An Explainable Ensemble Model for ZARDI-Wise Maize Yield Prediction Using Climate Data in Uganda

Abstract Accurate maize yield prediction is crucial for food security in countries like Uganda, where diverse agro-ecological zones with varying climates make agricultural planning challenging. This study proposes an explainable, zone-specific maize yield prediction framework tailored to the ten Zonal Agricultural Research and Development Institute (ZARDI) regions in Uganda. Recognizing the agro-ecological heterogeneity across these zones, the framework employs an optimized XGBoost regressor for each ZARDI, enabling localized modeling of climatic–yield relationships. Data preprocessing involved outlier removal based on a two-standard-deviation threshold to improve robustness, followed by mutual information-based feature selection to retain the five most informative climatic variables per zone. Hyperparameter optimization was conducted via grid search cross-validation to enhance model performance. The results indicate substantial variability in predictive accuracy across zones. Mbarara, Buginyanya, and Abi recorded low Mean Absolute Percentage Errors (MAPE) of 7.07%, 15.28%, and 32.41%, respectively. In comparison, zones such as Nabuin and Rwebitaba experienced higher errors exceeding 86%, due to data scarcity and climatic variability. LIME and SHAP analyses were integrated to provide global and local interpretability. They identified rainfall, temperature range, and soil moisture as dominant yield predictors, with their relative importance varying by zone. The proposed framework offers a transparent, data-driven decision-support tool for region-specific agricultural planning.