domain feature learning and uses a group-wise feature exchange mechanism to optimize bi-temporal interaction. The wavelet attention module decomposes features into high-low frequency components and emphasize important ones to improve edge-aware feature extraction. In the meanwhile, the group interaction strategy enables both channel-group and spatial-group feature exchange to capture the correlation between bi-temporal features while better protecting structural integrity, so that promotes more discriminative change representation. Experimental results on public LEVIR-CD and WHU-CD datasets show that WAGINet outperforms existing state-of-the-art methods, providing an effective solution for high-precision remote sensing image CD in complex scenarios. Code available at https://github.com/yizhilanmaodhh/WAGINet.

Referring remote sensing interpretation holds significant application value in various scenarios such as ecological protection, resource exploration, and emergency management. However, referring remote sensing expression comprehension and segmentation (RRSECS) faces critical challenges, including micro-target localization drift problem caused by insufficient extraction of boundary features in existing paradigms. Moreover, when transferred to remote sensing domains, polygon-based methods encounter issues such as contour-boundary misalignment and multi-task co-optimization conflicts problems. In this paper, we propose SeeFormer, a novel contour autoregressive paradigm specifically designed for RRSECS, which accurately locates and segments micro, irregular targets in remote sensing imagery. We first introduce a brain-inspired feature refocus learning (BIFRL) module that progressively attends to effective object features via a coarse-to-fine scheme, significantly boosting small-object localization and segmentation. Next, we present a language-contour enhancer (LCE) that injects shape-aware contour priors, and a corner-based contour sampler (CBCS) to improve mask-polygon reconstruction fidelity. Finally, we develop an autoregressive dual-decoder paradigm (ARDDP) that preserves sequence consistency while alleviating multi-task optimization conflicts. Extensive experiments on RefDIOR, RRSISD, and OPTRSVG datasets under varying scenarios, scales, and task paradigms demonstrate transformative performance gains: compared to the baseline PolyFormer, our proposed SeeFormer improves oIoU and mIoU by 27.58% and 39.37% for referring image segmentation and by 18.94% and 28.90% for visual grounding on the RefDIOR dataset.

The iterative search process of evolutionary algorithms (EAs) encapsulates optimization knowledge within historical populations and fitness evaluations. Effective utilization of this knowledge is crucial for facilitating knowledge transfer and online adaptation. However, current research typically addresses these goals in isolation and faces distinct limitations: evolutionary sequential transfer optimization often suffers from incomplete utilization of prior knowledge, while adaptive strategies, utilizing real-time knowledge, are limited to tailoring specific evolutionary operators. To simultaneously achieve these two capabilities, we introduce the Optimization Knowledge Adaptation Evolutionary Model (OKAEM), a unified learnable evolutionary framework capable of adaptively updating parameters based on available optimization knowledge. By parameterizing evolutionary operators via attention mechanisms, OKAEM enables learnable update rules that facilitate the utilization of optimization knowledge via two phases: pre-training to integrate extensive prior knowledge for efficient transfer, and adaptive optimization to dynamically update parameters based on real-time knowledge. Experimental results confirm that OKAEM significantly outperforms state-of-the-art sequential transfer methods across 12 transfer scenarios via pre-training, and surpasses advanced learnable EAs solely through its self-tuning mechanism in prior-free settings. Beyond demonstrating practical utility in prompt tuning for vision-language models, ablation studies validate the necessity of the learnable components, while visualization analyses reveal the model's capacity to autonomously discover interpretable evolutionary principles.

Matrix factorization is a fundamental characterization model in machine learning and is usually solved using mathematical decomposition reconstruction loss. However, matrix factorization is a data-driven model whose results depend on data quality, making it susceptible to noise. Inspired by physics, the law of conservation of energy is used to introduce physical laws into matrix factorization, which is called Physics-informed Matrix Factorization operator (PiMF). The PiMF operator uses the heat conduction equation to construct the energy objective function for matrix factorization, thereby retaining the mathematical model's decomposition meaning and satisfying the interpretability of physics. The PiMF follows the physical laws, thereby suppressing irregular or sudden noise signals that violate these physical principles. The solutions of the PiMF operator include more comprehensive knowledge of mathematics and physics, which improves the ability to generalize complex data, especially for noisy data. We demonstrate the consistency of the energy objective function and the mathematical model, which verifies the feasibility of matrix factorization using physical energy laws. In addition, the physical interpretability of the PiMF operator is proved from the perspective of energy decline. This study proposes two practical algorithms for PiMF in classification and clustering tasks, enhancing the practicability of matrix factorization by incorporating task-specific prior information constraints. The experimental results of PiMF for classification and clustering demonstrate the advantages of the proposed operator. The importance of physics-informed matrix factorization is verified, especially for noisy data.

Zero-shot captioning aims to generate descriptive captions for unseen image and video data by leveraging the potential of visual language models (VLMs) and language models (LMs) without requiring task-specific training. It has emerged as a critical task, but its performance is often hindered by the inherent gap between the training distribution and unseen test data. The fundamental challenge lies in the model’s strong dependence on the marginal distribution of the training data, which leads to biased predictions when handling test samples. To address this issue, we propose an Energy-aware Reinforcement Feedback Calibration (ERFC) framework to calibrate the distribution and predictions of caption models from a novel energy perspective. The calibration process of ERFC is divided into two key components: 1) We first construct an Energy Stabilizer (ES) based on the caption model, where energy is considered a measure of the affinity between the input sample and the model’s learned distribution. ES iteratively adjusts the embedding features of the input sample using Langevin Dynamics, reducing its energy to implicitly align the model’s distribution with the unseen target domain. 2) We deploy a Reinforcement Calibrator (RC) to refine and calibrate the generated captions through a reward-feedback mechanism. RC leverages the expert CLIP model as a reward signal to assess the quality of the generated captions and employs the policy gradient algorithm to reward or penalize the model, thereby improving its performance. By iteratively combining energy-based optimization and reward-driven calibration, ERFC achieves superior zero-shot generalization capabilities, as demonstrated on image benchmarks such as MSCOCO, Flickr30K, and NoCaps, as well as video benchmarks such as MSR-VTT and MSVD.

Textural details are useful for image super-resolution, but massive CNN methods ignored the high-frequency components and generated over-smoothed outputs. The knowledge-based contourlet inference network is proposed in this paper. Different from other CNN-based methods that are directly infer high-resolution (HR) images, our model learns to reconstruct the HR image through the series of corresponding contourlet coefficients. Specifically, first, we consider the low-pass subbands of the contourlet as the corresponding low-resolution (LR) image. Then, feed it to the embedding net with residual blocks to provide adequate information for the contourlet coefficients prediction. Finally, we innovatively convert the estimation of contourlet coefficients into the estimation of the generalized gaussian distribution (GGD) parameters, and design the corresponding loss function to ensure training stability, which explores the smoothness of the contour effectively and guarantees the general structure and details of images. Experiments on four remote sensing datasets, four natural scenes and human-made content datasets, and the outdoor dataset demonstrate the superiority of the proposed model quantitatively and qualitatively.

Panchromatic images (PANs) and multispectral (MS) images (MSs) are widely used for dual-source remote sensing image classification, gradually becoming a research hotspot. However, making the most of dual-source image information with insufficiently labeled samples is a significant challenge. This article proposes a progressive semi-distillation model (PSDM) to classify dual-source remote sensing images with insufficient samples. We design a framework of rookie teacher network (RTN)-teaching assistant system (TAS)-student grouping network (SGN) in the case of a traditional teacher network (TN) (i.e., rookie TN (RTN)) that does not provide excellent guidance to student network (SN) due to insufficient samples. The PSDM expands the samples and compresses the space through the RTN-SGN structure to cope with the dilemma of insufficient samples. To make RTN better guide the SGN, we design TAS, which can gradually guide SGN to learn the samples from easy to difficult. It can also further assist SGN training to improve the classification performance of SGN with insufficient samples. We design SGN and add cooperation and correction mechanism to better learn dual- source information. These strategies can eliminate SGN's over-dependence on the RTN, help SGN outperform the RTN, and achieve the effect of semi-distillation. Experimental results and theoretical analysis have sufficiently pointed out the proposed method's accuracy, efficiency, and robustness under insufficient sample situations. Our model is available at https://github.com/MarjordCpz/PSDM.

In recent years, multivariate time series anomaly detection has become an important research topic in the field of anomaly detection. In Industrial Internet of Things (IIoT) systems, the collected data may be affected by internal failures, external disturbances, or other adverse factors. In such cases, appropriate anomaly detection methods are required to ensure the stable operation of the system. However, existing methods based on reconstruction, prediction, or hybrid approaches often suffer performance degradation when anomalies are present in large amounts of training data, as these anomalies can negatively impact the training process. To address this challenge, we propose a dynamic masking strategy in both temporal and spatial dimensions. We develop a time series imputation framework based on a diffusion model that integrates Graph Neural Network (GNN) and Transformer architectures. This framework, termed Spatial-Temporal Dynamic Mask Diffusion for Anomaly Detection (STDMD-AD), incorporates a dynamic masking mechanism: temporally, reconstruction errors are used to mask data by randomly concealing values with higher errors; spatially, attention is applied to mask channels that are more likely to contain anomalies during training. Experiments on five real-world datasets demonstrate that the proposed method outperforms existing benchmarks and achieves state-of-the-art anomaly detection performance.

With the Transformer architecture achieving impressive results in the vision domain. It has become a current popular research to explore more potentials of Transformer mixed architectures and explore more suitable mixed combinations. In this paper, Transformer classification network is designed and explored by multi-task architecture search algorithm. A new paradigm for multi-task architecture is designed by combining convolution and Transformer. The designed search architecture can combine the respective advantages of convolution and Transformer and can obtain better performance. At the same time, corresponding knowledge-aware multi-task genetic operators are designed to generate offspring individuals. Inter-task and inter-experience knowledge-aware is utilised to facilitate evolutionary convergence. During the search process, the reference evaluation method is utilised to reduce the redundant computation and time during the search process. In the experimental section, the search results are compared with state-of-the-art architectures and search algorithms. The experimental results confirm the effectiveness and high generalisation of the searched architectures. The ablation experimental part proves the effectiveness of the proposed architectural paradigm, genetic operators and reference evaluation.

Recently, the zero-shot image captioning (zero-shot IC) method based on pre-trained visual language models (VLMs) and large language models (LLMs) has made significant progress. However, how to adapt it to the zero-shot video captioning (zero-shot VC) scenario (without video-text paired supervision) has not been well explored. Inspired by various recent test-time strategies (sacrificing additional test time to improve performance), we try to introduce a new paradigm of Test-time Reinforcement Polishing in zero-shot VC scenario. We take temporal dependency modeling as the starting point and propose a novel framework for Refinable Zero-shot VC, called RefZVC. RefZVC can greatly cover the long-term context of the video and continuously polish and refine the generated captions in a reward-feedback manner. We first design an Adaptive Frame Skipping module (AdaSkip) to skip redundant frames and select diverse keyframe sequences. Subsequently, we propose a Multi-granularity Reinforcement Polishing (MRP) mechanism, which iteratively polishes captions by leveraging Gaussian Kernel Cache (GKC) to capture temporal dynamics, store and reuse relevant historical context. In addition, MRP calculates rewards for generated captions at both the sentence-level and entity-level to achieve test-time polishing. With the MRP mechanism, RefZVC achieves superior zero-shot generalization performance, outperforming previous zero-shot VC methods on benchmarks such as MSVD, MSR-VTT, and VATEX.

Although significant advances have been achieved in SAR land-cover classification, recent methods remain predominantly focused on supervised learning, which relies heavily on extensive labeled datasets. This dependency not only limits scalability and generalization but also restricts adaptability to diverse application scenarios. In this paper, a general-purpose foundation model for SAR land-cover classification is developed, serving as a robust cornerstone to accelerate the development and deployment of various downstream models. Specifically, a Dynamic Instance and Contour Consistency Contrastive Learning (DI3CL) pre-training framework is presented, which incorporates a Dynamic Instance (DI) module and a Contour Consistency (CC) module. DI module enhances global contextual awareness by enforcing local consistency across different views of the same region. CC module leverages shallow feature maps to guide the model to focus on the geometric contours of SAR land-cover objects, thereby improving structural discrimination. Additionally, to enhance robustness and generalization during pre-training, a large-scale and diverse dataset named SARSense, comprising 460,532 SAR images, is constructed to enable the model to capture comprehensive and representative features. To evaluate the generalization capability of our foundation model, we conducted extensive experiments across a variety of SAR land-cover classification tasks, including SAR land-cover mapping, water detection, and road extraction. The results consistently demonstrate that the proposed DI3CL outperforms existing methods. Our code and pre-trained weights are publicly available at: https://github.com/SARpre-train/DI3CL.

Remote sensing image captioning is a multimodal foundation task for fine-grained understanding of remote sensing images. However, remote sensing images contain complex scenes and rich objects, it is very challenging to accurately describe the objects in the scene with their attributes and dependencies. To address these issues, the article proposes a novel scale-aware prompting with optimal transport (SPOT) to learn effective multiscale features under diverse scenes, and to build fine-grained cross-modal alignment between semantic features and linguistic words during caption generation. Specifically, a scale-aware prompt extractor is constructed to explore feature integrations in complex scenes through learning prompts that query multi-scale features, and to enhance the representation of attributes and dependencies for objects by embedding positional relations. Besides, a fine-grained cross-modal alignment is designed to dynamically match image feature representations and textual semantics through optimal transport. Through the above manner, the model can learn effective language-aligned feature representations for caption generation. Finally, a caption Transformer with causal self-attention is introduced to generate accurate captions for remote sensing scenes. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on three public datasets, with the superiority of the proposed method further demonstrated by ablating the role of each component.

The multi-scale geometric analysis is a great representation tool. It can be used to improve the feature representation and learning process of deep networks. In addition to extracting features, the multi-scale geometric prior knowledge can also be used for the structure improvement of deep networks. In this paper, we propose a multi-scale scattering representation learning network, abbreviated as MSRLN, for image classification tasks. The exploration of structure improvement can be made with multi-scale scattering operations. In this way, the better singularity representation learning process for networks can be achieved. Firstly, the filter banks and multi-scale scattering operator are introduced for non-linear and singularity representation. Secondly, the novel multi-scale scattering representation learning network structure is designed. The scaling- wise scattering process is deployed in the shallow layer as a non-linear layer. This structure essentially supplements deep networks with geometric prior knowledge. It can further improve the non-linear activation and singularity representation process. Thirdly, we put forward the multi-stage scattering representation strategy and the prior knowledge weakening mechanism. With flexible scaling factors and learning rates, the stepwise approximation and learning process of networks can be achieved. In sum, MSRLN is a kind of structural innovative, and the scattering singularity representation structure can be extended to other backbones or tasks. Extensive experimental results show that MSRLN can achieve better image classification accuracy. Finally, necessary convergence, insight, and adaptability analyses are provided in evaluation experiments.

Locality Preserving Projections (LPP) aims to find a projection matrix to map the high-dimensional data into a low-dimensional subspace while preserving the local manifold structure, which is a classical unsupervised subspace learning method. However, the lack of label guidance makes LPP not able to fully exploit the discriminative information of the data. To solve the problem, we propose a Self-Guided Discriminative LPP algorithm employing pseudo labels learned by K-Means to guide the subspace learning. In this way, it facilitates the discovery of discriminative cluster information while preserving inherent manifold structure. Besides, considering K-Means' sensitivity to selection of cluster centroids, we introduce a centerless K-Means method to improve robustness by eliminating the need of centroid initialization. We also discuss the internal relationship between K-Means and LPP, and prove that K-Means can be written in the form of LPP under certain conditions. Experiments on seven benchmark datasets demonstrate that our method greatly improves the clustering performance.

Foundational vision-language models (VLMs) like CLIP are redefining the vision domain with their exceptional generalization capabilities. Prompt-based learning methods adapt pre-trained VLMs to video action recognition tasks using task-specific learnable text tokens. However, these tokens often struggle to generalize to unseen categories, as they tend to forget general textual knowledge. To address this, we construct knowledge prompts composed of handcrafted and descriptive prompts and introduce a novel knowledge-guided context mapping to enhance the generalization of learnable prompts to unseen categories. This approach mitigates the forgetting of fundamental knowledge by reducing the discrepancy between learnable prompts and knowledge prompts while simultaneously allowing the prompts to extract rich contextual knowledge from LLM data. Then, incorporating the knowledge-guided context mapping into the contrastive loss enables zero-shot transfer of prompts to new categories and data, providing discriminative prompts for both seen and unseen tasks. In addition, we propose an advanced temporal aggregation method that refines uniform mean pooling by incorporating frame-level textual relevance scoring. Extensive evaluations on multiple benchmarks demonstrate that learning to prompt with refining text knowledge is an effective quick-tuning method, achieving superior sample generalization performance without increasing training parameters.

RGB-based object tracking is a fundamental task in computer vision, aiming to identify, locate, and continuously track objects of interest across sequential video frames. Despite the significant advancements in the performance of traditional RGB trackers, they still face challenges in maintaining accuracy and robustness in the presence of complex backgrounds, occlusions, and rapid movements. To tackle these challenges, combining visual auxiliary modalities has gained significant attention. Beyond this, integrating natural language information offers additional advantages by providing high-level semantic context, enhancing robustness, and clarifying target priorities, further elevating tracker performance. This work proposes theAdaptiveMulti-modalVisual Tracking with Dynamic Semantic Prompts (AMVTrack) tracker, which efficiently incorporates image descriptions and avoids text dependency during tracking to improve flexibility and adaptability. AMVTrack significantly reduces computational resource consumption by freezing the parameters of the image encoder, text encoder, and Box Head and only optimizing a few learnable prompt parameters. Additionally, we introduce the Adaptive Dynamic Semantic Prompt Generator (ADSPG), which dynamically generates semantic prompts based on visual features, and theVisual-LanguageFusionAdaptation (V-L FA) method, which integrates multi-modal features to ensure consistency and complementarity of information. Additionally, we partition the Image Encoder to conduct an in-depth investigation into the relationship between the importance of features across different depth and width regions. Experimental results demonstrate that AMVTrack achieves significant performance improvements on multiple benchmark datasets, proving its effectiveness and robustness in complex scenarios.

Satellite video tracking presents significant challenges due to unpredictable target variations, environmental disturbances, and occlusions. Existing approaches either rely on auxiliary modalities or require full fine-tuning of foundation models, resulting in excessive parameter sensitivity and poor generalization. Meanwhile, conventional prompt-based tuning only updates parameters at a single location, limiting its ability to adapt to complex appearance changes. To address these limitations, we propose Adaptive Visual Prompting for Effective Satellite Video Tracking (AVPTrack). Unlike conventional prompts, introduced Super Prompts dynamically refine the original template at multiple distinct positions. This multi-location adaptation allows for fine-grained representation learning, enabling the tracker to better capture target variations and resist environmental disturbances. Additionally, Dynamic Templates are introduced to mitigate tracking failures in highly challenging scenarios, such as occlusions and background clutter, ensuring robust target localization. Furthermore, the Template Selection Adapter (TSA) selects the most relevant templates in real-time, enhancing tracking efficiency. These components are optimized during training while keeping other parameters frozen, ensuring parameter efficiency. We also investigate the relationship between fine-tuning proportions and learning rates to optimize model performance. Extensive evaluations on the SV248S, SatSOT, and VISO datasets demonstrate the superior adaptability and robustness of AVPTrack compared to existing methods.

Remote sensing (RS) images are prone to various degradations, which poses challenges to downstream tasks. Although existing single-task remote sensing image restoration methods are effective, they lack generalizability across tasks. All-in-one methods can handle multiple degradation tasks, but they usually focus on spatial information, ignoring the physical properties of the degradation information. To address the above limitations, we propose a Multiscale Spatial-Frequency Degradation Decoupling framework for All-in-One remote sensing image restoration (SFD$^{2}$IR), which decouples degradation features across different tasks to guide the model in performing task-specific image restoration. Specifically, a task-specific instruction generator (TIG) is proposed first to transform degradation features into task-specific prompts. Then, a multi-scale multi-frequency enhancement (MME) module is designed to decouple degradation effects from both spatial and frequency perspectives, thus enhancing the model's adaptability to various degradation types. Finally, a prompt feature refinement (PFR) module is developed to further refine the model's response to degraded tasks. Extensive experiments demonstrate that the proposed method achieves excellent performance on different RSIR tasks, including cloud removal, deblurring, dehazing, and super-resolution. The source code will be publicly available at SFD$^{2}$IR.

Existing diffusion-based methods have achieved impressive results in human motion editing. However, these methods often exhibit significant ghosting and body distortion in unseen in-the-wild cases. In this paper, we introduce Edit-Your-Motion, a video motion editing method that tackles these challenges through one-shot fine-tuning on unseen cases. Specifically, firstly, we utilized DDIM inversion to initialize the noise, preserving the appearance of the source video and designed a lightweight motion attention adapter module to enhance motion fidelity. DDIM inversion aims to obtain the implicit representations by estimating the prediction noise from the source video, which serves as a starting point for the sampling process, ensuring the appearance consistency between the source and edited videos. The Motion Attention Module (MA) enhances the model's motion editing ability by resolving the conflict between the skeleton features and the appearance features. Secondly, to effectively decouple motion and appearance of source video, we design a spatio-temporal two-stage learning strategy (STL). In the first stage, we focus on learning temporal features of human motion and propose recurrent causal attention (RCA) to ensure consistency between video frames. In the second stage, we shift focus on learning the appearance features of the source video. With Edit-Your-Motion, users can edit the motion of humans in the source video, creating more engaging and diverse content. Extensive qualitative and quantitative experiments, along with user preference studies, show that Edit-Your-Motion outperforms other methods.

Weakly supervised video anomaly detection (WSVAD) aims at predicting frame-level anomaly scores by modeling training videos with video-level annotations. The category names of abnormal events contain high-level knowledge abstracted by humans about abnormalities, which is of great help in identifying abnormal events. To utilize the knowledge implicit in category names, based on the visual-language pretraining model, we introduce a learnable abnormal prompt from three aspects: learnable domain prompt, learnable category prompt, and nonlearnable category definition prompt. Based on the learnable abnormal prompt, we propose a novel fine-grained WSVAD method: PromptVAD, which exploits a learnable abnormal prompt to reduce the semantic gap between visual images and anomaly categories. Through a similarity measure and our proposed coarse-grained two-class prompt module, our PromptVAD jointly learns coarse-grained and fine-grained VAD. Extensive experimental results on the ShanghaiTech, University of Central Florida (UCF)-Crime, and XD-Violence datasets show that our method achieves state-of-the-art performance. Specifically, our method achieves an area under the curve (AUC) of 88.62% on the UCF-Crime dataset.

The contemporary research paradigm in remote sensing hyperspectral monitoring increasingly relies on multisatellite and multiplatform Earth observation. While the traditional hyperspectral research framework based on single-image processing (SIP) has facilitated the application of idealized scenarios and the development of standardized evaluation benchmarks, it inherently constrains the model's ability to generalize feature representations across varying spatial and temporal domains. As hyperspectral data applications grow in complexity and data requirements, the limitations of SIP in addressing the demands of modern remote sensing tasks become increasingly apparent. To overcome these research limitations, we utilize the decentralized nature and data security features of federated learning to propose a cross-image hyperspectral image (HSI) federated learning approach for classification tasks. We first develop a client-oriented self-guided knowledge-enhanced personalized learning method that enhances the personalization of the local learning process by leveraging relevant features from other clients, thereby improving the learning efficiency of each client. To address the issue of "bias" in global knowledge caused by uneven data distribution across the federated learning process, we introduce a multiscale semantic aligned dynamic aggregation method to ensure fairness in integrating global knowledge. To our knowledge, this article is the first to explore the joint learning of HSI classification using federated learning. Accordingly, we have constructed open-set and closed-set datasets tailored to this task and have demonstrated the effectiveness of our method on these datasets. The code is available at: https://github.com/Gallipaxi/FedHIC.

In feature learning (FL), structural information shows advantages in retaining information and maintaining stability. Graph diffusion, a graph learning method that can focus on neighborhood structure and transmit information, has great research potential. In this study, a novel dynamic diffusion graph nonnegative matrix factorization (DGNMF) method is proposed, which uses a diffusion graph to improve the performance of FL and further enhances the effectiveness and stability of downstream classification tasks. DGNMF aims to mine and retain structural information more deeply in FL to build a more powerful and stable FL method. First, the model embeds graph learning into FL to obtain features containing structural information. Second, dynamic diffusion graph learning is used to mine deeper and more global structural information. Finally, we construct an updateable indicator matrix to enhance the discriminability of features. The classification experimental results of DGNMF on six databases demonstrate its advantages, verify its effectiveness and stability, and prove the importance of diffusion graph in improving FL.

Matrix factorization (MF) is a fundamental problem in machine learning, which is usually used as a feature learning method in various fields. For complex data involving spatiotemporal interactions, MF that only handles 2-D data will disrupt spatial dependence or temporal dynamics, failing to effectively couple spatial information with temporal factors. According to Markov chain principle, the spatial information of the present time is related to the spatial state of the previous time. We propose a spatial-temporal diffusion model for MF (STDMF), which uses graph diffusion to couple spatial-temporal information. Then, MF is used to learn the joint feature of data and spatial-temporal diffusion graph. Specifically, STDMF utilizes the graph diffusion with physical laws to generate spatial-temporal structure information. It obtains the underlying core structure of complex systems from a global perspective, which enhances the generalization ability of MF in noisy time-series data. To learn the lowest rank subspace of MF in time-series data, STDMF uses structural learning to constrain the rank of the learned features. Finally, STDMF is applied to clustering and anomaly detection of dynamic graph. The effectiveness of this method is verified by sufficient experiments, especially for noisy data.

Deep neural networks have achieved promising progress in signal modulation classification (SMC), playing an essential role in a variety of applications such as cognitive radio networks, cyber defense, and electronic surveillance. However, most existing SMC methods still follow the traditional machine learning paradigm that trains on static closed datasets, lacking the ability to cope with the challenge of continuous data distribution shifts in real communication scenarios. Directly applying the model to a new environment may lead to severe degradation of classification performance on previous scenarios, i.e., catastrophic forgetting. To address this, this article proposes the first domain-incremental learning (DIL) paradigm for SMC and designs a parameter-efficient isolation DIL (PID) method, which enables SMC models to rapidly adjust to new scenarios by extending only a few parameters, while significantly retaining classification capabilities on previous scenarios. Specifically, we first propose a parameter space decomposition-based classifier (PSD), separating the model parameters into a set of bases and corresponding coefficients. By freezing the bases and fine-tuning the low-dimensional coefficients, the catastrophic forgetting problem can be efficiently eliminated. Furthermore, we design a scene-aware domain controller (SDC) to select the most suitable domain-specific coefficients for each sample, thereby maintaining the SMC model's classification capabilities across all domains. The extensive experimental results show the superiority of the proposed PID, which achieves state-of-the-art (SOTA) overall performance. The code will be available at: https://github.com/SMC-IL/PID.

The increasing proliferation of unmanned aerial vehicles (UAVs) poses significant challenges to airspace security, necessitating the development of effective detection technologies. Passive detection techniques, such as passive radar, offer key advantages including spectrum efficiency and covert operation. However, passive radars that rely on coherent integration are often computationally expensive and dependent on strong Doppler signatures, rendering them ineffective for detecting low-speed or hovering UAVs. To overcome these limitations, we explore the use of channel state information (CSI) time series to characterize UAV presence and propose a temporal variation network for detecting UAV states, including hovering conditions. Our method utilizes digital terrestrial multimedia broadcast (DTMB) signals, which have wide coverage and high transmission power. By capturing DTMB signals with a single receiver, we reduce the complexity of passive detection systems. First, we perform channel estimation on the received signal to obtain CSI, which is arranged in frame order to form a CSI time series. This enables the modeling of interference channels caused by UAVs. We then propose the Channel Estimation and Temporal Variation Network (CETVNet), which leverages an adaptive noise reduction module and a multi-period feature extraction module to process these series for passive UAV state detection. Finally, a real-world signal dataset is collected using a software-defined radio device to train and evaluate CETVNet. Experimental results demonstrate that CETVNet achieves superior performance compared to state-of-the-art methods.

Synthetic Aperture Radar (SAR) image registration is an essential upstream task in geoscience applications, in which pre-detected keypoints from two images are employed as observed objects to seek matched-point pairs. In general, the registration is regarded as a typical closed-set classification, which forces each keypoint to be classified into the given classes, but ignoring an essential issue that numerous redundant keypoints are beyond the given classes, which unavoidably results in capturing incorrect matched-point pairs. Based on this, we propose a Cross-Rejective Open-set SAR Image Registration (CroR-OSIR) method. In this work, these redundant keypoints are regarded as out-of-distribution (OOD) samples, and we formulate the registration as a special open-set task with two modules: supervised contrastive feature-tuning and cross-rejective open-set recognition (CroR-OSR). Unlike traditional open-set recognition, all samples, including OOD samples, are available in the CroR-OSR module. CroR-OSR conducts the closed-set classifications in individual open-set domains from two images, meanwhile employing the cross-domain rejection during training, to exclude these OOD samples based on confidence and consistency. Moreover, a new supervised contrastive tuning strategy is incorporated for feature-tuning. Especially, the cross-domain estimation labels obtained by CroR-OSR are fed back to the feature-tuning module for feature-tuning, to enhance feature discriminability. The experimental results illustrate that the proposed method achieves more precise registration than the state-of-the-art methods. The code is released at https://github.com/XDyaoshi/CroR-OSIR-main.

Multiple-instance learning (MIL) was initially proposed to identify key instances within a set (bag) of instances when only one bag-level label is provided. Current deep MIL models mostly solve multi-instance problem in feature space. Nevertheless, with the increasing complexity of data, we found this paradigm faces significant risks in representation learning stage, which could lead to algorithm degradation in deep MIL models. We speculate that the degradation issue stems from the persistent drift of instances in feature space during learning. In this paper, we propose a novel Probability-Space MIL network (PSMIL) as a countermeasure. In PSMIL, a self-training alignment strategy is introduced in probability space to cope with the drift problem in feature space, and the alignment target objective is proven mathematically optimal. Furthermore, we reveal that the widely-used attention-based pooling mechanism in current deep MIL models is easily affected by the perturbation in feature space and further introduce an alternative called probability-space attention pooling. It effectively captures the key instance in each bag from feature space to probability space, and further eliminates the impact of selection drift in the pooling stage. To summarize, PSMIL seeks to solve a MIL problem in probability space rather than feature space. Experimental results illustrate that PSMIL could potentially achieve performance close to supervised learning level in complex tasks (gap within 5\%), with the incremental alignment in propability space bring more than 19\% accuracy improvements for current existing mainstream models in simulated CIFAR datasets. For existing publicly available MIL benchmarks/datasets, attention in probability space also achieves competitive performance to the state-of-the-art deep MIL models. Codes are available at \url{https://github.com/LMBDA-design/PSAMIL}.

Visual grounding (VG) refers to detecting the specific objects in images based on linguistic expressions, and it has profound significance in the advanced interpretation of natural images. In remote sensing image interpretation, visual grounding is limited by characteristics such as the complex scenes and diverse object sizes. To solve this problem, we propose a novel remote sensing visual grounding (RSVG) framework, named language-guided hybrid representation learning Transformer (LGFormer). Specifically, we designed a multimodal dual-encoder Transformer structure called the adaptive multimodal feature fusion module. This structure innovatively integrates text and visual features as hybrid queries, enabling early-stage decoding queries to perceive the target position accurately. Then, the different modal information from the dual encoders is aggregated by hybrid queries to obtain the final object embedding for coordinate regression. Besides, a multi-scale cross-modal feature enhancement module (MSCM) is designed to enhance the self-representation of the extracted text and visual features and align them semantically. As for the hybrid queries, we use linguistic guidance to select visual features as the visual part and sentence-level features as the textual part. Finally, the LGFormer model we designed achieved the best results compared to existing models on the DIOR-RSVG and OPT-RSVG datasets.

Semi-supervised semantic segmentation (S4) has shown significant promise in reducing the burden of labor-intensive data annotation. However, existing methods mainly rely on pixel-level information, neglecting the strong region consistency inherent in remote sensing images (RSIs), which limits their effectiveness in handling the complex and diverse backgrounds of RSIs. To address this, we propose RegionMatch, a novel approach that leverages unlabeled data from a fresh object-level perspective, which is more tailored to the nature of semantic segmentation. We design the Pixel-Region Synergy Pseudo-Labeling strategy, which explicitly injects object-level contextual information into the S4 pipeline and promotes knowledge collaboration between pixel and region perspectives for generating high-quality pseudo-labels. In addition, we propose the Region Structure-Aware Correlation Consistency, which models object-level relationships by establishing inter-region correlations across images and pixel correlations within regions, providing more effective supervision signals for unlabeled data. Experimental results demonstrate that RegionMatch outperforms state-of-the-art methods on multiple authoritative remote sensing datasets, highlighting its superiority in the RSIs.

Multi-modal visual tracking leverages complementary sensor information to enhance robustness under challenging conditions. However, the security of multi-modal tracking systems remains largely unexplored. Existing attacks primarily target single-modal trackers or independently disrupt each modality, failing to exploit the inherent feature interactions and fusion mechanisms that define multi-modal tracking. As a result, these methods exhibit limited attack effectiveness and fail to assess multi-modal tracking systems' vulnerabilities accurately. Understanding these security risks is crucial, as adversarial threats could lead to severe failures in safety-critical applications. To address these challenges, a feature-aware adversarial attack, termed FA3T is proposed. It is designed to explicitly disrupt feature extraction and cross-modal alignment, thereby weakening the fusion process that multi-modal trackers rely on. To achieve this, a Frequency-Spatial Feature Separation (FSFS) module is constructed to perturb feature representations at multiple levels, weakening the modality-complementary advantages of multi-modal tracking. Furthermore, a Target Confusion Attack (TCA) module is devised to manipulate the target-background-template relationships, making it increasingly difficult for the tracker to distinguish the true target, significantly impairing tracking performance. Extensive experiments on five benchmark datasets (i.e., LasHeR, RGBT234, DepthTrack, VOT-RGBD2022, VisEvent) across three different modalities (RGB-T, RGB-D, and RGB-E) demonstrate that our attack substantially degrades state-of-the-art multi-modal trackers, exposing their susceptibility to adversarial threats.

Unsupervised anomaly detection in hyperspectral images (HSI), aiming to detect unknown targets from backgrounds, is challenging for earth surface monitoring. However, current studies are hindered by steep computational costs due to the high-dimensional property of HSI and dense sampling-based training paradigm, constraining their rapid deployment. Our key observation is that, during training, not all samples within the same homogeneous area are indispensable, whereas ingenious sampling can provide a powerful substitute for reducing costs. Motivated by this, we propose an Asymmetrical Consensus State Space Model (ACMamba) to significantly reduce computational costs without compromising accuracy. Specifically, we design an asymmetrical anomaly detection paradigm that utilizes region-level instances as an efficient alternative to dense pixel-level samples. In this paradigm, a low-cost Mamba-based module is introduced to discover global contextual attributes of regions that are essential for HSI reconstruction. Additionally, we develop a consensus learning strategy from the optimization perspective to simultaneously facilitate background reconstruction and anomaly compression, further alleviating the negative impact of anomaly reconstruction. Theoretical analysis and extensive experiments across eight benchmarks verify the superiority of ACMamba, demonstrating a faster speed and stronger performance over the state-of-the-art. Code is released at https://github.com/PURE-melo/ACMamba.

Signal Modulation Classification (SMC) is a fundamental technique in wireless communications. However, the prevalence of label noise in practical scenarios severely constrains the advancement of SMC technology. Existing SMC methods heavily rely on high-quality labeled data and often underutilize the inherent physical prior knowledge of signals. To address these issues, this article proposes a Physics-Aware Collaborative Framework with Prototype Consistency (PhyCo-PC), designed for noisy label environments and operating without requiring reliable labels. Firstly, the framework leverages co-teaching for noise identification and incorporates a collaborative consensus-guided module for prototype learning and pseudo-label generation. Secondly, it constructs physics-guided downstream decision module that fuses deep learning features with instantaneous physical signal characteristics to enhance decision robustness. Thirdly, a Domain Knowledge-guided Adaptive Sample Selection (DKASS) strategy is introduced. DKASS parameterizes the selection rate scheduling function, incorporates domain knowledge to constrain the search space, and utilizes automated search for optimization. This enables the model to adaptively determine the optimal training strategy for varying noise environments. Finally, experimental results demonstrate that PhyCo-PC significantly improves SMC classification performance under complex label noise scenarios on the RML2016.10a/04c datasets, exhibiting excellent robustness and significant advantages.

Change detection from synthetic aperture radar (SAR) imagery is critical in remote sensing research. Existing methods have made significant progress in the application of convolutional neural networks (CNNs) and attention mechanisms. However, traditional CNNs suffer the limitations in feature representation due to their depth and width constraints, and struggle to effectively capture complex interactions between image features. To address these issues, we propose a novel dynamic bilinear fusion network (DBFNet) for change detection in SAR imagery. First, to compensate for the lack of traditional convolutional representation capability, we design a dynamic shift convolution module that adaptively aggregates multiple convolution kernels and shifts pixels, enabling richer and more detailed features to be extracted. Second, a bilinear fusion module (BFM) is designed to generate the bilinear joint representation between parallel features by computing a matrix outer product of feature maps. The parallel features include both intraimage and interimage features, thereby effectively modeling the complex interactions and capturing the dependence relationship between spatiotemporal features. The experimental results on three real SAR datasets demonstrate the superior performance of DBFNet compared to existing state-of-the-art methods. The codes are available athttps://github.com/yizhilanmaodhh/DBFNet.

Zero-shot image captioning can harness the knowledge of pre-trained visual language models (VLMs) and language models (LMs) to generate captions for target domain images without paired sample training. Existing methods attempt to establish high-quality connections between visual and textual modalities in text-only pre-training tasks. These methods can be divided into two perspectives: sentence-level and entity-level. Although they achieve effective performance on some metrics, they suffer from hallucinations due to biased associations during training. In this paper, we propose a scene-relation-level pre-training task by considering relations as more valuable modal connection bridges. Based on this, we construct a novel Visual-Language Scene Relation Aware Captioner (SRACap), which expands the ability to predict scene relations while generating captions for images. In addition, SRACap possesses excellent cross-domain zero-shot generalization capability, which is driven by a well-designed scene reinforcement switching pipeline. We introduce a scene policy network to dynamically crop salient regions from images and feed them into a language model to generate captions. We integrate multiple expert CLIP models to form a mixture-of-rewards module (MoR) as a reward source, and deeply optimized SRACap through the policy gradient algorithm in the zero-shot inference stage. With the iteration of scene reinforcement switching, SRACap can gradually refine the generated caption details while maintaining high semantic consistency across visual-linguistic modalities. We conduct extensive experiments on multiple standard image captioning benchmarks, showing that SRACap can accurately understand scene structures and generate high-quality text, significantly outperforming other zero-shot inference methods.

Building fair deep neural networks (DNNs) is a crucial step towards achieving trustworthy artificial intelligence. Delving into deeper factors that affect the fairness of DNNs is paramount and serves as the foundation for mitigating model biases. However, current methods are limited in accurately predicting DNN biases, relying solely on the number of training samples and lacking more precise measurement tools. Here, we establish a geometric perspective for analyzing the fairness of DNNs, comprehensively exploring how DNNs internally shape the intrinsic geometric characteristics of datasets-the intrinsic dimensions (IDs) of perceptual manifolds, and the impact of IDs on the fairness of DNNs. Based on multiple findings, we propose Intrinsic Dimension Regularization (IDR), which enhances the fairness and performance of models by promoting the learning of concise and ID-balanced class perceptual manifolds. In various image recognition benchmark tests, IDR significantly mitigates model bias while improving its performance.

To address the challenges of long-tailed classification, researchers have proposed several approaches to reduce model bias, most of which assume that classes with few samples are weak classes. However, recent studies have shown that tail classes are not always hard to learn, and model bias has been observed on sample-balanced datasets, suggesting the existence of other factors that affect model bias. In this work, we first establish a geometric perspective for analyzing model fairness and then systematically propose a series of geometric measurements for perceptual manifolds in deep neural networks. Subsequently, we comprehensively explore the effect of the geometric characteristics of perceptual manifolds on classification difficulty and how learning shapes the geometric characteristics of perceptual manifolds. An unanticipated finding is that the correlation between the class accuracy and the separation degree of perceptual manifolds gradually decreases during training, while the negative correlation with the curvature gradually increases, implying that curvature imbalance leads to model bias. We thoroughly validate this finding across multiple networks and datasets, providing a solid experimental foundation for future research. We also investigate the convergence consistency between the loss function and curvature imbalance, demonstrating the lack of curvature constraints in existing optimization objectives. Building upon these observations, we propose curvature regularization to facilitate the model to learn curvature-balanced and flatter perceptual manifolds. Evaluations on multiple long-tailed and non-long-tailed datasets show the excellent performance and exciting generality of our approach, especially in achieving significant performance improvements based on current state-of-the-art techniques. Our work opens up a geometric analysis perspective on model bias and reminds researchers to pay attention to model bias on non-long-tailed and even sample-balanced datasets.

This paper conducted a thorough investigation into the primary difficulty of the cross-channel automatic modulation classification (AMC) task by examining data distribution and feature space of different channel conditions. We concluded that the disruption of the target channel feature space structure breakdown the mapping relationship across channels, serving as the main contributor to model performance degradation. Based on the above conclusion, in order to improve the performance of cross-channel AMC, we introduce the Auxiliary Feature Learning-Guided Network (AFLNet). This network improves the structure of the target feature space through two uniquely designed tasks and facilitates efficient cross-domain alignment via a collaborative alignment mechanism. Specifically, AFLNet integrates similarity-based and confidence-based auxiliary feature learning tasks to enhance the discriminability of the target feature space and maintain the correspondence of category structures across different channels, thereby reducing the difficulty of feature alignment. The collaborative alignment mechanism combines adversarial training-based and self-training-based feature alignment methods, leveraging their mutually reinforcing effect and complementary strengths in global alignment and class-level alignment to enhance overall alignment performance. We carried out extensive experiments across four scenarios characterized by substantial channel variations, verifying that AFLNet achieves state-of-the-art with accuracy improvement of up to 9.71%.

As a useful remote sensing (RS) scene interpretation technique, multi-label RS scene classification (RSSC) always attracts researchers’ attention and plays an important role in the RS community. To assign multiple semantic labels to a single RS image according to its complex contents, the existing methods focus on learning the valuable visual features and mining the latent semantic relationships from the RS images. This is a feasible and helpful solution. However, they are often associated with high computational costs due to the widespread use of Transformers. To alleviate this problem, we propose a Mamba-based efficient network based on the newly emerged state space model called MLMamba. In addition to the basic feature extractor (convolutional neural network and language model) and classifier (multiple perceptrons), MLMamba consists of two key components: a pyramid Mamba and a feature-guided semantic modeling (FGSM) Mamba. Pyramid Mamba uses multi-scale scanning to establish global relationships within and across different scales, improving MLMamba’s ability to explore RS images. Under the guidance of the obtained visual features, FGSM Mamba establishes associations between different land covers. Combining these two components can deeply mine local features, multi-scale information, and long-range dependencies from RS images and build semantic relationships between different surface covers. These superiorities guarantee that MLMamba can fully understand the complex contents within RS images and accurately determine which categories exist. Furthermore, the simple and effective structure and linear computational complexity of the state space model ensure that pyramid Mamba and FGSM Mamba will not impose too much computational burden on MLMamba. Extensive experiments counted on three benchmark multi-label RSSC data sets validate the effectiveness of MLMamba. The positive results demonstrate that MLMamba achieves state-of-the-art performance, surpassing existing methods in accuracy, model size, and computational efficiency. Our source codes are available athttps://github.com/TangXu-Group/ multilabelRSSC/tree/main/MLMamba.

Hyperspectral image (HSI) provides detailed spectral and spatial information, essential for precise earth observation and various applications. Deep learning has advanced HSI classification, but the scarcity of labeled data and large model parameters necessitate semi-supervised methods to enhance performance and generalization. In this paper, we propose a novel semi-supervised framework dubbed Knowledge-Aware Geometric Contourlet Semantic Learning (KGCSL), aiming to achieve high-precision HSI classification with limited samples leveraging geometric and semantic knowledge. Specifically, to fully leverage geometric knowledge, KGCSL incorporates multi-scale and multi-directional representations of the contourlet transform within the neural network, enhancing the robustness of feature extraction and interpretability. Furthermore, to fully utilize semantic knowledge, an entropy-weighted prototype loss function is designed that exploits the attribute relationships between labeled and unlabeled samples to guide the optimization of unlabeled samples, promoting comprehensive semantic learning. Comprehensive evaluations of the proposed KGCSL framework on three public HSI datasets show that it outperforms existing state-of-the-art HSI classification methods and exhibits excellent generalization capabilities in limited-sample scenarios. The source code is available athttps://github.com/ShirlySmile/KGCSL.

Few-Shot Class-Incremental Learning (FSCIL) faces a huge stability-plasticity challenge due to continuously learning knowledge from new classes with a small number of training samples without forgetting the knowledge of previously seen old classes. To alleviate this challenge, we propose a novel method called Prompt-based Concept Learning (PCL) for FSCIL, which generalizes conceptual knowledge learned from old classes to new classes by simulating human learning capabilities. In our PCL, in the base session, we simultaneously learn common basic concepts from the training data and the class-concept weight of each class in a prompt learning manner, and in each incremental session, class-concept weights between new classes and previously learned basic concepts are learned to achieve incremental learning. Furthermore, in order to avoid catastrophic forgetting, we propose a distribution estimation module to retain feature distributions of previously seen classes and a data replay module to randomly sample features of previously seen classes in incremental sessions. We verify the effectiveness of our PCL on widely used benchmarks, such as miniImageNet, CIFAR-100, and CUB-200. Experimental results show that our PCL achieves competitive results compared with other state-of-the-art methods, especially we achieve an average accuracy of 94.02% across all sessions on the miniImageNet benchmark.

Evolutionary computation (EC)-based neural architecture search (NAS) has achieved remarkable performance in the automatic design of neural architectures. However, the high computational cost associated with evaluating searched architectures poses a challenge for these methods, and a fixed form of learning rate (LR) schedule means greater information loss on diverse searched architectures. This paper introduces an efficient EC-based NAS method to solve these problems via an innovative meta-learning framework. Specifically, a meta-learning-rate (Meta-LR) scheme is used through pretraining to obtain a suitable LR schedule, which guides the training process with lower information loss when evaluating each individual. An adaptive surrogate model is designed through an adaptive threshold to select the potential architectures in a few epochs and then evaluate the potential architectures with complete epochs. Additionally, a periodic mutation operator is proposed to increase the diversity of the population, which enhances the generalizability and robustness. Experiments on CIFAR-10, CIFAR-100, and ImageNet1K datasets demonstrate that the proposed method achieves high performance comparable to that of many state-of-the-art peer methods, with lower computational cost and greater robustness.

Unsupervised hyperspectral change detection (UHCD), detecting subtle changes between bi-temporal images without manual annotations, is an essential but challenging task in the earth observation community. The current modus operandi often performs it in a feature comparison manner, which is limited by variations in imaging conditions. We observe that fully supervised paradigms using limited annotations are capable of overcoming this challenge. Based on this, we introduce a novel Observational Learning Paradigm (OraL) for UHCD by mimicking fully supervised paradigms. OraL comprises two sequential stages: Observation, which designs a spatial-temporal observation strategy (STO) that records the learning consistency of pixels under different training steps and views, to obtain reliable pseudo-labels. Reproduction, which retrains the model with these pseudo-labels and introduces a distribution-aware spectral learning strategy (DSL) to adaptively increase their learning difficulty according to spectral distributions, enhancing the robustness and generalization of the model. Extensive experiments on several public hyperspectral image datasets demonstrate its state-of-the-art performance and pluggability for previous unsupervised methods. Code will be made available.

The diversity of contextual information is of great importance for accurate semantic segmentation. However, most methods focus on single spatial contextual information, which results in an overlap of the semantic content of categories and a loss of contour information of objects. In this article, we propose a novel contour knowledge-aware perception learning network (CKPL-Net) to capture diverse contextual information by space-category aggregation module (SCAM) and contour-aware calibration module (CACM). First, SCAM is introduced to enhance intraclass consistency and interclass differentiation of features. By integrating space-aware and category-aware attention, SCAM reduces the redundancy of features from a categorical perspective while maintaining spatial correlation of pixels, substantially avoiding the overlap of the semantic content in categories. Second, CACM is designed to maintain the integrity of objects by perceiving contour contextual information. It develops a novel contour-aware knowledge and adaptively transforms the grid structure of convolutions for boundary pixels, which effectively calibrates the representation of features near boundaries. Finally, the quantitative and qualitative analyses on the three public datasets: ISPRS Potsdam dataset, ISPRS Vaihingen dataset, and WHDLD dataset, demonstrate that the proposed CKPL-Net achieves superior performance compared with prevalent methods, which indicates diverse contextual information is beneficial for accurate segmentation.

This paper studies the domain generalized remote sensing semantic segmentation (RSSS), aiming to generalize a model trained only on the source domain to unseen domains. Existing methods in computer vision treat style information as domain characteristics to achieve domain-agnostic learning. Nevertheless, their generalizability to RSSS remains constrained, due to the incomplete consideration of domain characteristics. We argue that remote sensing scenes have layout differences beyond just style. Considering this, we devise a joint style and layout synthesizing framework, enabling the model to jointly learn out-of-domain samples synthesized from these two perspectives. For style, we estimate the variant intensities of per-class representations affected by domain shift and randomly sample within this modeled scope to reasonably expand the boundaries of style-carrying feature statistics. For layout, we explore potential scenes with diverse layouts in the source domain and propose granularity-fixed and granularity-learnable masks to perturb layouts, forcing the model to learn characteristics of objects rather than variable positions. The mask is designed to learn more context-robust representations by discovering difficult-to-recognize perturbation directions. Subsequently, we impose gradient angle constraints between the samples synthesized using the two ways to correct conflicting optimization directions. Extensive experiments demonstrate the superior generalization ability of our method over existing methods.

In building extraction, collecting pixel-level annotations required by fully supervised methods is extremely costly. Image-level weakly supervised methods based on class activation maps (CAMs) effectively reduce the cost and have shown promising progress. However, generating high-quality CAMs remains challenging due to the supervision gap between classification and segmentation tasks. Specifically, image-level supervision causes CAMs to activate only the most discriminative regions, which compromises the integrity of CAMs. Meanwhile, the absence of pixel-level supervision leads to a depletion of spatial information, resulting in imprecise boundaries. In this study, we propose a novel image-level weakly supervised building extraction method based on view consistency, named ViewCAM, to generate high-quality CAMs. The view transformation module is designed to apply view transformations to remote sensing images and the high-dimensional features. Additionally, a feature affinity enhancement module (FAEM) is proposed to capture positional relationships between pixels and low-level features, such as edges and textures, improving boundary fineness. We integrate these two modules into a classification network and incorporate pixel-level supervision using view-consistency constraints. The entire network is then trained in an end-to-end manner, leading to improved integrity and boundary fineness of the seeds generated from CAMs. To verify the effectiveness and robustness of ViewCAM, we conduct experiments on two representative datasets, and the results demonstrate that our proposed method achieves superior CAM integrity and boundary fineness, outperforming state-of-the-art methods.

Remote sensing scene classification plays a vital role in identifying Earth’s surface features. However, traditional methods necessitate vast quantities of labeled training data, which significantly limits their applicability when labeled samples are scarce. To overcome this limitation, remote sensing fewshot classification (RSFSC) methods have arisen as a promising approach to maintaining classification accuracy with minimal data. In this paper, we present DBA-RMCL (Refined Metric Contrastive Learning Framework with Dual-Branch Attention), a novel framework that tackles two fundamental challenges in RSFSC: intra-class diversity and inter-class similarity. Our framework employs a Dual-Branch architecture consisting of Way and Shot branches to reduce intra-class variance while enhancing feature consistency and inter-class separability. The architecture incorporates feature fusion between branches to optimize feature learning and classification performance. We further enhance feature representations through our RSCLR (Refined Siamese Contrastive Learning Representation) module, which implements self-supervised contrastive learning. The framework also includes a CAP-Head(Combine Attention-based Prototypical Head) module with a compositional attention computation unit to model complex relationships between query sets and prototypes. Experimental results on six publicly available datasets demonstrate that DBA-RMCL achieves state-of-the-art performance compared to existing FSC methods. For 5-way 1-shot tasks, our framework attains accuracies of 77.15±0.34% on NWPU-RESISC45 and 87.33±0.22% on WHU-RS19. Additionally, we established the first benchmark for few-shot scene classification, MEET-FSL, using the MEET dataset. These outcomes not only underscore the efficacy of our Dual-Branch architecture but also represent a significant advancement in the field of few-shot learning for remote sensing classification.

Hyperspectral unmixing is an important issue in remote sensing field for identifying the constituent materials and their respective distributions in a scene. In recent years, a large number of data-driven HSI unmixing methods have been proposed. These methods reformulate the unmixing task as a transformation from spectral pixels to abundance. Diffusion models, a recently emerging class of generative models, have shown potential in converting noise into other meaningful representations. However, applying diffusion models to unsupervised hyperspectral unmixing faces the following challenges: stochasticity of the generative process, neglect of the quality and prior knowledge of the estimation of original data, and difficulty in solving the learning optimization problem. To address these issues, we propose a novel diffusion-based unsupervised hyperspectral unmixing framework, enabling the transformation from noise to abundance maps. First, we embed the diffusion process into an autoencoder architecture and introduce a denoising guidance generator to extract robust guidance from hyperspectral images, directing the reverse diffusion process to produce spatially aligned and physically plausible abundances. Second, we design an abundance evaluation module that supervises the original data, i.e., abundances in this context, using decoder-based reconstruction evaluation and unmixing prior knowledge. Additionally, we adopt a stage-wise freezing training strategy to progressively stabilize the model and address the learning optimization problem. Experimental results on both simulated and real-world datasets demonstrate that the proposed method effectively addresses the challenges of unsupervised hyperspectral unmixing, producing accurate, physically meaningful, and robust unmixing results.

With remote sensing (RS) images’ resolution increasing, a single scene label cannot adequately represent RS scenes’ contents. Therefore, multilabel RS scene classification (MLRSSC) is gradually attracting the researchers’ attention. Many methods have been proposed recently, and most use deep features or semantic connections to complete MLRSSC. However, they ignore the combination of these two aspects. In addition, the high interclass similarity and low intraclass similarity of RS images limit the robustness of these methods. In this article, we propose a semantic-assisted feature integration network (SFIN) to overcome the above limitations. It contains a dual-scale feature extractor module (DFEM), a local semantic enhance module (LSEM), a cross-scale interactive attention module (CIAM), and a classifier module (CM). DFEM utilizes the convolutional neural networks (CNNs) to extract multiscale features from RS images. LSEM extracts semantic information and establishes their relationships at different scales. CIAM enhances the feature representation by interacting with the clues across different scales. CM completes the prediction of classification (CLA) results. Integrating them into an end-to-end framework, SFIN can discover the diverse and complex land covers hidden in RS images. Furthermore, to ensure the accuracy of explored semantics and enhance the SFIN’s feature extraction ability, we design a semantic supervision (SS) loss and a semantic-based contrastive learning (SB-CL) loss. They are in charge of the correctness and discrimination of the mined semantics. Along with the typical CLA loss, SFIN can be adequately trained. Extensive experiments have been conducted on four MLRSSC datasets, and the positive results demonstrate that SFIN outperforms many existing methods in MLRSSC tasks. Our source codes are available at:https://github.com/TangXu-Group/multilabelRSSC/tree/main/SFIN.

Mainstreamed weakly supervised road extractors rely on highly confident pseudo-labels propagated from scribbles, and their performance often degrades gradually as the image scenes tend to vary. We argue that such degradation is due to the poor model’s invariance to scenes with different complexities, whereas existing solutions to this problem are commonly based on crafted priors that cannot be derived from scribbles. To eliminate the reliance on such priors, we propose a novel structure-aware mixup and invariance learning framework (SA-MixNet) for weakly supervised road extraction that improves the model invariance in a data-driven manner. Specifically, we design a structure-aware mixup (SA-Mix) scheme to paste road regions from one image onto another to create an image scene with increased complexity while preserving the road’s structural integrity. Then, an invariance regularization is imposed on the predictions of constructed and origin images to minimize their conflicts, which thus forces the model to behave consistently in various scenes. Moreover, a discriminator-based regularization is designed to enhance connectivity while preserving the structure of roads. Combining these designs, our framework demonstrates superior performance on the DeepGlobe, Wuhan, and Massachusetts datasets, outperforming the state-of-the-art techniques by 1.47%, 2.12%, and 4.09%, respectively, in IoU metrics, and showing its potential as a plug-and-play solution. Our source code is available athttps://github.com/xdu-jjgs.

This paper proposes a novel interpretable framework for fine-grained aircraft classification in high-stakes remote sensing applications. Our approach addresses three key challenges: small inter-class variance, large intra-class variance, and the need for model interpretability. Specifically, our framework is built on the Swin Transformer (SwinT) backbone and includes three main modules. First, we present the Dynamic Attention Fusion Module (DAFM), which adaptively fuses multi-stage attention maps from the SwinT backbone. By leveraging a dispersion-based weighting mechanism, DAFM balances the contributions of coarse and fine-grained features, capturing both global structures and localized details. Second, we propose the Adaptive Image Pair Interaction Module (AIPI), which dynamically adjusts feature interaction strategies based on intra-class and inter-class similarity, effectively enhancing informative regions and improving robustness. To further optimize discriminative power, we incorporate an AIPI loss function that enforces intra-class consistency and inter-class separability. Finally, we develop a Binary Neural Tree Module (BNTM) to hierarchically select and propagate informative image patches, enhancing both feature refinement and interpretability through explicit path-based decision-making. Extensive experiments on benchmark datasets demonstrate that our framework significantly improves classification accuracy and interpretability, making it well-suited for applications requiring transparent and reliable decision-making. The Code can be found at https://github.com/StarmanGzx/BNTM.

Time-series Synthetic Aperture Radar (SAR) offers significant potential for crop mapping due to its all-weather, weather-independent imaging capabilities. Existing crop mapping methods with time-series SAR data have achieved good performance. However, these methods often ignore the phenological diversity of the same crop in time-series data, and their limited performance significantly constrains their potential for application in large-scale crop mapping. To address these issues, this paper proposes a prototype-based time-series remote sensing crop mapping framework called Multi-Distribution time-series Prototype Learning (MDPL). The framework aims to learn multiple time-series prototypes for the same crop for various phenological distributions of crops, effectively capturing the complex and varied phenological characteristics of crops. Secondly, a phenological-invariant feature learning module is proposed to enhance the model’s generalization capability for large-scale crop mapping. Additionally, a new temporal metric is proposed to capture phenological differences. Experimental results on three benchmark datasets have demonstrated the effectiveness and superiority of MDPL compared with state-of-the-art time-series SAR crop mapping methods.

In the field of multimodal remote sensing image (MRSI) classification, self-supervised learning (SSL) algorithms have demonstrated significant advantages, particularly in scenarios with limited labeled samples. Existing SSL methods typically use auxiliary tasks within either contrastive or generative frameworks, focusing on discriminative or structural information separately. In this article, we propose a novel hybrid SSL paradigm, discriminative autoencoder with Gaussian selective patch (DAE-GSP) for MRSI classification. The DAE framework integrates contrastive learning with the masked image modeling (MIM) technique, allowing for simultaneous learning of structural information and discriminative representations from images. Furthermore, a cross-attention-based data-level fusion strategy is introduced during pretraining stage to enhance intermodal interactions, thereby improving the effectiveness of modality fusion. In addition, we propose a novel Gaussian selective patch (GSP) strategy, addressing the limitations of traditional square patch selection methods. Combined with self-supervised auxiliary tasks, this strategy facilitates the improved integration of multiple modalities and encourages the model to capture essential semantic information. Extensive experiments conducted on three public datasets (Houston2013, Augsburg, and Berlin) demonstrate the effectiveness of the proposed approach. With only ten labeled training samples per class, the proposed method achieves overall accuracy (OA) of 90.15%, 82.64%, and 71.03% on the Houston2013, Augsburg, and Berlin datasets, respectively, indicating improvements of 1.31%, 1.22%, and 1.48% over state-of-the-art methods.

Remote sensing image-text retrieval (RSITR) is critical for applications, including environmental monitoring and disaster management. The main challenge in this field is that the multi-scale feature of remote sensing images and the semantic differences of professional texts make it difficult to achieve accurate alignment. Existing coarse-grained methods struggle to address the inherent difference between images and text. In light of this, we propose the Fine-Grained Visual-Language Alignment (FGVLA) method. Our FGVLA employs a hybrid loss function that combines coarse-grained contrastive and triplet loss with novel fine-grained loss. Fine-grained loss includes spatial mask loss and fine-grained contrastive loss to enhance semantic alignment. The method also introduces an inference process that works cooperatively with fine-grained loss to explicitly align image patches with textual nouns. Extensive experiments on RSICD, RSITMD, and UCM-Caption datasets demonstrate that FGVLA outperforms existing methods, achieving superior retrieval performance. The code of our FGVLA has been released at https://github.com/Ji-Haoyang/FGVLA.

With the advancement of remote sensing (RS) technology and satellite observation, the task of fusing multisource data, such as multispectral (MS) and panchromatic (PAN) images, has become increasingly important. However, image fusion involving certain semantic differences can hinder the model’s ability to learn effective feature mappings. To reconstruct richer and more consistent features during fusion, we propose an adaptive dual-supervised cross-deep dependency network (ADCD-Net), which consists of two training stages. Stage I uses a semantic perceptual self-supervision strategy (SPS) to learn deep features across different modalities, thereby reducing semantic differences while mining its own non-singular features. Stage II uses the deep temporal Mamba module (DTM-Module) to interactively learn the output of each network layers, which are able to take part in the deep feature reinforcement and improve the classification performance of semantic information. Finally, to eliminate channel redundancy during the two-stage network training process while enhancing spatial location memory and feature discrimination in the 2-D features, we propose a deformable interactive attention module (DIA-Module) to further bolster feature representation capabilities. Additionally, we conduct comparative and transfer experiments on multiple RS datasets, achieving outstanding classification results. Our code is available athttps://github.com/ChenC1027/ADCD-Net.

In recent years, with the rapid advancements and widespread application of satellite photography technology, it has become increasingly possible to obtain high-quality panchromatic (PAN) and multispectral (MS) data, which has provided new opportunities and challenges for multisource information fusion and classification research. Remote sensing data have the characteristics of small interclass differences and large intraclass differences, which easily leads to category confusion in network learning. In addition, how to fully tap the advantages of multisource data, better align multisource features, improve classification accuracy, and achieve collaborative classification are key issues that need to be solved urgently. In this article, a dual-path prototype feature decoupling alignment network (DPFDA-Net) is designed to solve the above issues. The network consists of two components: a prototype feature embedding (PFE) module and a feature alignment module (FAM) based on prototype decoupling. In the feature extraction stage, the PFE module uses the prototype concept to learn the discriminative prototype features of each category of the dual-source data separately, making the boundaries between categories more obvious. The FAM operates at the dual-source prototype feature level and achieves feature alignment by decoupling single-source prototype features and performing feature transformation to supplement the missing information of another data source. Finally, we use the aligned features for classification. The results of the experiment demonstrate that our approach has made significant progress in improving classification precision. The code is available athttps://github.com/Xidian-AIGroup190726/DPFDANet.

Remote sensing (RS) scene classification (RSSC) is a prominent research topic in the RS community. Multilevel feature fusion is an important way of addressing RS scene classification, and many methods have been proposed in recent years. Although they succeed, current methods can still be improved, particularly in distinguishing the contributions of different multilevel features and fully and effectively fusing them. To address the above issues and fully exploit the potential of multilevel features for RS scene classification tasks, we propose a new model named multiscale sparse cross-attention network (MSCN). It not only focuses on the effectiveness of feature learning but also emphasizes the rationality of feature fusion. In detail, MSCN first extracts multilevel features using a pre-trained ResNet50. Also, these features are divided into high- and low-level features according to the clues they involved. Then, a multiscale sparse cross-attention (MSC) module is developed to cross-fuse the high-level feature with various low-level features, thereby effectively mining helpful information from multilevel features. In the fusion process, MSC not only explores the multiscale messages in RS scenes but also mitigates the negative impact of irrelevant information by employing sparse operations. Third, a group convolutional block attention module (CBAM) enhancer (GCE) is presented to enhance the representation of classification features. GCE detects local salient information within classification features using grouped CBAM and further enhances crucial details by readjusting the CBAM attention weights. This way, the classification features’ discrimination can be improved. We conducted extensive experiments on three public RS scene classification datasets. The exceptional experimental results indicate that our proposed MSCN achieves superior classification accuracy, surpassing many existing methods. Our source codes are available at https://github.com/TangXu-Group/Remote-Sensing-Images-Classification/tree/main/MSCN.

Referring Remote Sensing Image Segmentation (RRSIS) task aims to generate segmentation masks for target objects based on language descriptions. It requires precise localization while distinguishing between visually similar yet semantically distinct objects. Fusing vision-language features only during extraction causes information loss and semantic forgetting in the decoder, harming similar target distinction. Additionally, high-resolution remote sensing images present challenges, including complex backgrounds, diverse object scales, and intricate boundaries, limiting the effectiveness of previous methods. To address these issues, we propose the Long-term Semantic-guidance ConvFormer (LSCF) Network. First, we fuse multi-receptive-field local features extracted by the Multi-scale CoordConv (MCC) module with language-aware global features from the Cross-modal Attention (CA) module to obtain multi-modal representations. Second, the Sampling Attention (SA) module enables fine-grained vision context alignment under semantic guidance. Finally, the Global Language Fusion (GLF) module is incorporated in the decoder to maintain long-term vision-language alignment and mitigate semantic degradation. Experimental validation on the RefSegRS, RRSIS-D, and RISBench datasets demonstrates that LSCF achieves oIoU scores of 83.27%, 77.42%, and 74.88%, and mIoU scores of 77.44%, 64.25%, and 68.53%, respectively. On RefSegRS, LSCF surpasses the SOTA method FIANet by 5.53% (oIoU) and 9.58% (mIoU), while delivering competitive performance on RRSIS-D and RISBench. Code and experimental configurations will be released.

This study introduces a novel cross-modal spatial-spectral interaction Mamba (CMS2I-Mamba) for remote sensing image fusion classification. Unlike convolution-based models focusing on local details and Transformer-based models with high computational complexity, CMS2I-Mamba efficiently models global long-range dependencies in a linear complexity manner. First, multispectral (MS) and panchromatic (PAN) images each have unique advantages in the spectral and spatial attributes. Given this, this paper innovatively designs the multi-path selective-scan mechanism (MPS2M), which applies different path scanning strategies to deeply capture the global features from both spectral and spatial dimensions, aiming to enhance the robustness and complementarity of spatial-spectral features. Secondly, to overcome the characterization differences between images acquired by different sensors, this paper further introduces the channel interaction alignment module (CIAM). This module employs efficient former-last and oddeven channel interaction strategies to achieve precise semantic alignment of deep features between modalities. Finally, to leverage the shared fusion features to guide the unique singular features, this paper proposes a semantic-aware calibration module (SACM), which accurately constraints and calibrates the same semantic information in deep features. This not only enhances the model’s ability to understand scene semantics, but also promotes the deep fusion and utilization of information between different modalities. Through experimental verification on multiple datasets, the CMS2I-Mamba proposed in this paper shows excellent recognition performance and computational efficiency (parameter quantity and running speed) in fusion classification tasks. The code for CMS2I-Mamba is available at: https://github.com/ru-willow/CMSI-Mamba.

Pan-sharpening refers to fusing remote sensing multispectral (MS) and panchromatic (PAN) images to generate high-resolution multispectral (HR-MS) images. Recent advancements in deep learning-based pan-sharpening techniques have shown promising results. However, they face the following two issues. On one hand, there is a modality gap between MS and PAN images. Directly fusing them can lead to spectral and spatial distortions. On the other hand, the fusion process is prone to information loss, which can lead to image blurriness. To tackle these issues, we develop a Transformer-based model: FAFormer, which incorporates frequency analysis and focuses on the correlation and specificity of the PAN and MS images. Focusing on correlation can reduce the spectral and spatial distortions while focusing on specificity can reflect the specific information from MS and PAN images in the fusion result. We utilize the Discrete Wavelet Transform (DWT) to obtain the correlate and specific features. We introduce bijective functions based on the Transformer to design an Integrated Attention Block (IAB). As a critical component of the model, it effectively utilizes the correlation and specificity of the two images. In designing the model’s overall framework, we employ a Correlative Feature Attention Module (CFAM) to leverage the correlation between MS and PAN. We utilize a Specific Feature Attention Module (SFAM) to integrate specific information into fused features gradually. Experimental results show that our method improves pan-sharpening performance and has practical value. Codes are available at https://github.com/Xidian-AIGroup190726/FAFormer.

Prominent domain discrepancies in Remote Sensing Images (RSIs), such as sensor types, geographical patterns, and land usage, significantly hinder the research and practical applications of cross-scene land classification. Unsupervised Domain Adaptation (UDA) fully exploits domain invariance between labelled source and unlabelled target domain, which alleviates the challenge of inaccurate land classification due to lack of labels in RSIs. However, most existing UDA methods for RSIs semantic segmentation are insufficient in exploring cross-domain features and have difficulty in modelling fine-grained domain-invariant features between inter-class. In this paper, we propose a novel self-training UDA method named Hierarchical Prototype Learning (HPL), which learns the inherent domain-wise consistency and class-wise invariance through progressive exploration of the prototypy, significantly alleviating the ambiguity in uncertain regions. HPL mainly consists of Domain-wise Progressive Prototype Interaction (DPPI) and Class-wise Dynamic Prototype Collaboration (CDPC). DPPI and CDPC specialize in hierarchically building prototype interaction architectures tailored to domain-wise alignment and class-wise calibration, respectively. This design not only mitigates the sensitivity to cross-domain scenarios but also allows for precise correction of uncertain regions. Furthermore, CDPC exhibits the capacity for pixel-level category restoration and promotes the correct and fine-grained updating of pseudo-labels. Extensive Experiments on two public datasets and a private self-build dataset demonstrate the superiority of HPL over other state-of-the-art methods for UDA semantic segmentation of RSIs.

Domain-adaptive Remote Sensing Images (RSIs) semantic segmentation mitigates the overfitting problem that affects the effectiveness of segmentation, which results from the scarcity of high-quality labels and the cross-domain styles of ground objects. The effectiveness of domain adaptive segmentation remains suboptimal in complex scenarios due to inadequate exploitation of latent geographic knowledge. Consequently, inter-class ambiguity and boundary agnostic are further exacerbated under cross-domain transfer scenarios. To address this issue, we first devise a deep geospatio-semantic guided network named DSSAL, which comprehensively investigates the potential spatial relationship and semantic correlation between classes of RSIs by geospatial aware interaction and geosemantic aware interaction, respectively. To mitigate class-wise cognitive deviation in the unlabeled domain, DSSAL-DA is developed to further enhance the segmentation effect with the spatio-semantic domain alignment module in manifold cross domain tasks. Furthermore, a pseudo-labels consistency filter is developed for DSSAL-DA to ensure reliability in self-training through cross-view consistency verification. Extensive experiments on two public datasets and a private dataset demonstrate the superiority of DSSAL and DSSAL-DA over the state-of-the-art methods for UDA semantic segmentation of RSIs.

Deep learning methods has been widely applied to synthetic aperture radar (SAR) land cover classification. The complexity of SAR data and the limited availability of labeled samples greatly constrain the feature learning and the generalization ability of the model. Inspired by the excellent generative performance of Denoising Diffusion Probabilistic Models(DDPM) on complex data distributions, a self-supervised learning framework based on contrast-diffusion models (CDM) is proposed to expand the applicability to multiple broad scenarios with complex and varying imaging conditions under limited annotated data conditions. Specially, The proposed framework consists of the upstream CDM pre-training on all unlabeled samples and the downstream land cover classification with few labeled samples in each test scene. Concretely, in the upstream task, the features are captured through the generative learning of the DDPM. Following this, the Dimensionality Reduction and Resolution Expansion (DRRE) module is designed and embedded to reduce feature redundancy and align the feature granularity between layers and the input image. Finally, contrastive learning is employed to enforce semantic feature consistency across different steps. In the downstream task, the feature in pre-trained CDM is efficiently delivered in a single-step reverse diffusion process and then fine-tuned with few labeled samples from each test scene and finally output the predictions. Compared with several supervised and self-supervised methods, the proposed framework achieves superior classification and generalization performance on multiple broad scenes with complex and varying imaging conditions. For example, based on the average results from six test scenes, CDM shows improvements in overall accuracy (OA) of 41.42%, 34.40%, 7.70% ,8.53% and 9.60% compared to Deeplabv3+, CCNR, Segformer, MAE and DDPM respectively. The code is available at https://github.com/gosling123456/CDM.git.

Land cover classification (LCC) based on remote sensing image segmentation is a prominent task of remote sensing data interpretation. The commonly used optical data is susceptible to the weather, so it has the potential to utilize complementary features from the supplementary synthetic aperture radar (SAR) data to enhance segmentation performance. However, current multi-modal segmentation methods focus on the deep fusion of features, which usually ignores the significance of structural consistency information. In order to make use of the mutual correction and information exchange between multi-modal data, we propose DCIFNet, a dual-stream correction-interaction-fusion multi-modal LCC network. Specifically, we design a differential feature correction and enhancement module (DF-CEM) that leverages bidirectional differential features to correct multi-modal features. In addition, for corrected feature pairs, we deploy a parallel attention interaction module (PAIM) to focus on the pixel-level feature correlation and achieve effective information exchange in both channel and spatial dimensions. Through the expert fusion module (EFM), DCIFNet leverages the gate network to attain a flexible and compact feature fusion between multi-modal features. Experimental results show that our method achieves a superior performance compared with other multi-modal fusion segmentation methods on three optical-SAR datasets. The source code of DCIFNet is publicly available at https://gitee.com/asdwer2046/dcifnet.

With the exponential growth of satellite remote sensing data, land cover classification models must adapt continuously to new classes. However, conventional incremental learning methods face critical challenges: catastrophic forgetting degrades recognition of old classes, and the softmax function further suppresses old-class probabilities due to ”class crowding.” Existing distillation techniques also struggle to transfer features in irregular geospatial regions. To address these issues, we propose Soft Labels and Subregion Distillation (SLSRD). SLSRD mitigates class crowding by employing soft labels instead of hard labels, derived from a hybrid of softmax and sigmoid outputs that preserve richer probabilistic information. Concretely, the soft label is a convex combination of softmax- and sigmoid-based probabilities that preserves inter-class relations while relaxing over-confident exclusivity for newly introduced categories, and it supervises all pixels across stages. In parallel, a breadth-first search identifies subregions within each image, which are weighted by probability and size, and similarity between corresponding subregions of the old and new models is maximized. This dual strategy effectively transfers fine-grained knowledge and overcomes the limitations of conventional distillation methods, particularly for large-scale remote sensing imagery. Experiments on three benchmark datasets-Vaihingen, GID, and FBP-demonstrate that SLSRD outperforms traditional methods, significantly improving incremental land cover classification.

Hyperspectral image (HSI) classification is crucial in the remote sensing (RS) community. In recent years, Transformers have been popular in this field due to their global information modeling capabilities. However, the quadratic complexity limits their performance under limited computational resources. Fortunately, a selective structured state space model named Mamba emerges. Like Transformer, it is good at modeling the long-distance relationships hidden in the pending data. Unlike Transformer, its complexity remains at a linear level. Therefore, a growing number of studies have been proposed to explore the usefulness of Mamba in HSI classification. Nevertheless, most of them only apply Mamba to HSIs directly but do not consider the inherent characteristics of HSIs properly. To exploit the potential of Mamba in HSI classification deeply, this paper presents a new spatial-spectral complementary Mamba with a spatial spiral scan named SpiralMamba. It mainly encloses three main components: a spatial Mamba encoder (SpaME), a spectral Mamba encoder (SpeME), and a spatial-spectral complementary fusion module (SSCFM). SpaME focuses on understanding the spatial context within HSIs. To this end, instead of the common scanning, a spatial spiral scan strategy is introduced to address the sequence transformation of non-causal HSIs. SpeME aims to comprehensively extract valuable spectral features from HSIs. To achieve this goal, besides developing a spectral bidirectional scan strategy, a multilayer convolution (MLC) is also incorporated to capture local variations within spectral tokens. SSCFM concentrates on building the complex connections between spatial and spectral features and fusing them. For this purpose, a relationship learning block (RLB) and a threshold enhancement mechanism (TEM) are developed. Positive experimental results counted on three public HSI datasets demonstrate the effectiveness of SpiralMamba. Our source codes are available at https://github.com/TangXu-Group/Hyperspectral-Images-Classification/tree/main/SpiralMamba.

In recent years, the number and ways of acquiring panchromatic images (PAN) and multispectral images (MS) have increased, and manual labeling costs have also increased. Processing these data efficiently has become a challenge. In this paper, we propose a sample-level improved cross-source contrastive learning method for PAN and MS joint classification (SLCL), which aims to provide a self-supervised pre-training model using unlabeled samples for downstream joint classification using a small quantity of labeled samples. First, we propose a sample weighting and screening (SWS) strategy, which enables the model to learn inter- and intra-source sample representations, while balancing the interference from false samples so that the model learns true samples. It solves the problems of homologous similar features embedded far away and false negative samples bringing the wrong learning direction, which exist in existing contrastive learning methods. In addition, we design a hard sample learning (HSL) module for the problem of mining and optimization of hard samples. The module efficiently mines hard samples and uses a new loss function to make the model more focused on hard sample optimization. It further improves the accuracy of pre-training models for downstream tasks. Our method performs best on multiple datasets, and it is experimentally validated and analyzed. The code is available at: https://github.com/Xidian-AIGroup190726/SLCL.

Remote sensing image change detection plays a critical role in applications like video surveillance and geographic information systems. However, existing binary and semantic change detection methods often rely solely on visual information, neglecting language information, which limits interpretability and the ability to provide specific change details. This work proposes the Change Knowledge-Guided Vision-Language Remote Sensing Change Detection (CKCD) method to address these limitations. By introducing change knowledge as language information, CKCD enhances semantic understanding and change detail representation. A Cross-Modal Affinity (CMA) module is designed to effectively fuse visual and textual features, improving information complementarity and fusion coherence. CKCD further enhances data utilization efficiency by merging change area detection and change category information into a single output through endto- end learning. This design reduces redundant data representations and simplifies the detection process, leading to a more compact and efficient use of the input data without requiring additional branches or multiple output heads. Experimental results demonstrate consistent performance improvements over traditional methods across multiple change detection datasets.

Deep neural networks often inherit biases from training data, compromising generalization. In visual recognition, distinguishing foreground from background is insufficient, as models tend to rely on spurious correlations rather than learning essential causal patterns. To address this issue, this paper proposes a novel transformer architecture, termed dual causal-aware detection transformer (DCDT), specifically designed for object detection in optical remote sensing images from a causal perspective. Specifically, we begin by constructing a structural causal model to intuitively analyze the causal effects inherent in the overall visual patterns. Building on this foundation, DCDT introduces causal constraints at the attention level by embedding dynamic multi-scale causal prototypes into the attention mechanism. The derived causal priors are subsequently used to enhance features at the representation level, thereby enforcing feature-level causal modulation. This dual causal-aware strategy enables the precise extraction and reinforcement of causally relevant features, improving both robustness and discriminative capability in complex detection scenarios. In addition, a sparse kernel-region mask is incorporated to decouple local information from global representations, effectively strengthening the modeling of fine-grained structures. Extensive experiments conducted on two challenging public datasets, DIOR and HRRSD, demonstrate that DCDT consistently outperforms existing methods and baselines. These results validate the effectiveness of DCDT in capturing both global causal semantics and local fine-grained features, highlighting its practicality in complex remote sensing scenarios.

Applying vision foundation models to remote sensing change detection (CD) has attracted extensive research attention. These studies employ inherent general knowledge from vision foundation models to enhance CD performance. Existing methods explicitly employ the foundation model as a feature extractor while designing additional learnable modules to bridge the task gap. However, these methods substantially increase the computational burden and memory demand in the inference. This paper therefore focuses on addressing the core challenge of effectively leveraging the knowledge from vision foundation models to enhance CD performance while maintaining computational efficiency. Instead of explicitly utilizing the foundation model, we propose Burden-Free Distillation (BFD), an architecture-agnostic foundation model-based distillation framework for efficient CD. BFD transfers the general knowledge from foundation models to task-specific models, thereby eliminating the dependency on foundation models during inference. Specifically, BFD transfers the foundation model knowledge through Dual-temporal Feature Matching module (DFM). This module enables multi-level feature alignment by computing pixel-wise spatial similarity between the foundation models’ general features and the CD models’ task-specific features. Additionally, we leverage patch contrastive distillation, which transfers localized structural patterns to the CD model to further mitigate task discrepancies between foundation models and CD models. We conduct extensive experiments across multiple foundation models and CD architectures, experimental results demonstrate that BFD effectively adapts the knowledge of foundation models to CD tasks without additional computational burden. Compared to other foundation model-based CD methods, BFD reduces the model parameters by 80.3% and improves IoU by 1.78% on the S2Looking dataset. The code is available at https://github.com/Younger-hua/Burden-Free-Distillation.

Building change detection (BCD) is a hot topic in geoscience and remote sensing (RS) with widespread applications. However, most existing BCD methods only focus on areas where changes have occurred, but ignore the change statuses. To address this problem, a building fine-grained change detection (BFCD) task is further explored in this work, which aims to judge the time-related “disappeared”, “appeared”, and “rebuilt” change types of buildings. Meanwhile, a scale- and shape-aware network (S2Net) with prediction decoupling is designed. Firstly, a prediction decoupling framework with dual decoders is built to ensure the prediction consistency with the temporal order of bi-temporal images. Secondly, considering the rebuilt type is the changes between building instances, which are often reflected in the scale and shape differences of the buildings. Thereby, a scale-aware module (ScAM) and a shape-aware module (ShAM) are designed. These two modules help extract the discriminative features of buildings with different scales and shapes for subsequent change detection (CD). In addition, two BCD datasets widely used, LEVIR-CD+ and WHU-CD, are relabeled in this work to support the study of BFCD. Experimental results show that S2Net achieves competitive performance, and its effectiveness is confirmed. The code and datasets will be publicly available at https://github.com/ptdoge/S2Net.

The scarcity of labeled data and domain shift among polarimetric synthetic aperture radar (PolSAR) images degrades the performance of the supervised-learning-based algorithm. Some unsupervised domain adaptation (UDA) algorithms have been proposed to address this problem and achieve good performance. The existing UDA algorithms for PolSAR terrain classification focus on the feature distribution shift problem but ignore the label shift problem in UDA task. In addition, feature alignment-based algorithms generate pseudo labels for target domain which introduce label noise and compromising the UDA performance. To alleviate the problems above, we present a cross-domain feature fusion network (CDFNet) for PolSAR terrain classification. Specifically, a domain-balanced sampling (DBS) module is proposed to obtain a nearly balanced training dataset to alleviate the label shift problem. Then, a cross-domain feature fusion (CDF) module is presented to achieve class-wise feature alignment with no additional label noise introduction. Experimental results on four PolSAR datasets demonstrate that our algorithm outperforms state-of-the-art UDA algorithms in terms of target domain performance.

The land cover analysis using hyperspectral images (HSIs) remains an open problem due to their low spatial resolution and complex spectral information. Recent studies are primarily dedicated to designing Transformer-based architectures for spatial-spectral long-range dependencies modeling, which is computationally expensive with quadratic complexity. Selective structured state space model (SSM; Mamba), which is efficient for modeling long-range dependencies with linear complexity, has recently shown promising progress. However, its potential in HSI processing that requires handling numerous spectral bands has not yet been explored. In this article, we innovatively propose S2Mamba, a spatial-spectral SSM for HSI classification, to excavate spatial-spectral contextual features, resulting in more efficient and accurate land cover analysis. In S2Mamba, two selective structured SSMs through different dimensions are designed for feature extraction, one for spatial, and the other for spectral, along with a spatial-spectral mixture gate (SMG) for optimal fusion. More specifically, S2Mamba first captures spatial contextual relations by interacting each pixel with its adjacent through a patch cross scanning (PCS) module and then explores semantic information from continuous spectral bands through a bidirectional spectral scanning (BSS) module. Considering the distinct expertise of the two attributes in homogenous and complicated texture scenes, we realize the SMG by a group of learnable matrices, allowing for the adaptive incorporation of representations learned across different dimensions. Extensive experiments conducted on HSI classification benchmarks demonstrate the superiority and prospect of S2Mamba. The code will be made available at:https://github.com/PURE-melo/S2Mamba.

Contrastive learning is a promising paradigm in the multimodal few-shot domain, leveraging images captured from different sensors as diverse perspectives of the same ground objects. However, existing methods encounter two major challenges: relatively weak discriminative ability, complex features inherent in remote sensing data, and insufficient cross-modal interactions, as these methods often align features within individual modalities without fully exploiting their complementary information. These limitations hinder the overall performance of models in multi-modal remote sensing tasks. In this paper, we propose a novel framework to address these challenges: First, we introduce the Transformer as the encoder, at the same time, we replace traditional multilayer perceptrons with the Kolmogorov–Arnold Network (KAN) network for feature dimensionality reduction, the model reduces parameters slightly without affecting accuracy, achieving better efficiency. Second, we design a specialized projection layer to align and deeply integrate spatial and LiDAR modalities, effectively harnessing their complementary strengths and implementing a self-supervised decoding mechanism to reconstruct masked image patches generated during the data augmentation process, enforcing spatial and structural learning for robust multi-modal representations. Extensive experiments conducted on four public datasets demonstrate the superior classification performance of our proposed method, outperforming state-of-the-art approaches and confirming its efficacy in multi-modal remote sensing tasks.

With the gradual maturity of deep learning technology and its extensive application in the field of remote sensing, hyperspectral image classification technology has made tremendous progress. The existing methods can achieve excellent classification performance in a closed-set environment (CSE), where the class distribution of the training and test set is consistent. However, in the open environment of the real world, many samples in the test set belong to the classes that never appear in the training set. Therefore, when unknown samples are present, existing closed-set classification methods will predict them as one of the known classes and cannot identify the unknown samples. Open Set Recognition (OSR) aims to solve this problem by simultaneously identifying unknown classes and distinguishing known classes. In this paper, we propose a reconstruction-enhanced prototype network (RePro) for HSI open-set classification. Specifically, the prototypes are constructed for each class in the embedding space, and the classification loss is also introduced to make the samples closer to the prototype of each class. Then the reconstruction loss based onL1is adopted to fully exploit the spectral-spatial information of training samples to provide the auxiliary information to identify the samples of unknown classes. Finally,the reconstruction loss and prototype distance are combined to identify samples of unknown classes. Experiments on multiple datasets demonstrate the effectiveness of the proposed method.

Effective modeling of the relationship between local spectral details and global contextual information remains a core challenge in hyperspectral image (HSI) classification. In this paper, a local-global spectral feature-aware network (LGSFA-Net) is proposed, which achieves local and global spectral feature learning through synergistic integration of local convolutional inductive biases and global state-space models (SSMs). The architecture of LGSFA-Net comprises three sequential components, including an embedding stage using convolutions for fundamental feature extraction, an encoding stage that cascades standard Mamba blocks with specialized interactive Mamba (IMamba) blocks and an enhanced spatial-spectral feature fusion (ESSFF) module. The proposed IMamba blocks employ separable convolutions and feature interaction learning for explicitly modeling the cross-channel spectral correlations learning, which can be effective in awareness of the spectral feature. And then, the ESSFF module utilizes self-attention mechanisms to dynamically balance local and global spatial-spectral feature contributions. The final prediction stage incorporates a lightweight classification head for efficient inference. Experimental results validate the effectiveness of the proposed methods for HSI classification on four benchmark datasets, including the PaviaU, Houston, Honghu, and Hanchuan datasets. The proposed LGSFA-Net achieves approximately 1.48%-2.51% increased overall accuracy (OA), 1.34%-2.06% increased average accuracy(AA), and 1.37%-3.75% increased Kappa on the aforementioned four datasets, respectively, outperforming the contrasting methods. The code implementation will be available at https://github.com/yutinyang/LGSFA-Net.

Pan-sharpening is the process of fusing panchromatic (PAN) and multispectral (MS) images. Its critical focus lies in accurately capturing the contour information from the PAN image during the fusion process and presenting it at a high resolution. However, existing deep learning methods lack the precise capture of delicate and smooth contour information, resulting in contour diffusion that affects the fusion results. Therefore, we introduce contourlet decomposition to capture multiscale directional delicate contour features and construct multiscale graph structures for semantic mining of dual-source contour features, continually updated through dynamic learning. By incorporating global features, we guide the multihead attention mechanism with directional decoding, enabling the network to pay more attention to high-resolution contour features, thereby gaining an advantage in image reconstruction. Cross-decoding between modalities provides strong representational capabilities for the advantageous features of both modalities, effectively enhancing the sharpening effect. Our algorithm achieves state-of-the-art results, and its effectiveness and advantages have been thoroughly validated across multiple datasets, including GaoFen-2, WorldView2, WorldView3, etc. Our code is available athttps://github.com/Xidian-AIGroup190726/CDFInet.

Remote sensing image super-resolution (RSISR) techniques are crucial in practice as an economical approach to enhancing the resolution of remote sensing images (RSIs). The scale of structural information and the richness of texture details in RSIs far exceed those in natural images. Therefore, accurately restoring and preserving edge and detail information are a critical challenge in the super-resolution (SR) process. Currently, convolutional neural network (CNN)-based methods primarily rely on local feature extraction, which fails to effectively capture and integrate global contextual information. Generative adversarial network (GAN)-based methods, while improving the visual quality, often suffer from artifacts and training instability, adversely affecting image quality. Moreover, these approaches struggle to accurately represent high-frequency features, leading to blurriness or distortion when reconstructing fine details and edges. To address these limitations, we introduce the channel–spatial coherent transformer (CSCT). The core of CSCT includes the channel–spatial coherent attention (CSCA) and the frequency-gated feed-forward network (FGFN), which work synergistically to enhance edge and detail preservation while significantly improving overall image clarity. CSCA efficiently aggregates channel and spatial information, while FGFN adaptively adjusts frequency information to enhance high-frequency details and suppress low-frequency noise. Moreover, this article leverages advanced data augmentation methods that markedly boost RSISR performance, offering new avenues for further exploration. The empirical analysis across several remote sensing SR benchmark datasets reveals that our approach excels in detail restoration, effectively reduces artifacts and noise, and significantly enhances the quality of SR images.

Remote sensing image captioning plays an important role in advancing remote sensing image understanding with natural language generation. However, it is difficult to generate accurate semantic descriptions of crucial objects and their relationships, due to large coverage and abundant information in remote sensing images. To address these issues, this article proposes a novel interactive concept network enhanced transformer (ICNET) for remote sensing image captioning. First, multilevel visual features are extracted within a local and global feature extraction module. To comprehensively capture key objects in the local features, a concept mapping network (CMN) is constructed to project multiscale local features onto high-level semantic concepts of the objects. This allows for the integration of the relevant feature vectors in the visual feature mapping into multiple relatively independent word features, thus bridging the gap between visual features and semantic concepts. Subsequently, a global feature enhancement (GFE) module is introduced to boost the discrimination of global relationships and filter irrelevant content. Finally, to aggregate semantic concepts and global features, a transformer equipped with a concept interaction module (CIM) is designed to facilitate feature alignment and generate captions with proper categories and relationships. The experimental results on three remote sensing image captioning datasets demonstrate the superiority of the proposed method.

Remote sensing image captioning between visual images and natural language remains a long-standing challenge in the remote sensing community. Due to the wide coverage and large amount of information in remote sensing images, existing methods struggle to effectively utilize the relevant semantic information about objects and their attributes at different scales across samples to generate descriptions. To address these issues, the article proposes a novel Adaptive Scale-aware Semantic Memory Network (ASSMN) for remote sensing image captioning. First, to fully extract the semantic information in remote sensing images, multilevel feature enhancement is constructed to improve the feature representation extracted from the CLIP pre-training model. Subsequently, a scale-aware attention aggregator is introduced to further integrate the enhanced multi-scale image features into the high-level semantics of remote sensing images. Then, to fully exploit the semantic information of the joint observed samples, a semantic memory reinforcement is designed to strengthen the semantic representation of the current scene through the relevant semantics obtained from other training samples. Finally, a captioning decoder is performed to generate a comprehensive scene caption with accurate objects and attributes. In the experiments, the performance of the proposed ASSMN on three remote sensing image captioning datasets is evaluated and compared with well-designed baselines and state-of-the-art methods, and the superiority of the proposed method is demonstrated by ablating the role of each proposed component. The code will be available at https://github.com/zcsisiyao/ASSMN.

Recently, cross-domain few-shot learning (FSL) has achieved remarkable performance in hyperspectral image classification (HSIC). However, current prototype-based FSL methods overlook the redundancy between prototypes, which may lead to intra-domain redundancy dependence and interfere with the effective alignment of data distributions between the source domain (SD) and target domain (TD). To address these issues, we propose a prototype mask contrast and cross-stage feature refinement (PMC-CFR) method for few-shot HSIC. Specifically, a neighbor-resistant multidimensional feature fusion network (NRMFF) is designed to fuse features inclined toward different dimensional information after random neighboring-pixel removal. Next, the parameter-free adaptive mask generator (PAMG) identifies potential redundant components based on the statistical distribution of intra-domain prototypes to form negative pairs, while treating other components as key features to form positive pairs. Contrastive pairs are further processed through graph convolution (GC) for supervised contrastive learning (SCL), which mitigates the dependence on intra-domain redundancy and enhances the learning of key features. To extract more representative features, multi-level representations are formed depending on whether the shared multilayer perceptron (MLP) is applied to focus on key components in each domain by fitting positive pairs. Meanwhile, the directional calibration (DC) separates inter-class prototypes across levels by balancing direction and distance. The synergy between MLP and DC in hierarchical feature modeling and semantic separation enables cross-stage feature refinement. Experimental results on four public HSI datasets demonstrate that the PMC-CFR outperforms many advanced contrastive methods, and feature visualizations further support its effectiveness in cross-domain feature distribution alignment.

In recent years, in order to alleviate the performance degradation problem caused by domain drift in scene classification tasks, some unsupervised domain adaptation methods have been introduced into the field of remote sensing images. Such methods require simultaneous access to both source and target domain data during training. However, the large storage and transmission costs of remote sensing images limit the further application of these methods. To address this challenge, we investigate the task of source-free cross-domain scene classification for remote sensing images. In the model adaptation process, only the target domain dataset and the trained source domain model are used. Our approach consists of two parts: a distribution alignment strategy based on source domain model statistics matching and a noise adaptation strategy. In order to fully utilize the knowledge of the pre-trained source domain model, we fix the classifier to get the feature distribution of the source domain, so that the target domain feature distribution is close to the feature distribution of the source domain. The noise adaptation layer is inserted after the classifier in order to improve the robustness of the model to noise-containing pseudo-labels, and the sample-wise noise transfer matrix is learned. Experimental results on 12 transfer tasks on the cross-scene dataset, and 2 transfer tasks on the cross-sensor dataset, to validate the effectiveness of our approach. Compared to traditional unsupervised domain adaptive methods, our method is able to achieve better performance under the condition of not accessing the source domain data.

Automated building footprint segmentation is crucial in remote sensing with widespread applications in various fields. Semi-supervised semantic segmentation methods are gaining traction in the remote sensing community as they significantly reduce the need for labor-intensive pixel-level annotations in training segmentation models. However, these methods typically rely on individual pixel-level supervision for unlabeled data, neglecting the contextual relationships between pixels. This limits their potential to exploit unlabeled data. To bridge this gap, this paper proposes a novel Context Priors Guided Semi-Supervised Building Footprint Segmentation method that leverages contextual relationships among numerous unlabeled pixels to build supervisory signals that extend beyond individual pixel-level guidance for learning on unlabeled data. The CPG comprises two main components: Spatial context priors-guided pseudo-label regularization and semantic context priors-guided representation learning. By integrating contextual knowledge from both pixel spatial locations and semantic representation spaces, these components capture comprehensive class semantic attributes and enable the model to learn complete shapes of building footprints. Our approach achieves state-of-the-art performance on three publicly available building footprint segmentation datasets, validating its effectiveness.

Although modern Remote Sensing Object Detection (RSOD) methods have achieved advanced performance, they heavily rely on a large amount of annotated data. This paper explores semi-supervised RSOD to mitigate annotation costs, leveraging recent extensive research in generic Semi-Supervised Object Detection (SSOD) based on the self-training paradigm. Current SSOD methods encounter challenges in adapting to remote sensing images due to the complexity and variability of RSIs. Two key issues remain underexplored: the noise in pseudo-labels caused by model instability and the difficulty in distinguishing similar categories. This paper introduces the Temporal-Feedback Self-Training (TST) framework, a novel approach to tackle these challenges in semi-supervised RSOD. TST consists of two components: Temporal Consistency Based Pseudo-labels Certainty Estimation (TCE) and Temporal Self-Feedback Feature Refinement (TSF). TCE addresses pseudo-label noise during training by evaluating the stability of pseudo-label classification and localization over time series to assess the quality of pseudo-labels. On the other hand, TSF enhances pseudo-label quality by dynamically identifying the models confusing categories as feedback for feature refinement. Both components facilitate the progression of the self-training-based RSOD during training. We conducted extensive experiments on two challenging public datasets, DOTA and DIOR. The results demonstrate that the proposed TST and TCE components significantly improve the baseline models performance, surpassing the state-of-the-art generic SSOD method. This suggests that our approach is more effective than generic SSOD methods in addressing the challenges posed by remote sensing images.

Automatic Modulation Classification (AMC)-oriented Deep Neural Networks (ADNNs) have received much attention in recent years for their wide range of applications. However, they are vulnerable to attacks. Adversarial Examples (AEs) of modulation signals with added weak perturbations can easily fool ADNNs. The study of AEs on AMC, on one side, can enhance the security of wireless communication systems; on the other side, it can provide an effective defence against potential attacks. Nevertheless, most existing attack methods generate AEs with low transferability. In this paper, we propose a Multiscale Discriminative Attack Method (MDAM) for modulated signals. The method strives to alleviate such transferability issue by destroying discriminative features in multi-layer. Specifically, we utilize interpretable class activation maps to distinguish the discriminative regions, ignoring the noise and focusing on the interference of the discriminative features. Beyond that, we propose a multi-layer activation disruption loss to constrain activations in the middle layers. In so doing, the AEs do not erroneously retain deep features of the original signal. We conduct extensive experiments on RadioML datasets and the local area network (LAN) communication dataset we collected to evaluate the effectiveness of MDAM in both white-box and black-box attack scenarios. The results show that MDAM outperforms existing methods.

Ballistocardiograms (BCG) is a passive, non-contact heart rate detection technology that requires no action on the part of the individual. However, during the BCG signal acquisition process, the surface pressure generated by cardiac contraction is easily disturbed by external factors, and as people's health deteriorates, the j-peak (the main peak of the BCG signal) is no longer prominent. Our aim is to establish a non-contact, self-supervised heart rate detection method based on physiological information, to improve the accuracy and robustness of BCG heart rate detection under wider and more adverse conditions. The algorithm is guided by the heart rate estimation based on BCG itself, thereby reconstructing a signal with physiological significance. We also propose a heartbeat mapping algorithm based on Bidirectional Long Short-Term Memory Network (BiLSTM) for extracting global deep features, achieving real-time heartbeat prediction, and eliminating local deviations brought about by reconstruction. To verify the effectiveness of the proposed method, this paper evaluated 40 young subjects and 4 elderly subjects. Compared with the existing state-of-the-art methods, beat-to-beat heart rate estimation and heartbeat detection both performed excellently, surpassing most methods using precise labels. The experimental results show that the proposed method achieves effective heartbeat detection, demonstrating robustness and effectiveness in the face of unavoidable noise and variations.

The long-tailed data distribution frequently occurs in the real-world scenarios, whereas deep learning is not effective enough for such distribution. In order to improve the effectiveness for the long-tailed data, data augmentation is widely used to balance the distribution of classes by generating new samples. However, most existing studies are designed from the perspective of the class-independence assumption by default, ignoring the effect of interrelation among classes for data augmentation, which causes that some generated samples may be unrepresentative and useless for balancing the class-distribution. Inspired by this, we propose a new data augmentation method based the sparse class-correlation exploitation in this paper, which can generate more representative samples by utilizing the class-correlation, to effectively balance the class-distribution for the long-tailed data. In the proposed method, a sparse class-correlation exploration module is first proposed to explore the potential correlations among multiple classes for boosting the classification performance. Based on the class-correlations, the pivotal seed-samples are generated by maximizing the sparse representation of challenging samples. Meanwhile, an ambiguity-filtered translation module is designed to generate more representative new samples for the target classes based the obtained seed-samples by enhancing the class-consistency and suppressing the deviation from the target classes. In addition, we introduce the self-supervised feature and fuse it with the discriminative feature to explore more accurate class-correlations. Experimental results illustrate that the proposed method obtains better performance only with a small number of generated samples than the state-of-the-art methods.

Due to the presence of redundancy and interference, frame sampling is a promising but challenging solution to mitigate the expensive computation of video action recognition. Although the motion prior has shown great potential for frame selection, existing motion-based strategies suffer from limitations in terms of robustness and interpretive evaluation. In this paper, we devise a robust frame sampling strategy called positive motion guided sampler (PMGSampler). It consists of two procedures, local motion capture and global motion statistics. At the local level, we propose two concepts about inter-frame motion amplitude and motion continuity, which helps to perceive the movement of subjects and identify abnormal events that may generate negative pseudo-motion information. Then, through a global analysis of the obtained local motions, the sampler becomes more sensitive to informative frames and robust to outliers. The proposed sampler can be applied to most existing models for improving recognition accuracy. We conduct extensive experiments on four widely-used benchmarks to demonstrate the superiority of our PMGSampler over other methods of the same type. In addition, to analyse how sampled frames influence action recognition, we present a visual interpretation method for video models, termed as spatio-temporal class activation map (STCAM). By introducing spatial and temporal branches, our STCAM is able to visualise the salience of spatio-temporal features. With the help of STCAM, we can further intuitively evaluate the performance of different sampling strategies.

Video anomaly detection (VAD) is an important intelligent system application, but most current research views it as a coarse binary classification task that lacks a fine-grained understanding of abnormal video sequences. We explore a new task for video anomaly analysis called Comprehensive Video Anomaly Caption (CVAC), which aims to generate comprehensive textual captions (containing scene information such as time, location, anomalous subject, anomalous behavior, etc.) for surveillance videos. CVAC is more consistent with human understanding than VAD, but it has not been well explored. We constructed a large-scale benchmark CVACBench to lead this research. For each video clip, we provide 6 fine-grained annotations, including scene information and abnormal keywords. A new evaluation metric Abnormal-F1 (A-F1) is also proposed to more accurately evaluate the caption generation performance of the model. We also designed a method called Anomaly-Led Generating Prompting Transformer (AGPFormer) as a baseline. In AGPFormer, we introduce an anomaly-led language modeling mechanism (Anomaly-Led MLM, AMLM) to focus on anomalous events in videos. To achieve more efficient cross-modal semantic understanding, we design the Interactive Generating Prompting (IGP) module and Scene Alignment Prompting (SAP) module to explore the divide between video and text modalities from multiple perspectives, and to improve the model's performance in understanding and reasoning about the complex semantics of videos. We conducted experiments on CVACBench by using traditional caption metrics and the proposed metrics, and the experimental results demonstrate the effectiveness of AGPFormer in the field of anomaly caption.

View synthesis of aerial scenes has gained attention in the recent development of applications such as urban planning, navigation, and disaster assessment. This development is closely connected to the recent advancement of the Neural Radiance Field (NeRF). However, when autonomousaerial vehicles(AAVs) encounter constraints such as limited perspectives or energy limitations, NeRF degrades with sparsely sampled views in complex aerial scenes. On this basis, we aim to solve this problem in a few-shot manner. In this paper, we propose Uncertainty Guided Perception NeRF (UPNeRF), an uncertainty-guided perceptual learning framework that focuses on applying and improving NeRF in few-shot aerial view synthesis (FSAVS). First, simply optimizing NeRF in complex aerial scenes with sparse input can lead to overfitting in training views, resulting in a collapsed model. To address this, we propose a progressive learning strategy that utilizes the uncertainty present in sparsely sampled views, enabling a gradual transition from easy to hard learning. Second, to take advantage of the inherent inductive bias in the data, we introduce an uncertainty-aware discriminator. This discriminator leverages convolutional capabilities to capture intricate patterns in the rendered patches associated with uncertainty. Third, direct optimization of NeRF lacks prior knowledge of the scene. This, coupled with a reduction in training views, can result in unrealistic rendering. To overcome this, we present a perceptual regularizer that incorporates prior knowledge through prompt tuning of a self-supervised pre-trained vision transformer. In addition, we adopt a sampled scene annealing strategy to enhance training stability. Finally, we conducted experiments with two public datasets, and the positive results indicate our method is effective.

Self-attention learns capturing the long-range dependencies between embeddings (e.g., image pixels). However, the memory overhead and computation cost are prohibitive due to being quadratic in term of the spatial resolution. The structure analysis reveals two crucial roles in the attention: the correlation-based dependency structure and feature normalization. In this work, an efficacious Local-Global Semantics (LGS) module is proposed to alleviate the above issues by modeling the local semantic aggregation and global semantic interaction. Our LGS module contains a group convolution and an Efficient Global Semantic Attention (EGSA). Firstly, the group convolution aggregates local semantics. Secondly, considering a feature map as a sequence of 2-D channel representations, EGSA formulates a general model for the global semantic interaction. The linear correlation is computed between global semantics. LGS has the linear memory overhead and computation cost in term of the spatial resolution. The LGS module can be smoothly incorporated into object detection frameworks. The experiment results verify its effectiveness on two popular detection datasets: the MS COCO and PASCAL VOC.

Transformer displays the impressive capabilities on vision tasks. The built-in self-attention retains the quadratic computation burden in respect of the spatial resolution of image features. The traditional downsampling (e.g., average pooling) can reduce the resolution. Nonetheless, it may suffer from the dropping of detailed information. In this work, we propose an Efficient Wavelet Attention (EWA), which injects the wavelet transform and a Mean GELU (MGELU) function. Firstly, the wavelet transform enables the detailed information to participate in the efficient interaction modeling. Secondly, MGELU regards the statistical mean as reference and loosely passes the high relative responses. Building upon EWA, we present an effective Semantic-aware Wavelet Transformer (SWFormer), which is then employed for pyramid learning, including CNN feature hierarchy or Region of Interest (RoI) features. For the feature hierarchy, a Pyramid SWFormer (PSWFormer) incorporates SWFormer at each level to fit the bidirectional features. For RoIs, a Recognition-Localization SWFormer (RLSWFormer) is inserted into the head to fit their features from all levels. The effectiveness of our SWFormer is displayed experimentally on the MS COCO detection dataset and the Pascal VOC dataset. When exploiting Swin-small backbone, our SWFormer-based method acquires AP of 52.1 in the single-scale evaluation on the COCO test-dev set. This work will have the codes athttps://github.com/TimeIsFuture/Dt2_SWFormer.

Visual transformers have achieved great success in representation learning. This is mainly due to efficient token dependency modeling via self-attention. However, the computational burden increases sharply as the input pixels increase. Although recent Fourier-based global frequency-domain mixing methods attempt to improve the efficiency of transformers for high-resolution image inputs, the Fourier operator has limited ability to capture the local geometric structure. Complex wavelets can perform local attention in both the spatial domain and the frequency domain. Therefore, we propose the complex wavelet informed transformer operator that uses the real and imaginary wavelets of the dual-tree complex wavelet transform to simulate the interaction in the attention kernel. In order to further reduce the computational burden of operators, we introduce an adaptive local block shared attention mechanism in the channel domain for our wavelet informed operators. Further, we construct the deep multi-head operator network consisting of a hybrid stack of complex wavelet informed transformer operators and self-attention layers. This enables the Transformer to more sparsely capture multi-scale and multi-directional structured features in the process of learning dependencies. Extensive experimental results show that our adaptive complex wavelet informed transformer operator under the Transformer architecture achieves highly competitive accuracy performance on multiple image classification benchmark datasets. And the proposed operators can be flexibly and effectively migrated to vision tasks in dynamic video scenarios.

Deep learning based remote sensing (RS) image segmentation significantly impacts several real application scenarios. Behind its success, massive labeled data plays an important role. However, annotating high-resolution RS images requires time-consuming and relevant expertise efforts. To address it, many works dive into semi-supervised learning which utilizes raw information embedded in unlabeled data to improve the segmentation model. Nevertheless, previous studies ignore the integrity and effectiveness of the potential context information hidden in RS data. In this work, we propose an uncertainty-aware masked consistency learning (U-MCL) framework that contains an uncertainty-aware masked denoising (U-MD) module and an uncertainty-aware masked image consistency (U-MIC) module. U-MCL initially generates a patch-wise uncertainty map for each unlabeled image during each training iteration, which is then used to derive an adaptive mask ratio for pseudo-label denoising in U-MD. Simultaneously, the uncertainty map is adopted to model a masked unlabeled image for reasoning unseen areas in U-MIC. Consequently, U-MCL is capable of enhancing model performance by engaging in accurate and stable consistency learning while preserving the integrity of the context and employing the context to infer the predictions of the masked regions safely. Extensive experiments on six RS datasets, i.e., ISPRS Vaihingen, FloodNet, MiniFrance, LoveDA, MER, and MSL, demonstrate the superiority of our U-MCL over recent most advanced methods, achieving new state-of-the-art performance under all benchmarks.

As imaging sensor technology in remote sensing has advanced quickly, multimodal fusion classification has become an important research direction in land cover and urban planning classification tasks. While generative models and image classification have greatly benefited from diffusion models, the present ones primarily concentrate on single-modality-driven diffusion processes. Therefore, this paper presents a 3D self-awareness diffusion network (3DSA-DiffNet) for multispectral (MS) and panchromatic (PAN) image fusion classification, which would make it easier to classify heterogeneous data from various sensors. First, in order to model the relationship between multi-channel spectra and multi-pixel spatial distributions as well as samples, respectively, a spatial-spectral joint denoising network (S$^{2}$JD-Net) is proposed. It can incorporate the diffusion process into the neural network to enhance the quality of diffusion features. Secondly, to imitate the brain's spatial-spectral coexistence learning mechanism, this work offers a 3D self-awareness module (3DSA-Module) that can learn the weight of each pixel in 3D space, resulting in extraordinarily high feature representation capabilities. Finally, experimental verification demonstrates that the 3D self-awareness diffusion fusion network driven by brain inspiration outperforms more sophisticated approaches on the Xi'an, Huhhot, and Muufl datasets.

In satellite video object tracking, the individual frame analysis method is usually used for target localization, ignoring informative cues of the dynamic scene. Temporal information could contribute to identifying the target from distractors. In this work, a novel dynamic scene learning reasoning tracker is proposed for satellite videos, which reasons over temporal dynamic information to derive the target location. It is inspired by the tracking pattern through human perception and reasoning. First, static-dynamic united analysis is designed to construct dynamic scenes by concatenating the static searching results along the temporal dimension. Second, the information of each response object is aggregated by wavelet transforms. Meanwhile, these scenes are projected into low-frequency and high-frequency subspaces, which could imitate different levels of perceptions of humans for scenes. Third, an object-aware reasoning transformer is proposed to utilize the temporal dynamics of input response objects. In each subspace, it models the mutual interactions between dynamic objects and further learns the intrinsic property of each object for target reasoning. Finally, to obtain the current reasoning result, inverse wavelet transforms are utilized to integrate the results of low-frequency and high-frequency subspaces. The effectiveness of the proposed method is validated on three public satellite video datasets, including SV248S, SkySat, and VISO. Qualitative and quantitative experimental results show that the proposed tracker outperforms 22 popular approaches in seven challenging tracking satellite scenarios.

Change detection has essential significance for the region's development, in which pseudo-changes between bitemporal images induced by imaging environmental factors are key challenges. Existing transformation-based methods regard pseudo-changes as a kind of style shift and alleviate it by transforming bitemporal images into the same style using generative adversarial networks (GANs). However, their efforts are limited by two drawbacks: 1) Transformed images suffer from distortion that reduces feature discrimination. 2) Alignment hampers the model from learning domain-agnostic representations that degrades performance on scenes with domain shifts from the training data. Therefore, oriented from pseudo-changes caused by style differences, we present a generalizable domain-agnostic difference learning network (DonaNet). For the drawback 1), we argue for local-level statistics as style proxies to assist against domain shifts. For the drawback 2), DonaNet learns domain-agnostic representations by removing domain-specific style of encoded features and highlighting the class characteristics of objects. In the removal, we propose a domain difference removal module to reduce feature variance while preserving discriminative properties and propose its enhanced version to provide possibilities for eliminating more style by decorrelating the correlation between features. In the highlighting, we propose a cross-temporal generalization learning strategy to imitate latent domain shifts, thus enabling the model to extract feature representations more robust to shifts actively. Extensive experiments conducted on three public datasets demonstrate that DonaNet outperforms existing state-of-the-art methods with a smaller model size and is more robust to domain shift.

Many research works have shown that the traditional alternating direction multiplier methods (ADMMs) can be better understood by continuous-time differential equations (DEs). On the other hand, many unfolded algorithms directly inherit the traditional iterations to build deep networks. Although they achieve superior practical performance and a faster convergence rate than traditional counterparts, there is a lack of clear insight into unfolded network structures. Thus, we attempt to explore the unfolded linearized ADMM (LADMM) from the perspective of DEs, and design more efficient unfolded networks. First, by proposing an unfolded Euler LADMM scheme and inspired by the trapezoid discretization, we design a new more accurate Trapezoid LADMM scheme. For the convenience of implementation, we provide its explicit version via a prediction-correction strategy. Then, to expand the representation space of unfolded networks, we design an accelerated variant of our Euler LADMM scheme, which can be interpreted as second-order DEs with stronger representation capabilities. To fully explore this representation space, we designed an accelerated Trapezoid LADMM scheme. To the best of our knowledge, this is the first work to explore a comprehensive connection with theoretical guarantees between unfolded ADMMs and first- (second-) order DEs. Finally, we instantiate our schemes as (A-)ELADMM and (A-)TLADMM with the proximal operators, and (A-)ELADMM-Net and (A-)TLADMM-Net with convolutional neural networks (CNNs). Extensive inverse problem experiments show that our Trapezoid LADMM schemes perform better than well-known methods.

How to learn and accurately distinguish new concepts from few samples, as humans do, is a long-standing concern in artificial intelligence (AI). Studies in brain science and neuroscience have shown that human brain perception is based on nonlinear manifolds, and high-dimensional manifolds can facilitate concept learning in neural circuits. Based on this inspiration, in this paper, we propose a heterogeneous Riemannian few-shot learning network (HRFL-Net), which is the first few-shot learning method to perform end-to-end deep learning on heterogeneous Riemannian manifolds. Specifically, to enhance the geometric invariance of the image representation, the image features are projected into three heterogeneous Riemannian manifold spaces. Then, the implicit Riemannian kernel function maps the manifolds to the separable high-dimensional reproducing Hilbert space. It is assumed that the embedded kernel features of the complementary manifolds are mapped to the same common subspace. Thus, a novel neural network-based Riemannian metric learning method is designed to solve the subspace feature vectors by imposing orthogonal normalized projection, which overcomes the data extension limitation of the Riemannian metric. Finally, with the optimization objective of increasing the interclass distance and decreasing the intraclass distance in Hilbert space, the HRFL-Net is trained with end-to-end stochastic optimization, and the optimal aggregation subspace is learned during the gradient descent process. Thus, the proposed HRFL-Net can be easily generalized to challenging nonconvex data. The evaluation of four public datasets shows that the proposed HRFL-Net has significant superiority and also achieves competitive results compared with the state-of-the-art methods.

Unsupervised domain adaptation (UDA) techniques, extensively studied in hyperspectral image (HSI) classification, aim to use labeled source domain data and unlabeled target domain data to learn domain invariant features for cross-scene classification. Compared to natural images, numerous spectral bands of HSIs provide abundant semantic information, but they also increase the domain shift significantly. In most existing methods, both explicit alignment and implicit alignment simply align feature distribution, ignoring domain information in the spectrum. We noted that when the spectral channel between source and target domains is distinguished obviously, the transfer performance of these methods tends to deteriorate. Additionally, their performance fluctuates greatly owing to the varying domain shifts across various datasets. To address these problems, a novel shift-sensitive spatial-spectral disentangling learning (S4DL) approach is proposed. In S4DL, gradient-guided spatial-spectral decomposition (GSSD) is designed to separate domain-specific and domain-invariant representations by generating tailored masks under the guidance of the gradient from domain classification. A shift-sensitive adaptive monitor is defined to adjust the intensity of disentangling according to the magnitude of domain shift. Furthermore, a reversible neural network is constructed to retain domain information that lies not only in semantic but also the shallow-level detailed information. Extensive experimental results on several cross-scene HSI datasets consistently verified that S4DL is better than the state-of-the-art UDA methods. Our source code will be available athttps://github.com/xdu-jjgs/IEEE_TNNLS_S4DL.

Multisource optical remote sensing (RS) image classification has obtained extensive research interest with demonstrated superiority. Existing approaches mainly improve classification performance by exploiting complementary information from multisource data. However, these approaches are insufficient in effectively extracting data features and utilizing correlations of multisource optical RS images. For this purpose, this article proposes a generalized spatial-spectral relation-guided fusion network (S2RGF-Net) for multisource optical RS image classification. First, we elaborate on spatial- and spectral-domain-specific feature encoders based on data characteristics to explore the rich feature information of optical RS data deeply. Subsequently, two relation-guided fusion strategies are proposed at the dual-level (intradomain and interdomain) to integrate multisource image information effectively. In the intradomain feature fusion, an adaptive de-redundancy fusion module (ADRF) is introduced to eliminate redundancy so that the spatial and spectral features are complete and compact, respectively. In interdomain feature fusion, we construct a spatial-spectral joint attention module (SSJA) based on interdomain relationships to sufficiently enhance the complementary features, so as to facilitate later fusion. Experiments on various multisource optical RS datasets demonstrate that S2RGF-Net outperforms other state-of-the-art (SOTA) methods.

The progress of brain cognition and learning mechanisms has provided new inspiration for the next generation of artificial intelligence (AI) and provided the biological basis for the establishment of new models and methods. Brain science can effectively improve the intelligence of existing models and systems. Compared with other reviews, this article provides a comprehensive review of brain-inspired deep learning algorithms for learning, perception, and cognition from microscopic, mesoscopic, macroscopic, and super-macroscopic perspectives. First, this article introduces the brain cognition mechanism. Then, it summarizes the existing studies on brain-inspired learning and modeling from the perspectives of neural structure, cognitive module, learning mechanism, and behavioral characteristics. Next, this article introduces the potential learning directions of brain-inspired learning from four aspects: perception, cognition, understanding, and decision-making. Finally, the top-ten open problems that brain-inspired learning, perception, and cognition currently face are summarized, and the next generation of AI technology has been prospected. This work intends to provide a quick overview of the research on brain-inspired AI algorithms and to motivate future research by illuminating the latest developments in brain science.

Recently, the multiscale problem in computer vision has gradually attracted people's attention. This article focuses on multiscale representation for object detection and recognition, comprehensively introduces the development of multiscale deep learning, and constructs an easy-to-understand, but powerful knowledge structure. First, we give the definition of scale, explain the multiscale mechanism of human vision, and then lead to the multiscale problem discussed in computer vision. Second, advanced multiscale representation methods are introduced, including pyramid representation, scale-space representation, and multiscale geometric representation. Third, the theory of multiscale deep learning is presented, which mainly discusses the multiscale modeling in convolutional neural networks (CNNs) and Vision Transformers (ViTs). Fourth, we compare the performance of multiple multiscale methods on different tasks, illustrating the effectiveness of different multiscale structural designs. Finally, based on the in-depth understanding of the existing methods, we point out several open issues and future directions for multiscale deep learning.

Attention-based transformer networks have recently played an increasingly important role in computer vision tasks. However, since pixel-by-pixel attention multiplication does not involve constraint assumptions such as spatial invariance, the computational complexity grows quadratically with the increase of input pixels. Therefore, this article proposes a complex pyramid scattering Transformer in dense scale space, which introduces sparse scattering constraints with a small number of wavelet basis parameters. It enhances the Transformer's flexibility and sparsity in multiscale space and, to a certain extent, slows down the increase in computational complexity caused by multiresolution input. In addition, compared with the general single-tree real wavelet transform, the dual-tree complex scattering method improves the aliasing of the scattering attention layer and helps obtain a more robust feature representation. At the same time, the multihead stepwise pyramid scattering coupling mechanism helps increase the abundance of directional priors. We conduct experiments in image classification and video tracking scenarios and verify the reliability and superiority of our dual-tree complex pyramid scattering Transformer for visual tasks with different scale requirements. The performance is better than that of the baseline Transformer and other advanced wavelet scattering networks at the same parameter scale. The code is available at https://github.com/Dawn5786/CPSTFormer.

The grounded situation recognition (GSR) task aims to recognize the structured semantics of an image to achieve "human-like" event understanding. Most previous studies primarily focus on the visual features of the situation, overlooking the step-by-step cognitive reasoning process that humans employ in complex task settings. Recently, the emergence of multimodal large language models (MLLMs) has provided novel directions for addressing complex problems. However, directly deploying MLLMs on the GSR task is suboptimal due to their tendency to exhibit "hallucination" issues. Additionally, fine-tuning MLLMs for the GSR task incurs high training costs. To address these challenges, inspired by human cognitive theory and the chain-of-thought (CoT) strategy, we propose the chain-of-situation progressive inference learning (CoS-PIL) framework, a lightweight approach that progressively completes verb prediction, noun prediction, and role grounding. The prediction of each step depends on the historical information of the previous step. Specifically, we first design situation prompts tailored to the GSR task and utilize MLLMs to analyze the input image and language prompts, generating heuristic response text for the current situation in the image. Instead of fine-tuning the MLLM, we activate the reasoning capabilities of the frozen MLLM and adapt its generated responses into three lightweight modules: CoS-Verb, CoS-Noun, and CoS-Ground. Considering that MLLMs may generate redundant content, we carefully design the chain-of-interest predictor (CoI-Predictor) to extract key information from the extensive response text and inject it into the model as prompts to enhance the performance. Extensive experiments on the challenging SWiG benchmark demonstrate that CoS-PIL outperforms other state-of-the-art methods. The code is publically available at https://github.com/XDLiuyyy/CoS-PIL.

The Transformer-convolutional neural network (CNN) hybrid learning approach is gaining traction for balancing deep and shallow image features for hierarchical semantic segmentation. However, they are still confronted with a contradiction between comprehensive semantic understanding and meticulous detail extraction. To solve this problem, this article proposes a novel Transformer-CNN hybrid hierarchical network, dubbed contourlet transformer (CoT). In the CoT framework, the semantic representation process of the Transformer is unavoidably peppered with sparsely distributed points that, while not desired, demand finer detail. Therefore, we design a deep detail representation (DDR) structure to investigate their fine-grained features. First, through contourlet transform (CT), we distill the high-frequency directional components from the raw image, yielding localized features that accommodate the inductive bias of CNN. Second, a CNN deep sparse learning (DSL) module takes them as input to represent the underlying detailed features. This memory- and energy-efficient learning method can keep the same sparse pattern between input and output. Finally, the decoder hierarchically fuses the detailed features with the semantic features via an image reconstruction-like fashion. Experiments demonstrate that CoT achieves competitive performance on three benchmark datasets: PASCAL Context [57.21% mean intersection over union (mIoU)], ADE20K (54.16% mIoU), and Cityscapes (84.23% mIoU). Furthermore, we conducted robustness studies to validate its resistance against various sorts of corruption. Our code is available at: https://github.com/yilinshao/CoT-Contourlet-Transformer.

The cross-modal remote sensing image-text retrieval (CMRSITR) is a lively research topic in the remote sensing (RS) community. Benefiting from the large pretrained image-text models, many successful CMRSITR methods have been proposed in recent years. Although their performance is attractive, there are still some challenges. First, fine-tuning large pretrained models requires a significant amount of computational resources. Second, most large models are pretrained by natural images, which reduces their effectiveness in processing RS images. To tackle these challenges, we propose a new CMRSITR network named context and uncertainty-aware prompt (CUP). First, prompt tuning theory is introduced into CUP to eliminate the burden of optimization resources. By training the prompt tokens rather than all parameters, the large model's knowledge can be transferred to CMRSITR tasks with small trainable parameters. Second, considering the differences between natural-image-based prior clues and RS images, apart from adopting the free-prompt tokens, we develop a prompt generation module (PGM) to produce the RS-oriented prompt tokens. The specific prompt tokens are rich in object-level messages of RS images, which help CUP narrow the gaps between natural large models and RS images. Third, we further design an uncertainty estimation module (UEM) to whittle down the uncertainties caused by the model and data. This way, can not only the semantic misalignment and intraclass diversity imbalance problems be mitigated but also the RS clues can be deeply explored. Competitive experimental results counted on three public benchmark datasets demonstrate that our CUP can achieve competitive performance in the CMRSITR task compared with many existing methods. Our source codes are available at: https://github.com/TangXu-Group/Cross-modal-remote-sensing-image-and-text-retrieval-models/tree/main/CUP.

Existing efforts are dedicated to designing many topologies and graph-aware strategies for the graph Transformer, which greatly improve the model's representation capabilities. However, manually determining the suitable Transformer architecture for a specific graph dataset or task requires extensive expert knowledge and laborious trials. This article proposes an evolutionary graph Transformer architecture search (EGTAS) framework to automate the construction of strong graph Transformers. We build a comprehensive graph Transformer search space with the micro-level and macro-level designs. EGTAS evolves graph Transformer topologies at the macro level and graph-aware strategies at the micro level. Furthermore, a surrogate model based on generic architectural coding is proposed to directly predict the performance of graph Transformers, substantially reducing the evaluation cost of evolutionary search. We demonstrate the efficacy of EGTAS across a range of graph-level and node-level tasks, encompassing both small-scale and large-scale graph datasets. Experimental results and ablation studies show that EGTAS can construct high-performance architectures that rival state-of-the-art manual and automated baselines.

Weakly supervised object detection (WSOD) and semantic segmentation with image-level annotations have attracted extensive attention due to their high label efficiency. Multiple instance learning (MIL) offers a feasible solution for the two tasks by treating each image as a bag with a series of instances (object regions or pixels) and identifying foreground instances that contribute to bag classification. However, conventional MIL paradigms often suffer from issues, e.g., discriminative instance domination and missing instances. In this article, we observe that negative instances usually contain valuable deterministic information, which is the key to solving the two issues. Motivated by this, we propose a novel MIL paradigm based on negative deterministic information (NDI), termed NDI-MIL, which is based on two core designs with a progressive relation: NDI collection and negative contrastive learning (NCL). In NDI collection, we identify and distill NDI from negative instances online by a dynamic feature bank. The collected NDI is then utilized in a NCL mechanism to locate and punish those discriminative regions, by which the discriminative instance domination and missing instances issues are effectively addressed, leading to improved object- and pixel-level localization accuracy and completeness. In addition, we design an NDI-guided instance selection (NGIS) strategy to further enhance the systematic performance. Experimental results on several public benchmarks, including PASCAL VOC 2007, PASCAL VOC 2012, and MS COCO, show that our method achieves satisfactory performance. The code is available at: https://github.com/GC-WSL/NDI.

This article considers a worst and most challenging scene in domain generalization (DG), where a model aims to generalize well on unseen domains while only one single domain is available for training. Existing randomization-based methods achieve this goal by enriching the style of the training data. However, they fail to guarantee the diversity of newly generated data required for generalization and thus lead to insufficient expansion of the training distribution. Thus, we propose a novel single DG (SDG) framework, unseen style seeking-based meta-learning (USSML). In USSML, multiple plausible domains with various styles are first constructed from a single source domain and the combination is performed across generated domains to emulate unseen images, extending the distribution boundaries of the source domain. The domain combination is performed at two levels, i.e., global and instance, to meet the generalization challenge in semantic segmentation. Then, the generated diverse domains are further exploited to force the model to optimize in an unbiased manner across all domains by relearning regions lacking domain-invariant representation capability, driving the model toward domain invariance. A point worth mentioning is that the proposed method is easily integrated into existing segmentation methods with little computational cost to improve their generalization. Extensive experiments are conducted on five popular segmentation datasets and the results have verified the effectiveness of USSML in improving the model's generalization and the superiority of USSML over existing works.

Partial point cloud registration plays a crucial role in computer vision and has widespread applications in 3D map construction, pose estimation, and high-precision localization. However, the collected point clouds often contain missing data due to hardware limitations and complex environments. Various partial registration algorithms have been proposed, most of which rely on estimating overlap regions. However, a significant proportion of these algorithms rely heavily on ground truth labels. Manual labeling is both time-consuming and labor-intensive, whereas algorithmic automatic labeling lacks sufficient accuracy. To tackle this issue, we present PSEudo Label learning for unsupervised partial point cloud registration (PSEL). This method utilizes complementary tasks to learn reliable pseudo labels for overlap regions and correspondences without depending on ground truth labels. The key idea is to use the complementarity between overlap estimation and registration to generate two types of pseudo labels based on the nearest points in pairs of aligned point clouds. These pseudo labels are then employed to supervise the learning of overlap regions and correspondences, gradually enhancing their accuracy throughout the learning process and ultimately establishing an unsupervised learning framework. PSEL consists of an overlap estimation module and a correspondence filtering module. The pseudo labels generated after registration are used to supervise both modules. Notably, the correspondence filtering module has two pipelines. The similarity and difference of the corresponding point features are used to eliminate false correspondences during the training and inference stages, respectively, with only the latter being optimized with pseudo labels. To validate the effectiveness of our registration method, we conducted experiments using the synthetic dataset ModelNet40, the indoor dataset 3DMatch, and the outdoor dataset KITTI.

Learning from a limited number of samples in automatic modulation classification (AMC) has garnered considerable attention. However, existing few-shot AMC works solely focus on single-domain conditions where the training and testing data share the same data distribution, which overlook the potential domain differences. In practice, the complex and variable communication channels, along with different radio frequency (RF) devices, may result in significant data distribution differences, which can be defined as cross-domain conditions. The neglect of such cross-domain conditions may leads to a significant decline in the performance of existing few-shot AMC models. To consider a more general situation, this paper unifies single-domain and cross-domain few-shot AMC into one task, named SaC-FSL. We propose the Paired Samples Relationship Network (PSRNet) as a solution. PSRNet does not require additional network structure design for domain shifts. It distinguishes categories by learning the relationships between sample pairs rather than directly learning the features of samples. To achieve this, we randomly pair the samples to construct different relationships between different classes and domains, and learn these relationships through classification task. Extensive experiments conducted on multiple datasets have demonstrated the superiority of our PSRNet, which can achieve considerable improvements in both single-domain and cross-domain conditions.

Pre-trained vision-language(V-L) models such as CLIP have demonstrated impressive Zero-Shot performance in many downstream tasks. Since adopting contrastive video-text pairs methods like CLIP to video tasks is limited by its high cost and scale, recent approaches focus on efficiently transferring the image-based CLIP to the video domain. A major finding is that fine-tuning the pre-trained model to achieve strong fully supervised performance leads to low zero shot, few shot, and base to novel generalization. Instead, freezing the backbone network to maintain generalization ability weakens fully supervised performance. Otherwise, no single prompt tuning branch consistently performs optimally. In this work, we proposed a multimodal prompt learning scheme that balances supervised and generalized performance. Our prompting approach contains three sections: 1) Independent prompt on both the vision and text branches to learn the language and visual contexts. 2) Inter-modal prompt mapping to ensure mutual synergy. 3) Reducing the discrepancy between the hand-crafted prompt (a video of a person doing [CLS]) and the learnable prompt, to alleviate the forgetting about essential video scenarios. Extensive validation of fully supervised, zero-shot, few-shot, base-to-novel generalization settings for video recognition indicates that the proposed approach achieves competitive performance with less commute cost.

Neural Radiance Fields (NeRF) have been successfully applied in various aerial scenes, yet they face challenges with sparse views due to limited supervision. The acquisition of dense aerial views is often prohibitive, as unmanned aerial vehicles (UAVs) may encounter constraints in perspective range and energy constraints. In this work, we introduce Multiplane Prior guided NeRF (MPNeRF), a novel approach tailored for few-shot aerial scene rendering-marking a pioneering effort in this domain. Our key insight is that the intrinsic geometric regularities specific to aerial imagery could be leveraged to enhance NeRF in sparse aerial scenes. By investigating NeRF's and Multiplane Image (MPI)'s behavior, we propose to guide the training process of NeRF with a Multiplane Prior. The proposed Multiplane Prior draws upon MPI's benefits and incorporates advanced image comprehension through a Swin V2 Transformer, pre-trained via SimMIM. Our extensive experiments demonstrate that MPN-eRF outperforms existing state-of-the-art methods applied in non-aerial contexts, by tripling the performance in SSIM and LPIPS even with three views available. We hope our work offers insights into the development of NeRF-based applications in aerial scenes with limited data.

We study source-free unsupervised domain adaptation (SFUDA) for semantic segmentation, which aims to adapt a source-trained model to the target domain without accessing the source data. Many works have been proposed to address this challenging problem, among which uncertainty-based self-training is a predominant approach. However, without comprehensive denoising mechanisms, they still largely fall into biased estimates when dealing with different domains and confirmation bias. In this paper, we observe that pseudo-label noise is mainly contained in unstable samples in which the predictions of most pixels undergo significant variations during self-training. Inspired by this, we propose a novel mechanism to denoise unstable samples with stable ones. Specifically, we introduce the Stable Neighbor Denoising (SND) approach, which effectively discovers highly correlated stable and unstable samples by nearest neighbor retrieval and guides the reliable optimization of unstable samples by bi-level learning. Moreover, we compensate for the stable set by object-level object paste, which can further eliminate the bias caused by less learned classes. Our SND enjoys two advantages. First, SND does not require a specific segmentor structure, endowing its universality. Second, SND simultaneously addresses the issues of class, domain, and confirmation biases during adaptation, ensuring its effectiveness. Extensive experiments show that SND consistently outperforms state-of-the-art methods in various SFUDA semantic segmentation settings. In addition, SND can be easily integrated with other approaches, obtaining further improvements. The source code is available at https://github.com/DZhaoXd/SND.

Extensive pre-trained visual-language alignment models, such as Contrastive Language-Image Pre-training (CLIP), have demonstrated significant potential for learning representations transferable to domain generation tasks. In hyperspectral image (HSI) classification, a major challenge in deploying such models lies in prompt engineering, which requires particular expertise and substantial time investment. Moreover, existing methods ignore correlation information cross spectral bands. To address these issues, a novel method named learnable prompts-based Transformer (LPFormer) is proposed in this paper. In LPFormer, cross-band correlation information is extracted by self-attention of the transformer, which converted into positional embedding within the transformer framework to obtain the visual features. Subsequently, prompt words are modeled using learnable parameters that turn into efficient expertise. Finally, contrast learning method is used to align visual and textual features. Experimental results on two HSI datasets shows that the proposed LPFormer outperforms other domain adaptation methods.

Remote sensing visual question answering (RSVQA) aims to answer the questions based on the content in remote sensing (RS) images. Due to the complexity of RS images, it is challenging to focus on regions relevant to the questions in the RS images. To this end, we propose a channel-selective multi-scale cross-attention (CSCa) model for RSVQA tasks. Specifically, we design a text-driven multi-scale feature extractor to extract question-related features in RS images. To obtain the cross-attention map in this extractor, we design a novel channel selection mechanism to capture more accurate question-related regions in RS images and develop a channel-wise contrastive learning task to align the semantics between image and text features. We set up experiments on RSVQA-LR and RSIVQA datasets. Experiment results show that our CSCa achieves excellent performance.

Vision Transformer (ViT) has achieved great success in the field of computer vision since it was proposed, and there have been many works applying ViT based models to remote sensing scene classification (RSSC) tasks. The proposal of Pyramid Vision Transformer (PVT) greatly reduces the calculation amount of the ViT while maintaining accuracy. But PVT did not utilize multi-scale information in remote sensing (RS) scenes, which is crucial for RSSC. This paper proposes a multi-scale sparse transformer (MST) based on PVT. MST enables the network to learn multi-scale representations of RS scenes through spatial reduction implementations at different scales. In addition, we employ sparse operations to adaptively guide the model’s attention towards semantically relevant regions during self-attention computation, thereby reducing interference from semantically irrelevant areas. Experiments conducted on the UCM and AID datasets demonstrate the outstanding performance of the proposed MST.

Flood rapid mapping products play an important role in informing flood emergency response and management. To this end, the 2024 IEEE GRSS Data Fusion Contest Track 2 (DFC24-T2) establishes a multimodal benchmark for the segmentation of flood areas from Sentinel-2 multispectral images. However, the problems of imbalanced data distribution, data scarsity, and inter-modal differences severely inhibit the performance of deep-learning-based segmentation networks. In this work, we propose an end-to-end Multimodal Transformer-based Segmentation Network (MTSN) for accurate flood rapid mapping. MTSN first employs two Siamese encoders with shared parameters to accept multimodal inputs and output their respective hierarchical multiscale features, which are then enriched by several channel attention blocks. Subsequently, a Cross-modal Feature Fusion Module (CFFM) based on a gated mechanism is proposed to efficiently integrate the benefits of multimodal features, and generate informative representations. Finally, the fused features are decoded by a lightweight pure multilayer perception decoder to quickly generate mapping results of flood areas. Moreover, we introduce offline data augmentation, semi-supervised learning, test-time augmentation, and multimodal post-process to further boost the performance and generalization of our MTSN. Experimental results and extensive ablations show the effectiveness of our method. Code is available at https://github.com/xiaoqiang-lu/MMSegFormer.

Multi-modal remote sensing image template matching is a meaningful and crucial topic in remote sensing image processing. However, due to different imaging mechanisms, there are significant nonlinear radiometric variations among multi-modal remote sensing images, increasing the matching challenge and leading to poor matching performances. To tackle this issue, this paper proposes a Fourier Domain Adaptive Network (FDANet) for multi-modal remote sensing image matching. Firstly, FDANet randomly swaps the low-frequency spectrum information between multi-modal images through the Fourier transform to reduce differences among multi-modal images, enhancing network adaptability to different image modalities and improving the multi-modal image matching performance. Secondly, FDANet extracts domain-invariant features from the transformed images through a deep Siamese network. After that, FDANet performs template matching and achieves high-precision multi-modal remote sensing image matching. In addition, we adopt the contrastive learning loss to optimize the FDANet. Extensive experiments on multi-modal remote sensing image matching demonstrate the effectiveness and advantages of the proposed FDANet.

The oriented target detection algorithm based on deep learning has made significant progress and has been widely applied in various fields, including remote sensing. However, existing methods still face challenges in large-sized targets and targets with similar backgrounds, leading to unsatisfactory detection performance in these scenarios. To address these issues, this paper proposes two modules on the basis of the feature pyramid: the Multi-Scale Feature Fusion Module and the Feature Compensation Module. The Multi-Scale Feature Fusion Module effectively integrates features from different levels, filters out noise introduced during the fusion process, and allows the network to focus more on target regions. The Feature Compensation Module provides semantic information compensation for the highest-level feature map, enhancing the feature representation capability. Extensive experiments were conducted on the DOTA and DIOR-R datasets. The experimental results demonstrate that the introduction of these two modules significantly improves the detection accuracy of the baseline algorithm.

Instance segmentation in the field of remote sensing imagery is recognized as a complex and difficult task. Previous approaches suffer from inadequate feature fusion, insufficient learning of shape information, and lack of segmentation of object edges. To address these challenges, we introduce FEA-Net(Fusion Edge-Aware Instance Segmentation Network), a multiple information fusion model for remote sensing image instance segmentation. Our model makes the predicted instance masks more accurate and can effectively improve the instance segmentation performance of high-resolution remote sensing images. We have evaluated our method on two datasets, NWPU VHR-10, and the iSAID. The experimental results demonstrate the effectiveness of our method, showing strong performance.

In remote sensing (RS), Few-Shot Novel View Synthesis (FS-NVS) focuses on creating images of unobserved viewpoints using limited training images. Recently, 3D Gaussian Splatting (3DGS) has drawn scholars’ attention by its increasing rendering speeds and providing an explicit neural representation for 3D scenes. However, 3DGS tends to overfit limited training data. To tackle this challenge, we propose a Pseudoview Regularized 3DGS (PR3DGS) FSNVS method for RS scenarios. Our PR3DGS method introduces a pseudo-views regularization module to discriminate synthetic RS images generated from training- or pseudo-viewpoints. Therefore, our PR3DGS method can effectively mitigate overfitting in seen views and enhance the model’s capability to generate more realistic RS images from novel viewpoints. Besides, the excellent experimental results on the LEVIR-NVS dataset demonstrate the effectiveness of our method in RS FSNVS.

Learning object detectors for remote sensing images commonly requires for a huge number of annotated boundary boxes, which are not available without enormous manual efforts in annotating. Alternatively, points can indicate the existence of the objects of interests with reduced labeling cost. Existing Point-supervised object detection (PSOD) methods predominantly employ a two-stage training strategy, which involves propagating point annotations to pseudo boxes at the first stage then training an object detector with these pseudo boxes in a fully supervised manner. However, such paradigm substantially impedes the end-to-end flow of training gradients. In this work, we propose a novel dual-branch network (DBNet) for end-to-end weakly supervised object detection on remote sensing images. Firstly, a pseudo box generation network is attached to the object detector as a sibling branch, which produces semantic response maps for the objects of interest then extracts pseudo boxes by examining their spatial connectivity. Then, instead of training this pseudo box generation network separately, we jointly adjust the pseudo box generation network and the detection network through a multi-task loss. Experimental results on the DOTA-v1.0 dataset demonstrate the effectiveness of our proposed method, achieving an average precision (mAP50) of 32.3%.

Synthetic aperture radar (SAR)iamge change detection is an important and challenging task. Existing methods mainly extract features from the difference map and original images in a parallel structure, ignoring the correlations between features. In this paper, we propose a novel SAR image change detection method based onmulti-scale and co-attention learning, called the Multi-scale Co-attention Network (MCNet). Specifically, we extract features from the difference map and original images using co-attention mechanism, which establishes a relationship between the two kinds of features in the feature space while retaining key information from the original images. With this attention mechanism, the model pays more attention to the changed areas in SAR images, and the features contain richer information. Furthermore, we propose a multi-scale feature fusion method to combine high-level and low-level semantic information, improving the generality and robustness of the features. Finally, the performance of the proposed method has been validated for its effectiveness.

In recent years, as a very important part of remote sensing interpretation, change detection (CD) has developed rapidly with deep learning and remote sensing interpretation. However, most of the existing CD methods focus on how to extract the spatial features of the bi-temporal image pairs, ignore the importance of the temporal features. To solve this problem, a new time-guided network (TG-Net) using temporal information to guide feature extraction is proposed in this paper. In TG-Net, we use a newly proposed time-guided feature fusion (TGF2) block that uses temporal information to guide spatial feature fusion to extract temporal and spatial information comprehensively. We conducted experiments on two publicly available remote sensing datasets, LEVIR-CD and WHU, and compared them with three common CD methods.

Benefiting from the thousands of continuous band information in hyperspectral images (HSIs), the task of HSI classification has become an indispensable part of the field of remote sensing. With the development of deep learning, deep learning techniques such as convolutional neural networks have been widely introduced into HSI classification research. However, most of these methods do not fully consider the potential relationship between the central pixel and surrounding neighborhoods. Therefore, we introduce a novel center mask self-attention network (CMSAN) to enable the model to effectively capture the association between the central pixel and its neighbors for better feature extraction. We conduct experiments on two publicly available HSI datasets. The positive results on both datasets fully demonstrate the effectiveness of our proposed method.

With the development of deep learning, many PolSAR image classification methods based on deep learning have shown impressive performance. Most of these methods rely on manually designed neural networks, which requires a lot of professional knowledge. In response to this issue, this study proposes a complex-valued convolutional neural network architecture search based on evolutionary algorithm for PolSAR image classification. The complex-valued convolutional neural network can extract the complex-valued characteristics of PolSAR images directly. The architecture search for complex-valued convolutional neural network is transformed into an evolutionary population-based optimization, where classification accuracy is the fitness function. Furthermore, a two-steps selection strategy is designed to reduce the architected complexity while ensuring high classification accuracy. The experimental results on two different PolSAR datasets show that the automatic-designed complex-valued convolutional neural network have superior classification performance.

Recent advances in vision-language pre-trained models like CLIP have greatly enhanced general domain image-text retrieval performance. This success has led scholars to develop methods for applying CLIP to Specific Domain Image-Text Retrieval (SDITR) tasks such as Remote Sensing Image-Text Retrieval (RSITR) and Text-Image Person Re-identification (TIReID). However, these methods for SDITR often neglect two critical aspects: the enhancement of modal-level distribution consistency within the retrieval space and the reduction of CLIP's computational cost during inference. To address these issues, this paper presents a novel framework, Accurate and lightweight learning for specific domain Image-text Retrieval (AIR), based on the CLIP. AIR incorporates a Modal-Level distribution Consistency Enhancement regularization (MLCE) loss and a Self-Pruning Distillation Strategy (SPDS) to improve retrieval precision and computational efficiency. The MLCE loss harmonizes the sample distance distributions within image and text modalities, fostering a retrieval space closer to the ideal state. SPDS employs a strategic knowledge distillation process to transfer deep multimodal insights from CLIP to a shallower level, maintaining only the essential layers for inference, thus achieving model light-weighting. Comprehensive experiments across various datasets in RSITR and TIReID reveal that MLCE loss secures optimal retrieval, while SPDS achieves a favorable balance between accuracy and computational demand during testing.

With the advancement of Internet of Things technology, the need for sophisticated signal modulation classification has intensified, ensuring seamless communication and bolstering security among interconnected devices. In the contemporary complex channel environment, the difficult lies in dealing with a multitude of modulation schemes that exhibit subtle distinctions. Prior knowledge-guided and deep learning methods have complementary strengths in the current context of signal modulation classification. To synthesize the advantages of these two methods, we propose an integrated method of prior knowledge and contrast feature for signal modulation classification, called APFS. APFS integrates prior knowledge from the modulation task with feature information acquired through contrastive learning. Feature extraction guided by prior knowledge accurately captures the key patterns in modulated signals. Contrastive learning reveals the inherent distinctions among various modulation modes by comparing different samples. In the joint feature extraction approach for prior knowledge, each form of prior knowledge is first analyzed independently, and then jointed to extract information from its temporal sequence. The contrast features surpass the constraints of labeling and unearth deeper implicit information. In experiments, we systematically compared the performance of our method with various baselines, as well as combinations of prior knowledge and contrast feature. The results demonstrate the superior performance of our method.

Advances in communication technology have enabled signal modulation classification (SMC) to be widely used in noncooperative identification situations, such as spectrum detection, electronic countermeasures, and target identification. In the face of complex electromagnetic environments and various classification tasks, the class imbalance phenomenon in modulated signal data sets has become a problem that cannot be ignored. For the SMC based on machine learning, the unbalanced training data set will cause the actual decision boundary to shift, thereby reducing the prediction accuracy of minority signals. And for SMC based on deep learning, unbalanced data will lead to distortion of the feature space and affect the extraction of discriminative features. However, the existing modulation classification methods cannot effectively deal with the imbalance problem. This study introduces an oversampling method tailored for modulation signals. Our method balances the data set by synthesizing new samples according to the distribution of signal samples and the distance between samples, which will effectively reduce the impact of the imbalance problem on the classifier. For modulated signals, experimental results show that our method performs better than other oversampling methods. In addition to the SMC task, we test the performance of the proposed method for individual identification of radiation sources on the aircraft communications addressing and reporting system data set. Compared with other comparison methods, our method improves the classification performance the most.

Seamless Internet of Things (IoT) connections expose many vulnerabilities in wireless networks, and IoT devices inevitably face many malicious active attacks. automatic modulation recognition (AMR) is an effective way to combat IoT physical layer threats. In the field of noncollaborative communication, feature representation learning for unlabeled signals is an important task of AMR. However, due to the unavailability of a priori knowledge and the influence of interference during signal transmission, the intercepted unlabeled signals are difficult to perform efficient feature representation. In this article, we propose cooperative contrast learning for unlabeled modulation signal Cooperative Contrast Learning for modulation Signals (CoCL-Sig). Specifically, the CoCL-Sig is trained using both sequence and constellation diagram modalities, and is divided into two parts: 1) modal-level feature representation and 2) instance-level auxiliary feature representation. In modal-level feature representation, two modal projections are matched in the same hyperplane space. To ensure the stability of the feature representation, a sequence auxiliary branch is added to form an instance-level feature representation of the sequence. In addition, the feature representations obtained by the CoCL-Sig can be applied to modulation signals for semi-supervised classification and clustering tasks. We have conducted extensive experiments on two widely used modulation signal data sets, RML2016.10A and RML2016.04C. The results demonstrate the effectiveness of our method in modulation signal feature representation and its superiority compared to other methods.

This work delves into unveiling the singularity issue latent in global attention-based Transformers. Empirical and theoretical analyses elucidate that interrelationships among token channels lead to singularities, impeding the training of attention weights. Concretely, the similar neighbor pixels within image patches can form intercorrelated channels after being flattened. Images that one color dominates can possess correlated channels . Furthermore, the fixed global connection architecture retains correlation relationships, contributing to the persistence of singularities. High singularity risks reducing Transformers’ performance and robustness. Based on the singularity analysis, we propose the Token Singularity Removal (TSR) strategy. It incorporates the Dual-Tree Complex Wavelet Transform (DTCWT) stem and Feature Decorrelation (FD) loss, aiming to encourage Transformers to learn tokens with unrelated channels and eliminate singularities. Experimental validation across various image classification datasets and corruption image data sets demonstrate improved accuracy and robustness of Transformers utilizing the TSR strategy. Our code is publicly available at https://github.com/wdanc/TSR .

The visual question answering (VQA) method applied to remote sensing images (RSIs) can complete the interaction of image information and text information, which avoids professional barriers in different RSIs processing fields. The current methods face challenges in both fully utilizing the global and local information of the image to interact with the question information and addressing the issue of inter-class interference. To address these challenges, this paper proposes a remote sensing visual question answering (RSVQA) model based on the mask-based dual-stream feature mutual attention network (MADNet). First, the dual-stream feature extraction module of the image is used to obtain image features, and the deep and shallow layer feature encoding module is used to obtain question features. Second, the attention mechanism is introduced and combined with the pointwise multiplication method to utilize of the dual-stream features that were extracted in the earlier step. Finally, an answer relevance modulation module based on a binary mask vector is implemented to filter out irrelevant answers. In the experiments, the performance of the proposed strategy is evaluated using two datasets collected by aerial and Sentinel-2 sensors. In our study, we propose a model that outperforms previous approaches, achieving a 6.89% increase in overall accuracy (OA) over the baseline. This enhancement is notable for its persistence, even when the training data is reduced by half, as evidenced by our experiments on the low-resolution dataset.

Significant progress has been achieved in the development of oriented target detection algorithms based on deep learning, which have found widespread application in various fields, including remote sensing. However, existing methods struggle with adjusting model size and often exhibit unsatisfactory detection performance for targets that overlap, are large, or have similar backgrounds. To address these challenges, this letter proposes an oriented target detection algorithm called Oriented you only look once X (YOLOX), which integrates several optimization techniques. Specifically, to meet the requirements of oriented detection while enhancing feature extraction, we introduce a new network architecture that includes an orientation detection branch and a multiscale feature fusion module (MSFFM). An MSFFM based on attention weights is proposed to integrate features across scales while minimizing noise. In addition, to mitigate the impact of the number of positive samples on the original loss function and focus the network’s attention on learning challenging targets, an object-aware reweighted loss function is introduced in this study. This approach dynamically adjusts the loss contribution for each target. Two models of different sizes are developed using the Oriented YOLOX scaling strategy to cater to scenarios prioritizing either accuracy or speed. Extensive experiments on the dataset for object detection in aerial images (DOTA) and object detection in optical remote sensing images (DIOR-R) datasets demonstrate that Oriented YOLOX performs better in detecting challenging targets. Compared with other oriented target detection methods, this approach not only achieves higher detection accuracy but also reduces parameter counts, improving inference speed.

Hyperspectral images provide plentiful latent information that requires exploration for ground object recognition, where self-supervised learning is efficient and independent of manual labeling. However, the severe spectral uncertainty poses a significant challenge in discriminative and generalizable representation by self-supervision. This letter proposes a variational generative transformer with momentum contrastive supervision (ConVaT) to alleviate the problem. ConVaT contains two branches: a variational generative branch and a contrastive learning branch—the former guides informative data representation via an encoder-decoder transformer with variational inference; the latter encourages the representation with discriminability by distinguishing positive anchors from negative ones. Significantly, to facilitate a more generalizable latent representation, we reconstruct data with reparameterized tokens sampled multiple times from the global anchor, instead of the latent representation of unmasking data. Extensive experiments on three public datasets show that ConVaT is superior in data representation with intra-class clustering and inter-class distinction, and it achieves considerable improvements over present methods under linear probing, especially for the Indian pines dataset with intense spectral uncertainty. Our code will be available at https://github.com/liuzuo-byte/ConVaT.

In previous works, powerful CNN backbones are typically used for one- or two-stage detectors to facilitate multi-categories object classification. Unfortunately, continuous convolution and pooling operations tend to weaken the detailed information. We propose an end-to-end Multi-content Collaboration Detector (MCDet) to improve object recognition accuracy. First, we summarize the reasons for the disappearance of detailed features in traditional feature extraction backbone networks, and propose a Shallow Clue Refinement (SCR) module, which helps us to retain more critical local detail information in the downsampling process. Second, to receive more suitable contextual information, we design a Self-dilating Spatial Pooling (SSP) module, it adaptively learns a contextual reception field, thereby alleviating the mismatch between the theoretical receptive field of the network design and the practical requirements. Finally, extensive experiments on the NWPU VHR-10 and DIOR datasets have shown that the proposed MCDet significantly improves detection accuracy. Our code is available at https://github.com/Xidian-AIGroup190726/RS-objectdetection-MCDet.

Land cover classification (LCC) is an important application in remote sensing data interpretation. As two common data sources, SAR images can be regarded as an effective complement to optical images, which will reduce the influence caused by single-modal data. But common LCC methods are focusing on designing advanced network architectures to process single-modal remote sensing data. Few works have been oriented toward improving segmentation performance through fusing multi-modal data. In order to deeply integrate SAR and optical features, we propose SwinTFNet, a dual-stream deep fusion network. Through the global context modeling capability of Transformer structure, SwinTFNet models teleconnections between pixels in other regions and pixels in cloud regions for better prediction in cloud regions. In addition, a Cross-Attention Fusion Module (CAFM) is proposed to fuse features from optical and SAR data. Experimental results show that our method improves greatly in the classification of clouded images compared with other excellent segmentation methods and achieves the best performance on multi-modal data.

Model-based target decomposition method has been widely applied due to its clear physical scattering significance. However, after establishing decomposition basis, the process of solving the scattering components and parameters is usually underdetermined, which will lead to the issues such as component negative power and overestimation. For this problem, this letter examines the target decomposition task from the perspective of deep learning and proposes an adversarial decomposition feature learning (ADFL) model. This model could learn decomposition features suitable for current terrain characteristics according to input data. At the same time, the model-based adversarial feature prior is embedded in ADFL to maintain the physical scattering meanings. On real PolSAR datasets, the learned features of proposed model are well correlated with real terrain scattering characteristics. Further, it avoids negative decomposition features and make more accurate fitting of scattering components, effectively alleviating the above problems.

Deep learning-based approaches play important roles and achieve impressive performances in remote sensing image change detection. Most of the networks are designed by the researchers with rich experiences. It is difficult to design the fixed networks with universal and good performance on various datasets. Regarding this issue, this study presents a transformer architecture search for remote sensing image change detection. The transformer architectures can be designed automatically by a two-stage transformer architecture search. The first stage can search for the effective combinations of attention mechanisms, while the second stage can identify the suitable combinations of multiscale modules. A gradient-based optimization method is employed for enabling the stable and efficient transformer architecture search. The experiments on various change detection datasets can verify the effectiveness of this work.

Deep neural network models based on vision transformer (ViT) have shown unprecedented performance in the field of remote sensing image object detection. However, those models often require massive training data, which cost a lot of time to train and greatly prevent the research progress. Distributed training is a common way to accelerate the training period. In this letter, we propose a large batch object detector named LargeRSDet for remote sensing image object detection task, which can train with a batch size up to 1024 with only a little acceptable performance loss. Using the LargeRSDet, we can effectively utilize at most 1024 GPUs and greatly improve the training speed, which enables several benefits that not only help our model converge in a faster way but also provide the ability to reach a higher accuracy. Experimental results demonstrate that our method can finish training DIOR remote sensing image dataset in less than 5 min, and finally, the model achieves 75% mAP at 0.5.

Global shapes and local joints are significant cues to learn skeleton representations for human action recognition. However, most shapes-based methods ignore local joints information and thus perform less well when recognizing similar actions. In this letter, we propose a two-stream method of global shapes and salient joints features learning to address the above issues. Firstly, the global sparse shapes modeling stream (GSSM) explores the global skeletal shapes in Kendall space and codes the nonlinear shapes by a sparse coding and dictionary learning method. Then, the salient enhancing joints generation stream (SEJG) explores local joints information by extracting discriminative salient joints and generates frame-enhancing action sequences with invariant length by a bilinear frame interpolation module. To improve the temporal modeling ability of our method, we use similar multi-scale LSTM methods in both GSSM and SEJG, which explore global-local features in static, short-term, and long-term temporal scales. Extensive experiments on four challenging datasets verify the effectiveness of our proposed method, which achieves competitive results compared to state-of-the-art (SOTA) methods.

As jamming systems continue to evolve, jammers are becoming increasingly complex. They modify their strategies in response to changes in radar status, thereby creating a dynamic confrontation with the radar. Currently, deep reinforcement learning (DRL) methods are widely employed in the decision-making process for radar anti-jamming strategies; however, most of these methods overlook the learning capabilities of the jammer. This letter proposes a novel multi-agent deep reinforcement learning (MADRL) method that leverages the robust capabilities of the neural fictitious self-play (NFSP) method and extensive-form game (EFG) frameworks to address multi-agent interactions and partial information observability challenges. The letter primarily discusses the model design, which includes the construction of the EFG framework, hierarchical optimization of the action space, and the design of the reward function. Comparative results with other state-of-the-art methods demonstrate that this approach achieves significant accuracy and rapid convergence in anti-jamming decision-making, thereby effectively enhancing the intelligent anti-jamming capabilities of radar systems.

Object Tracking in satellite videos is a challenging task due to the small target size, low spatial resolution, limited appearance and texture information, and the potential for background confusion. While current state-of-the-art tracking methods perform well on natural images, they often produce unsatisfactory results when applied to satellite videos. In this paper, we address these challenges by leveraging location prompts and refining the feature extractor and bounding box refinement module. Furthermore, we integrate motion features to effectively handle illumination variations that frequently arise in satellite videos, thereby enhancing the overall robustness of the tracker. Our proposed approach, abbreviated as SVLPNet, has been thoroughly evaluated through extensive experiments conducted on two authentic satellite video datasets. The obtained results unequivocally showcase the promising potential of SVLPNet in facilitating object tracking on satellite videos. The source code and raw results will be released at https://github.com/Wprofessor/SVLPNet.

Cross-domain classification of hyperspectral data is a critical challenge in remote sensing, especially when labels are unavailable in the target domain. Deep learning-based domain adaptation (DA) methods have been widely used in recent years. However, curren methods primarily focus on the global domain structure of the source and target domains when considering domain adaptation, neglecting the subdomain structure within each class. Additionally, current methods directly employ predicted outputs without further exploring the confidence level of the target domain samples. These limitations lead to confusion in domain adaptation and hinder effective feature selection in neural networks. In this paper, we propose the Pseudo-Label-Assisted Subdomain Adaptation (PASDA) method, which addresses these limitations by jointly considering the subdomain structure of the source and target domains and adopting a sample selection strategy. PASDA aligns the subdomains while learning domain-invariant features as a foundation. Furthermore, it selects high-quality pseudo-labeled samples from the target domain to enhance the learning of domain-invariant features. For generating pseudo-labels in the target domain, we employ the Reweighted Pruning Label Propagation (RPLPA) strategy to reweight the output of the predicted target domain. Finally, the high-confidence samples with pseudo-labels are selected to finetune the network. The entropy regularized dual classifier constraint is introduced to enhance the discriminative feature extraction ability for the target domain. Extensive experiments on three public HSI cross-domain datasets, Pavia, Houston, and HyRANK, using overall accuracy (OA), average accuracy (AA) and kappa coefficient (Kappa) as the evaluation indicators of classification performance, demonstrate the superiority of our method. Compared with the existing state-of-the-art (SOTA) unsupervised domain adaptation (UDA) methods, our method improves OA by 2% and AA by 4%.

Satellite video object tracking involves tracking a specified tiny object within a wide scene. The insufficient appearance features of these tiny objects pose significant challenges to appearance-based object trackers, particularly in situations involving occlusion, target blur, and similar interferences. In this paper, a novel Graph Association MOtion-aware tracker (GAMO) is proposed for tiny object in satellite videos, which integrates motion and spatial relationship information. First, a Gaussian motion estimator is proposed that decouples motion into velocity and direction, rather than using traditional x-y movement modeling. This estimator predicts the object’s position and estimates motion uncertainty with a directional motion probability map. Furthermore, the estimated motion serves as a prior to guide the proposal sampling. A probabilistic proposal sampling module is designed that samples candidate bounding boxes according to the directional motion probability map, focusing on the region where the target is most likely to appear. Additionally, we implement a graph association module to model and propagate the spatial relationships between the target and neighboring objects over time. This relationship information assists the appearance features in distinguishing the target from similar interferences. Experiments on the Skysat-1, SV248S, and VISO datasets demonstrate the superiority of the proposed tracker. GAMO leverages motion and surrounding information, resulting in significant improvements with minimal computational overhead. The code and results will be publicly available inhttps://github.com/Midkey/GAMO.

With the development of multimedia technology, events are usually presented in multimedia forms, thus multimedia event extraction (MEE) has become more and more important. Existing MEE works usually use simple strategies to align two modalities, making it difficult to precisely extract events and arguments in complex multimedia documents. To address this problem, we propose a novel Multi-grained Gradual Inference Model (MGIM) that focuses on inferring and interpreting events in complex multimedia structures in a coarse-to-fine manner. To efficiently integrate textual and visual modalities, we design a Coarse-grained Alignment (CA) module, which represents the two modalities in a graph structure and performs coarse-grained alignment. Based on the CA module, we further propose a Fine-grained Inference module (FI) that fine-grained aligns text and image by performing multiple rounds of gradual inference. MGIM provides a comprehensive interpretation of multimedia events at two information granularities (coarse and fine). Extensive experiments on the M2E2 dataset demonstrate the effectiveness of MGIM.

Recently, semi-supervised semantic segmentation methods based on weak-to-strong consistency learning have achieved the most advanced performance. The key to such a technique lies in strong perturbations and multi-objective co-training. However, CutMix, the most commonly used data augmentation in this field, limits the strength of perturbations as it only focuses on single random local context. Besides, complex optimization targets also reduce computational efficiency. In this work, we propose an efficient consistency learning based framework. Specifically, a novel unsupervised data augmentation strategy, EntropyMix, is present for semi-supervised semantic segmentation. Patches of unlabeled data from multi-view augmentations are combined into new training samples based on their prediction entropy, which provides more informative and powerful perturbations for consistency regularization and impels the model to focus on cross-view local context. On this basis, we further propose Self Pseudo Entropy Knowledge Distillation (SPEED) to learn global pixel relations from multi- and cross-view perturbations by optimizing a linear combination of feature-and logit-level distillation loss, enhancing model performance without additional auxiliary segmentation heads or a complex pre-trained teacher model. The collocation of the two ideas above is a plug-and-play technique without additional modification. Extensive experimental results on PASCAL VOC and Cityscapes datasets under various training settings demonstrate the superiority of the proposed data augmentation strategy and self-distillation loss, achieving new state-of-the-art performance. Remarkably, our method reaches mIoU of 75.16% using only 0.87% labeled data on PASCAL VOC and mIoU of 76.98% using only 6.25% labeled data on Cityscapes. The code is available at https://github.com/xiaoqiang-lu/SPEED.

A growing number of earth observation satellites are able to simultaneously gather multimodal images of the same area due to the expanding availability and resolution of satellite remote sensing data. This paper proposes a novel multimodal balanced self-learning interaction network (MBSI-Net) for the classification task. It involves a dual-branch teacher-student network that enables knowledge interaction and transfer between the multimodalities. Firstly, in order to introduce statistical information in addition to local and global structural information, a texture feature equalization module (TFE-Module) is proposed. This can enhance the texture information of features through histogram equalization and further improve the representation ability of features. Secondly, to enable the student network to provide timely feedback questions, the paper proposes a feature fusion module (F2-Module) that models and enhances teacher features through the student network. This helps to raise the classification’s accuracy by incorporating information from multimodal images. Finally, the paper proposes a loss function based on structural similarity analysis to ensure balanced self-learning between the student and the teacher networks. Taking the multispectral (MS) and the panchromatic (PAN) images of the same scene as examples, through experimental verification, the proposed method can achieve good results on multiple datasets compared with other methods. Therefore, it offers an effective method for classifying and fusing multimodal data.

The abundant spectral signatures and spatial contexts are effectively utilized as the key to hyperspectral image (HSI) classification. Existing convolutional neural networks (CNNs), only focus on locally spatial context information and lack the ability to learn global spectral sequence representations, whereas the transformer performs well in learning the global dependence of sequential data. To solve this issue, inspired by the transformer, we propose an interactive global spectral and local spatial feature fusion transformer called ISSFormer. Specifically, we achieve an elegant integration of self-attention and convolution in a parallel design, i.e., the multi-head self-attention mechanism (MHSA) and the local spatial perception mechanism (LSP). ISSFormer can learn both local spatial feature representation and global spectral feature representation simultaneously. More significantly, we propose a bi-directional interaction mechanism (BIM) of features across the parallel branch to provide complementary clues. The local spatial features and the global spectral features interact through the BIM which could emphasize the local spatial details and add spatial constraints to overcome spectral variability, and can further improve classification performance. With extensive experiments on three benchmark datasets, including Indian Pines, Pavia University, and WHU-Hi-HanChuan, ISSFormer can accomplish superior classification accuracy and visualization performance.

Despite the extensive research on RGBT object tracking, there are still several challenges and issues in practical applications, such as modality differences, lighting variations and disappearance of the target, and changes in viewpoint. Existing methods mostly address these issues by fusing image features, while neglecting a significant amount of target label information. To address these challenges, this paper introduces text to drive the alignment of visible and infrared image features, transforming features from different modalities into the same feature space and fully using complementary features between different modalities. Furthermore, inspired by the success of prompt learning in various tasks, we utilize prior boxes and language as prompts to further guide the model in tracking the target. Extensive experiments demonstrate that the proposed VLCTrack tracker has excellent potential in RGBT object tracking. Compared to previous methods developed for this purpose, our approach achieves state-of-the-art performance on three benchmark datasets.

Visual object tracking in natural scenes is a popular but challenging task, owing to the difficulties of feature representation from various changes of the targets, such as size change, deformation, illumination change, rotations, motion blur, background clutter, etc. High-speed hyperspectral imaging systems capture hyperspectral videos (HSVs) in wide spectral ranges and provide abundant spectral and spatial information to tell targets apart from backgrounds, alleviating the model drift in appearance-based tracking methods. However, different hyperspectral imagers, such as near-infrared (NIR), red-to-near-infrared (RedNIR), and visible (VIS), obtain heterogeneous types of data that could not be handled by common object trackers. In this paper, a domain adaptive Transformer framework is proposed for hyperspectral object tracking. Considering the HSVs are from different types of sensors, their heterogeneous features are learned in an adversarial way by domain label reverse learning with a gradient reversed layer. To fully utilize the spectral information in HSV frames, a band-wise spatial attention module (BSAM) is designed to emphasize the salient area near the target of interest. We adopt a Siamese-like Transformer tracker as the main structure for tracking. Our tracker outperforms top-ranking methods on a hyperspectral object tracking benchmark dataset containing three types, 87 hyperspectral videos in total. The comparison experiments validate the effectiveness of the proposed method. The source code and trained models of this work will be publicly available soon at https://github.com/LianYi233/Trans-DAT.

Existing wavelet pooling methods discard the high-frequency sub-bands, which can improve the noise-robustness of convolutional neural networks (CNNs) but lose the essential detailed features. Besides, most of them depend on different wavelets, which is not adaptive. In this paper, a novel efficient lifting-based wavelet pooling (LWPooling) is proposed to alleviate the problems above. Firstly, wavelet pooling is rethought based on the equivalence of 2D discrete wavelet transform (DWT) and standard average pooling (SAP), which suggests the lack of detailed information on traditional wavelet pooling. Secondly, the efficient LWPooling module is proposed to adaptively capture and preserve the critical high-frequency features via lifting-based wavelets. It can constrain the features linear independence, which efficiently makes important features salient. Thirdly, the lifting-based wavelet collaborative network (LWCNet) is constructed for classification and segmentation tasks based on the efficient LWPooling module. Experiments are validated on Cifar10, Cifar100, and ADE20K datasets. It suggests that the efficient LWPooling can enhance CNN’s representation and achieve a particular performance advantage compared to average, maximum, and original wavelet pooling. Besides, the proposed LWCNet shows the potential for scene parsing. The code implementation will be available at https://github.com/yutinyang/LWCNet.

Solving the complex challenges of sophisticated terrain and multi-scale targets in remote sensing (RS) images requires a synergistic combination of Transformer and convolutional neural network (CNN). However, crafting effective CNN architectures remains a major challenge. To address these difficulties, this study introduces the knowledge guided evolutionary Transformer for RS scene classification (Evo RSFormer). It amalgamates adaptive evolutionary CNN (Evo CNN) with Transformers in a hybrid strategy synergistically, which combines fine-grained local feature extraction of CNNs with long-range contextual dependency modeling of Transformers. Furthermore, for the development of Evo CNN blocks, this paper presents a knowledge-guided adaptive efficient multi-objective evolutionary neural architecture search (MOE2-NAS) strategy. This approach markedly diminishes the labor-intensive characteristics associated with traditional CNN design, striking a balance for both accuracy and compactness. Additionally, by leveraging domain knowledge from natural scene analysis into the RS field, MOE2-NAS facilitates the efficiency of classical NAS. It utilizes a priori knowledge to generate promising initial solutions and constructs a surrogate model for efficient search. The effectiveness of the proposed Evo RSFormer has been rigorously tested on various benchmark RS datasets, including UC Merced, NWPU45, and AID. Empirical results strongly support the superiority of Evo RSFormer over existing methods. Furthermore, experiments on MOE2-NAS have been studied to confirm the important role of knowledge guidance in improving the efficiency of NAS.

Vision transformers (ViTs) are rapidly evolving and are widely used in computer vision. However, high-performance ViTs require many computations, which limit their further development in the vision field. In this article, a novel evolutionary dual-stream transformer (E-DST) model is proposed to alleviate the computational resource demand problem. A hybrid attention mechanism structure is proposed for a DST model. The DST model uses a dual-branch structure to fuse convolutional and transformer features. Combining the features learned by the transformer and convolution effectively saves model computational resources. In addition, an evolutionary optimizer is proposed to optimize the parameters of the model. The excellent search ability of the evolutionary algorithm is utilized to optimize the transformer model parameters. The convergence of the evolutionary optimizer is proved in this article. In addition, the proposed E-DST model is experimentally compared with a variety of classic models and their deformations based on three datasets. And, the evolutionary optimizer proves its generality in convolutional and recurrent neural networks. The experimental results show that the E-DST model can effectively reduce computational resources and that the evolutionary optimizer can solve large-scale optimization problems. In conclusion, our proposed method is feasible and effective.

The construction of machine learning models involves many bi-level multiobjective optimization problems (BL-MOPs), where upper-level (UL) candidate solutions must be evaluated via training weights of a model in the lower level (LL). Due to the Pareto optimality of subproblems and the complex dependency across UL solutions and LL weights, a UL solution is feasible if and only if the LL weight is Pareto optimal. It is computationally expensive to determine which LL Pareto weight in the LL Pareto weight set is the most appropriate for each UL solution. This article proposes a bi-level multiobjective learning framework (BLMOL), coupling the above decision-making process with the optimization process of the upper-level MOP (UL-MOP) by introducing LL preference$\boldsymbol {r}$. Specifically, the UL variable and$\boldsymbol {r}$are simultaneously searched to minimize multiple UL objectives by evolutionary multiobjective algorithms. The LL weight with respect to$\boldsymbol {r}$is trained to minimize multiple LL objectives via gradient-based preference multiobjective algorithms. In addition, the preference surrogate model is constructed to replace the expensive evaluation process of the UL-MOP. We consider a novel case study on multitask graph neural topology search. It aims to find a set of Pareto topologies and their Pareto weights, representing different tradeoffs across tasks at UL and LL, respectively. The found graph neural network is employed to solve multiple tasks simultaneously, including graph classification, node classification, and link prediction. Experimental results demonstrate that BLMOL can outperform some state-of-the-art algorithms and generate well-representative UL solutions and LL weights.

Video object tracking has been a popular area in the field of computer vision. As video data evolves, more special perspectives and challenging video data are constantly kept up to date. This poses challenges for object tracking tasks and places higher demands on the generalization capabilities of the models. In this article, we propose a novel quantum evolutionary learning tracker (QELT) for video. The model combines quantum evolution with deep networks for tracking video objects. The model uses a QELT to generate a reliable population of candidate regions and a deep network for classification. In particular, the quantum evolutionary predictor predicts the object motion state through rotation operator and trajectory inference, and provides motion state information for the tracker. The predictor can incorporate object history contextual information and can provide stable candidate estimation populations for the model in case of failure of appearance features. Both quantum evolution and deep networks are combined to form an end-to-end online video object tracker. In addition, we propose a new video object tracking evaluation algorithm, Balanced Intersection over Union. The evaluation algorithm uses aspect ratios to balance the share of overlap and distance. Finally, we test the model on the OTB 2015 dataset for natural video and on the SV248A10-SOT dataset for satellite video. The performance of the proposed model is also analyzed and validated by comparing it with more than 20 classical tracker models. The experimental results show that our model has high generalization ability and robustness.

With the development of deep learning and the increase in the amount of data, general artificial intelligence models have become a popular research area nowadays. When facing a new application scenario, a pretraining general model can often show better performance than models trained with new data on its own. However, because of the specificity of the differences in hyperspectral image data bands, the current hyperspectral image classification (HSIC) field has not proposed a better general model training solution, and it is difficult to utilize the information of the existing hyperspectral datasets for model training in the face of a new scenario. In order to solve this problem, this article proposes a generalized hyperspectral classification model training method, which effectively completes the training of hyperspectral classification models across datasets by adaptive channel module and masked self-supervised pretraining method, and can pretrain and fine-tune hyperspectral classification models using multiple datasets. The adaptive channel module is able to solve the band difference problem of using hyperspectral datasets across datasets, and the masked self-supervised learning method solves the label difference and labeling difficulties of training models across datasets. Experimental results on multiple datasets show that the method proposed in this article can effectively use a large amount of data to complete the pretraining of hyperspectral classification models, and the fine-tuning results on downstream datasets have certain advantages relative to current advanced deep learning methods.

Recently, self-supervised learning (SSL) has gained great prominence in hyperspectral image classification (HSIC) due to its powerful capability to alleviate data-hunger problem. The generative-based method and the contrastive-based method have become two main streams in the field of SSL. To fully integrate the merits of both of them, we propose an efficient hybrid SSL method, that is, mask-enhanced contrastive learning (MECL). Essentially, MECL remains a prototypical contrastive-learning (CL) method, but incorporates the masking-and-predicting idea of the generative-based method. When the sample gradually matches the prototypes, feature reconstruction is implicitly performed in MECL step by step. Furthermore, we design a spatial-spectral multimasking mechanism for hyperspectral data and also propose two complementary strategies to prevent MECL from collapsing. To demonstrate the performance of MECL, a wide range of experiments are carried out in our work. On the one hand, the results of ablation experiments prove that HSIC does benefit from the fusion of SSL methods. On the other hand, the classification results on three hyperspectral datasets confirm that our MECL is an effective SSL method with high stability and strong reliability.

Deep learning methods have gradually developed into the mainstream methods of moving vehicle detection in satellite videos. However, these methods require labor-intensive and time-consuming box-level annotations to predict accurate locations and sizes, which is challenging for large-scale satellite video datasets with hundreds of vehicles. To address this problem, a novel coarse-to-fine dynamic refinement framework (CFDRM) is proposed for moving vehicle detection in satellite videos only under the supervision of point-level annotations. CFDRM generates initial proposal boxes and performs spatio-guided matching with point annotations to obtain coarse box-level pseudo annotations. The initial priority of these coarse annotations is calculated by leveraging locally-consistent prior tailored to satellite videos. Then, a dynamic refinement detector is constructed to transfer coarse annotations to fine annotations with prior and predictive collaborative curriculum refinement. During the curriculum learning process, the coarse annotations are sequentially learned with a certain priority, where the priority is inferred by considering the prior knowledge from the locally-consistent prior and the knowledge itself from the predicted detector. Ultimately, a novel ambiguity-aware loss is designed to optimize the dynamic refinement detector from coarse annotations to fine annotations in an adaptively-weighted fashion. Extensive experiments have been conducted on the Jilin-1 and SkySat satellite video datasets demonstrate the superiority of CFDRM.

The task of hyperspectral image (HSI) classification is fundamental and crucial in HSI processing. Currently, domain adaptive methods have become a research hotspot in HSI classification. However, most domain adaptive methods ignore the class alignment in different domains. Additionally, HSIs have the characteristics of category imbalance and complex spatial-spectral distribution, which restricts the adaptation performance in HSIs. To address these problems, a class-aligned and class-balancing generative domain adaptation (CCGDA) method is proposed for HSI classification. The architecture of CCGDA is designed by using the classifier, domain discriminator, sampler and two weight-sharing generators. In the classifier, split-level capsule network is constructed by extracting rich spatial information of shallow layer and spectral features of deep layer with equivariant characteristic. Then, the classifier provides the pseudo label of samples in the target domain. To prevent the generators from mode collapse caused by category imbalance, the sampler is designed. It samples and re-samples the samples of the target domain in an adaptive proportion according to the statistical calculation through confidence and distribution of pseudo labels. Finally, a novel class-aligned domain adversarial loss is defined to jointly optimize the generators and discriminator. It incorporates the class shift adjusting and adaptive sampling for the samples of the target domain to better adapt the discriminant boundary of the classifier to the target domain. Experiments on benchmark HSI datasets verify the superiority of the proposed method for domain adaptive classification.

The task of fine-grained aircraft recognition is crucial in the field of remote sensing. Despite some progress achieved by traditional deep learning methods in addressing this challenge, they are often perceived as a “black box,” lacking transparent explanations for model decisions. Current interpretable methods based on attention mechanisms, although providing some interpretability, do not align with human thought logic. Therefore, we propose a multiscale rotation-invariant prototype network (MSRIP-Net). Our approach simulates the intuitive reasoning process of humans in identifying objects by segmenting them into multiple components. Importantly, MSRIP-Net has the capability to automatically recognize rigid components on aircraft targets without relying on additional part annotations, using only image-level class labels. In addition, our approach effectively addresses challenges presented by noise, deformations, and multiscale variations in remote sensing targets and has been comprehensively evaluated on datasets FAIR1M1.0 and Rareplane. Our results demonstrate that MSRIP-Net achieves higher accuracy compared with existing fine-grained recognition methods. Furthermore, we provide insights into the model’s decision-making process to illustrate the interpretability of our approach.

To overcome the inherent domain gap between natural images and remote sensing images (RSIs), it is highly desirable to develop pretraining methods specifically for RSIs. Considering the lack of widely recognized large-scale benchmarks like ImageNet in the RSI community and limited computational resources, this article proposes multilayer perceptron (MLP)-guided convolutional neural network (CNN) (MGC), a method that employs an MLP to guide the pretraining of a CNN from small-scale datasets for RSIs. MGC has two encoders, each consisting of a CNN branch and an MLP branch. We first contrast pairwise samples from the same type of branches or different types of branches across the encoders and employ a positive-pair guidance strategy to explore consistency. Due to the inherent locality issue of shallow layers in a CNN, the CNN branches often do not attend to correct foreground regions such as objects, regions of interest, and land coverage. Therefore, we further propose an attention guidance strategy to guide the CNN branches to focus on foreground regions and learn discriminative representations effectively. The proposed MGC method is validated by pretraining a CNN model using the MGC and applying it to different downstream tasks including scene classification, rotated object detection, semantic segmentation, and change detection on ten datasets. Results have confirmed the effectiveness of the proposed MGC. Our code will be released at:https://github.com/benesakitam/MGC.

Aerial image objects are usually orientated arbitrarily, with a large scale range, and densely distributed. Traditional horizontal bounding box (HBB) detectors tend to filter out densely distributed objects leading to missed detections, such as ship (SH) and vehicle. Therefore, oriented object detection has become a mainstream solution in recent years. The 2-D Gaussian distribution representation of the oriented bounding boxes (OBBs) solves the problem of angular discontinuity and boundary discontinuity and thus gets more attention. However, as the aspect ratio of the object gradually decreases, its predicted angular performance continues to decrease. We find that the angular gradient of an object decreases sharply as the aspect ratio decreases, resulting in a large gradient gap between a small aspect ratio object (SARO) and a large aspect ratio object (LARO). It makes the detector prefer to ignore SARO during training, which weakens the high precision performance of SARO. We call this phenomenon shape imbalance. To solve the problem, we proposed a simple gradient rebalancing strategy named shape balance. Since the shape imbalance is only related to the aspect ratio of the object, we designed a modulation function with an inverse aspect ratio to calculate the balance coefficient. The principle of the function is that the larger the aspect ratio, the smaller the balance coefficient; the smaller the aspect ratio, the larger the balance coefficient. We aim to get the balance coefficients for objects with different aspect ratios. Location loss multiplied by a balance coefficient can directly adjust the gradient gap between objects with different aspect ratios to achieve a rebalancing effect. Extensive experiments conducted on DOTA-v1.0 dataset and DIOR-R dataset verify the effectiveness of our proposed method. Our method improves the detection performance of Gaussian location loss by an average of 2.08%/1.01%(AP75/mAP) metrics on the DOTA-v1.0 dataset and 1.17%/0.82%(AP75/mAP) improvements for DIOR-R dataset.

Few-shot segmentation aims to segment specific objects in a query image based on a few densely annotated images and has been extensively studied in recent years. In remote sensing, image segmentation faces challenges such as less training data, large intraclass diversity, and low foreground-background contrast. In this work, we propose a novel few-shot segmentation method in remote sensing imagery based on mask-guided correlation learning (MGCL) to alleviate the above challenges. In our MGCL, a novel mask-guided feature enhancement (MGFE) module is proposed, which makes features have intramask consistency by leveraging oversegmented masks. In order to enhance the contrast between foreground and background, a novel foreground-background correlation (FBC) module is proposed, which enhances background correlation representation by learning foreground correlation and background correlation separately. Furthermore, a novel mask-guided correlation decoder (MGCD) module is proposed to guide the decoder to focus on the consistency within the mask, thereby learning how to segment complete objects and improving segmentation accuracy. Sufficient experiments on the iSAID-$5^{i}$and DLRSD-$5^{i}$datasets show that our MGCL outperforms all comparative methods. In particular, in the one-shot setting of the iSAID-$5^{i}$dataset, we achieve an mIoU of 39.92 based on ResNet50, which is an improvement of 4.25 over the state-of-the-art (SOAT) method. The visualization of features before and after the MGFE module further concretely demonstrates the motivation and advantages of our MGCL. The code is available athttps://github.com/LiShuo1001/MGCL.

Remote sensing image change captioning (RSICC) is more challenging than remote sensing change detection task, which requires extracting occurred changes in similar remote sensing image (RSI) pairs while generating change caption. However, few works have been investigated on RSICC, the main challenges come from how to learn abundant change clues and face the modality gap. To handle these problems, we rethink this task from the perspective of obtaining and aligning symmetrical change features for temporal RSIs. In this work, the proposed intertemporal interaction and symmetric difference learning network are cascaded through several multitemporal integration units to model differences from coarse to fine representations. Specifically, we design a cross-temporal attention (CTA) mechanism to probe direct interaction between bi-temporal RSIs for motivating information coupling between intralevel representations and suppressing irrelevant interferences. To learn robust change features, a symmetric difference transformer (SDT) module is devised to guarantee temporal symmetry between the “before-to-after” and “after-to-before” change representations. Besides, the bi-directional triplet ranking loss is adopted to guide the network to learn strongly discriminative and temporal-symmetric change representation. Extensive experimental results on Dubai-CC and LEVIR-CC datasets demonstrate that our framework with the proposed components can achieve excellent performance and surpass recent state-of-the-art methods.https://github.com/romanticLYP/TISDNet

Deep learning models (algorithms) have demonstrated their superior performance in interpreting Earth science and remote sensing data. However, adversarial examples generated with perturbations imperceptible to humans could render deep learning algorithms ineffective. This significant vulnerability of deep learning models, thus, inspires the exploration of defense methods resistible to adversarial examples. Although numerous countermeasures against adversarial examples have been proposed, the design of a universally applicable defense method across multiple scenarios still remains to be explored. In this study, we propose an effective denoising diffusion-based defense restore network (D3R-Net) based on the denoising diffusion model from the perspective of adversarial restoration, which transforms the adversarial examples into clean samples. Utilizing a highly effective denoising diffusion probabilistic model (DDPM), our D3R-Net transforms input adversarial examples into a state of noise, where diverse forms of adversarial noise transition into Gaussian noise. Subsequently, it captures semantic information through a series of iterative denoising steps. The pixel distribution of adversarial examples is restored in the proposed network to match the original distribution, enabling the classifier to identify adversarial examples correctly. Furthermore, we introduce a combined filtering module to preserve the semantic information of the original image, thereby further enhancing the defensive performance. Instead of modifying the model structure or excluding suspected samples, the proposed method restores the adversarial examples, making it simple yet effective and applicable to a broader range of scenarios. Extensive experiments are conducted on four benchmark datasets, and the results demonstrate that D3R-Net has significant defense capabilities against known and unknown attacks. Our source code is available athttps://github.com/SIM-xidian/D3R-Net.

Change detection involves analysis of discrepancies between two phases. However, when the unchanged elements are known, the changed features to be identified become straightforward. In addition, remote sensing image is constrained by limited spectral information, which leads to blurred boundaries between different semantics. Based on these two prior knowledge, in this artical, we introduce a novel change detection framework, named the mutually reinforcing similarity network (MutSimNet). This architecture aims to minimize false alarms along changing boundaries and reduce misjudgment rates among outliers. First, similarity learning is applied to change detection. The relationship between the two phases is considered when deriving the change feature maps. Second, we devise a mutually reinforcing loss function that integrates initial features with final features. Third, a self-attention module is connected in the feature pyramid network. This design mitigates information loss during the down-sampling process. Fourth, an attention feature fusion strategy is proposed for the integration of multi-layer features. This strategy takes into account the interaction between layer-by-layer features. Fifth, experimental results validate MutSimNet’s efficiency, particularly its ability to focus on edge contour learning. The MutSimNet also achieves superior performance on two benchmark datasets and predicts positive samples with higher probability. The codebase is accessible at https://github.com/ly-yu/MutSimNet.

Change detection (CD) is one of the important research topics in remote sensing (RS) image processing. Recently, convolutional neural networks (CNNs) have dominated the RSCD community. Many successful CNN-based models have been proposed, and they achieved cracking performance. Nevertheless, influenced by the limited receptive field, the CNN-based models are not good at capturing long-distance context dependencies within RS images, negatively impacting their performance. With the appearance of the visual transformer, the above problems have been mitigated. However, the high time costs of the transformer-based models limit their applicability. In addition, previous CD networks (whether CNN-based or transform-based) do not pay attention to the edges of changed areas, reducing the quality of change maps. To overcome the shortcomings discussed above, we propose a new CD method named edge-assisted adaptive transformer detector (EATDer). EATDer consists of a Siamese encoder and an edge-aware decoder. Each branch in the Siamese encoder encloses three self-adaption vision transformer (SAVT) blocks, which aim to capture the local and global information within RS images. Also, two branches are connected by full-range fusion modules (FRFMs), which focus on mining the temporal clues among bi-temporal RS images and pointing out the changed/unchanged messages. The edge-aware decoder first integrates the multiscale features obtained by the encoder using a restoring block. Then, it enhances the combined features by a refining block. Finally, based on the refined features, both the change and edge detection results can be produced. Along with a joint loss function, we can get high-quality change maps in which the changed areas are correct and have clear and smooth edges. The usefulness of our EATDer is validated by extensive experiments conducted on three popular RSCD datasets. Our source codes are available athttps://github.com/TangXu-Group/Remote-Sensing-Image-Change-Detection/tree/main/EATDer

Recently, remote sensing technology has developed faster and faster, and obtaining high-quality panchromatic (PAN) and multispectral (MS) images has become more accessible. The complementarity between them provides new opportunities in multisource remote sensing image classification. However, solving the problem of the semantic gap between multisource high-level features and, at the same time, utilizing the complementary properties between them to reduce intersource information redundancy is still a challenge. This article constructs an$I^{3}$RL-Net for the multisource remote sensing image classification task. Specifically, we design a cross-source interactive enhanced fusion module (CIEF-Module). For multilevel multisource features, by strengthening the dependencies of intrasource features and conducting intersource enhanced fusion, intrasource correlation features are refined, and the problem of the intersource semantic gap can be effectively alleviated. During the cross-source interaction process, we design a complementary representation supervised learning strategy (CRSL-Strategy). According to the similarities and differences of multisource features, it can adaptively promote complementary feature learning, thus generating a nonredundant multisource representation. The method has been verified to be effective on multiple RS datasets. The code is open source at:https://github.com/Xidian-AIGroup190726/Ping-Pie-I3RL-Net.git.

In recent years, small object detection has remained challenging in remote sensing tasks. Firstly, small objects inherently have fewer pixels, making them susceptible to interference from prominently featured larger objects during feature extraction. Secondly, existing detection methods solely based on the Intersection over Union (IOU) loss are disadvantageous for small object detection and fail to leverage the rich prior information in remote sensing images. Based on these observations, we propose a significant feature elimination and sample assessment network for small object detection called SESA-Net, based on the Facet derivative model. SESA-Net introduces prior information to the network through the directional derivatives characteristic of remote sensing images. The overall network comprises the ADM module and SIA strategy. The ADM module eliminates significant responses from shallow large objects, directing the network’s focus towards the features of shallow small objects. The Sample Importance Assessment (SIA) strategy addresses the limitations of the IOU loss function by using high-quality positive samples generated by ADM to provide an evaluation strategy for different positive samples of small objects. This enables the network to focus more on high-quality positive samples, thereby improving the accuracy of small object detection. The effectiveness of the proposed algorithm has been validated on multiple datasets. Our code is available at https://github.com/Xidian-AIGroup190726/RS-objectdetection-SESANet.

With the development of remote sensing technology, remote sensing object detection has been widely applied in various fields, but it still faces some thorny challenges, such as the following: 1) the complexity of object scale changes in remote sensing images makes it difficult to improve the performance of small object detection and 2) remote sensing images have complex backgrounds and densely arranged small and weak objects, which pose a serious problem of feature interference. To alleviate these challenges, we propose an end-to-end adaptive feature separation network called AFSNet, which includes a scale-aware module (SAM) and a class-aware module (CAM). The SAM mainly enables feature maps of different resolutions to detect objects of different scales. Shallow feature maps mainly suppress the features of large objects they contain to focus on small object detection, while deep feature maps increase the detailed features of large objects they contain to focus on large object detection. The CAM is mainly used to distinguish the features in the feature map by category, separating the features of different categories into different channels, thus mitigating the problem of inter class feature interference, and blocking background interference. The effectiveness of this article has been proven on the NWPU VHR-10, IPIU-M, DIOR, and DOTA2.0 datasets. It can be widely applied in civilian, military, and other fields. Through experimental verification, our AFSNet achieved 97.70% mAP on the NWPU VHR-10 dataset, 78.9% mAP on the DIOR dataset, and 58.22% mAP on the DOTA2.0 dataset. Our code is available at:https://github.com/Xidian-AIGroup190726/AFSNet.

Hyperspectral image (HSI) classification is a hot topic in remote sensing. A large number of studies have been proposed and achieved excellent performance. Most of them rely on accurate annotations. However, this requirement cannot always be met. Due to the complex contents within HSIs and the uncontrollable external interference factors, incorrect labels are inevitable. Thus, the study of noisy HSI classification is boomed. Some attempts have been made, and their central ideas are to filter the noisy samples from the training set. Although feasible, this would result in information loss, i.e., the contents covered by the removed samples are ignored. Besides, the characteristics of HSIs are not fully considered in many models. To overcome the above limitations, we develop a spatial pooling transformer network (SPTNet) and a noise-tolerant learning algorithm in this paper. SPTNet first uses a spectral feature extraction (SFE) module to capture the rich spectral information from HSI patches. Then, three spatial pooling transformers (SPTs) are constructed and stacked to explore the spatial knowledge and depress confusing clues caused by the HSI patch division. Finally, a standard transformer encoder is used to enhance the obtained spectral-spatial features for the downstream classification. To use SPTNet to handle noisy HSI classification, the noise-tolerant learning algorithm is designed. It encloses two parts, i.e., a data partition scheme and a label-independent similarity regularization. The data partition scheme divides the training data into clean and noisy sets. Then, the clean samples are used to train SPTNet with the classification loss function. At the same time, similarity regularization helps SPTNet to comprehensively understand HSIs by analyzing the resemblance between clean and noisy samples. Integrating two parts into a co-training framework, SPTNets can be trained under a noisy scenario. Four popular HSI datasets are selected to testify to our methods. The positive results demonstrate that the combination of SPTNet and the noise-tolerant learning algorithm is helpful to the noisy HSI classification. Our source codes are available at https://github.com/TangXu-Group/Hyperspectral-Images-Classification/tree/main/SPTNet-NTLA.

Deep feature learning methods have shown significant advantages over handcrafted feature-based methods in remote sensing image matching and registration. Existing deep learning methods usually introduce complex modules into the deep convolutional network for more robust feature learning. However, they usually require high computation and memory resources for the computing device and have expensive time costs for image registration. As a basic image-processing task, it is crucial to build a lightweight matching network (LM-Net) for fast and accurate image matching and registration. Unfortunately, the image-matching performance will decrease significantly when we directly compress the deep model to a lightweight one. This article proposes an LM-Net based on the knowledge distillation (KD) learning framework for remote sensing image matching and registration. We first build an LM-Net with three convolutional layers. Then, this article proposes an effective KD approach for network optimization, which transfers the effective knowledge from the deep matching network to LM-Net to improve image-matching performances. Specifically, this article considers the useful information in the instance samples and the relation information between samples. It designs the feature and feature relation distillation learning for LM-Net training. Extensive experimental results and analysis have shown the effectiveness and advantages of the proposed LM-Net. LM-Net can reduce the number of parameters and computational complexity of the matching network. Meanwhile, LM-Net can significantly decrease the time cost and achieve results comparable to those of the deep model. It reduces the average image registration time by 42% on remote sensing image matching and registration. Additionally, LM-Net generalizes well on other multimodal remote sensing images.

Multimodal remote sensing image registration is crucial for multimodal information fusion and applications. The significant nonlinear appearance difference between multimodal images caused by the various imaging mechanisms dramatically increases the challenge of image registration. This article proposes an adaptive frequency feature filtering network (F3Net) for cross-modal remote sensing image registration. On the one hand, F3Net explicitly explores the useful frequency components across modal images based on multilevel deep features. On the other hand, F3Net can take advantage of the nonlocal receptive fields by frequency modulation for feature learning and boosting image registration performances. F3Net inserts frequency feature filtering (F3) modules in multilevel deep features. Specifically, F3Net first performs the fast Fourier transform (FFT) for deep features. Then, F3Net designs a frequency attention (FA) module to adaptive enhance the shared and discriminative frequency features between multimodal images while suppressing the frequency components that hinder the cross-modal image registration. In addition, F3Net adopts multiscale frequency filtering fusion to facilitate discriminative feature learning, including global frequency feature filtering (GF3) based on the global image spectrum and local frequency feature filtering (LF3) based on the spectrum of stacked image regions. Experimental results on many remote sensing images have demonstrated the efficiency of the F3Net on multimodal image registration.

Effective feature representation is the key to SAR image terrain classification. Limited by the abstract appearance and the scarcity of high-quality labeled data in this field, the features learned by current methods, especially deep learning models, do not have enough directivity and applicability, which hampers the performance. This paper proposes Multi-image Factor Self-Supervised Learning(MFSSL) to achieve directional feature learning and obtain generalized features with few patch-level labeled data. The framework consists of an upstream multi-factor image style transfer task and a downstream terrain classification task. In the upstream task, the goal of feature learning is first set up by multiple SAR image factors, including the observation region, the terrain category, and the imaging parameters. And then, different styles of SAR terrain images are generated and reconstructed under this goal. Through this bidirectional generative learning, the low-level external appearance of the terrain is removed, while the essential and discriminative feature representation is retained and shared across different factors. Finally, the downstream model inherits the general feature from the upstream model and implements the terrain classification task using a small amount of labeled data. Experiments conducted on three broad SAR scenes with different image factors demonstrate that the proposed framework can improve pixel-level terrain classification only with a few patch-level labeled data.

Convolutional neural networks (CNNs) and self-attention (SA) are highly effective techniques used for the fusion of multisource remote sensing (RS) data, and they have found extensive application in Earth observation (EO) tasks. Nevertheless, CNNs are insufficient for the comprehensive extraction of contextual information and the representation of the sequential properties of spectral features. Furthermore, the loss of edge geometry information is often a consequence of information mining, which limits its application in RS. To address the abovementioned limitations, we propose a method called “multiscale geometric prior aware cross-modal network (MGPACNet)” for RS image fusion classification. First, a geometric prior feature enhanced residual module (GPFEResM) is created to extract shallow multiscale geometric edge prior features and detailed information from multimodal RS data to enhance feature boundary information. Second, a multiscale global-local spatial-spectral feature extraction module (MG-LS2FEM) uses multiscale spatial modeling and global-local spectral modeling to perceive rich semantic information in the spatial-spectral domain. Finally, a dual attention fusion module (DAFM) is designed to use pixel-level SA and cross-attention between heterogeneous data to achieve deep aggregation and cross-focusing of cross-modal information in two branches, and enhance the complementarity of heterogeneous data. A comprehensive examination of public RS data (hyperspectral-synthetic aperture radar (HS-SAR) Augsuburg/Berlin, hyperspectral-light detection and ranging (HS-LiDAR) Trento/MUUFL) from four distinct modalities (HS/SAR/LiDAR) has revealed that our method outperforms alternative models.

In recent years, change detection (CD) has emerged as an increasingly intricate research domain. However, in natural images, the orientation of objects is often aligned with the image boundaries, whereas in RS images, the imaging angles are random. As a result, existing CD methods encounter limitations when effectively representing vector features. In this article, we propose a rotation-invariant CD architecture named RFormer. It effectively utilizes direction-sensitive position embedding (DSPE) to represent features in RS images. To address the challenge of the quadratic growth in attention mechanism complexity with sequence length, we introduce low-cost cross attention (LC2A) to reduce its complexity to$1/{C^{2}}$. Furthermore, we employ the implicit timing extraction process (TEP) to represent interframe bitemporal features. TEP plays a crucial role in mitigating prediction biases caused by seasonal changes in land cover and prevents overconfident discrimination by the classifier in CD tasks. Experimental results demonstrate that RFormer achieves competitive performance on WHU, deeply supervised image fusion network (DSIFN)-CD, CDD, and LEVIR-CD datasets.

Remote sensing (RS) image-text retrieval (RSITR) aims to retrieve relevant texts (RS images) based on the content of a given RS image (text). Existing methods are used to employing the convolutional neural network (CNN) and recurrent neural network (RNN) as encoders to learn visual and textual features for retrieval. Although feasible, the global information hidden in different data does not receive the attention it deserves. To mitigate this problem, transformers have been introduced. Nevertheless, the complexity of RS images present challenges in directly introducing Transformer-based architectures to multimodal learning in RS scenes, particularly in visual feature extraction and cross-modal interaction. In addition, the textual captions are always simpler than the complex RS images, leading to a semantic description appearing in different images. This typical false-negative (FN) sample problem increases the difficulty of RSITR tasks. To address the above limitations, we propose a new RSITR model named prior-experience-based RS vision-language (PERSVL). First, the specific visual and text encoders are used to extract features from RS images and texts. Also, a high-level feature complement (HFC) module is developed based on the self-attention mechanism (SAM) for the visual encoder to explore the complex contents from RS images fully. Second, a dual-branch multimodal fusion encoder (DBMFE) is designed to complete the cross-modal learning. It comprises a dual-branch multimodal interaction (DBMI) module and a branch fusion module. DBMI is designed to fully explore the relationships between different modalities, enriching visual and textual features. The branch fusion module integrates the cross-modal features and utilizes a classification head to generate matching scores for retrieval. Finally, a learning from prior experiences (LPEs) module is designed to reduce the influence of FN samples by analyzing the historical data produced in the model training process. Experiments are conducted on three popular datasets, and the positive results show that our PERSVL model achieves superior performance compared with previous methods. By integrating the advantages of natural language and RS images, our PERSVL can be applied in various applications, such as environmental monitoring, disaster evaluation, and urban planning. Our source codes are available at:https://github.com/TangXu-Group/Cross-modal-remote-sensing-image-and-text-retrieval-models/tree/main/PERSVL.

Joint classification of hyperspectral image (HSI) and light detection and ranging (LiDAR) can simultaneously utilize rich spectral information and elevation information and has become a hot research topic in remote sensing (RS). Although many works have been proposed for this task, their performance cannot reach what we expected due to inadequate cross-modal feature learning and simple feature fusion. This article proposes a multiple information collaborative fusion network (MICF-Net) to overcome those limitations, which aims to leverage the essentially consistent spatial relationships and high-level semantic information in multimodal data to guide the extraction of multimodal fusion features. Specifically, MICF-Net first uses a simple two-branch convolutional neural network (CNN) for preliminary feature extraction. Then, a dual-branch cross-modal attention fusion transformer (CMAFT) is developed to mine global contextual content. By fusing the attention maps of two modalities and limiting their similarity, CMAFT can retain modality-specific information while achieving information interaction based on spatial relationships. Next, an adaptive mask modulation (AMM) module is designed to dynamically balance the learning rate of each modality to ensure the effectiveness of the features of all modalities. Finally, to mine the complementary information of HSI and LiDAR data, a semantic-guided feature fusion (SGFF) module is introduced. It achieves mutual guided learning by exchanging semantic information between two modalities. Positive experimental results counted on three popular HSI and LiDAR datasets demonstrate the effectiveness of the proposed MICF-Net. Our source codes are available athttps://github.com/TangXu-Group/Hyperspectral-Images-Classification/tree/main/MICF-Net.

Remote sensing image (RSI) captioning is a vision-language multimodal task concentrating on both image comprehension and sentence generation. Several studies suggest that encoder–decoder-based methods have achieved success in RSI captioning. However, existing encoder–decoder-based methods may not fully explore image representations for RSI captioning and suffer from a lack of additional prompt information for sentence generation. In this article, a novel multi-view feature fusion and prompt (MVP)-based model is proposed to obtain better RSI representations and enhance language model performance in RSI captioning. Specifically, we design an attention-based feature fusion module to dynamically fuse multi-view visual features, which are extracted from the fine-tuned vision-language pretraining (VLP) model and the vision-task pretraining (VP) model. Then, a flexible visual prefix mapping module is proposed to transform images into visual prefixes, providing semantic information for the subsequent sentence generation. Finally, a BERT-based caption generator is applied to generate accurate descriptions based on the fused visual features and the visual prefixes, which are both outputs from our designed modules. Extensive experiments are conducted on three well-known benchmark datasets, demonstrating that our method achieves state-of-the-art (SOTA) performance. The relevant code is available athttps://github.com/QiaoLing-Lin/MVP.

Multitask learning has been widely applied in visual learning to significantly enhance the performance. The combination of hyperspectral change detection (HCD) and band reweighting can achieve discriminative feature enhancement for improving detection performance. However, existing multitask models for these two tasks are unidirectional, with band reweighting unable to learn from task guidance. To address this challenge, a multitask HCD (MHCD) framework with differential band reweighting and unbalanced contrastive learning is proposed. MHCD consists of a differential band reweighting network (DBRN) and a Siamese detection network. DBRN extracts discriminative information for HCD by analyzing the differential spatial-spectral information across time states, whose optimization is under the guidance of HCD. Furthermore, a multitemporal interaction module and multidomain fusion module are inserted into the Siamese detection network. They hierarchically connect cross-temporal features and fuse features from spatial, spectral, and temporal domains, providing complementary clues in these different domains. Considering the sample imbalance and enormous variation within a class in binary HCD, an unbalanced contrastive learning method based on multiple prototypes (UCLM) tailored has been considered. It estimates multiple prototypes to flexibly adjust the contribution of different classes of samples to the loss. The proposed method has been validated using three public benchmark datasets, demonstrating improvements in multiple metrics for change detection. The code of our paper is available at:https://github.com/jiefeng0109/MHCD.

Remote sensing image captioning (RSIC) is an interesting but challenging cross-modal reasoning task for computer vision and natural language processing. Most of the recent popular approaches for RSIC utilize encoder-decoder architectures, which focus on visual features captured by convolutional neural network (CNN)-based encoder and semantic information by recurrent neural network (RNN)-based or long short-term memory (LSTM)-based decoder, but encounter difficulties with multiscale, multicategories, and direction ambiguity challenges. To make the most of semantic understanding ability of Transformers, in this article, we propose a new attention-based visual-linguistic reasoning framework with dual Transformer for RSIC. Specifically, Swin Transformer (SwinT) encoder with shifted window partitioning scheme is introduced for multiscale visual feature extraction to discover the intrinsic relationship in the objects, and then, a Transformer language model (TLM) with self-attention and cross attention is designed as the decoder to generate a well-formed sentence for the image. Extensive experiments are conducted on the public RSIC benchmark datasets, including UCM-Captions, Sydney-Captions, and RSICD. The impressive performance verifies the effectiveness and superiority of the proposed method. In addition, the source code and models of this work are publicly available athttps://github.com/LianYi233/TrTr-CMR.

In the domain of remote sensing image classification, single sensors are constrained by their sensing angles and information dimensions, rendering them incapable of fully capturing the intricate characteristics of ground objects. Different sensors can provide complementary information, significantly enhancing the performance of object classification. However, due to their unique physical observation principles and varying spatial-spectral resolutions, different modalities capture heterogeneous features of ground objects, resulting in a semantic gap issue when integrating modalities. This article designs a multimodal contrastive learning framework, starting with the preprocessing of hyperspectral image (HSI) to obtain its spatial and spectral modalities, and then combining these with the light detection and ranging (LiDAR) modality, forming a ternary modality contrastive learning framework that achieves deep semantic alignment between different modalities. Furthermore, to enhance the model’s generalization ability, based on the neighborhood semantic similarity of HSI, we propose a spectral selection data augmentation method. Extensive experiments on four public datasets show that our method outperforms several other state-of-the-art (SOTA) methods in classification performance.

Surrounding contexts are generally perceived as interfering with object tracking in satellite videos, leading to model drift. From another perspective, they can also be seen as reference objects of the tracked target, the dynamic interactions between them could provide essential information. In this article, a high-order relation learning transformer (HRLT) is proposed for satellite video object tracking, which not only models the high-order interactions of different target-context pairs but also reasons the associations between these high-order relations across multiple frames. First, a spatial high-order relation reasoning (SHR2) module is designed to model the high-order interactions between the target and scene contexts. Second, a temporal high-order relation reasoning (THR2) module is proposed to associate and reason these spatial high-order relations across multiple frames. Third, historical high-order relations are collected to provide more reasoning bases for the current frame prediction. Finally, qualitative and quantitative evaluations are performed on the SV248S, SkySat, and VISO datasets. The results show that HRLT outperforms 20 popular methods in different challenging scenarios.

In deep learning, convolutional neural networks (CNNs) and transformers have gained excellent achievements in remote sensing scene parsing. Strong feature representation ability is still a challenge for them. Besides, the complex scenes are still essential challenges for deep learning in remote sensing scene parsing. In this article, an efficient local–global lifting transformer (LGLFormer) framework is proposed to ease the challenges above. It effectively combines CNNs, transformer, and wavelet transform to build a strong local–global (LG) feature representation network. Besides, global feature learning driven by LG adaptive features is proposed based on the 2-D LG adaptive feature extractor (LGAFE) and refined global feature attention module. The 2-D LG lifting feature extractor is inspired by the lifting scheme, which introduces local and global dependency. Furthermore, two LG lifting schemes are proposed, including the series and parallel modes, which can effectively learn LG relations between pixels. Finally, experiments are validated on three remote sensing benchmark datasets. The proposed LGLFormer achieves the state-of-the-art with 99.02%, 99.2%, and 99.48% overall accuracy (OA) on AID, WHU-RS19, and UCM datasets, respectively. In addition, LGLFormer shows good convergence with competitive parameters. The experimental code will be available athttps://github.com/yutinyang/LGLFormer.

Tiny objects in satellite videos are usually not independent individuals, there exist rich semantic and temporal relations with each other. Thus, modeling and reasoning the variation of such intrinsic relationships can be beneficial for tiny object tracking. In this paper, a relation learning reasoning method is proposed for tiny object tracking in satellite videos. The core of the proposed is the relation reasoning network that consists of a key context module, a global semantic module, and a relation reasoning module sequentially. First, the key context module exploits global key contexts which explicitly or implicitly contribute to the target object, modeling the intrinsic relations with the target. Second, to reason the contribution, the global semantic module analyses the interaction between them in the same frame. Third, the relation reasoning module deduces the target based on the variation of the semantic relations among different frames. Such a relation learning reasoning approach which takes the target as the core is aligned with the satellite tiny object tracking task, significantly improves the identification performance in dense similarity scenes and the retrieval ability after completely occluded. Furthermore, the proposed method is shown to report improved qualitative and quantitative results on Jilin-1 and SkySat satellite video datasets.

Semi-supervised learning aims to exploit the potential of unlabeled data to enhance model performance, which makes it suitable for addressing the challenge of limited labeled data. As a popular technology, pseudo-label is widely applied in many semi-supervised remote sensing (RS) change detection methods. However, when facing limited labeled data, abundant low-quality pseudo-labels from a poorly-performing model hinder the effective enhancement of model performance. To address this issue, we propose a novel semi-supervised strategy, named ensemble cross pseudo supervision (ECPS). The utilization of ensemble learning to merge outputs from several change detection models enhances pseudo-label quality, leading to more accurate change information and a significant boost in model performance, even with limited labeled data. In this method, adopting crosswise supervision ensures that no additional inference costs caused by ensemble learning are consumed. This provides both high efficiency and effectiveness for identifying land-cover changes. On the other hand, a simple yet effective ensemble strategy is proposed, which allows to manually adjust the model’s tendency towards higher precision or recall for satisfying practical requirements. We conduct extensive experiments on four public RS change detection datasets, and the promising results demonstrate the superiority of the proposed method across various numbers of labeled samples. Our source codes are available at https://github.com/TangXu-Group/ECPS.

As an important topic in the remote sensing (RS) image processing community, change detection has attracted much attention from researchers, which aims to distinguish land-cover changes in a geographic position. This is a challenging task because the visual representations of land cover captured from RS images at different periods would vary widely and considerably, resulting in significant differences in feature representations. To alleviate this problem, many existing deep-based methods employ the parameter-shared strategy to map RS images into a common feature space for detecting the changes. Although they are feasible, the simple and single visual information learned by deep models is still not sophisticated enough for satisfactory results. To address this problem, we propose a forward dictionary learning (DL) model named forward DL detector (FDLdet) in this article. Besides the common visual features, our FDLdet takes into account the essential information, e.g., element composition and land-cover category, for change detection. FDLdet consists of a feature extractor, a coefficient generator, and a deep dictionary. Specifically, first, the feature extractor is used to extract shared deep features from RS images. Second, the coefficient generator transforms these deep features into word coefficients. Third, words within the deep dictionary are combined by word coefficients to generate the dictionary features with essential information. Finally, the dictionary features are used instead of deep features to detect land-cover changes. Extensive experiments are conducted on two public large-scale datasets, i.e., season-varying change detection (SVCD), Sun Yat-sen University change detection (SYSU-CD), and LEVIR change detection (LEVIR-CD). Experimental results demonstrate the effectiveness of the proposed FDLdet. Our source codes are available athttps://github.com/TangXu-Group/FDLdet.

Remote sensing image-text retrieval (RSITR) is pivotal for knowledge services and data mining in the remote sensing (RS) domain. Considering the multiscale representations in image content and text vocabulary can enable the models to learn richer representations and enhance retrieval. Current multiscale RSITR approaches typically align multiscale fused image features with text features but overlook aligning image-text pairs at distinct scales separately. This oversight restricts their ability to learn joint representations suitable for effective retrieval. We introduce a novel multiscale alignment (MSA) method to overcome this limitation. Our method comprises three key innovations: 1) a multiscale cross-modal alignment transformer (MSCMAT), which computes cross-attention between single-scale image features and localized text features, integrating global textual context to derive a matching score matrix within a mini-batch; 2) a multiscale cross-modal semantic alignment loss (MSCMA loss) that enforces semantic alignment across scales; and 3) a cross-scale multimodal semantic consistency loss (CSMMC loss) that uses the matching matrix from the largest scale to guide alignment at smaller scales. We evaluated our method across multiple datasets, demonstrating its efficacy with various visual backbones and establishing its superiority over existing state-of-the-art methods. The GitHub URL for our project ishttps://github.com/yr666666/MSA.

The continual progression and expanding applications of Hyperspectral Imaging (HSI) technology necessitate the development of lightweight HSI classification models that are capable of lifelong learning. However, the computationally-demanding task of training and updating HSI classification models, exacerbated by the substantial number of trainable parameters in feature extractors, remains a substantial challenge. This paper proposes an Attention-based Reservoir Computing (ARC) model to overcome these hurdles. The ARC model utilizes a cross-slicing operation to generate multi-directional inputs, treating the HSI dataset as spatial sequence for processing within a reservoir. For every target pixel, four spatial sequences from various directions are introduced into the reservoir, generating four corresponding outputs. A voting mechanism then evaluates these outputs to yield the final prediction. Additionally, we design an attention-based leaky function for reservoir computing to capture the spatial correlation inherent in HSI data accurately. The attention-based leaky function enables the reservoir state to weigh less on the pixels outside the region of interest (ROI) and have a longer memory for pixels inside the ROI. The ARC was tested on widely used HSI datasets: Indian Pines, PaviaU, and Salinas. It demonstrated competitive lightweight classification performance against state-of-the-art lightweight models by maintaining comparable training time while achieving superior accuracy. Furthermore, the model’s lifelong learning accuracy also showed superior performance compared to existing lifelong learning models, with a one thousand times reduction of parameter-to-be-updated. This work makes the ARC model an effective contender for HSI classification tasks, excelling in both lightweight classification and lifelong learning capacities. The source codes are available publicly at: https://github.com/Waterman-Ann/ARC.

Existing mainstream MOT methods can be categorised into two frameworks including two-stage and one-stage ones. Two-stage ones divide MOT task into object detection and association tasks which usually achieve high accuracy. One-stage ones train a joint model to achieve both detection and tracking. So their advantage usually lies in the high tracking efficiency. In this paper, we inherit the advantages of the two types frameworks and propose the object knowledge distilled joint detection and tracking framework (OKD-JDT) to achieve accurate as well as efficient tracking. Firstly, the performance of two-stage methods largely depends on the highly performed detection network. So, we treat the detection network as the teacher network to guide the discriminative object feature learning in one-stage methods by using knowledge distillation. Then, in distillation learning, we design the adaptive attention learning to learn the discriminative features from teacher network to student network. In addition, with the similar appearance and uniform moving behaviour of objects in satellite videos, we propose to use joint center point distance and intersection-over-onion (IOU) to generate tracklets. Experiments on JiLin-1 satellite videos with different objects demonstrate the effectiveness and the state-of-the-art performance of the proposed method.

Multilabel classification in remote sensing (RS) images aims to correctly predict multiple object labels in an RS image with the primary challenge of mining correlations among multiple labels. In this context, we argue that a scene can be treated as a high-level depiction of the interactions among multiple interconnected objects within the image. However, hierarchical relationships between the scene and local objects are often neglected in other state-of-the-art approaches. In this article, we consider multilabel classification as a global-to-local prediction process, whereas the scene of an image is first identified, followed by recognition of local objects in the image. To achieve this, we propose a novel hierarchical knowledge graph (HKG)-based framework for multilabel classification in RS images (ML-HKG). Specifically, we first construct a hierarchical KG to depict label correlations between scenes and objects and represent the hierarchical knowledge as interrelated scene- and object-level label embeddings. Subsequently, we generate a scene-aware enhanced feature map by recognizing scene categories in an image under the guidance of scene-level knowledge embeddings. Afterward, object-level embeddings are used to derive category-specific visual representations for final multilabel prediction. Extensive experiments on the UCM and AID datasets demonstrate the effectiveness of our framework.

Satellite video tracking tasks are often characterized by blurred foreground boundaries in vast scenes, a wide range of targets varying in scale, and irregular changes in appearance. These challenges significantly impact the optimization of robust tracker performance. Therefore, it is imperative to extract diverse features with dynamic adaptive learning capabilities for the target being tracked in each sequence. In this article, we explore a novel adaptive multi-scale Transformer (MT) tracker for satellite videos to explore the potential spatiotemporal information of the target effectively. Specifically, a multi-scale spatial Transformer (MSST) is designed to leverage stage-by-stage spatial reduction and channel doubling, thereby enhancing the representation capabilities for the tracked target. In dynamic feature learning, an adaptive temporal Transformer (ATT) is then introduced based on multiple cross attentions, which analyzes the adaptive learning capacity for the dynamic target. It analyzes the weight proportion of different attentions automatically in the specific sequence through the learnable parameters. Finally, a multi-scale feature (MSF) regression module is crafted to improve the positioning accuracy of targets with low pixel counts in satellite scenes. This module accomplishes precise annotation of target boxes by effectively fusing features from diverse stages. We evaluate the proposed tracker performance on several public satellite datasets, including SatSOT, SV248S, and VISO. Experimental results show that the performance of our model can be comparable to the state-of-the-art trackers.

Change detection (CD) based on deep learning has achieved excellent performance in recent years. However, these models exhibit limited capability in complete temporal modeling or face problems with fine-grained spatial features being overshadowed by the temporal context. Pure CNN-based CD pipelines also struggle to establish long-range connections. In this article, a globally aware continuous-time redistribution network (GCRNet) is proposed for RSCD. First, a boundary extraction branch is designed to preserve the semantic invariance of objects within the same boundary. This is achieved by providing boundary attention to adaptively guide the integration of temporal and spatial information. Then, a globally aware operator (GAO) is developed to obtain global interaction features. GAO utilizes the convolution theorem, which combines the Fourier transform and inverse Fourier transform, achieving it with low computational costs. Finally, an adaptive feature redistribution (AFR) module is designed to increase the distance between positive and negative samples in the latent space with change perception. It alleviates the effects of the severe class imbalance issue. Experimental results demonstrate that our proposed GCRNet surpasses 13 state-of-the-art CD methods. It achieves F1-score 0.33%, 0.62%, 0.84%, 0.17%, and 1.54% higher than the second-best model on the LEVIR-CD, LEVIR-CD+, WHU, CDD, and DSIFN datasets. The code of GCRNet is available athttps://github.com/XiaowenZhang-kuku/GCRNet.

Robust feature learning has always been a research hotspot in dynamic temporal tasks. It makes the model almost unaffected by some challenging properties. The sequential nature of the transformer means attractive for temporal learning tasks, making it perform well in the video field. It is a current research hotspot for learning effective features by utilizing the target motion trends in satellite videos with multiple attributes, such as similar objects (SOBs) interference and occlusion. In this article, a novel discriminative temporal learning transformer tracker (DTLTracker) is introduced to characterize the dynamic target information for satellite videos. A discriminative transformer (DT) is proposed to comprehensively explore the dynamic target features with multiple attention mechanisms. It focuses on the primary information of the search area, making the target more discriminative. A fast convergence (FC) filter is designed to accelerate the weights convergence in calculating the target correlation operation, thereby ensuring the efficiency of model learning. The effectiveness and convergence have been demonstrated for the proposed optimization method. Additionally, a motion prior correction (MPC) module is constructed to utilize temporal information for target tracklet prediction, assisting the tracker in predicting the correct target. Numerous experiments are performed on three satellite videos to verify the effectiveness and feasibility of the proposed DTLTracker. It shows robustness compared to the state-of-the-art trackers on some challenging properties.

Satellite imagery, due to its long-range imaging, brings with it a variety of scale-preferred tasks, such as the detection of tiny/small objects, making the precise localization and detection of small objects of interest a challenging task. In this article, we design a knowledge discovery network (KDN) to implement the renormalization group theory in terms of efficient feature extraction (FE). Renormalized connection (RC) on the KDN enables “synergistic focusing” of multiscale features. Based on our observations of KDN, we abstract a class of RCs with different connection strengths, called$n21$C, and generalize it to feature pyramid network (FPN)-based multibranch detectors. In a series of FPN experiments on the scale-preferred tasks, we found that the “divide-and-conquer” idea of FPN severely hampers the detector’s learning in the right direction due to the large number of large-scale negative samples and interference from background noise. Moreover, these negative samples cannot be eliminated by the focal loss function. The RCs extends the multilevel feature’s “divide-and-conquer” mechanism of the FPN-based detectors to a wide range of scale-preferred tasks, and enables synergistic effects of multilevel features on the specific learning goal. In addition, interference activations in two aspects are greatly reduced and the detector learns in a more correct direction. Extensive experiments of 17 well-designed detection architectures embedded with$n21$Cs on five different levels of scale-preferred tasks validate the effectiveness and efficiency of the RCs. Especially the simplest linear form of RC—E421C performs well in all tasks, and it satisfies the scaling property of renormalization group theory. All experiments can be trained and tested on a graphics card with 8 GB of video memory, which greatly enhances the applicability of our methodology. We hope that our approach will transfer a large number of well-designed detectors from the computer vision community to the remote sensing community. Datasets and codes will be available at:https://github.com/rabbitme/

Semantic segmentation of high-resolution remote sensing images is essential in many fields. Nevertheless, in practical applications, constrained by limited computational resources and complex network structures, many advanced models on semantic segmentation often fail to show efficient performance, prompting research on lightweight models. For lightweight semantic segmentation models, the two-branch architecture has been shown to work well in speed and performance. However, such two-branch architectures usually do not utilize enough information for shallow structures to efficiently provide richer multiscale information for the two branches. The lightweight modules it uses are difficult to extract the global context information of the features effectively. Compared with the current advanced semantic segmentation models, lightweight models still have some differences in performance. In order to solve these problems, we propose a new lightweight dual-branch architecture efficient shallow-deep interaction network (ESDINet), which can quickly extract low-level spatial and high-level semantic information of images through the detail branch and semantic branch. Specifically, we have constructed an efficient double-branch structure with shallow and deep different interactions to achieve multiscale information interaction. At the same time, we optimize the semantic branch and propose a new linear attention block to effectively improve the global perception of the semantic branch. We performed extensive experiments and the results show that our model achieves a good balance between segmentation accuracy and inference speed. In particular, ESDINet achieves 82.03% mean intersection over union (mIoU) on the Vaihingen test set, while the proposed model achieves an inference speed of 116 frames/s (FPS) for$512\times512$inputs on a single NVIDIA GTX 2080Ti GPU.

With the rapid advances in deep learning techniques, remote sensing object detection has achieved remarkable achievements in recent years. However, tiny object detection remains unsatisfactory and suffers from two main drawbacks, including (1) the high sensitivity of IoU for location deviation in tiny objects and (2) the poor-quality feature representations of tiny objects. To address the aforementioned problems, we propose a Multi-stage Enhancement Network (MENet) that achieves the instance-level and feature-level enhancement of tiny objects from different stages of the detector. Since the IoU-based label assignment drastically deteriorates the positive samples for tiny objects, we first propose a Central Region-based (CR) label assignment to substitute it in the Region Proposal Network (RPN). The CR label assignment regards the anchors that fall into the central region of ground-truth boxes as positive samples, which provides more positive samples for tiny objects. Then, we design a Gated Context Aggregation (GCA) module that selectively aggregates valuable context information to enhance the feature representation of tiny objects. Additionally, we devise a positive RoI feature (pRoI) generator in the Region Convolutional Neural Network (R-CNN) to generate a rich diversity of high-quality positive RoI features for tiny objects. We conduct extensive experiments on AI-TOD and SODA-A datasets, and the results demonstrate the effectiveness of our proposed method.

Image registration and change detection are crucial for multitemporal remote sensing image analysis. The images should be registered before the change information detection. Existing deep learning methods have shown significant advantages in image registration and change detection tasks. They usually design two independent task-specific deep networks for image registration and change detection, respectively. These independent deep networks will learn from scratch and rely on many task-specific labeled training datasets. This article finds that image registration and change detection have similar learning mechanisms, which focus on extracting discriminative features. Inspired by this, we propose a Unified image Registration and Change detection Network (URCNet) that can perform image alignment and change information detection through a single network. Additionally, this article proposes various deep collaborative learning methods for URCNet optimization, which enforce that the URCNet can effectively support remote sensing image registration and change detection simultaneously. Extensive experiments demonstrate the effectiveness of the proposed URCNet for image registration and change detection, which can achieve comparable and better results with task-specific and more complex deep networks. The proposed URCNet can support multitasks based on the same scene images, different scene images, and even multimodal images. Moreover, URCNet shows significant advantages over other deep networks in change detection under limited labeled datasets.

The joint classification of multispectral (MS) and panchromatic (PAN) images aims to provide a more detailed and accurate interpretation of land features. Although deep-learning-based methods have achieved remarkable success in this task, the generalization performance of networks is compromised when labeled samples are insufficient. In this study, we explore the possibility of leveraging unlabeled remote sensing images (RSIs) through contrastive learning and demonstrate the challenges associated with directly applying contrastive learning to RSIs. To end this, we propose a cross-source contrastive learning method for few-shot MS and PAN joint classification (CrossCLMP), which aims to learn sufficient transferable representations in a self-supervised contrastive manner so as to provide a robust pretrained model for fine-tuning the downstream joint classification task. Specifically, we design: 1) intersource and intrasource alignment loss (ER-Align) to achieve self-supervised feature extraction and alignment; 2) the source-unique feature adaptive separation (SUAS) strategy to model source-unique information explicitly; and 3) the auxiliary contrastive learning (ACL) strategy to mitigate the adverse impact of numerous false-negative samples in the pretraining stage. The experimental results and the theoretical analyses on multiple popular datasets comprehensively demonstrate the effectiveness and robustness of the proposed method under few-shot. Our code is available at:https://github.com/Xidian-AIGroup190726/CrossCLMP.

With the rapid development of remote sensing technology, satellites can easily acquire multispectral (MS) and panchromatic (PAN) images. It is challenging to utilize their complementarity to effectively combine each other’s advantages and mitigate the differences between different modes. In this article, we propose a high-low-frequency progressive-guided diffusion model. It is used to generate an image with the advantages of both MS and PAN, which can be complementary to MS and PAN and, thus, can better reduce the modal differences between them. Therefore, we use the fusion image as an auxiliary mode and an intermediate bridge, which can better connect the characteristics between various sources. First, we design guidance information that contains the advantages of MS and PAN, and some operations can make this information better guide the generation stage. In addition, we design a high-low-frequency progressive guidance strategy; by using this strategy, we can first ensure the overall structure and layout of the image in the generation stage and then refine the local details and features of the image. This dramatically improves the quality of the generated image. Finally, we use mathematical knowledge to explain the rationality of the strategy. We validate our method on multiple datasets and achieve the best performance. Our code ishttps://github.com/Xidian-AIGroup190726/HLF-GDiffusion.

As a hot research topic in remote sensing, effectively integrating the advantageous features of multispectral and panchromatic images is the main challenge for fusing these two remote sensing images. This article proposes a multiscale frequency fusion network based on ConvGRU. To address the underutilization of texture features, we extract multiscale bandpass and low-pass sub-bands representing texture and content features through Contourlet decomposition. Multiscale bandpass sub-bands contain more comprehensive and concentrated texture details. Then, by proposing a multiscale frequency feature extractor based on ConvGRU, we effectively integrate and enhance sub-bands of different scales and frequencies, fully utilizing the characteristics of multispectral and panchromatic images and scale transmission. With these enhanced sub-band features, we obtain more comprehensive scale-enhanced texture features. Simultaneously, content features are also preserved as dual-source image features. Moreover, to reduce redundancy between fused features and make more efficient use of the obtained enhanced features, we designed an Inver-band integrator (IBI) module. It can fuse enhanced features at different scales, improve the complementarity between features, and thus achieve effective fusion. Experimental results demonstrate the effectiveness and robustness of our model on multiple datasets. Our codes are available athttps://github.com/Xidian-AIGroup190726/GMFnet.

In recent years, high-resolution (HR) remote sensing images (RSIs) segmentation has received growing attention. The huge number of pixels poses a challenge to the semantic segmentation algorithm, which is limited by the storage of GPUs, so the current methods for processing HR RSIs are categorized into two main categories, i.e., global methods and local methods. The former downsamples the original image and loses a lot of feature details. The latter crops the original image and fails to obtain global contextual information. Both types of methods lead to limited segmentation accuracy. In this article, we propose an end-to-end framework, called global injection network (GINet), which explores two levels of feature distribution and feature relationship to achieve tradeoff between global context and local details. In concrete terms, we propose the global-guided normalization (GGN) module, which injects global context information into local branch and modulates local features using global features to enhance the global perception of local branch. In addition, to constrain the spatial consistency of two branches, inspired by the knowledge distillation technique, we propose local affinity distillation (LAD) loss, which distills the relations in local features into global features to keep the similarity of the relationships corresponding to patches in the two branches. The comprehensive experimental results on three large-scale land-cover classification datasets, DeepGlobe ($2448 \times 2448$), Inria Aerial ($5000 \times 5000$), and GID-15 ($7200 \times 6800$), confirm the effectiveness and superiority of our method in HR semantic segmentation tasks.

In recent years, panchromatic (PAN) images and multispectral (MS) images, as a type of multimodal remote sensing data, are attracting increasingly more attention to their classification problems. However, effectively representing size variations of targets in remote sensing images and reducing redundant representations of different modalities’ deep semantic features to enhance classification accuracy remains a challenge. In this article, we propose a semantically nonredundant continuous-scale feature network (SNCF-Net) for PAN and MS classification, consisting of two modules: the texture-enhanced continuous scale input generation module and the cross-modal feature Kernel interaction (CMKI) module. By simulating the human eye’s adjustment of distance to observe objects of different sizes, we employ 3-D convolution to extract continuous-scale images generated by the texture-enhanced continuous-scale input generation (TCIG) module, enabling optimal feature representation of objects in remote sensing images. Additionally, the texture enhancement (TE) strategy in the TCIG module alleviates texture diffusion in scale space, enhancing the network’s ability to represent texture features. Subsequently, the CMKI module utilizes the response differences between different features to generate convolution kernels from deep feature maps, enabling feature interaction between the PAN modal and MS modal. This reduces redundant representations of essential image content information in deep features of two modalities, facilitating a better mapping between dual-modal features and categories. Our results achieve state-of-the-art performance on multiple datasets. The code is available athttps://github.com/Xidian-AIGroup190726/SNCFNet.

Semi-supervised change detection (SSCD) has experienced rapid development, with numerous semi-supervised methods being proposed to reduce the reliance on labeled data in change detection. Existing approaches typically rely on manually set high-confidence thresholds to select robust pseudo-labels. However, the single-pixel threshold filtering method for pseudo-labels (STFP) lacks context correlation, cannot eliminate high-confidence false positive samples, and leads to erroneously filtering out low-confidence true positive samples. To address this issue, we propose robust instance-based semi-supervised learning change detection (RISL) for remote sensing images. RISL evaluates the reliability of each instance object by linking the semantic information of the context, thereby generating robust pseudo-labels. In RISL, firstly, a simple boundary trimming module (BT) as a preprocessing method for change prediction map is introduced. BT can effectively remove low-confidence false positive samples while avoiding confusion in the category of instance objects, thereby improving the quality of instance objects. Then, we propose a reliable instance evaluation module (RIEM) to evaluate the reliability of each instance object. RIEM combines the semantic information of the entire instance and establishes correlations between sample contexts to determine the reliability of the instance, effectively eliminating high false positive samples. In addition, the consistency regularization (CR) is integrated into RISL, and a new strategy suitable for RIEM is constructed. This strategy enhances the model’s generalization ability by mining and hiding semantic information from different views of unlabeled data. Experimental results on the challenging WHU-CD, LEVIR-CD, and CDD-CD datasets show that the proposed method achieves 89.80%, 90.01%, and 87.56% F1 scores on labeled data with 5% distribution. RISL achieves state-of-the-art performance compared to other methods.

The electromagnetic environments have become more complex with the development of wireless communication technology. Signal modulation classification has attracted extensive attention due to its application in electronic countermeasures and physical layer security threat prevention under complex electromagnetic environments. Excellent classification performance requirements challenge the adaptability of the method and the ability to extract modulation characteristics. This paper proposes an automated machine learning framework, AutoSMC, for signal modulation classification. An adaptive signal augmentation method is proposed to adapt to the network changes during the search process. In order to extract the modulation features effectively, an scalable convolutional random fourier feature block is proposed. Moreover, the initial search space of the framework is given. The Bayesian Optimization is used to drive hyperparameter optimization to achieve AutoSMC and obtain the optimal method state. Great experiments were carried out on RADIOML 2016.10A and RADIOML 2016.10B. Experimental evaluations on these datasets show that our approach AutoSMC achieves state-of-the-art results compared to the most relevant signal modulation classification methods.

Connections between visual components are ubiquitous. Graphs, as a highly flexible data structure, not only allow imposing relational induction bias on data, but can provide a completely distinct learning perspective for regular image data. In this paper, we propose a hierarchical dynamic graph clustering network (HDGCN) for visual feature learning. We construct hierarchical graph representations in graph domain in an adaptive, data-adaptive and task-adaptive manner. First, the initial graph is constructed in high-dimensional feature domain of images. To mine the hierarchical geometric features in latent graph space, adaptive clustering network (ClusterNet) is performed to learn discriminative clusters and generates cluster-based coarse graph. Then, graph convolutional networks (GCNs) are used to diffuse, transform and aggregate information among clusters. So, the intra-class and inter-class information is fully explored to increase the discriminativity of graph representations. Next, coarsened graph representations are mapped to grid based on its affinity with linear projection features. To further improve the task adaptation of clusters and hierarchical graph representations, ClusterNet and GCNs are fused in the same framework for end-to-end training and clusters is updated dynamically. We have conducted extensive experiments on classification and segmentation tasks. The experimental results fully validate the robustness of the proposed algorithm.

Multisource remote sensing image fusion classification aims to produce accurate pixel-level classification maps by combining complementary information from different sources of remote sensing data. Existing methods based on Convolutional Neural Networks (CNN-based) utilize a patch-based learning framework, which has a high computational cost, leading to poor real-time performance. In contrast, methods based on Fully Convolutional Networks (FCN-based) can process the entire image directly, achieving fast inference. However, FCN-based methods require high computational resources and exhibit shortcomings in feature fusion, hindering practical applications. In this paper, a lightweight FCN-based Progressive Hierarchical Fusion Network (PHFNet) is tailored for multisource remote sensing image classification. PHFNet comprises a pyramid dual-path encoder and a pyramid decoder. In the encoder, cross-source features are hierarchically fused via the adaptive modulation fusion module (AMF), which leverages style calibration for cross-source alignment and promotes the complementarity of the fusion feature. In the decoder, we introduced an improved convolutional gated recurrent unit (iConvGRU) to progressively integrate the semantic and detailed information of hierarchical features, producing a context-enhanced global representation. In addition, we consider the relation between the channel number, convolutional kernel size, and parameter count to make the model as lightweight as possible. Comprehensive evaluations on three multisource remote sensing datasets demonstrate that PHFNet improves overall accuracy by 1.5% to 2.8% with a low computational overhead compared to state-of-the-art methods. The source code is avaliable athttps://github.com/ShirlySmile/PHFNet.

Unsupervised cross-domain scene segmentation approach adapts the source model to the target domain, which utilizes two-stage strategies to minimize the inter-domain and intra-domain gap. However, the accumulation of errors in the previous stages affects the training of the subsequent stages. In this paper, a framework called statistical and structural domain adaptation (SSDA) is proposed to optimize inter-domain and intra-domain adaptation jointly. Firstly, the statistical inter-domain adaptation (StaIA) is proposed to model dynamic subdomains, which continuously adjust seed samples during the process of domain adaptation to mitigate error accumulation. The dynamic subdomains are modeled by exploring Bayesian uncertainty statistics and global balance statistics, which alleviate the imbalance problem in uncertainty estimation. StaIA encourages the model to transfer comprehensive and genuine knowledge through the seed loss for inter-domain adaptation. Secondly, the structural intra-domain adaptation (StrIA) is proposed to align the intra-domain gap among dynamic subdomains by the structural priors. Specifically, the StrIA models structural priors by truncated conditional random field (TruCRF) loss within the neighborhood, which constrains intra-domain semantic consistency to reduce the intra-domain gap. Experimental results demonstrate the effectiveness of the proposed cross-domain scene segmentation approaches on two commonly-used unsupervised domain adaptation benchmarks. The code is available at https://github.com/ChicalH/SSDA.

Constructing effective proxy data is one of the core challenges in data-free knowledge distillation. The existing models ignore the influence of the category entanglement of the generated data on the distillation. To alleviate this issue, imitating the human learning process, a new category-aware curriculum learning mechanism is proposed in this paper to perform data-free knowledge distillation, called CCL-D. The main ideology of this category-aware curriculum learning mechanism is to provide a new learning mode for data generation and network training, which enables the model to realize the knowledge distillation process from easy to difficult through automated curriculum learning. In this novel learning mechanism, a category-aware monitoring module is proposed to constrain the category attribute of generated data. Based on this monitoring module, the curriculum learning process for data generation and network training is designed and applied. Initially, the generator is guided to obtain new data with clear category features. The utilization of data with apparent category features is easy for student network training, and it enables the student network to learn clear and significant category features at the early training stage. Subsequently, the generator is guided to generate data with category entanglement. Utilizing these new data with category entanglement problems can improve the recognition ability of the student network to interclass interference and enhance network robustness. The effectiveness of the CCL-D is verified on the six benchmark experimental datasets (MNIST, CIFAR-10, CIFAR-100, SVHN, Caltech-101, Tiny-Imagenet).

For accurate segmentation, effective feature extraction has always been a challenging problem, since the variability of appearance and the fuzziness of object boundaries. Convolutional neural networks have recently gained recognition in feature representation learning. However, it is only conducted in the spatial domain, and lacks effective representation of directionality, singularity and regularity in the spectral domain for anomaly detection of images. This is the key to feature learning representation of high-order singularity. To solve this problem, a multi-scale contourlet knowledge guide learning network is proposed in this paper. It is novel in this sense that, different from the CNNs in the spatial domain, the proposed method learns the multi-scale contourlet sparse representation to obtain more effective and sparse features in multi-scales and multi-directions. Furthermore, the contourlet knowledge guide learning can enhance the representation of spectral domain features. It is shown that the proposed network can learn the multi-level discriminative features and capture the more accurate object boundaries. The segmentation ability in theoretical analysis and experiments on five polyp segmentation datasets (CVC-ColonDB, CVC-ClinicDB, Kvasir-SEG, ETIS-LaribPolypDB, EndoSceneStill) and two building datasets (Massachusetts, WHU) are compared with developed methods. It must be emphasized that there is potential in effective feature learning representation and the generalization capability of the proposed method in deep learning, recognition and interpretation.

Inspired by the sparse and hierarchical features representation in the ventral stream of the human visual system, the biologically inspired multi-scale contourlet attention network (BMCAnet) is proposed to extract robust discriminative features. First, we constructed the multi-scale contourlet filter banks as a population of neurons in the primary visual cortex (V1), and extracted sparse features in a multi-scale and multi-direction way. It simulated a simple cell in V1 that responds to stimuli in a specific direction. Second, in order to refine contourlet features adaptively, the Shannon block attention module (SBAM) is introduced by integrating Shannon entropy as the third branch of the channel attention module (CAM), thus the weights of contourlet coefficients can be learned adaptively. Third, the responses of the spatial and spectral features are pooled by the proposed contourlet pooling layer to obtain the invariant structure features with the specified rules, which roughly stimulate the pooling process of complex cells in the V1 area. Last, the combination of global average pooling (GAP) and full connection (FC) is used for classification. The competitive results on eight databases demonstrate that the BMCAnet can effectively extract sparse and effective features for the classification tasks.

Currently, existing action recognition methods mainly use a data-driven method to extract spatio-temporal representations of actions for recognition. However, this method may face performance bottlenecks. At the same time, existing action recognition methods are easily affected by the bias of scene information and object information in videos. In order to explore the essential causal relationship between factors and remove bias in action recognition, we introduce the theory of causal inference into the field of action recognition and propose a Knowledge-based Hierarchical Causal Inference Network (KHCIN) to help us step toward a new direction of inference in action recognition. First, we construct a Knowledge-based Hierarchical Causal Graph (KHCG) to structurally represent the scene, object and motion knowledge of a video. Then, in the model inference stage, we perform factual causal inference on a video on the constructed KHCG, and then deploy counterfactual inference on the Direct Content Hierarchy (DCH) and Indirect Interaction Hierarchy (IIH) in the KHCG. For DCH, we intervene in the model at the decision level to highlight bias errors in the model predictions. For the IIH, we focus on intervening in the feature modelling process. The biased interactions are revealed by interrupting the information communication in the feature space. By comparing the results of factual and counterfactual inference, we can easily expose the biased information in the original representations and eliminate them. Driven by counterfactual causal inference, our approach can significantly improve the performance of action recognition while improving model explainability. Extensive experiments demonstrate the effectiveness of this method. We hope that KHCIN can provide some new ideas for better introduction of causal inference theory in the action recognition community in the future.

Real-world data typically follows a long-tailed distribution. When a small sample of tail classes does not cover the underlying distribution well, methods such as class re-balancing strategies and decoupled training are difficult to work, and additional knowledge needs to be introduced to recover the underlying distribution of the tail classes. In this work, we observe that the similarity between the variances of the feature distributions increases with the class similarity. Then, we also find that well-represented feature distributions typically contain multiple subcenters, which allows for denser samples at the edges of the distribution and promotes model learning to more robust decision bounds. Based on these observations, we propose to calibrate the feature distribution of the tail class by transferring the variance of the feature distribution of the head class, and then sample from the calibrated tail class distribution to generate augmented samples. To coordinate with the tail class calibration method, we also propose label-aware noise suppression (LANS) for reducing the generation of noisy samples and a three-stage training scheme for reshaping decision boundaries and compacting feature learning. Experimental results on iNaturalist2018, ImageNet-LT, CIFAR-10-LT, and CIFAR-100-LT show that our method achieves state-of-the-art performance in most metrics compared to similar approaches.

In recent years, object localization and detection methods in remote sensing images (RSIs) have received increasing attention due to their broad applications. However, most previous fully supervised methods require a large number of time-consuming and labor-intensive instance-level annotations. Compared with those fully supervised methods, weakly supervised object localization (WSOL) aims to recognize object instances using only image-level labels, which greatly saves the labeling costs of RSIs. In this article, we propose a self-directed weakly supervised strategy (SD-WSS) to perform WSOL in RSIs. To specify, we fully exploit and enhance the spatial feature extraction capability of the RSIs' classification model to accurately localize the objects of interest. To alleviate the serious discriminative region problem exhibited by previous WSOL methods, the spatial location information implicit in the classification model is carefully extracted by GradCAM++ to guide the learning procedure. Furthermore, to eliminate the interference from complex backgrounds of RSIs, we design a novel self-directed loss to make the model optimize itself and explicitly tell it where to look. Finally, we review and annotate the existing remote sensing scene classification dataset and create two new WSOL benchmarks in RSIs, named C45V2 and PN2. We conduct extensive experiments to evaluate the proposed method and six mainstream WSOL methods with three backbones on C45V2 and PN2. The results demonstrate that our proposed method achieves better performance when compared with state-of-the-arts.

The task of hyperspectral image (HSI) classification has attracted extensive attention. The rich spectral information in HSIs not only provides more detailed information but also brings a lot of redundant information. Redundant information makes spectral curves of different categories have similar trends, which leads to poor category separability. In this article, we achieve better category separability from the perspective of increasing the difference between categories and reducing the variation within category, thus improving the classification accuracy. Specifically, we propose the template spectrum-based processing module from spectral perspective, which can effectively expose the unique characteristics of different categories and reduce the difficulty of model mining key features. Second, we design an adaptive dual attention network from spatial perspective, where the target pixel can adaptively aggregate high-level features by evaluating the confidence of effective information in different receptive fields. Compared with the single adjacency scheme, the adaptive dual attention mechanism makes the ability of target pixel to combine spatial information to reduce variation more stable. Finally, we designed a dispersion loss from the classifier's perspective. By supervising the learnable parameters of the final classification layer, the loss makes the category standard eigenvectors learned by the model more dispersed, which improves the category separability and reduces the rate of misclassification. Experiments on three common datasets show that our proposed method is superior to the comparison method.

Modeling contextual relationships in images as graph inference is an interesting and promising research topic. However, existing approaches only perform graph modeling of entities, ignoring the intrinsic geometric features of images. To overcome this problem, a novel multiresolution interpretable contourlet graph network (MICGNet) is proposed in this article. MICGNet delicately balances graph representation learning with the multiscale and multidirectional features of images, where contourlet is used to capture the hyperplanar directional singularities of images and multilevel sparse contourlet coefficients are encoded into graph for further graph representation learning. This process provides interpretable theoretical support for optimizing the model structure. Specifically, first, the superpixel-based region graph is constructed. Then, the region graph is applied to code the nonsubsampled contourlet transform (NSCT) coefficients of the image, which are considered as node features. Considering the statistical properties of the NSCT coefficients, we calculate the node similarity, i.e., the adjacency matrix, using Mahalanobis distance. Next, graph convolutional networks (GCNs) are employed to further learn more abstract multilevel NSCT-enhanced graph representations. Finally, the learnable graph assignment matrix is designed to get the geometric association representations, which accomplish the assignment of graph representations to grid feature maps. We conduct comparative experiments on six publicly available datasets, and the experimental analysis shows that MICGNet is significantly more effective and efficient than other algorithms of recent years.

Most existing methods that cope with noisy labels usually assume that the classwise data distributions are well balanced. They are difficult to deal with the practical scenarios where training samples have imbalanced distributions, since they are not able to differentiate noisy samples from tail classes' clean samples. This article makes an early effort to tackle the image classification task in which the provided labels are noisy and have a long-tailed distribution. To deal with this problem, we propose a new learning paradigm which can screen out noisy samples by matching between inferences on weak and strong data augmentations. A leave-noise-out regularization (LNOR) is further introduced to eliminate the effect of the recognized noisy samples. Besides, we propose a prediction penalty based on the online classwise confidence levels to avoid the bias toward easy classes which are dominated by head classes. Extensive experiments on five datasets including CIFAR-10, CIFAR-100, MNIST, FashionMNIST, and Clothing1M demonstrate that the proposed method outperforms the existing algorithms for learning with long-tailed distribution and label noise.

The feature representation learning process greatly determines the performance of networks in classification tasks. By combining multiscale geometric tools and networks, better representation and learning can be achieved. However, relatively fixed geometric features and multiscale structures are always used. In this article, we propose a more flexible framework called the multiscale dynamic curvelet scattering network (MSDCCN). This data-driven dynamic network is based on multiscale geometric prior knowledge. First, multiresolution scattering and multiscale curvelet features are efficiently aggregated in different levels. Then, these features can be reused in networks flexibly and dynamically, depending on the multiscale intervention flag. The initial value of this flag is based on the complexity assessment, and it is updated according to feature sparsity statistics on the pretrained model. With the multiscale dynamic reuse structure, the feature representation learning process can be improved in the following training process. Also, multistage fine-tuning can be performed to further improve the classification accuracy. Furthermore, a novel multiscale dynamic curvelet scattering module, which is more flexible, is developed to be further embedded into other networks. Extensive experimental results show that better classification accuracies can be achieved by MSDCCN. In addition, necessary evaluation experiments have been performed, including convergence analysis, insight analysis, and adaptability analysis.

Domain generalization (DG) is one of the critical issues for deep learning in unknown domains. How to effectively represent domain-invariant context (DIC) is a difficult problem that DG needs to solve. Transformers have shown the potential to learn generalized features, since the powerful ability to learn global context. In this article, a novel method named patch diversity Transformer (PDTrans) is proposed to improve the DG for scene segmentation by learning global multidomain semantic relations. Specifically, patch photometric perturbation (PPP) is proposed to improve the representation of multidomain in the global context information, which helps the Transformer learn the relationship between multiple domains. Besides, patch statistics perturbation (PSP) is proposed to model the feature statistics of patches under different domain shifts, which enables the model to encode domain-invariant semantic features and improve generalization. PPP and PSP can help to diversify the source domain at the patch level and feature level. PDTrans learns context across diverse patches and takes advantage of self-attention to improve DG. Extensive experiments demonstrate the tremendous performance advantages of the PDTrans over state-of-the-art DG methods.

This article addresses the problem of the building an out-of-the-box deep detector, motivated by the need to perform anomaly detection across multiple hyperspectral images (HSIs) without repeated training. To solve this challenging task, we propose a unified detector [anomaly detection network (AUD-Net)] inspired by few-shot learning. The crucial issues solved by AUD-Net include: how to improve the generalization of the model on various HSIs that contain different categories of land cover; and how to unify the different spectral sizes between HSIs. To achieve this, we first build a series of subtasks to classify the relations between the center and its surroundings in the dual window. Through relation learning, AUD-Net can be more easily generalized to unseen HSIs, as the relations of the pixel pairs are shared among different HSIs. Secondly, to handle different HSIs with various spectral sizes, we propose a pooling layer based on the vector of local aggregated descriptors, which maps the variable-sized features to the same space and acquires the fixed-sized relation embeddings. To determine whether the center of the dual window is an anomaly, we build a memory model by the transformer, which integrates the contextual relation embeddings in the dual window and estimates the relation embeddings of the center. By computing the feature difference between the estimated relation embeddings of the centers and the corresponding real ones, the centers with large differences will be detected as anomalies, as they are more difficult to be estimated by the corresponding surroundings. Extensive experiments on both the simulation dataset and 13 real HSIs demonstrate that this proposed AUD-Net has strong generalization for various HSIs and achieves significant advantages over the specific-trained detectors for each HSI.

Recently, the excellent performance of transformer has attracted the attention of the visual community. Visual transformer models usually reshape images into sequence format and encode them sequentially. However, it is difficult to explicitly represent the relative relationship in distance and direction of visual data with typical 2-D spatial structures. Also, the temporal motion properties of consecutive frames are hardly exploited when it comes to dynamic video tasks like tracking. Therefore, we propose a novel dynamic polar spatio-temporal encoding for video scenes. We use spiral functions in polar space to fully exploit the spatial dependences of distance and direction in real scenes. We then design a dynamic relative encoding mode for continuous frames to capture the continuous spatio-temporal motion characteristics among video frames. Finally, we construct a complex-former framework with the proposed encoding applied to video-tracking tasks, where the complex fusion mode (CFM) realizes the effective fusion of scenes and positions for consecutive frames. The theoretical analysis demonstrates the feasibility and effectiveness of our proposed method. The experimental results on multiple datasets validate that our method can improve tracker performance in various video scenarios.

Considering a wide range of applications of nonnegative matrix factorization (NMF), many NMF and their variants have been developed. Since previous NMF methods cannot fully describe complex inner global and local manifold structures of the data space and extract complex structural information, we propose a novel NMF method called dual-graph global and local concept factorization (DGLCF). To properly describe the inner manifold structure, DGLCF introduces the global and local structures of the data manifold and the geometric structure of the feature manifold into CF. The global manifold structure makes the model more discriminative, while the two local regularization terms simultaneously preserve the inherent geometry of data and features. Finally, we analyze convergence and the iterative update rules of DGLCF. We illustrate clustering performance by comparing it with latest algorithms on four real-world datasets.

Deep learning-based clustering methods usually regard feature extraction and feature clustering as two independent steps. In this way, the features of all images need to be extracted before feature clustering, which consumes a lot of calculation. Inspired by the self-organizing map network, a self-supervised self-organizing clustering network ( [Formula: see text]OCNet) is proposed to jointly learn feature extraction and feature clustering, thus realizing a single-stage clustering method. In order to achieve joint learning, we propose a self-organizing clustering header (SOCH), which takes the weight of the self-organizing layer as the cluster centers, and the output of the self-organizing layer as the similarities between the feature and the cluster centers. In order to optimize our network, we first convert the similarities into probabilities which represents a soft cluster assignment, and then we obtain a target for self-supervised learning by transforming the soft cluster assignment into a hard cluster assignment, and finally we jointly optimize backbone and SOCH. By setting different feature dimensions, a Multilayer SOCHs strategy is further proposed by cascading SOCHs. This strategy achieves clustering features in multiple clustering spaces. [Formula: see text]OCNet is evaluated on widely used image classification benchmarks such as Canadian Institute For Advanced Research (CIFAR)-10, CIFAR-100, Self-Taught Learning (STL)-10, and Tiny ImageNet. Experimental results show that our method significant improvement over other related methods. The visualization of features and images shows that our method can achieve good clustering results.

Deep neural networks (DNNs) play key roles in various artificial intelligence applications such as image classification and object recognition. However, a growing number of studies have shown that there exist adversarial examples in DNNs, which are almost imperceptibly different from the original samples but can greatly change the output of DNNs. Recently, many white-box attack algorithms have been proposed, and most of the algorithms concentrate on how to make the best use of gradients per iteration to improve adversarial performance. In this article, we focus on the properties of the widely used activation function, rectified linear unit (ReLU), and find that there exist two phenomena (i.e., wrong blocking and over transmission) misguiding the calculation of gradients for ReLU during backpropagation. Both issues enlarge the difference between the predicted changes of the loss function from gradients and corresponding actual changes and misguide the optimized direction, which results in larger perturbations. Therefore, we propose a universal gradient correction adversarial example generation method, called ADV-ReLU, to enhance the performance of gradient-based white-box attack algorithms such as fast gradient signed method (FGSM), iterative FGSM (I-FGSM), momentum I-FGSM (MI-FGSM), and variance tuning MI-FGSM (VMI-FGSM). Through backpropagation, our approach calculates the gradient of the loss function with respect to the network input, maps the values to scores, and selects a part of them to update the misguided gradients. Comprehensive experimental results on ImageNet and CIFAR10 demonstrate that our ADV-ReLU can be easily integrated into many state-of-the-art gradient-based white-box attack algorithms, as well as transferred to black-box attacks, to further decrease perturbations measured in the -norm.

As a typical label-limited task, it is significant and valuable to explore networks that enable to utilize labeled and unlabeled samples simultaneously for synthetic aperture radar (SAR) image scene classification. Graph convolutional network (GCN) is a powerful semisupervised learning paradigm that helps to capture the topological relationships of scenes in SAR images. While the performance is not satisfactory when existing GCNs are directly used for SAR image scene classification with limited labels, because few methods to characterize the nodes and edges for SAR images. To tackle these issues, we propose a contrastive learning-based dual dynamic GCN (DDGCN) for SAR image scene classification. Specifically, we design a novel contrastive loss to capture the structures of views and scenes, and develop a clustering-based contrastive self-supervised learning model for mapping SAR images from pixel space to high-level embedding space, which facilitates the subsequent node representation and message passing in GCNs. Afterward, we propose a multiple features and parameter sharing dual network framework called DDGCN. One network is a dynamic GCN to keep the local consistency and nonlocal dependency of the same scene with the help of a node attention module and a dynamic correlation matrix learning algorithm. The other is a multiscale and multidirectional fully connected network (FCN) to enlarge the discrepancies between different scenes. Finally, the features obtained by the two branches are fused for classification. A series of experiments on synthetic and real SAR images demonstrate that the proposed method achieves consistently better classification performance than the existing methods.

The multispectral (MS) and the panchromatic (PAN) images belong to different modalities with specific advantageous properties. Therefore, there is a large representation gap between them. Moreover, the features extracted independently by the two branches belong to different feature spaces, which is not conducive to the subsequent collaborative classification. At the same time, different layers also have different representation capabilities for objects with large size differences. In order to dynamically and adaptively transfer the dominant attributes, reduce the gap between them, find the best shared layer representation, and fuse the features of different representation capabilities, this article proposes an adaptive migration collaborative network (AMC-Net) for multimodal remote-sensing (RS) images classification. First, for the input of the network, we combine principal component analysis (PCA) and nonsubsampled contourlet transformation (NSCT) to migrate the advantageous attributes of the PAN and the MS images to each other. This not only improves the quality of images themselves, but also increases the similarity between the two images, thereby reducing the representational gap between them and the pressure on the subsequent classification network. Second, for the interaction on the feature migrate branch, we design a feature progressive migration fusion unit (FPMF-Unit) based on the adaptive cross-stitch unit of correlation coefficient analysis (CCA), which can make the network automatically learn the features that need to be shared and migrated, aiming to find the best shared-layer representation for multifeature learning. And we design an adaptive layer fusion mechanism module (ALFM-Module), which can adaptively fuse features of different layers, aiming to clearly model the dependencies among multiple layers for different sized objects. Finally, for the output of the network, we add the calculation of the correlation coefficient to the loss function, which can make the network converge to the global optimum as much as possible. The experimental results indicate that AMC-Net can achieve competitive performance. And the code for the network framework is available at: https://github.com/ru-willow/A-AFM-ResNet.

Feature extraction is a key step for deep-learning-based point cloud registration. In the correspondence-free point cloud registration task, the previous work commonly aggregates deep information for global feature extraction and numerous shallow information which is positive to point cloud registration will be ignored with the deepening of the neural network. Shallow information tends to represent the structural information of the point cloud, while deep information tends to represent the semantic information of the point cloud. In addition, fusing information of different dimensions is conducive to making full use of shallow information. Inspired by this, we verify shallow information in the middle layers can bring a positive impact on the point cloud registration task. We design various architectures to combine shallow information and deep information to extract global features for point cloud registration. Experimental results on the ModelNet40 dataset illustrate that feature extractors that incorporate shallow information will bring positive performance.

The differentiable neural architecture search (NAS) framework has obtained extensive attention and achieved remarkable performance due to its search efficiency. However, most existing differentiable NAS methods still suffer from issues of model collapse, degenerated search-evaluation correlation, and inefficient hardware deployment, which causes the searched architectures to be suboptimal in accuracy and cannot meet different computation resource constraints (e.g., FLOPs and latency). In this article, we propose a novel resource-constrained NAS (ReCNAS) method, which can efficiently search high-performance architectures that satisfy the given constraints, and deal with the issues observed in previous differentiable NAS methods from three aspects: search space, search strategy, and resource adaptability. First, we introduce an elastic densely connected layerwise search space, which decouples the architecture depth representation from the search of candidate operations to alleviate the aggregation of skip connections and architecture redundancies. Second, a scheme of group annealing and progressive pruning is proposed to improve the efficiency and bridge the search-evaluation gap, which steadily forces the architecture parameters close to binary distribution and progressively prunes the inferior operations. Third, we present a novel resource-constrained architecture generation method, which prunes the redundant channel throughout the search based on dynamic programming, making the searched architecture scalable to different devices and requirements. Extensive experimental results demonstrate the efficiency and search stability of our ReCNAS, which is capable of discovering high-performance architectures on different datasets and tasks, surpassing other NAS methods, while tightly meeting the target resource constraints without any tuning required. Besides, the searched architectures show strong generalizability to other complex vision tasks.

This article focuses on end-to-end image matching through joint key-point detection and descriptor extraction. To find repeatable and high discrimination key points, we improve the deep matching network from the perspectives of network structure and network optimization. First, we propose a concurrent multiscale detector (CS-det) network, which consists of several parallel convolutional networks to extract multiscale features and multilevel discriminative information for key-point detection. Moreover, we introduce an attention module to fuse the response maps of various features adaptively. Importantly, we propose two novel rank consistent losses (RC-losses) for network optimization, significantly improving image matching performances. On the one hand, we propose a score rank consistent loss (RC-S-loss) to ensure that the key points have high repeatability. Different from the score difference loss merely focusing on the absolute score of an individual key point, our proposed RC-S-loss pays more attention to the relative score of key points in the image. On the other hand, we propose a score-discrimination RC-loss to ensure that the key point has high discrimination, which can reduce the confusion from other key points in subsequent matching and then further enhance the accuracy of image matching. Extensive experimental results demonstrate that the proposed CS-det improves the mean matching result of deep detector by 1.4%-2.1%, and the proposed RC-losses can boost the matching performances by 2.7%-3.4% than score difference loss. Our source codes are available at https://github.com/iquandou/CS-Net.

For complex data, high dimension and high noise are challenging problems, and deep matrix factorization shows great potential in data dimensionality reduction. In this article, a novel robust and effective deep matrix factorization framework is proposed. This method constructs a dual-angle feature for single-modal gene data to improve the effectiveness and robustness, which can solve the problem of high-dimensional tumor classification. The proposed framework consists of three parts, deep matrix factorization, double-angle decomposition, and feature purification. First, a robust deep matrix factorization (RDMF) model is proposed in the feature learning, to enhance the classification stability and obtain better feature when faced with noisy data. Second, a double-angle feature (RDMF-DA) is designed by cascading the RDMF features with sparse features, which contains the more comprehensive information in gene data. Third, to avoid the influence of redundant genes on the representation ability, a gene selection method is proposed to purify the features by RDMF-DA, based on the principle of sparse representation (SR) and gene coexpression. Finally, the proposed algorithm is applied to the gene expression profiling datasets, and the performance of the algorithm is fully verified.

Recently, deep multi-view clustering (MVC) has attracted increasing attention in multi-view learning owing to its promising performance. However, most existing deep multi-view methods use single-pathway neural networks to extract features of each view, which cannot explore comprehensive complementary information and multilevel features. To tackle this problem, we propose a deep structured multi-pathway network (SMpNet) for multi-view subspace clustering task in this brief. The proposed SMpNet leverages structured multi-pathway convolutional neural networks to explicitly learn the subspace representations of each view in a layer-wise way. By this means, both low-level and high-level structured features are integrated through a common connection matrix to explore the comprehensive complementary structure among multiple views. Moreover, we impose a low-rank constraint on the connection matrix to decrease the impact of noise and further highlight the consensus information of all the views. Experimental results on five public datasets show the effectiveness of the proposed SMpNet compared with several state-of-the-art deep MVC methods.

Accurately extracting buildings from aerial images has essential research significance for timely understanding human intervention on the land. The distribution discrepancies between diversified unlabeled remote sensing images (changes in imaging sensor, location, and environment) and labeled historical images significantly degrade the generalization performance of deep learning algorithms. Unsupervised domain adaptation (UDA) algorithms have recently been proposed to eliminate the distribution discrepancies without re-annotating training data for new domains. Nevertheless, due to the limited information provided by a single-source domain, single-source UDA (SSUDA) is not an optimal choice when multitemporal and multiregion remote sensing images are available. We propose a multisource UDA (MSUDA) framework SPENet for building extraction, aiming at selecting, purifying, and exchanging information from multisource domains to better adapt the model to the target domain. Specifically, the framework effectively utilizes richer knowledge by extracting target-relevant information from multiple-source domains, purifying target domain information with low-level features of buildings, and exchanging target domain information in an interactive learning manner. Extensive experiments and ablation studies constructed on 12 city datasets prove the effectiveness of our method against existing state-of-the-art methods, e.g., our method achieves 59.1% intersection over union (IoU) on Austin and Kitsap → Potsdam, which surpasses the target domain supervised method by 2.2%. The code is available at https://github.com/QZangXDU/SPENet.

In recent years, convolutional neural networks (CNNs)-based methods achieve cracking performance on hyperspectral image (HSI) classification tasks, due to its hierarchical structure and strong nonlinear fitting capacity. Most of them, however, are supervised approaches that need a large number of labeled data to train them. Conventional convolution kernels are fixed shape of rectangular with fixed sizes, which are good at capturing short-range relations between pixels within HSIs but ignore the long-range context within HSIs, limiting their performance. To overcome the limitations mentioned above, we present a dynamic multiscale graph convolutional network (GCN) classifier (DMSGer). DMSGer first constructs a relatively small graph at region-level based on a superpixel segmentation algorithm and metric-learning. A dynamic pixel-level feature update strategy is then applied to the region-level adjacency matrix, which can help DMSGer learn the pixel representation dynamically. Finally, to deeply understand the complex contents within HSIs, our model is expanded into a multiscale version. On the one hand, by introducing graph learning theory, DMSGer accomplishes HSI classification tasks in a semi-supervised manner, relieving the pressure of collecting abundant labeled samples. Superpixels are generally in irregular shapes and sizes which can group only similar pixels in a neighborhood. On the other hand, based on the proposed dynamic-GCN, the pixel-level and region-level information can be captured simultaneously in one graph convolution layer such that the classification results can be improved. Also, due to the proper multiscale expansion, more helpful information can be captured from HSIs. Extensive experiments were conducted on four public HSIs, and the promising results illustrate that our DMSGer is robust in classifying HSIs. Our source codes are available at https://github.com/TangXu-Group/DMSGer.

The emergence of neural architecture search (NAS) algorithms has removed the constraints on manually designed neural network architectures, so that neural network development no longer requires extensive professional knowledge, trial and error. However, the extremely high computational cost limits the development of NAS algorithms. In this article, in order to reduce computational costs and to improve the efficiency and effectiveness of evolutionary NAS (ENAS) is investigated. In this article, we present a fast ENAS framework for multiscale convolutional networks based on evolutionary knowledge transfer search (EKTS). This framework is novel, in that it combines global optimization methods with local optimization methods for search, and searches a multiscale network architecture. In this article, evolutionary computation is used as a global optimization algorithm with high robustness and wide applicability for searching neural architectures. At the same time, for fast search, we combine knowledge transfer and local fast learning to improve the search speed. In addition, we explore a multiscale gray-box structure. This gray box structure combines the Bandelet transform with convolution to improve network approximation, learning, and generalization. Finally, we compare the architectures with more than 40 different neural architectures, and the results confirmed its effectiveness.

Due to the uncertainty of non-cooperative communication channels, the received signals often contain various impairment factors, leading to a significant decline in the performance of existing deep learning (DL)-based automatic modulation classification (AMC) models. Several preliminary works utilize domain adaptation (DA) to alleviate this issue, however, they are constrained by singular domain difference factor, whereas in practice, these factors often manifest cumulatively. Therefore, this paper introduce a more realistic task named superimposed DA, where multiple domain difference factors are overlaid, reflecting the cumulative nature of them. We propose the SigDA as a solution framework, which adopts adversarial training to align the data distribution in different domains. Two technical modules, Multi-task based Masked Signal Feature Extractor (M2SFE) and Signal Feature Pyramid Aggregation (SFPA), are innovatively designed in SigDA. M2SFE utilizes mask and reconstruction task to enhance feature extraction and achieves discriminative feature selection through the design of feature mapping layers, while SFPA can solve the problem of inconsistent signal length in superimposed DA and can aggregate the features of signals into the same dimension. We consider and superimpose various typical signal domain difference factors, comprehensive experiments demonstrate that the proposed framework can achieve significant performance improvement in various communication channels.

Recently, contrastive learning (CL) has exhibited considerable advantages for automatic modulation classification (AMC) with a scarcity of labeled samples. Nevertheless, the majority of the available CL-based AMC methods use the simple signal augmentation strategy and suffer from interference from false negatives. To explore the more generalizable global temporal semantics within signals, a novel masked contrastive learning with hard negatives (MCLHN) method is proposed in this paper. MCLHN first strategically incorporates semantic-preserving data augmentation, ensuring the diversity and semantic invariance of signals. Second, MCLHN adopts an encoder with temporal masking to enable robust temporal modeling. Moreover, a debiased hardness-weighted contrastive (DHWC) loss is designed to balance the adverse impact of the debiased strategy and the advantage of hard negatives. Extensive experiments are conducted on several benchmark datasets, and the experimental results demonstrate the superior performance and generalization capability of MCLHN to other methods. Significantly, the performance of MCLHN with only one labeled sample per modulation under each signal-to-noise ratio (SNR) rivals that of other methods with five to twenty times the number of labeled samples.

To address the challenges of long-tailed classification, researchers have proposed several approaches to reduce model bias, most of which assume that classes with few samples are weak classes. However, recent studies have shown that tail classes are not always hard to learn, and model bias has been observed on sample-balanced datasets, suggesting the existence of other factors that affect model bias. In this work, we systematically propose a series of geometric measurements for perceptual manifolds in deep neural networks, and then explore the effect of the geometric characteristics of perceptual manifolds on classification difficulty and how learning shapes the geometric characteristics of perceptual manifolds. An unanticipated finding is that the correlation between the class accuracy and the separation degree of perceptual manifolds gradually decreases during training, while the negative correlation with the curvature gradually increases, implying that curvature imbalance leads to model bias. Therefore, we propose curvature regularization to facilitate the model to learn curvature-balanced and flatter perceptual manifolds. Evaluations on multiple long-tailed and non-long-tailed datasets show the excellent performance and exciting generality of our approach, especially in achieving significant performance improvements based on current state-of-the-art techniques. Our work opens up a geometric analysis perspective on model bias and reminds researchers to pay attention to model bias on non-long-tailed and even sample-balanced datasets. The code and model will be made public.

Unsupervised domain adaptation (UDA) in semantic segmentation transfers the knowledge of the source domain to the target one to improve the adaptability of the segmentation model in the target domain. The need to access labeled source data makes UDA unable to handle adaptation scenarios involving privacy, property rights protection, and confidentiality. In this paper, we focus on unsupervised model adaptation (UMA), also called source-free domain adaptation, which adapts a source-trained model to the target domain without accessing source data. We find that the online self-training method has the potential to be deployed in UMA, but the lack of source domain loss will greatly weaken the stability and adaptability of the method. We analyze two reasons for the degradation of online self-training, i.e. inopportune updates of the teacher model and biased knowledge from the source-trained model. Based on this, we propose a dynamic teacher update mechanism and a training-consistency based resampling strategy to improve the stability and adaptability of online self-training. On multiple model adaptation benchmarks, our method obtains new state-of-the-art performance, which is comparable or even better than state-of-the-art UDA methods. The code is available at https://github.com/DZhaoXd/DT-ST.

Domain adaptive semantic segmentation aims to adapt a model trained on labeled source domain to unlabeled target domain. Self-training shows competitive potential in this field. Existing methods along this stream mainly focus on selecting reliable predictions on target data as pseudo-labels for category learning, while ignoring the useful relations between pixels for relation learning. In this paper, we propose a pseudo-relation learning framework, Relation Teacher (RTea), which can exploitable pixel relations to efficiently use unreliable pixels and learn generalized representations. In this framework, we build reasonable pseudo-relations on local grids and fuse them with low-level relations in the image space, which are motivated by the reliable local relations prior and available low-level relations prior. Then, we design a pseudo-relation learning strategy and optimize the class probability to meet the relation consistency by finding the optimal sub-graph division. In this way, the model’s certainty and consistency of prediction are enhanced on the target domain, and the cross-domain inadaptation is further eliminated. Extensive experiments on three datasets demonstrate the effectiveness of the proposed method. The code will be available at https://github.com/DZhaoXd/RTea.

For GAN-based Person Re-identification (Re-ID), the key is to generate pedestrian images with higher identity consistency and meanwhile larger intra-class diversity. Generally, the main discriminative parts focus on some local regions from the foreground of each pedestrian image for Re-ID, and they should be irrelevant to the background. Whereas, most existing methods generate pedestrian images only based on global features, which difficultly achieves emphasizing crucial local regions and weakening the background. Based on this, we propose a CAM-enhancing generative Re-ID method in which the global and local features are jointly used. In the proposed method, an adaptive CAM-enhancing local encoder is designed to explore the significance of local appearances and enhance the effect of crucial local features in generations, where the foreground is divided into multiple local parts and separated from the background by pedestrian segmentation. Moreover, a new generation loss is proposed to supervise the identity consistency by reducing the inconsistency of crucial regions in foregrounds and meanwhile enrich the intra-class diversity by generating variant backgrounds. Experimental results indicate that the proposed method obtains better generation images and Re-ID performance than other methods.

Feature descriptor-based methods have demonstrated remarkable performance in remote sensing image patch matching tasks and are usually optimized using contrastive loss and triplet loss. However, these optimization losses focus on calculating the distance between samples, ignoring the rich information of higher-order feature relationships between multiple image patches. The latter provides valuable information that can be used to improve task performance. Inspired by the superior performance of second-order relations in graph matching and clustering tasks, we aim to exploit the rich information available from high-order relations fully. This paper proposes a high-order relationship (HOR) learning method for remote sensing image patch matching. This method combines low-order feature relations between image patch pairs and high-order feature relations between multiple patches to enhance image matching performance. Extensive experimental results on a multimodel remote sensing image dataset, SEN 1-2, consisting of optical and SAR images, demonstrate that the proposed HOR learning method can improve the performance of remote sensing image patch matching.

The change detection task of remote sensing images is a basic scientific problem, and has been further widely used in real life. Recently, transformer model has shown strong learning and representation abilities in visual interpretation. In this article, Inspired by the success of the Vision Transformer and its variants, we propose a novel change detection model for remote sensing images, named Swin ResNet Transformers (Swin ResNet). Different from other methods, the proposed Swin ResNet architecture uses a Swin transform encoder, which extracts feature representations of multiple resolutions through a shift window mechanism to calculate self-attention. On three datasets, the proposed model showed good performance, and demonstrate that the Swin transformer has a strong ability to learn long-term dependencies of multi-scale context representation.

Fine-grained building classification provides a solid basis for the comparison of city morphologies and the investigation of urban planning. To this aim, the DFC23 establishes a large-scale and multi-modal benchmark for the classification of building roof types. However, the problems of long-tailed distribution, data insufficient, inter-class similarity, and intra-class difference severely inhibit the performance of the detector. In this work, we build a strong vision transformer adapter fine-tuned on the cropped building instances to enhance the capacity of feature extraction and design a cross-modal fusion (CMF) module to effectively aggregate features from RGB and SAR data. When transferring to building instance segmentation, we construct a robust training pipeline and a two-stage test-time results ensemble scheme. Furthermore, we introduce self-training with two key denoising techniques, global average filtering (GAF) and intra-class adaptive thresholding (IAT), to boost the generalization of the model. Experimental results show the effectiveness of our method, ranking 2nd in the test phase of the contest.

Building extraction and height estimation provide solid fundamentals for reconstructing city morphologies and investigating urban planning. To this aim, the DFC23 establishes a large-scale and multi-modal benchmark for multi-task learning of building reconstruction. However, the problems of data limitation and fore-background confusion severely inhibit the performance of the model. In this work, we propose a novel trident cooperation network (TCNet) to perform end-to-end building extraction and height estimation using RGB and SAR data. Specifically, to enrich the feature representation and generalization of the shared backbone, we introduce a vision transformer adapter to inject vision-specific inductive biases and design a cross-modal fusion (CMF) module to effectively aggregate features from multi-modal data. For downstream visual tasks, we construct trident decoders including a detector, a lightweight MLP segmentation head, and a pixel-wise regression head. Moreover, to highlight the foreground object, we use the binary mask predicted by the MLP head to cooperate with the height estimation map predicted by the estimator. And the weighted sub-task losses are gathered to optimize our TCNet. Experimental results show the effectiveness of our method, ranking 2nd in the test phase of the contest.

Multi-modal images can obtain complementary and rich information images, which are more widely used in various applications. However, due to the different imaging mechanisms of different sensors, there are significant domain distribution differences between multi-modal images. In multi-modal image matching, existing deep learning methods should deal with the image content difference caused by rotation transformation and the domain distribution difference caused by different sensors, which are very difficult for the deep network. To address this issue, we propose to combine an instance comparison and a batch comparison to deal with image content differences and domain distribution differences, respectively. We design a new domain distribution alignment method to explicitly constrain the sample domain distribution of the multi-modal images are consistent through the domain distribution alignment loss. Extensive multi-modal remote sensing image patch matching experiments have shown the effectiveness of the proposed method. Furthermore, the proposed multi-modal domain distribution alignment method has more obvious advantages when there are significant content differences and distribution differences.

Unsupervised domain adaptation (UDA) has been widely used in hyperspectral image (HSI) classification. Domain adversarial learning methods and self-training methods are two major UDA methods. Most existing methods use these two methods independently, which limit their capacity to knowledge transferability and robustness of classification. Therefore, a novel domain adversarial debiased self-training (DADST) is proposed to combine these two methods, where domain adversarial learning reduces the domain discrepancy and debiased self-training updates model via a self-paced curriculum policy. To this end, we first apply debiased self-training into HSI classification. Then domain adversarial learning module is combined to align joint distribution between the source domain and the target domain. Experimental results on two cross-scene HSI datasets demonstrate that the proposed DADST method outperforms other domain adaptation approaches.

The recent explosive on Transformer has suggested its potential to become the mainstream model for feature representation and classification of remote sensing. Although Transformer has excellent global modeling capabilities, it lacks inductive bias. In contrast, CNNs show excellent performance in computer vision due to their strong inductive bias. To solve the above issue, and benefit from combining the Transformer model with CNNs, a dense cross-scale Transformer is proposed for remote sensing scene classification. Firstly, the attention aggregation based feature pyramid network (A2-FPN) is adopted to obtain the multi-scale features. And then, the multi-scale features are input into the multi-scale global features learning with channel attention (MSCA) module to obtain the global features. Besides, the multi-scale features are input into the dense cross-scale attention (DCSA) module to learn the multi-level cross-scale features. Finally, the outputs of these two modules are considered for computing the final class score. Experimental results obtained from the three public datasets indicate that the proposed method surpasses other remote sensing classification methods.

Due to the powerful feature extraction capabilities of deep neural networks, traditional approaches are gradually replaced by deep learning approaches for image matching tasks. For multi-modal image patch matching, the deep model should mainly learn the modality-invariant features. For multi-modal images with rotation transformation (RT), the deep model should learn the modality-invariant features and rotation-invariant features simultaneously. However, the performance of the latter trained model is degraded for the former task. The main reason is that the modality invariance of the features degenerates. This paper proposes a deep multi-modal remote sensing image matching network (DCMNet) that combines descriptor learning and continuous learning to solve this problem. Firstly, DCMNet is trained for learning modality-invariant features in multi-modal image patch matching. Then, DCMNet is optimized for multi-modal image patch matching with RT. In the later learning process, we reduce the change of important parameters for the modality-invariant features learning. Experiments demonstrate the effectiveness and robustness of DCMNet in alleviating the modal invariance degradation problem of features.

Object detectors based on convolutional neural networks have been widely-applied to detect moving vehicles in satellite videos. However, many detectors render superior detection accuracy at the expense of increased computational complexity and decreased inference speed. This prevents these detectors from being deployed into mobile devices. In this paper, an efficient 3D motion-guided lightweight network (3D-MGLNet) is proposed. Specifically, 3D-MGLNet constructs a motion-guided module based on 3D convolution to extract motion cues from spatial-temporal information. This module uses model compression strategies to detect moving vehicles in real-time while following the principle of "fewer channels, smaller convolution kernels," significantly reducing the number of parameters and computational complexity. Extensive experiments are conducted on the Jilin-1 and SkySat satellite video datasets. The results demonstrate that 3D-MGLNet gains strong performance by striking an excellent tradeoff between resource and accuracy, resulting in the fewest parameters (0.35M) and fastest speed (66.84 fps) compared to other popular models.

In scenarios with long-tailed distributions, the model's ability to identify tail classes is limited due to the under-representation of tail samples. Class rebalancing, information augmentation, and other techniques have been proposed to facilitate models to learn the potential distribution of tail classes. The disadvantage is that these methods generally pursue models with balanced class accuracy on the data manifold, while ignoring the ability of the model to resist interference. By constructing noisy data manifold, we found that the robustness of models trained on unbalanced data has a long-tail phenomenon. That is, even if the class accuracy is balanced on the data domain, it still has bias on the noisy data manifold. However, existing methods cannot effectively mitigate the above phenomenon, which makes the model vulnerable in long-tailed scenarios. In this work, we propose an Orthogonal Uncertainty Representation (hOUR) of feature embedding and an end-to-end training strategy to improve the long-tail phenomenon of model robustness. As a general enhancement tool, OUR has excellent compatibility with other methods and does not require additional data generation, ensuring fast and efficient training. Comprehensive evaluations on long-tailed datasets show that our method significantly improves the long-tail phenomenon of robustness, bringing consistent performance gains to other long-tailed learning methods.

To enable machines to mimic human cognitive abilities and alleviate the catastrophic forgetting problem in cross-modal image-text retrieval (CMITR), this paper proposes a novel continual learning method, Knowledge Decomposition and Replay (KDR), which emulates the process of knowledge decomposition and replay exhibited by humans in complex and changing environments. KDR has two components: a feature Decomposition-based CMITR Model (DCM) and a cross-task Generic Knowledge Replay strategy (GKR). DCM decomposes text and image features into task-specific and generic knowledge features, mimicking the human cognitive process of knowledge decomposition. Specifically, it employs a generic knowledge features extraction module for all tasks and a task-specific module for each task with a few trainable fully connected layers. Similarly, GKR emulates the human behavior of knowledge replay by utilizing the image-text similarity matrix output from the old task model with inputting the previous samples to induce the learning of the image-text similarity matrix output from the current task model with inputting the previous samples, using knowledge distillation technology. To demonstrate the effect of KDR, we adapted a continual learning dataset Seq-COCO from MSCOCO. Extensive experiments on Seq-COCO showed that KDR reduces catastrophic forgetting and consolidates general knowledge, improving the model's learning ability in CMITR.

In video analysis, an important challenge is insufficient annotated data due to the rare occurrence of the critical patterns, and we need to provide discriminative frame-level representation with limited annotation in some applications. Multiple Instance Learning (MIL) is suitable for this scenario. However, many MIL models paid attention to analyzing the relationships between instance representations and aggregating them, but neglecting the critical information from the MIL problem itself, which causes difficultly achieving ideal instance-level performance compared with the supervised model. To address this issue, we propose the $\textbf{\textit{Regressor-Guided MIL network} (RGMIL)}$, which effectively produces discriminative instance-level representations in a general multi-classification scenario. In the proposed method, we make full use of the $\textit{regressor}$ through our newly introduced $\textit{aggregator}$, $\textbf{\textit{Regressor-Guided Pooling} (RGP)}$. RGP focuses on simulating the correct inference process of humans while facing similar problems without introducing new parameters, and the MIL problem can be accurately described through the critical information from the $\textit{regressor}$ in our method. In experiments, RGP shows dominance on more than 20 MIL benchmark datasets, with the average bag-level classification accuracy close to 1. We also perform a series of comprehensive experiments on the MMNIST dataset. Experimental results illustrate that our $\textit{aggregator}$ outperforms existing methods under different challenging circumstances. Instance-level predictions are even possible under the guidance of RGP information table in a long sequence. RGMIL also presents comparable instance-level performance with S-O-T-A supervised models in complicated applications. Statistical results demonstrate the assumption that a MIL model can compete with a supervised model at the instance level, as long as a structure that accurately describes the MIL problem is provided. The codes are available on $\url{https://github.com/LMBDA-design/RGMIL}$.

Precise crop classification plays a significant role in the agriculture field. An appropriate data source for precise crop classification is high spatial resolutions hyperspectral imagery (H2 imagery) acquired by unmanned aerial vehicle (UAV). However, for imagery with many different classes of crops, crop classification of UAV H2 imagery is a huge challenge. The significant spectral diversity, spatial heterogeneity and nonlinear data structure of UAV H2 imagery results in poor spectral discriminability. To improve the discriminability, a kernel tensor slice sparse coding-based classifier (KTSSCC) is proposed for precise crop classification of UAV H2 imagery in this research. The kernel tensor representation mechanism in KTSSCC can reduce the nonlinear separation while well preserving the spectral characteristics and spatial constraints of land-covers, and thus the discriminability is greatly improved. Furthermore, this paper puts forward the kernel tensor slice sparse orthogonal matching pursuit (KTSSOMP) algorithm to optimize kernel tensor slice sparse coding in the spectral space, which greatly reduces the computation cost. Moreover, there are very few parameters to be tuned in our proposed model. We assess the performance of KTSSCC on two real UAV hyperspectral imagery datasets, and find that, based on visual and quantitative results, it provides satisfactory crop classification results and outperforms the state-of-the-art approaches.

When a top-down method is taken to the task of human pose estimation, the accuracy of joint point localization is often limited by the accuracy of human detection. In addition, conventional algorithms commonly encode the image to generate a heat map before processing, but the systematic error in decoding the heat map back to the original image has an impact on the positioning. Therefore, to address the two problems, we propose an algorithm that uses multiple angle models to generate the human boxes and then performs lightweight decoding to recover the image. The new boxes can better fit humans and the recovery error can be reduced. First, we split the backbone network into three sub-networks, the first sub-network is responsible for generating the original human box, the second sub-network is responsible for generating a coarse pose estimation in the boxes, and the third sub-network is responsible for a high-precision pose estimation. In order to make the human box fit the human body better, with only a small number of interfering pixels inside the box, models of the human boxes with multiple rotation angles are generated. The results from the second sub-network are used to select the best human box. Using this human box as input to the third sub-network can significantly improve the accuracy of the pose estimation. Then to reduce the errors arising from image decoding, we propose a lightweight unbiased decoding strategy that differs from traditional methods by combining multiple possible offsets to select the direction and size of the final offset. On the MPII dataset and the COCO dataset, we compare the proposed algorithm with 11 state-of-the-art algorithms. The experimental results show that the algorithm achieves a large improvement in accuracy for a wide range of image sizes and different metrics.

Specific emitter identification (SEI) is crucial for attacking and defending Internet of Things (IoT) devices in untrusted scenarios or battlefield environments. However, existing SEI methods usually require annotation information, which is often unavailable in noncooperative communications and untrusted scenarios. In this article, we propose a signal contrastive self-supervised clustering (SCSC) method for unsupervised SEI applications. First, we propose SCSC with 1-D fingerprint pyramid feature extractor (1D-FPFE) for obtaining hierarchical subtle features of emitter signals. Then, we propose a bit-pulse selection (BPS) strategy and several signal data augmentation methods. By constructing signal positive and negative instance pairs through data augmentation, our approach generates cluster preference representations in a contrastive self-supervised learning manner. Extensive experimental results based on communication burst emitter dataset show that SCSC achieves an accuracy improvement of about 26% over the current best communication signal clustering algorithm. Moreover, SCSC also exhibits good performance and generalization for 30 emitter clustering and few-shot unlabeled signal clustering.

Internet of Things (IoT) networks are often subject to many malicious attacks in untrusted environments, and automatic modulation classification (AMC) is an effective way to combat IoT physical-layer threats. However, most existing AMC methods assume sufficient labeled signals and invariant signal distribution, which is often impossible in untrusted environments. In this article, a new meta-learning method is proposed for a few-shot AMC with distribution bias. First, a multi-frequency octave ResNet (MFOR) is constructed to learn coarse (low-frequency) and fine (high-frequency) features, which can efficiently identify the modulation type of the signal while saving computational resources. Second, a large number of classification-related meta-tasks are established for training MFOR to explore general knowledge in signal classification, and then transfer it to the AMC. Different with deep neural networks (DNNs) that learn a mapping by multiple instances, the MFOR with meta-learning (denoted as M-MFOR) can improve the generalization ability of new AMC tasks with very few instances and distribution bias. Furthermore, we find that the distribution bias between data can be reduced by adjusting the normalized distribution and propose a class-related mixup. Extensive experiments are taken on several datasets to investigate the effectiveness of M-MFOR. The results show its feasibility and superiority over existing methods.

Most existing classification methods design complicated and large deep neural network (DNN) model to deal with the ubiquitous spectral variability and nonlinearity of hyperspectral images (HSIs). However, their application is blocked by limited training samples and considerable computational costs in real scenes. To solve these problems, we propose a simple spectral hierarchical feature fusion and selection network (HFFSNet). Specifically, we apply 1-D grouped convolution for dimensionality reduction and multilevel feature extraction, then the multilevel features are fused to assist the adaptive feature selection of different layer features via the soft attention mechanism, and finally the selected features are fused to further enhance the feature representation. Extensive experimental results on three hyperspectral datasets demonstrate the effectiveness of the proposed network.

Synthetic aperture radar (SAR) image change detection is a hot but challenging task due to SAR images’ complex contents and inherent speckle noises. The expected change detection methods should reduce the influence of speckle noises, obtain the discriminative feature representations, and generate accurate change maps simultaneously. To these ends, we propose a new SAR image change detection method named feature fusion of information transfer network (FFITN). First, we develop a hybrid convolution block to depress the speckle noise impacts and explore the valuable information from SAR images. Thus, the feature extraction module (FEM) is constructed to obtain the multi-level features. Then, an information transfer module (ITM) is proposed to capture the salient regions from various aspects. Also, the salient knowledge is transferred among features at different levels to enhance their discrimination. Next, a self-attention-based feature fusion module (SAFFM) is introduced to fuse various features. Finally, a change map generation module (CMGM) with the clustering algorithm and specific loss functions is designed to produce the pseudo labels and change maps. Experimental results on three public SAR data sets demonstrate the model’s effectiveness. Our source codes are available at https://github.com/TangXu-Group/FFITN.

In recent years, the joint classification approach of hyperspectral image (HSI) and light detection and ranging (LiDAR) data based on deep learning (DL) has received increasing attention. However, existing methods either lack interaction between heterogeneous features during feature extraction or treat them equally during the interaction, inevitably resulting in redundant information stacking and reaching the performance bottleneck. To this end, we propose a novel discrepant bi-directional interaction fusion network (DBIFNet) for the collaborative classification of HSI and LiDAR data. First, a discrepant bi-directional interaction module (DBDIM) is designed to establish correlations between heterogeneous features to enhance the respective feature learning. Furthermore, a cross-modal attention fusion module (CAFM) is developed to dynamically fuse multi-modal features, which can further improve classification performance. Extensive experiments on the Houston and Trento datasets demonstrate that the proposed DBIFNet can achieve competitive classification performance.

In this letter, a kernel tensor sparse coding model (KTSCM) is proposed for precise crop classification of unmanned aerial vehicle (UAV) hyperspectral image (HSI). Benefited from the kernel tensor representation mechanism in KTSCM, which can not only improve the linear separation but also well preserving the spatial-spectral structures of land-covers, the discriminability of UAV HSI is greatly improved. The L1-norm based tensor sparsity makes the tensor operation in KTSCM can be equivalently converted to matrix operation, which greatly reduces the computation cost. Furthermore, the analytical solution to KTSCM allows it be well optimized with very few iterations. The performance of KTSCM is assessed on two real UAV HSIs. The experimental results indicate that KTSCM can provides rapid and accurate crop classification results with limited labeled pixels and outperforms the related counterparts.

Bilinear pooling is a feature aggregation step applied after the convolutional layers of a deep network and encodes a matrix of local features into a fixed-size bilinear representation. It improves performance in many image classification tasks. Since its emergence, this pooling has seen two major improvements: Compact Bilinear Pooling (CBP) and square-root normalization. Recently, the combination of these two elements has been widely studied. However, due to the lack of good normalization solutions, existing combination approaches showed less efficiency when they are plugged into different networks and less compatibility when they work with existing CBP techniques. To solve this problem, in this paper, we propose to apply Newton iterations, a fast square-root normalization method, to produce a new normalized matrix calledpseudo square-rooted matrix. Subsequently, the new matrix allows a CBP technique to encode itself into a compact and normalized bilinear representation. In order to further accelerate the normalization process, our approach has two variants which can handle feature matrix extracted by different networks. Tested on three fine-grained image classification datasets, it provides competitive classification performance while consuming less computational time than other prior works.

Existing two-dimensional principal component analysis methods can only handle second-order tensors (i.e., matrices). However, with the advancement of technology, tensors of order three and higher are gradually increasing. This brings new challenges to dimensionality reduction. Thus, a multilinear method called MPCA was proposed. Although MPCA can be applied to all tensors, using the square of the F-norm makes it very sensitive to outliers. Several two-dimensional methods, such as Angle 2DPCA, have good robustness but cannot be applied to all tensors. We extend the robust Angle 2DPCA method to a multilinear method and propose Cosine Multilinear Principal Component Analysis (CosMPCA) for tensor representation. Our CosMPCA method considers the relationship between the reconstruction error and projection scatter and selects the cosine metric. In addition, our method naturally uses the F-norm to reduce the impact of outliers. We introduce an iterative algorithm to solve CosMPCA. We provide detailed theoretical analysis in both the proposed method and the analysis of the algorithm. Experiments show that our method is robust to outliers and is suitable for tensors of any order.

The concealed object detection in millimeter-wave human body images is a challenging task due to the noise and dim-small objects. Exploiting the spatial dependencies to mine the difference between the object and the noise is vital for the discrimination of objects. However, most approaches ignore the context around the object. In this paper, a concealed object detection framework based on structural context is proposed to suppress noise interference and refine localizable semantic features. The framework consists of two subnetworks, structural region-based multi-scale weakly supervised feature refinement and local context-based concealed object detection. The multi-scale weakly supervised feature refinement is constructed to learn position-aware semantics of objects of various sizes while suppressing background noises in structural regions. Specifically, a multi-scale pooling method is proposed to better localize objects of different sizes, and an object-activated region enhancement module is designed to strengthen object semantic representations and suppress the background interference. Moreover, an adaptive local context aggregation module is designed to integrate the local context around the bounding box in the concealed object detection, which improves the discrimination of the model for the dim-small objects. Experimental results on the AMMW and the PMMW datasets demonstrate that the proposed approach improves detection performance with lower false alarm rates.

Global average pooling (GAP) plays an important role in traditional channel attention. However, there is the disadvantage of insufficient information to use the result of GAP as the channel scalar. At the same time, the existing spatial attention models focus on the areas of interest using average pooling or convolutional networks, but there is a loss of feature information and neglect of the structural feature. In this paper, dual wavelet attention is proposed, which can effectively alleviate the aforementioned problems and enhance the representation ability of CNNs. Firstly, the equivalence between the sum of the low-frequency subband coefficients of 2D DWT (Haar) and GAP is proved. On this basis, the statistical characteristics of low-frequency and high-frequency subbands are effectively combined to obtain the channel scalars, which can better measure the importance of each channel. In addition, 2D DWT can effectively capture the approximate and detailed structural features. Thus, wavelet spatial attention is proposed, which can effectively focus on the key spatial structural features. Different from traditional spatial attention, it can better curve the structural and spatial attention for different channels. The experiments are verified on four natural image data sets and three remote sensing scene classification data sets, which shows the effectiveness and versatility of the proposed methods. The code of this paper will be available athttps://github.com/yutinyang/DWAN.

During target tracking process, the state of the target is usually unpredictable. In theory, it is often beneficial to automatically assign suitable features to describe the specific target in each frame. Inspired by this, in this paper, we propose a novel dynamic feature-adaptive tracking framework (DFAT) which automatically assigns appropriate features to the consecutive frames during tracking process to boost the tracking performance. To implement DFAT, a large pool consisting of trackers/experts based on correlation filtering (CF) is constructed which is called candidate pool (CandPool). The diversity of the experts lies in their feature configurations and we call them candidate experts (CandExp). In this way, different features can be assigned for continuously changed scenarios and the target. Then to assign suitable experts, for each frame, we design the dynamic tracking process as the following three steps: (1) Several experts which are called executive experts (ExeExp) are selected from the CandPool according to CandExps’ past performance. (2) The ExeExps generate the tracking results and the performance of them are evaluated via a novel evaluation mechanism. (3) The selection rate of each CandExp in the CandPool is updated according to the performance evaluation and the final tracking result is selected. To better evaluate the CandExp, we propose two novel criteria: (1) content similarity weighted intra-evaluation, and (2) response confidence based self-evaluation. Compared with traditional post-event ensemble trackers that use fixed experts, the proposed method learns to dynamically assign appropriate ExeExps selected from a large CandPool which leads to adaption to different cases. Moreover, overfitting caused by fixed experts can also be mitigated via dynamic tracking. Experiments on both public available general and satellite videos based data sets demonstrate the superiority of the proposed method.

Achieving reliable intervehicle positioning is one of the most fundamental elements for many vehicular applications, including collision avoidance and autonomous driving. Vehicle position is generally provided by a global navigation satellite system (GNSS), which unfortunately suffers from inaccuracy to varying degrees in challenging environments, for example, GNSS outages. In this article, a reliable fusion technique, called non-Gaussian Redheffer weighted least squares ( n GRWLSs), is proposed. This new approach highlights the intervehicle positioning estimation in multi-GNSS outage environments, such as complete, partial, and free GNSS pseudorange outages. The proposed method combines, on the one hand, the benefits of the Gaussian dynamical matrix principle and the Redheffer distribution function for the sparse property in complete GNSS pseudorange outages and, on the other hand, the use of the optimal window size to regulate the data flow generated by both the inertial navigation systems (INSs) and GNSS during a partial GNSS pseudorange outage. During the free GNSS pseudorange outage, the process ignores data from the INS, and instead, GNSS pseudorange information alone will be considered to compute the intervehicle positioning information. Consequently, weighted least squares is used as an intervehicle positioning estimator. To address the pseudorange uncommon and INS measurement noises, the generalized error distribution (GED) is used to estimate the non-Gaussian densities. Finally, road-test experiments are implemented to evaluate the consistency of the proposed approach. The experimental results show that the proposed n GRWLS can accurately estimate the intervehicle positioning under various conditions (free, partial, and complete GNSS pseudorange outages).

Attention-based deep multiple-instance learning (MIL) has been applied to many machine-learning tasks with imprecise training labels. It is also appealing in hyperspectral target detection, which only requires the label of an area containing some targets, relaxing the effort of labeling the individual pixel in the scene. This article proposes an L1 sparsity-regularized attention multiple-instance neural network (L1-attention MINN) for hyperspectral target detection with imprecise labels that enforces the discrimination of false-positive instances from positively labeled bags. The sparsity constraint applied to the attention estimated for the positive training bags strictly complies with the definition of MIL and maintains better discriminative ability. The proposed algorithm has been evaluated on both simulated and real-field hyperspectral (subpixel) target detection tasks, where advanced performance has been achieved over the state-of-the-art comparisons, showing the effectiveness of the proposed method for target detection from imprecisely labeled hyperspectral data.

With the increasing accessibility of remote sensing videos, remote sensing tracking is gradually becoming a hot issue. However, accurately detecting and tracking in complex remote sensing scenes is still a challenge. In this article, we propose a collaborative learning tracking network for remote sensing videos, including a consistent receptive field parallel fusion module (CRFPF), dual-branch spatial-channel co-attention (DSCA) module, and geometric constraint retrack strategy (GCRT). Considering the small-size objects of remote sensing scenes are difficult for general forward networks to extract effective features, we propose a CRFPF-module to establish parallel branches with consistent receptive fields to separately extract from shallow to deep features and then fuse hierarchical features adaptively. Since the objects and their background are difficult to distinguish, the proposed DSCA-module uses the spatial-channel co-attention mechanism to collaboratively learn the relevant information, which enhances the saliency of the objects and regresses to precise bounding boxes. Considering the interference of similar objects, we designed a GCRT-strategy to judge whether there is a false detection through the estimated motion trajectory and then recover the correct object by weakening the feature response of interference. The experimental results and theoretical analysis on multiple datasets demonstrate our proposed method's feasibility and effectiveness. Code and net are available at https://github.com/Dawn5786/CoCRF-TrackNet.

Supervised salient object detection (SOD) methods achieve state-of-the-art performance by relying on human-annotated saliency maps, while unsupervised methods attempt to achieve SOD by not using any annotations. In unsupervised SOD, how to obtain saliency in a completely unsupervised manner is a huge challenge. Existing unsupervised methods usually gain saliency by introducing other handcrafted feature-based saliency methods. In general, the location information of salient objects is included in the feature maps. If the features belonging to salient objects are called salient features and the features that do not belong to salient objects, such as background, are called nonsalient features, by dividing the feature maps into salient features and nonsalient features in an unsupervised way, then the object at the location of the salient feature is the salient object. Based on the above motivation, a novel method called learning salient feature (LSF) is proposed, which achieves unsupervised SOD by LSF from the data itself. This method takes enhancing salient feature and suppressing nonsalient features as the objective. Furthermore, a salient object localization method is proposed to roughly locate objects where the salient feature is located, so as to obtain the salient activation map. Usually, the object in the salient activation map is incomplete and contains a lot of noise. To address this issue, a saliency map update strategy is introduced to gradually remove noise and strengthen boundaries. The visualization of images and their salient activation maps show that our method can effectively learn salient visual objects. Experiments show that we achieve superior unsupervised performance on a series of datasets.

Existing fusion-based local community detection algorithms have achieved good results. However, when assigning a node to a community, similarity functions are sometimes used, which only use node information, while ignoring connection information within the community. These algorithms sometimes fail to find influential nodes, which eventually leads to the failure to find a complete local community. To address these problems, a new local community detection algorithm is proposed in this article. Two strategies, of strong fusion followed by weak fusion, are used alternately to fuse nodes. Compared with using two fusion strategies alone, the alternating loop method can improve the solution of the algorithm in each stage. In strong fusion, we propose a new membership function that considers both node information and connection information in the local community. This improves the quality of the fused node while preserving the structure of the current community. In weak fusion, we propose a parameter-based similarity measure, which can detect influential nodes for a local community. We also propose a local community evaluation metric, which does not require true division to determine the optimal local community under different parameters. Experiments, compared to six state-of-the-art algorithms, show that the proposed algorithm improves accuracy and stability, and also demonstrate the effectiveness of the new local community evaluation metrics in parameter selection.

Model quantization can reduce the model size and computational latency, it has been successfully applied for many applications of mobile phones, embedded devices, and smart chips. Mixed-precision quantization models can match different bit precision according to the sensitivity of different layers to achieve great performance. However, it is difficult to quickly determine the quantization bit precision of each layer in deep neural networks under some constraints (for example, hardware resources, energy consumption, model size, and computational latency). In this article, a novel sequential single-path search (SSPS) method for mixed-precision model quantization is proposed, in which some given constraints are introduced to guide the searching process. A single-path search cell is proposed to combine a fully differentiable supernet, which can be optimized by gradient-based algorithms. Moreover, we sequentially determine the candidate precisions according to the selection certainties to exponentially reduce the search space and speed up the convergence of the searching process. Experiments show that our method can efficiently search the mixed-precision models for different architectures (for example, ResNet-20, 18, 34, 50, and MobileNet-V2) and datasets (for example, CIFAR-10, ImageNet, and COCO) under given constraints, and our experimental results verify that SSPS significantly outperforms their uniform-precision counterparts.

Understanding the private car aggregation effect is conducive to a broad range of applications, from intelligent transportation management to urban planning. However, this work is challenging, especially on weekends, due to the inefficient representations of spatiotemporal features for such aggregation effect and the considerable randomness of private car mobility on weekends. In this article, we propose a deep learning framework for a spatiotemporal attention network (STANet) with a neural algorithm logic unit (NALU), the so-called STANet-NALU, to understand the dynamic aggregation effect of private cars on weekends. Specifically: 1) we design an improved kernel density estimator (KDE) by defining a log-cosh loss function to calculate the spatial distribution of the aggregation effect with guaranteed robustness and 2) we utilize the stay time of private cars as a temporal feature to represent the nonlinear temporal correlation of the aggregation effect. Next, we propose a spatiotemporal attention module that separately captures the dynamic spatial correlation and nonlinear temporal correlation of the private car aggregation effect, and then we design a gate control unit to fuse spatiotemporal features adaptively. Further, we establish the STANet-NALU structure, which provides the model with numerical extrapolation ability to generate promising prediction results of the private car aggregation effect on weekends. We conduct extensive experiments based on real-world private car trajectories data. The results reveal that the proposed STANet-NALU outperforms the well-known existing methods in terms of various metrics, including the mean absolute error (MAE), root mean square error (RMSE), Kullback-Leibler divergence (KL), and R2.

Metaheuristic algorithms have been investigated intensively to address highly complex combinatorial optimization problems. However, most metaheuristic algorithms have been designed manually by researchers of different expertise without a consistent framework. This article proposes a general search framework (GSF) to formulate in a unified way a range of different metaheuristics. With generic algorithmic components, including selection heuristics and evolution operators, the unified GSF aims to serve as the basis of analyzing algorithmic components for automated algorithm design. With the established new GSF, two reinforcement learning (RL)-based methods, deep$Q$-network based and proximal policy optimization-based methods, have been developed to automatically design a new general population-based algorithm. The proposed RL-based methods are able to intelligently select and combine appropriate algorithmic components during different stages of the optimization process. The effectiveness and generalization of the proposed RL-based methods are validated comprehensively across different benchmark instances of the capacitated vehicle routing problem with time windows. This study contributes to making a key step toward automated algorithm design with a general framework supporting fundamental analysis by effective machine learning.

Hyperspectral image classification (HSIC) has attracted widespread attention due to its important application in environment alterations and geophysical disaster monitoring. However, surface cultivation is not static as time passes, which leads to different hyperspectral images information collected from the same area at different time periods. Therefore, researchers are currently eager to construct a HSIC model that continuously acquires new classes of data. During the continuous learning process, the model is expected to not only effective in extracting unique spatial-spectral features of the hyperspectral image, but also ensures the ability to maintain the old classes knowledge while learning new data. To achieve this purpose, we propose a method which based on geometric spatial-spectral feature integration network with class incremental learning (GS2FIN-CIL) framework in continuous learning to make the model adaptable to new classes data and not overly forgetting the old classes knowledge during the training process. We conduct extensive experiments with the proposed GS2FIN-CIL method on widely-used hyperspectral datasets including Indian Pines, PaviaU and Salinas. The experimental results show that our GS2FIN-CIL method can achieve significantly improved results compared to current state-of-the-art class incremental learning methods, allowing for efficient adaptation and utilization of spatial-spectral features in processing new classes of hyperspectral images and alleviating the problem of catastrophic forgetting of learned old classes knowledge. The GS2FIN-CIL method could be successfully applied to the challenge of adding new classes data in HSIC task.

Recent years, in order to solve the problem of lacking accurately labeled hyperspectral image data, self-supervised learning has become an effective method for hyperspectral image classification. The core idea of self-supervised learning is to define a pretext task which helps to train the model without the labels. By exploiting both the information of the labeled and unlabeled samples, self-supervised learning shows enormous potential to handle many different tasks in the field of hyperspectral image processing. Among the vast amount of self-supervised methods, contrastive learning and masked autoencoder are well known because of their impressive performance. This article proposes a Transformer based masked autoencoder using contrastive learning (TMAC), which tries to combine these two methods and improve the performance further. TMAC has two branches, the first branch has an encoder-decoders structure, it has an encoder to capture the latent image representation of the masked hyperspectral image and two decoders where the pixel decoder aims to reconstruct the hyperspectral image at pixel-level and the feature decoder is built to extract the high-level feature of the reconstructed image. The second branch consists of a momentum encoder and a standard projection head to embed the image into the feature space. Then, by combining the output of feature decoder and the embedding vectors via contrastive learning to enhance the model’s classification performance. According to the experiments, our model shows powerful feature extraction capability and gets outstanding results on hyperspectral image datasets.

The idea of spatial correlation has been used in polarimetric synthetic aperture radar (PolSAR) classification for many years. It is common that the bigger the spatial correlation, the more information it contains. Though the recent advances in deep learning for PolSAR classification have achieved remarkable progress, the number of training samples is still a problem we have to face. Aiming to solve this problem, it is valuable to explore a greater spatial correlation of terrains as a priori for classification. Considering the correlated regions of the training sample as a prior, rather than a small neighborhood, the paper proposed Wishart locally constrained expansion algorithm. Based on this, a PolSAR classification algorithm is designed. The whole proposed PolSAR classification algorithm includes 3 parts: Wishart locally constrained expansion algorithm, convolution neural network training algorithm, and Markov random field post-processing algorithm. Supervised by the cluster knowledge from Wishart classifier, Wishart locally constrained expansion algorithm expands samples iteratively from the regions related to the training samples with a region-expanding technique, where very few labeled samples turn to a larger training set. Convolution neural network algorithm is designed to train convolution neural network with 3 types of training samples for a more accurate result. Convolution neural network is trained firstly with the numerous training samples generated through Wishart locally constrained expansion algorithm, and then the samples from consistency extraction and raw samples are involved to fine-tune the convolution neural network to get the improved result. Finally, Markov random field prior is used to smooth the result. Several benchmark datasets are adopted to evaluate the effectiveness of the proposed algorithm. The experimental results show that the new semi-supervised classification algorithm outperforms the state-of-the-art semi-supervised classification algorithms and classical supervised classification algorithms.

Band selection is an effective method to deal with the difficulties in image transmission, storage, and processing caused by redundant and noisy bands in hyperspectral images (HSIs). Existing band selection methods usually need to learn a specific model for each HSI dataset, which ignores the inherent correlation and common knowledge among different band selection tasks. Meanwhile, these methods lead to a huge waste of computation. In this article, a novel zero-shot band selection method, called MR-Selection, is proposed for HSI classification. It formalizes zero-shot band selection as a metalearning problem, where advantage actor–critic algorithm-based reinforcement learning (A2C-RL) is designed to extract the metaknowledge in the band selection tasks of various seen hyperspectral datasets through a shared agent. To learn a consistent representation among different tasks, a dynamic structure-aware graph convolutional network is constructed to build a shared agent in A2C-RL. In A2C-RL, the state is tailored in a feasible way and easy to adapt to various tasks. Meanwhile, the reward is defined according to an efficient evaluation network, which can evaluate each state effectively without any fine-tuning. Furthermore, a two-stage optimization strategy is designed to coordinate optimization directions of a shared agent from different tasks effectively. Once the shared agent is optimized, it can be directly applied to unseen HSI band selection tasks without any available samples. Experimental results demonstrate the effectiveness and efficiency of the MR-Selection on the band selection of unseen HSI datasets.

In the last decade, generative adversarial network (GAN) and its variants provide a powerful training mechanism for hyperspectral image (HSI) classification. In HSIs, the distribution of samples is more complicated due to the existence of abundant spatial-spectral information and multi-scale information. The single generation pattern of GANs is prone to modal collapse for the sample generation of HSIs. Moreover, the promotion of the generator only relies on adversarial learning with the discriminator, which limits the generator’s performance. To address these problems, a multi-complementary GANs with contrastive learning (CMC-GAN) is proposed. CMC-GAN consists of two groups of GANs, where coarse-grained GAN adopts the structure in encoder-decoder form for hidden fine-scale and coarse-scale generation, and another fine-grained GAN is responsible for fine-scale generation. In fine-grained GAN, the discriminator is constructed to distinguish the fine-scale samples from different generators, which enforces the joint optimization of these two groups of GANs and makes GANs generate diverse multi-scale samples. Furthermore, a novel contrastive learning constraint is added into GANs, where a unidirectional contrastive loss guarantees the generators to extract intra-class invariant representation and a class-specific contrastive loss urges the discriminators to learn more discriminative features for classification. Finally, both discriminators are adaptively-fused to extract complementary multi-scale spatial-spectral features for classification under the guidance of diverse generated samples. The experimental results demonstrate CMC-GAN has superior classification performance, especially for small sample classification.

Recently, multi-modal remote sensing image (MRSI) classification has attracted increasing attention of researchers. However, classification of MRSI with limited labeled instances is still a challenging task. In this paper, a novel self-supervised cross-modal contrastive learning method is proposed for MRSI classification. Joint intra- and cross-modal contrastive learning are used to better mine multi-modal feature representations during pre-training, and the intra- and cross-modal contrastive learning objectives are jointly optimized, whereby it encourages the learned representation to be semantically consistent within and between modalities simultaneously. Moreover, a simple but effective hybrid cross-modal fusion module (HCFM) is designed in the fine-tuning stage, which could better compactly integrate complementary information across these modalities for more accurate classification. Extensive experiments are taken on four benchmark datasets (i.e., Houston 2013, Augsburg, Trento, and Berlin), and the results show that the proposed method outperforms state-of-the-art methods.

Multisource remote sensing images provide complementary multidimensional information for reliable and accurate classification. However, gaps in imaging mechanisms result in heterogeneity between multiple source images. During fusion, this heterogeneity causes the generated multisource representations may be redundant and ignore discriminative uni-source information, which significantly hampers the fusion classification performance. To address this challenge, we introduce a novel multisource joint representation learning method for remote sensing image fusion classification, termed Multisource Information Bottleneck Fusion Network (MIBF-Net). Based on the Information Bottleneck principle, MIBF-Net employs mutual information constraints to effectively integrate multisource information, generating a comprehensive and non-redundant multisource representation. Specifically, MIBF-Net first introduces an attribution-driven noise adaptation layer to dynamically balance the speed of feature learning across sources for extracting discriminative uni-source intrinsic information. Furthermore, a cross-source relationship encoding module is designed to fully explore cross-source complex dependencies for enhancing the richness of fused representations. Finally, we design an information bottleneck fusion module to fuse uni-source semantic information and cross-source information while reducing redundancy. In particular, we employ variational inference techniques to effectively address the mutual information optimization problem and provide theoretical derivations. Extensive experimental results on three heterogeneous multisource remote sensing data benchmarks show that the model significantly outperforms the state-of-the-art methods.

The rapid progress in remote sensing technology has made it convenient for satellites to capture both multispectral (MS) and panchromatic (PAN) images. MS has more spectral information, and PAN has higher spatial resolution. How to exploit the complementarity between MS and PAN images, and effectively combine their respective advantageous features while alleviating mode differences, has become a crucial research task. This paper designs a Style Separation and Mode Unification network (SSMU-Net) for MS and PAN image classification from a novel and effective perspective. The network can be divided into two stages: style separation and mode unification. In the style separation stage, we use wavelet decomposition and techniques similar to generative adversarial networks to preliminarily separate the information of MS and PAN into different components. These components better preserve complete information from the original data and have their own advantages in style and content. Then we propose a Symmetrical Triplet Traction module to perform style traction on different components, making style features more unique and content features more unified, achieving feature separation and purification. In the mode unification stage, we design an encoder-decoder model to reduce the impact of mode differences. The experimental results from multiple datasets validate the effectiveness of our proposed method. Our overall accuracy improved by approximately 4% on the Shanghai and Beijing datasets, and it has exceeded 99.28% on the Hohhot and Vancouver datasets. Our code is available at: https://github.com/proudpie/SSMU-Net.

The overwhelming majority of models for remote sensing image (RSI) scene classification generally require the weights pre-trained on natural images for initialization before formal training. However, differences in imaging mechanisms lead to huge discrepancies between natural images and RSIs, and the strong visual representation learned from massive natural images limits the performance of models when inferencing RSIs. To address this issue, the well-established self-supervised contrastive learning paradigm in the natural image field is introduced to the RSI field. We propose a contrastive learning method based on multi-scale hard features, MHCL, which aims to use finite RSIs to learn sufficient visual representations in an unsupervised contrastive manner, thus provide a powerful upstream pre-trained model for fine-tuning downstream scene classification task. Multi-level features extracted by intermediate layers of each encoder’s backbone are first gathered, and then a hard features transformation method is proposed to create hard positive features and diverse queues that save hard negatives, thereby enriching the finite scene information in small-scale RSIs. Furthermore, we redesign the multi-scale hard features joint contrastive loss to boost the model to explore sufficient invariant representations by additionally pulling hard positive pairs closer and pushing hard negative pairs farther away in the embedding space. Extensive experiments demonstrate that the upstream pre-training model generated by MHCL achieves competitive transferred performance on three popular scene classification datasets, outperforming the traditional model pre-trained on ImageNet and models pre-trained by other state-of-the-art contrastive learning methods. Our code will be released at: https://github.com/benesakitam/MHCL.

Localization regression in oriented object detection tasks has long faced boundary discontinuity and angular discontinuity problems induced by periodic angles. These problems were successfully resolved by using a 2d Gaussian distribution to modelling the oriented bounding box (OBB). However, the angular information of square-like objects will be lost when they are converted to 2d Gaussian distribution, forming a systematic problem. Its fundamental reason is that when the aspect ratio of the object tends to 1, the equiprobability curve of 2d Gaussian distribution degenerates from an ellipse to a circle, thus losing the orientation information of the rotated object. This results in the bounding boxes of such square-like objects not being learned effectively. To resolve this problem, we used the Lamé curve (or superellipse) to modify the existing 2d Gaussian function and designed a super-Gaussian distribution. This distribution can maintain anisotropy at arbitrary aspect ratios, thus preserving the angular information of the oriented object. We used the Kullback-Leibler (KL) divergence to measure the distance between two super-Gaussian distributions and convert it into a localization loss (SGKLD) by a function. SGKLD is an improved version of KLD loss. By modifying the form of the probability distribution, we elegantly fix the angle missing problem of the traditional Gaussian distribution. We validated the effectiveness of the proposed algorithm on several datasets and obtained the performance of SOTA. Our algorithm achieves a mean average precision (mAP) of 80.07, 76.59, 62.27, and 90.55/98.13 on the DOTA-v1.0, DOTA-v1.5, DOTA-v2.0, and HRSC2016 datasets, respectively.

In aerial image scenes, the objects have properties of arbitrary orientation, large-scale range, and dense distribution. Thus, the object detector uses oriented bounding box (OBB) to locate objects, which is more complex and challenging than horizontal bounding box (HBB) detector. Mainstream OBB detectors mostly use one-to-many label assignment strategy to predict multiple bounding boxes for the same object, and filter out repeat predictions by non-maximum suppression (NMS). NMS ranks with confidence and drops the detection box with IoU higher than the threshold, which is easy to get the local optimum result. The clustered synthesis method gets more accurate results than the original NMS, but applying it to the OBB detector leads to border shift, which arises from the angular discontinuity problem. Therefore, we use Gaussian OBB (G-OBB) to deal with the angular discontinuity and thus eliminate the offset generated by direct synthesis. G-OBB is not an easy to understand and describe representation. For this reason, we analyze the properties of G-OBB, and design a decoding method to convert a G-OBB to a rotated rectangular box, further discussing its conditions. Based on the decoding method, we propose a Gaussian synthesis algorithm (GauS), which transforms the OBB into Gaussian space, followed by synthesis, and finally transforms the synthesis result back into a new OBB. We have derived the synthesis and decoding methods, and further verified their effectiveness. The extensive experiments on several existing models show that GauS takes very little computation and improves detector’s high-precision performance. Extensive experiments verify the effectiveness, stability, and universality of the proposed algorithm. In addition, The RTMDet using GauS achieves a performance of 81.61 AP50and gains a 0.39% improvement in mAP, which achieves the SOTA performance. Our implementation is available at: https://github.com/lzh420202/GauS.

Contrastive learning with learnable examples performs outstandingly in data representation. However, when dealing with hard samples, instance-level alignment with excessive uniformity may descend into trivial clusters, especially when confronted with inter-class similarity and intra-class diversity in hyperspectral images. To solve this problem, we regard prototypical contrastive learning as tracing the potential probability density distribution. Then, a novel pre-training method, Learnable Sparse Contrastive Sampling (LSCoSa), is proposed for discriminative representation learning, containing sparse positive sampling and multiple positives learning. Specifically, on the basis of cooperative-adversarial contrastive learning, we first exert a KL divergence regularizer on the average activation probability of the prototypes, suppressing fake density prototypes for sparse positive sampling. Furthermore, we propose multiple positives learning, in which the top-k potential positives are retrieved and dynamically weighted for contrastive supervision, to avoid trivial clusters and cover satisfying semantic variations. Comprehensive experiments on three HSI benchmark datasets demonstrate that LSCoSa achieves significant advantages over other HSIC methods. The code is available at https://github.com/sakurashine/LSCoSa.

Ship detection in aerial images remains an active yet challenging task due to its arbitrary object orientation and various aspect ratios from the bird’s-eye perspective. Most existing oriented objection detection methods rely on angular prediction or predefined anchor boxes, making these methods highly sensitive to unstable angular regression and excessive hyper-parameter setting. To address these issues, we replace the angular-based object encoding with an anchor-and-angle-free paradigm, and propose a novel detector deploying a center and four midpoints for encoding each oriented object, namely MidNet. Moreover, MidNet designs a novel symmetrical deformable convolution for enhanceing the features of midpoints, then the center and midpoints for an identical ship are adaptively matched by predicting corresponding centripetal shift and matching radius. Finally, a concise analytical geometry algorithm is proposed to calculate the ship orientation and refine the keypoints step-wisely for building precise oriented bounding boxes. On two public ship detection datasets, HRSC2016 and FGSD2021, MidNet outperforms the state-of-the-art detectors by achieving APs of 90.52% and 86.50%.

Supervised remote sensing (RS) image segmentation has achieved remarkable success with large amounts of manually labeled data, which may be difficult to acquire in some practical application scenarios. Semisupervised RS image segmentation can efficiently utilize the knowledge embedded in unlabeled data to improve recognition performance, which is of great significance for the generalization application of segmentation models. In this work, we propose an end-to-end semisupervised RS image segmentation method based on weak-to-strong consistency learning, denoted as WSCL. Specifically, a common strong data augmentation technique for image segmentation is introduced to provide powerful input perturbation to decouple self-biased cognition. By forcing weakly augmented, and strongly augmented perspectives from the same sample to be consistent, WSCL not only enables the model to steadily learn knowledge contained in unlabeled data but also alleviates overfitting. In addition, a novel sparse dual-view cross-sample image generation method is presented to generate new training samples, which helps provide a more comprehensive diversity of perturbations. Furthermore, an adaptive re-weighting strategy based on the entropy maps of the outputs of strongly perturbed samples is proposed to suppress noise, guiding the training process in a positive direction. Extensive experiments demonstrate the significant advantage of WSCL over other advanced methods, achieving new state-of-the-art under several evaluation metrics on DFC22, iSAID, MER, MSL, Vaihingen, and GID-15 datasets. The source code is open-sourced at https://github.com/xiaoqiang-lu/WSCL.

Labeling data in the field of remote sensing is time-consuming and labor-intensive, making domain adaptation between different domains an urgently needed solution. To address the domain gap between diverse datasets in the remote sensing domain, numerous methods tailored for domain adaptation in high-resolution remote sensing imagery have emerged. Some of the existing methods focus on reducing the domain gap at either the feature level or the pixel level, often overlooking their underlying connection. To tackle this issue, we introduce a prototype-wise contrastive feature alignment paradigm (PCFA) aimed at bridging the representations between the feature and pixel levels. By dynamically updating, we acquire prototype information encompassed by different mini-batches and employ an optimal transport mechanism to reasonably apply the prototype feature distribution in guiding the learning of target domain features. We conduct extensive domain adaptation semantic segmentation (DASS) experiments on the ISPRS Vaihingen and Potsdam datasets, achieving an improvement about 4%~5% in mIoU (mean Intersection over Union) compared to previous methods using the DeepLabV2 framework.

In the annotation of remote sensing images (RSIs), the effectiveness of common object detection methods trained on only a few samples decreases instantly, which has prompted increasing research on the few-shot problem in remote sensing. RSIs often exhibit suboptimal performance in few-shot scenarios due to the intricate nature of scene information interference and the high degree of cosine similarity, both of which present significant challenges to their effectiveness. In this paper, a two-stage detection framework based on fine-tuning is selected to deal with the common problems in the few-shot task of remote sensing domain. Considering the excessive scale variation of instances in remote sensing datasets, we introduce an automatically learned aug-aware search module to provide an intelligent data augmentation solution for Faster R-CNN using different optimal augmentation policies searched by the network to fit the current dataset. We introduce a contrastive RoI branch to better classify novel class proposal features that are easily confused by the base class. We named our work AACE and conducted extensive experiments on two common object detection datasets in remote sensing, NWPU VHR 10 and DIOR, on which AACE achieved about 2.30% and 2.61% improvement, respectively, in the number of shots listed in the paper, compared to other algorithms.

As a content management technique, remote sensing (RS) scene classification (RSSC) always attracts researchers’ attention. In the past decades, many successful methods have been proposed. Nevertheless, their prerequisite is that there are large labeled data sets, which is a strict demand in practice. To resolve this contradiction, developing RSSC models with the help of few-shot learning (FSL) has become popular. Due to lacking prior knowledge, most of the existing few-shot RSSC models pay attention to the learning algorithm. However, they do not attach importance to the complex contents within RS scenes and the intricate inter-/intra-class relations between RS scenes. This would influence their performance negatively. In this paper, we propose a new few-shot RSSC model named multi-pretext-task prototypes guided dynamic contrastive learning network (MPCL-Net). MPCL-Net consists of a multi-pretext tasks generation sub-module, a deep feature learning sub-module, and a joint optimization sub-module. First, two RS-oriented pretext tasks are constructed under the self-supervised learning (SSL) framework in the multi-pretext tasks generation sub-module, which aim to explore multi-scale and rotation-invariant information from RS scenes. Second, a simple convolutional neural network (CNN) is developed in the deep feature learning sub-module to transform the RS scenes into visual features. Third, three loss functions are formulated and integrated in the joint optimization sub-module. Their goals are to fully capture the diverse land covers within RS scenes and compact/separate the intra-/inter-class samples with limited supervision. Finally, our MPCL-Net can be trained in a meta way. The positive results counted on the three public RS scene data sets confirm that our MPCL-Net is helpful to RSSC tasks under the few-shot scenario. Our source codes are available at https://github.com/TangXu-Group/Remote-Sensing-Images-Classification/tree/main/MPCL.

Considering that deep learning achieves the prominent performance, it has been applied to synthetic aperture radar (SAR) image registration to improve the registration accuracy. In most methods, a deep registration model is constructed to classify matched points and unmatched points, in which SAR image registration is regarded as a supervised two-classification problem. However, it is difficult to annotate massive matched points manually in practice, which limits the performance of deep networks. Besides, inevitable differences among SAR images easily cause that some training and testing samples are inconsistent, which probably brings negative effects for training a robust registration model. To address these problems, we propose an adaptive self-supervised SAR image registration method, where SAR image registration is regarded as a self-supervised task rather than the supervised two-classification task. Inspired by self-supervised learning, we consider each point on SAR images as a category-independent instance, which mitigates the requirement of manual annotations. Based on key points from images, a self-supervised model is constructed to explore the latent feature of each key point, and then, pairs of match points are sought via evaluating similarities among key points and used to calculate the alignment transformation matrix. Meanwhile, to enhance the consistency of samples, we design a new strategy that constructs multiscale samples by transforming key points from one image into another, which avoids inevitable diversities between two images effectively. In particular, the constructed samples feeding to the self-supervised model are adaptively updated with the modification of the transformation matrix in iterations. Moreover, the similarity of maximal public areas (MPAS) indicator is proposed to assist in estimating the transformation. Finally, experimental results illustrate that the proposed method achieves more accurate registrations than other compared methods.

Multi-modal remote sensing images with large rotation transformation (RT) are challenging to be registered. It needs to deal with the global geometric deformation caused by great RT and significant local appearance differences caused by different imaging mechanisms. Existing deep learning methods mainly use a single deep descriptor learning (DDL) network to extract invariant features for identifying matching samples and discriminative feature descriptors for separating non-matching samples. However, it is difficult to extract local invariant feature descriptors to RT and modality change through a single DDL network. This paper proposes a novel coarse-to-fine deep learning image registration framework for multi-modal remote sensing images based on two task-specific deep models. Specifically, in the coarse registration stage, this paper designs an effective deep ordinal regression (DOR) network for rotation correction, which can reduce the difficulty of multi-modal image registration and boost image registration. The proposed DOR network transforms the rotation correction task into a rotation ordinal regression problem, which can exploit the potential relationship between the rotation ordinals to improve the accuracy of rotation estimation. In the fine registration stage, we adopt the DDL network to deal with the image modality change based on the rotation-corrected images. Extensive experimental results on multi-modal image datasets demonstrate the significant advantages of the proposed coarse-to-fine deep learning registration framework. The DOR network achieves higher rotation correction accuracy, which can significantly improve the multi-modal image registration performances.

During incremental learning tasks, catastrophic forgetting occurs when old models are updated with new information. To address this issue, we propose a novel method called label strategy and adaptive weights (LSAW) that improves the incremental learning process. The label strategy introduces the old classes and solves the problem of how to reasonably use the wrong samples predicted by the old model. In the cross-entropy (CE) loss, we apply a threshold to filter the pseudolabels predicted by the old model. Subsequently, we merge the pixel samples with high probability with the current label. The probability here refers to the probability that the pixel belongs to the true class. This process enables the introduction of information from old classes that are not directly accessible in the current stage. Moreover, this information is relatively reliable, and the model exhibits confidence in its accuracy. For the remaining pixels, we retain all classes’ information through label smoothing. In the distillation function, the old class and background pixel samples are selected for distillation according to the prediction map of the old classes. The weights of the classes are adaptively updated and adjusted using specific label information from each batch and the different stages of incremental learning. As demonstrated by the results of our experiment, on three remote sensing image datasets: China Computer Federation (CCF), Potsdam, and Vaihingen, our method achieves the best results.

The complementary and heterogeneous properties fusion of multimodal data (such as hyperspectral, lidar, and synthetic aperture radar data) can significantly improve the accuracy of remote sensing (RS) images joint classification. Thus, we propose a spatial-spectral bilinear representation fusion network (S2BRFNet), which captures long-range dependencies cross-modality and within the same modality to achieve the final joint classification. Firstly, a cross-modal spatial-spectral representation module (S2RM) is designed, it utilizes spatial-spectral attention and self-attention between heterogeneous data to enhance the characterization capabilities of cross-modal complementary properties and spatial-spectral features of single-source data. Secondly, a semantic space-guided bilinear feature fusion module (S2BFM) is developed, which uses deep and shallow features to regain fine-grained features. It uses shallow location details to improve the semantic prediction of deep features. Furthermore, it uses the different representation capabilities of different layers for objects with obvious feature differences to enhance the feature advantages. Therefore, rich global context information is obtained. Finally, the semantic space re-weight strategy is used to guide the outer product fusion of heterogeneous features, which enhances the ability of the network to identify similar features. Classification experiments are carried out on four common datasets of different modality combinations (HS-SAR-DSM Augsburg, Berlin, Trento, and Muufl), and this can prove the superiority of the S2BRFNet.

Deep Learning-based (DL) methods have dominated the task of semantic segmentation of remote sensing images. However, the sizes of different objects vary widely, and there is a great deal of label-noise due to the inevitable shadows. Therefore, there is an urgent need for a method that can precisely handle complex ground data. In this paper, we propose an Inter-Class Enhanced Network (ICEN) for representing features of varying sizes. It comprises two branches: Sparse Representation Network (SPN) and Feature Extraction Network (FEN). Then, a Class-Perception Block is inserted between the two branches to instruct the SPN’s low-level semantic features to be merged into the deeper network. Such a block can reduce label-noise in remote sensing image segmentation. In addition, the proposed EIRI provides a more precise classification process for target edges containing many misclassified points without requiring excessive computational overhead. The experimental results of our proposed Class-Perception Network (C-PNet) achieve competitive performance on the Vaihingen, Potsdam, LoveDA, and UAVid datasets.

Cross-modal remote sensing image-text retrieval (CMRSITR) is a challenging topic in the remote sensing (RS) community. It has gained growing attention because it can be flexibly used in many practical applications. In the current deep era, with the help of deep convolutional neural networks (DCNNs), many successful CMRSITR methods have been proposed. Most of them first learn valuable features from RS images and texts respectively. Then, the obtained visual and textual features are mapped into a common space for the final retrieval. The above operations are feasible, however, two difficulties are still to be solved. One is that the semantics within the visual and textual features are misaligned due to the independent learning manner. The other one is that the deep links between RS images and texts cannot be fully explored by simple common space mapping. To overcome the above challenges, we propose a new model named interacting-enhancing feature transformer (IEFT) for CMRSITR, which regards the RS images and texts as a whole. First, a simple feature embedding module (FEM) is developed to map images and texts into the visual and textual feature spaces. Second, an information interacting-enhancing module (IIEM) is designed to simultaneously model the inner relationships between RS images and texts and enhance the visual features. IIEM consists of three feature interacting-enhancing (FIE) blocks, each of which contains an inter-modality relationship interacting (IMRI) sub-block and a visual feature enhancing (VFE) sub-block. The duty of IMRI is to exploit the hidden relations between cross-modal data, while the responsibility of VFE is to improve the visual features. By combining them, semantic bias can be mitigated, and the complex contents of RS images can be studied. Finally, the retrieval module (RM) is constructed to generate the matching scores for deciding the search results. Extensive experiments are conducted on four public RS data sets. The positive results demonstrate that our IEFT can achieve superior retrieval performance compared with many existing methods. Our source codes are available at https://github.com/TangXu-Group/Cross-modal-remote-sensing-image-and-text-retrieval-models/tree/main/IEFT.

Change detection (CD) is a hot research topic in the remote sensing (RS) community. With the increasing availability of high-resolution (HR) RS images, there is a growing demand for CD models with high detection accuracy and generalization ability. In other words, the CD models are expected to work well for various HRRS images. Convolutional neural networks (CNNs) have been dominated in HRRS image CD due to their excellent information extraction and nonlinear fitting capabilities. However, they are not skilled in modeling long-range contexts hidden in HRRS images, which limits their performance in CD tasks more or less. Recently, the Transformer, which is good at extracting global context dependencies, has become popular in the RS community. Nevertheless, detailed local knowledge receives insufficient emphasis in common Transformers. Considering the above discussion, we combine CNN and Transformer and propose a new W-shaped dual Siamese branch hierarchical network for HRRS image CD named WNet. WNet first incorporates a Siamese CNN and a Siamese Transformer into a dual-branch encoder to extract multi-level local fine-grained features and global long-range contextual dependencies. Also, we introduce deformable ideas into the Siamese CNN and Transformer to make WNet understand the critical and irregular areas within HRRS images. Second, the difference enhancement module (DEM) is developed and embedded into the encoder to produce the difference feature maps at different levels. Using simple pixel-wise subtraction and channel-wise concatenation, the changes of interest and irrelevant changes can be highlighted and suppressed in a learnable manner. Next, the multi-level difference feature maps are fused stage by stage by CNN-Transformer fusion modules (CTFMs), which are the basic units of the decoder in WNet. In CTFM, the local, global, and cross-scale clues are taken into account to ensure the integrity of information. Finally, a simple classifier is constructed and added at the top of the decoder to predict the change maps. Positive experimental results counted on four public datasets demonstrate that the proposed WNet is helpful in HRRS image CD tasks. Our source codes are available at https://github.com/TangXu-Group/Remote-Sensing-Image-Change-Detection/tree/main/WNet.

Most current change detection methods require a large amount of labeled data to train huge parameters. To break this limitation, this paper proposes a novel semi-supervised learning framework for remote sensing change detection, named a semi-dual change detection network (SDCDNet). The SDCDNet consists of a dual shared network and dual branching networks. The dual shared network is designed to exploit the full potential of the data, and the dual branching network is proposed to differentiate the kinds of annotated data and eliminate the disturbance between different types of data. In addition, the adaptive weighting module (AWM) enhances the features of weak branching, and the mask constraint module (MCM) is proposed to increase the ability of the network to extract foreground features. To solve the complex problem of data labeling, a patch-based weak label construction method is proposed to build super-weak labels. Experiments show that the proposed SDCDNet achieves excellent results on two remote sensing image change detection datasets.

Oriented object detection is a challenging task in remote sensing, where the detected objects can be represented by oriented bounding boxes (OBBs). Angle prediction in oriented object detection has been widely studied, due to its crucial role in object detection. However, the precision of angle prediction is severely limited by misalignments in most of the existing methods, including representation-, evaluation-, and optimization-based misalignments. To alleviate these misalignments, this paper presents a novel angle prediction method, called Angle Quality Estimation (AQE). Specifically, our proposed AQE transforms the angle prediction task into a distribution estimation task to address the representation misalignment problem and implicitly measure the quality of the predicted angles. Based on the estimated angle quality, we then propose a new metric to comprehensively evaluate the quality of OBBs. Then we propose an object aspect ratio based loss function to optimize angle prediction for addressing the optimization misalignment. Our proposed AQE is a plug-and-play method, which can be embedded on any existing oriented object detector. Experimental results on three public benchmarks, including DOTA, HRSC2016, and ICDAR2015 datasets, show that our method achieves better performance than the other state-of-the-art.

Most popular visual trackers for natural scenarios always adopt handcraft features or deep features to track the target in a video. However, they face with difficulties in discriminative feature representation and usually suffer from severe model drift for satellite videos, especially when encountering challenges of dim and small targets, low contrast or similar target interference. To overcome these difficulties, we propose a Contourlet-based Siamese Learning Tracker (CSLT), which mainly aims at tracking dim and small objects in satellite videos. In contrast to conventional methods, the contourlet transform enriches directional multi-resolution information which is crucial to discriminative feature representation for dim and small targets in satellite video frames that lack distinguishable appearance features. We jointly use multi-resolution features with deep features by spatial-attention fusion strategy and then track the targets by a Siamese structure network. To further improve the accuracy and robustness, a model drift alarm and calibration module, including translation drifting penalty and rotation drifting penalty, is employed during tracking. We conduct extensive comparisons with 16 popular state-of-the-art trackers on three satellite video datasets. The experimental results validate the effectiveness of the proposed tracker.

To minimize the feature redundancy of multispectral (MS) and panchromatic (PAN) images and maximize the complementary advantages of PAN and MS, a Dual-Stream Transformer with Diff-attention (DSTD)-Net is proposed for PAN and MS classification in this paper. Firstly, in terms of feature extraction, we use Self-attention and Co-attention (SCA) block to extract both specific advantageous features and common essential features. Based on that, a self-attention module strengthened by diff-attention (SSDA) that pays attention to the difference between two specific advantageous features is designed to reduce the essential redundancy in specific features. It can take advantage of the difference between two specific features and reduce the essential redundancy of the specific advantageous features, making them purer and better for classification. Finally, since the specific features and common features of multispectral (MS) and panchromatic (PAN) images make different contributions to classification, a Multi-stage Gated Fusion (MGF) strategy is used. The MGF strategy mainly uses Gated multisource units (GMU) to adapt the weight of different features and fuse them. So, our MGF strategy can strengthen the specific advantageous features beneficial for classification. Above all, the several experiment results verify our proposed networks’ effectiveness and robustness. Our code is available at: https://github.com/blackkiring/DSTD.

Contrastive self-supervised learning (CSSL) is a promising method in extracting effective features from unlabeled data. It performs well in image-level tasks, such as image classification and retrieval. However, the existing CSSL methods are not suitable for pixel-level tasks, e.g., change detection (CD), since they ignore the correlation between local patches or pixels. In this paper, we firstly propose a multi-cue contrastive self-supervised learning (MC-CSSL) method to derive dense features for change detection. Besides data augmentation, the MC-CSSL takes advantage of more cues based on the semantic meaning and temporal correlation of local patches. Specially, the positive pair is built from local patches with the similar semantic meaning or temporal ones with the same geographic location. The assumption is that local patches belonging to the same kind of land-covering tend to share similar features. Secondly, the affinity matrix is truncated and introduced to extract change information between two temporal patches obtained from different types of sensors. As a result, some initial unchanged pixels are selected to serve as the supervision for mapping the dense features into a consistent space. Based on the distance between all bi-temporal pixels in the consistent space, a difference image (DI) is generated and more unchanged pixels can be available. The dense feature mapping and unchanged pixel updating proceed alternately. The proposed CD method is evaluated in both homogeneous and heterogeneous cases and the experimental results demonstrate its effectiveness and priority after comparison with some existing state-of-the-art methods. The source code will be available at https://github.com/Yang202308/ChangeDetection_CSSL.

Deep convolutional neural networks (CNNs) are significant in remote sensing. Due to the strong local representation learning ability, CNNs have excellent performance in remote sensing scene classification. However, CNNs focus on location-sensitive representations in the spatial domain and lack contextual information mining capabilities. Meanwhile, remote sensing scene classification still faces challenges, such as complex scenes and significant differences in target sizes. To address the problems and challenges above, more robust feature representation learning networks are necessary. In this paper, a novel and explainable spatial-frequency multi-scale Transformer framework, SF-MSFormer, is proposed for remote sensing scene classification. It mainly comprises spatial-domain and frequency-domain multi-scale Transformer branches, which consider the spatial-frequency global multi-scale representation features. Besides, the texture-enhanced encoder is designed in the frequency-domain multi-scale Transformer branch, which is adaptive to capture the global texture features. In addition, an adaptive feature aggregation module is designed to integrate the spatial-frequency multi-scale feature for final recognition. The experimental results verify the effectiveness of SF-MSFormer and show better convergence. It achieves state-of-the-art results (98.72%, 98.6%, 99.72%, and 94.83% overall accuracies, respectively) on the AID, UCM, WHU-RS19, and NWPU-RESISC45 datasets. Besides, the feature visualizations evaluate the explainability of the texture-enhanced encoder. The code implementation of this article will be available at https://github.com/yutinyang/SF-MSFormer.

For remote sensing image segmentation, the boundaries of objects are difficult to distinguish, which is ignored by most methods. Therefore, it is challenging how to excavate and recover the boundaries of objects accurately. In this article, we propose a boundary-aware multi-scale network (BMNet) to solve this problem. The key components of BMNet include the scale attention module (SA-module) and boundary guidance module (BG-module). Specifically, SA-module is proposed to guide the refinement of multi-scale features in a context-aware way. It enhances the discriminability of multi-scale features by establishing contextual dependencies, which enables the refinement of the prediction of objects. Then, BG-module is proposed to enable networks to distinguish the boundary of objects. It utilizes manifold information of features to generate boundary guidance maps and forces the network to focus more on the boundary of objects. The effectiveness of the proposed BMNet is demonstrated on two public remote sensing datasets: ISPRS 2-D semantic labeling Potsdam dataset and Vaihingen dataset, where BMNet achieves better segmentation than prevalent methods. Finally, the experimental results indicate that BMNet can produce sharper boundaries of objects to reconstruct more detailed segmentation results.

Few-shot remote sensing scene classification aims to identify unseen classes using only a small number of labeled samples. Considering the large intra-class variances and inter-class similarity of remote sensing scenes, most existing methods focus on feature extraction, ignoring the overfitting problem caused by insufficient samples. To this end, we propose a novel few-shot learning framework, called multibranch feature generation networks (MFGNets), which solves the few-shot scene classification from the source by online sample generation at the representation space. Specifically, we first build a feature generation net to transform the few-shot classification into a regular classification problem, in which the generated samples are achieved by combining the class-specific features with the sampled intra-class features. Then, to ensure the quality of the generated samples, we introduce two novel regularization terms: the intra-class diversity loss (ID-Loss) and the inter-class consistency loss (IC-Loss), which aid the model in generating more diverse samples. Furthermore, we introduce a scale-angle aware self-supervised pretext to learn scale-invariant and rotation-invariant features, improving the model’s feature representation capability in remote sensing scenes. We evaluate the proposed method on three publicly available datasets, namely UC_Merced, NWPU-RESISC45, and AID. Our approach has achieved state-of-the-art performance, with an improvement of more than 3.31%, 2.64%, and 6.86% on the most challenging 1-shot tasks, respectively.

For visual question answering on remote sensing (RSVQA), current methods scarcely consider geospatial objects typically with large-scale differences and positional sensitive properties. Besides, modeling and reasoning the relationships between entities have rarely been explored, which leads to one-sided and inaccurate answer predictions. In this article, a novel method called spatial hierarchical reasoning network (SHRNet) is proposed, which endows a remote sensing (RS) visual question answering (VQA) system with enhanced visual–spatial reasoning capability. Specifically, a hash-based spatial multiscale visual representation module is first designed to encode multiscale visual features embedded with spatial positional information. Then, spatial hierarchical reasoning is conducted to learn the high-order inner group object relations across multiple scales under the guidance of linguistic cues. Finally, a visual-question (VQ) interaction module is employed to learn an effective image–text joint embedding for the final answer predicting. Experimental results on three public RS VQA datasets confirm the effectiveness and superiority of our model SHRNet.

Convolutional neural networks (CNNs) have superior feature learning capabilities with large numbers of labeled samples. The reality is that labeling these samples is costly in terms of human labor. Existing data augmentation methods alleviate the scarcity of labeled samples. However, these methods are not suitable for synthetic aperture radar (SAR) images, owing to special imaging mechanisms and observational objects. The generative SAR images by existing augmented methods show structure distortion. To address this issue, we introduce a curvelet adversarial augmented neural network (CA2NN) for SAR image classification. Specifically, an$\text{A}^{2}$NN is established, which consists of two generative streams and one discriminative stream. In the generative stream, through the mutual transformation between the whole and partial images, more new samples with structural consistency are generated to augment the limited labeled data. In the discriminative stream, these generated samples show certain appearance variations after adversarial training based on the novel joint discriminant criterion. Simultaneously, given the multiscale and multidirectional nature of SAR images, we construct discretized curvelet in 2-D space, aiming to extract the singularity features and avoid overfitting. By integrating curvelet kernels into$\text{A}^{2}$NN, CA2NN can automatically generate more representative features adapting to complex terrain, while greatly reducing the complexity of the network. Experiments are conducted on the SAR images with large-scale and complex scenes, suggesting that the proposed approach significantly improves the classification performance with few labeled samples.

Hyperspectral image change detection (HSI-CD) aims to detect subtle changes on the Earth’s surface through approximately continuous spectral information, which has gradually become a very important research hotspot in the field of remote sensing (RS). In recent years, convolutional neural networks (CNNs) based HSI-CD methods have shown strong feature extraction capabilities. However, due to the simple fusion of spectral information in the channel dimension by CNN, the medium and long-term sequence properties of spectral features cannot be well mined and represented. Most previous studies mainly extract semantic features from images at different times, ignoring the temporal correlation between features, which cannot fully extract and effectively utilize temporal-spatial-spectral features. To this end, this paper proposes a cascade spectral aware transformer (CAST) for HSI-CD. First, we propose a temporal-spatial transformer (TS-Former) to enhance the temporal correlation and spatial global relationship of extracted features, thereby addressing the insufficient consideration of temporal correlation. Second, a spectral awareness transformer (SA-Former) is designed to better mine and represent the sequence properties of spectral features, especially the medium and long-term dependencies. Finally, we observe a spectral distortion in the process of extracting temporal-spatial features and based on this present a spectral constraint module (SCM) to preserve the sequence properties of spectral features and reduce the distortion of the spectrum. Extensive experiments on three challenging hyperspectral datasets demonstrate that our method achieves state-of-the-art results. The code is available at: https://github.com/tianshunli/CAST.

Recently, Hyperspectral Image Classification (HIC) with noisy labels is attracting increasing interest. However, existing methods usually neglect to explore feature-dependent knowledge to reduce label noise, and thus perform poorly when the noise ratio is high or the clean samples are limited. In this paper, a novel Triple Contrastive Representation Learning (TCRL) framework is proposed from a deep clustering perspective for robust HIC with noisy labels. The TCRL explores the cluster-level, instance-level, and structure-level representation of HIC by defining triple learning loss. First, the strong and weak transformation are defined for hyperspectral data augmentation. Then, a simple yet effective lightweight Spectral Prior Attention-based Network (SPAN) is presented for spatial-spectral feature extraction of all augmented samples. Additionally, cluster-level and instance-level contrastive learning are performed on two projection subspaces for clustering and distinguishing samples respectively. Meanwhile, structure-level representation learning is employed to maximize the consistency of data after different projections. Taking the feature-dependent information learned by triple representation learning, our proposed end-to-end TCRL can effectively alleviate the overfitting of classifier to noisy labels. Extensive experiments have been taken on three public datasets with various noise ratios and two types of noise. The results show that the proposed TCRL could provide more robust classification results when training on noisy datasets compared with state-of-the-art methods, especially when clean samples are limited. The code will be available at https://github.com/Zhangxy1999.

Multiple object tracking (MOT) in satellite videos requires to detect all objects belonging to specified categories and identify each object, which plays a basic and necessary role in automatic driving, traffic surveillance, and smart city. The traditional MOT methods in satellite videos mostly follow the detection–association framework. However, the detection–association framework works under a strict assumption that all objects are correctly localized by the detector. In practice, MOT in satellite videos faces challenges such as low resolution, tiny objects, and the wide field of view, which leads to the degradation of detector performance. In order to reduce the impact of detector degradation, we propose a bidirectional MOT framework based on trajectory criteria (BMTC) in satellite videos. In BMTC, the single object tracking (SOT) tracker carries out locating the objects between consecutive frames and the detector is just used for finding new objects. Therefore, it is less dependent on the detector performance. According to the characteristics of satellite videos, the trajectory criteria are designed to control the state of the tracker, which includes trajectory density, the limit of consecutive virtual motion predictions, and trajectory similarity measurement. Invalid fragment trajectory backtracking is implemented to alleviate the misalignment caused by the above subsection trajectory criteria. The method is validated on the VISO benchmark and SkySat-1 dataset. The experimental results show the improvement of completeness and accuracy, and the proposed tracker achieves the state-of-the-art performance.

Cross-domain networks can solve the problem of insufficient labeled samples, especially for hyperspectral images (HSIs) where obtaining labeled samples is time-consuming and laborious. Most of the current methods rely on the spatial information to achieve domain alignment, without considering the rich spectral information of HSIs. Furthermore, the methods based on convolutional neural network (CNN) cannot get the spatial information of irregular image regions, resulting in poor classification results of object edges. Therefore, we design a spectral-spatial distribution consistent network (SSDC) based on meta-learning. Firstly, to improve the feature extraction ability of the cross-domain classification model, we introduce a feature pre-extraction module, which uses the spectral attention mechanism and the alternating meta-learning method to obtain the general features of the source domain and the discriminative features of the target domain, so as to obtain the spectral weight matrix for subsequent processing. Secondly, we propose a spectral consistent module based on singular value decomposition, which increases the difference between different classes of features by penalizing the singular values of the feature matrix to achieve data distribution alignment in the spectral dimension. Finally, aiming at the low classification accuracy of irregular image regions, we propose a spatial consistent module to obtain non-local spatial topological information through stacked cross modules and graph sample and aggregate networks, which can reduce domain shift. The experiments of SSDC on four classical HSI datasets show that the proposed method can obtain competitive results with other methods based on CNN and cross-domain.

Deep representation learning has improved automatic remote change detection (RSCD) in recent years. Existing methods emphasize primarily convolutional neural networks (CNNs) or Transformer-based networks. However, most of them neither effectively combine CNNs and Transformer nor use prior geometric information to refine regions. In this paper, a novel geometric representation Transformer (GeoFormer) is proposed for high-resolution RSCD. GeoFormer utilizes convolutional information to guide the Transformer by employing geometric prior knowledge. Specifically, the proposed GeoFormer consists of three carefully designed components: the geometric-based Swin Transformer (Geo-Swin Transformer) encoder, the Laplace attention fusion (LAFusion) module, and the UNet++CD decoder. Firstly, Geo-Swin Transformer is a novel designed non-local Siamese encoder that combines geometric convolution with Transformer to provide local geometric representation information for remote contextual features. Then, a LAFusion module is proposed to achieve robust bi-temporal feature fusion, which is founded on attention mechanism and edge information. Finally, UNet++CD decodes fine-grained information from the fused features by dense multiscale upsampling process. Experimental results demonstrate that the proposed GeoFormer performs better than benchmark methods on four change detection datasets (LEVIR-CD, WHU-CD, DSIFN-CD, and CDD) and is able to detect the edges of change regions more precisely. Our code is available at https://github.com/Jiaxzhao/GeoFormer.

Building footprint extraction plays an important role in the analysis of remote sensing images and has an extensive range of applications. Obtaining precise boundaries of buildings remains a challenge in existing building extraction methods. Some previous works have made notable efforts to address this concern. However, most of these methods require cumbersome and expensive post-processing steps. Moreover, they ignored the correlation between building semantics and contours, which we believe is crucial for building footprint extraction. To mitigate this issue, our paper presents an intuitive and effective framework that explores semantic and contour cues of buildings and fully excavates their correlation. Specifically, we construct an interactive dual-stream decoder. The Intermediate connections within this decoder interactively transmit features between branches, contributing to learning correlations between semantics and contours. We propose the Semantic Collaboration Module (SCM) to strengthen the connection between the two branches. To further boost performance, we build the Multi-Scale Semantic Context Fusion Module (MSCF) to fuse semantic information from the higher and lower layers of the network, allowing the network to obtain superior feature representations. The experimental results on the WHU, INRIA, and Massachusetts building datasets demonstrate the superior performance of our method.

In satellite videos, moving vehicles are extremely small-sized and densely clustered in vast scenes. Anchor-free detectors offer great potential by predicting the keypoints and boundaries of objects directly. However, for dense small-sized vehicles, most anchor-free detectors miss the dense objects without considering the density distribution. Furthermore, weak appearance features and massive interference in the satellite videos limit the application of anchor-free detectors. To address these problems, a novel semantic-embedded density adaptive network (SDANet) is proposed. In SDANet, the cluster-proposals, including a variable number of objects, and centers are generated parallelly through pixel-wise prediction. Then, a novel density matching algorithm is designed to obtain each object via partitioning the cluster-proposals and matching the corresponding centers hierarchically and recursively. Meanwhile, the isolated cluster-proposals and centers are suppressed. In SDANet, the road is segmented in vast scenes and its semantic features are embedded into the network by weakly supervised learning, which guides the detector to emphasize the regions of interest. By this way, SDANet reduces the false detection caused by massive interference. To alleviate the lack of appearance information on small-sized vehicles, a customized bi-directional conv-RNN module extracts the temporal information from consecutive input frames by aligning the disturbed background. The experimental results on Jilin-1 and SkySat satellite videos demonstrate the effectiveness of SDANet, especially for dense objects.

The scene classification of remote sensing (RS) images plays an essential role in the RS community, aiming to assign the semantics to different RS scenes. With the increase of spatial resolution of RS images, high-resolution RS (HRRS) image scene classification becomes a challenging task because the contents within HRRS images are diverse in type, various in scale, and massive in volume. Recently, deep convolution neural networks (DCNNs) provide the promising results of the HRRS scene classification. Most of them regard HRRS scene classification tasks as single-label problems. In this way, the semantics represented by the manual annotation decide the final classification results directly. Although it is feasible, the various semantics hidden in HRRS images are ignored, thus resulting in inaccurate decision. To overcome this limitation, we propose a semantic-aware graph network (SAGN) for HRRS images. SAGN consists of a dense feature pyramid network (DFPN), an adaptive semantic analysis module (ASAM), a dynamic graph feature update module, and a scene decision module (SDM). Their function is to extract the multi-scale information, mine the various semantics, exploit the unstructured relations between diverse semantics, and make the decision for HRRS scenes, respectively. Instead of transforming single-label problems into multi-label issues, our SAGN elaborates the proper methods to make full use of diverse semantics hidden in HRRS images to accomplish scene classification tasks. The extensive experiments are conducted on three popular HRRS scene data sets. Experimental results show the effectiveness of the proposed SAGN. Our source codes are available at https://github.com/TangXu-Group/SAGN.

The convolutional neural network has achieved remarkable results in most medical image seg- mentation applications. However, the intrinsic locality of convolution operation has limitations in modeling the long-range dependency. Although the Transformer designed for sequence-to-sequence global prediction was born to solve this problem, it may lead to limited positioning capability due to insufficient low-level detail features. Moreover, low-level features have rich fine-grained information, which greatly impacts edge segmentation decisions of different organs. However, a simple CNN module is difficult to capture the edge information in fine-grained features, and the computational power and memory consumed in processing high-resolution 3D features are costly. This paper proposes an encoder-decoder network that effectively combines edge perception and Transformer structure to segment medical images accurately, called EPT-Net. Under this framework, this paper proposes a Dual Position Transformer to enhance the 3D spatial positioning ability effectively. In addition, as low-level features contain detailed information, we conduct an Edge Weight Guidance module to extract edge information by minimizing the edge information function without adding network parameters. Furthermore, we verified the effectiveness of the proposed method on three datasets, including SegTHOR 2019, Multi-Atlas Labeling Beyond the Cranial Vault and the re-labeled KiTS19 dataset called KiTS19-M by us. The experimental results show that EPT-Net has significantly improved compared with the state-of-the-art medical image segmentation method.

Fusion and interaction of multimodal features are essential for video question answering. Structural information composed of the relationships between different objects in videos is very complex, which restricts understanding and reasoning. In this paper, we propose a quaternion hypergraph network (QHGN) for multimodal video question answering, to simultaneously involve multimodal features and structural information. Since quaternion operations are suitable for multimodal interactions, four components of the quaternion vectors are applied to represent the multimodal features. Furthermore, we construct a hypergraph based on the visual objects detected in the video. Most importantly, the quaternion hypergraph convolution operator is theoretically derived to realize multimodal and relational reasoning. Question and candidate answers are embedded in quaternion space, and a Q&A reasoning module is creatively designed for selecting the answer accurately. Moreover, the unified framework can be extended to other video-text tasks with different quaternion decoders. Experimental evaluations on the TVQA dataset and DramaQA dataset show that our method achieves state-of-the-art performance.

Data-free knowledge distillation further broadens the applications of the distillation model. Nevertheless, the problem of providing diverse data with rich expression patterns needs to be further explored. In this paper, a novel dynastic data-free knowledge distillation ($D^{3}K$) model is proposed to alleviate this problem. In this model, a dynastic supernet generator (D-SG) with a flexible network structure is proposed to generate diverse data. The D-SG can adaptively alter architectural configurations and activate different subnet generators in different sequential iteration spaces. The variable network structure increases the complexity and capacity of the generator, and strengthens its ability to generate diversified data. In addition, a novel additive constraint based on the differentiable dhash (D-Dhash) is designed to guide the structure parameter selection of the D-SG. This constraint forces the D-SG to constantly jump out of the fixed generation mode and generate diverse data in semantics and instance. The effectiveness of the proposed model is verified on the experimental benchmark datasets (MNIST, CIFAR-10, CIFAR-100, and SVHN).

Multimodal Image fusion is becoming urgent in multi-sensor information utilization. However, existing end-to-end image fusion frameworks ignore a priori knowledge integration and long-distance dependencies across domains, which brings challenges to the network convergence and global image perception in complex scenes. In this paper, a conditional generative adversarial network with transformer (TCGAN) is proposed for multimodal image fusion. The generator is to generate a fused image with the source images content. The discriminators are adopted to distinguish the differences between the fused image and the source images. Adversarial training makes the final fused image to maintain the structural and textural details in the cross-modal images simultaneously. In particular, a wavelet fusion module makes the inputs contain image content from different domains as much as possible. The extracted convolutional features interact in the multiscale cross-modal transformer fusion module to fully complement the associated information. It makes the generator to focus on both local and global context. TCGAN fully considers the training efficiency of the adversarial process and the integrated retention of redundant information. Various experimental results of TCGAN have highlighted targets, rich details, and fast convergence properties on public datasets.

Feature representation has received more and more attention in image classification. Existing methods always directly extract features via convolutional neural networks (CNNs). Recent studies have shown the potential of CNNs when dealing with images' edges and textures, and some methods have been explored to further improve the representation process of CNNs. In this article, we propose a novel classification framework called the multiscale curvelet scattering network (MSCCN). Using the multiscale curvelet-scattering module (CCM), image features can be effectively represented. There are two parts in MSCCN, which are the multiresolution scattering process and the multiscale curvelet module. According to multiscale geometric analysis, curvelet features are utilized to improve the scattering process with more effective multiscale directional information. Specifically, the scattering process and curvelet features are effectively formulated into a unified optimization structure, with features from different scale levels being efficiently aggregated and learned. Furthermore, a one-level CCM, which can essentially improve the quality of feature representation, is constructed to be embedded into other existing networks. Extensive experimental results illustrate that MSCCN achieves better classification accuracy when compared with state-of-the-art techniques. Eventually, the convergence, insight, and adaptability are evaluated by calculating the trend of loss function's values, visualizing some feature maps, and performing generalization analysis.

Deep learning (DL) has made breakthroughs in many computer vision tasks and also in visual tracking. From the beginning of the research on the automatic acquisition of high abstract feature representation, DL has gone deep into all aspects of tracking to date, to name a few, similarity metric, data association, and bounding box estimation. Also, pure DL-based trackers have obtained the state-of-the-art performance after the community's constant research. We believe that it is time to comprehensively review the development of DL research in visual tracking. In this article, we overview the critical improvements brought to the field by DL: deep feature representations, network architecture, and four crucial issues in visual tracking (spatiotemporal information integration, target-specific classification, target information update, and bounding box estimation). The scope of the survey of DL-based tracking covers two primary subtasks for the first time, single-object tracking and multiple-object tracking. Also, we analyze the performance of DL-based approaches and give meaningful conclusions. Finally, we provide several promising directions and tasks in visual tracking and relevant fields.

With the development of remote sensing technology, panchromatic images (PANs) and multispectral images (MSs) can be easily obtained. PAN has higher spatial resolution, while MS has more spectral information. So how to use the two kinds of images' characteristics to design a network has become a hot research field. In this article, a multi-scale progressive collaborative attention network (MPCA-Net) is proposed for PAN and MS's fusion classification. Compared to the traditional multi-scale convolution operations, we adopt an adaptive dilation rate selection strategy (ADR-SS) to adaptively select the dilation rate to deal with the problem of category area's excessive scale differences. For the traditional pixel-by-pixel sliding window sampling strategy, the patches which are generated by adjacent pixels but belonging to different categories contain a considerable overlap of information. So we change original sampling strategy and propose a center pixel migration (CPM) strategy. It migrates the center pixel to the most similar position of the neighborhood information for classification, which reduces network confusion and increases its stability. Moreover, due to the different spatial and spectral characteristics of PAN and MS, the same network structure for the two branches ignores their respective advantages. For a certain branch, as the network deepens, characteristic has different representations in different stages, so using the same module in multiple feature extraction stages is inappropriate. Thus we carefully design different modules for each feature extraction stage of the two branches. Between the two branches, because the strong mapping methods of directly cascading their features are too rough, we design collaborative progressive fusion modules to eliminate the differences. The experimental results verify that our proposed method can achieve competitive performance.