Germany, with a focus on Brandenburg, where the drought after 2018 caused a strong decline in groundwater storage and increased pressure on regional water resources. This study evaluates the performance of the Global Gravity-based Groundwater Product, G3P, derived from GRACE and GRACE FO satellite gravimetry. It also assesses a downscaling framework that converts coarse groundwater storage anomalies from 0.5 ° to 1 km resolution. Two modelling approaches, Multiscale Geographically Weighted Regression (MGWR) and Random Forest (RF), were used with high-resolution hydroclimatic and land surface variables as predictors. Model performance was validated against in situ observations. Additionally, temporal trends (2002–2020) and seasonal dynamics were analysed to assess long-term groundwater changes. Results show that RF achieved higher predictive accuracy and better spatial representation than MGWR. The downscaled estimates improved cross correlation with in situ observations by 24.0% for RF and 21.0% for MGWR compared with the original GRACE-based groundwater storage anomalies. Trend analysis indicates a persistent decline in groundwater storage anomalies from 2002 to 2020, with stronger depletion after 2018. Seasonal analysis shows that wet season anomalies deepened from −90 mm in 2002 to −160 mm in 2020, while dry season anomalies increased from −55 mm to −178 mm. The high-resolution groundwater storage anomalies produced by the Random Forest framework provide improved spatial detail and support regional groundwater assessment and distributed hydrological modelling in Germany. • The framework downscaled GRACE data effectively. Random Forest outperformed Multiscale Geographically Weighted Regression. • The Global Gravity-based Groundwater Product (G3P) leverages GRACE/FO satellite gravimetry to improve groundwater monitoring. • Downscaled GWS anomalies from GRACE showed significant agreement with in situ wells data. • Overall, a decreasing trend in GWSA was observed across Germany, with some inconsistencies in the estimated time series.
Raza et al. (Tue,) studied this question.