Analyses of crop yield dynamics and the development of a multimodal neural network prediction model with G×E×M interactions.

This study investigated how genotype, environment, and management (G×E×M) interactions influence yield and highlight the importance of accurate, early yield predictions for effective farm management and enhancing food security. We developed a yield prediction model capable of determining field-level outputs based on comprehensive data inputs, including genotype, spatial, temporal, environmental, and management factors. Among tested models-LASSO, Random Forest, XGBoost, single-modal CNN-DNN, and multimodal CNN-DNN-the multimodal CNN-DNN ensembled with XGBoost demonstrated superior performance. Applied to the G2F dataset covering 21 states from 2014 to 2021 across various treatments (i.e., standard, drought, irrigation, disease trials), the model excelled particularly in stable historical yield settings (RMSE 2.36 Mg/ha for standard treatment) with an overall RMSE of 2.45 Mg/ha. Additionally, we introduced an empirical tool for identifying high-yield hybrids suitable for standard and challenging conditions. Exploratory analysis confirmed that crop yields vary greatly by hybrid and location interaction and that late planting generally yields less than standard timing. Customized management strategies based on specific local and hybrid conditions are crucial for optimal yield outcomes.