[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-SAR":3},{"total":4,"page":4,"page_size":5,"items":6},1,100,[7],{"id":8,"title":9,"url":10,"summary":11,"summary_zh":12,"content":13,"source_name":14,"source_url":10,"published_at":15,"category":16,"cover_url":13,"hotness":17,"is_selected":18,"score":19,"score_detail":20,"sources":25,"tags":27,"view_count":33,"doi":34,"paper":35,"created_at":75},2786,"SAF-CropNet: Spatially adaptive fusion of SAR and optical imagery for semantic segmentation in operational cropland mapping across regions","https:\u002F\u002Fdoi.org\u002F10.1016\u002Fj.isprsjprs.2026.08.029","Accurate cropland mapping from remote sensing imagery is essential for agricultural monitoring, food security assessment, and land resource management. Although optical and synthetic aperture radar (SAR) observations provide complementary information for cropland extraction, their local contributions vary across imaging conditions, land-cover backgrounds, parcel morphologies, and regional agricultural systems. This variability limits direct concatenation and static fusion, as modality-specific errors may propagate into segmentation outputs in the absence of spatially adaptive weighting. This study proposes SAF-CropNet, a spatially adaptive SAR–optical fusion framework for binary cropland semantic segmentation. Specifically, SAF-CropNet first extracts modality-specific SAR structural features and optical spectral-textural features through separate encoder branches. The Paired Reciprocal Inter-modal Selective Modulator enhances complementary SAR–optical interactions, while the Contextual State-space Modeling module captures long-range parcel organization. A spatially adaptive fusion head then estimates per-pixel contribution weights for SAR, optical, and contextual features before decoding. Evaluation was conducted on a seven-area SAR–optical benchmark spanning China, Germany, and France. The benchmark integrates GF1\u002FGF3 and Sentinel-1\u002F2 imagery with reference labels derived from field surveys and RapidCrops products and covers fragmented smallholder systems, water-rich and peri-urban mosaics, and large mechanized cropland. Within-region five-fold validation shows that SAF-CropNet achieves an average F1 score of 0.8550 and an mIoU of 0.8068, outperforming the evaluated segmentation baselines while remaining lightweight, with 3.38 M parameters and 6.21 GFLOPs for 256 × 256 SAR–optical inputs. In leave-one-area-out transfer experiments, stratified few-shot adaptation increased the average F1 score from 0.4986 under zero-shot transfer to 0.7698. Grouped cross-continental and cross-sensor experiments further showed that direct transfer remains constrained under compound shifts in sensor characteristics, spatial resolution, landscape structure, and reference-label conventions. These findings indicate that SAF-CropNet combines strong within-region segmentation accuracy and computational efficiency with substantial gains from limited target-domain adaptation, while direct zero-shot generalization under severe compound domain shifts remains limited.","从遥感影像中准确提取耕地信息对农业监测、粮食安全评估和土地资源管理至关重要。尽管光学与合成孔径雷达（SAR）观测为耕地提取提供了互补信息，但其局部贡献会随成像条件、地表覆盖背景、地块形态和区域农业系统的不同而变化。这种变异性限制了直接拼接和静态融合的效果，因为在缺乏空间自适应加权的情况下，模态特有的误差可能传播至分割输出中。本研究提出SAF-CropNet，一种用于二分类耕地语义分割的空间自适应SAR–光学融合框架。具体而言，SAF-CropNet首先通过独立的编码器分支提取模态特有的SAR结构特征和光学光谱-纹理特征。成对互反模态间选择性调制器增强SAR与光学之间的互补交互，而上下文状态空间建模模块则捕获长距离地块组织信息。随后，空间自适应融合头在解码前估计SAR、光学和上下文特征的逐像素贡献权重。评估在一个涵盖中国、德国和法国的七区域SAR–光学基准上进行。该基准整合了GF1\u002FGF3和Sentinel-1\u002F2影像，参考标签来源于实地调查和RapidCrops产品，覆盖破碎化小农系统、富水与城郊镶嵌景观以及大规模机械化耕地。区域内五折验证表明，SAF-CropNet平均F1分数达到0.8550，mIoU为0.8068，优于所评估的分割基线，同时保持轻量级，对于256 × 256的SAR–光学输入仅需3.38 M参数和6.21 GFLOPs。在留一区域迁移实验中，分层少样本自适应将平均F1分数从零样本迁移下的0.4986提升至0.7698。分组跨大陆和跨传感器实验进一步表明，在传感器特性、空间分辨率、景观结构和参考标签惯例的复合变化下，直接迁移仍然受到限制。这些发现表明，SAF-CropNet兼具较强的区域内分割精度和计算效率，并能从有限的目标域自适应中获得显著增益，而在严重复合域偏移下的直接零样本泛化能力仍然有限。",null,"ISPRS Journal of Photogrammetry and Remote Sensing","2026-09-17T00:00:00Z","论文",10,false,82,{"impact":21,"substance":22,"depth":21,"authority":23,"freshness":17,"relevant":4,"comment":24},18,22,14,"提出空间自适应SAR-光学融合分割框架，跨七区域基准验证，兼顾精度与轻量化，对耕地遥感制图有实质方法贡献。",[26],{"name":14,"url":10},[28,29,30,31,32],"农业人工智能","遥感","耕地监测","作物分类","SAR",0,"10.1016\u002Fj.isprsjprs.2026.08.029",{"doi":34,"openalex_id":36,"authors":37,"venue":14,"cited_by_count":33,"oa_url":10,"card":68,"direction":72,"ingested_from":74},"W7213463520",[38,41,43,46,49,52,55,57,59,61,63,65],{"name":39,"orcid":40},"minghui chang","https:\u002F\u002Forcid.org\u002F0009-0008-7954-5308",{"name":42,"orcid":13},"Shuaifeng Peng",{"name":44,"orcid":45},"Tao Xu","https:\u002F\u002Forcid.org\u002F0000-0002-7855-4199",{"name":47,"orcid":48},"Yi Yuan","https:\u002F\u002Forcid.org\u002F0009-0006-9481-4938",{"name":50,"orcid":51},"洋一 馬目","https:\u002F\u002Forcid.org\u002F0000-0003-0226-5558",{"name":53,"orcid":54},"Jie Bai","https:\u002F\u002Forcid.org\u002F0000-0003-2426-2358",{"name":56,"orcid":13},"Fugui Luo",{"name":58,"orcid":13},"Jingyu Zhang",{"name":60,"orcid":13},"Xiaoyu Xiao",{"name":62,"orcid":13},"Yu Mu",{"name":64,"orcid":13},"Yong Wang",{"name":66,"orcid":67},"Shihua Li","https:\u002F\u002Forcid.org\u002F0000-0003-4807-5012",{"tldr":69,"method":70,"finding":71,"direction":72,"opportunity":73},"提出SAF-CropNet，用空间自适应融合SAR与光学影像做跨区域耕地语义分割。","双分支编码器提取SAR结构\u002F光学纹理特征，结合互模态调制与状态空间建模，在七区域","区域内F1达0.8550且轻量；零样本迁移F1仅0.4986，少样本适配可升至0.7698。","农业遥感与作物表型","复合域偏移下零样本泛化差，可研究跨传感器、跨区域的无监督域自适应融合分割方法。","openalex","2026-09-17T23:30:34.424579Z"]