An AI-Based GRACE Terrestrial Water Storage Data Assimilation Improves Hydrological Simulation

Terrestrial water storage (TWS) is a key variable in the water cycle, and accurate estimation of TWS is crucial for understanding hydrological processes and improving hydrological prediction. In this study, we develop an AI-based data assimilation method for GRACE TWS observations, aiming to integrate the advantages of satellite observations and land surface models. The assimilation adopts the ResUnet model combined with a self-supervised learning strategy. Specifically, the ResUnet model is used to extract large-scale variation information from GRACE TWS observations and high-resolution information from the land surface model. This assimilation system is applied to the NoahMP land surface model for long-term simulation, and the performance is compared with the nudging method. Results show that the AI-based assimilation method is more conducive to depicting fine-scale hydrological processes. Quantitative evaluation indicates that the assimilation effect of the proposed method is superior to that of the nudging. In addition, validation against in-situ observations confirms the rationality and reliability of the proposed method, as it can more accurately estimate terrestrial water storage and related hydrological variables. In the future, this AI-based assimilation method can be extended to the assimilation of more hydrological variables and multi-source observations, which is expected to further improve the estimation capability of land surface hydrological variables and provide more reliable data support for water resource management.