Pub Date : 2026-01-12DOI: 10.1109/JSTARS.2026.3651900
Enyu Zhao;Yu Shi;Nianxin Qu;Yulei Wang;Hang Zhao
Infrared small target detection is focused on accurately identifying tiny targets with low signal-to-noise ratio against complex backgrounds, representing a critical challenge in the field of infrared image processing. Existing approaches frequently fail to retain small target information during global semantic extraction and struggle with preserving detailed features and achieving effective feature fusion. To address these limitations, this article proposes a morphology-edge enhanced triple-cascaded network (MEETNet) for infrared small target detection. The network employs a triple-cascaded architecture that maintains high resolution and enhances information interaction between different stages, facilitating effective multilevel feature fusion while safeguarding deep small-target characteristics. MEETNet integrates an edge-detail enhanced module (EDEM) and a detail-aware multi-scale fusion module (DMSFM). These modules introduce edge-detail enhanced features that amalgamate contrast and edge information, thereby amplifying target saliency and improving edge representation. Specifically, EDEM augments target contrast and edge structures by integrating edge-detail-enhanced features with shallow details. This integration improves the discriminability capacity of shallow features for detecting small targets. Moreover, DMSFM implements a multireceptive field mechanism to merge target details with deep semantic insights, enabling the capture of more distinctive global contextual features. Experimental evaluations conducted using two public datasets—NUAA-SIRST and NUDT-SIRST—demonstrate that the proposed MEETNet surpasses existing state-of-the-art methods for infrared small target detection in terms of detection accuracy.
{"title":"MEETNet: Morphology-Edge Enhanced Triple-Cascaded Network for Infrared Small Target Detection","authors":"Enyu Zhao;Yu Shi;Nianxin Qu;Yulei Wang;Hang Zhao","doi":"10.1109/JSTARS.2026.3651900","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3651900","url":null,"abstract":"Infrared small target detection is focused on accurately identifying tiny targets with low signal-to-noise ratio against complex backgrounds, representing a critical challenge in the field of infrared image processing. Existing approaches frequently fail to retain small target information during global semantic extraction and struggle with preserving detailed features and achieving effective feature fusion. To address these limitations, this article proposes a morphology-edge enhanced triple-cascaded network (MEETNet) for infrared small target detection. The network employs a triple-cascaded architecture that maintains high resolution and enhances information interaction between different stages, facilitating effective multilevel feature fusion while safeguarding deep small-target characteristics. MEETNet integrates an edge-detail enhanced module (EDEM) and a detail-aware multi-scale fusion module (DMSFM). These modules introduce edge-detail enhanced features that amalgamate contrast and edge information, thereby amplifying target saliency and improving edge representation. Specifically, EDEM augments target contrast and edge structures by integrating edge-detail-enhanced features with shallow details. This integration improves the discriminability capacity of shallow features for detecting small targets. Moreover, DMSFM implements a multireceptive field mechanism to merge target details with deep semantic insights, enabling the capture of more distinctive global contextual features. Experimental evaluations conducted using two public datasets—NUAA-SIRST and NUDT-SIRST—demonstrate that the proposed MEETNet surpasses existing state-of-the-art methods for infrared small target detection in terms of detection accuracy.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4748-4765"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11340625","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146082064","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}
Deep forest-based models for synthetic aperture radar (SAR) image change detection are generally challenged by noise sensitivity and high feature redundancy, which significantly degrade the prediction performance. To address these issues, this article proposes a structure-constrained and feature-screened deep forest, abbreviated as SC-FS-DF, for SAR image change detection. In preclassification, a fuzzy multineighborhood information C-means clustering is proposed to generate high-quality pseudo-labels. It introduces the edge information, the nonlocal and intrasuperpixel neighborhoods into the objective function of fuzzy local information C-means, thus suppressing the speckle noise and constraining structures of targets. In the sample learning and label prediction module, a feature-screened deep forest (FS-DF) framework is constructed by combining feature importance and redundancy analysis with a dropout strategy, thus screening out the noninformative features and meanwhile retaining the informative ones for learning at each cascade layer. Finally, a novel energy function fusing the nonlocal and superpixel information is derived for refining the detection map generated by FS-DF, further preserving fine details and edge locations. Extensive comparison and ablation experiments on five real SAR datasets verify the effectiveness and robustness of the proposed SC-FS-DF, and demonstrate that the SC-FS-DF can well screen the high-dimensional features in change detection and constrain the structures of targets.
{"title":"Feature-Screened and Structure-Constrained Deep Forest for Unsupervised SAR Image Change Detection","authors":"Wanying Song;Ruijing Zhu;Jie Wang;Yinyin Jiang;Yan Wu","doi":"10.1109/JSTARS.2026.3651534","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3651534","url":null,"abstract":"Deep forest-based models for synthetic aperture radar (SAR) image change detection are generally challenged by noise sensitivity and high feature redundancy, which significantly degrade the prediction performance. To address these issues, this article proposes a structure-constrained and feature-screened deep forest, abbreviated as SC-FS-DF, for SAR image change detection. In preclassification, a fuzzy multineighborhood information C-means clustering is proposed to generate high-quality pseudo-labels. It introduces the edge information, the nonlocal and intrasuperpixel neighborhoods into the objective function of fuzzy local information C-means, thus suppressing the speckle noise and constraining structures of targets. In the sample learning and label prediction module, a feature-screened deep forest (FS-DF) framework is constructed by combining feature importance and redundancy analysis with a dropout strategy, thus screening out the noninformative features and meanwhile retaining the informative ones for learning at each cascade layer. Finally, a novel energy function fusing the nonlocal and superpixel information is derived for refining the detection map generated by FS-DF, further preserving fine details and edge locations. Extensive comparison and ablation experiments on five real SAR datasets verify the effectiveness and robustness of the proposed SC-FS-DF, and demonstrate that the SC-FS-DF can well screen the high-dimensional features in change detection and constrain the structures of targets.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4056-4068"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11339914","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026482","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 : 2026-01-12DOI: 10.1109/JSTARS.2026.3651075
Yu Yao;Hengbin Wang;Xiang Gao;Ziyao Xing;Xiaodong Zhang;Yuanyuan Zhao;Shaoming Li;Zhe Liu
High-resolution remote sensing images provide crucial data support for applications such as precision agriculture and water resource management. However, super-resolution reconstructions often suffer from over-smoothed textures and structural distortions, failing to accurately recover the intricate details of ground objects. To address this issue, this article proposes a remote sensing image super-resolution network (DTWSTSR) that combines the Dual-Tree Complex Wavelet Transform and Swin Transformer, which enhances the ability of texture detail reconstruction by fusing frequency-domain and spatial-domain features. This model includes a Dual-Tree Complex Wavelet Texture Feature Sensing Module (DWTFSM) for integrating frequency and spatial features, and a Multiscale Efficient Channel Attention mechanism to enhance attention to multiscale and global details. In addition, we design a Kolmogorov–Arnold Network based on a branch attention mechanism, which improves the model’s ability to represent complex nonlinear features. During the training process, we investigate the impact of hyperparameters and propose the two-stage SSIM&SL1 loss function to reduce structural differences between images. Experimental results show that DTWSTSR outperforms existing mainstream methods under different magnification factors (×2, ×3, ×4), ranking among the top two in multiple metrics. For example, at ×2 magnification, its PSNR value is 0.64–2.68 dB higher than that of other models. Visual comparisons demonstrate that the proposed model achieves clearer and more accurate detail reconstruction of target ground objects. Furthermore, the model exhibits excellent generalization ability in cross-sensor image (OLI2MSI dataset) reconstruction.
{"title":"DTWSTSR: Dual-Tree Complex Wavelet and Swin Transformer Based Remote Sensing Images Super-Resolution Network","authors":"Yu Yao;Hengbin Wang;Xiang Gao;Ziyao Xing;Xiaodong Zhang;Yuanyuan Zhao;Shaoming Li;Zhe Liu","doi":"10.1109/JSTARS.2026.3651075","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3651075","url":null,"abstract":"High-resolution remote sensing images provide crucial data support for applications such as precision agriculture and water resource management. However, super-resolution reconstructions often suffer from over-smoothed textures and structural distortions, failing to accurately recover the intricate details of ground objects. To address this issue, this article proposes a remote sensing image super-resolution network (DTWSTSR) that combines the Dual-Tree Complex Wavelet Transform and Swin Transformer, which enhances the ability of texture detail reconstruction by fusing frequency-domain and spatial-domain features. This model includes a Dual-Tree Complex Wavelet Texture Feature Sensing Module (DWTFSM) for integrating frequency and spatial features, and a Multiscale Efficient Channel Attention mechanism to enhance attention to multiscale and global details. In addition, we design a Kolmogorov–Arnold Network based on a branch attention mechanism, which improves the model’s ability to represent complex nonlinear features. During the training process, we investigate the impact of hyperparameters and propose the two-stage SSIM&SL1 loss function to reduce structural differences between images. Experimental results show that DTWSTSR outperforms existing mainstream methods under different magnification factors (×2, ×3, ×4), ranking among the top two in multiple metrics. For example, at ×2 magnification, its PSNR value is 0.64–2.68 dB higher than that of other models. Visual comparisons demonstrate that the proposed model achieves clearer and more accurate detail reconstruction of target ground objects. Furthermore, the model exhibits excellent generalization ability in cross-sensor image (OLI2MSI dataset) reconstruction.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4730-4747"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11329193","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146082033","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 : 2026-01-12DOI: 10.1109/JSTARS.2026.3651639
Yang Liu;Kun Zhang;Chun-Yi Song;Zhi-Wei Xu
In high-resolution maritime radar working in scanning mode, the classification and identification of ships require the recovery of the ship’s high-resolution range profiles (HRRPs) from radar returns. The return signal from the ship is a complex sparse signal interfered by non-Gaussian sea clutter. In this article, three sparse optimization methods matching the non-Gaussian characteristics of sea clutter, i.e., the sparse optimization matching K-distribution method, the sparse optimization matching generalized Pareto distribution method, the sparse optimization matching CGIG distribution method, are proposed to estimate complex HRRPs of ships. The compound Gaussian model is used to describe the non-Gaussianity of sea clutter, and the sparsity of ships’ complex HRRPs is constrained by the random distribution with one parameter. In the three methods, the Anderson–Darling test is used to search the parameters of the sparse constraint model. Besides, the non-Gaussian characteristics of sea clutter depend on the marine environment parameters and radar operating parameters. For different scenarios, the minimal criterion of the Kolmogorov–Smirnov distance is used to select the best model from the three compound Gaussian models, and then select the corresponding proposed methods. Simulated and measured radar data are used to evaluate the performance of the proposed methods and the results show that the proposed methods obtain better estimates of ship HRRPs compared to the recent SRIM method and the classical SLIM method.
{"title":"Estimation of Ships’ Complex High-Resolution Range Profiles Based on Sparse Optimization Method in Non-Gaussian Sea Clutter","authors":"Yang Liu;Kun Zhang;Chun-Yi Song;Zhi-Wei Xu","doi":"10.1109/JSTARS.2026.3651639","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3651639","url":null,"abstract":"In high-resolution maritime radar working in scanning mode, the classification and identification of ships require the recovery of the ship’s high-resolution range profiles (HRRPs) from radar returns. The return signal from the ship is a complex sparse signal interfered by non-Gaussian sea clutter. In this article, three sparse optimization methods matching the non-Gaussian characteristics of sea clutter, i.e., the sparse optimization matching K-distribution method, the sparse optimization matching generalized Pareto distribution method, the sparse optimization matching CGIG distribution method, are proposed to estimate complex HRRPs of ships. The compound Gaussian model is used to describe the non-Gaussianity of sea clutter, and the sparsity of ships’ complex HRRPs is constrained by the random distribution with one parameter. In the three methods, the Anderson–Darling test is used to search the parameters of the sparse constraint model. Besides, the non-Gaussian characteristics of sea clutter depend on the marine environment parameters and radar operating parameters. For different scenarios, the minimal criterion of the Kolmogorov–Smirnov distance is used to select the best model from the three compound Gaussian models, and then select the corresponding proposed methods. Simulated and measured radar data are used to evaluate the performance of the proposed methods and the results show that the proposed methods obtain better estimates of ship HRRPs compared to the recent SRIM method and the classical SLIM method.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"3998-4013"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11339885","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026530","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 : 2026-01-12DOI: 10.1109/JSTARS.2026.3652368
Jinlong Hu;Ming Zhao;Biying Liu;Xing Chen
Infrared target detection plays a critical role in military reconnaissance, forest-fire prevention, and search-and-rescue operations, owing to its all-weather capability and strong penetration performance. Nevertheless, detecting ultra-small infrared targets in remote sensing and monitoring scenarios remains exceedingly challenging due to the targets’ diminutive size and the long standoff distances involved. In this work, we address this problem by introducing IR-DETR, a lightweight, high-precision infrared small-target detection model built upon the real-time detection Transformer (RT-DETR) framework. First, we propose the multiscale edge-aware convolution module (LDConv), which integrates parallel multiscale Laplacian-of-Gaussian filtering and dilated convolutions within the backbone’s shallow layers, augmented by lightweight channel attention, to markedly enhance the extraction of weak-texture features and boost target–background contrast. Second, we devise the MSCShiftCSP multiscale fusion module: by orchestrating parallel multiscale convolutional branches and parameter-free channel shifting within a CSP structure (replacing the standard RepC3 unit) to strengthen spatial–channel interactions and global context fusion while preserving model efficiency. Third, we replace the original P5 large-target detection head with a high-resolution P2 branch head, halving the channel count from 256 to 128 to capture edges and fine details of targets as small as 6–7 pixels while substantially reducing model complexity. Extensive ablation studies on four public datasets (IRSTD-1 K, SIRST-UAVB, SIRST-v1, and NUDT-SIRST) show that IR-DETR achieves up to a 22% increase in mAP$_{50text{-}95}$ over the RT-DETR baseline with only 9.5 million parameters; notably, on SIRST-UAVB (average target size 6–7pixels), mAP$_{75}$ improves by 47%. Compared against 23 mainstream detectors (including Faster R-CNN, SSD, YOLOv5/8/10/11, and various DETR variants), IR-DETR consistently attains the highest detection accuracy across benchmarks. In summary, IR-DETR delivers a powerful yet lightweight solution for real-time detection of diverse small infrared targets in complex environments, achieving superior accuracy without significant parameter overhead and advancing the state of the art in infrared small-target detection.
{"title":"IR-DETR: An Infrared Small Object Detector Combining Edge-Aware Mechanism and Multiscale Feature Fusion","authors":"Jinlong Hu;Ming Zhao;Biying Liu;Xing Chen","doi":"10.1109/JSTARS.2026.3652368","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3652368","url":null,"abstract":"Infrared target detection plays a critical role in military reconnaissance, forest-fire prevention, and search-and-rescue operations, owing to its all-weather capability and strong penetration performance. Nevertheless, detecting ultra-small infrared targets in remote sensing and monitoring scenarios remains exceedingly challenging due to the targets’ diminutive size and the long standoff distances involved. In this work, we address this problem by introducing <italic>IR-DETR</i>, a lightweight, high-precision infrared small-target detection model built upon the real-time detection Transformer (RT-DETR) framework. First, we propose the multiscale edge-aware convolution module (LDConv), which integrates parallel multiscale Laplacian-of-Gaussian filtering and dilated convolutions within the backbone’s shallow layers, augmented by lightweight channel attention, to markedly enhance the extraction of weak-texture features and boost target–background contrast. Second, we devise the MSCShiftCSP multiscale fusion module: by orchestrating parallel multiscale convolutional branches and parameter-free channel shifting within a CSP structure (replacing the standard RepC3 unit) to strengthen spatial–channel interactions and global context fusion while preserving model efficiency. Third, we replace the original P5 large-target detection head with a high-resolution P2 branch head, halving the channel count from 256 to 128 to capture edges and fine details of targets as small as 6–7 pixels while substantially reducing model complexity. Extensive ablation studies on four public datasets (IRSTD-1 K, SIRST-UAVB, SIRST-v1, and NUDT-SIRST) show that IR-DETR achieves up to a 22% increase in mAP<inline-formula><tex-math>$_{50text{-}95}$</tex-math></inline-formula> over the RT-DETR baseline with only 9.5 million parameters; notably, on SIRST-UAVB (average target size 6–7pixels), mAP<inline-formula><tex-math>$_{75}$</tex-math></inline-formula> improves by 47%. Compared against 23 mainstream detectors (including Faster R-CNN, SSD, YOLOv5/8/10/11, and various DETR variants), IR-DETR consistently attains the highest detection accuracy across benchmarks. In summary, IR-DETR delivers a powerful yet lightweight solution for real-time detection of diverse small infrared targets in complex environments, achieving superior accuracy without significant parameter overhead and advancing the state of the art in infrared small-target detection.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"7264-7279"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11339883","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147299583","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 : 2026-01-12DOI: 10.1109/JSTARS.2026.3652404
Ali Caglayan;Nevrez Imamoglu;Toru Kouyama
Self-supervised pretraining has emerged as a powerful approach for learning transferable representations from large-scale unlabeled data, significantly reducing reliance on task-specific labeled datasets. Although masked autoencoders (MAEs) have shown considerable success in optical remote sensing, such as RGB and multispectral imagery, their application to synthetic aperture radar (SAR) data remains underexplored due to their unique imaging characteristics, including speckle content and intensity variability. In this work, we investigate the effectiveness of MAEs for SAR pretraining, specifically applying MixMAE [Liu, et al.,(2023)] to Sentinel-1 SAR imagery. We introduce SAR-W-MixMAE, a domain-aware self-supervised learning approach that incorporates an SAR-specific pixelwise weighting strategy into the reconstruction loss, mitigating the effects of speckle content and high-intensity backscatter variations. Experimental results demonstrate that SAR-W-MixMAE consistently improves baseline models in multilabel SAR image classification and flood detection tasks, extending the state-of-the-art performance on the popular BigEarthNet dataset. Extensive ablation studies reveal that pretraining duration and fine-tuning dataset size significantly impact downstream performance. In particular, early stopping during pretraining can yield optimal downstream task accuracy, challenging the assumption that prolonged pretraining enhances results. These insights contribute to the development of foundation models tailored for SAR imagery and provide practical guidelines for optimizing pretraining strategies in remote sensing applications.
自监督预训练已经成为一种从大规模未标记数据中学习可转移表征的强大方法，显著减少了对特定任务标记数据集的依赖。尽管掩膜自动编码器（MAEs）在光学遥感（如RGB和多光谱成像）中取得了相当大的成功，但由于其独特的成像特性（包括散斑含量和强度可变性），它们在合成孔径雷达（SAR）数据中的应用仍未得到充分探索。在这项工作中，我们研究了MAEs在SAR预训练中的有效性，特别是将MixMAE [Liu， et .,(2023)]应用于Sentinel-1 SAR图像。我们引入了SAR-W-MixMAE，这是一种领域感知的自监督学习方法，它将sar特定的像素加权策略纳入重建损失，减轻了散斑内容和高强度后向散射变化的影响。实验结果表明，SAR- w - mixmae在多标签SAR图像分类和洪水检测任务中不断改进基线模型，扩展了流行的BigEarthNet数据集的最先进性能。广泛的消融研究表明，预训练时间和微调数据集大小显著影响下游性能。特别是，在预训练期间提前停止可以产生最佳的下游任务准确性，挑战了延长预训练可以提高结果的假设。这些见解有助于开发适合SAR图像的基础模型，并为优化遥感应用中的预训练策略提供实用指南。
{"title":"SAR-W-MixMAE: Polarization-Aware Self-Supervised Pretraining for Masked Autoencoders on SAR Data","authors":"Ali Caglayan;Nevrez Imamoglu;Toru Kouyama","doi":"10.1109/JSTARS.2026.3652404","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3652404","url":null,"abstract":"Self-supervised pretraining has emerged as a powerful approach for learning transferable representations from large-scale unlabeled data, significantly reducing reliance on task-specific labeled datasets. Although masked autoencoders (MAEs) have shown considerable success in optical remote sensing, such as RGB and multispectral imagery, their application to synthetic aperture radar (SAR) data remains underexplored due to their unique imaging characteristics, including speckle content and intensity variability. In this work, we investigate the effectiveness of MAEs for SAR pretraining, specifically applying MixMAE [Liu, et al.,(2023)] to Sentinel-1 SAR imagery. We introduce SAR-W-MixMAE, a domain-aware self-supervised learning approach that incorporates an SAR-specific pixelwise weighting strategy into the reconstruction loss, mitigating the effects of speckle content and high-intensity backscatter variations. Experimental results demonstrate that SAR-W-MixMAE consistently improves baseline models in multilabel SAR image classification and flood detection tasks, extending the state-of-the-art performance on the popular BigEarthNet dataset. Extensive ablation studies reveal that pretraining duration and fine-tuning dataset size significantly impact downstream performance. In particular, early stopping during pretraining can yield optimal downstream task accuracy, challenging the assumption that prolonged pretraining enhances results. These insights contribute to the development of foundation models tailored for SAR imagery and provide practical guidelines for optimizing pretraining strategies in remote sensing applications.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"5590-5601"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11344788","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175697","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 : 2026-01-12DOI: 10.1109/JSTARS.2026.3651577
Wei Huang;JiaLu Li;Qiqiang Chen;Junru Yin;Jiqiang Niu;Le Sun
In recent years, the integration of convolutional neural networks and Transformers has significantly advanced hyperspectral image (HSI) classification by jointly capturing local and global features. However, most existing methods primarily focus on the fusion of spectral–spatial features while neglecting the complementary information contained in frequency-domain features. To address this issue, we propose a spatial–frequency cross-attention fusion network (SFCFNet) that jointly models spectral, spatial, and frequency-domain features for HSI classification. The framework consists of three core modules: first, the multiscale spectral–spatial feature learning module extracts joint spectral spatial features using multiscale 3-D and 2-D convolutions. Next, the triple-branch representation module employs three branches to capture global spatial features of large-scale structures, local spatial features of fine-grained textures, and multiscale frequency features based on Haar wavelet decomposition, providing complementary multidomain representations for subsequent deep fusion. Finally, the dual-domain feature cross-attention fusion module achieves effective fusion of spatial structures and frequency-domain textures, enhancing the model’s ability to separate complex backgrounds from fine-grained targets and thereby improving classification performance. Compared with other methods, SFCFNet achieves higher overall accuracy on the Salinas, Houston2013, WHU-Hi-LongKou, and Xuzhou datasets, reaching 99.05%, 98.07%, 98.76%, and 98.18%, respectively.
{"title":"SFCFNet: A Spatial–Frequency Cross-Attention Fusion Network for Hyperspectral Image Classification","authors":"Wei Huang;JiaLu Li;Qiqiang Chen;Junru Yin;Jiqiang Niu;Le Sun","doi":"10.1109/JSTARS.2026.3651577","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3651577","url":null,"abstract":"In recent years, the integration of convolutional neural networks and Transformers has significantly advanced hyperspectral image (HSI) classification by jointly capturing local and global features. However, most existing methods primarily focus on the fusion of spectral–spatial features while neglecting the complementary information contained in frequency-domain features. To address this issue, we propose a spatial–frequency cross-attention fusion network (SFCFNet) that jointly models spectral, spatial, and frequency-domain features for HSI classification. The framework consists of three core modules: first, the multiscale spectral–spatial feature learning module extracts joint spectral spatial features using multiscale 3-D and 2-D convolutions. Next, the triple-branch representation module employs three branches to capture global spatial features of large-scale structures, local spatial features of fine-grained textures, and multiscale frequency features based on Haar wavelet decomposition, providing complementary multidomain representations for subsequent deep fusion. Finally, the dual-domain feature cross-attention fusion module achieves effective fusion of spatial structures and frequency-domain textures, enhancing the model’s ability to separate complex backgrounds from fine-grained targets and thereby improving classification performance. Compared with other methods, SFCFNet achieves higher overall accuracy on the Salinas, Houston2013, WHU-Hi-LongKou, and Xuzhou datasets, reaching 99.05%, 98.07%, 98.76%, and 98.18%, respectively.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4994-5008"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11340627","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146082022","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 : 2026-01-12DOI: 10.1109/JSTARS.2026.3651847
Kirk M. Scanlan;Anja Rutishauser;Sebastian B. Simonsen
The spatiotemporal properties of the Greenland Ice Sheet firn layer are an important factor when assessing overall ice sheet mass balance and internal meltwater storage capacity. Increasingly a target for the satellite remote sensing community, this study investigates the recovery of vertical firn density heterogeneity over a ten-year period from the synthesis of passive microwave and active radar altimetry measurements. The mismatch between ESA SMOS observations and a passive microwave forward model, initialized with surface densities estimated from the backscatter strength of ISRO/CNES SARAL and ESA CryoSat-2, serves as a proxy for vertical density variability. Validated with in situ measurements, the results demonstrate clear long-term patterns in Greenland firn heterogeneity characterized by spatially expansive sharp increases in firn heterogeneity following extreme melt seasons that require multiple quiescent years to rehabilitate. The results demonstrate that by the start of the 2023 melt season (i.e., the end of the timeframe considered), the Greenland firn layer had reached its most heterogeneous state of the preceding decade. Continued investigation into the synthesis of different remote sensing datasets represents a pathway toward generating novel insights into the spatiotemporal evolution of Greenland Ice Sheet surface conditions.
{"title":"Spatiotemporal Heterogeneity in Greenland Firn From the Synthesis of Satellite Radar Altimetry and Passive Microwave Measurements","authors":"Kirk M. Scanlan;Anja Rutishauser;Sebastian B. Simonsen","doi":"10.1109/JSTARS.2026.3651847","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3651847","url":null,"abstract":"The spatiotemporal properties of the Greenland Ice Sheet firn layer are an important factor when assessing overall ice sheet mass balance and internal meltwater storage capacity. Increasingly a target for the satellite remote sensing community, this study investigates the recovery of vertical firn density heterogeneity over a ten-year period from the synthesis of passive microwave and active radar altimetry measurements. The mismatch between ESA SMOS observations and a passive microwave forward model, initialized with surface densities estimated from the backscatter strength of ISRO/CNES SARAL and ESA CryoSat-2, serves as a proxy for vertical density variability. Validated with in situ measurements, the results demonstrate clear long-term patterns in Greenland firn heterogeneity characterized by spatially expansive sharp increases in firn heterogeneity following extreme melt seasons that require multiple quiescent years to rehabilitate. The results demonstrate that by the start of the 2023 melt season (i.e., the end of the timeframe considered), the Greenland firn layer had reached its most heterogeneous state of the preceding decade. Continued investigation into the synthesis of different remote sensing datasets represents a pathway toward generating novel insights into the spatiotemporal evolution of Greenland Ice Sheet surface conditions.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4085-4098"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11339888","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026391","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}
Remote sensing change detection aims to identify changes on the Earth's surface from remote sensing images acquired at different times. However, the identification of changed areas is often hindered by pseudochanges in similar objects, leading to inaccurate identification of change boundaries. To address this issue, we propose a novel network named boundary-guided semantic context network (BSCNet), which decouples features to improve the feature representation ability for changing objects. Specifically, we design a selective context fusion module that selectively fuses semantically rich features by computing the similarity between features from adjacent stages of the backbone network, thereby preventing detailed features from being overwhelmed by contextual information. In addition, to enhance the ability to perceive changes, we design a context fast aggregation module that leverages a pyramid structure to help the model simultaneously extract and fuse detailed and semantic information at different scales, enabling more accurate change detection. Finally, we design a boundary-guided feature fusion module to aggregate edge-level, texture-level, and semantic-level information, which enables the network to represent change regions more comprehensively and precisely. Experimental results on the WHU-CD, LEVIR-CD, and SYSU-CD datasets show that BSCNet achieves F1 scores of 94.92%, 92.19%, and 82.55%, respectively.
{"title":"Learning Boundary-Aware Semantic Context Network for Remote Sensing Change Detection","authors":"Weiran Zhou;Guanting Guo;Huihui Song;Xu Zhang;Kaihua Zhang","doi":"10.1109/JSTARS.2026.3651696","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3651696","url":null,"abstract":"Remote sensing change detection aims to identify changes on the Earth's surface from remote sensing images acquired at different times. However, the identification of changed areas is often hindered by pseudochanges in similar objects, leading to inaccurate identification of change boundaries. To address this issue, we propose a novel network named boundary-guided semantic context network (BSCNet), which decouples features to improve the feature representation ability for changing objects. Specifically, we design a selective context fusion module that selectively fuses semantically rich features by computing the similarity between features from adjacent stages of the backbone network, thereby preventing detailed features from being overwhelmed by contextual information. In addition, to enhance the ability to perceive changes, we design a context fast aggregation module that leverages a pyramid structure to help the model simultaneously extract and fuse detailed and semantic information at different scales, enabling more accurate change detection. Finally, we design a boundary-guided feature fusion module to aggregate edge-level, texture-level, and semantic-level information, which enables the network to represent change regions more comprehensively and precisely. Experimental results on the WHU-CD, LEVIR-CD, and SYSU-CD datasets show that BSCNet achieves F1 scores of 94.92%, 92.19%, and 82.55%, respectively.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4177-4187"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11339892","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026466","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 : 2026-01-12DOI: 10.1109/JSTARS.2026.3650961
Shixin Liu;Pingyu Liu;Xiaofei Wang
The lack of prior knowledge is a challenging issue in target detection tasks for hyperspectral remote sensing images. In this article, we propose an effective network for object detection in hyperspectral remote sensing images. First, through spectral data augmentation methods, all surrounding pixels within a data block are encoded as the transformed spectral signature of the central pixel, thereby constructing a sufficient number of training sample pairs. Subsequently, a backbone network (PyramidMamba) was designed to establish long-term dependencies across the frequency domain and multiscale dimensions using the Mamba residual module and pyramid wavelet transform module. A residual self-attention module is further developed, integrating self-attention with convolutional operations to enhance feature extraction while improving the network's depth and stability. A backbone network was employed to extract representative vectors from augmented sample pairs, which were then optimized through a spectral contrast head to enhance the distinction between target and background features. Experimental results demonstrate that compared to mainstream algorithms, the proposed algorithm achieves higher detection accuracy and computational efficiency. It successfully learns deep nonlinear feature representations with stronger discriminative power, enabling effective separation of targets from background and delivering state-of-the-art performance.
{"title":"PyramidMamba: An Effective Hyperspectral Remote Sensing Image Target Detection Network","authors":"Shixin Liu;Pingyu Liu;Xiaofei Wang","doi":"10.1109/JSTARS.2026.3650961","DOIUrl":"https://doi.org/10.1109/JSTARS.2026.3650961","url":null,"abstract":"The lack of prior knowledge is a challenging issue in target detection tasks for hyperspectral remote sensing images. In this article, we propose an effective network for object detection in hyperspectral remote sensing images. First, through spectral data augmentation methods, all surrounding pixels within a data block are encoded as the transformed spectral signature of the central pixel, thereby constructing a sufficient number of training sample pairs. Subsequently, a backbone network (PyramidMamba) was designed to establish long-term dependencies across the frequency domain and multiscale dimensions using the Mamba residual module and pyramid wavelet transform module. A residual self-attention module is further developed, integrating self-attention with convolutional operations to enhance feature extraction while improving the network's depth and stability. A backbone network was employed to extract representative vectors from augmented sample pairs, which were then optimized through a spectral contrast head to enhance the distinction between target and background features. Experimental results demonstrate that compared to mainstream algorithms, the proposed algorithm achieves higher detection accuracy and computational efficiency. It successfully learns deep nonlinear feature representations with stronger discriminative power, enabling effective separation of targets from background and delivering state-of-the-art performance.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4163-4176"},"PeriodicalIF":5.3,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11329180","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026481","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：