Improving the Spatial Resolution of GRACE-Derived GFZ G3P Groundwater Storage Anomaly Product Through Unsupervised Deep Learning Downscaling

Abstract

While groundwater is the second largest freshwater reservoir on Earth, it remains difficult to monitor, relying mainly on sparse and often unavailable well measurements. The GRACE/GRACE–FO missions enabled the estimation of regional terrestrial water storage changes, although their coarse spatial resolution remains a major limitation. In this study, an unsupervised deep learning approach was exploited to downscale groundwater storage anomalies (GWSA) from the GRACE-derived GFZ G3P product from 0.5° to 0.1° spatial resolution, covering from 2003 to 2023. Thirteen ERA5 climatic variables and a digital elevation model were used as predictors. The framework was tested over France and validated in the Paris Basin using piezometric level information from the IGRAC GGMN open database. Results show good agreement between the downscaled and the original G3P product, with average temporal and spatial correlations of r = 0.99. When compared with in situ measurements, the downscaled product outperforms the original G3P data at the basin level (r = 0.67 vs. r = 0.62), while maintaining the accuracy at the local well scale (r = 0.39 vs. r = 0.38, respectively). The results highlight the potential of machine learning for GRACE super–resolution, while the main limitations are discussed.