Abstract
Monitoring groundwater storage anomalies (GWSA) at high spatial resolutions is critical for sustainable water resource management and hydrological applications. The Gravity Recovery and Climate Experiment (GRACE) satellite provides unique insights into changes in terrestrial water storage but is limited by its coarse spatial resolution. In this study, GRACE observations, in situ groundwater levels, and Global Land Data Assimilation System (GLDAS)-derived hydrological variables were integrated to downscale GWSA from 1° × 1° to 0.25° × 0.25° in the Breede Water Management Area, South Africa (2002–2022). Three machine learning models, namely, random forest (RF), support vector regression (SVR), and artificial neural networks (ANN), were tested. RF consistently outperformed the others ( R 2 = 0.86, Nash–Sutcliffe efficiency (NSE) = 0.75, root mean square error (RMSE) = 0.19) and was selected for downscaling. The downscaled GWSA revealed spatial heterogeneity, with higher recharge in the eastern mountainous regions and persistent depletion in the western lowlands. Temporal patterns captured major droughts (2002–2006, 2015–2021) and recovery phases (2008–2010, 2020–2022), consistent with the observed hydroclimatic variability. Validation against in situ boreholes (NSE up to 0.55, RMSE = 19–25 mm, R 2 = 0.77–0.81) confirmed strong agreement, although some discrepancies reflected aquifer heterogeneity and effects of groundwater abstractions. Uncertainties remain due to GRACE errors, GLDAS predictor biases, and limited borehole coverage, which constrain model generalization. Despite these limitations, this study demonstrates the potential of RF-based downscaling to improve GRACE groundwater monitoring in data-scarce regions and to inform adaptive water management under increasing climate variability.