Hyperspectral image (HSI) classification plays a critical role in remote sensing by enabling precise land-cover identification through rich spectral information. While deep learning has led to significant progress, over 90 % of existing methods rely on small patch-based networks, which suffer from two key limitations: (1) restricted receptive fields (e.g., 7 × 7 , 9 × 9 ) that hinder structural awareness and result in noisy misclassifications within homogeneous regions; and (2) undefined optimal patch sizes that lead to coarse label predictions and degraded accuracy. Inspired by large-scale image segmentation techniques such as U-Net architectures—known for their strong boundary delineation and spatial coherence—this study explores their adaptation to HSI classification. However, such adaptations remain underutilized due to challenges including performance concerns with large patches, abundant unlabeled regions, and input-shape mismatches. To address these gaps, we propose HSIseg, a segmentation-adapted framework for HSI classification. HSIseg incorporates three novel components—Dynamic Shifted Regional Transformer (DSRT), Discriminative Feature Selection (DFS), and Cross Feature Interaction (CFI)—to enhance feature representation and fusion. A progressive learning strategy with adaptive pseudo-labeling is employed to leverage unlabeled data, while multi-source data collaboration further strengthens the model’s capability. Extensive experiments on five public datasets demonstrate the effectiveness of HSIseg in overcoming the limitations of traditional patch-based approaches. Code is available at https://github.com/zhouweilian1904/HSI_Segmentation .

Hyperspectral image (HSI) classification plays a crucial role in remote sensing (RS) applications, enabling the precise identification of materials and land cover based on spectral information. This supports tasks such as agricultural management and urban planning. While sequential neural models like Recurrent Neural Networks (RNNs) and Transformers have been adapted for this task, they present limitations: RNNs struggle with feature aggregation and are sensitive to noise from interfering pixels, whereas Transformers require extensive computational resources and tend to underperform when HSI datasets contain limited or unbalanced training samples. To address these challenges, Mamba architectures have emerged, offering a balance between RNNs and Transformers by leveraging lightweight, parallel scanning capabilities. Although models like Vision Mamba (ViM) and Visual Mamba (VMamba) have demonstrated improvements in visual tasks, their application to HSI classification remains underexplored, particularly in handling land-cover semantic tokens and multi-scale feature aggregation for patch-wise classifiers. In response, this study introduces the Mamba-in-Mamba (MiM) architecture for HSI classification, marking a pioneering effort in this domain. The MiM model features: (1) a novel centralized Mamba-Cross-Scan (MCS) mechanism for efficient image-to-sequence data transformation; (2) a Tokenized Mamba (T-Mamba) encoder that incorporates a Gaussian Decay Mask (GDM), Semantic Token Learner (STL), and Semantic Token Fuser (STF) for enhanced feature generation; and (3) a Weighted MCS Fusion (WMF) module with a Multi-Scale Loss Design for improved training efficiency. Experimental results on four public HSI datasets—Indian Pines, Pavia University, Houston2013, and WHU-Hi-Honghu—demonstrate that our method achieves an overall accuracy improvement of up to 3.3%, 2.7%, 1.5%, and 2.3% over state-of-the-art approaches (i.e., SSFTT, MAEST, etc.) under both fixed and disjoint training-testing settings. • A novel multi-scale pyramid Mamba model for efficient HSI classification. • Tokenized Mamba encoder enhances Mamba’s suitability for visual tasks. • Centralized Mamba-Cross-Scan improves patch-wise HSI sequential classifiers. • Satisfying classification performance with fixed and disjoint training-testing samples.

Currently, many target classification and recognition methods for synthetic aperture radar (SAR) images rely on extracting common features of targets from images at different azimuth angles. These methods are highly dependent on the quantity of target image data and are sensitive to feature variations. Specifically, when changes in incidence angle cause image differences, the recognition accuracy of such methods fails to obtain satisfying results. To address the above issues, a solution based on simulating human visual characteristics is proposed to facilitate feature interaction between images taken from different incident angles, thereby identifying the variation patterns of target features. By integrating convolutional gated recurrent units, weighted coupling units, and capsule networks, a target recognition network is designed that takes multi-incident angle images as input. Additionally, to enhance the correlation between consecutive results and improve recognition accuracy, a decision attention mechanism (DAM) is introduced that optimizes multistep classification decisions. The output capsule vectors from multiple passes through the network are enhanced in a specialized reinforcement manner to strengthen classification features and correct misclassification issues during the decision-making process. The experiments are conducted on simulated multi-incidence angle vehicle and ship targets, followed by fine-tuning experiments on a small set of real target samples using the model trained on simulated data. The results demonstrate that the proposed method exhibits superiority and robustness on both simulated and real samples.

Unsupervised low-rank matrix decomposition (LRMD) theory, leveraging inherent structural features of images, has demonstrated significant potential for synthetic aperture radar (SAR) image change detection (CD) under label scarcity. However, conventional methods rely on static sparsity priors, which inadequately model the spatial diversity and dynamic complexity of changes, leading to great decomposition errors. To address this limitation, we redefine CD within a multiview framework and propose a structurally diversified self-representation learning model. By jointly enforcing low-rank and sparse constraints on the coefficient matrix, our approach enhances the global consistency and local continuity of change information, effectively rectifying representation errors in complex change patterns. Furthermore, a robust background suppression mechanism is integrated to mitigate speckle noise and pseudo-changes, improving the discriminability between changed and unchanged regions. To refine ambiguous boundaries, an unsupervised classification refinement module is developed, calibrating transitional samples via nonlinear regression-based feature projection without labeled data. The entire framework operates without supervision, eliminating dependence on annotated samples. Extensive experiments on six bitemporal and three extended multitemporal datasets validate the effectiveness and superiority of the proposed method. The source code will be made available athttps://github.com/95xiaoli/S3CD

Deep learning technologies have significantly improved performance in the field of synthetic aperture radar (SAR) image target recognition compared to traditional methods. However, the inherent “black box" property of deep learning models leads to a lack of transparency in decision-making processes, making them difficult to be widespread applied in practice. This is especially true in SAR applications, where the credibility and reliability of model predictions are crucial. The complexity and insufficient explainability of deep networks have become a bottleneck for their application. To tackle this issue, this study proposes a physically explainable framework for complex-valued SAR image recognition, designed based on the physical process of microwave propagation. This framework utilizes complex-valued SAR data to explore the amplitude and phase information and its intrinsic physical properties. The network architecture is fully parameterized, with all learnable parameters endowed with clear physical meanings. Experiments on both the complex-valued MSTAR dataset and a self-built Qilu-1 complex-valued dataset were conducted to validate the effectiveness of framework. The de-overlapping capability of EMWaveNet enables accurate recognition of overlapping target categories, whereas other models are nearly incapable of performing such recognition. Against 0dB forest background noise, it boasts a 20% accuracy improvement over traditional neural networks. When targets are 60% occluded by noise, it still outperforms other models by 9%. An end-to-end complex-valued synthetic aperture radar automatic target recognition (SAR-ATR) algorithm is constructed to perform recognition tasks in interference SAR scenarios. The results demonstrate that the proposed method possesses a strong physical decision logic, high physical explainability and robustness, as well as excellent de-aliasing capabilities. Finally, a perspective on future applications is provided.

Synthetic aperture radar (SAR) target recognition frequently encounters the emergence of novel targets. Updating and iterating existing models with newly arriving data constitutes an incremental learning task. However, training with only new data often causes catastrophic forgetting, leading to a substantial loss of performance on previously learned classes, while retraining the entire model is prohibitively costly. This necessitates the development of algorithms with strong continual learning capabilities. Moreover, compared with traditional optical scenarios, SAR incremental tasks pose even greater challenges: due to the imaging mechanism, target characteristics, and application scenarios, they are highly sensitive to azimuth variations, exhibit low inter-class variability, and demand lightweight solutions. To address these issues, this paper proposes a scattering-aware adaptive dynamic node generation (SDG) algorithm for class-incremental learning. First, the attributed scattering center model (ASCM) is integrated to extract representative prototype vectors, which enhance spatial attention and calibrate the parameters of the expanded backbone network. Second, a gradient-constrained and indicator-driven dynamic network expansion strategy, combined with parameter minimization, ensures both model compactness and efficiency. Finally, a multi-teacher knowledge distillation approach is introduced, where different backbones serve as distinct teachers to mitigate the low inter-class variability inherent in SAR images. Extensive experiments on the MSTAR, SAR-AIRcraft, and FUSAR-Ship datasets demonstrate that SDG achieves significant improvements in both recognition accuracy and parameter efficiency. This study thus provides a lightweight and effective solution for real-time SAR target recognition on resource-constrained platforms.

In recent years, supervised learning has seen notable progress in research on aircraft target detection using synthetic aperture radar (SAR) images, but its performance is heavily restricted by the amount of labeled data. Self-supervised learning (SSL) can address this problem by pre-training on unlabeled data to extract generalizable features. However, existing SSL methods are predominantly developed based on approaches designed for natural image processing, making them unsuitable for fully understanding SAR images. Therefore, this study proposes a self-supervised contrastive learning method (SARDet-CL) that integrates feature enhancement and imaging mechanism constraints, tailored to the characteristics and requirements of aircraft target detection tasks. Specifically, SARDet-CL enhances the model’s ability to represent spatiotemporal features through spatial feature masking and multi-temporal consistency learning, and robustly extracts target features using a dynamic threshold quantization strategy. In addition, structural information is inferred based on the SAR imaging mechanism and incorporated as an auxiliary self-supervisory signal to guide the model toward a more accurate understanding of SAR images. Experimental results based on various SAR datasets show that the proposed method outperforms other pre-training techniques in down-stream SAR target detection tasks.

Compared to optical sensors, Synthetic Aperture Radar (SAR) can acquire remote sensing imagery under all-weather and all-time conditions. This technology is extensively applied in maritime vessel detection. Recently, deep learning approaches have shown promising performance in ship detection in SAR images. However, accurate detection still poses challenges in coastal scenes, especially in areas with densely arranged ships. To address this issue, this paper presents a ship target detection method for nearshore areas based on cluster detection. This method employs a cluster detection module to focus the detector on densely arranged ship areas and performs refined secondary detection. Subsequently, local and global results are strategically combined to derive the final outcome. Experiments conducted on the public datasets SSDD and RSDD-SAR validate the efficiency of the proposed method in detecting ships in densely arranged areas.

Although Convolutional Neural Networks (CNNs) have achieved success in traditional Closed-Set Recognition (CSR), their lack of robustness when faced with unknown categories in open environments is evident. In this context, Open Set Recognition (OSR) becomes particularly crucial, as it not only effectively classifies known categories but also accurately identifies unknown categories as "unknown," which is pivotal for the algorithm’s robustness in real-world applications. To address this challenge, we propose a prototype learning network based on attribute scattering center (ASC-CPL) to tackle open-set recognition issues in SAR image analysis. Firstly, this framework retains the representational capabilities of CNN while discarding the closed-world Softmax assumption, opting instead for an open-world and human-like prototype model. Secondly, to better handle the discreteness and sensitive pose characteristics of SAR image targets, this approach integrates attribute scattering centers into the convolutional kernel design, significantly enhancing the network’s ability to model scattering features. Experimental verification on the MSTAR dataset demonstrates the effectiveness and feasibility of this method, providing robust support for the future development of open-set recognition in the field.

Synthetic aperture radar (SAR) images are widely utilized for the detection and recognition of aircraft targets. Unlike optical images, SAR images possess the advantage of being applicable in all weather conditions and at all times of the day. However, in contrast to optical images, aircraft typically appear as discrete points in SAR images, and their outlines are not distinctly clear. To effectively use the scatter information of aircraft in SAR images, a convolutional modulation scattering feature network (CMSF) is proposed in this paper. Firstly, a scattering feature extraction module is introduced to make full advantage of scattering information. Secondly, following convolution processing, the convolution modulation module is employed to generate a similar fraction matrix. Thirdly, the fusion of scattering features and convolution features is achieved through convolutional modulation and matrix multiplication. Finally, a four-stage convolutional processing is employed to recognize the aircraft target. Extensive experiments conducted on the SARAircraft-1.0 dataset demonstrate the effectiveness of the convolutional modulated scattering feature network for aircraft target classification in SAR images.

Image registration plays a pivotal role in various image processing applications, which is widely used in image fusion and change detection. However, The presence of speckle noise in SAR images causes a primary reduction in registration accuracy and existing algorithms have not achieved high-precision registration while maintaining low computational complexity. This paper proposes an image registration algorithm based on edge intersections and deep learning descriptors. An edge-directed voting mechanism is introduced to identify corner points, and a custom SAR image dataset is constructed to retrain the HardNet descriptor network. Experimental results validate the superiority of the proposed method in terms of robustness and accuracy, achieving SAR image registration on a self-constructed dataset with an RMSE of 0.38, showcasing the utmost registration accuracy, while maintaining lower computational complexity than traditional approaches.

Currently, with the increasing number of satellites and the expansion of applications, the amount of remote sensing image data is growing geometrically. Due to limitations, such as bandwidth and power, it is less efficient to transmit the large amounts of data to ground servers for further processing. Real-time onboard image inference is clearly a better choice. However, previous efficient neural networks are mostly optimized for parameters rather than inference speed. To alleviate the speed bottleneck, an efficient backbone MobileShuffle combining hardware-friendly modules and structural reparameterization technique is proposed in this letter. In the proposed network, large amounts of depthwise separable convolution (DSC) are used to accelerate the inference process. A new structural reparameterization method is applied to enable the network better focus on the critical part of SAR images. Visualized heatmaps show that the learning focus of the network is optimized by multibranch architecture during training. The smallest variant has similar latency to the state-of-the-art resource efficient network MobileOne-S0, but with a 0.9% improvement in accuracy. Results on optical and SAR datasets show that MobileShuffle has excellent accuracy-speed tradeoff compared with other networks. Code and models are available athttps://github.com/2474137474/MobileShuffle.

With the rapid development and widespread deployment of radar technology, the interpretation of the vast amount of synthetic aperture radar (SAR) imagery obtained daily has emerged as a hot topic. The complexity of the electromagnetic scattering mechanisms contained within radar images makes SAR image interpretation a challenging task. Current methodologies for SAR image interpretation primarily involve feature extraction-based techniques, categorized into image-domain and frequency-domain algorithms. However, these methods are faced with issues, such as rough segmentation in images, high-computational complexity, and poor robustness, presenting significant challenges in the field. In this article, a novel framework for SAR image interpretation is proposed, leveraging reinforcement learning (RL) for the extraction of electromagnetic scattering features and the inversion of parameters. Within this framework, a nonsparse reward function, combined with curriculum learning, is introduced as the supervisory information. It enables more efficient policy updates through a structured two-stage training approach. In addition, an algorithm that integrates a four-neighbor breadth-first search (BFS) with the watershed segmentation process is proposed, aiming to enhance the accuracy of scattering center analysis in SAR imagery. The attribute scattering center model (ASCM) is utilized as a prototype for conducting algorithmic research and experimentation. Experimental results on both simulation data and measured data have indicated that the proposed method significantly improves efficiency while ensuring accuracy, demonstrating its capability to extract parameters from measured data in most scenarios.

Synthetic aperture radar (SAR) has emerged as a vital tool for ship monitoring due to its all-weather, all-day high-resolution imaging capabilities. In practical operations, the wide coverage and sparse ship distribution in very large full-scene SAR images pose challenges in terms of low efficiency and high false alarm rates. Traditional methods perform poorly in complex scenarios, while deep learning (DL) methods have high computation cost. This study proposes a fast progressive detection algorithm for ship targets in large SAR images, combining the advantages of traditional Non-DL methods and DL approaches. First, at a global scale, image preprocessing operations based on traditional methods are designed to quickly extract candidate regions. Then, at regional scale, an oriented ship detector is designed for refined ship detection within candidate regions. Finally, at individual-target scale, a false alarm discrimination network is constructed to further remove false alarms. Experimental results on GF-3 full-scene SAR images demonstrate that the proposed method can achieve minutes-level detection efficiency in images of billion-pixel-level size, while achieving high detection accuracy.

Superpixel analysis is showing great potential for high-resolution synthetic aperture radar (SAR) image change detection, as it uses a larger detection granularity and enhances computational efficiency. However, some deficiencies still exist. It is difficult for previous methods to extract the complete change regions from cluttered backgrounds. Meanwhile, the local spatial correlation and structure consistency are not well represented. To address the above problems and achieve better separation of changed and unchanged superpixels in complicated scenarios, we design a novel unsupervised change detection framework from the perspective of low-rank matrix decomposition (LRMD) theory. The entire framework is carried out in two stages. Firstly, the ℓ1-norm sparsity constraint LRMD model is constructed to decompose change features into a low-rank component associated with background and a sparse component representing changed regions. Then, the local spatial correlation and structure consistency constraint are explicitly modeled by introducing a Laplacian regularization term. The unified model smooths the local similarity superpixels and enlarges the distance between changed regions and the background in the feature subspace. In this stage, the saliency difference image (DI) is generated to indicate the change probabilities of each superpixel. Furthermore, a classification refining module is designed to learn the projection from the change feature matrix to the saliency DI, which can further fine-tune such obscure regions and boost the binary classification. Extensive experiments on five challenging datasets from the TerraSAR-X sensor demonstrate the effectiveness and superiority of the proposed method.

Target recognition for synthetic aperture radar (SAR) images has been a longstanding hotspot. However, using polarimetric information to enhance recognition performance is under-researched. In this paper, we develop a visualization approach to analyze and highlight the contributions of polarimetric elements, and propose a simple polarimetric correlation feature for target recognition. In the visualization method, a channel-wise convolutional structure is well developed, which serves as a proxy of the polarimetric elements. By using the gradients of the target class flowing into each developed convolutional channel with normalization, we obtain activation maps indicating the contribution of each polarimetric element. Then a comprehensive quantitative evaluation of polarimetric element contribution is also performed. We demonstrate again that polarimetric information maintains significant advantages over single-polarization intensity, and the correlations between cross- and co-polarization emerge as key components for target recognition. Exactly inspired by these insights, the straightforward feature is defined as the correlation between the target and typical scatterers, seamlessly fusing intensities and correlations. This polarimetric correlation feature aptly encapsulates the physical scattering of the target, shows a clear mapping relationship with the geometric structure, and captures the differences between categories, achieving intra-category compactness and inter-category separation. Experimental results on both synthetic and real datasets underscore the effectiveness and superiority of this proposed feature.

Amidst the landscape of contemporary remote sensing technology, the endeavor to detect and recognize aircraft within remote sensing images (RSIs) assumes pivotal strategic and practical significance. The complex nature of fine-grained aircraft recognition is a result of the intricate interplay between aircraft and their background environments, alongside category imbalance, which collectively lead to the emergence of a long-tail distribution within the dataset. However, experts proficient in RSIs interpretation can effectively address these challenges through the application of prior knowledge. This paper introduces the Aircraft Reasoning Network (ARNet), a framework tailored for aircraft detection and fine-grained recognition in RSIs, building upon prior knowledge employed in expert interpretation. Specifically, the Knowledge Reasoning Module (KRM) introduces a knowledge graph that incorporates both common and expert knowledge into the end-to-end network. Additionally, the network encompasses a Spatial Context Module (SCM) and an Airport Facility Relationship Module (AFRM). These components facilitate highly accurate detection and recognition of fine-grained aircraft in diverse environmental contexts by employing adaptive prior knowledge reasoning and optimizing target spatial location. Furthermore, an independent Aircraft Component Discrimination Module (ACDM) distinguishes aircraft based on their predominant component features, contributing to improved classification performance in both the few-shot and easily confused categories. Moreover, this paper introduces the AR-RSI dataset, a compilation of RSIs capturing fine-grained aircraft targets from diverse locations. The effectiveness and superiority of ARNet are exemplified on AR-RSI, achieving a minimum of 3.7 percentage higher mAP than the mainstream aircraft detection framework.

This study introduces an innovative approach in hyperspectral imaging (HSI) classification by integrating convolution, recurrence, and self-attention mechanisms in a 3D configuration. We address several challenges such as the 1) disruption of spectral continuity by traditional dimensionality reduction methods like PCA, 2) the overlooking of band-to-band continuous features in existing spatial-only 2D multiscanning strategy, and 3) the limitations in model design by simply cascading recurrent neural networks (RNNs) with Transformers for HSI analysis. Our solution involves three core components: 1) sub-band grouping with group-wise convolution for refined dimension reduction, 2) a novel cubed 3D-multiscanning technique enabling thorough multi-directional analysis in both spectral and spatial domains, and 3) the development of a Cubic-Net framework with a specially designed Segmented Recurrent Transformer (SRT). This SRT is tailored to effectively utilize spectral continuity along with spatial contextual features, overcoming common sequential data analysis challenges seen in RNNs and Transformers. Furthermore, our feature fusion strategy successively integrates ‘short-term’ and ‘long-term’ SRT features, thereby enhancing the model’s ability to process both spectral and spatial features effectively. Experimental results from three public HSI datasets indicate our method’s improved performance over existing baselines and state-of-the-art methods. This research offers a new perspective in 3D sequential HSI classification.

High-resolution SAR images can be modeled as the superposition of attributed scattering centers (ASC). Extracting ASC features from SAR images is considered to be a high-dimensional, nonlinear, and non-convex model parameter optimization problem. In early works, ASC parameter estimation is solved by performing scattering center segmentation from image domain SAR data with an iterative optimization algorithm, or performing Bayesian learning and dictionary-based learning from frequency domain SAR data. However, almost all these parameter estimation methods lack an efficient parameter update strategy, and the large parameter space brings great algorithm complexity. In this paper, it is proposed to extract ASCs by reinforcement learning by modeling the repetitive iterative process of parameter optimization as the interaction process between the agent and the environment, to improve the parameter update efficiency in the inference stage.

Detecting infrared targets in complex backgrounds is a demanding task, especially for small targets. Infrared small target is usually difficult to be detected accurately because of the complex background and the lack of rich color information and edge information. Based on this difficulty, a Multiscale Interactive Attention Network (MIA-Net) is proposed to accurately detect small infrared targets in complex backgrounds. Specifically, the feature skip interactive module (FSIM) is firstly designed for realizing multiscale feature information transmission and fusion, which solves the problem of losing infrared small target features after multiple convolutions. Based on this, a multiscale edge reconstruction (MER) block is designed to learn the edge features of small targets and achieve optimal reconstruction of their edges. Extensive experiments were conducted on the SIRST dataset to evaluate the proposed MIA-Net and the results show that our MIA-Net performs better in terms of IoU, Pd, and Fa.

The problem of insufficient samples has been limiting the performance of intelligent interpretation in Synthetic Aperture Radar (SAR) images. Humans have the ability to recognize new instances with only a few samples, indicating that attributes play a crucial role in recognition. Attributes can be shared across categories and provide a distinctive representation. Motivated by this fact, this paper proposes an attribute-guided network consisting of a base classifier (BC) and an attribute classifier (AC). Firstly, we design binary attributes for SAR objects to enable more distinct feature representations. Secondly, the images are mapped into a semantic embedding space by embedding the attribute vectors. Finally, the performance of few-shot classification in SAR images is improved by jointly optimizing the loss in both attribute space and deep feature space. Our attribute-based framework is validated through ablation experiments on the MSTAR dataset and a self-built SAR aircraft dataset.

Change detection has played an increasingly important role in multitemporal remote sensing applications recently. Long time series analysis is providing new information of land cover changes and improving the quality and accuracy of the change information being derived from remote sensing. The purpose of this study is to dig for more change temporal information and change pattern information from synthetic aperture radar (SAR) image time series (ITS), which is of great significance for monitoring urban area changes, conducting land use surveys, and renovating illegal constructions. In the study, a novel unified framework for long time series SAR image change detection and change pattern analysis (SAR-TSCC) was proposed for land cover change mapping. To obtain the most notable change time rapidly, a fast SAR ITS change point search method based on pruned exact linear time (SAR-PELT) algorithm was adopted. Meanwhile, the deep time series classification network, named SAR time series transformer (SAR-TST), was implemented to recognize the change patterns, which is based on time series transformer (TST) architecture. Considering the lack of real training data, a novel synthetic data generation method is developed. The combination of the synthetic and real data enhanced the generalization of the classifiers. The proposed framework was used for monitoring a large urbanization area in the northwest of Hong Kong, China. The Cosmo Skymed (CSK) time series data acquired from 2013 to 2020 were exploited for land cover change analysis. Experiment results showed that our approach achieved the state-of-the-art performance, as the time accuracy reached 86% and the classification accuracy on the four main change patterns (impulse, step, cycle, and complex) is over 99%. In particular, the proposed SAR-TST model showed remarkable advantages in the presence of insufficient real data.

The goal of hyperspectral image (HSI) classification is to assign land-cover labels to each HSI pixel in a patch-wise manner. Recently, sequential models, such as recurrent neural networks (RNN), have been developed as HSI classifiers which need to scan the HSI patch into a pixel-sequence with the scanning order first. However, RNNs have a biased ordering that cannot effectively allocate attention to each pixel in the sequence, and previous methods that use multiple scanning orders to average the features of RNNs are limited by the validity of these orders. To solve this issue, it is naturally inspired by Transformer and its self-attention to discriminatively distribute proper attention for each pixel of the pixel-sequence and each scanning order. Hence, in this study, we further develop the sequential HSI classifiers by a specially designed RNN-Transformer (RT) model to feature the multiple sequential characters of the HSI pixels in the HSI patch. Specifically, we introduce a multiscanning-controlled positional embedding strategy for the RT model to complement multiple feature fusion. Furthermore, the RT encoder is proposed for integrating ordering bias and attention re-allocation for feature generation at the sequence-level. Additionally, the spectral-spatial-based soft masked self-attention is proposed for suitable feature enhancement. Finally, an additional Fusion Transformer is deployed for scanning order-level attention allocation. As a result, the whole network can achieve competitive classification performance on four accessible datasets than other state-of-the-art methods. Our study further extends the research on sequential HSI classifiers.

Deep learning (DL) has been fast developed and widely used in Synthetic Aperture Radar (SAR) image target detection with the booming of AI technique. Many DL models have been proposed. However, most of them cannot work in embedded devices because of the limited hardware resource. In this paper, a post-quantification method for neural networks is proposed to reduce computational costs. Embedded devices usually support best for Float16 and Int8 low-bit data operations. A training procedure to preserve end-to-end model accuracy is designed, so that the bit depth of the model is compressed without significant loss of model accuracy, and can efficiently reduce the complexity of the model space, reducing the complexity of inference time. To verify our study, the quantified model is deployed on embedded devices to perform inference on 1024*1024 SAR images. The speed can reach 26FPS, which is 4 times faster than the inference speed before quantization.

Aircraft detection in synthetic aperture radar (SAR) image is a special case because all the targets are located in the airport. Comparing with the whole scene SAR image, the area of an airport is relatively small, therefore, this information can be utilized to speed up the algorithm. This paper proposes a centroid network detection method based on the combination of geographic coordinate information and subscene classification. Firstly, the airport area is detected based on the priori geographic coordinate information. Secondly, to further narrow down the scope of detection and extract the regions containing valid targets, the subscene is fed into the ResNet50 network which incorporates Squeeze and Excitation (SE) to separate the aircraft area from the background area. The method is validated in ablation experiments on the GaoFen-3(GF3) datasets to reveal the impact of each factor. The results show that the proposed method can achieve a reduction in false alarm rate around 6% and time cost around 30%.

Data augmentation is beneficial when the measured training data are insufficient to train a robust deep model. One of the promising techniques is to use simulated data generated by physics-based engines. For example, few-shot learning of synthetic aperture radar (SAR) targets could be benefited from simulated SAR images. However, the characteristics of the simulated training data significantly affect the performance of the trained model. Therefore, it is of great significance to evaluate the utility of simulated data objectively and effectively. A recognition rate versus substitution rate curve (RSC)-based assessment criterion is proposed, consisting of substitution rate (SR)-based dataset allocation stage and RSC-based evaluation stage. First, the differential dataset allocation is performed under a progressive SR to obtain paired reference and comparison training sets. Then, the reference and comparison classifiers are trained under different SRs using the same network and parameter configuration in the RSC criterion-based evaluation stage. AconvNet and AlexNet are selected as the backbones of the evaluation network. Especially, k-fold cross-validation is applied to alleviate selection bias. The difference between the integrals of RSCs is defined as the RSC score for the simulated dataset. Experiments conducted on the measured and simulated moving and stationary target acquisition and recognition (MSTAR) database demonstrate the rationality and validity of the proposed RSC criterion. Specifically, multisource simulated datasets are adopted, including the adversarial autoencoder-generated and electromagnetic simulation datasets. The proposed RSC criterion shows promising utility evaluation ability, flexibility, and extensibility compared with traditional full-reference image-quality assessment criteria.

Most methods based on the convolutional neural network show satisfying performance for hyperspectral image (HSI) classification. However, the spatial dependence among different pixels is not well learned by CNNs. A recurrent neural network (RNN) can effectively establish the dependence of nonadjacent pixels and ensure that each feature activation in its output is an activation at the specific location concerning the whole image, in contrast to the usual local context window in the CNNs. However, recent limited conversion schemes in RNN-based methods for HSI classification cannot fully capture the complete spatial dependence of an HSI patch. In this study, a novel multiscanning strategy with RNN is proposed to feature the sequential character of the HSI pixel and fully consider the spatial dependence in the HSI patch. By investigating different scanning forms, eight scanning orders are considered spatially, which flattens one local HSI patch into eight neighboring continuous pixel sequences. Moreover, considering that eight scanning orders complement one local patch with correlative dependence, the concatenated features from all scanning orders are fed into the RNN again for complementarity. As a result, the network can achieve competitive classification performance on three publicly accessible datasets using fewer parameters than other state-of-the-art methods.

Aircraft detection in synthetic aperture radar (SAR) images is a challenging task because of the discreteness, variability, and interference of aircraft scattering features. This article proposes a new hybrid approach of scattering information enhancement (SIE) and an attention pyramid network (APN). It first extracts strong scattering points (SSPs) of aircraft via an adapted Harris-Laplace detector. These SSPs are then clustered into candidate scattering regions by density-based spatial clustering of applications with noise (DBSCAN) and are then modeled with a Gaussian mixture model (GMM). Target scattering clusters are discriminated from background clutter by measuring the Kullback-Leibler divergence (KLD) to the known target templates. These target scattering clusters are enhanced in the preprocessing stage. All the SIE-preprocessed images are then fed into the APN for training and testing. It is composed of the multiscale feature pyramid network (FPN) and the modified convolutional block attention module (CBAM) to cope with the discreteness and variability of aircraft. In addition, focal loss (FL) is adopted to deal with the issue of unbalanced sample distribution and the interference from hard samples. Experiments conducted on the Gaofen-3 and TerraSAR-X data sets demonstrate the effectiveness of the proposed method with an average precision (AP) of 83.25%.

Target detection is an important application in remote sensing. In this paper, an end-to-end aircraft detection algorithm is proposed for large scene spaceborne synthetic aperture radar (SAR) imagery. Due to the diversity and variability of scattering mechanism, representative features including edge information and depth characteristics are utilized in the algorithm. Firstly, the accurate airport mask is extracted via Otsu algorithm and adaptive identification operator (AIO) with airport morphological features. Secondly, edge-detection based on Canny operator and k-means clustering algorithm are adopted to generate candidate areas. Finally, aircraft targets are discriminated from candidate areas via ResNet-based convolutional neural network (CNN). Experiments are conducted on collected spaceborne SAR imagery, and the results indicate that the proposed algorithm can extract airport area precisely and detect aircraft accurately with low false alarm.

Terrain surface classification is probably the most common application of polarimetric SAR (PolSAR) data. Methods for PolSAR terrain classification can be divided into either supervised or unsupervised. In this paper, PolSAR image classification algorithms are reviewed from traditional polarimetric methods such as alpha-H-, Freeman-, Yamaguchi-decomposition, to deep learning, and then a general deep learning algorithm is proposed to PolSAR data classification. The suitability and potential of deep convolutional neural network in supervised terrain classification of PolSAR images has been investigated. The results show that deep learning based method can be used for PolSAR terrain classification.

In this paper, a semi-physical system of SAR jamming based on electromagnetic scattering simulation was designed and constructed by NI and xPC platform. The basic idea is transmitting simulated target scattering echoes to realize jamming for real SAR system. Scattering signal is simulated by our laboratory developed software, and a hardware system was designed to receive SAR signal and transmit simulated echoes. A software radio ADALM-Pluto is used as radar, and jammer adopts NI PXle-1082 and NI PXle-5840. The experimental results show that the designed semiphysical system can reach the goal of jamming by transmitting simulated images. The proposed system can be used to verify SAR jamming and anti-jamming algorithm for researchers.

With growing needs of the global land cover information, high resolution datasets have been released, such as 30m-resolution GLC30 and NLCD 2011. However, these datasets cannot update in time when the ground-truth changes. In this paper, an automatic algorithm is proposed to generate land cover based on convolutional neural network (CNN) using optical image. A land cover translation framework is designed by utilizing the techniques of the fully convolutional network, and it can provide pixel-to-pixel translation of the image. The experiment is carried out on high resolution optical image sourcing from Google earth, and the results demonstrate that the proposed method is able to generate robust and reasonable prediction of the land cover. By testing at several different areas, this methods achieve the average accuracy at 70.2% for 8 types land cover comparing with NLCD 2011 datasets.

Artificial intelligence such as deep learning has become the dominant approach in computer vision area. It has great potential in improving the performance of SAR automatic target recognition (ATR) as well. In this paper, we present a framework for intelligence SAR ship recognition and a preliminary implementation as well as a demonstration with the ALOS2 data.

A microwave/mmw imaging method for moving objects of non-rigid body using 2D sparse MIMO (multiple-input multiple-output) array is proposed in this paper. To achieve high-resolution 3D image, two orthogonal linear arrays are introduced to form a 2D sparse MIMO array and wide-band signals are transmitted in this method. Subsequently, a space-time trajectory model is introduced to describe the motion of the non-rigid body target. The joint use of segmental and joint-estimation leads to the effective movement compensation of each component. Finally, all the images of components are merged into an image of the whole target and the 3D high-resolution image of the target is reconstructed.

Following the great success of deep convolutional neural networks (CNNs) in computer vision, this paper proposes a complex-valued CNN (CV-CNN) specifically for synthetic aperture radar (SAR) image interpretation. It utilizes both amplitude and phase information of complex SAR imagery. All elements of CNN including input-output layer, convolution layer, activation function, and pooling layer are extended to the complex domain. Moreover, a complex backpropagation algorithm based on stochastic gradient descent is derived for CV-CNN training. The proposed CV-CNN is then tested on the typical polarimetric SAR image classification task which classifies each pixel into known terrain types via supervised training. Experiments with the benchmark data sets of Flevoland and Oberpfaffenhofen show that the classification error can be further reduced if employing CV-CNN instead of conventional real-valued CNN with the same degrees of freedom. The performance of CV-CNN is comparable to that of existing state-of-the-art methods in terms of overall classification accuracy.

Inspired by nonlocal filtering, this paper proposes a saliency detector based on pattern recurrence. In visual attention and saliency detection framework, people are extracting patches that are not redundant, which would more likely to attracting attention. Hence, target detection or saliency detection in SAR image could also be done using the dissimilarity as an indicator of saliency, meaning that the interesting target or saliency target is different from background around it. Similarity of two pixels can be defined together with their local neighbors, then calculate the cross-correlation of two normalized patches. To analyze results better, a normalized version of cross-correlation is used. Experimental results on SAR image are shown to test the effectiveness of the proposed method. The experimental results compared with CFAR on SAR images also prove the effectiveness in saliency detection on SAR images.

Deep convolutional neural networks have achieved great success in computer vision and many other areas. They automatically extract translational-invariant spatial features and integrate with neural network-based classifier. This letter investigates the suitability and potential of deep convolutional neural network in supervised classification of polarimetric synthetic aperture radar (POLSAR) images. The multilooked POLSAR data in the format of coherency or covariance matrix is first converted into a normalized 6-D real feature vector. The six-channel real image is then fed into a four-layer convolutional neural network tailored for POLSAR classification. With two cascaded convolutional layers, the designed deep neural network can automatically learn hierarchical polarimetric spatial features from the data. Two experiments are presented using the AIRSAR data of San Francisco, CA, and Flevoland, The Netherlands. Classification result of the San Francisco case shows that slant built-up areas, which are conventionally mixed with vegetated area in polarimetric feature space, can now be successfully distinguished after taking into account spatial features. Quantitative analysis with respect to ground truth information available for the Flevoland test site shows that the proposed method achieves an accuracy of 92.46% in classifying the considered 15 classes. Such results are comparable with the state of the art.

The algorithm of synthetic aperture radar automatic target recognition (SAR-ATR) is generally composed of the extraction of a set of features that transform the raw input into a representation, followed by a trainable classifier. The feature extractor is often hand designed with domain knowledge and can significantly impact the classification accuracy. By automatically learning hierarchies of features from massive training data, deep convolutional networks (ConvNets) recently have obtained state-of-the-art results in many computer vision and speech recognition tasks. However, when ConvNets was directly applied to SAR-ATR, it yielded severe overfitting due to limited training images. To reduce the number of free parameters, we present a new all-convolutional networks (A-ConvNets), which only consists of sparsely connected layers, without fully connected layers being used. Experimental results on the Moving and Stationary Target Acquisition and Recognition (MSTAR) benchmark data set illustrate that A-ConvNets can achieve an average accuracy of 99% on classification of ten-class targets and is significantly superior to the traditional ConvNets on the classification of target configuration and version variants.

In this paper, a new All-Convolutional Networks (A-ConvNets) is proposed and applied to Moving and Stationary Target Acquisition and Recognition (MSTAR) data. Conventional deep learning algorithms, especially the deep convolutional networks (ConvNets) have achieved many success state-of-art results. However, directly applying ConvNets to SAR data will yield severe overfitting because of limited data availability. The proposed A-ConvNets can significantly reduce the number of free parameters and the degree of overfitting. Average accuracy of 99.1% on classification of 10-class targets was obtained by applying A-ConvNets to MSTAR datasets.

Man-made building objects mostly with vertical wall structures may present distinct scattering patterns, e.g., wall/roof upfront scattering, wall-ground double scattering, etc., along the range dimension in high-resolution synthetic aperture radar (SAR) images. In this letter, a 1-D detector, referred to as the “range detector,” is presented for building detection, which operates only along the range direction. Experiments show that this range detector can effectively detect and extract the footprint of the illuminated wall of a cuboid building, with which the outline of the building image can be captured by marching the footprint toward radar. This approach is applied to an airborne Pi-SAR image of Sendai, Japan, and more than 80% of the buildings can be identified. The building height and length are also estimated, and the errors are found around 4-5 m based on optical image.

Deep learning algorithms such as convolutional neural networks (CNN) have been successfully applied in computer vision. This paper attempts to adapt the optical camera-oriented CNN to its microwave counterpart, i.e. synthetic aperture radar (SAR). As a preliminary study, a single layer of convolutional neural network is used to automatically learn features from SAR images. Instead of using the classical backpropagation algorithm, the convolution kernel is trained on randomly sampled image patches using unsupervised sparse auto-encoder. After convolution and pooling, an input SAR image is then transformed into a series of feature maps. These feature maps are then used to train a final softmax classifier. Initial experiments on MSTAR public data set show that an accuracy of 90.1% can be achieved on three types of targets classification task, and an accuracy of 84.7% is achievable on ten types of targets classification task.

In this paper, interferometric synthetic aperture radar (InSAR) data are utilized to extract the forestry damage information caused by typhoon over Tomakomai, Hokkaido, Japan. By computing the interferograms of two single-pass X-band SAR data before and after the typhoon, the change of their heights could be detected even the fallen tress were not cleared, and therefore, the damage level can then be evaluated. The results show that the main damaged areas detected by InSAR analysis are in agreement with the ground truth data.

The purpose of this study is to extract the forests destroyed by typhoons and to quantitatively estimate the damage levels by using high-resolution polarimetric synthetic aperture radar (SAR) data. The study area is located in Tomakomai, Hokkaido, Japan [1]–[4]. Two sets of data were acquired before and after the typhoon by the L-band airborne Pi-SAR (Polarimetric-interferometric SAR) with 3m × 3m resolution (4-look in azimuth direction). It was found that the values of RCS (Radar Cross Section) averaged over the whole image after the typhoon damage changed by −0.47 dB, 0.05 dB, and 0.64 dB at HH-, HV-, and VV-polarization respectively in comparison with those before the damage. To fully utilize the data, a scattering model of the linear combination of the cross- and co-polarization RCS changes was developed to estimate the damage levels. Similar analytical approaches were also applied using the three-component decomposition analysis. The changes in RCS of double-, volume- and surfacescattering mechanisms after the damage were respectively 27.5 dB, −0.20 dB and −20.3 dB. Finally, by comparing the results with the ground survey data, the accuracies of 64.1% and 77.7% were obtained for the RCS and decomposition analyses respectively.

A simple method is described to estimate forest biomass by high-resolution polarimetric synthetic aperture radar (SAR). The method is based on the regression analysis between the measured biomass from the ground survey and the second intensity moment of the non-Gaussian texture in the cross-polarized L-band SAR images. The SAR data used in the analysis were acquired by the airborne polarimetric interferometric SAR over the coniferous forest in Hokkaido, Japan. The regression analysis was first carried out, and a model function was derived to relate the intensity moment and the measured biomass in 19 forest stands. Using this model function, the biomass values were estimated and compared with those of 21 different stands with known biomass. The average accuracy of the moment model was found to be 85%, which is similar to that of the previous K -distribution model. The advantage of this method over the distribution-based model is that there is no need to search a specific distribution function which fits best to the image texture.

In our previous regression model for estimating forest biomass, it was shown that non-Gaussian amplitude fluctuations in high-resolution polarimetric synthetic aperture radar (SAR) data of coniferous forests can accurately be described by the -distribution and that the order parameter of the -distribution can be useful in estimating the tree biomass of coniferous forests from L-band cross-polarization amplitude images in a wider range than the conventional method using the radar cross section alone. The result was based on the analysis of the ldquoground-truthrdquo biomass data of 19 forest stands and airborne polarimetric interferometric SAR L-band data over the Tomakomai forests in Hokkaido, Japan. From this relation, an empirical regression model was developed to estimate forest biomass from SAR data. In this paper, we report the results on further analyses of this regression model. The validity of the -distribution is first reconfirmed using the Akaike information criterion, followed by the description on the accuracy of the model. To examine model accuracy, we carried out further field measurements on 22 forest stands in 2005, and the ground survey was made in 2006 to find out the causes of several anomalous data. Based on a comparison of the model-based biomass and the ground-truth data, the accuracy of the model was found to be approximately 86%. The regression model was then updated for practical application in estimating the biomass of the Hokkaido forests by including the ground-truth data of all 41 forest stands.

JERS-1/SAR images taken in 1990’s and ALOS/PALSAR images taken in 2000’s are used and examined the deforestation status during ∼14 years. Many trees had been fallen by a typhoon hit in 2004 in a test site, Tomakomai. There is also a plantation area and the site show active change during the term. The forest stands could be roughly classified for four types, forest stands, vacant, where almost all trees were carried out, and two types of transitional stands from a forest to a vacant. Many deforested areas are easily detected by using the difference of backscattering coefficients between two images, if fallen trees have been carried out from the stands. The change from the normal forest to the vacant stands causes 3.1dB decrease in the σ0HH. On the other hands, transitional stands show almost same backscattering as the normal forest stands, although. Threecomponent scattering model shows surface scattering component accounts for 50% over the vacant stands, while volume scattering component accounts for ∼60% over the forest stands. But the model doesn’t show the clear difference between transitional forest site and normal forest. The temporal changes of the forest during 14 years are also examined for the plantation area. One stand show gradual increase of σ0and the values seem to be saturated around 17.4 tons/ha (∼5m in average height).

The purpose of this letter is to present the results on the study of searching effective parameters that describe the relation between high-resolution synthetic aperture radar (SAR) images and forest parameters. The study is based on the non-Gaussian texture analysis of the polarimetric airborne Pi-SAR data over coniferous forests in Hokkaido, Japan. The radar cross section (RCS) in terms of a forest biomass is first analyzed. It is found that the$hboxL$-band RCS increases steadily with the biomass and saturates at approximately 40 tons/ha. These results are similar to the previous studies. The probability density function of the image amplitude is then investigated, and among Rayleigh, log-normal, Weibull, and$K$-distributions, the$K$-distribution is found to fit best to the$hboxL$-band data of all polarizations, although the Weibull distribution fits equally well. Further, the correlation between the tree biomass and the order parameter of the$K$-distribution in the cross-polarization images is found to be very high, and the order parameter increases consistently with the biomass to approximately 100 tons/ha, which is well beyond the saturation limit of the$hboxL$-band RCS. Thus, the order parameter of the$K$-distribution can be a promising new parameter to estimate the forest biomass from high-resolution polarimetric SAR data in a much wider range than the conventional RCS method.

A general expression is derived and discussed for the interlook cross-correlation function (CCF) of speckle patterns in synthetic aperture radar (SAR) multilook images of dynamic ocean surfaces processed with partially overlapped sub-reference signals at arbitrary Doppler center times. Provided that the backscattered field is statistically “white”, the CCF of the speckle intensity is given by the squared modulus of the autocorrelation function of the amplitude weighting functions of sub-reference signal where the time lag is the center time difference. For such Gaussian speckle patterns, the CCF of the interlook speckle patterns is independent of the surface coherence time of randomly moving dynamic sea surfaces. From the integral equation, an analytic expression is derived for a rectangular sub-reference weighting. Comparisons of the theoretical CCFs with those estimated from the RADARSAT-1 C-band and JERS-1 L-band images of the coastal waters of Japan show good agreement.

In the present paper, a general integral expression is derived and discussed for the cross-correlation function (CCF) of speckle patterns in synthetic aperture radar (SAR) images processed by using partially overlapped subapertures of arbitrary Doppler center frequencies (or equivalent azimuth times). It is shown that, under the white noise approximation for the backscattered field, the CCF of the interlook speckle intensity patterns is given by the squared modulus of the autocorrelation function of the amplitude weighting function of subapertures where the time lag is the center time difference. It is also shown that the CCF of the interlook speckle patterns is independent of the surface coherence time of sea surface. The integral expression for the intensity CCF is then evaluated for a rectangular weighting function, and comparison is made with Japanese Earth Resources Satellite-1 (JERS-1) L-band and RADARSAT-1 C-band SAR images of sea surface to test the theory. The CCFs computed from the JERS-1 SAR data show excellent agreement with the theory, and good agreement is obtained with the RADARSAT-1 data.