Vision-Language Models (VLMs) have demonstrated great potential in interpreting remote sensing (RS) images through language-guided semantic. However, the effectiveness of these VLMs critically depends on high-quality image-text training data that captures rich semantic relationships between visual content and language descriptions. Unlike natural images, RS lacks large-scale interleaved image-text pairs from web data, making data collection challenging. While current approaches rely primarily on rule-based methods or flagship VLMs for data synthesis, a systematic framework for automated quality assessment of such synthetically generated RS vision-language data is notably absent. To fill this gap, we propose a novel score model trained on large-scale RS vision-language preference data for automated quality assessment. Our empirical results demonstrate that fine-tuning CLIP or advanced VLMs (e.g., Qwen2-VL) with the top 30% of data ranked by our score model achieves superior accuracy compared to both full-data fine-tuning and CLIP-score-based ranking approaches. Furthermore, we demonstrate applications of our scoring model for reinforcement learning (RL) training and best-of-N (BoN) test-time scaling, enabling significant improvements in VLM performance for RS tasks. Our code, model, and dataset are publicly available.

The absence of temporal information in single-temporal satellite remote sensing images presents a substantial challenge for target state discrimination. In this letter, a pioneering Remote Sensing Airplane State Discrimination Network (RSASDNet) is introduced, by leveraging the relationship between targets and their backgrounds in single-temporal high-resolution remote sensing images. To facilitate the study, we take airplane state discrimination as an example, and a Remote Sensing Airport Panoptic Segmentation with Airplane States Dataset (RSAPS-ASD) is constructed. RSASDNet incorporates two key innovations: 1) a scene knowledge graph generation module that constructs scene knowledge representation by capturing spatial relationships between airplane instances and their surrounding environment (e.g., taxiways and hangars); and 2) a novel graph-image hybrid convolution discrimination module that synergistically integrates structural knowledge and spatial semantic information through dedicated dual-branch learning. The effectiveness of the proposed method is validated using RSAPS-ASD, with experimental results demonstrating that RSASDNet achieves an impressive accuracy of 73.95% in airplane state discrimination.

In the context of global climate change and frequent extreme weather events, forecasting future geospatial vegetation states under these conditions is of significant importance. The vegetation change process is influenced by the complex interplay between dynamic meteorological variables and static environmental variables, leading to high levels of uncertainty. Existing deterministic methods are inadequate in addressing this uncertainty and fail to accurately model the impact of these variables on vegetation, resulting in blurry and inaccurate forecasting results. To address these issues, VegeDiff is proposed for the geospatial vegetation forecasting task. To our best knowledge, VegeDiff is the first to employ a diffusion model to probabilistically capture the uncertainties in vegetation change processes, enabling the generation of clear and accurate future vegetation states. VegeDiff also separately models the global impact of dynamic meteorological variables and the local effects of static environmental variables, thus accurately modeling the impact of these variables. Extensive experiments on geospatial vegetation forecasting tasks demonstrate the effectiveness of VegeDiff. By capturing the uncertainties in vegetation changes and modeling the complex influence of relevant variables, VegeDiff outperforms existing deterministic methods, providing clear and accurate forecasting results of future vegetation states. Interestingly, this study demonstrate the potential of VegeDiff in applications of forecasting future vegetation states from multiple aspects and exploring the impact of meteorological variables on vegetation dynamics. The code of this work will be available at https://github.com/walking-shadow/Official VegeDiff.

It is time-consuming to collect a huge number of pixel-level annotations for accurately extracting buildings by deep neural networks. Supported by class activation map (CAM), weakly supervised semantic segmentation (WSSS) methods with image-level annotations serve as an efficient solution for building extraction. However, it is a great challenge to generate highquality CAM heatmaps for buildings from high-resolution remote sensing images. On one hand, image-level labels lack spatial information, resulting in partial integrity and hollow phenomenon for building extraction. On the other hand, complex backgrounds in remote sensing images can lead to inaccurate extraction of building boundaries. In this study, we propose a novel weakly supervised building extraction method called FlipCAM to deal with these challenges. The Flip module based on feature-level flipping augmentation is designed to improve the integrity of CAM heatmaps by fusing the original and flipped feature maps. In addition, by combining Flip module with slice and merge (SAM) module based on consistency architecture, FlipCAM is able to generate high-quality CAM heatmaps with both boundary fineness and internal integrity in an end-to-end manner, which also alleviates special difficulties for building extraction, including adhesions in dense buildings and confusions with background and shadows, providing reliable pixel-level pseudo masks for training segmentation network to extract buildings. Extensive experiments on three high-resolution datasets show that FlipCAM achieves excellent performance and outperforms other weakly supervised methods in terms of effectiveness and robustness capabilities. Our code is public at https://github.com/NJU-LHRS/FlipCAM-master.

Snow-covered forests are widely distributed in middle- and high-latitude regions of the Northern Hemisphere and have significant impacts on global climate change and albedo feedback. However, knowledge of the radiative transfer mechanism in snow–canopy–atmosphere systems is insufficient. Existing bidirectional reflectance models often oversimplify, assuming that snow only persists on the floor, and neglect the interactions between the atmosphere and snow-covered forests. In our previous snow-covered forest bidirectional reflectance (SFBR) model, we considered ground snow, the concentration of soot pollution, needle leaf characteristics, and discontinuous canopy distributions. Furthermore, this study proposed an improved snow-covered forest bidirectional reflectance (SFBR2) model by considering canopy-intercepted snow (CIS) and atmospheric effects. Specifically, the SFBR2 model was constructed by a series of analytical models for CIS, ground snow (asymptotic radiative transfer (ART) snow model), soil [brightness shape moisture (BSM)], needle leaf [leaf incorporating biochemistry exhibiting reflectance and transmittance yields (LIBERTY)], canopy [two-layer version of four-stream scattering by arbitrarily inclined leaves (4SAIL2)], and atmosphere [simplified method for atmospheric correction (SMAC)], in which the CIS is parameterized by snow optical properties and two-stream theory, and the interaction between the atmosphere and snow-covered forests is optimized by a four-stream theory. It makes the model able to simulate the reflectance at the top of the canopy/atmosphere (TOC/TOA). Validations against 3-D large-scale remote sensing data and image simulation framework over heterogeneous 3-D scenes (LESS) model, unmanned aerial vehicle (UAV) observations, and MODIS data indicated a good consistency with SFBR2 model in the reflectance simulation. The established model has the ability to simulate the bidirectional reflectance of different snow-covered forest scenes whether CIS exists or not. Potential applications include satellite signal simulation, radiation mechanism analysis, and parameter inversion in snow-covered forests.

Context modeling is critical for remote sensing image dense prediction tasks. Nowadays, the growing size of very-high-resolution (VHR) remote sensing images poses challenges in effectively modeling context. While transformer-based models possess global modeling capabilities, they encounter computational challenges when applied to large VHR images due to their quadratic complexity. The conventional practice of cropping large images into smaller patches results in a notable loss of contextual information. To address these issues, we propose the remote sensing Mamba (RSM) for dense prediction tasks in large VHR remote sensing images. RSM is specifically designed to capture the global context of remote sensing images with linear complexity, facilitating the effective processing of large VHR images. Considering that the land covers in remote sensing images are distributed in arbitrary spatial directions due to characteristics of remote sensing over-head imaging, the RSM incorporates an omnidirectional selective scan module (OSSM) to globally model the context of images in multiple directions, capturing large spatial features from various directions. We designed simple yet effective models based on RSM, achieving state-of-the-art performance on dense prediction tasks in VHR remote sensing images without fancy training strategies. Extensive experiments on semantic segmentation (SS) and change detection (CD) tasks across various land covers demonstrate the effectiveness of the proposed RSM. Leveraging the linear complexity and global modeling capabilities, RSM achieves better efficiency and accuracy than transformer-based models on large remote sensing images. Interestingly, we also demonstrated that our model generally performs better with a larger image size on dense prediction tasks.

Although introducing self-attention mechanisms is beneficial to establish long-range dependencies and explore global context information in the task of remote sensing image semantic segmentation, it results in expensive computation and large memory cost. In this letter, we address this dilemma by proposing a lightweight dual-range context aggregation network (LDCANet) for efficient remote sensing image semantic segmentation. First, a dual-range context aggregation module (DCAM) is designed to aggregate the local features and the global semantic context acquired by convolutions and self-attention, respectively, where self-attention is implemented easily by applying two cascaded linear layers to reduce the computational complexity. Furthermore, a simple and lightweight decoder is employed to combine information from different levels, in which a multilayer perceptron (MLP)-based efficient linear block (ELB) is proposed to yield a strong and efficient representation. Experiments conducted on the International Society for Photogrammetry and Remote Sensing (ISPRS) Vaihingen dataset and the Gaofen Image dataset (GID) prove that our LDCANet achieves an excellent trade-off between segmentation accuracy and computational efficiency. In particular, our method achieves 74.12% mean intersection over union (mIoU) on the ISPRS Vaihingen dataset and 61.42% mIoU on the GID with only 4.98-M parameter size.

Semantic segmentation of remote sensing images is effective for large-scale land cover mapping, which heavily relies on a large amount of training data with laborious pixel-level labeling. Weakly supervised semantic segmentation (WSSS) based on image-level labels has attracted intensive attention due to its easy availability. However, existing image-level WSSS methods for remote sensing images mainly focus on binary segmentation, which are difficult to apply to multiclass scenarios. This study proposes a comprehensive framework for image-level multiclass WSSS of remote sensing images, consisting of appropriate image-level label generation, high-quality pixel-level pseudo mask generation, and segmentation network iterative training. Specifically, a training sample filtering method, as well as a dataset cooccurrence evaluation metric, is proposed to demonstrate proper image-level training samples. Leveraging multiclass class activation maps, an uncertainty-driven pixel-level weighted mask is proposed to relieve the overfitting of labeling noise in pseudo masks when training the segmentation network. Extensive experiments demonstrate that the proposed framework can achieve high-quality multiclass WSSS performance with image-level labels, which can attain 94.23% and 90.77% of the IoUs from pixel-level labels for the ISPRS Potsdam and Vaihingen datasets, respectively. Beyond that, for the DeepGlobe dataset with more complex landscapes, the WSSS framework can achieve an accuracy close to 99% of the fully supervised case. Additionally, we further demonstrate that compared to adopting multiple binary WSSS models, directly training a multiclass WSSS model can achieve better results, which can provide new thoughts to achieve WSSS of remote sensing images for multiclass application scenarios. Our code is public at https://github.com/NJU-LHRS/OME.

Self-supervised learning (SSL) has gained widespread attention in the remote sensing (RS) and earth observation (EO) communities owing to its ability to learn task-agnostic representations without human-annotated labels. Nevertheless, most existing RS SSL methods are limited to learning either global semantic separable or local spatial perceptible representations. We argue that this learning strategy is suboptimal in the realm of RS, since the required representations for different RS downstream tasks are often varied and complex. In this study, we proposed a unified SSL framework that is better suited for RS images representation learning. The proposed SSL framework, Contrastive Mask Image Distillation (CMID), is capable of learning representations with both global semantic separability and local spatial perceptibility by combining contrastive learning (CL) with masked image modeling (MIM) in a self-distillation way. Furthermore, our CMID learning framework is architecture-agnostic, which is compatible with both convolutional neural networks (CNN) and vision transformers (ViT), allowing CMID to be easily adapted to a variety of deep learning (DL) applications for RS understanding. Comprehensive experiments have been carried out on four downstream tasks (i.e. scene classification, semantic segmentation, object-detection, and change detection) and the results show that models pre-trained using CMID achieve better performance than other state-of-the-art SSL methods on multiple downstream tasks. The code and pre-trained models will be made available at https://github.com/NJU-LHRS/official-CMID to facilitate SSL research and speed up the development of RS images DL applications.

Change detection is a critical task in earth observation applications. Recently, deep-learning-based methods have shown promising performance and are quickly adopted in change detection. However, the widely used multiple encoders and single decoder (MESD) as well as dual-encoder–decoder (DED) architectures still struggle to effectively handle change detection well. The former has problems of bitemporal feature interference in the feature-level fusion, while the latter is inapplicable to intraclass change detection (ICCD) and multiview building change detection (MVBCD). To solve these problems, we propose a new strategy with an exchanging DED (EDED) structure for binary change detection with semantic guidance and spatial localization. The proposed strategy solves the problems of bitemporal feature inference in MESD by fusing bitemporal features in the decision level and the inapplicability in DED by determining changed areas using bitemporal semantic features. We build a binary change detection model based on this strategy and then validate and compare it with 18 state-of-the-art change detection methods on six datasets in three scenarios, including ICCD datasets (CDD and SYSU), single-view building change detection (SVBCD) datasets (WHU, LEVIR-CD, and LEVIR-CD+), and an MVBCD dataset (NJDS). The experimental results demonstrate that our model achieves superior performance with high efficiency and outperforms all benchmark methods with F1-scores of 97.77%, 83.07%, 94.86%, 92.33%, 91.39%, and 74.35% on CDD, SYSU, WHU, LEVIR-CD, LEVIR-CD+, and NJDS datasets, respectively. The code of this work will be available athttps://github.com/NJU-LHRS/official-SGSLN.

Because of the anisotropy of snow surface reflectance, it is essential to select a proper snow bidirectional reflectance model for extracting snow cover and inverting snow properties from remote sensing image, especially during the period of snow rapidly changed. In this letter, the ability of reproducing snow bidirectional reflectance by three semiempirical bidirectional reflectance distribution function (BRDF) models (Ross–Li, Roujean, and Raman–Pinty–Verstraete) and the asymptotic radiative transfer theory (ART) model was evaluated using the polarization and directionality of the earth reflectance (POLDER) data. In addition, the ART model was compared with the bicontinuous geometric optics (bic-GORT) model based on field measurements. The results indicated that the root mean square errors (RMSEs) are small and similar for all models during the stable-snow period. The physical models perform better than the semiempirical models in capturing the bidirectional signatures of snow during the periods of snow rapidly changed. The bic-GORT model achieves higher accuracy, while the ART model holds the advantages of simple and efficient to be used.

It is promising to improve MODIS snow cover mapping by training effective machine learning models. However, considering the strong spatiotemporal heterogeneity of snow cover, the representativeness of training samples becomes a critical issue to ensure the model effectiveness over large areas and long terms. To deal with this issue, we propose them-day dynamic training strategy. The core of this strategy is to divide a long-term snow cover mapping task into multiple short-term tasks with consecutivemdays (mis greater or equal to the number of days that higher-resolution remote sensing images used as equivalent ground references cover the study area once). It can ensure the spatial representativeness of training samples by utilizing all available equivalent ground references data, and reduce the issue caused by snow cover changing over time. We apply this strategy to random forest for both binary (BSC) and fractional (FSC) snow cover mapping using Landsat-8 data as equivalent ground reference and validate its effect on the Tibetan Plateau. Results show that this strategy achieves higher accuracies than other training strategies, withF1of 0.9596 for BSC mapping,RandMSEof 0.9037 and 0.0260 for FSC mapping, respectively. In addition, spatiotemporal analysis of results further demonstrates that this strategy holds advantages for snow cover mapping in areas of complex terrain and in periods when snow cover changes rapidly. In conclusion, this strategy can effectively ensure the representativeness of training samples generated from Landsat-8 data and thus improve MODIS snow cover mapping through machine learning.

Learning effective visual representations without human supervision is a critical problem for the task of semantic segmentation of remote sensing images (RSIs), where pixel-level annotations are difficult to obtain. Self-supervised learning (SSL), which learns useful representations by creating artificial supervised learning problems, has recently emerged as an effective method to learn from unlabelled data. Current SSL methods are generally trained on ImageNet through image-level prediction tasks. We argue that this is suboptimal for application in semantic segmentation of RSIs since it does not take into account spatial position information between objects, which is critical for segmentation of RSIs characterized by multi-object. In this study, we propose a novel self-supervised dense representation learning method, IndexNet, for the semantic segmentation of RSIs. On the one hand, considering the multi-object characteristics of RSIs, IndexNet learns pixel-level representations by tracking object positions while maintaining sensitivity to object position changes to ensure that no mismatches are caused. On the other hand, by combining image-level contrast and pixel-level contrast, IndexNet can learn spatiotemporal invariant features. Experimental results show that our method works better than ImageNet pre-training and outperforms state-of-the-art (SOTA) self-supervised learning methods. Code and pre-trained models will be available at: https://github.com/pUmpKin-Co/offical-IndexNet.

Snow in mountainous areas is a major source of surface water and groundwater recharge in the world. The water balance in mountainous regions is controlled by the interactions between the climate, cryospheric, and hydrological systems. Surface snow grain is a sensitive thermodynamic indicator of snowpack and plays an important role in the snow albedo. In mountainous regions, the complicated terrain conditions may introduce errors in the snow grain size estimated from satellite imagery. In this letter, an effective method based on the Snow Grain-Size and Pollution (SGSP) amount algorithm is proposed to estimate the surface snow grain size with careful topographic correction, using spectral reflectance data in Channel 5 (1.24 μm) of the Moderate Resolution Imaging Spectrometer. The SGSP-estimated snow grain size was validated with in situ measurements collected from field campaigns in mountainous areas of Manasi River Basin, China, during the periods of snow accumulation and ablation from 2011 to 2015. The R2value of 0.90 and root mean squared error of 80.42 μm were obtained.

High-resolution remote sensing (HR) images provide geometric details of land surface, allowing detailed comparison of geographic objects for change detection. Object-based change detection (OBCD) has become the principle approach to detect changes from HR images. Within OBCD, the separate segmentation strategy has the potential of revealing specific object-to-object changes by providing distinct multi-temporal segments for comparison. Three separate segmentation strategies, named SIISeg, SAISeg, and SAOSeg, are compared to show the importance of imbedding association cues in separate segmentation procedure for successive spatial correspondence establishment to reveal the geometric changes. Buildings in urban area are used as examples to show the change detection results and a pair of aerial images is used in the experiment to show the comparison result.

This paper presents a cosegmentation-based method for building change detection from multitemporal high-resolution (HR) remotely sensed images, providing a new solution to object-based change detection (OBCD). First, the magnitude of a difference image is calculated to represent the change feature. Next, cosegmentation is performed via graph-based energy minimization by combining the change feature with image features at each phase, directly resulting in foreground as multitemporal changed objects and background as unchanged area. Finally, the spatial correspondence between changed objects is established through overlay analysis. Cosegmentation provides a separate and associated, rather than a separate and independent, multitemporal image segmentation method for OBCD, which has two advantages: 1) both the image and change features are used to produce foreground segments as changed objects, which can take full advantage of multitemporal information and produce two spatially corresponded change detection maps by the association of the change feature, having the ability to reveal the thematic, geometric, and numeric changes of objects and 2) the background in the cosegmentation result represents the unchanged area, which naturally avoids the problem of matching inconsistent unchanged objects caused by the separate and independent multitemporal segmentation strategy. Experimental results on five HR datasets verify the effectiveness of the proposed method and the comparisons with the state-of-the-art OBCD methods further show its superiority.

Snow cover in cold and arid regions is a key factor controlling regional energy balances, hydrological cycle, and water utilization. Interferometric synthetic aperture radar (InSAR) technology offers the ability to monitor snow cover in all weather. In this letter, a support vector machine (SVM) method for extracting snow cover based on SAR and optical data in rugged mountain terrain is introduced. In this method, RadarSat-2 InSAR interferometric coherence images are analyzed, adopting snow-covered and snow-free areas obtained from GF-1 satellite observations as the “ground truth.” The analysis results indicate that the coherence in copolarizations is clearly correlated with the underlying surface type and local incidence angle. These two factors, combined with training samples from GF-1 wide field viewer data, were used to build an SVM to classify coherence images in HH polarization. The classification results demonstrate that snow cover extraction using this method can achieve mean accuracies of 83.8% and 77.5% in areas with low and high vegetation coverage, respectively. These accuracies are significantly higher than those achieved by the typical thresholding algorithm (72.7% and 69.2%, respectively).

A novel spectral similarity measure approach, which is named spectral frequency spectrum difference (SFSD), is proposed for hyperspectral image classification based on the frequency spectrum of spectral signature using the Fourier transform. Many important characteristics of spectral signature can be clearly reflected in the frequency spectrum. Therefore, the spectral similarity is defined as the frequency spectrum's difference between the target and reference signatures. The frequency spectrum analysis in this study suggests that the magnitude values of the first few low-frequency components for spectral signature can effectively represent the spectral similarity. To balance the difference between the low- and high-frequency components, the frequency spectrum of the target spectral signature is taken as the normalized factor in the SFSD method. Next, the U.S. Geological Survey spectral data and two hyperspectral remote sensing images were employed as test data in our validation experiments. The new SFSD proposed here was compared with the leading approaches in terms of the spectral discriminability and classification accuracy. Results show that the SFSD exhibits a relatively better performance and has more robust applications for hyperspectral image classification.

Object-based analysis of high spatial resolution remote sensing images addresses the matter of multiscale segmentation. However, existing segmentation evaluation methods mainly focus on single-scale segmentation. In this paper, we examine the issue of supervised multiscale segmentation evaluation and propose two discrepancy measures to determine the manner in which geographic objects are delineated by multiscale segmentations. A QuickBird scene in Hangzhou, China, is used to conduct the evaluation. The results reveal the effectiveness of the proposed measures, in terms of method comparison and parameter optimization, for multiscale segmentation of high spatial resolution images. Moreover, meaningful indications for selecting suitable multiple segmentation scales are presented. The proposed measures are applicable to performance evaluation and parameter optimization for multiscale segmentation algorithms.

Image segmentation is a critical step in the analysis of high-spatial-resolution remotely sensed imagery using object-based image analysis. The segmentation quality is extremely important to the subsequent analysis. This letter proposes an improved unsupervised method to evaluate the segmentation quality for remotely sensed imagery. The evaluation criteria take into account global intrasegment homogeneity and intersegment heterogeneity measures, which can be useful for the comparison of segmentation results produced by a single segmentation method. The proposed method is compared with other two mature unsupervised evaluation methods on two segmentation methods: region growing and mean shift. QuickBird images are used for the comparative study. The effectiveness of the proposed method is validated through comparing with the supervised evaluation method Rand Index and visual analysis.