Climate-driven maize yield prediction across Uganda’s ZARDI zones using pooled learning and uncertainty-aware modeling

Abstract

Accurate maize yield prediction is essential for agricultural planning and food security, yet remains challenging in data-scarce, heterogeneous agroecological settings. This study investigates climate-driven maize yield prediction across Uganda’s Zonal Agricultural Research and Development Institute (ZARDI) regions using a pooled modeling framework that mitigates the problem of extremely small zone specific sample sizes. Seasonal climatic variables, including rainfall, temperature, solar radiation, and soil moisture, collected from 2018 to 2021, were integrated with zone indicators and evaluated under a strict leave-one-year-out (LOYO) cross-validation protocol to prevent temporal information leakage. Regularized linear regression (Ridge) achieved the most reliable generalization performance, with a root mean squared error (RMSE) of approximately 0.98 t ha−1 and mean absolute error (MAE) of 0.57 t ha−1, marginally outperforming global-mean (RMSE ≈ 0.99 t ha−1) and ZARDI-mean (RMSE ≈ 1.05 t ha−1) baselines. More complex models, including a constrained XGBoost regressor, did not yield consistent performance gains under LOYO evaluation. Zone-specific analysis revealed substantial heterogeneity, with some regions exhibiting moderate predictive skill (MAE < 0.25 t ha−1), while others showed weak climate–yield coupling and negative coefficients of determination. To address overconfidence in the face of data scarcity, prediction uncertainty was explicitly quantified using conformal prediction intervals based on out-of-fold residuals. Global prediction intervals achieved empirical coverage of 91.7% and 96.3% at the 90% and 95% nominal levels, respectively, indicating well-calibrated uncertainty estimates. Zone-specific interval widths varied markedly, ranging from approximately 0.84 t ha−1 in better-performing regions to over 4.6 t ha−1 in the most uncertain zones. Model interpretability analysis using Shapley Additive explanations identified rainfall, soil moisture, solar radiation, and seasonal timing as the dominant drivers of yield predictions, while zone identifiers contributed minimally. The proposed framework provides a transparent, defensible basis for yield assessment and underscores the need to integrate additional agronomic and remotely sensed data to improve predictive skill in regions with weak climate–yield relationships.