[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"item-3159":3,"related-3159":61},{"id":4,"title":5,"url":6,"summary":7,"summary_zh":8,"content":9,"source_name":10,"source_url":6,"published_at":11,"category":12,"cover_url":9,"hotness":13,"is_selected":14,"score":15,"score_detail":16,"sources":23,"tags":25,"search_phrases":31,"slug":34,"view_count":35,"doi":36,"paper":37,"created_at":60},3159,"A multi-agent multi-modal LLM with Monte Carlo tree search for interpretable plant disease classification","https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffpls.2026.1937679","In plant disease classification tasks, existing task specific models for plant disease classification typically achieve high accuracy by employing tailored network architectures. However, these models are unable to provide reasoning chains, which reduces interpretability and undermines trustworthiness in practical applications. In addition, although multimodal large language models (MLLMs) have made significant progress in smart agriculture, their disease classification accuracy lags behind that of task-specific models due to insufficient domain knowledge. While fine-tuning can enhance domain-specific knowledge, relying solely on unimodal data or simple multimodal information remains insufficient for adequately improving classification accuracy. Specifically, the simple multimodal information in this context refers to a single global description of the entire image produced by an MLLM. To address these issues, this paper proposes a multi-agent collaborative classification framework that integrates Monte Carlo Tree Search (MCTS) with multimodal enhancement. With an MLLM as the core, the framework first introduces an MCTSguided multi-agent collaboration mechanism that directs the agents to progressively generate region-specific descriptive textual information, which together with the image forms complex multimodal information and thereby boosts classification performance. Meanwhile, the search trajectories generated by MCTS are leveraged to construct reasoning processes for enhanced interpretability. Finally, a self-training mechanism is adopted to iteratively optimize the model, further improving overall performance. Experimental results show that the framework improves the baseline MLLM’s accuracy while providing interpretable reasoning chains, achieving gains from 0.6701 to 0.9965 on 3-class potato, 0.5340 to 0.9651 on 5-class potato, 0.5387 to 0.9467 on tomato, and 0.4705 to 0.9948 on grape.","在植物病害分类任务中，现有的植物病害分类专用模型通常通过采用定制化的网络架构来实现高准确率。然而，这些模型无法提供推理链，这降低了可解释性，并在实际应用中削弱了可信度。此外，尽管多模态大语言模型（MLLMs）在智慧农业领域取得了显著进展，但由于领域知识不足，其病害分类准确率仍落后于专用模型。虽然微调可以增强领域特定知识，但仅依赖单模态数据或简单的多模态信息仍不足以充分提升分类准确率。具体而言，此处的简单多模态信息是指MLLM对整个图像生成的单一全局描述。为解决上述问题，本文提出了一种融合蒙特卡洛树搜索（MCTS）与多模态增强的多智能体协同分类框架。该框架以MLLM为核心，首先引入MCTS引导的多智能体协同机制，引导各智能体逐步生成区域特定的描述性文本信息，这些信息与图像共同构成复杂的多模态信息，从而提升分类性能。同时，利用MCTS生成的搜索轨迹构建推理过程，以增强可解释性。最后，采用自训练机制迭代优化模型，进一步提升整体性能。实验结果表明，该框架在提升基线MLLM准确率的同时提供了可解释的推理链，在三分类马铃薯上从0.6701提升至0.9965，在五分类马铃薯上从0.5340提升至0.9651，在番茄上从0.5387提升至0.9467，在葡萄上从0.4705提升至0.9948。",null,"Frontiers in Plant Science","2026-09-21T00:00:00Z","论文",10,false,80,{"impact":17,"substance":18,"depth":17,"authority":19,"freshness":20,"relevant":21,"comment":22},18,22,14,8,1,"提出多智能体+MCTS+多模态增强的可解释植物病害分类框架，准确率大幅提升，方法新颖且结论可靠，对智慧农业AI应用有参考价值。",[24],{"name":10,"url":6},[26,27,28,29,30],"智慧农业","农业人工智能","可解释AI","植物病害识别","多模态大模型",[32,33],"多智能体 蒙特卡洛树搜索 植物病害分类","多模态大模型 马铃薯 番茄 葡萄 病害识别","多智能体蒙特卡洛树搜索植物病害分类-3159",0,"10.3389\u002Ffpls.2026.1937679",{"doi":36,"openalex_id":38,"authors":39,"venue":10,"cited_by_count":35,"oa_url":6,"card":52,"direction":58,"ingested_from":59},"W7213924931",[40,43,45,47,50],{"name":41,"orcid":42},"Ridong Wu","https:\u002F\u002Forcid.org\u002F0000-0003-0729-3099",{"name":44,"orcid":9},"Hua Zou",{"name":46,"orcid":9},"Zhijie Li",{"name":48,"orcid":49},"Qian Zhou","https:\u002F\u002Forcid.org\u002F0000-0001-7964-8130",{"name":51,"orcid":9},"一子 江﨑",{"tldr":53,"method":54,"finding":55,"direction":56,"opportunity":57},"提出多智能体多模态LLM结合蒙特卡洛树搜索，实现可解释的植物病害分类。","MCTS引导多智能体生成区域描述文本，构建复杂多模态信息并自训练优化。","分类准确率大幅提升，如马铃薯三分类从0.6701升至0.9965，并提供推理链。","农业人工智能与决策模型","可探索将MCTS多智能体框架迁移至其他作物病害或田间复杂场景，提升可解释性与泛化性。","智慧农业 \u002F 农业物联网","openalex","2026-09-22T23:30:11.045394Z",{"total":62,"page":21,"page_size":62,"items":63},6,[64,92,126,160,208,268],{"id":65,"title":66,"url":67,"summary":68,"summary_zh":9,"content":9,"source_name":69,"source_url":9,"published_at":70,"category":12,"cover_url":9,"hotness":13,"is_selected":14,"score":71,"score_detail":72,"sources":75,"tags":77,"search_phrases":79,"slug":82,"view_count":35,"doi":9,"paper":83,"created_at":91},2901,"AgriScope:面向农业图像的像素级多模态理解统一框架,arXiv 2609.20325(预印本)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.20325","Boudiaf、Alanssari、Hussain、Javed提出AgriScope,一个统一的像素级多模态农业图像理解框架,联合支持图像级、区域级、像素级理解,可实现接地描述生成、指代表达分割、多轮多模态交互等任务。集成生物专用语义表征、密集空间表征与像素解码;引入大规模像素级农业多模态指令调优数据集AgriGround,包含50万+图像和1100万+指令跟随样本,涵盖植物病害分析、作物与杂草识别、昆虫识别、细粒度植物理解。实验表明AgriScope在多项农业视觉语言任务上有效。","arXiv (preprint)","2026-09-17T00:00:00Z",75,{"impact":17,"substance":18,"depth":17,"authority":20,"freshness":73,"relevant":21,"comment":74},9,"提出统一像素级农业多模态理解框架并开源50万图像、1100万指令样本的大规模数据集，方法新颖、数据规模突出，但为arXiv预印本、未经同行评审，权威性有限，值得作为前沿技术动态精选。",[76],{"name":69,"url":67},[26,27,78,29,30],"农业遥感",[80,81],"AgriScope 农业图像 多模态","AgriGround 像素级 农业数据集","AgriScope农业图像多模态-2901",{"doi":9,"openalex_id":9,"authors":84,"venue":9,"cited_by_count":35,"oa_url":9,"card":85,"direction":56,"ingested_from":90},[],{"tldr":86,"method":87,"finding":88,"direction":56,"opportunity":89},"提出AgriScope统一框架，实现农业图像像素级多模态理解与多任务交互。","构建AgriGround数据集（50万+图像、1100万+指令样本），融合语义与","AgriScope在接地描述、指代分割、多轮交互等农业视觉语言任务上有效。","可探索像素级多模态模型在田间实时病害诊断与精准施药决策中的落地与轻量化。","agent","2026-09-19T00:06:08.678379Z",{"id":93,"title":94,"url":95,"summary":96,"summary_zh":97,"content":9,"source_name":98,"source_url":95,"published_at":99,"category":12,"cover_url":9,"hotness":13,"is_selected":14,"score":100,"score_detail":101,"sources":104,"tags":106,"search_phrases":108,"slug":111,"view_count":35,"doi":112,"paper":113,"created_at":125},2178,"Hybrid explainable DeiT-based framework for plant disease classification and severity estimation","https:\u002F\u002Fdoi.org\u002F10.1007\u002Fs44163-026-02191-2","The detection of plant diseases is essential for preserving agricultural productivity and food security; however, existing approaches often suffer from limited interpretability and generalization capability. This study proposes a hybrid deep learning framework based on Data-efficient Image Transformers (DeiT) for plant disease classification and severity estimation. The framework employs DeiT-Base, DeiT-Small, and DeiT-Tiny models to capture global contextual dependencies in plant leaf images. To improve interpretability, a hybrid Explainable Artificial Intelligence (XAI) module is introduced by combining Gradient-weighted Class Activation Mapping (Grad-CAM) for local feature attribution with Attention Rollout for global dependency visualization. In addition, HSV-based segmentation is applied after classification to isolate disease-relevant regions for damage ratio computation, severity estimation, and explanation refinement. The damage ratio is further integrated with Hybrid XAI attention maps to estimate disease severity. Experiments were conducted on the New Plant Diseases Dataset (Augmented), comprising 70,295 training images and 17,572 validation images across 38 disease classes and 14 plant species. The proposed DeiT-Base model achieved a maximum classification accuracy of 99.13%, outperforming several CNN architectures, including ResNet50, DenseNet121, MobileNetV3, EfficientNet-B4, and InceptionV3. Furthermore, the proposed Hybrid XAI framework demonstrated superior interpretability performance in terms of Focus Score, Background Noise, Signal-to-Noise Ratio (SNR), and Entropy compared with individual explanation methods. Overall, the proposed framework improves classification accuracy, enhances model transparency, and provides meaningful disease severity estimation, making it a promising solution for intelligent precision agriculture.","植物病害检测对于保障农业生产力和粮食安全至关重要，然而现有方法往往存在可解释性有限和泛化能力不足的问题。本研究提出了一种基于数据高效图像Transformer（Data-efficient Image Transformers，DeiT）的混合深度学习框架，用于植物病害分类和严重程度估计。该框架采用DeiT-Base、DeiT-Small和DeiT-Tiny模型来捕获植物叶片图像中的全局上下文依赖关系。为提高可解释性，引入了一种混合可解释人工智能（Explainable Artificial Intelligence，XAI）模块，将用于局部特征归因的梯度加权类激活映射（Gradient-weighted Class Activation Mapping，Grad-CAM）与用于全局依赖可视化的注意力展开（Attention Rollout）相结合。此外，在分类之后应用基于HSV的分割来分离病害相关区域，以进行损伤比率计算、严重程度估计和解释优化。损伤比率进一步与混合XAI注意力图相结合以估计病害严重程度。实验在新植物病害数据集（增强版）（New Plant Diseases Dataset (Augmented)）上进行，该数据集包含70,295张训练图像和17,572张验证图像，涵盖38个病害类别和14种植物物种。所提出的DeiT-Base模型达到了99.13%的最高分类准确率，优于多种CNN架构，包括ResNet50、DenseNet121、MobileNetV3、EfficientNet-B4和InceptionV3。此外，所提出的混合XAI框架在聚焦分数（Focus Score）、背景噪声（Background Noise）、信噪比（Signal-to-Noise Ratio，SNR）和熵（Entropy）方面表现出优于单一解释方法的可解释性性能。总体而言，所提出的框架提高了分类准确率，增强了模型透明度，并提供了有意义的病害严重程度估计，使其成为智能精准农业的一种有前景的解决方案。","Discover Artificial Intelligence","2026-09-10T00:00:00Z",78,{"impact":17,"substance":18,"depth":17,"authority":102,"freshness":20,"relevant":21,"comment":103},12,"基于DeiT与混合可解释AI的植物病害分类与严重度估计研究，在7万张图像、38类病害上取得99.13%准确率，方法新颖、数据扎实，对智慧农业病害智能诊断有参考价值。",[105],{"name":98,"url":95},[26,27,28,107,29],"精准农业",[109,110],"农业人工智能 植物病害识别 智慧农业 精准农业","农业人工智能 植物病害识别","农业人工智能植物病害识别智慧农业精准农业-2178","10.1007\u002Fs44163-026-02191-2",{"doi":112,"openalex_id":114,"authors":115,"venue":98,"cited_by_count":35,"oa_url":95,"card":120,"direction":56,"ingested_from":59},"W7212179384",[116,118],{"name":117,"orcid":9},"Muskan Batra",{"name":119,"orcid":9},"Pooja Sharma",{"tldr":121,"method":122,"finding":123,"direction":56,"opportunity":124},"提出基于DeiT的混合可解释框架，实现植物病害分类与严重度估计。","DeiT-Base\u002FSmall\u002FTiny结合Grad-CAM与Attention","DeiT-Base分类准确率达99.13%，混合XAI可解释性指标优于单一方法。","可探索轻量化DeiT在边缘设备部署及多模态数据融合的病害严重度实时估计。","2026-09-11T23:30:47.756846Z",{"id":127,"title":128,"url":129,"summary":130,"summary_zh":131,"content":9,"source_name":132,"source_url":129,"published_at":11,"category":12,"cover_url":9,"hotness":13,"is_selected":14,"score":133,"score_detail":134,"sources":136,"tags":138,"search_phrases":141,"slug":144,"view_count":35,"doi":145,"paper":146,"created_at":159},3195,"Segmentation of Cactus Diseases Using Machine and Deep Learning)","https:\u002F\u002Fdoi.org\u002F10.37394\u002F232018.2026.14.50","Artificial intelligence and machine learning play a critical role in plant disease detection and management. This study proposes an approach for diagnosing and segmenting cactus diseases using Random Forest (RF) and compares its performance with SVM, ANN, KNN, CNN, and U-Net models. A dataset of 20 cactus images (10 cochineal and 10 black bacterial soft rot), resized to 720×720 pixels, was used. Experimental results show that U-Net achieved the highest accuracy, reaching 98% for cochineal and 94% for black bacterial soft rot. RF also demonstrated strong performance, achieving 92% and 91%, respectively, outperforming several traditional machine learning models. These results confirm the effectiveness of U-Net for accurate segmentation and the suitability of RF for reliable disease detection with limited datasets, supporting their application in precision agriculture.","人工智能与机器学习在植物病害检测与管理中发挥着关键作用。本研究提出了一种利用随机森林（Random Forest, RF）诊断和分割仙人掌病害的方法，并将其性能与SVM、ANN、KNN、CNN和U-Net模型进行了比较。研究使用了20张仙人掌图像数据集（10张胭脂虫病害和10张黑细菌性软腐病），图像尺寸调整为720×720像素。实验结果表明，U-Net取得了最高准确率，胭脂虫病害达到98%，黑细菌性软腐病达到94%。随机森林同样表现出强劲性能，分别达到92%和91%，优于多种传统机器学习模型。这些结果证实了U-Net在精确分割方面的有效性，以及随机森林在有限数据集下进行可靠病害检测的适用性，支持其在精准农业中的应用。","WSEAS TRANSACTIONS ON COMPUTER RESEARCH",62,{"impact":20,"substance":17,"depth":19,"authority":102,"freshness":13,"relevant":21,"comment":135},"方法对比扎实但数据集仅20张图像，属细分作物病害识别的技术验证，产业影响有限。",[137],{"name":132,"url":129},[26,27,29,139,140],"图像分割","仙人掌",[142,143],"仙人掌 病害 图像分割","U-Net 胭脂虫 黑腐病","仙人掌病害图像分割-3195","10.37394\u002F232018.2026.14.50",{"doi":145,"openalex_id":147,"authors":148,"venue":132,"cited_by_count":35,"oa_url":129,"card":154,"direction":56,"ingested_from":59},"W7213945610",[149,151],{"name":150,"orcid":9},"Aissam El Ibrahimi",{"name":152,"orcid":153},"Nabil El Akchioui","https:\u002F\u002Forcid.org\u002F0000-0001-9808-932X",{"tldr":155,"method":156,"finding":157,"direction":56,"opportunity":158},"该研究用随机森林和多种深度学习模型对仙人掌病害进行分割诊断，并比较性能。","使用20张仙人掌图像，对比RF、SVM、ANN、KNN、CNN和U-Net模型。","U-Net分割准确率最高（胭脂虫98%，黑细菌软腐病94%），RF在有限数据下表现良好。","可探索小样本条件下传统机器学习与深度学习融合的病害分割方法，并扩展至更多作物和病害类型。","2026-09-22T23:30:39.786818Z",{"id":161,"title":162,"url":163,"summary":164,"summary_zh":165,"content":9,"source_name":98,"source_url":163,"published_at":166,"category":12,"cover_url":9,"hotness":13,"is_selected":14,"score":71,"score_detail":167,"sources":171,"tags":173,"search_phrases":176,"slug":179,"view_count":35,"doi":180,"paper":181,"created_at":207},3153,"Explainable optimized deep learning and generative AI based framework for finger millet disease detection in smart agriculture","https:\u002F\u002Fdoi.org\u002F10.1007\u002Fs44163-026-02222-y","Abstract Eleusine coracana, locally known as finger millet (ragi), is a wholesome, climate-resilient crop. It is a staple food in semi-dry and dry areas of Asia and Africa. The productivity of this crop can be severely reduced by diseases such as downy, mottle, seedling, smut, and wilt. Conventional diagnosis methods are often time-consuming and unsuitable for large-scale farming. Therefore,Deep Learning (DL) offers an alternative for smart commercial farming that can streamline disease screening and improve productivity. This work offers an AI-integrated framework for early disease intervention combining DL, Explainable AI (XAI), and an image-based decision-support interface. The finger millet (ragi) dataset from Kaggle was used to train and evaluate custom CNN, VGG16, and ResNet50 models. To reduce feature redundancy and improve generalization, Grey Wolf Optimizer (GWO) based feature selection was applied to the penultimate-layer features of each model. Custom CNN, VGG16, and ResNet50 achieved accuracies of 89.40%, 83.11%, and 91.83%, respectively. Further, the use of GWO improved the accuracies to 97.71%, 90.84%, and 98.36%, respectively. ResNet50 achieved the highest accuracy 98.36% and was selected as the final backbone. The final ResNet50 model is integrated with the Gemini API to provide disease-specific recommendations based on the predicted class and user queries. To increase prediction transparency, Grad-CAM was used to highlight image regions that influenced the model output. The integration of DL, XAI, and generative AI provides a scalable approach for detecting and managing finger millet diseases. The proposed work boosts precision agriculture by providing AI-based and XAI-supported decision-making, while fostering sustainable agricultural practices.","摘要 穇子（Eleusine coracana），当地称为指黍（ragi），是一种有益健康且气候适应性强的作物。它是亚洲和非洲半干旱及干旱地区的主食。霜霉病、斑驳病、苗枯病、黑穗病和枯萎病等病害可严重降低该作物的产量。传统诊断方法往往耗时且不适合大规模种植。因此，深度学习（DL）为智能商业农业提供了一种替代方案，可简化病害筛查并提高生产力。本研究提出了一种人工智能集成框架，用于早期病害干预，结合了深度学习、可解释人工智能（XAI）和基于图像的决策支持界面。使用来自Kaggle的指黍（ragi）数据集训练和评估了自定义CNN、VGG16和ResNet50模型。为减少特征冗余并提高泛化能力，将灰狼优化器（GWO）基于特征选择应用于每个模型的倒数第二层特征。自定义CNN、VGG16和ResNet50分别达到了89.40%、83.11%和91.83%的准确率。此外，使用GWO将准确率分别提高到97.71%、90.84%和98.36%。ResNet50达到了最高准确率98.36%，并被选为最终骨干网络。最终的ResNet50模型与Gemini API集成，根据预测类别和用户查询提供针对特定病害的建议。为提高预测透明度，使用Grad-CAM突出显示影响模型输出的图像区域。深度学习、可解释人工智能和生成式人工智能的集成提供了一种可扩展的方法，用于检测和管理指黍病害。所提出的工作通过提供基于人工智能和可解释人工智能支持的决策，促进了精准农业，同时推动了可持续农业实践。","2026-09-22T00:00:00Z",{"impact":168,"substance":169,"depth":17,"authority":102,"freshness":13,"relevant":21,"comment":170},15,20,"将深度学习、可解释AI与生成式AI结合用于指状粟病害识别，方法新颖、数据充分，对智慧农业有参考价值。",[172],{"name":98,"url":163},[26,27,28,174,175],"病害识别","小米作物",[177,178],"ResNet50 GWO 病害识别","农业人工智能 小米作物 智慧农业 病害识别","ResNet50GWO病害识别-3153","10.1007\u002Fs44163-026-02222-y",{"doi":180,"openalex_id":182,"authors":183,"venue":98,"cited_by_count":35,"oa_url":163,"card":202,"direction":58,"ingested_from":59},"W7213978567",[184,187,189,191,193,196,199],{"name":185,"orcid":186},"Sunil Kumar Mohapatra","https:\u002F\u002Forcid.org\u002F0000-0002-5865-095X",{"name":188,"orcid":9},"A. Sanjib Kumar Patro",{"name":190,"orcid":9},"Lulen Kumar Sahu",{"name":192,"orcid":9},"Chinmaye Dora",{"name":194,"orcid":195},"Sujata Chakravarty","https:\u002F\u002Forcid.org\u002F0000-0002-1293-5378",{"name":197,"orcid":198},"Kshira Sagar Sahoo","https:\u002F\u002Forcid.org\u002F0000-0002-6435-5738",{"name":200,"orcid":201},"Byomakesh Mahapatra","https:\u002F\u002Forcid.org\u002F0000-0002-9126-1729",{"tldr":203,"method":204,"finding":205,"direction":56,"opportunity":206},"提出融合深度学习、可解释AI与生成式AI的指状粟病害检测框架，实现高精度识别与决策支持。","用Kaggle指状粟图像训练CNN、VGG16、ResNet50，并用灰狼优化做","GWO特征选择显著提升精度，ResNet50达98.36%，结合Grad-CAM与Gemini AP","可探索轻量化模型与多作物泛化，并将XAI与生成式建议在田间移动端实时验证。","2026-09-22T23:30:10.255790Z",{"id":209,"title":210,"url":211,"summary":212,"summary_zh":213,"content":9,"source_name":214,"source_url":211,"published_at":11,"category":12,"cover_url":9,"hotness":13,"is_selected":14,"score":215,"score_detail":216,"sources":219,"tags":221,"search_phrases":224,"slug":227,"view_count":35,"doi":228,"paper":229,"created_at":267},3137,"From patterns to prediction: detecting crop sequence and interpreting crop selection with explainable AI at field level in Denmark","https:\u002F\u002Fdoi.org\u002F10.1016\u002Fj.agsy.2026.104982","CONTEXT Crop sequences are important for sustainable agriculture, yet we know relatively little about how they recur across individual fields or which information best predicts the crop grown each year. Denmark's national field records allow both questions to be examined over time. OBJECTIVE We identified recurrent crop sequences across Denmark and tested how well annual crop-family selection could be predicted from crop history, farm characteristics, management, prices, soil and climate. METHODS Using 10 years (2011−2020) of national field-level crop data (∼600,000 fields yr −1 ), we developed a heuristic algorithm to detect recurring crop sequence patterns. We then trained machine learning (LightGBM) and deep learning (TabNet) models to predict annual crop choices based on preceding crops and lagged management and farm structure predictors, and pedoclimatic conditions. Models were evaluated through forward-chaining validation and temporal holdout tests, and SHAP values were used to examine how LightGBM used each predictor. RESULTS AND DISCUSSION Crop sequences were largely dominated by cereals, with little diversification even in longer sequences. Crop selection patterns were strongly associated with the previous crops and farm typology, but pedoclimatic conditions and previous management played a minor role. The DL model achieved a slightly higher overall accuracy, particularly for dominant crops, while the ML model provided a more balanced performance across different crops and enabled interpretation through SHAP. SIGNIFICANCE The analysis separates two distinct tasks: identifying multi-year crop sequences and predicting the crop family grown each year. The sequence analysis shows why diversification cannot be assessed from sequence length or crop counts alone. The predictive models could help generate locally plausible sequences for agri-environmental simulations, reducing reliance on standard rotations that poorly reflect observed field histories.","背景 作物序列对可持续农业至关重要，但我们对它们在单个田块中如何重复出现，以及哪些信息最能预测每年种植的作物知之甚少。丹麦的国家田块记录使这两个问题得以在长时间尺度上加以考察。 目标 我们识别了丹麦全国范围内重复出现的作物序列，并检验了年度作物科选择能在多大程度上由作物历史、农场特征、管理、价格、土壤和气候来预测。 方法 利用10年（2011−2020年）的国家田块级作物数据（每年约60万块田），我们开发了一种启发式算法来检测重复出现的作物序列模式。随后，我们训练了机器学习（LightGBM）和深度学习（TabNet）模型，基于前茬作物以及滞后的管理和农场结构预测因子及土壤气候条件来预测年度作物选择。模型通过前向链式验证和时间留出测试进行评估，并使用SHAP值考察LightGBM如何利用每个预测因子。 结果与讨论 作物序列在很大程度上以谷类作物为主，即使在较长的序列中也几乎没有多样化。作物选择模式与前茬作物和农场类型密切相关，而土壤气候条件和先前管理的作用较小。深度学习模型取得了略高的总体准确率，尤其是在优势作物上，而机器学习模型在不同作物之间提供了更均衡的表现，并可通过SHAP进行解释。 意义 该分析区分了两个不同的任务：识别多年作物序列和预测每年种植的作物科。序列分析表明，为什么不能仅凭序列长度或作物计数来评估多样化。预测模型有助于为农业环境模拟生成局部合理的序列，减少对不能很好反映观测田块历史的标准轮作方式的依赖。","Agricultural Systems",79,{"impact":217,"substance":18,"depth":17,"authority":19,"freshness":73,"relevant":21,"comment":218},16,"基于丹麦十年全国田块数据，用可解释AI预测作物选择，方法新颖、数据规模大，对农业信息化与轮作模拟有参考价值。",[220],{"name":214,"url":211},[26,27,28,222,223],"作物轮作","丹麦农业",[225,226],"丹麦 作物序列 机器学习","LightGBM 作物选择 预测","丹麦作物序列机器学习-3137","10.1016\u002Fj.agsy.2026.104982",{"doi":228,"openalex_id":230,"authors":231,"venue":214,"cited_by_count":35,"oa_url":211,"card":262,"direction":56,"ingested_from":59},"W7213919617",[232,234,237,240,242,244,246,248,251,253,256,259],{"name":233,"orcid":9},"João G. Serra",{"name":235,"orcid":236},"Edwin Haas","https:\u002F\u002Forcid.org\u002F0000-0003-0664-642X",{"name":238,"orcid":239},"Diego Ábalos","https:\u002F\u002Forcid.org\u002F0000-0002-4189-5563",{"name":241,"orcid":9},"Søren Kolind Hvid",{"name":243,"orcid":9},"Tommy Dalgaard",{"name":245,"orcid":9},"Lars Uldall-Jessen",{"name":247,"orcid":9},"Franca Giannini-Kurina",{"name":249,"orcid":250},"Meshach Ojo Aderele","https:\u002F\u002Forcid.org\u002F0009-0007-9381-6445",{"name":252,"orcid":9},"Bo Thiesson",{"name":254,"orcid":255},"Jørgen Eivind Olesen","https:\u002F\u002Forcid.org\u002F0000-0002-6639-1273",{"name":257,"orcid":258},"Klaus Butterbach‐Bahl","https:\u002F\u002Forcid.org\u002F0000-0001-9499-6598",{"name":260,"orcid":261},"Jaber Rahimi","https:\u002F\u002Forcid.org\u002F0000-0002-2754-2358",{"tldr":263,"method":264,"finding":265,"direction":56,"opportunity":266},"用丹麦十年田块数据识别轮作模式并预测年度作物选择。","60万田块年数据，启发式序列检测+LightGBM\u002FTabNet，SHAP解释。","作物序列以谷物为主、多样化低；作物选择主要由前茬和农场类型决定。","可结合可解释AI生成区域化可行轮作方案，替代脱离实际的标准化轮作假设。","2026-09-22T23:30:05.518307Z",{"id":269,"title":270,"url":271,"summary":272,"summary_zh":273,"content":9,"source_name":274,"source_url":271,"published_at":11,"category":12,"cover_url":9,"hotness":13,"is_selected":14,"score":275,"score_detail":276,"sources":280,"tags":282,"search_phrases":285,"slug":288,"view_count":35,"doi":289,"paper":290,"created_at":320},3134,"DAC-Grad-CAM: Defect-aware explainable AI for automated mango quality classification","https:\u002F\u002Fdoi.org\u002F10.1016\u002Fj.compag.2026.112443","Deep learning models achieve high fruit classification accuracy but lack interpretability; standard Grad-CAM produces diffuse attention without defect-specific localization. Existing explainable AI (XAI) methods for fruit grading rely on qualitative visual inspection, offering no quantitative validation of defect alignment, thereby limiting trust and regulatory acceptance in commercial quality control. This study presents an explainable mango classification framework that integrates classification performance with quantitative interpretability analysis through Defect-Aware Contrastive Grad-CAM (DAC-Grad-CAM). A new method that multiplicatively fuses gradient-based attention with an unsupervised LAB color-space bruise prior. A two-stage pipeline combining YOLO-11n detection with four classification architectures (ResNet18, MobileNetV3, VGG16, and a custom CNN). Evaluated on a three-class mango dataset (premium, class 1, class 2). Transfer-learning architectures substantially outperformed the custom CNN (ResNet18: 97.8%, VGG16: 97.6%, MobileNetV3: 96.5% vs 75.4%), a 22–23 percentage-point gap, with premium-class precision > 0.995. DAC-Grad-CAM produced the most concentrated explanations (entropy 10.02 versus 10.27 for Grad-CAM and 10.18 for SHAP). Validated against 50 manually annotated pixel-level bruise masks, it improved peak localization over standard Grad-CAM (pointing-game accuracy 0.561 versus 0.439) and, on the transfer-learning backbones, exceeded SHAP (0.659 versus 0.615) while requiring 34 ms per region against 4220 ms, a 124-fold reduction in cost. The Defect Focus Score (DFS) enabled task-specific validation, and the agreement between gradient and perturbation-based explanations (r = 0.68–0.75, p \u003C 0.001) confirmed that the localized features are task-genuine rather than method-specific artifacts. Under controlled perturbations emulating varied lighting, resolution, defocus, compression, and occlusion, defect-focused attention remained stable while classification accuracy degraded, and end-to-end throughput reached 17.4 fruit s⁻ 1 . The framework supports a confidence-based human-AI workflow (margin > 0.95, DFS > 0.37, r > 0.60) for trustworthy automated quality control in post-harvest operations.","深度学习模型虽能实现高精度水果分类，却缺乏可解释性；标准Grad-CAM产生的注意力分布弥散，无法实现缺陷特异性定位。现有用于水果分级的可解释人工智能（XAI）方法依赖定性视觉检查，无法对缺陷对齐进行定量验证，从而限制了其在商业质量控制中的可信度与法规接受度。本研究提出了一种可解释芒果分类框架，通过缺陷感知对比Grad-CAM（DAC-Grad-CAM）将分类性能与定量可解释性分析相结合。该方法将基于梯度的注意力与无监督LAB色彩空间瘀伤先验进行乘法融合。采用两阶段流水线，将YOLO-11n检测与四种分类架构（ResNet18、MobileNetV3、VGG16及自定义CNN）相结合。在三类芒果数据集（特级、一级、二级）上进行评估。迁移学习架构显著优于自定义CNN（ResNet18：97.8%，VGG16：97.6%，MobileNetV3：96.5% vs 75.4%），差距达22–23个百分点，特级类精确率> 0.995。DAC-Grad-CAM产生了最集中的解释（熵10.02，对比Grad-CAM的10.27和SHAP的10.18）。针对50个手动标注的像素级瘀伤掩膜进行验证，其峰值定位优于标准Grad-CAM（指向游戏准确率0.561 vs 0.439），且在迁移学习骨干网络上超过SHAP（0.659 vs 0.615），同时每区域仅需34 ms，而SHAP需4220 ms，成本降低124倍。缺陷聚焦分数（DFS）实现了任务特异性验证，基于梯度与基于扰动的解释之间的一致性（r = 0.68–0.75，p \u003C 0.001）证实了所定位特征是任务真实的，而非方法特异性伪影。在模拟不同光照、分辨率、散焦、压缩和遮挡的受控扰动下，缺陷聚焦注意力保持稳定，而分类准确率下降，端到端吞吐量达到17.4果·秒⁻¹。该框架支持基于置信度的人机协作工作流（margin > 0.95，DFS > 0.37，r > 0.60），可用于采后作业中可信赖的自动化质量控制。","Computers and Electronics in Agriculture",81,{"impact":217,"substance":277,"depth":278,"authority":19,"freshness":73,"relevant":21,"comment":279},23,19,"方法新颖、量化验证扎实的芒果品质分级可解释AI研究，对采后智能质检有直接参考价值。",[281],{"name":274,"url":271},[26,27,28,283,284],"采后处理","水果分级",[286,287],"芒果 品质分级 深度学习","DAC-Grad-CAM 芒果 缺陷检测","芒果品质分级深度学习-3134","10.1016\u002Fj.compag.2026.112443",{"doi":289,"openalex_id":291,"authors":292,"venue":274,"cited_by_count":35,"oa_url":271,"card":315,"direction":56,"ingested_from":59},"W7213923596",[293,295,297,299,301,303,306,308,311,313],{"name":294,"orcid":9},"Arshed Ahmed",{"name":296,"orcid":9},"Zhao Zhang",{"name":298,"orcid":9},"Ming Li",{"name":300,"orcid":9},"Shahram Hamza Manzoor",{"name":302,"orcid":9},"Zhanjiang Zhu",{"name":304,"orcid":305},"Rashed Ahmed","https:\u002F\u002Forcid.org\u002F0009-0000-7170-2200",{"name":307,"orcid":9},"Noor Gul",{"name":309,"orcid":310},"Mustafa Mhamed","https:\u002F\u002Forcid.org\u002F0000-0002-3106-669X",{"name":312,"orcid":9},"Bing Liu",{"name":314,"orcid":9},"Mahmoud A. Abdelhamid",{"tldr":316,"method":317,"finding":318,"direction":56,"opportunity":319},"提出DAC-Grad-CAM可解释框架，实现芒果缺陷感知分类与定量验证。","融合梯度注意力与LAB色空间瘀伤先验，YOLO-11n检测加四种分类网络。","ResNet18达97.8%精度，定位优于Grad-CAM与SHAP且快124倍。","可将缺陷感知XAI扩展至多水果多缺陷，并构建人机协同质控标准。","2026-09-22T23:30:02.407104Z"]