Prediction of Maize Grain Yield across Environments Using Machine Learning and Weather-derived Covariates

Accurate prediction of maize grain weight across diverse environments remains a major challenge because of complex interactions between genotype performance and variable climatic conditions. This study evaluated several machine learning approaches for predicting maize grain weight across environments by integrating phenotypic observations with weather-derived covariates. Data from multi-environment trials were analysed using Support Vector Regression, Random Forest, Artificial Neural Networks, Light Gradient Boosting Machine and Extreme Gradient Boosting models. Weather-based indices, including cumulative growing degree days at key phenological stages, were incorporated to account for environment-specific thermal variability. Model performance was assessed using error metrics across environments. The model performance ranking was XGBoost > LightGBM > RF > SVR > ANN. Among the approaches tested, XGBoost performed best on the training and testing sets, with MSE values of 0.011 and 0.057 and RMSE values of 0.109 and 0.239, respectively. XGBoost outperformed LightGBM, suggesting a greater capacity to capture nonlinear relationships and interactions between weather covariates and grain weight in the present dataset. The findings indicate that incorporating weather-derived information into boosting-based machine learning frameworks may improve grain-weight prediction across the environments examined. These results highlight the potential relevance of such models to maize breeding under variable climatic conditions.