Assimilating external observations of crop state in cropping system models is essential for making spatially explicit predictions of crop variables relevant in precision agriculture. Satellite-based leaf area index (LAI) estimates have been the most frequent variable used as a proxy of actual crop growth. However, additional information beyond LAI, like canopy N content, water content, and structure, can be retrieved from satellite observations. Including such variables by data assimilation directly is difficult because many crop models do not have corresponding state variables or the relationship between the observations and the process that regulates crop growth is unclear. Therefore, other approaches are required to include such information. In this study, we investigate the improvement in the predicted yield and feature impact on model outputs by using a hybrid approach that combines observations from Sentinel-1 and 2 time-series with the outputs from a process-based model embedded in a data assimilation framework and uses the Gradient-boosted trees regressor (GBTR) as predictive model. We used two regions with soybean fields: the US (13 K points) and Uruguay (400 K points). We found an advantage when using the GBTR as the predictive model (reduced RRMSE by ∼16%) compared to data assimilation. Adding the vegetation indices had a marginal improvement (reduced RRMSE by ∼1%), while the impact of adding reflectance and backscatter values was negative. The satellite-based features had a very small importance score, while features' impact on prediction was predominantly unclear, explaining the marginal predictive power added by satellite-based features. We found that features from the reproductive stages had the highest importance, while the importance of an index related to drought stress (NMDI) across the growing season provided insights for further improvement of data assimilation methods. However, more studies are required to better disentangle pathways towards further improvement in constraining crop models by ingesting satellite observations.

中文翻译：

---

超越叶面积指数的同化：利用机器学习的附加光谱信息进行特定地点的大豆产量预测

在种植系统模型中吸收作物状态的外部观测对于对精准农业中相关作物变量进行空间明确的预测至关重要。基于卫星的叶面积指数（LAI）估计值是最常用作实际作物生长指标的变量。然而，除 LAI 之外的其他信息，如冠层氮含量、含水量和结构，可以从卫星观测中检索。直接通过数据同化包含这些变量是很困难的，因为许多作物模型没有相应的状态变量，或者观测结果与调节作物生长的过程之间的关系不清楚。因此，需要其他方法来包含此类信息。在本研究中，我们通过使用混合方法来研究预测产量和特征对模型输出的影响的改进，该方法将 Sentinel-1 和 2 时间序列的观察结果与嵌入数据同化框架中的基于过程的模型的输出相结合并使用梯度增强树回归器（GBTR）作为预测模型。我们使用了两个有大豆田的地区：美国（13 K 点）和乌拉圭（400 K 点）。与数据同化相比，我们发现使用 GBTR 作为预测模型具有优势（RRMSE 降低约 16%）。添加植被指数有一定的改善（RRMSE 降低约 1%），而添加反射率和后向散射值的影响是负面的。基于卫星的特征的重要性得分非常小，而特征对预测的影响主要不清楚，这解释了基于卫星的特征增加的边际预测能力。我们发现生殖阶段的特征具有最高的重要性，而与整个生长季节的干旱胁迫（NMDI）相关的指数的重要性为进一步改进数据同化方法提供了见解。然而，需要更多的研究来更好地理清通过吸收卫星观测进一步改进约束作物模型的途径。