[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"item-2505":3},{"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":24,"tags":26,"view_count":32,"doi":33,"paper":34,"created_at":70},2505,"A continually evolving knowledge-guided deep learning framework for daily maize yield formation","https:\u002F\u002Fdoi.org\u002F10.1016\u002Fj.agsy.2026.104981","CONTEXT As multi-source agricultural datasets expand, accurate and interpretable simulation of crop yield formation is increasingly feasible and important for food security. Process-based models are widely used for yield forecasting but require laborious regional calibration. Artificial intelligence models learn physiological patterns from large datasets but lack physiological consistency and generalize poorly under data-scarce conditions. OBJECTIVE We developed AgroEvoDeep-Yield, a knowledge-guided model that encodes mechanistic understanding through its neural architecture, enabling explicit and interpretable simulation of daily maize yield formation from flowering to maturity. We further established a framework that supports continual model evolution as new data become available. METHODS The model was pretrained on APSIM-simulated grain number and daily grain biomass sequences to impose physiological consistency. Following pretraining, the model evolved through two continual learning stages. First, field-level continual learning using 796 field observations reduced biases inherited from APSIM. The model was then adapted to regional-scale multi-source inputs using long-term records from 945 meteorological stations and expert-constrained cultivar-parameter optimization. RESULTS AND CONCLUSIONS After field-level continual learning, AgroEvoDeep-Yield reduced grain number and yield simulation errors by 29% and 24%, respectively, relative to APSIM, and outperformed conventional machine learning models in out-of-region prediction. Region-level continual learning reduced yield RMSE to 1.09 t\u002Fha, and expert-constrained cultivar-parameter optimization further reduced it to 0.80 t\u002Fha. SIGNIFICANCE This study demonstrates how a knowledge-guided model can be repeatedly adapted as new data become available, providing a practical framework for continual model evolution and a basis for future agricultural decision support.","背景 随着多源农业数据集的不断扩展，准确且可解释地模拟作物产量形成过程日益可行，对粮食安全也愈发重要。基于过程的模型广泛用于产量预测，但需要耗费大量精力进行区域校准。人工智能模型能够从大规模数据集中学习生理模式，但缺乏生理一致性，在数据稀缺条件下泛化能力较差。目标 我们开发了AgroEvoDeep-Yield，这是一种知识引导模型，通过其神经网络架构编码机理认识，能够对玉米从开花到成熟的逐日产量形成过程进行显式且可解释的模拟。我们还建立了一个框架，以支持在新数据可用时模型的持续演化。方法 该模型在APSIM模拟的粒数和逐日籽粒生物量序列上进行预训练，以施加生理一致性约束。预训练后，模型通过两个持续学习阶段进行演化。首先，利用796个田间观测数据进行田块级持续学习，以减少继承自APSIM的偏差。随后，利用945个气象站的长期记录和专家约束的品种参数优化，将模型适配到区域尺度多源输入。结果与结论 经过田块级持续学习后，AgroEvoDeep-Yield相对于APSIM将粒数和产量模拟误差分别降低了29%和24%，并在区域外预测中优于传统机器学习模型。区域级持续学习将产量RMSE降至1.09 t\u002Fha，专家约束的品种参数优化进一步将其降至0.80 t\u002Fha。意义 本研究表明，知识引导模型可以在新数据可用时被反复适配，为模型持续演化提供了实用框架，并为未来农业决策支持奠定了基础。",null,"Agricultural Systems","2026-09-14T00:00:00Z","论文",10,false,85,{"impact":17,"substance":18,"depth":19,"authority":20,"freshness":21,"relevant":22,"comment":23},22,23,18,14,8,1,"知识引导的深度学习框架实现玉米逐日产量形成模拟，误差显著低于APSIM且支持持续学习，方法新颖、数据规模扎实，对智慧农业决策支持有参考价值。",[25],{"name":10,"url":6},[27,28,29,30,31],"智慧农业","农业人工智能","产量预测","玉米","作物模型",0,"10.1016\u002Fj.agsy.2026.104981",{"doi":33,"openalex_id":35,"authors":36,"venue":10,"cited_by_count":32,"oa_url":6,"card":63,"direction":67,"ingested_from":69},"W7212532256",[37,39,41,44,47,50,52,55,57,60],{"name":38,"orcid":9},"Junji Ou",{"name":40,"orcid":9},"Wenyao Yan",{"name":42,"orcid":43},"Fangzheng Chen","https:\u002F\u002Forcid.org\u002F0009-0006-5061-6227",{"name":45,"orcid":46},"Tao Ye","https:\u002F\u002Forcid.org\u002F0000-0002-5037-8410",{"name":48,"orcid":49},"Ke Liu","https:\u002F\u002Forcid.org\u002F0000-0002-8343-0449",{"name":51,"orcid":9},"Matthew Tom Harrison",{"name":53,"orcid":54},"William D. Batchelor","https:\u002F\u002Forcid.org\u002F0000-0002-3881-6246",{"name":56,"orcid":9},"Yong Chen",{"name":58,"orcid":59},"Kelin Hu","https:\u002F\u002Forcid.org\u002F0000-0001-9321-0821",{"name":61,"orcid":62},"Puyu Feng","https:\u002F\u002Forcid.org\u002F0000-0003-4845-9876",{"tldr":64,"method":65,"finding":66,"direction":67,"opportunity":68},"提出知识引导的深度学习框架AgroEvoDeep-Yield，实现玉米逐日产量形成的可解释模拟与持续","基于APSIM模拟数据预训练，结合796个田间观测和945个气象站数据持续学习。","田间持续学习使粒数和产量误差比APSIM降低29%和24%，区域优化后RMSE降至0.80 t\u002Fha","农业人工智能与决策模型","可探索将知识引导与持续学习框架迁移至其他作物，并融合实时遥感与物联网数据实现动态产量预测。","openalex","2026-09-15T23:30:04.611750Z"]