Groundwater recharge modeling and management implications using WetSpass and random forest in Kobo Valley, Ethiopia

Abstract

ABSTRACT Accurate estimation of groundwater recharge is essential for sustainable water resources management, particularly in data-scarce regions. This study develops a computationally efficient groundwater recharge estimation framework by integrating a physically based water balance model with a machine-learning surrogate to enable rapid simulation under limited data availability. Dynamic predictors, including soil moisture (Soil Moisture Active Passive), seasonal precipitation (Climate Hazards Group InfraRed Precipitation with Station), and vegetation dynamics (Sentinel-2 NDVI), were combined with static variables such as topography, soil type, and land use to drive the random forest model. Uncertainty was quantified using Monte Carlo simulations to improve the robustness of recharge estimates. The simulated long-term mean recharge was 28.37 mm during the dry season (October–May) and 25.75 mm during the wet season (June–September). The random forest surrogate achieved high predictive performance (R2 = 0.99; root mean square error = 0.96) and reduced computation time from 410 seconds to 205 seconds in the dry-season simulation. These results demonstrate that the surrogate model can efficiently approximate physically based simulations while maintaining accuracy, enabling rapid identification of groundwater recharge hotspots. This integrated approach provides a scalable tool for groundwater assessment and supports water resource planning and climate adaptation in data-limited aquifer systems.