How Can GRACE Satellite Observations Be Downscaled into Fine-Scale Groundwater Anomalies? A Hybrid Machine Learning and Uncertainty Quantification Framework

The Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow-On (GRACE-FO) missions provide terrestrial water storage anomalies (TWSA) at resolutions too coarse for local groundwater assessment. We developed a two-stage framework to downscale monthly GRACE/GRACE-FO TWSA from 0.5° to 0.01° over Rachna Doab, Pakistan, for 2002-2024. Seasonal-Trend decomposition using LOESS (STL) bridged the mission gap. Extreme Gradient Boosting (XGBoost) first refined TWSA to 0.1°, followed by a feed-forward neural network (FNN) at 0.01°. Monthly domain-mean adjustment and native-scale reaggregation retained basin-scale consistency, while bootstrap and Monte Carlo dropout ensembles quantified pixel-wise uncertainty. Of seven tested configurations, XGBoost-FNN performed best, with Nash-Sutcliffe and Kling-Gupta efficiencies of 0.90 and 0.89, respectively, and median pixel uncertainty of 0.32 cm. Seasonal comparison with monitoring wells yielded a mean correlation of 0.87, and the downscaled fields resolved sub-basin storage contrasts not visible in native GRACE.