Earth observation satellites are crucial in regional-scale monitoring, but the effectiveness of optical satellite is often severely restricted by cloud cover. Most existing satellite scheduling methods at regional-scale ignore the impact of clouds, resulting in a waste of observation resources and degraded data quality. To address this problem, this study proposes a regional-scale multisatellite scheduling framework that considers a historical cloud cover for the first time. The core of the framework is a novel pixel-region integrated method, which quantifies the functional relationship between observation success probability (OSP) and observation success fraction (OSF) using historical cloud products (MOD35) and satellite overpass information. We construct an optimization model to maximize OSF under a user-preset OSP (e.g., 95%) and use a genetic algorithm to solve it. Through experiments in three different climate and terrain regions in China, we compare the proposed method with two baseline strategies, namely, “nadir observation” and “maximum coverage”. The results show that the proposed method can not only achieve 100% geometric coverage, but also far exceeds the baseline strategy in terms of effectiveness in considering cloud effects. For example, at 95% OSP confidence level, the OSFs obtained by this method for the three regions are on average 14% higher than those of the maximum coverage strategy. In addition, through Monte Carlo simulation verification, the central limit theorem approximation method we rely on improves computational efficiency by hundreds of times while ensuring accuracy. This framework can provide mission planners with decision-making solutions with clear probabilistic guarantees derived from long-term historical cloud statistics, significantly improving the efficiency of satellite resource utilization and the ability to obtain effective data under the influence of cloud uncertainty.
{"title":"Multisatellite Scheduling Method Based on a Historical Cloud Coverage at Regional Scale","authors":"Rongguang Ni;Guangjian Yan;Xihan Mu;Tian Xie;Wenhao Jiang;Si Gao;Donghui Xie","doi":"10.1109/JSTARS.2026.3665349","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3665349","url":null,"abstract":"Earth observation satellites are crucial in regional-scale monitoring, but the effectiveness of optical satellite is often severely restricted by cloud cover. Most existing satellite scheduling methods at regional-scale ignore the impact of clouds, resulting in a waste of observation resources and degraded data quality. To address this problem, this study proposes a regional-scale multisatellite scheduling framework that considers a historical cloud cover for the first time. The core of the framework is a novel pixel-region integrated method, which quantifies the functional relationship between observation success probability (OSP) and observation success fraction (OSF) using historical cloud products (MOD35) and satellite overpass information. We construct an optimization model to maximize OSF under a user-preset OSP (e.g., 95%) and use a genetic algorithm to solve it. Through experiments in three different climate and terrain regions in China, we compare the proposed method with two baseline strategies, namely, “nadir observation” and “maximum coverage”. The results show that the proposed method can not only achieve 100% geometric coverage, but also far exceeds the baseline strategy in terms of effectiveness in considering cloud effects. For example, at 95% OSP confidence level, the OSFs obtained by this method for the three regions are on average 14% higher than those of the maximum coverage strategy. In addition, through Monte Carlo simulation verification, the central limit theorem approximation method we rely on improves computational efficiency by hundreds of times while ensuring accuracy. This framework can provide mission planners with decision-making solutions with clear probabilistic guarantees derived from long-term historical cloud statistics, significantly improving the efficiency of satellite resource utilization and the ability to obtain effective data under the influence of cloud uncertainty.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"8359-8371"},"PeriodicalIF":5.3,"publicationDate":"2026-02-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11397348","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440502","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-16DOI: 10.1109/JSTARS.2026.3665394
Morteza Rahimpour;Taha B. M. J. Ouarda
Precipitable water vapor (PWV) is a critical variable for monitoring extreme weather events, particularly precipitation. Accurate retrieval of PWV is therefore essential in hydrology and meteorology. This study proposes an applied machine learning framework, termed PWV-ML, to enhance PWV estimation. Three machine learning methods, artificial neural networks, convolutional neural networks, and random forest (RF), were evaluated using data from Radiosonde and GNSS observation stations between 2019 and 2023 across the United States and Canada. The models incorporated multiple auxiliary variables derived from AMSR2 brightness temperature data, MODIS products (MOD21A1N, MOD13A2), and seven ERA5 reanalysis variables. Results show that PWV-ML substantially outperforms traditional multiple linear regression. Among the tested methods, RF achieved slightly superior accuracy. Variable selection proved critical, with dewpoint temperature (Td) and total column water vapor (TCWV) from ERA5 providing the strongest contributions. An optimized minimal input configuration, consisting of BT89V, cloud liquid water, the microwave atmospheric water vapor index, altitude, and Td, yielded strong agreement with ground observations (R = 0.92; RMSE = 4.77 mm; KGE = 0.87 for Radiosonde, and R = 0.91; RMSE = 2.82 mm; KGE = 0.76 for GNSS). Incorporating additional ERA5 variables further improved performance, achieving near-perfect correlations (R = 0.99). These findings demonstrate that the PWV-ML framework can retrieve PWV with high accuracy under diverse climatic conditions.
{"title":"Enhanced Retrieval of Precipitable Water Vapor Over Land: An Approach for Algorithmic Improvement and Performance Assessment","authors":"Morteza Rahimpour;Taha B. M. J. Ouarda","doi":"10.1109/JSTARS.2026.3665394","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3665394","url":null,"abstract":"Precipitable water vapor (PWV) is a critical variable for monitoring extreme weather events, particularly precipitation. Accurate retrieval of PWV is therefore essential in hydrology and meteorology. This study proposes an applied machine learning framework, termed PWV-ML, to enhance PWV estimation. Three machine learning methods, artificial neural networks, convolutional neural networks, and random forest (RF), were evaluated using data from Radiosonde and GNSS observation stations between 2019 and 2023 across the United States and Canada. The models incorporated multiple auxiliary variables derived from AMSR2 brightness temperature data, MODIS products (MOD21A1N, MOD13A2), and seven ERA5 reanalysis variables. Results show that PWV-ML substantially outperforms traditional multiple linear regression. Among the tested methods, RF achieved slightly superior accuracy. Variable selection proved critical, with dewpoint temperature (T<sub>d</sub>) and total column water vapor (TCWV) from ERA5 providing the strongest contributions. An optimized minimal input configuration, consisting of BT89V, cloud liquid water, the microwave atmospheric water vapor index, altitude, and T<sub>d</sub>, yielded strong agreement with ground observations (R = 0.92; RMSE = 4.77 mm; KGE = 0.87 for Radiosonde, and R = 0.91; RMSE = 2.82 mm; KGE = 0.76 for GNSS). Incorporating additional ERA5 variables further improved performance, achieving near-perfect correlations (R = 0.99). These findings demonstrate that the PWV-ML framework can retrieve PWV with high accuracy under diverse climatic conditions.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"8396-8409"},"PeriodicalIF":5.3,"publicationDate":"2026-02-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11397434","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440542","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Accurate and timely mapping of maize distribution is crucial for national food security and sustainable agricultural development. However, maize classification in diverse agricultural landscapes is often hindered by subjective thresholds and limited feature representativeness in time-series remote sensing data. To overcome these limitations, we introduce an optimized feature selection time-weighted dynamic time warping (OFS-TWDTW) method that integrates multidimensional feature selection from Sentinel-2 imagery. Our approach begins by constructing robust standard maize curves through phenological and morphological sample screening to ensure sample reliability. We then apply the Relief algorithm to evaluate feature importance, followed by separability analysis and autocorrelation removal to select an optimal set of discriminative phenological features, enhancing classification efficiency. Finally, we employ TWDTW with an adaptive minimum distance classification to eliminate reliance on subjective thresholds. By conducting extensive evaluation across three climatically diverse regions in China (Dezhou, Pingliang, and Nenjiang), OFS-TWDTW achieved overall accuracies of 95.36%, 93.62%, and 90.59%, respectively. Notably, it demonstrated superior robustness over the traditional NDVI-based baseline, particularly in resolving spectral confusion between maize and soybean in complex landscapes. This method reduces misclassification and omission errors, offering a scalable, high-accuracy solution for large-scale crop mapping with broader applicability to other crops.
{"title":"Improved Maize Mapping Through Optimizing Spatio-Temporal Feature Selection","authors":"Shuangxi Miao;Yuhan Jiang;Jing Yao;Fuqiang Shen;Zhewei Zhang;Zhongxiang Xie;Xuecao Li;Huiying Li;Jianxi Huang","doi":"10.1109/JSTARS.2026.3664709","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3664709","url":null,"abstract":"Accurate and timely mapping of maize distribution is crucial for national food security and sustainable agricultural development. However, maize classification in diverse agricultural landscapes is often hindered by subjective thresholds and limited feature representativeness in time-series remote sensing data. To overcome these limitations, we introduce an optimized feature selection time-weighted dynamic time warping (OFS-TWDTW) method that integrates multidimensional feature selection from Sentinel-2 imagery. Our approach begins by constructing robust standard maize curves through phenological and morphological sample screening to ensure sample reliability. We then apply the Relief algorithm to evaluate feature importance, followed by separability analysis and autocorrelation removal to select an optimal set of discriminative phenological features, enhancing classification efficiency. Finally, we employ TWDTW with an adaptive minimum distance classification to eliminate reliance on subjective thresholds. By conducting extensive evaluation across three climatically diverse regions in China (Dezhou, Pingliang, and Nenjiang), OFS-TWDTW achieved overall accuracies of 95.36%, 93.62%, and 90.59%, respectively. Notably, it demonstrated superior robustness over the traditional NDVI-based baseline, particularly in resolving spectral confusion between maize and soybean in complex landscapes. This method reduces misclassification and omission errors, offering a scalable, high-accuracy solution for large-scale crop mapping with broader applicability to other crops.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"8031-8043"},"PeriodicalIF":5.3,"publicationDate":"2026-02-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11395563","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147362386","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-12DOI: 10.1109/JSTARS.2026.3664318
Imen Ziadi;Nejla Essaddi;Mongi Besbes
Predicting whether an earthquake will generate a tsunami is critical for early warning systems and disaster mitigation. In this study, we present an AI-driven approach to classify earthquakes as tsunami-generating or nontsunami events. We utilize three machine learning models—random forest, support vector machine, and logistic regression—trained on USGS earthquake data from 2015 to 2025, considering features, such as magnitude, depth, latitude, and longitude. Our exploratory data analysis highlights key correlations and feature distributions, while model evaluation demonstrates high classification performance, with random forest achieving up to 91% accuracy. We further investigate feature importance and provide ROC curves and confusion matrices for comparative analysis. The results show that AI-driven classification can effectively support early warning systems, offering a scalable and data-informed tool for seismic hazard assessment.
{"title":"AI-Driven Classification of Tsunami-Generating Earthquakes: Harnessing Random Forest, SVM, and Logistic Regression for Early Detection","authors":"Imen Ziadi;Nejla Essaddi;Mongi Besbes","doi":"10.1109/JSTARS.2026.3664318","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3664318","url":null,"abstract":"Predicting whether an earthquake will generate a tsunami is critical for early warning systems and disaster mitigation. In this study, we present an AI-driven approach to classify earthquakes as tsunami-generating or nontsunami events. We utilize three machine learning models—random forest, support vector machine, and logistic regression—trained on USGS earthquake data from 2015 to 2025, considering features, such as magnitude, depth, latitude, and longitude. Our exploratory data analysis highlights key correlations and feature distributions, while model evaluation demonstrates high classification performance, with random forest achieving up to 91% accuracy. We further investigate feature importance and provide ROC curves and confusion matrices for comparative analysis. The results show that AI-driven classification can effectively support early warning systems, offering a scalable and data-informed tool for seismic hazard assessment.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"8441-8447"},"PeriodicalIF":5.3,"publicationDate":"2026-02-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11395265","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147440489","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Mangrove ecosystems act as highly efficient carbon sinks and provide critical ecological services for climate regulation; however, they are increasingly threatened by anthropogenic pressures and climate change. Recent advances in annual satellite-based Earth Observation embedding (EM) datasets have enabled the integration of multisource data fusion, improving land use and land cover analyses. Nevertheless, EM datasets have rarely been applied for the monitoring of mangrove ecosystems. Within this context, the present study compares the performance of EM with combined Sentinel-1, Sentinel-2, and Landsat 8/9 datasets (S2S1L8/9) through the integration of a pretrained ResNet152–U-Net architecture and random forest (RF) modeling to assess spatiotemporal patterns of mangrove intactness and degradation from 2017 to 2024. A ResNet152 encoder pretrained on ImageNet was employed to train a U-Net model using the Global Mangrove Dataset (2020) for the mapping of potential mangrove extent. The EM dataset outperformed the S2S1L8/9 combination, achieving validation Intersection over Union scores of 0.85 and 0.84, and Dice coefficients of 0.89 and 0.88, respectively. Spatial comparison further indicated that EM yielded a lower root mean square error (6.51 ha) compared to S2S1L8/9 (7.27 ha). The RF model, trained on intact and nonintact mangrove samples across multiple years, confirmed the superior performance of EM in delineating intact mangrove areas along coastlines, with an overall accuracy of 0.97, an F1-score of 0.97, and a Matthews Correlation Coefficient of 0.93. Degraded mangrove areas were identified by masking intact regions, and gain–loss analysis revealed a decline of approximately 3% in 2021 and 2022 relative to the 2017 baseline. These findings demonstrate that EM provides a more accurate and spatially consistent approach than conventional multisource datasets for mapping mangrove intactness and degradation. By minimizing classification errors and improving coastline delineation, EM establishes a robust framework for large-scale monitoring of mangrove dynamics, supporting conservation planning, carbon accounting, and climate resilience strategies across diverse coastal and terrestrial systems at global scales.
{"title":"A Scalable Potential of Alpha Earth and Multisensor Datasets for Assessing Mangrove Intactness and Degradation Using Deep and Machine Learning Algorithms","authors":"Akkarapon Chaiyana;Filippo Sarvia;Narissara Nuthammachot;Jaturong Som-ard","doi":"10.1109/JSTARS.2026.3664013","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3664013","url":null,"abstract":"Mangrove ecosystems act as highly efficient carbon sinks and provide critical ecological services for climate regulation; however, they are increasingly threatened by anthropogenic pressures and climate change. Recent advances in annual satellite-based Earth Observation embedding (EM) datasets have enabled the integration of multisource data fusion, improving land use and land cover analyses. Nevertheless, EM datasets have rarely been applied for the monitoring of mangrove ecosystems. Within this context, the present study compares the performance of EM with combined Sentinel-1, Sentinel-2, and Landsat 8/9 datasets (S2S1L8/9) through the integration of a pretrained ResNet152–U-Net architecture and random forest (RF) modeling to assess spatiotemporal patterns of mangrove intactness and degradation from 2017 to 2024. A ResNet152 encoder pretrained on ImageNet was employed to train a U-Net model using the Global Mangrove Dataset (2020) for the mapping of potential mangrove extent. The EM dataset outperformed the S2S1L8/9 combination, achieving validation Intersection over Union scores of 0.85 and 0.84, and Dice coefficients of 0.89 and 0.88, respectively. Spatial comparison further indicated that EM yielded a lower root mean square error (6.51 ha) compared to S2S1L8/9 (7.27 ha). The RF model, trained on intact and nonintact mangrove samples across multiple years, confirmed the superior performance of EM in delineating intact mangrove areas along coastlines, with an overall accuracy of 0.97, an F1-score of 0.97, and a Matthews Correlation Coefficient of 0.93. Degraded mangrove areas were identified by masking intact regions, and gain–loss analysis revealed a decline of approximately 3% in 2021 and 2022 relative to the 2017 baseline. These findings demonstrate that EM provides a more accurate and spatially consistent approach than conventional multisource datasets for mapping mangrove intactness and degradation. By minimizing classification errors and improving coastline delineation, EM establishes a robust framework for large-scale monitoring of mangrove dynamics, supporting conservation planning, carbon accounting, and climate resilience strategies across diverse coastal and terrestrial systems at global scales.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"8346-8358"},"PeriodicalIF":5.3,"publicationDate":"2026-02-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11394719","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147362392","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-12DOI: 10.1109/JSTARS.2026.3664428
Libo Cheng;Wenlin Du;Zhe Li;Xiaoning Jia
Hyperspectral remote sensing images (HSIs) inevitably suffer from multiple sources of noise during acquisition, including Gaussian noise, impulse noise, and stripe noise, which severely degrade the reliability of subsequent interpretation tasks. To address the limitations of existing denoising methods, including their reliance on paired clean reference data and the difficulty in simultaneously preserving global structures and recovering local details, we propose a self-supervised conditional diffusion-based denoising network termed DM-WAN. Specifically, DM-WAN incorporates a detail feature prompting strategy into the diffusion denoising process, in which a time-dependent feature modulation mechanism is introduced to dynamically enhance edge and texture information. Moreover, a multihead wavelet attention mechanism is designed to exploit the band-separation capability of wavelet decomposition, enabling the model to capture global structures and local details across different frequency bands, thereby achieving collaborative multiscale feature fusion and structure-preserving reconstruction. During training, a self-supervised learning paradigm is adopted by constructing pseudo-supervised pairs from degraded images, which eliminates the dependence on paired ground-truth supervision. Experimental results demonstrate that the proposed method exhibits favorable performance in terms of detail preservation, structural consistency, and denoising robustness under various synthetic and real-noise scenarios, validating the effectiveness of DM-WAN for HSI denoising.
{"title":"A Self-Supervised Conditional Diffusion Network With Multihead Wavelet Attention for Remote Sensing Image Denoising","authors":"Libo Cheng;Wenlin Du;Zhe Li;Xiaoning Jia","doi":"10.1109/JSTARS.2026.3664428","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3664428","url":null,"abstract":"Hyperspectral remote sensing images (HSIs) inevitably suffer from multiple sources of noise during acquisition, including Gaussian noise, impulse noise, and stripe noise, which severely degrade the reliability of subsequent interpretation tasks. To address the limitations of existing denoising methods, including their reliance on paired clean reference data and the difficulty in simultaneously preserving global structures and recovering local details, we propose a self-supervised conditional diffusion-based denoising network termed DM-WAN. Specifically, DM-WAN incorporates a detail feature prompting strategy into the diffusion denoising process, in which a time-dependent feature modulation mechanism is introduced to dynamically enhance edge and texture information. Moreover, a multihead wavelet attention mechanism is designed to exploit the band-separation capability of wavelet decomposition, enabling the model to capture global structures and local details across different frequency bands, thereby achieving collaborative multiscale feature fusion and structure-preserving reconstruction. During training, a self-supervised learning paradigm is adopted by constructing pseudo-supervised pairs from degraded images, which eliminates the dependence on paired ground-truth supervision. Experimental results demonstrate that the proposed method exhibits favorable performance in terms of detail preservation, structural consistency, and denoising robustness under various synthetic and real-noise scenarios, validating the effectiveness of DM-WAN for HSI denoising.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"7498-7518"},"PeriodicalIF":5.3,"publicationDate":"2026-02-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11395387","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147299586","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-11DOI: 10.1109/JSTARS.2026.3657404
Xin He;Junjie Li;Tianyu Song;Xiang Chen
Remote sensing images are frequently affected by adverse weather conditions, such as haze and raindrops, which degrade image quality and subsequently impair the performance of downstream vision tasks. Currently, mainstream methods for remote sensing image restoration primarily rely on Convolutional Neural Networks (CNNs) and Transformer architectures. However, CNNs face limitations in handling long-range dependencies, while Transformers are constrained by computational efficiency, making it difficult to strike a balance between performance and efficiency. To address these issues, we propose a hierarchical state-space model for remote sensing image restoration, termed hierarchical RS Mamba (Hi-RSMamba). This model enhances contextual modeling by integrating multiscale representations through both global and local state-space models. Specifically, we introduce a hierarchical state-space model (HSSM) that improves the original state-space module by incorporating hierarchical feature representations while preserving local 2-D dependencies, thus enabling better aggregation of rich local and global information. Furthermore, given the complementary nature of global and local dependencies, we design a gated feedforward network to adaptively assist the broadcaste of these multiscale features, allowing for high-quality image background reconstruction.Extensive experimental results demonstrate that Hi-RSMamba exhibits significant advantages on widely used benchmark datasets in remote sensing image restoration tasks.
{"title":"Hi-RSMamba: Hierarchical Mamba for Remote Sensing Image Restoration Under Adverse Weather","authors":"Xin He;Junjie Li;Tianyu Song;Xiang Chen","doi":"10.1109/JSTARS.2026.3657404","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3657404","url":null,"abstract":"Remote sensing images are frequently affected by adverse weather conditions, such as haze and raindrops, which degrade image quality and subsequently impair the performance of downstream vision tasks. Currently, mainstream methods for remote sensing image restoration primarily rely on Convolutional Neural Networks (CNNs) and Transformer architectures. However, CNNs face limitations in handling long-range dependencies, while Transformers are constrained by computational efficiency, making it difficult to strike a balance between performance and efficiency. To address these issues, we propose a hierarchical state-space model for remote sensing image restoration, termed hierarchical RS Mamba (Hi-RSMamba). This model enhances contextual modeling by integrating multiscale representations through both global and local state-space models. Specifically, we introduce a hierarchical state-space model (HSSM) that improves the original state-space module by incorporating hierarchical feature representations while preserving local 2-D dependencies, thus enabling better aggregation of rich local and global information. Furthermore, given the complementary nature of global and local dependencies, we design a gated feedforward network to adaptively assist the broadcaste of these multiscale features, allowing for high-quality image background reconstruction.Extensive experimental results demonstrate that Hi-RSMamba exhibits significant advantages on widely used benchmark datasets in remote sensing image restoration tasks.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"7373-7388"},"PeriodicalIF":5.3,"publicationDate":"2026-02-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11393606","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147299626","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-11DOI: 10.1109/JSTARS.2026.3663742
Dehua Huo;Weida Zhan;Yueyi Han;Jinxin Guo;Depeng Zhu;Yu Chen;YiChun Jiang;Deng Han
State-space models, particularly Mamba, have garnered significant attention from researchers owing to their efficient balance between computational efficiency and model performance, and they also provide a performance breakthrough for accurate land cover classification of hyperspectral images (HSIs). However, HSI data feature high spectral dimensionality and dense spatial information, so efficiently constructing Mamba sequences that align with the structural characteristics of HSI data remains a critical unresolved issue. To address this challenge, we propose an efficient deformable spatial–spectral Mamba network (DSS-Mamba), which aims to efficiently mine valuable spatial–spectral features from HSIs. To achieve this goal, we design a two-branch architecture consisting of a spatial deformable Mamba and a spectral bidirectional Mamba: the spatial deformable scanning Mamba integrates a spatial deformable state-space model and adaptively focuses on salient feature regions through a dynamic scanning strategy, while the spectral bidirectional scanning Mamba incorporates a spectral bidirectional state-space model and fully exploits spectral dimensional information via a bidirectional scanning mechanism. During feature processing, we propose a spatial–spectral complementary fusion module, which refines feature weights by means of a dual-threshold enhancement unit to realize efficient processing and dynamic fusion of spatial–spectral features. Extensive experiments demonstrate that deformable spatialspectral Mamba network (DSS-Mamba) effectively balances the fine-grained capture of spatial–spectral features and computational efficiency through the adaptive capture of spatial features by the deformable structure, the in-depth mining of spectral information via bidirectional scanning, and the complementary fusion of the spatial–spectral module. Consequently, it achieves superior classification accuracy in land cover classification.
{"title":"DSS-Mamba: Deformable Spatial–Spectral State-Space Model for Hyperspectral Land Cover Classification","authors":"Dehua Huo;Weida Zhan;Yueyi Han;Jinxin Guo;Depeng Zhu;Yu Chen;YiChun Jiang;Deng Han","doi":"10.1109/JSTARS.2026.3663742","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3663742","url":null,"abstract":"State-space models, particularly Mamba, have garnered significant attention from researchers owing to their efficient balance between computational efficiency and model performance, and they also provide a performance breakthrough for accurate land cover classification of hyperspectral images (HSIs). However, HSI data feature high spectral dimensionality and dense spatial information, so efficiently constructing Mamba sequences that align with the structural characteristics of HSI data remains a critical unresolved issue. To address this challenge, we propose an efficient deformable spatial–spectral Mamba network (DSS-Mamba), which aims to efficiently mine valuable spatial–spectral features from HSIs. To achieve this goal, we design a two-branch architecture consisting of a spatial deformable Mamba and a spectral bidirectional Mamba: the spatial deformable scanning Mamba integrates a spatial deformable state-space model and adaptively focuses on salient feature regions through a dynamic scanning strategy, while the spectral bidirectional scanning Mamba incorporates a spectral bidirectional state-space model and fully exploits spectral dimensional information via a bidirectional scanning mechanism. During feature processing, we propose a spatial–spectral complementary fusion module, which refines feature weights by means of a dual-threshold enhancement unit to realize efficient processing and dynamic fusion of spatial–spectral features. Extensive experiments demonstrate that deformable spatialspectral Mamba network (DSS-Mamba) effectively balances the fine-grained capture of spatial–spectral features and computational efficiency through the adaptive capture of spatial features by the deformable structure, the in-depth mining of spectral information via bidirectional scanning, and the complementary fusion of the spatial–spectral module. Consequently, it achieves superior classification accuracy in land cover classification.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"7973-7990"},"PeriodicalIF":5.3,"publicationDate":"2026-02-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11391558","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147362389","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-11DOI: 10.1109/JSTARS.2026.3663646
Hengren Tang;Yaxuan Jia;Jiacheng Cheng;Yang Mu;Yi Wu;Xiwen Yao
The core task of semantic segmentation is to assign predefined category labels to each pixel in an image, thereby distinguishing between different objects and backgrounds. Few-shot semantic segmentation (FSS) is a specialized semantic segmentation task that aims to accurately segment pixel-level targets of novel classes in query images, relying only on a limited number of annotated support samples to enable rapid adaptation to unseen categories without extensive labeled data. FSS in remote sensing imagery is a critical yet challenging task, primarily due to two intrinsic data characteristics: extreme scale variations among target objects and significant intraclass heterogeneity. These challenges severely degrade the performance of existing FSS methods, which often rely on single, global prototypes and are not explicitly designed for such variability. To address these limitations, we propose GlocalDualNet, a novel FSS framework tailored for remote sensing applications. GlocalDualNet integrates two core technical contributions. First, a multiscale support prototype extraction module generates a set of heterogeneous local prototypes in addition to a conventional global prototype. This approach mitigates the spatial detail loss associated with global-only representations and provides a more comprehensive feature signature for matching. Second, a dual-branch segmentation network is designed to explicitly disentangle the feature learning process for large- and small-scale targets, thereby improving segmentation accuracy across disparate scales. Experimental validation on the iSAID-5i benchmark dataset demonstrates that our proposed modules yield a notable 2.13% improvement in segmentation accuracy, establishing the efficacy of the GlocalDualNet framework.
{"title":"GlocalDualNet: Disentangling Scale and Representation for Few-Shot Remote Sensing Segmentation","authors":"Hengren Tang;Yaxuan Jia;Jiacheng Cheng;Yang Mu;Yi Wu;Xiwen Yao","doi":"10.1109/JSTARS.2026.3663646","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3663646","url":null,"abstract":"The core task of semantic segmentation is to assign predefined category labels to each pixel in an image, thereby distinguishing between different objects and backgrounds. Few-shot semantic segmentation (FSS) is a specialized semantic segmentation task that aims to accurately segment pixel-level targets of novel classes in query images, relying only on a limited number of annotated support samples to enable rapid adaptation to unseen categories without extensive labeled data. FSS in remote sensing imagery is a critical yet challenging task, primarily due to two intrinsic data characteristics: extreme scale variations among target objects and significant intraclass heterogeneity. These challenges severely degrade the performance of existing FSS methods, which often rely on single, global prototypes and are not explicitly designed for such variability. To address these limitations, we propose GlocalDualNet, a novel FSS framework tailored for remote sensing applications. GlocalDualNet integrates two core technical contributions. First, a multiscale support prototype extraction module generates a set of heterogeneous local prototypes in addition to a conventional global prototype. This approach mitigates the spatial detail loss associated with global-only representations and provides a more comprehensive feature signature for matching. Second, a dual-branch segmentation network is designed to explicitly disentangle the feature learning process for large- and small-scale targets, thereby improving segmentation accuracy across disparate scales. Experimental validation on the iSAID-5<sup>i</sup> benchmark dataset demonstrates that our proposed modules yield a notable 2.13% improvement in segmentation accuracy, establishing the efficacy of the GlocalDualNet framework.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"8018-8030"},"PeriodicalIF":5.3,"publicationDate":"2026-02-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11391662","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147362313","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-10DOI: 10.1109/JSTARS.2026.3663388
Baokai Zu;Zhengrui Yang;Yafang Li;Hongyuan Wang;Jianqiang Li;Ziping He
Hyperspectral image (HSI) classification is challenging due to their high spectral complexity and the need to model long-range dependencies. Although contrastive learning has shown promise, most existing approaches rely on convolutional neural networks or transformers, which are inefficient for modeling long sequences. Recently, the Mamba state space model has emerged as a scalable alternative, offering linear complexity and strong sequence modeling capabilities. However, current HSI classification methods still depend on discrete class labels, which lack rich semantic information and limit the effectiveness of representation learning. To address this limitation, we propose text-guided contrastive mamba (Mamba-TGC), a novel text-guided contrastive learning framework for HSI classification. In Mamba-TGC, each contrastive branch integrates a Mamba encoder that separately models spectral and spatial patch embeddings, which are then fused to capture rich spectral–spatial representations. These features are further aligned with category-level textual descriptions through contrastive loss, enabling the model to learn more discriminative and semantically informative representations. Extensive evaluations of three benchmark HSI datasets demonstrate that Mamba-TGC consistently surpasses existing methods, highlighting the effectiveness of combining Mamba-based spectral–spatial modeling with text-guided supervision for robust hyperspectral representation learning.
{"title":"Text-Guided Contrastive Mamba for Hyperspectral Image Classification","authors":"Baokai Zu;Zhengrui Yang;Yafang Li;Hongyuan Wang;Jianqiang Li;Ziping He","doi":"10.1109/JSTARS.2026.3663388","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3663388","url":null,"abstract":"Hyperspectral image (HSI) classification is challenging due to their high spectral complexity and the need to model long-range dependencies. Although contrastive learning has shown promise, most existing approaches rely on convolutional neural networks or transformers, which are inefficient for modeling long sequences. Recently, the Mamba state space model has emerged as a scalable alternative, offering linear complexity and strong sequence modeling capabilities. However, current HSI classification methods still depend on discrete class labels, which lack rich semantic information and limit the effectiveness of representation learning. To address this limitation, we propose text-guided contrastive mamba (Mamba-TGC), a novel text-guided contrastive learning framework for HSI classification. In Mamba-TGC, each contrastive branch integrates a Mamba encoder that separately models spectral and spatial patch embeddings, which are then fused to capture rich spectral–spatial representations. These features are further aligned with category-level textual descriptions through contrastive loss, enabling the model to learn more discriminative and semantically informative representations. Extensive evaluations of three benchmark HSI datasets demonstrate that Mamba-TGC consistently surpasses existing methods, highlighting the effectiveness of combining Mamba-based spectral–spatial modeling with text-guided supervision for robust hyperspectral representation learning.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"8144-8159"},"PeriodicalIF":5.3,"publicationDate":"2026-02-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11389133","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147362312","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：