Spatio-Temporal Downscaling of GRACE/GRACE-FO Groundwater Storage Anomalies Using XGBoost Reveals Sub-Regional Heterogeneity

Abstract

This study was carried out in East Java Province, Indonesia (47,800 km², 7°12'–8°48' S, 111°00'–114°04' E), a densely populated tropical region with over 40 million inhabitants. The province is characterized by a complex hydrogeology, including volcanic, alluvial, and karst aquifers, alongside intensive agriculture covering 70% of land irrigated by groundwater, and significant water stress from droughts, urbanization, and a variable monsoon climate. This study aimed to evaluate spatio-temporal downscaling of GRACE/GRACE-FO terrestrial water storage anomalies (TWSA) at 0.5°- 0.25° resolution using XGBoost machine learning method. The framework was integrated with Global Land Data Assimilation System (GLDAS) hydrological variables, including soil moisture, canopy storage, and precipitation, for the period 2005–2024. Furthermore, gap-filling during the GRACE-FO transition and residual correction were applied to ensure temporal consistency and spatial enhancement. Groundwater storage anomalies (GWSA) were derived using a mass balance residual method. The results of downscaling showed that there was sub-regional heterogeneity in GWSA. Northern alluvial region experienced dry-season depletion (–5 to –10 cm) with slow recharge, while southern volcanic region showed stable to moderate trends (+0.82 cm/year overall). Model performance (CC=0.89, NSE=0.78, RMSE=5.19 cm, RMSE*=0.46) maintained the integrity of GRACE signal and identified a lagged GWSA response to monsoon rainfall of 2–3 months. Additional evaluation using temporal data splitting indicates reduced predictive performance under extrapolation conditions, highlighting limitations for long-term forecasting while confirming robustness for historical reconstruction. These results provided a robust basis for targeted groundwater management in data-scarce East Java, where extraction remained unsustainable.