As severe convective weather exerts growing influence on public safety, enhancing forecast accuracy has become critically important. However, the predictive capability remains limited due to insufficient observational coverageenlrg in certain regions or variables, as well as the inadequate representation of the fine-scale physical processes responsible for local convective development. In response to these challenges, this study proposes a physically embedded neural network based on heterogeneous meteorological data, which utilizes satellite multispectral images and atmospheric temperature and humidity profile synergistically retrieved from space-based and ground-based infrared spectral observations, to forecast local convective initiation (CI) within a 6-hour lead time. The core innovation of this study lies in the development of a physically consistent model that explicitly embeds the convective available potential energy equation into the network architecture. By embedding physical information, the model enables the atmospheric thermodynamic feature extraction module to generate physically consistent feature tensors, thereby enhancing the representation of key convective processes. We trained the network using the pretraining and fine-tuning approach, then validated its effectiveness with reanalysis and actual observational data. The results demonstrate that incorporating the retrieved atmospheric profile data leads to a 40% improvement in the 6-hour average critical success index (CSI), increasing from 0.44 to 0.62 relative to forecasts without atmospheric profile input. Furthermore, in validation experiments using reanalysis data and radar observations, the proposed atmospheric profile feature extraction module consistently improves the model’s average forecast CSI by more than 29% compared to models utilizing purely data-driven profile extraction modules.
{"title":"A Deep Learning-Based Model for Nowcasting of Convective Initiation Using Infrared Observations","authors":"Huijie Zhao;Xiaohang Ma;Guorui Jia;Jialu Xu;Yihan Xie;Yujun Zhao","doi":"10.1109/JSTARS.2025.3650686","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3650686","url":null,"abstract":"As severe convective weather exerts growing influence on public safety, enhancing forecast accuracy has become critically important. However, the predictive capability remains limited due to insufficient observational coverageenlrg in certain regions or variables, as well as the inadequate representation of the fine-scale physical processes responsible for local convective development. In response to these challenges, this study proposes a physically embedded neural network based on heterogeneous meteorological data, which utilizes satellite multispectral images and atmospheric temperature and humidity profile synergistically retrieved from space-based and ground-based infrared spectral observations, to forecast local convective initiation (CI) within a 6-hour lead time. The core innovation of this study lies in the development of a physically consistent model that explicitly embeds the convective available potential energy equation into the network architecture. By embedding physical information, the model enables the atmospheric thermodynamic feature extraction module to generate physically consistent feature tensors, thereby enhancing the representation of key convective processes. We trained the network using the pretraining and fine-tuning approach, then validated its effectiveness with reanalysis and actual observational data. The results demonstrate that incorporating the retrieved atmospheric profile data leads to a 40% improvement in the 6-hour average critical success index (CSI), increasing from 0.44 to 0.62 relative to forecasts without atmospheric profile input. Furthermore, in validation experiments using reanalysis data and radar observations, the proposed atmospheric profile feature extraction module consistently improves the model’s average forecast CSI by more than 29% compared to models utilizing purely data-driven profile extraction modules.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4188-4202"},"PeriodicalIF":5.3,"publicationDate":"2026-01-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11328812","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026535","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 : 2026-01-01DOI: 10.1109/JSTARS.2025.3650394
Shun Zhang;Xuebin Zhang;Yaohui Xu;Ke Wang
Few-shot object detection (FSOD) in remote sensing imagery faces two critical challenges compared to general methods trained on large datasets, first, only a few labeled instances leveraged as the training set significantly limit the feature representation learning of deep neural networks; second, Remote sensing image data contain complicated background and multiple objects with greatly different sizes in the same image, which leads the detector to large numbers of false alarms and miss detections. This article proposes a FSOD framework (called DeCL-Det) that applies self-training to generate high-quality pseudoannotations from unlabeled target domain data. These refined pseudolabels are iteratively integrated into the training set to expand supervision for novel classes. An auxiliary network is introduced to mitigate label noise by rectifying misclassifications in pseudolabeled regions, ensuring robust learning. For multiscale feature learning, we propose a gradient-decoupled framework, GCFPN, combining feature pyramid networks (FPN) with a gradient decoupled layer (GDL). FPN is to extract multiscale feature representations, and GDL is to decouple the modules between the region proposal network and RCNN head into two stages or tasks through gradients. The two modules, FPN and GDL, train Faster R-CNN in a decoupled way to facilitate the multiscale feature learning of novel objects. To further enhance the classification ability, we introduce a supervised contrastive learning head to enhance feature discrimination, reinforcing robustness in FSOD. Experiments on the DIOR dataset indicate that our method performs better than several existing approaches and achieves competitive results.
{"title":"Few-Shot Object Detection on Remote Sensing Images Based on Decoupled Training, Contrastive Learning, and Self-Training","authors":"Shun Zhang;Xuebin Zhang;Yaohui Xu;Ke Wang","doi":"10.1109/JSTARS.2025.3650394","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3650394","url":null,"abstract":"Few-shot object detection (FSOD) in remote sensing imagery faces two critical challenges compared to general methods trained on large datasets, first, only a few labeled instances leveraged as the training set significantly limit the feature representation learning of deep neural networks; second, Remote sensing image data contain complicated background and multiple objects with greatly different sizes in the same image, which leads the detector to large numbers of false alarms and miss detections. This article proposes a FSOD framework (called DeCL-Det) that applies self-training to generate high-quality pseudoannotations from unlabeled target domain data. These refined pseudolabels are iteratively integrated into the training set to expand supervision for novel classes. An auxiliary network is introduced to mitigate label noise by rectifying misclassifications in pseudolabeled regions, ensuring robust learning. For multiscale feature learning, we propose a gradient-decoupled framework, GCFPN, combining feature pyramid networks (FPN) with a gradient decoupled layer (GDL). FPN is to extract multiscale feature representations, and GDL is to decouple the modules between the region proposal network and RCNN head into two stages or tasks through gradients. The two modules, FPN and GDL, train Faster R-CNN in a decoupled way to facilitate the multiscale feature learning of novel objects. To further enhance the classification ability, we introduce a supervised contrastive learning head to enhance feature discrimination, reinforcing robustness in FSOD. Experiments on the DIOR dataset indicate that our method performs better than several existing approaches and achieves competitive results.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"3983-3997"},"PeriodicalIF":5.3,"publicationDate":"2026-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11321270","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026503","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 : 2025-12-31DOI: 10.1109/JSTARS.2025.3649816
Siqi Lai;Mingliang Tao;Yanyang Liu;Lei Cui;Jia Su;Ling Wang
Radio frequency interference (RFI) may degrade the quality of remote sensing images acquired by spaceborne synthetic aperture radar (SAR). In the interferometric wide-swath mode of the Sentinel-1 satellite, the SAR receiver may capture multiple types of RFI signals within a single observation period, which is referred to as heterogeneous RFI, increasing the complexity of interference detection and mitigation. This article proposes a heterogeneous interference mitigation method based on subimage segmentation and local spectral features analysis. The proposed method divides the original single look complex image into multiple subimages along the range direction, enhancing the representation of interference features in the range frequency domain. Spectral analysis is then performed on each subimage to detect and mitigate interference. Finally, the image after RFI mitigation is reconstructed by stitching the subimages together. Experiments were conducted using simulated interference data generated from LuTan-1 and measured interference data from Sentinel-1. The results demonstrate that the proposed method can effectively mitigate RFI artifacts in various typical interference scenarios and restore the obscured ground object information in the images.
{"title":"Heterogeneous RFI Mitigation in Image-Domain via Subimage Segmentation and Local Frequency Feature Analysis","authors":"Siqi Lai;Mingliang Tao;Yanyang Liu;Lei Cui;Jia Su;Ling Wang","doi":"10.1109/JSTARS.2025.3649816","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3649816","url":null,"abstract":"Radio frequency interference (RFI) may degrade the quality of remote sensing images acquired by spaceborne synthetic aperture radar (SAR). In the interferometric wide-swath mode of the Sentinel-1 satellite, the SAR receiver may capture multiple types of RFI signals within a single observation period, which is referred to as heterogeneous RFI, increasing the complexity of interference detection and mitigation. This article proposes a heterogeneous interference mitigation method based on subimage segmentation and local spectral features analysis. The proposed method divides the original single look complex image into multiple subimages along the range direction, enhancing the representation of interference features in the range frequency domain. Spectral analysis is then performed on each subimage to detect and mitigate interference. Finally, the image after RFI mitigation is reconstructed by stitching the subimages together. Experiments were conducted using simulated interference data generated from LuTan-1 and measured interference data from Sentinel-1. The results demonstrate that the proposed method can effectively mitigate RFI artifacts in various typical interference scenarios and restore the obscured ground object information in the images.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4069-4084"},"PeriodicalIF":5.3,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11320316","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026506","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 : 2025-12-31DOI: 10.1109/JSTARS.2025.3647616
Cong Wang;Yunfeng Wang;Yu Wang;Mingming Xu;Leiquan Wang
In hyperspectral anomaly detection (HAD), anomalous pixels typically exhibit a sparsely distributed spatial pattern. Existing deep models often generate backgrounds by reconstructing spectral vectors, yet fail to capture the inherent spatial characteristics of the image. To overcome the semantic and structural information loss caused by neglecting spatial features, we propose the granularity-inconsistent transformer (GIFormer) for unsupervised HAD. Specifically, the interaction between the spatial and spectral dimensions is leveraged to enhance the spatial-spectral feature representation of hyperspectral images, highlighting the differences between background and anomaly features. The GIFormer performs multilevel background reconstruction to detect anomalies. In the encoder, patch-level anomaly elimination masks are applied to reconstruct background features, where spatial correlations of anomalies are utilized to suppress anomalous patterns spanning multiple pixels. The decoder operates at the pixel level, using fine-grained receptive fields for global attention modeling, which enables the model to refine local details that may have been aggregated by the encoder in larger patches, ensuring the final reconstruction retains the intricate structure of the original hyperspectral data. Furthermore, adaptive weight loss is incorporated to guide network training. Extensive experimental results confirm the superior performance of GIFormer.
{"title":"Granularity-Inconsistent Transformer for Unsupervised Hyperspectral Anomaly Detection","authors":"Cong Wang;Yunfeng Wang;Yu Wang;Mingming Xu;Leiquan Wang","doi":"10.1109/JSTARS.2025.3647616","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3647616","url":null,"abstract":"In hyperspectral anomaly detection (HAD), anomalous pixels typically exhibit a sparsely distributed spatial pattern. Existing deep models often generate backgrounds by reconstructing spectral vectors, yet fail to capture the inherent spatial characteristics of the image. To overcome the semantic and structural information loss caused by neglecting spatial features, we propose the granularity-inconsistent transformer (GIFormer) for unsupervised HAD. Specifically, the interaction between the spatial and spectral dimensions is leveraged to enhance the spatial-spectral feature representation of hyperspectral images, highlighting the differences between background and anomaly features. The GIFormer performs multilevel background reconstruction to detect anomalies. In the encoder, patch-level anomaly elimination masks are applied to reconstruct background features, where spatial correlations of anomalies are utilized to suppress anomalous patterns spanning multiple pixels. The decoder operates at the pixel level, using fine-grained receptive fields for global attention modeling, which enables the model to refine local details that may have been aggregated by the encoder in larger patches, ensuring the final reconstruction retains the intricate structure of the original hyperspectral data. Furthermore, adaptive weight loss is incorporated to guide network training. Extensive experimental results confirm the superior performance of GIFormer.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"2850-2863"},"PeriodicalIF":5.3,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11320282","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982143","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}
Robust detection of small objects in remote sensing imagery remains a significant challenge due to complex backgrounds, scale variation, and modality inconsistency. In this article, we propose DARFNet, a novel multispectral detection framework that effectively integrates RGB and infrared information for accurate small object localization. DARFNet employs a dual-branch architecture with a dynamic attention-based fusion mechanism to adaptively enhance complementary features. In addition, we incorporate lightweight yet expressive modules–ODConv and ConvNeXtBlock–to boost detection performance while maintaining computational efficiency. Extensive experiments on three widely-used benchmarks, including VEDAI, NWPU, and DroneVehicle, demonstrate that DARFNet outperforms state-of-the-art methods in both accuracy and efficiency. Notably, our model shows superior performance in detecting small and densely distributed targets under complex remote sensing conditions.
{"title":"DARFNet: A Divergence-Aware Reciprocal Fusion Network for Multispectral Feature Alignment and Fusion","authors":"Junyu Huang;Jiawei Chen;Renbo Luo;Yongan Lu;Jinxin Yang;Zhifeng Wu","doi":"10.1109/JSTARS.2025.3647819","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3647819","url":null,"abstract":"Robust detection of small objects in remote sensing imagery remains a significant challenge due to complex backgrounds, scale variation, and modality inconsistency. In this article, we propose DARFNet, a novel multispectral detection framework that effectively integrates RGB and infrared information for accurate small object localization. DARFNet employs a dual-branch architecture with a dynamic attention-based fusion mechanism to adaptively enhance complementary features. In addition, we incorporate lightweight yet expressive modules–ODConv and ConvNeXtBlock–to boost detection performance while maintaining computational efficiency. Extensive experiments on three widely-used benchmarks, including VEDAI, NWPU, and DroneVehicle, demonstrate that DARFNet outperforms state-of-the-art methods in both accuracy and efficiency. Notably, our model shows superior performance in detecting small and densely distributed targets under complex remote sensing conditions.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4779-4789"},"PeriodicalIF":5.3,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11320323","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146082028","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 : 2025-12-30DOI: 10.1109/JSTARS.2025.3649266
Jinshan Zhu;Yu Wang;Ruifu Wang;Yuquan Wen;Cong Jiao;Yina Han;Bopeng Liu
Bathymetry is a crucial topographic element in shallow water. When retrieving bathymetry using the semianalytical (SA) model, issues of too many unknown parameters and difficulty of obtaining substrate spectrum should be addressed. In this article, a Case-I water semianalytical bathymetry retrieval model assisted by a pixel substrate spectrum (SBM-P) is proposed for multispectral images to retrieve bathymetry without prior information. First, the substrate spectrum of each pixel is obtained with the assistance of the Ice, Cloud, and land Elevation Satellite-2 and Sentinel-2 data. Second, the SA bathymetry retrieval model for Case- I water is reparametrized. Third, the optimal objective function is selected in the process of numerical optimization. Finally, the performance of the proposed SBM-P model is evaluated. Two datasets, Oahu Island (OI) and Vieques Island (VI), are prepared for experiments. Results show that the minimum distance and spectral angle matching (MS) objective function has better performance compared to the minimum distance (MD). For example, in the bright substrate of OI case, compared to MD, the root mean square error ($text{RMSE}$), mean absolute error ($text{MAE}$), and mean relative error ($text{MRE}$) of MS decrease by 2.07 m, 1.61 m, 43.1%, respectively. Compared to the generally used semianalytical bathymetry retrieval model using a fixed substrate spectrum (SBM-F), the SBM-P demonstrates improvements in the evaluation metrics: for the OI case, the $text{RMSE}$ decreases by 0.75 m, the $text{MAE}$ by 0.67 m, and the $text{MRE}$ by 40.1% ; similarly for the VI case, the $text{RMSE}$$text{MAE}$, and $text{MRE}$ reduce by 0.84, 0.7, and 19% . In conclusion, the proposed SBM-P model is effective and can achieve higher accuracy compared to the generally used SBM-F model.
水深测量是浅水地形的重要组成部分。当使用半解析（SA）模型检索水深测量时，应解决未知参数过多和难以获得底物光谱的问题。本文提出了一种基于像元底谱辅助的Case-I型水半解析水深检索模型，用于无先验信息的多光谱图像水深检索。首先，在冰、云和陆地高程卫星2号和哨兵2号数据的帮助下，获得每个像元的基底光谱。其次，对Case- I水的SA水深反演模型进行了重新参数化。第三，在数值优化过程中选择最优目标函数。最后，对所提出的SBM-P模型进行了性能评价。准备了瓦胡岛（OI）和别克斯岛（VI）两个数据集进行实验。结果表明，最小距离与光谱角匹配目标函数（MS）比最小距离匹配目标函数（MD）具有更好的性能。例如，在OI情况下，与MD相比，MS的均方根误差（$text{RMSE}$）、平均绝对误差（$text{MAE}$）和平均相对误差（$text{MRE}$）分别降低了2.07 m、1.61 m和43.1%。与使用固定底物谱（SBM-F）的常用半解析测深检索模型相比，SBM-P在评价指标上有所改进：对于OI情况，$text{RMSE}$降低了0.75 m, $text{MAE}$降低了0.67 m, $text{MRE}$降低了40.1%；同样，对于VI的情况，$text{RMSE}$ $text{MAE}$和$text{MRE}$分别减少了0.84、0.7和19%。综上所述，与常用的SBM-F模型相比，所提出的SBM-P模型是有效的，可以达到更高的精度。
{"title":"Sentinel-2 Multispectral Imagery Case-I Water Semianalytical Bathymetry Retrieval Model Assisted by Satellite-Derived Pixel Substrate Spectrum","authors":"Jinshan Zhu;Yu Wang;Ruifu Wang;Yuquan Wen;Cong Jiao;Yina Han;Bopeng Liu","doi":"10.1109/JSTARS.2025.3649266","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3649266","url":null,"abstract":"Bathymetry is a crucial topographic element in shallow water. When retrieving bathymetry using the semianalytical (SA) model, issues of too many unknown parameters and difficulty of obtaining substrate spectrum should be addressed. In this article, a Case-I water semianalytical bathymetry retrieval model assisted by a pixel substrate spectrum (SBM-P) is proposed for multispectral images to retrieve bathymetry without prior information. First, the substrate spectrum of each pixel is obtained with the assistance of the Ice, Cloud, and land Elevation Satellite-2 and Sentinel-2 data. Second, the SA bathymetry retrieval model for Case- I water is reparametrized. Third, the optimal objective function is selected in the process of numerical optimization. Finally, the performance of the proposed SBM-P model is evaluated. Two datasets, Oahu Island (OI) and Vieques Island (VI), are prepared for experiments. Results show that the minimum distance and spectral angle matching (MS) objective function has better performance compared to the minimum distance (MD). For example, in the bright substrate of OI case, compared to MD, the root mean square error (<inline-formula><tex-math>$text{RMSE}$</tex-math></inline-formula>), mean absolute error (<inline-formula><tex-math>$text{MAE}$</tex-math></inline-formula>), and mean relative error (<inline-formula><tex-math>$text{MRE}$</tex-math></inline-formula>) of MS decrease by 2.07 m, 1.61 m, 43.1%, respectively. Compared to the generally used semianalytical bathymetry retrieval model using a fixed substrate spectrum (SBM-F), the SBM-P demonstrates improvements in the evaluation metrics: for the OI case, the <inline-formula><tex-math>$text{RMSE}$</tex-math></inline-formula> decreases by 0.75 m, the <inline-formula><tex-math>$text{MAE}$</tex-math></inline-formula> by 0.67 m, and the <inline-formula><tex-math>$text{MRE}$</tex-math></inline-formula> by 40.1% ; similarly for the VI case, the <inline-formula><tex-math>$text{RMSE}$</tex-math></inline-formula> <inline-formula><tex-math>$text{MAE}$</tex-math></inline-formula>, and <inline-formula><tex-math>$text{MRE}$</tex-math></inline-formula> reduce by 0.84, 0.7, and 19% . In conclusion, the proposed SBM-P model is effective and can achieve higher accuracy compared to the generally used SBM-F model.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4133-4150"},"PeriodicalIF":5.3,"publicationDate":"2025-12-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11319153","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026426","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 : 2025-12-30DOI: 10.1109/JSTARS.2025.3649267
Zilin Xie;Jinbao Jiang;Kangning Li;Xiaojun Qiao;Jinzhong Yang
An open-pit mine semantic change detection (SCD) using high-resolution remote sensing images is a critical task for both mineral resource management and environmental protection. Traditional approaches usually rely on land cover classification to indirectly SCD, a process that often introduces cumulative errors and consequently limits their performance and robustness. While advanced SCD methods using multitask architectures have demonstrated strong performance in other domains, their application to open-pit mines remains unexplored. Moreover, these methods face challenges, including inference conflicts among subtasks, a lack of semantic segmentation labels for unchanged areas, and insufficient exploration of model lightweighting. Therefore, a novel lightweight semantic reasoning and fusion network (LSRFNet) is introduced for open-pit mine SCD. LSRFNet leverages a lightweight convolutional backbone within a multitask framework. Moreover, an improved multitask fusion architecture is proposed, building upon existing multitask frameworks to explicitly optimize the final SCD output by fusing subtask predictions at the decision level, thereby mitigating inference conflicts. Furthermore, a semantic reasoning loss is designed based on pseudolabeling semisupervised learning and the local semantic consistency of land cover. By generating pseudolabels and applying local semantic consistency constraints, LSRFNet can iteratively self-train and progressively infer semantic information in unchanged areas. Experiments confirm that LSRFNet achieves state-of-the-art performance on the open-pit mine SCD task, with OA, mIoU, Sek, and Fscd values of 98.12%, 83.86%, 10.89%, and 82.13%, respectively, with only about 1/50 of the parameters and inference time compared with mainstream SCD methods. LSRFNet shows significance for open-pit mine SCD in high-resolution remote sensing images.
{"title":"A Lightweight Semantic Reasoning and Fusion Network for Open-Pit Mine Semantic Change Detection in High-Resolution Remote Sensing Images","authors":"Zilin Xie;Jinbao Jiang;Kangning Li;Xiaojun Qiao;Jinzhong Yang","doi":"10.1109/JSTARS.2025.3649267","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3649267","url":null,"abstract":"An open-pit mine semantic change detection (SCD) using high-resolution remote sensing images is a critical task for both mineral resource management and environmental protection. Traditional approaches usually rely on land cover classification to indirectly SCD, a process that often introduces cumulative errors and consequently limits their performance and robustness. While advanced SCD methods using multitask architectures have demonstrated strong performance in other domains, their application to open-pit mines remains unexplored. Moreover, these methods face challenges, including inference conflicts among subtasks, a lack of semantic segmentation labels for unchanged areas, and insufficient exploration of model lightweighting. Therefore, a novel lightweight semantic reasoning and fusion network (LSRFNet) is introduced for open-pit mine SCD. LSRFNet leverages a lightweight convolutional backbone within a multitask framework. Moreover, an improved multitask fusion architecture is proposed, building upon existing multitask frameworks to explicitly optimize the final SCD output by fusing subtask predictions at the decision level, thereby mitigating inference conflicts. Furthermore, a semantic reasoning loss is designed based on pseudolabeling semisupervised learning and the local semantic consistency of land cover. By generating pseudolabels and applying local semantic consistency constraints, LSRFNet can iteratively self-train and progressively infer semantic information in unchanged areas. Experiments confirm that LSRFNet achieves state-of-the-art performance on the open-pit mine SCD task, with OA, mIoU, Sek, and F<sub>scd</sub> values of 98.12%, 83.86%, 10.89%, and 82.13%, respectively, with only about 1/50 of the parameters and inference time compared with mainstream SCD methods. LSRFNet shows significance for open-pit mine SCD in high-resolution remote sensing images.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"5618-5633"},"PeriodicalIF":5.3,"publicationDate":"2025-12-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11319156","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175856","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 monitoring of wetland vegetation inundation is crucial for maintaining regional ecological balance and conserving biodiversity, serving as a fundamental prerequisite for wetland environmental monitoring and protection. The complex scattering characteristics of vegetation under different inundation conditions, combined with spatial and seasonal heterogeneity, pose significant challenges to precise vegetation inundation state identification. Therefore, this study proposes a novel approach named the linear-exponential model, shapelets, and multirocket integration (LESMI), for monitoring the inundation state and temporal changes of wetland vegetation using radar backscatter variation patterns. First, a new linear-exponential model is developed to characterize the backscatter-water depth relationship and represent the inundation state characteristics of wetland vegetation. Second, based on the typical inundated state of historical stages determined by the linear-exponential model, LESMI method innovatively combines the Shapelets with multirocket classification to efficiently extract multivariate key time periods features for inundation state identification and achieve large-scale, near real-time inundation state classification. Experimental results in the Dongting Lake wetland show that the proposed method achieves inundation recognition accuracies of 96.84% for reeds and 92.59% for grassland, outperforming traditional methods and LSTM deep learning by average margins of 12.95% and 1.87%, respectively. The linear-exponential model significantly enhances identification performance, improving accuracy by 5.64% and 3.83% compared to linear and normal distribution models. Monitoring from 2019 to 2021 demonstrates that LESMI effectively captures flood peak impacts on vegetation inundation and provides detailed classification of noninundated, shallow inundated, and deep inundated states, offering reliable technical support for dynamic wetland ecosystem monitoring and refined management.
{"title":"LESMI: Integrating Linear-Exponential Model, Shapelets, and Multirocket for Wetland Vegetation Inundation Monitoring With Time Series SAR","authors":"Yuanye Cao;Xiuguo Liu;Yuannan Long;Hui Yang;Shixiong Yan;Qihao Chen","doi":"10.1109/JSTARS.2025.3649200","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3649200","url":null,"abstract":"Accurate monitoring of wetland vegetation inundation is crucial for maintaining regional ecological balance and conserving biodiversity, serving as a fundamental prerequisite for wetland environmental monitoring and protection. The complex scattering characteristics of vegetation under different inundation conditions, combined with spatial and seasonal heterogeneity, pose significant challenges to precise vegetation inundation state identification. Therefore, this study proposes a novel approach named the linear-exponential model, shapelets, and multirocket integration (LESMI), for monitoring the inundation state and temporal changes of wetland vegetation using radar backscatter variation patterns. First, a new linear-exponential model is developed to characterize the backscatter-water depth relationship and represent the inundation state characteristics of wetland vegetation. Second, based on the typical inundated state of historical stages determined by the linear-exponential model, LESMI method innovatively combines the Shapelets with multirocket classification to efficiently extract multivariate key time periods features for inundation state identification and achieve large-scale, near real-time inundation state classification. Experimental results in the Dongting Lake wetland show that the proposed method achieves inundation recognition accuracies of 96.84% for reeds and 92.59% for grassland, outperforming traditional methods and LSTM deep learning by average margins of 12.95% and 1.87%, respectively. The linear-exponential model significantly enhances identification performance, improving accuracy by 5.64% and 3.83% compared to linear and normal distribution models. Monitoring from 2019 to 2021 demonstrates that LESMI effectively captures flood peak impacts on vegetation inundation and provides detailed classification of noninundated, shallow inundated, and deep inundated states, offering reliable technical support for dynamic wetland ecosystem monitoring and refined management.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4242-4256"},"PeriodicalIF":5.3,"publicationDate":"2025-12-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11319178","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026501","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 : 2025-12-30DOI: 10.1109/JSTARS.2025.3649548
Xuan Liu;Lina Cai;Jiahua Li;Tianle Mao
This study examined a novel harmful algal bloom (HAB) inversion model (HABI) using domestic Chinese ocean color and temperature scanner multispectral data from HY-1C/D satellites. The model achieves the dual capabilities of HAB presence detection and density quantification, a key advancement over conventional binary classification models that lack the ability to delineate HAB density gradients. Key findings of this article include the following. 1) The HABI model uses spectral bands at 443, 490, and 565 nm, demonstrating superior performance in quantifying HAB density gradients compared to existing methods, with design adaptability to sensors featuring similar spectral configurations. 2) HABI achieved high inversion accuracy (R2 = 0.8682, RMSE = 0.09195, Recall = 0.9300, Precision = 0.949, F1-score = 0.939), showing strong consistency with Bulletin of China Marine Disaster and in situ HAB measurements in the Waters near the Yangtze River Estuary. 3) The distribution of HABs takes on obvious temporal and spatial change characteristics, with high density clusters localized in coastal zones, peaking in spring/summer, and changed seasonally. Their seasonal factors contributing to the change of HAB mainly include Yangtze River freshwater discharge and coastal upwelling, and modulated by physical (e.g., sea surface temperature), anthropogenic (e.g., industrial wastewater), and biogeochemical factors (e.g., dissolved inorganic nitrogen) as well as biodiversity. These findings are conceptually integrated in Fig. 14, synthesizing the model mechanics and spatio-temporal dynamics. The HABI algorithm proposed in this article can effectively applied for HAB monitoring and quantification, providing a technical support for near-shore ecological assessment and management.
{"title":"Study on Harmful Algal Blooms in the Waters Near the Yangtze River Estuary Based on Twin Satellites HY-1C/D COCTS Data","authors":"Xuan Liu;Lina Cai;Jiahua Li;Tianle Mao","doi":"10.1109/JSTARS.2025.3649548","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3649548","url":null,"abstract":"This study examined a novel harmful algal bloom (HAB) inversion model (HABI) using domestic Chinese ocean color and temperature scanner multispectral data from HY-1C/D satellites. The model achieves the dual capabilities of HAB presence detection and density quantification, a key advancement over conventional binary classification models that lack the ability to delineate HAB density gradients. Key findings of this article include the following. 1) The HABI model uses spectral bands at 443, 490, and 565 nm, demonstrating superior performance in quantifying HAB density gradients compared to existing methods, with design adaptability to sensors featuring similar spectral configurations. 2) HABI achieved high inversion accuracy (<italic>R</i><sup>2</sup> = 0.8682, RMSE = 0.09195, Recall = 0.9300, Precision = 0.949, F1-score = 0.939), showing strong consistency with Bulletin of China Marine Disaster and in situ HAB measurements in the Waters near the Yangtze River Estuary. 3) The distribution of HABs takes on obvious temporal and spatial change characteristics, with high density clusters localized in coastal zones, peaking in spring/summer, and changed seasonally. Their seasonal factors contributing to the change of HAB mainly include Yangtze River freshwater discharge and coastal upwelling, and modulated by physical (e.g., sea surface temperature), anthropogenic (e.g., industrial wastewater), and biogeochemical factors (e.g., dissolved inorganic nitrogen) as well as biodiversity. These findings are conceptually integrated in Fig. 14, synthesizing the model mechanics and spatio-temporal dynamics. The HABI algorithm proposed in this article can effectively applied for HAB monitoring and quantification, providing a technical support for near-shore ecological assessment and management.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"4872-4886"},"PeriodicalIF":5.3,"publicationDate":"2025-12-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11319152","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146082029","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 : 2025-12-26DOI: 10.1109/JSTARS.2025.3648330
Mengmeng Wang;Xu Lin;Yuanxin Ye;Wenhui Wu;Bai Zhu;Yanshuai Dai
Change detection (CD) is a fundamental task that is pivotal in understanding surface changes. Recently, CD methods have advanced rapidly and attained impressive results, driven by deep learning technology. However, existing methods generally employ fusion modules with the same design for multilevel features, overlooking the inherent distinctions between low-level spatial features and deep-level semantic features generated by deep networks. To overcome this limitation, this article proposes a novel CD network, referred to as DACNet. This method introduces a divide-and-conquer fusion strategy designed to fuse multilevel features using different fusion strategies. Specifically, the widely used MobileNetV2 is employed within a dual-branch architecture to extract multilevel features from bitemporal images. Subsequently, the proposed divide-and-conquer fusion strategy comprises two specialized modules: the change region localization module and the edge complementarity module, which are tailored to fuse deep-level semantic features and low-level spatial features, respectively. In addition, to mitigate the unnecessary noise introduced by the conventional UNet architectures, attention gates are introduced into the UNet decoder to enhance the changed information and suppress background noises. Extensive experiments are conducted on three available CD datasets: LEVIR-CD, Google-CD, and MSRS-CD. The proposed network achieved favorable results compared to the nine state-of-the-art methods across all experiments, improving the F1 score by 0.93%, 1.10%, and 0.81% on the LEVIR-CD, Google-CD, and MSRS-CD datasets, respectively.
{"title":"A Novel Network for Change Detection Based on a Divide-and-Conquer Fusion Strategy","authors":"Mengmeng Wang;Xu Lin;Yuanxin Ye;Wenhui Wu;Bai Zhu;Yanshuai Dai","doi":"10.1109/JSTARS.2025.3648330","DOIUrl":"https://doi.org/10.1109/JSTARS.2025.3648330","url":null,"abstract":"Change detection (CD) is a fundamental task that is pivotal in understanding surface changes. Recently, CD methods have advanced rapidly and attained impressive results, driven by deep learning technology. However, existing methods generally employ fusion modules with the same design for multilevel features, overlooking the inherent distinctions between low-level spatial features and deep-level semantic features generated by deep networks. To overcome this limitation, this article proposes a novel CD network, referred to as DACNet. This method introduces a divide-and-conquer fusion strategy designed to fuse multilevel features using different fusion strategies. Specifically, the widely used MobileNetV2 is employed within a dual-branch architecture to extract multilevel features from bitemporal images. Subsequently, the proposed divide-and-conquer fusion strategy comprises two specialized modules: the change region localization module and the edge complementarity module, which are tailored to fuse deep-level semantic features and low-level spatial features, respectively. In addition, to mitigate the unnecessary noise introduced by the conventional UNet architectures, attention gates are introduced into the UNet decoder to enhance the changed information and suppress background noises. Extensive experiments are conducted on three available CD datasets: LEVIR-CD, Google-CD, and MSRS-CD. The proposed network achieved favorable results compared to the nine state-of-the-art methods across all experiments, improving the F1 score by 0.93%, 1.10%, and 0.81% on the LEVIR-CD, Google-CD, and MSRS-CD datasets, respectively.","PeriodicalId":13116,"journal":{"name":"IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing","volume":"19 ","pages":"2891-2904"},"PeriodicalIF":5.3,"publicationDate":"2025-12-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11316246","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145982153","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：