Bayesian Gaussian Process Regression for Interpretable Groundwater Prediction in a Climate-Sensitive Region of Malawi

Predicting groundwater levels in regions with limited hydrogeological data remains a significant challenge. This study proposes two Gaussian Process Regression (GPR) models with different input features to reflect diverse hydrogeological conditions. To assess the influence of spatial, geophysical, and climatic factors, a Random Forest (RF) model is used to analyse the residuals. This integrated machine learning framework enhances understanding of prediction uncertainty by combining GPR-based forecasting with RF-driven residual diagnostics. Applied to borehole and multi-year rainfall data from Chikwawa District, Malawi, the models not only predict groundwater levels but also reveal how spatial, hydrogeological, and temporal factors influence prediction errors. Notably, rainfall during specific periods shows strong correlations with residual patterns, suggesting links to aquifer recharge dynamics. In addition, preliminary groundwater chemistry observations indicate possible impacts of rainfall variability on water quality, though these associations were statistically inconclusive. By incorporating temporally explicit rainfall variables, the study introduces a novel perspective for interpreting model behaviour in climate-sensitive, data-scarce environments. This approach demonstrates how residual analysis can uncover critical sources of uncertainty, offering practical insights to support decision-making and improve sustainable groundwater management strategies in data-scarce, climate-sensitive regions.