CSA | 甘肃农业大学殷文教授基于深度学习揭示小麦-玉米间作产量的非线性气候驱动机制

Understanding how climatic variability influences crop yield is essential for improving agricultural resilience in climate-vulnerable regions. This study develops an interpretable deep learning framework to analyse wheat and maize yield responses in a long-term conservation-based strip intercropping system. By integrating daily meteorological sequences with phenology-aligned climatic summaries, the model captures both short-term weather variability and cumulative stage-specific climate effects while maintaining high predictive accuracy. The results revealed distinct climatic sensitivity structures for the two crops. Wheat yield was most strongly influenced by precipitation during flowering and radiation during jointing to booting, with additional sensitivity to thermal accumulation and hot days during grain filling. In contrast, maize yield was primarily controlled by precipitation during rapid vegetative growth to pre-tasselling and by growing degree days during tasselling, silking, and grain filling. Attention, SHapley Additive exPlanations value, and perturbation analyses consistently showed that climate–yield relationships were strongly nonlinear. Yield responses exhibited clear threshold and saturation behaviours, with diminishing benefits under excessive moisture and rapid declines once thermal conditions exceeded critical limits. These findings demonstrate that yield formation in this intercropping system is governed by stage-specific climatic thresholds rather than simple linear responses. The framework provides a practical tool for identifying vulnerability windows, climatic thresholds, and adaptation limits, offering useful guidance for climate-resilient management in dryland and data-scarce agroecosystems.

了解气候变率如何影响作物产量对于提高气候脆弱地区的农业韧性至关重要。本研究开发了一个可解释的深度学习框架，用于分析长期保护性条带间作系统中小麦和玉米的产量响应。该模型将每日气象序列与物候期气候概要相结合，在保持高预测精度的同时，既能捕捉短期天气变率，又能捕捉各生长阶段的累积气候效应。结果揭示了两种作物截然不同的气候敏感性结构。小麦产量主要受开花期降水和拔节至孕穗期辐射的影响，灌浆期对热量积累和高温日数也较为敏感。相比之下，玉米产量主要受快速营养生长至抽穗前期降水和抽穗、吐丝及灌浆期生长积温的控制。注意力分析、SHapley加性解释值分析和扰动分析均一致表明，气候与产量之间的关系具有显著的非线性特征。产量响应表现出明显的阈值和饱和特性，水分过多时收益递减，一旦温度条件超过临界值，产量便迅速下降。这些发现表明，该间作系统中的产量形成受特定阶段的气候阈值控制，而非简单的线性响应。该框架为识别脆弱性窗口、气候阈值和适应极限提供了一种实用工具，可为干旱地区和数据匮乏的农业生态系统中的气候适应型管理提供有益指导。

图1 可解释深度学习框架的工作流程。CMA：中国气象局；IDW：反距离加权；IQR：四分位距；Z-score：Z-score标准化；GDD：生长积温；BiLSTM：双向长短期记忆网络；Adam：自适应矩估计，Adam通过估计梯度的一阶矩和二阶矩动态调整每个参数的学习率，从而实现高效稳定的收敛；SHAP：Shapley加性解释。

本研究构建了一种专为小样本农业生态数据量身定制的可解释深度学习方法论。在对气象与产量数据进行严格的时空对齐与特征去噪后，研究创新性地提出了“双轨制”特征工程：将包含每日生长度日（GDD）的动态气象序列，与15个基于作物关键物候期的静态气候摘要（如均温、累计降水等）相融合。基于此，开发了双输入BiLSTM-Attention模型（DBLstmAM）。

本研究提出了一种创新的可解释深度学习框架，通过将高分辨率逐日气象序列与基于关键物候期的气候特征相融合，突破了传统的单纯产量预测，实现了对小麦-玉米间作系统气候胁迫“时间、地点与机制”的精准诊断。研究证实，在数据受限地区仅依赖气候输入即可获得高度鲁棒的预测结果。更重要的是，解释性分析揭示了间作系统的产量受控于阶段特异性的脆弱窗口、非线性的气候敏感性以及跨季节的气候记忆效应，证明了系统的气候适应性受限于动态的临界阈值，而非无限的线性缓冲。本研究成功将可解释人工智能（XAI）从单纯的预测技术重塑为区域农业可持续性的“诊断引擎”。未来的研究与应用应进一步利用该框架，精准界定复杂农业生态系统的“适应极限”与“气候安全空间”。这不仅为干旱和数据稀缺地区应对日益加剧的气候不确定性提供了一条将复杂气候-产量关系转化为实用决策知识的路径，也将有力支撑未来气候智慧型农业与资源节约型管理策略的科学规划与落地。