基于可解释性深度学习的遥感玉米估产研究

Remote Sensing-based Maize Yield Estimation via Explainable Deep Learning

摘要:

准确获取区域玉米产量对制定农业政策和促进国民经济发展具有重要意义。长短期记忆网络（Long short-term memory，LSTM）和Transformer模型具有处理时序遥感数据优势，广泛应用于农作物产量估测。为兼顾LSTM对局部时序信息提取和Transformer对全局依赖的高效表达，本文将遥感多参数和玉米单产作为输入数据，分别构建Transformer Encoder-LSTM（TFEL）、Transformer-LSTM（TFL）和Transformer估产模型，利用贝叶斯优化算法获取模型最佳隐藏层数量、学习率等超参数，进而估测山西省太原盆地和上党盆地的玉米单产;基于沙普利值的加性解释方法（Shapley additive explanations，SHAP）对混合深度学习模型进行全局事后可解释性分析。结果表明，相较于TFL模型（R2=0.62，P<0.01，RMSE为974.14 kg/hm2，MAPE为13.50%，NRMSE为15.14%）和Transformer模型（R2=0.53，P<0.01，RMSE为1 028.76 kg/hm2，MAPE为19.13%，NRMSE为16.16%），TFEL模型估产精度较高（R2=0.72，P<0.01，RMSE为756.43 kg/hm2，MAPE为10.58%，NRMSE为11.86%）;区域单产估测结果呈现南北部高产、东西部低产的空间特征，估测产量与统计产量间呈现较好的线性相关性，表明TFEL模型具有良好的鲁棒性和泛化能力;在TFEL模型和TFL模型中，双波段增强植被指数（Two-band enhanced vegetation index，EVI2）和绿色叶绿素植被指数（Green chlorophyll vegetation index，GCVI）特征重要性排序位于前列;相较于TFL模型，TFEL模型更有效聚焦关键遥感参数，稳定识别和量化关键遥感参数对产量估测的贡献度，具有较高产量估测精度。综上所述，基于Transformer和LSTM的混合模型在玉米产量估测中展现良好的应用潜力，可为区域尺度作物估产提供理论与方法参考。

Abstract:

Accurate acquisition of maize yield information was recognized as critical for formulating agricultural policies and supporting national economic development. Long short-term memory (LSTM) networks and Transformer models were employed for crop yield estimation due to their respective strengths in processing remote sensing time series data. To balance LSTM's capacity for capturing local temporal dependencies with the Transformer's efficiency in modeling global relationships, multi-source remote sensing parameters and maize yield were used as inputs to construct three yield estimation models: Transformer Encoder-LSTM (TFEL), Transformer-LSTM (TFL), and pure Transformer. A Bayesian optimization algorithm was applied to determine the optimal combinations of hidden layer size, learning rate, and other hyperparameters. These models were then used to estimate county-scale maize yields in the Taiyuan and Shangdang Basins of Shanxi Province. The Shapley additive explanations (SHAP) method was employed to quantify the contribution of each remote sensing feature within the hybrid models. The TFEL model demonstrated superior estimation accuracy (R2 was 0.72, P<0.01, RMSE was 756.43 kg/hm2, MAPE was 10.58%, NRMSE was 11.86%) compared with both the TFL model (R2 was 0.62, P<0.01, RMSE was 974.14 kg/hm2, MAPE was 13.50%, NRMSE was 15.14%) and the Transformer model (R2 was 0.53, P<0.01, RMSE was 1 028.76 kg/hm2, MAPE was 19.13%, NRMSE was 16.16%). The spatial distribution of estimated yields showed higher values in the northern and southern regions and lower values in the eastern and western regions. A strong linear relationship was observed between estimated and statistical yields, confirming the TFEL model's generalization capability. Results revealed that the two-band enhanced vegetation index (EVI2) and green chlorophyll vegetation index (GCVI) provided the greatest contributions to maize yield estimation in both the TFEL and TFL frameworks. Compared with the TFL model, the TFEL model focused more effectively on key remote sensing parameters, and consistently identified and quantified their contributions to yield estimation, thereby achieving higher estimation accuracy. In summary, the hybrid model based on Transformer and LSTM showed promising application potential in maize yield estimation, which can provide theoretical and method reference for regional crop yield assessment.