Machine-Learning Downscaling of GRACE/GRACE-FO for Basin-Scale Groundwater Storage Assessment in a Semi-Arid Basin

The Gravity Recovery and Climate Experiment (GRACE/GRACE-FO) missions provide terrestrial water storage anomalies (TWSA) at coarse spatial resolution (300 km), limiting their application in medium-sized basins. This study develops a machine-learning framework to enhance the spatial interpretability of GRACE mascon TWSA within the 48,000 km² Hadejia-Jama’are River Basin, Nigeria. Hydroclimatic predictors derived from TerraClimate, Global Land Data Assimilation System (GLDAS), and Climate Hazards Group InfraRed Precipitation with Stations (CHIRPS) were integrated within a unified 4 km spatial framework. Four machine learning models were evaluated, including Random Forest (RF), Gradient Boosting, Histogram Gradient Boosting, and a Multi-Layer Perceptron. The RF model achieved the highest skill in reproducing mascon-scale TWSA (R² = 0.937; NSE = 0.937; RMSE = 4.36 cm). Aggregation of the 4 km fields back to the mascon scale preserved basin-integrated mass (R² = 0.94), confirming consistency with the original GRACE signal. The resulting groundwater storage anomaly (GWSA) fields resolve sub-basin spatial gradients and seasonal recharge-depletion cycles that are not discernible in the native product. Validation against 31 monitoring wells yielded moderate temporal agreement (Pearson correlation coefficient, r = 0.656), with magnitude discrepancies attributable primarily to scale mismatch and hydrogeological heterogeneity. While not a substitute for in-situ monitoring, the downscaled product enhances basin-scale groundwater assessment in data-scarce semi-arid regions. The framework is transferable to comparable basins and supports regional drought monitoring and water-resource management.