Lagged Response of Groundwater Storage to Extreme Precipitation Using Machine-Learning-Downscaled GRACE Data at the Watershed Scale

Understanding how groundwater storage responds to extreme precipitation is essential for assessing aquifer resilience under climate variability. In this study, we developed a 1 km groundwater storage anomaly (GWSA) dataset for the Baiyangdian Watershed from 2002 to 2024 by downscaling GRACE observations with a Light Gradient Boosting Machine (LightGBM) model. The downscaled GWSA showed good consistency with independent hydrological datasets, including GLDAS and groundwater-level anomalies. Based on the downscaled product, we characterized long-term groundwater changes and quantified GWSA responses to extreme precipitation events (EPEs). Groundwater storage exhibited three distinct phases: rapid depletion before 2014 (−1.35 cm/yr), a slower decline during 2014–2019 (−1.04 cm/yr), and marked recovery after 2020 (+3.45 cm/yr). Spatially, GWSA generally increased from the southwest to the northeast of the watershed. Composite analysis of 11 EPEs revealed a delayed groundwater response, with the strongest signal occurring approximately two months after precipitation. Monthly effective precipitation was more closely associated with GWSA recovery than short-duration daily precipitation extremes, and the response was stronger in plains than in mountainous areas. These findings indicate that EPEs provide episodic recharge pulses, while sustained groundwater recovery depends on cumulative climatic inputs and human water-management influences.