A machine-learning framework to improve GRACE/GRACE-FO terrestrial water storage anomaly in China

This is the authors' abstract. We don't add key points for this paper.

China mainland. Five machine-learning models, including Random Forest (RF), XGBoost, Decision Tree, Stochastic Gradient Descent (SGD), and K-Nearest Neighbors (KNN), were developed to downscale and reconstruct missing GRACE Total Water Storage Anomaly (TWSA) data from 0.5° to 0.25° spatial resolution. The RF model exhibited the best performance, with an R 2 of 0.982 and an RMSE of 2.167 cm, and it successfully reconstructed 34 missing monthly TWSA records from April 2002 to December 2022. The downscaled dataset shows a 4.1% improvement in spatial detail and an approximately 32.33% increase in correlation with groundwater monitoring, while maintaining a mean deviation of about 1 cm from the original GRACE data. Further analysis indicates that TWSA exhibits strong seasonal variability and marked spatial differences across climatic regions. In two representative areas, the North China Plain (NCP) and the Qinghai-Tibet Plateau (QTP), TWSA decreases at a rate of −1.368 cm/yr in the NCP, whereas the QTP shows an increasing trend of 0.408 cm/yr. These contrasting trends are driven by the combined influences of regional climate and human activities. • The Random Forest model achieved high accuracy in the downscaling task, with an R² of 0.982 and an RMSE of only 2.167 cm. • Downscaled TWSA enhances spatial gradients by 4.1% and improves correlation with groundwater observations by 32.33%. • Nationwide maps reveal climate-zone-dependent patterns in terrestrial water storage. • Regional analyses show natural-driven variability on the Tibetan Plateau and human-driven change in the North China Plain.