南京农业大学 | “深度学习+作物机理”预测小麦抽穗期

图1. 三个群体中小麦基因型来源的地理分布（A）以及分别选定的欧洲（B）和中国（C）小麦主要种植区的代表性地点。

Fig. 1. Geographic distribution of wheat genotypes sources in three populations (A) and representative sites of wheat main planting areas selected in Europe (B) and China (C), respectively. 

图2. 使用集成 WheatGrow-LSTM 方法预测小麦抽穗期的技术流程图。

Fig. 2. Technical flow chart for predicting wheat heading date using an integrated WheatGrow-LSTM method. 

图3.  来自三个数据集的 19 个试验环境中的小麦抽穗期分布。

Fig. 3. Distribution of wheat heading dates across 19 experimental environments from three datasets.

图4. 基于 APSIM-Wheat 模拟，对所有环境 (E1–E19) 进行 LOEO 评估，分析五个生长阶段 (GS1–GS5) 的关键气象变量分布。

Fig. 4. Distributions of key meteorological variables across five growth stages (GS1–GS5) during LOEO evaluation, based on APSIM-Wheat simulations for all environments (E1–E19). 

图5. 基于 WheatGrow 模拟，对所有环境 (E1–E19) 进行 LOEO 评估，五个生长阶段 (GS1–GS5) 的关键气象变量分布情况。

Fig. 5.Distributions of key meteorological variables across five growth stages (GS1–GS5) during LOEO evaluation, based on WheatGrow simulations for all environments (E1–E19). 

图6. 使用 APSIM-Wheat 模型 (A–C) 和 APSIM-Wheat-LSTM 模型 (D–F) 对三个数据集的预测小麦抽穗日期与观察到的小麦抽穗日期进行比较。

Fig. 6. Predicted versus observed wheat heading dates across three datasets using the APSIM-Wheat model (A–C) and the APSIM-Wheat-LSTM model (D–F).

图7. 使用 WheatGrow 模型 (A–C) 和 WheatGrow-LSTM 模型 (D–F) 对三个数据集预测的小麦抽穗日期与观察到的小麦抽穗日期进行比较。

Fig. 7.Predicted versus observed wheat heading dates across three datasets using the WheatGrow model (A–C) and the WheatGrow-LSTM model (D–F).

图8. 通过 LOEO 对三个群体（AGIS、GABI 和 iWheat）在所有测试环境下的小麦抽穗期预测准确性进行比较。

Fig. 8. Comparison of prediction accuracy for wheat heading date across all testing environments via the LOEO for the three populations (AGIS, GABI, and iWheat).

图9. 在 LOEO 交叉验证下，APSIM-Wheat-LSTM 模型在五个生长阶段（GS1 至 GS5）中通过 GradientSHAP 导出的特征归因值。

Fig. 9. Feature attribution values derived from GradientSHAP for the APSIM-Wheat-LSTM model under LOEO cross-validation across five growth stages (GS1 to GS5). 

图10. 在 LOEO 交叉验证下，WheatGrow-LSTM 模型在五个生长阶段（GS1 至 GS5）中通过 GradientSHAP 导出的特征归因值。

Fig. 10. Feature attribution values derived from GradientSHAP for the WheatGrow-LSTM model under LOEO cross-validation across five growth stages (GS1 to GS5).