Machine Learning-based Downscaling of GRACE Data to Enhance Assessment of Spatiotemporal Evolution of Coastal Plain Groundwater Storage

Groundwater storage is vital for regional water security, particularly in hydrologically complex coastal plains. However, the coarse spatial resolution of GRACE satellite data limits its application in local-scale groundwater studies. This study presents a machine learning–based framework to downscale GRACE-derived terrestrial water storage anomalies (TWSA) using high-resolution (4 km × 4 km) TerraClimate datasets over the Hang-Jia-Hu (HJH) Plain in eastern China. Three machine learning models—Random Forest (RF), XGBoost (XGB), and LightGBM (LGBM) —were evaluated, with RF showing the best performance. The downscaled groundwater storage anomaly (GWSA) data were validated against 68 in-situ groundwater level (GWL) time series, resulting in an average correlation improvement of 16% at 45% of unconfined aquifer stations and 23% at 69% of confined aquifer stations. The analysis revealed a sharp GWS decline before 2011 due to over-extraction, followed by gradual recovery after groundwater regulation. Shapely Additive Explanations (SHAP) analysis identified the Palmer Drought Severity Index (PDSI), actual evapotranspiration (AET), and vapor pressure (VAP) as key model drivers. A leakage intensity index was developed using downscaled GWSA and GWL data, highlighting spatial variations in aquifer connectivity and identifying zones with strong vertical recharge that may buffer seawater intrusion. These results demonstrate that the proposed downscaling approach not only enhances the spatial resolution and accuracy of GRACE data but also enables improved assessment of groundwater system behavior in coastal regions, supporting more informed water resource management.