基于时空注意力机制的小麦气象产量预测LSTM模型改进

An Improved LSTM Model for Wheat Meteorological Yield Prediction Based on Spatial-Temporal Attention Mechanism

摘要: 针对市域小麦气象产量预测中数据稀疏导致的过拟合及生育期先验知识利用不足的问题，本文提出一种基于时空双层次注意力机制的LSTM改进模型(ST-Attention LSTM)。该模型在标准LSTM主干网络中嵌入生育期级与特征级双层注意力模块，通过分生育期独立学习的权重矩阵显式编码物候学先验，实现“先选生育期、再选气象指标”的递进式特征提纯，以可控参数增量换取显著性能提升。基于豫北地区6个地市2000~2019年共108组样本开展留一市交叉验证，实验结果表明：所提模型在预测精度上优于XGBoost、Transformer等基准方法；消融实验证实移除注意力机制将导致显著性能退化；权重可视化揭示的生育期梯度分布与气象因子主导性客观反映小麦生理规律。该模型兼具高精度、可解释性与训练稳定性，为市域尺度气候风险评估提供了轻量化、可落地的智能解决方案。

Abstract: To address the overfitting caused by data sparsity and insufficient utilization of phenological prior knowledge in urban-scale wheat meteorological yield prediction, this paper proposes an improved LSTM model based on spatial-temporal dual-level attention mechanism (ST-Attention LSTM). The model embeds growth-stage-level and intra-feature attention modules into the standard LSTM backbone, explicitly encoding phenological priors through independently learned weight matrices for each growth stage, achieving progressive feature refinement via a “select growth stage first, then meteorological factors” hierarchy with controllable parameter increment. Leave-one-city-out cross-validation was conducted on 108 samples from 6 cities in northern Henan Province during 2000~2019. Experimental results show that the proposed model outperforms benchmark methods such as XGBoost and Transformer in prediction accuracy; ablation studies confirm that removing the attention mechanism leads to significant performance degradation; weight visualizations reveal gradient distributions across growth stages and dominant meteorological factors that objectively reflect wheat physiological patterns. The model achieves high accuracy, interpretability, and training stability, providing a lightweight and practical intelligent solution for urban-scale climate risk assessment.