Pub Date : 2025-11-20DOI: 10.1109/LGRS.2025.3635413
Brent Martin;Norman W. H. Mason;James D. Shepherd;Jan Schindler
The New Zealand Land Use Carbon Analysis System Land Use Map (LUCAS LUM) is a series of land use layers that map land use classes, including both exotic and native forest, dating back to 1990 and updated every four years since 2008. This map is a rich resource, but the significant effort required to update it means errors may creep in without detection. We trialed whether a deep learning model could be trained on this imperfect data. We found the model predicts exotic forestry nationally to a higher level of accuracy than previously achieved. The resulting layer was used to detect and correct missed exotic forest plantations in the current LUCAS LUM. We also demonstrate that the exotic forestry prediction is sufficiently sensitive to detect wilding conifer infestations and estimate infestation density. Our results highlight the effectiveness of weakly supervised learning, enabling accurate and scalable national land use and land cover mapping while drastically reducing manual labeling efforts.
{"title":"Improving New Zealand’s Vegetation Mapping Using Weakly Supervised Learning","authors":"Brent Martin;Norman W. H. Mason;James D. Shepherd;Jan Schindler","doi":"10.1109/LGRS.2025.3635413","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3635413","url":null,"abstract":"The New Zealand Land Use Carbon Analysis System Land Use Map (LUCAS LUM) is a series of land use layers that map land use classes, including both exotic and native forest, dating back to 1990 and updated every four years since 2008. This map is a rich resource, but the significant effort required to update it means errors may creep in without detection. We trialed whether a deep learning model could be trained on this imperfect data. We found the model predicts exotic forestry nationally to a higher level of accuracy than previously achieved. The resulting layer was used to detect and correct missed exotic forest plantations in the current LUCAS LUM. We also demonstrate that the exotic forestry prediction is sufficiently sensitive to detect wilding conifer infestations and estimate infestation density. Our results highlight the effectiveness of weakly supervised learning, enabling accurate and scalable national land use and land cover mapping while drastically reducing manual labeling efforts.","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145674859","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-19DOI: 10.1109/LGRS.2025.3634593
HengYu Li;Bo Huang;JianYong Lv
Driven by the increasing demand for intelligent Earth observation and large-scale scene understanding, remote sensing object detection has gained significant academic and practical importance. Despite notable progress in feature extraction and computational efficiency, many recent approaches still struggle to effectively handle issues such as detecting objects at multiple scales and preserving small targets. In this letter, an efficient remote sensing object detector called multiscale and feature-preserving YOLO with gated attention (YOLO-MFG) is proposed to address these challenges. First, a multiscale group shuffle attention (MGSA) module is introduced to adaptively aggregate multiscale spatial features, improving the model’s sensitivity to objects of diverse sizes. Second, the use of feature-preserving downsampling (FPD) enhances the downsampling process by introducing a triple-branch fusion mechanism that mitigates aliasing while jointly preserving semantics, saliency, and geometry. Finally, gated enhanced attention (GEA) is integrated to capture long-range dependencies and contextual cues crucial for remote sensing scenarios. The experimental results demonstrate that the proposed YOLO-MFG achieves a 2.9% improvement in mean average precision at an intersection over union (IoU) threshold of 0.5 (mAP50) on the optical remote sensing dataset SIMD compared with YOLO11. In addition, the mAP50 of detection results is improved by 1.4% and 4.2% on the DIOR and NWPU VHR-10 datasets, respectively.
{"title":"YOLO-MFG: Multiscale and Feature-Preserving YOLO With Gated Attention for Remote Sensing Object Detection","authors":"HengYu Li;Bo Huang;JianYong Lv","doi":"10.1109/LGRS.2025.3634593","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3634593","url":null,"abstract":"Driven by the increasing demand for intelligent Earth observation and large-scale scene understanding, remote sensing object detection has gained significant academic and practical importance. Despite notable progress in feature extraction and computational efficiency, many recent approaches still struggle to effectively handle issues such as detecting objects at multiple scales and preserving small targets. In this letter, an efficient remote sensing object detector called multiscale and feature-preserving YOLO with gated attention (YOLO-MFG) is proposed to address these challenges. First, a multiscale group shuffle attention (MGSA) module is introduced to adaptively aggregate multiscale spatial features, improving the model’s sensitivity to objects of diverse sizes. Second, the use of feature-preserving downsampling (FPD) enhances the downsampling process by introducing a triple-branch fusion mechanism that mitigates aliasing while jointly preserving semantics, saliency, and geometry. Finally, gated enhanced attention (GEA) is integrated to capture long-range dependencies and contextual cues crucial for remote sensing scenarios. The experimental results demonstrate that the proposed YOLO-MFG achieves a 2.9% improvement in mean average precision at an intersection over union (IoU) threshold of 0.5 (mAP50) on the optical remote sensing dataset SIMD compared with YOLO11. In addition, the mAP50 of detection results is improved by 1.4% and 4.2% on the DIOR and NWPU VHR-10 datasets, respectively.","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145830887","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Forest tree species classification has great significance for sustainable development of forest resource. Multisource remote sensing data provide abundant temporal, spatial, and spectral information for tree species classification. However, there lacks tree species classification methods, which comprehensively capture and fuse spatio–temporal–spectral information. Therefore, a tree species classification method based on deep ensemble learning of multisource spatio–temporal–spectral remote sensing data is proposed. First, multitemporal, high-resolution, and hyperspectral data are utilized for training temporal, spatial, and spectral deep networks. Furtherly, deep ensemble learning is developed for the fusion of spatio–temporal–spectral network outputs, where weighted fusion is implemented via dynamic weight optimization based on the spatio–temporal–spatial features. Experimental results indicate that the importance of temporal features is higher than that of spatial information, and spectral networks perform best among all network structures. After the spatio–temporal–spectral ensemble learning, the performance of tree species classification is further improved, and the overall accuracy (OA) of the proposed method reaches above 90%. The proposed algorithm realizes precise and fine-scale tree species classification and provides technique support for the monitoring and conservation of forest resource.
{"title":"Forest Tree Species Classification Based on Deep Ensemble Learning by Fusing High-Resolution, Multitemporal, and Hyperspectral Multisource Remote Sensing Data","authors":"Dengli Yu;Lilin Tu;Ziqing Wei;Fuyao Zhu;Chengjun Yu;Denghong Wang;Jiayi Li;Xin Huang","doi":"10.1109/LGRS.2025.3634553","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3634553","url":null,"abstract":"Forest tree species classification has great significance for sustainable development of forest resource. Multisource remote sensing data provide abundant temporal, spatial, and spectral information for tree species classification. However, there lacks tree species classification methods, which comprehensively capture and fuse spatio–temporal–spectral information. Therefore, a tree species classification method based on deep ensemble learning of multisource spatio–temporal–spectral remote sensing data is proposed. First, multitemporal, high-resolution, and hyperspectral data are utilized for training temporal, spatial, and spectral deep networks. Furtherly, deep ensemble learning is developed for the fusion of spatio–temporal–spectral network outputs, where weighted fusion is implemented via dynamic weight optimization based on the spatio–temporal–spatial features. Experimental results indicate that the importance of temporal features is higher than that of spatial information, and spectral networks perform best among all network structures. After the spatio–temporal–spectral ensemble learning, the performance of tree species classification is further improved, and the overall accuracy (OA) of the proposed method reaches above 90%. The proposed algorithm realizes precise and fine-scale tree species classification and provides technique support for the monitoring and conservation of forest resource.","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145612128","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-19DOI: 10.1109/LGRS.2025.3634759
Shengyi Wang;Xuehua Chen;Cong Wang;Junjie Liu;Xin Luo
High-resolution time–frequency analysis is crucial for seismic interpretation. Conventional sparse time–frequency transforms, such as the sparse generalized S transform (SGST), are not adaptive to the intrinsic characteristics of the signal. To address this limitation, we propose a sparse adaptive generalized S transform (SAGST). This method incorporates the signal amplitude spectrum into the Gaussian window function, allowing the window to adapt dynamically to the signal characteristics. This adaptive mechanism enables the construction of wavelet bases that are better matched to the signal. We apply the SAGST to the time–frequency analysis of both synthetic signal and field seismic data. The synthetic signal test shows that the SAGST achieves higher energy concentration, superior computational efficiency, and enhanced weak signal extraction compared with the sparse adaptive S transform (SAST) and SGST. A field example demonstrates that the SAGST can be used to indicate low-frequency shadow associated with hydrocarbon reservoirs.
{"title":"The Sparse Adaptive Generalized S Transform","authors":"Shengyi Wang;Xuehua Chen;Cong Wang;Junjie Liu;Xin Luo","doi":"10.1109/LGRS.2025.3634759","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3634759","url":null,"abstract":"High-resolution time–frequency analysis is crucial for seismic interpretation. Conventional sparse time–frequency transforms, such as the sparse generalized S transform (SGST), are not adaptive to the intrinsic characteristics of the signal. To address this limitation, we propose a sparse adaptive generalized S transform (SAGST). This method incorporates the signal amplitude spectrum into the Gaussian window function, allowing the window to adapt dynamically to the signal characteristics. This adaptive mechanism enables the construction of wavelet bases that are better matched to the signal. We apply the SAGST to the time–frequency analysis of both synthetic signal and field seismic data. The synthetic signal test shows that the SAGST achieves higher energy concentration, superior computational efficiency, and enhanced weak signal extraction compared with the sparse adaptive S transform (SAST) and SGST. A field example demonstrates that the SAGST can be used to indicate low-frequency shadow associated with hydrocarbon reservoirs.","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145612170","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-18DOI: 10.1109/LGRS.2025.3634350
Wenqiang Ding;Changying Ma;Xintong Dong;Xuan Li
The heterogeneity of subsurface media induces multipath scattering and dielectric loss in ground penetrating radar (GPR) signal propagation, which results in wavefront distortion and signal attenuation. These effects degrade B-scan profiles by blurring target signatures, hindering automated feature extraction, and reducing the clarity of regions of interest (ROI). To address these issues, we propose the adaptive region target enhancement algorithm (ARTEA), a multistage preprocessing framework. ARTEA integrates dynamic range compression, continuous-scale normalization guided by adaptive sigma maps, and a frequency-domain refinement step. By dynamically adjusting parameters according to local signal characteristics, ARTEA is designed to achieve an effective tradeoff between artifact suppression and target preservation. Experiments on both synthetic and field GPR data demonstrate that ARTEA can enhance target contrast and structural fidelity while suppressing artifacts and preserving essential target features.
{"title":"ARTEA: A Multistage Adaptive Preprocessing Algorithm for Subsurface Target Enhancement in Ground Penetrating Radar","authors":"Wenqiang Ding;Changying Ma;Xintong Dong;Xuan Li","doi":"10.1109/LGRS.2025.3634350","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3634350","url":null,"abstract":"The heterogeneity of subsurface media induces multipath scattering and dielectric loss in ground penetrating radar (GPR) signal propagation, which results in wavefront distortion and signal attenuation. These effects degrade B-scan profiles by blurring target signatures, hindering automated feature extraction, and reducing the clarity of regions of interest (ROI). To address these issues, we propose the adaptive region target enhancement algorithm (ARTEA), a multistage preprocessing framework. ARTEA integrates dynamic range compression, continuous-scale normalization guided by adaptive sigma maps, and a frequency-domain refinement step. By dynamically adjusting parameters according to local signal characteristics, ARTEA is designed to achieve an effective tradeoff between artifact suppression and target preservation. Experiments on both synthetic and field GPR data demonstrate that ARTEA can enhance target contrast and structural fidelity while suppressing artifacts and preserving essential target features.","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145830889","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Remote sensing (RS) image scene classification has wide applications in the field of RS. Although the existing methods have achieved remarkable performance, there are still limitations in feature extraction and lightweight design. Current multibranch models, although performing well, have large parameter counts and high computational costs, making them difficult to deploy on resource-constrained edge devices, such as uncrewed aerial vehicles (UAVs). On the other hand, lightweight models like StarNet, having less parameter, but rely on elementwise multiplication to generate features and lack the capture of explicit long-range spatial feature, resulting in insufficient classification accuracy. To address these issues, this letter proposes a lightweight mamba-based hybrid network, namely LMHMamba, whose core is an innovative lightweight multifeature hybrid Mamba (LMHM) module. This module combines the advantage of StarNet in implicitly generating high-dimensional nonlinear features, introduces a lightweight state-space module to enhance spatial feature learning capabilities, and then uses local and global attention modules to emphasize local and global features. This enables effective multidimensional feature fusion while maintaining low parameter. We validate the performance of LMHMamba model on three RS scene classification datasets and compare it with mainstream lightweight models and the latest methods. Experimental results show that LMHMamba achieves advanced levels in both classification accuracy and computational efficiency, significantly outperforming the existing lightweight models, providing an efficient solution for edge deployment. Code is available at https://github.com/yizhilanmaodhh/LMHMamba
{"title":"A Lightweight Multifeature Hybrid Mamba for Remote Sensing Image Scene Classification","authors":"Huihui Dong;Jingcao Li;Zongfang Ma;Zhijie Li;Mengkun Liu;Xiaohui Wei;Licheng Jiao","doi":"10.1109/LGRS.2025.3634398","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3634398","url":null,"abstract":"Remote sensing (RS) image scene classification has wide applications in the field of RS. Although the existing methods have achieved remarkable performance, there are still limitations in feature extraction and lightweight design. Current multibranch models, although performing well, have large parameter counts and high computational costs, making them difficult to deploy on resource-constrained edge devices, such as uncrewed aerial vehicles (UAVs). On the other hand, lightweight models like StarNet, having less parameter, but rely on elementwise multiplication to generate features and lack the capture of explicit long-range spatial feature, resulting in insufficient classification accuracy. To address these issues, this letter proposes a lightweight mamba-based hybrid network, namely LMHMamba, whose core is an innovative lightweight multifeature hybrid Mamba (LMHM) module. This module combines the advantage of StarNet in implicitly generating high-dimensional nonlinear features, introduces a lightweight state-space module to enhance spatial feature learning capabilities, and then uses local and global attention modules to emphasize local and global features. This enables effective multidimensional feature fusion while maintaining low parameter. We validate the performance of LMHMamba model on three RS scene classification datasets and compare it with mainstream lightweight models and the latest methods. Experimental results show that LMHMamba achieves advanced levels in both classification accuracy and computational efficiency, significantly outperforming the existing lightweight models, providing an efficient solution for edge deployment. Code is available at <uri>https://github.com/yizhilanmaodhh/LMHMamba</uri>","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145612119","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-18DOI: 10.1109/LGRS.2025.3634345
Bo Zhang;Yaxiong Chen;Ruilin Yao;Shengwu Xiong
The core of hyperspectral change detection lies in accurately capturing spectral feature differences across different temporal phases to determine whether surface objects have changed. Since spectral variations of different ground objects often manifest more prominently in specific wavelength bands, we design a weighted cascaded encoder–decoder network (WCEDNet) based on spatial–spectral difference features for hyperspectral change detection. First, unlike conventional change detection frameworks based on siamese networks, our proposed single-branch approach focuses more intensively on extracting spatial–spectral difference features. Second, the weighted cascaded structure introduced in the encoder stage enables differential attention to different bands, enhancing focus on spectral bands with high responsiveness. Furthermore, we have developed a spatial–spectral cross-attention (SSCA) module to model intrafeature correlations within spatial and spectral domains. Our method was evaluated on three challenging hyperspectral change detection datasets, and experimental results demonstrate its superior performance compared to competitive models. The detailed code has been open-sourced at https://github.com/WUTCM-Lab/WCEDNet
{"title":"WCEDNet: A Weighted Cascaded Encoder–Decoder Network for Hyperspectral Change Detection Based on Spatial–Spectral Difference Features","authors":"Bo Zhang;Yaxiong Chen;Ruilin Yao;Shengwu Xiong","doi":"10.1109/LGRS.2025.3634345","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3634345","url":null,"abstract":"The core of hyperspectral change detection lies in accurately capturing spectral feature differences across different temporal phases to determine whether surface objects have changed. Since spectral variations of different ground objects often manifest more prominently in specific wavelength bands, we design a weighted cascaded encoder–decoder network (WCEDNet) based on spatial–spectral difference features for hyperspectral change detection. First, unlike conventional change detection frameworks based on siamese networks, our proposed single-branch approach focuses more intensively on extracting spatial–spectral difference features. Second, the weighted cascaded structure introduced in the encoder stage enables differential attention to different bands, enhancing focus on spectral bands with high responsiveness. Furthermore, we have developed a spatial–spectral cross-attention (SSCA) module to model intrafeature correlations within spatial and spectral domains. Our method was evaluated on three challenging hyperspectral change detection datasets, and experimental results demonstrate its superior performance compared to competitive models. The detailed code has been open-sourced at <uri>https://github.com/WUTCM-Lab/WCEDNet</uri>","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145612164","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-17DOI: 10.1109/LGRS.2025.3633285
Weihua Shen;Yalin Li;Xiaohua Chen;Chunzhi Li
There are multiple challenges in small object detection (SOD), including limited instances, insufficient features, diverse scales, uneven distribution, ambiguous boundaries, and complex backgrounds. These issues often lead to high false detection rates and hinder model generalization and convergence. This study proposes a multiscale object detection algorithm that enhances the detection of subtle features by improving the change detection to DH throughout and incorporating a minimum point distance intersection-over-union loss. The enhanced DH improves target representation, enabling more precise localization and classification of small objects. Meanwhile, the new loss (NL) function stabilizes bounding box regression by adaptively adjusting auxiliary bounding box scales. Evaluations on two benchmark datasets demonstrate that our method achieves a 2.6% increase in mAP50 and a 1.8% improvement in mAP50:95 on the satellite imagery multivehicles dataset (SIMD) and a 1.9% increase in mAP50:95 on the DIOR dataset. Furthermore, the model reduces the number of parameters by 2.5% and the computational cost by 1.4%, demonstrating its potential for real-time detection applications.
{"title":"A Multiscale Feature Refinement Detector for Small Objects With Ambiguous Boundaries","authors":"Weihua Shen;Yalin Li;Xiaohua Chen;Chunzhi Li","doi":"10.1109/LGRS.2025.3633285","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3633285","url":null,"abstract":"There are multiple challenges in small object detection (SOD), including limited instances, insufficient features, diverse scales, uneven distribution, ambiguous boundaries, and complex backgrounds. These issues often lead to high false detection rates and hinder model generalization and convergence. This study proposes a multiscale object detection algorithm that enhances the detection of subtle features by improving the change detection to DH throughout and incorporating a minimum point distance intersection-over-union loss. The enhanced DH improves target representation, enabling more precise localization and classification of small objects. Meanwhile, the new loss (NL) function stabilizes bounding box regression by adaptively adjusting auxiliary bounding box scales. Evaluations on two benchmark datasets demonstrate that our method achieves a 2.6% increase in mAP50 and a 1.8% improvement in mAP50:95 on the satellite imagery multivehicles dataset (SIMD) and a 1.9% increase in mAP50:95 on the DIOR dataset. Furthermore, the model reduces the number of parameters by 2.5% and the computational cost by 1.4%, demonstrating its potential for real-time detection applications.","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145612169","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-17DOI: 10.1109/LGRS.2025.3633718
Kwonyoung Kim;Jungin Park;Kwanghoon Sohn
Parameter-efficient fine-tuning (PEFT) adapts large pretrained foundation models to downstream tasks, such as remote sensing scene classification, by learning a small set of additional parameters while keeping the pretrained parameters frozen. While PEFT offers substantial training efficiency over full fine-tuning (FT), it still incurs high inference costs due to reliance on both pretrained and task-specific parameters. To address this limitation, we propose a novel PEFT approach with model truncation, termed truncated parameter-efficient fine-tuning (TruncPEFT), enabling efficiency gains to persist during inference. Observing that predictions from final and intermediate layers often exhibit high agreement, we truncate a set of final layers and replace them with a lightweight attention module. Additionally, we introduce a token dropping strategy to mitigate interclass interference, reducing the model’s sensitivity to visual similarities between different classes in remote sensing data. Extensive experiments on seven remote sensing scene classification datasets demonstrate the effectiveness of the proposed method, significantly improving training, inference, and GPU memory efficiencies while achieving comparable or even better performance than prior PEFT methods and full FT.
{"title":"Geospatial Domain Adaptation With Truncated Parameter-Efficient Fine-Tuning","authors":"Kwonyoung Kim;Jungin Park;Kwanghoon Sohn","doi":"10.1109/LGRS.2025.3633718","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3633718","url":null,"abstract":"Parameter-efficient fine-tuning (PEFT) adapts large pretrained foundation models to downstream tasks, such as remote sensing scene classification, by learning a small set of additional parameters while keeping the pretrained parameters frozen. While PEFT offers substantial training efficiency over full fine-tuning (FT), it still incurs high inference costs due to reliance on both pretrained and task-specific parameters. To address this limitation, we propose a novel PEFT approach with model truncation, termed truncated parameter-efficient fine-tuning (TruncPEFT), enabling efficiency gains to persist during inference. Observing that predictions from final and intermediate layers often exhibit high agreement, we truncate a set of final layers and replace them with a lightweight attention module. Additionally, we introduce a token dropping strategy to mitigate interclass interference, reducing the model’s sensitivity to visual similarities between different classes in remote sensing data. Extensive experiments on seven remote sensing scene classification datasets demonstrate the effectiveness of the proposed method, significantly improving training, inference, and GPU memory efficiencies while achieving comparable or even better performance than prior PEFT methods and full FT.","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145674822","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-11-17DOI: 10.1109/LGRS.2025.3633588
Yanyan Zhang;Akira Hirose;Ryo Natsuaki
Advanced Land Observing Satellite-4 (ALOS-4) is a spaceborne high-resolution and wide-swath synthetic aperture radar (HRWS-SAR) that uses a variable pulse repetition interval (VPRI) technique to achieve continuous wide imaging. In some ALOS-4 images, azimuth fractional ambiguity caused by the VPRI is observed, and it differs from the usual integer ambiguity, resulting from interchannel errors in that it occurs at smaller intervals. In this letter, we propose a sequential Doppler offset (SDO) method for locating the original target (OT) that causes azimuth fractional ambiguity. First, the ratio of the interval of integer ambiguity to that of fractional ambiguity is obtained, which is used to generate SAR images with different Doppler center frequencies. Second, the coherence between the sum image of the generated images and the image with a zero Doppler center frequency is calculated. Third, some points with coherence greater than a threshold are selected based on the coherence. Finally, the final OT is obtained by detecting the filtered selected points. Some experiments are conducted based on ALOS-4 L1.2 data, and the results demonstrate that the method locates the OT accurately. In short, the proposed method provides a starting point for fractional ambiguity suppression in HRWS-SAR.
{"title":"A Sequential Doppler Offset (SDO) Method for Locating Targets Causing Azimuth Fractional Ambiguity in Spaceborne HRWS-SAR","authors":"Yanyan Zhang;Akira Hirose;Ryo Natsuaki","doi":"10.1109/LGRS.2025.3633588","DOIUrl":"https://doi.org/10.1109/LGRS.2025.3633588","url":null,"abstract":"Advanced Land Observing Satellite-4 (ALOS-4) is a spaceborne high-resolution and wide-swath synthetic aperture radar (HRWS-SAR) that uses a variable pulse repetition interval (VPRI) technique to achieve continuous wide imaging. In some ALOS-4 images, azimuth fractional ambiguity caused by the VPRI is observed, and it differs from the usual integer ambiguity, resulting from interchannel errors in that it occurs at smaller intervals. In this letter, we propose a sequential Doppler offset (SDO) method for locating the original target (OT) that causes azimuth fractional ambiguity. First, the ratio of the interval of integer ambiguity to that of fractional ambiguity is obtained, which is used to generate SAR images with different Doppler center frequencies. Second, the coherence between the sum image of the generated images and the image with a zero Doppler center frequency is calculated. Third, some points with coherence greater than a threshold are selected based on the coherence. Finally, the final OT is obtained by detecting the filtered selected points. Some experiments are conducted based on ALOS-4 L1.2 data, and the results demonstrate that the method locates the OT accurately. In short, the proposed method provides a starting point for fractional ambiguity suppression in HRWS-SAR.","PeriodicalId":91017,"journal":{"name":"IEEE geoscience and remote sensing letters : a publication of the IEEE Geoscience and Remote Sensing Society","volume":"23 ","pages":"1-5"},"PeriodicalIF":4.4,"publicationDate":"2025-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145612123","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","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：