Accurate and timely crop yield prediction is crucial for ensuring food security and maintaining stable agricultural markets, yet leveraging high-temporal-resolution, multi-source data for this purpose presents a big challenge in effectively capturing intricate variable interactions across time, especially during periods of extreme weather events. To address this issue, this study introduces an Attention and Graph Isomorphism Network-enhanced Bidirectional Long Short-Term Memory network (AGB-LSTM), which is specifically designed for estimating countylevel soybean yield in the United States. The proposed model effectively integrated diverse high-temporalresolution time-series data (5-days), including Near-Infrared Reflectance of Vegetation (NIRv), Solar-Included chlorophyll Fluorescence (SIF), and Gross Primary Productivity (GPP), along with environmental covariates. The AGB-LSTM model achieves an accuracy of R2 = 0.67 and rRMSE = 14.46 %, which outperforms traditional and advanced machine learning methods tested, such as Random Forest (RF) (R2 = 0.52, rRMSE = 17.36 %) and Transformer (R2 = 0.60, rRMSE = 15.80 %). Furthermore, sensitivity analysis revealed the capability of the model for accurate and stable yield prediction 1 to 2 months before harvest. Notably, utilizing data with finer temporal resolution (5-day and monthly composite data) continuously improves performance. Specifically, the model performance is R2 = 0.67 and rRMSE = 14.46 % for the 5-day data, and R2 = 0.55 and rRMSE = 16.81 % for the 30-day data. We also highlight the robustness of the model under extreme climate events, maintaining strong performance with R2 = 0.50 and rRMSE = 21.32 %. Finally, the validation against USDA reported yields for major North American soybean regions in 2023 show that AGB-LSTM well captures the spatial patterns of soybean yield. These findings underscore the AGB-LSTM model as a promising and effective method for yield estimation, showcasing large potential for global crop yield forecasting.
更多
原文链接
分享
引用
