[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"item-464":3,"related-464":51},{"id":4,"title":5,"url":6,"summary":7,"summary_zh":8,"content":9,"source_name":10,"source_url":8,"published_at":11,"category":12,"cover_url":8,"hotness":13,"is_selected":14,"score":15,"score_detail":16,"sources":24,"tags":26,"primary_topic":8,"secondary_topics":31,"classification":32,"search_phrases":45,"slug":48,"view_count":49,"doi":8,"paper":8,"created_at":50},464,"Comparative Analysis of Automatic Mask Generation Methods for Agricultural Seed Instance Segmentation","https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41598-026-66297-4","本文针对农业图像分析中像素级标注耗时费力的问题,比较了5种无需领域定制的自动掩码生成方法用于豌豆种子实例分割:Cellpose、EfficientSAM、SAM2、StarDist、FastSAM。研究使用164张图像数据集(100张手工标注含33462个实例,64张空白用于下游训练),固定测试集20张含6849个实例。结果显示,Cellpose在重叠豌豆尺寸颗粒上整体表现最佳(F1:0.957,Recall:0.973);EfficientSAM精度最高(mean IoU:0.970);SAM2召回高(0.965)但精度低(0.815);StarDist因形状约束在重叠颗粒上表现较差;FastSAM漏检密集簇中的颗粒。基于Cellpose自动标注的144张图像训练YOLO11m-seg模型与80张手工标注训练的模型性能接近(mAP50:0.848 vs 0.860),证明自动标注是密集农业影像中人工标注的实用替代品。",null,"*   Article\n*   [Open access](https:\u002F\u002Fwww.springernature.com\u002Fgp\u002Fopen-science\u002Fabout\u002Fthe-fundamentals-of-open-access-and-open-research)\n*   Published: 08 August 2026\n\n*   [Yavuz Unal](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41598-026-66297-4#auth-Yavuz-Unal-Aff1)[1](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41598-026-66297-4#Aff1)&\n*   [Yonis Gulzar](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41598-026-66297-4#auth-Yonis-Gulzar-Aff2)[ORCID: orcid.org\u002F0000-0002-6515-1569](https:\u002F\u002Forcid.org\u002F0000-0002-6515-1569)[2](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41598-026-66297-4#Aff2)\n\n[_Scientific Reports_](https:\u002F\u002Fwww.nature.com\u002Fsrep) (2026) [Cite this article](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41598-026-66297-4#citeas)\n\nWe are providing an unedited version of this manuscript to give early access to its findings. Before final publication, the manuscript will undergo further editing. Please note there may be errors present which affect the content, and all legal disclaimers apply.\n\n## Abstract\n\nIn agricultural image analysis, training deep learning models relies on large amounts of pixel-level labeled data. However, individually labeling densely stacked and overlapping objects, such as pea seeds, is both time-consuming and laborious. This study compares five automated mask-generation methods for instance segmentation of pea seeds (Pisum sativum) that do not require domain-specific customization: Cellpose, EfficientSAM, SAM2, StarDist, and FastSAM. A dataset of 164 images was used in the study. 100 of these images were manually annotated by a trained annotator using the LabelMe tool, containing a total of 33,462 instance labels. The remaining 64 images were left unlabeled and evaluated for automated labeling in downstream model training. All methods were evaluated on a fixed testset of 20 images containing 6,849 instances using precision, recall, F1 score, and mean IoU metrics. The results show that Cellpose, developed for biomedical cell segmentation, achieved the highest overall performance on overlapping pea-sized particles (F1: 0.957, Recall: 0.973). EfficientSAM produced the most precise pixel-level masks (mean IoU: 0.970), while SAM2 exhibited high recall (0.965) but low precision (0.815). StarDist struggled with overlapping particles due to shape constraints (mean IoU: 0.866), while FastSAM missed many particles in dense clusters, achieving high precision (0.976) but low recall (0.725). Furthermore, the YOLO11m-seg model, trained on 144 images automatically labeled with Cellpose tags, performed nearly as well as the model trained on only 80 manually labeled images (mAP50: 0.848 vs. 0.860). These findings suggest that automated labeling could be a practical alternative to manual labeling, even in challenging agricultural imagery containing overlapping objects.\n\n## Funding\n\nThis work was supported by the Deanship of Scientific Research, the Vice Presidency for Graduate Studies and Scientific Research, King Faisal University, Saudi Arabia, under the project number KFU263615.\n\n## Author information\n\n### Authors and Affiliations\n\n1.   Faculty of Engineering and Architecture, Sinop University, Sinop, Turkey\n\nYavuz Unal\n\n2.   Department of Management Information Systems, College of Business Administration, King Faisal University, Al-Ahsa, 31982, Saudi Arabia\n\nYonis Gulzar\n\nAuthors\n\n1.   Yavuz Unal\n2.   Yonis Gulzar\n\n### Corresponding authors\n\nCorrespondence to [Yavuz Unal](mailto:yunal@sinop.edu.tr) or [Yonis Gulzar](mailto:ygulzar@kfu.edu.sa).\n\n## Ethics declarations\n\n### Competing interests\n\nThe authors declare no competing interests.\n\n### Ethics statement\n\nThis study does not involve human participants or human subjects. The annotation process was performed by a trained annotator on image data of agricultural seeds. Therefore, informed consent was not required.\n\n## Additional information\n\n### Publisher’s note\n\nSpringer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.\n\n## Rights and permissions\n\n**Open Access** This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit [http:\u002F\u002Fcreativecommons.org\u002Flicenses\u002Fby-nc-nd\u002F4.0\u002F](http:\u002F\u002Fcreativecommons.org\u002Flicenses\u002Fby-nc-nd\u002F4.0\u002F).\n\n[Reprints and permissions](https:\u002F\u002Fs100.copyright.com\u002FAppDispatchServlet?title=Comparative%20analysis%20of%20automatic%20mask%20generation%20methods%20for%20agricultural%20seed%20instance%20segmentation&author=Yavuz%20Unal%20et%20al&contentID=10.1038%2Fs41598-026-66297-4&copyright=The%20Author%28s%29&publication=2045-2322&publicationDate=2026-08-08&publisherName=SpringerNature&orderBeanReset=true&oa=CC%20BY-NC-ND)\n\n## About this article\n\n[![Image 1: Check for updates. Verify currency and authenticity via CrossMark](blob:http:\u002F\u002Flocalhost\u002Fdf43c82f6cb6dcd2a87db38f317bf9f2)](https:\u002F\u002Fcrossmark.crossref.org\u002Fdialog\u002F?doi=10.1038\u002Fs41598-026-66297-4)\n\n### Cite this article\n\nUnal, Y., Gulzar, Y. Comparative analysis of automatic mask generation methods for agricultural seed instance segmentation. _Sci Rep_ (2026). https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41598-026-66297-4\n\n[Download citation](https:\u002F\u002Fcitation-needed.springer.com\u002Fv2\u002Freferences\u002F10.1038\u002Fs41598-026-66297-4?format=refman&flavour=citation)\n\n*   Received: 20 May 2026\n\n*   Accepted: 06 August 2026\n\n*   Published: 08 August 2026\n\n*   DOI: https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41598-026-66297-4\n\n### Keywords","Nature Scientific Reports 2026-08-08","2026-08-08T00:00:00Z","论文",10,false,68,{"impact":17,"substance":18,"depth":19,"authority":20,"freshness":21,"relevant":22,"comment":23},12,20,18,14,4,1,"论文比较了多种自动掩膜生成方法在农业种子实例分割中的性能，数据详实，结论具有实用价值，但时效性稍差。",[25],{"name":10,"url":6},[27,28,29,30],"实例分割","自动标注","农业图像","SAM",[],{"status":33,"locked":14,"version":34,"method":35,"evidence":36},"unclassified","2026-10-22.2","rules",{"candidates":37,"agricultural_evidence":43,"reason":44},[38],{"code":39,"score":22,"supported":14,"title":40,"abstract":41,"tags":42},"I2",[],[],[27],true,"缺少足够的主题证据",[46,47],"农业图像 实例分割 自动标注 SAM","农业图像 实例分割","农业图像实例分割自动标注SAM-464",3,"2026-08-12T23:57:24.231880Z",{"total":52,"page":22,"page_size":52,"items":53},2,[54,129],{"id":55,"title":56,"url":57,"summary":58,"summary_zh":59,"content":8,"source_name":60,"source_url":57,"published_at":61,"category":12,"cover_url":8,"hotness":13,"is_selected":14,"score":62,"score_detail":63,"sources":69,"tags":71,"primary_topic":76,"secondary_topics":81,"classification":82,"search_phrases":107,"slug":110,"view_count":111,"doi":112,"paper":113,"created_at":128},5075,"An integrated instance segmentation and deep feature fusion model for robust leaf species classification in natural environments","https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffpls.2026.1893779","Introduction The recognition of leaf segments and species in nature is a difficult process due to the changes in brightness, too much clutter in the background, and the overlapping of leaves due to which the accuracy falls down for 20-40% when using laboratory scans instead of field-ready applications. Methods This research presents a unified framework of deep learning that includes Mask R-CNN -based instance segmentation and deep feature fusion in order to overcome this problem. The process of extracting a number of leaves took place through Detectron2 and the excessive stream extractor that included ResNet-50 and VGG-16 used in conjunction with other descriptive methods allowed providing a joint representation of 6,148 dimensions. The classifier was trained according to three methods: Baseline, PSO (particle swarm optimization), and Grid search. Results The masked fusion system cut off the noise in the background while retaining the important botanical features with the help of achieving the mean Intersection-over-Union equal to 0.752, detection precision of 98.39%, and mAP@0.5 of 0.8291. The ResNet-50 model enhanced with PSO techniques has the best performance among other technologies based on ResNet-50 (90% accuracy, F1 score 0.9002), while the technology of Grid search revealed a little higher result for the special techniques such as ResNet-50 plus traditional features (F1 score = 0.9135) which means that the efficiency of the methodology is contingent to the type of the features not the fact if the technology is better than the other. Discussion Comparative and ablation data analysis proved the efficiency of the instance segmentation used in combination with multi-mode fusion especially when used in occlusion-prone conditions. The system offers a scalable application for smart agriculture and biodiversity monitoring while being a good example of bridging the gap in automated environmental intelligence.","引言 由于亮度变化、背景杂乱以及叶片重叠等因素，在自然环境中识别叶片片段和物种是一项困难的过程，当使用实验室扫描而非野外应用时，准确率会下降20-40%。方法 本研究提出了一个统一的深度学习框架，包括基于Mask R-CNN（掩码区域卷积神经网络）的实例分割和深度特征融合，以克服这一问题。通过Detectron2和包含ResNet-50与VGG-16的过度流提取器，结合其他描述性方法，提取了多个叶片，提供了6，148维的联合表示。分类器根据三种方法进行训练：基线、PSO（粒子群优化）和网格搜索。结果 掩码融合系统在保留重要植物学特征的同时消除了背景噪声，实现了平均交并比为0.752、检测精度为98.39%、mAP@0.5为0.8291。经PSO技术增强的ResNet-50模型在基于ResNet-50的其他技术中表现最佳（准确率90%，F1分数0.9002），而网格搜索技术在ResNet-50加传统特征等特定技术中显示出略高的结果（F1分数=0.9135），这意味着该方法的效率取决于特征类型，而非某种技术是否优于其他技术。讨论 比较和消融数据分析证明了实例分割与多模态融合相结合的有效性，尤其是在易遮挡条件下。该系统为智能农业和生物多样性监测提供了可扩展的应用，同时是弥合自动化环境智能差距的良好范例。","Frontiers in Plant Science","2026-10-15T00:00:00Z",74,{"impact":64,"substance":18,"depth":65,"authority":66,"freshness":67,"relevant":22,"comment":68},15,17,13,9,"方法新颖、数据扎实的叶片识别论文，对智慧农业与生物多样性监测有参考价值，但属细分技术进展，非重大突破。",[70],{"name":60,"url":57},[72,73,27,74,75],"智慧农业","农业人工智能","叶片识别","生物多样性监测",{"code":39,"slug":77,"name":78,"path":79,"parent_code":80},"agricultural-ai","农业AI与大模型","\u002Ftopic\u002Fdigital-agriculture\u002Fagricultural-ai","I",[],{"status":83,"locked":14,"version":34,"method":35,"evidence":84},"classified",{"candidates":85,"agricultural_evidence":43},[86,93,101],{"code":39,"score":87,"supported":43,"title":88,"abstract":89,"tags":92},8,[],[90,91],"深度学习","deep learning",[73,27],{"code":94,"score":95,"supported":43,"title":96,"abstract":97,"tags":100},"H6",6,[],[98,99],"生物多样性","biodiversity",[],{"code":102,"score":21,"supported":43,"title":103,"abstract":104,"tags":106},"I1",[],[105],"smart agriculture",[72],[108,109],"Mask R-CNN 叶片识别","ResNet-50 粒子群优化 叶片分类","MaskR-CNN叶片识别-5075",0,"10.3389\u002Ffpls.2026.1893779",{"doi":112,"openalex_id":114,"authors":115,"venue":60,"cited_by_count":111,"oa_url":57,"card":120,"direction":126,"ingested_from":127},"W7222005311",[116,118],{"name":117,"orcid":8},"K. Arunkumar",{"name":119,"orcid":8},"S. Leninisha",{"tldr":121,"method":122,"finding":123,"direction":124,"opportunity":125},"提出实例分割与深度特征融合框架，在自然环境下实现鲁棒叶片物种分类。","Mask R-CNN分割叶片，ResNet-50与VGG-16融合成6148维特","掩膜融合去背景噪声，mAP@0.5达0.8291，ResNet-50+PSO准确率90%，F1达0.","农业遥感与作物表型","可探索轻量化分割与多模态融合在移动端实时叶片识别及遮挡场景下的泛化能力。","智慧农业 \u002F 农业物联网","openalex","2026-10-16T23:30:07.962310Z",{"id":130,"title":131,"url":132,"summary":133,"summary_zh":8,"content":134,"source_name":135,"source_url":8,"published_at":136,"category":12,"cover_url":8,"hotness":13,"is_selected":14,"score":137,"score_detail":138,"sources":140,"tags":142,"primary_topic":8,"secondary_topics":146,"classification":147,"search_phrases":160,"slug":163,"view_count":67,"doi":164,"paper":8,"created_at":165},177,"基于改进YOLOv11n-seg的胡麻倒伏区域识别方法——兼顾精度与轻量化的智能化倒伏监测","https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010","针对胡麻倒伏区域识别难题，研究构建改进YOLOv11n-seg模型。改进后的模型在精确度和平均精度均值上分别达到92.6%和95.2%，较原始YOLOv11n-seg模型分别提升3.7个百分点和2.1个百分点。模型参数量仅为1.73 M，计算量为8.0 GFLOPs，模型体积仅为3.8 MB，显示出较好的轻量化优势。该方法兼顾精度与轻量化，为智能化胡麻倒伏监测与实际田间设备部署提供了技术基础。","欢迎您访问《智慧农业（中英文）》官方网站！[English](https:\u002F\u002Fwww.smartag.net.cn\u002FEN\u002F2096-8094\u002Fhome.shtml)\n\n[![Image 1](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002F2096-8094\u002Fimages\u002Fbanner.png)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F2096-8094\u002Fhome.shtml)\n\n[![Image 2](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002F2096-8094\u002Fimages\u002Fbanner_mobile.png)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F2096-8094\u002Fhome.shtml)\n\n[![Image 3](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002F2096-8094\u002Fimages\u002Fyyzkj.png)](https:\u002F\u002Fpano.bookan.com.cn\u002F?id=60704&cid=352)\n\nToggle navigation[智慧农业](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010)\n\n*   [网站首页](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F2096-8094\u002Fhome.shtml)\n*   [期刊介绍](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fcolumn\u002Fcolumn133.shtml)\n*   [编委会](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fcolumn\u002Fcolumn153.shtml)\n*   [投稿指南](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fcolumn\u002Fcolumn135.shtml)\n*   [期刊服务](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fcolumn\u002Fcolumn124.shtml)\n*   [下载中心](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fcolumn\u002Fcolumn126.shtml)\n*   [联系我们](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fcolumn\u002Fcolumn127.shtml)\n\n[Smart Agriculture](https:\u002F\u002Fwww.smartag.net.cn\u002F) ›› [2026](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Farticle\u002FshowTenYearVolumnDetail.do?nian=2026), [Vol. 8](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Farticle\u002FshowTenYearVolumnDetail.do?nian=2026) ›› [Issue (2)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fvolumn\u002Fvolumn_838.shtml): 118-132.doi: [10.12133\u002Fj.smartag.SA202510010](https:\u002F\u002Fdoi.org\u002F10.12133\u002Fj.smartag.SA202510010)\n\n• 专题--多源遥感驱动数字农业创新与实践 •[上一篇](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fabstract\u002Fabstract22643.shtml)[下一篇](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Fabstract\u002Fabstract22641.shtml)\n\n### 面向云遮挡与复杂地块场景下的跨模态注意力多源遥感农作物分类\n\n巫晨旭, 左浩龙, 李刚([![Image 4](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Femail.png)](mailto:35987336@qq.com))\n\n1.   黑龙江工程学院测绘工程学院，黑龙江 哈尔滨 150050，中国\n\n*   `收稿日期:`2025-10-11 `出版日期:`2026-03-30 \n\n*   `作者简介:`巫晨旭，硕士研究生，研究方向为时空大数据。E-mail：[15589720597@163.com](mailto:15589720597@163.com) \n*   `通信作者:`李 刚，博士，讲师，研究方向为时空大数据。E-mail：[35987336@qq.com](mailto:35987336@qq.com) \n\n### Cross-Modal Attention for Multi-Source Remote Sensing Crop Classification under Cloud Occlusion and Complex Field Scenarios\n\nWU Chenxu, ZUO Haolong, LI Gang([![Image 5](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Femail.png)](mailto:35987336@qq.com))\n\n1.   School of Geomatics Engineering, Heilongjiang Institute of Technology, Harbin 150050, China\n\n*   `Received:`2025-10-11 `Online:`2026-03-30 \n*   `Foundation items:`Heilongjiang Province Double First-Class Discipline Coordinated Innovation Achievement Project(LJGXCG 2025-P18) \n*   `About author:`WU Chenxu, E-mail: [15589720597@163.com](mailto:15589720597@163.com) \n*   `Corresponding author:`LI Gang, E-mail: [35987336@qq.com](mailto:35987336@qq.com) \n\n## [在线阅读 24](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)\n\n## [知网下载](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010)\n\n## [本地下载 48](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#1)\n\n[_![Image 6](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Fknowledge\\_map1.png)_ ## 可视化 0](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Flexeme\u002FshowArticleByLexeme.do?articleID=22683)\n\n#### [摘要\u002FAbstract](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010)\n\n**摘要：**\n\n**【目的\u002F意义】** 本研究旨在突破传统光学遥感在云遮天气与耕地破碎地形下的局限性，通过构建一种具备优秀跨模态融合能力与泛化性能的深度网络模型来提升农业遥感分类精度。 **【方法】** 提出一种基于注意力机制的3D卷积神经网络（3D Convolutional Neural Network Based on Attention Mechanism, Attention-3DCNN）模型：其通过3D卷积+2D卷积结构处理时序哨兵二号多光谱影像，在空间与时间维度上提取丰富特征；同时以深度可分离卷积形式处理来自哨兵一号的合成孔径雷达（Synthetic Aperture Radar, SAR）数据，实现对全天候可获取信息的高效抽取；进一步，模型引入“通道-时间-空间”三重注意力机制与残差连接策略，对两个模态的特征进行动态加权与深度融合，使得在光学数据缺失或遮挡严重的情形下，SAR数据能够有效补偿并维持分类性能。 **【结果和讨论】** 为全面评价模型性能，选取法国全景农业卫星时序数据集，以及山东省沂水县实测数据集进行对比实验：在法国数据上模型达到97.5%的整体准确率，在沂水县数据上获得93%的准确率，均显著优于对照基线模型；同时，通过对注意力分布的可视化分析可见，模型聚焦的关键物候期与当地农业实地记录高度一致，其高权重光谱波段亦符合农学机理，这体现出模型在判别机制层面的可解释性。 **【结论】** 综上，Attention-3DCNN模型在耕地破碎、云遮影响严重的山区条件下显著提升了作物分类精度，具有良好的推广前景与应用价值。\n\n**关键词:**[农作物分类,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[深度学习,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[3D卷积神经网络,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[注意力机制,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[遥感,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[合成孔径雷达](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)\n\n**Abstract:**\n\n**[Objective]** Accurate and timely crop mapping is fundamental for agricultural management, yield forecasting, and food security assessment. However, in mountainous and hilly regions characterized by frequent cloud cover and highly fragmented farmland, crop classification methods relying solely on optical remote sensing data are severely constrained. Persistent cloud contamination introduces data gaps and temporal inconsistencies in optical image time series, significantly degrading classification accuracy and robustness. To address these limitations, a robust and adaptive deep learning framework is developed capable of effectively integrating multi-modal remote sensing data. The primary objective is to enhance crop classification accuracy and stability under complex conditions where optical observations are scarce or unreliable, thereby supporting reliable agricultural monitoring in cloudy and fragmented landscapes. **[Methods]** A novel deep neural network architecture named 3D convolutional neural network based on attention mechanism (Attention-3DCNN) was proposed, designed to jointly exploit multi-temporal optical and synthetic aperture radar (SAR) observations. The model integrated Sentinel-2 multispectral time-series imagery with weather-insensitive Sentinel-1 SAR data through a dedicated cross-modal fusion strategy driven by a triple-attention mechanism. The network adopted a dual-branch feature extraction architecture. For the Sentinel-2 data, a hybrid module combining three-dimensional and two-dimensional convolutional neural networks (3D-CNN and 2D-CNN) was employed to capture discriminative spatiotemporal features and crop phenological dynamics across the growing season. This design enabled effective modeling of the spectral-temporal interactions inherent in crop development. For the Sentinel-1 SAR data, depthwise separable convolutions were utilized to efficiently extract spatial and textural features related to crop structure and surface scattering characteristics while reducing computational complexity. Features extracted from both modalities were subsequently integrated using a custom-designed attention-based fusion module. This module consisted of three complementary attention mechanisms: channel attention, temporal attention, and spatial attention. Residual connections were incorporated throughout the network to facilitate stable training and effective gradient propagation. The proposed model was evaluated on two datasets to assess both its performance and generalizability. The first was the publicly available panoptic agricultural satellite time series (PASTIS) benchmark dataset from France, which contained dense time-series observations and multiple crop classes. The second was a real-world dataset constructed for Yishui county, Shandong province, China, which was characterized by high cloud frequency (approximately 33%), highly fragmented farmland (average parcel size \u003C 0.5 hm 2), and a relatively simple crop rotation system. Comparative experiments were conducted against several state-of-the-art models, including 3D-ConvSTAR, UNet++, Self-Attention 3D, CNN-LSTM dual-stream network, and TGF-Net. Ablation studies were also performed to quantify the contribution of each attention component. **[Results and Discussions]** Experimental results demonstrated that Attention-3DCNN consistently outperformed all baseline methods on both datasets. On the PASTIS benchmark, the model achieved an overall accuracy (OA) of 97.5%, confirming its strong classification capability under favorable observation conditions. On the more challenging Yishui county dataset, Attention-3DCNN attained an OA of 93%, outperforming the other comparison models. Ablation experiments confirmed the effectiveness of the proposed triple-attention mechanism, as removing any attention component resulted in a clear reduction in classification performance. Under heavy cloud coverage, Attention-3DCNN exhibited the smallest accuracy degradation, with an OA drop of only 3.6 percentage points, indicating its ability to adaptively rely on SAR information when optical data quality deteriorated. In regions with highly fragmented farmland, the proposed model also maintained the highest accuracy and the smallest performance decline (2.8 percentage points), benefiting from the spatial attention mechanism. Moreover, attention visualization provided meaningful interpretability. Temporal attention peaks aligned with key crop phenological stages, while channel attention highlighted spectrally and physically informative optical bands and SAR polarizations, which was consistent with established agronomic and remote sensing knowledge. **[Conclusions]** This study presents the Attention-3DCNN model for accurate and robust crop classification in regions affected by persistent cloud cover and fragmented agricultural landscapes. By fusing Sentinel-2 optical and Sentinel-1 SAR time-series data through a channel-temporal-spatial triple-attention mechanism, the proposed framework enables adaptive integration of complementary multi-modal information. The model achieves outstanding performance on both benchmark and real-world datasets, demonstrates strong robustness under adverse conditions, and offers enhanced interpretability. Overall, the proposed approach provides a reliable and practical solution for crop mapping in complex agricultural environments.\n\n**Key words:**[crop classification,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[deep learning,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[3D convolutional neural network,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[attention mechanism,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[remote sensing,](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)[SAR](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)\n\n**中图分类号:**\n\n*   [S127](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)\n\n*   [TP18](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#)\n\n#### 引用本文\n\n巫晨旭, 左浩龙, 李刚. 面向云遮挡与复杂地块场景下的跨模态注意力多源遥感农作物分类[J]. 智慧农业(中英文), 2026, 8(2): 118-132.\n\nWU Chenxu, ZUO Haolong, LI Gang. Cross-Modal Attention for Multi-Source Remote Sensing Crop Classification under Cloud Occlusion and Complex Field Scenarios[J]. Smart Agriculture, 2026, 8(2): 118-132.\n\n#### 使用本文\n\n**[[收藏文章](javascript:;)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#) \u002F[推荐](javascript:void(null))**\n**导出引用管理器**[EndNote](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Farticle\u002FgetTxtFile.do?fileType=EndNote&id=22683)|[Ris](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Farticle\u002FgetTxtFile.do?fileType=Ris&id=22683)|[BibTeX](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002Farticle\u002FgetTxtFile.do?fileType=BibTeX&id=22683)\n\n**链接本文:**[https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010)\n\n[https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002FY2026\u002FV8\u002FI2\u002F118](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002FY2026\u002FV8\u002FI2\u002F118)\n\n#### [图\u002F表 23](javascript:;)\n\n[![Image 7](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F001.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass13989)\n\n**图1**\n\n山东省沂水县研究区域示意图 注：该图基于自然资源部标准地图服务网站下载的审图号为GS（2019）3333号标准地图制作，底图无修改。\n\n![Image 8](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F001.jpg)\n\n 图1 \n\n[![Image 9](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Ftable-icon.gif)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#inline_content13990)\n\n**表1**\n\nPASTIS数据集与Sentinel-2波段对应关系及分辨率对比表\n\n| PASTIS 波段 | Sentinel-2 波段 | 原始分辨率\u002Fm | 重采样后分辨率\u002Fm |\n| :---: | :---: | :---: | :---: |\n| 1—4 | B2—B4， B8 | 10 | 10 |\n| 5—7 | B5—B7 | 20 | 10 |\n| 8 | B8A | 20 | 10 |\n| 9—10 | B11—B12 | 20 | 10 |\n\n 表1 \n\n[![Image 10](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F002.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass13991)\n\n**图2**\n\nPASTIS数据集 a. 光学影像 b. SAR数据 c. 分类标签\n\n![Image 11](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F002.jpg)\n\n 图2 \n\n[![Image 12](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Ftable-icon.gif)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#inline_content13992)\n\n**表2**\n\nPASTIS与沂水县数据集核心参数的对比\n\n| 参数 | PASTIS（法国） | 沂水县（中国） | 对迁移的潜在影响 |\n| :---: | :---: | :---: | :---: |\n| 云覆盖率\u002F% | 28 | 33 | 沂水县需更强的抗云干扰能力 |\n| 时相数量（4—9月） | 32个时相 | 23个时相 | 沂水县时序信息更稀疏，考验时间注意力 |\n| 平均地块面积\u002Fhm 2 | 1.2 | \u003C0.5 | 沂水县需更高的空间细节捕捉能力 |\n\n 表2 \n\n[![Image 13](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F003.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass13993)\n\n**图3**\n\n沂水县主要农作物分类数据集\n\n![Image 14](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F003.jpg)\n\n 图3 \n\n[![Image 15](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F004.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass13994)\n\n**图4**\n\nAttention-3DCNN模型总架构图 注：DSConv为深度可分离卷积层；3D-CNN、2D-CNN分别是3D卷积层和2D卷积层；Ad-Pooling为自适应池化。\n\n![Image 16](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F004.jpg)\n\n 图4 \n\n[![Image 17](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F005.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass13995)\n\n**图5**\n\n多光谱分支模型图 注：Conv3D（3×3×3，s=1，p=1，C_out=64），Conv2D（3×3，s=1，p=1，C_out=128），Ad-Pooling（1×1，s=1，p=0，C_out= 128）；s为步长，p为填充，C_out为输出大小；spectral、time、spatial分别代表通道、时间、空间注意力机制。\n\n![Image 18](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F005.jpg)\n\n 图5 \n\n[![Image 19](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F006.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass13996)\n\n**图6**\n\nSAR分支模型架构图 注：BN为批量归一化；LeakyReLu为LeakyReLU激活函数DSConv（3×3，s=1，p=1，C_out=C_in）；Ad-Pooling（1×1，s=1，p=0，C_out=128）；s为步长；p为填充；C_out为输出大小；C_in为输入大小。\n\n![Image 20](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F006.jpg)\n\n 图6 \n\n[![Image 21](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F007.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass13997)\n\n**图7**\n\n注意力机制流程图\n\n![Image 22](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F007.jpg)\n\n 图7 \n\n[![Image 23](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F008.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass13998)\n\n**图8**\n\n通道注意力机制模型图\n\n![Image 24](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F008.jpg)\n\n 图8 \n\n[![Image 25](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Ftable-icon.gif)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#inline_content13999)\n\n**表3**\n\nAttention-3DCNN模型与其他模型在不同数据集上的表现\n\n| 模型名称 | OA（PASTIS）\u002F% | 宏平均 _F_ 1 分数（PASTIS） | Kappa（PASTIS） | OA（沂水县）\u002F% | 宏平均 _F_ 1 分数（沂水县） | Kappa（沂水县） |\n| :---: | :---: | :---: | :---: | :---: | :---: | :---: |\n| 标准3D-CNN | 94.2 | 0.932 | 0.915 | 87.2 | 0.845 | 0.822 |\n| 光学-SAR简单融合模型 | 95.3 | 0.945 | 0.928 | 89.5 | 0.876 | 0.858 |\n| 仅SAR 3D-CNN模型 | 91.6 | 0.902 | 0.889 | 88.3 | 0.861 | 0.842 |\n| 注意力双分支融合模型 | 95.1 | 0.938 | 0.935 | 89.9 | 0.882 | 0.870 |\n| 通道注意力双分支融合模型 | 96.5 | 0.953 | 0.942 | 90.0 | 0.898 | 0.890 |\n| 时间注意力双分支融合模型 | 95.8 | 0.945 | 0.942 | 89.5 | 0.894 | 0.887 |\n| 空间注意力双分支融合模型 | 96.9 | 0.959 | 0.956 | 91.0 | 0.906 | 0.900 |\n| Attention-3DCNN | 97.5 | 0.970 | 0.965 | 93.0 | 0.920 | 0.910 |\n\n 表3 \n\n[![Image 26](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F009.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass14000)\n\n**图9**\n\n在消融实验中Attention-3DCNN与其他模型在沂水县数据集上的结果\n\n![Image 27](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F009.jpg)\n\n 图9 \n\n[![Image 28](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F010.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass14001)\n\n**图10**\n\n在消融实验中Attention-3DCNN模型与其他模型在PASTIS数据集上的结果\n\n![Image 29](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F010.jpg)\n\n 图10 \n\n[![Image 30](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Ftable-icon.gif)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#inline_content14002)\n\n**表4**\n\nAttention-3DCNN模型与其他模型的对比\n\n| 模型名称 | 核心方法简述 | OA（沂水县）\u002F% | Kappa（沂水县） | 参数量\u002FM | 计算量\u002FGFLOPs | 推理时间\u002F（ms\u002F景） |\n| :---: | :---: | :---: | :---: | :---: | :---: | :---: |\n| 3D-ConvSTAR | 3DCNN多源融合（固定权重） | 89.5 | 0.858 | 45.2 | 128.3 | 156 |\n| Self-Attention 3D | 自注意力机制+3DCNN | 90.5 | 0.872 | 52.7 | 145.6 | 183 |\n| UNet++ | 编码器-解码器结构，多尺度特征融合 | 88.0 | 0.835 | 68.9 | 212.4 | 245 |\n| CNN-LSTM-DS | 光学影像时序+纹理特征融合 | 86.5 | 0.815 | 43.0 | 98.7 | 132 |\n| TGF-Net | 基于Transformer与卷积（CNN）架构 | 90.5 | 0.900 | 105.3 | 285.1 | 312 |\n| Attention-3DCNN | 跨模态三重注意力（通道-时间-空间） | 93.5 | 0.910 | 41.3 | 97.2 | 133 |\n\n 表4 \n\n[![Image 31](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbnail\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F011.jpg)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#figureClass14003)\n\n**图11**\n\n在对比实验中Attention-3DCNN模型与其他模型在沂水县数据集上的对比结果\n\n![Image 32](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002F1DDA0D33-B017-4e44-93B5-E54B9C48FD2F-F011.jpg)\n\n 图11 \n\n[![Image 33](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Ftable-icon.gif)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#inline_content14004)\n\n**表5**\n\n在高云量的条件下Attention-3DCNN模型与其他模型在沂水县数据集的对比\n\n| 模型名称 | 主要依赖信息 | OA\u002F% | 宏平均 _F_ 1 | Kappa |\n| :---: | :---: | :---: | :---: | :---: |\n| 3D-ConvSTAR | 光学时序为主 | 83.6 | 0.802 | 0.775 |\n| Self-Attention 3D | 光学 + 时序注意力 | 85.1 | 0.821 | 0.796 |\n| UNet++ | 光学空间特征 | 82.4 | 0.789 | 0.761 |\n| CNN-LSTM-DS | 光学时序 + 手工特征 | 81.9 | 0.781 | 0.754 |\n| TGF-Net | Transformer + CNN融合 | 86.3 | 0.836 | 0.812 |\n| Attention-3DCNN | 跨模态三重注意力（S2+S1） | 89.4 | 0.872 | 0.846 |\n\n 表5 \n\n[![Image 34](https:\u002F\u002Fwww.smartag.net.cn\u002Fimages\u002Ftable-icon.gif)](https:\u002F\u002Fwww.smartag.net.cn\u002FCN\u002F10.12133\u002Fj.smartag.SA202510010#inline_content14005)\n\n**表6**\n\n多源遥感农作物分类研究中不同模型在高云量条件下的下降幅度\n\n| 模型 | 常规OA\u002F% | 高云量OA\u002F% | 下降幅度\u002F百分点 |\n| :---: | :---: | :---: | :---: |\n| 3D-ConvSTAR | 89.5 | 83.6 | 5.9 |\n| Self-Attention 3D | 90.5 | 85.1 | 5.4 |\n| UNet++ | 88.0 | 82.4 | 5.6 |\n| CNN-LSTM-DS | 86.5 | 81.9 | 4.6 |\n| TGF-Net | 90.0 | 86.3 | 3.7 |\n| Attention-3DCNN | 93.0 | 89.4 | 3.6 |\n\n 表6 \n\n[![Image 35](https:\u002F\u002Fwww.smartag.net.cn\u002Ffileup\u002F2096-8094\u002FFIGURE\u002F2026-8-2\u002FImages\u002F2096-8094-2026-8-2-118\u002Fthumbn","智慧农业(中英文) 2026年第8卷 第2期","2026-07-08T00:00:00Z",67,{"impact":64,"substance":18,"depth":19,"authority":17,"freshness":52,"relevant":22,"comment":139},"核心期刊论文，提出改进YOLOv11n-seg模型用于胡麻倒伏识别，方法新颖且兼顾精度与轻量化，但发表于数月前，时效性较低。",[141],{"name":135,"url":132},[72,143,144,145,27],"YOLOv11n-seg","胡麻倒伏","轻量化",[],{"status":148,"locked":14,"version":34,"method":35,"evidence":149},"needs_review",{"candidates":150,"agricultural_evidence":14,"reason":159},[151,155],{"code":102,"score":22,"supported":14,"title":152,"abstract":153,"tags":154},[],[],[72],{"code":39,"score":22,"supported":14,"title":156,"abstract":157,"tags":158},[],[],[27],"农业相关性需要复核",[161,162],"实例分割 智慧农业 胡麻倒伏 轻量化","实例分割 智慧农业","实例分割智慧农业胡麻倒伏轻量化-177","10.12133\u002Fj.smartag.sa202510010","2026-08-01T23:53:44.590710Z"]