Reconstructing GRACE Terrestrial Water Storage Anomalies With Machine Learning for Drought Monitoring in Southwest China

In recent years, the southwestern region of China has experienced frequent drought disasters, causing severe impacts on the local economy and environment. The GRACE (Gravity Recovery and Climate Experiment)/GRACE-FO(Follow-on) gravity satellites can invert Terrestrial Water Storage Anomalies (TWSA), providing a new technological approach for drought monitoring. However, the data gap between the GRACE and GRACE-FO has affected the completeness and continuity of TWSA data, posing challenges for long-term drought monitoring and analysis. This study employs four machine learning models: Extreme Learning Machine (ELM), Nonlinear Autoregressive with External Input (NARX), Long Short-Term Memory (LSTM), and Extreme Gradient Boosting (XGBoost), combined with the Empirical Orthogonal Function (EOF) method. The key innovation of this study lies in integrating EOF decomposition with machine learning models to enhance the accuracy and reliability of TWSA gap-filling and drought monitoring under limited monthly-scale data conditions. The approach effectively bridges the data gap between GRACE and GRACE-FO and compares the results with reanalyzed/simulated TWSA and recently generated TWSA prediction products. The results indicate that all machine learning models can effectively reconstruct the interannual variations of TWSA, with the LSTM model performing the best. Compared to the TWSA prediction products provided by recent studies, the LSTM model offers more accurate TWSA predictions. Additionally, this study constructs the GRACE WSDI (Water Storage Deficit Index) based on the reconstructed TWSA data, and by comparing it with the Standardized Precipitation Evapotranspiration Index (SPEI), finds a correlation coefficient of 0.80 at a 6-month timescale, indicating that GRACE WSDI can effectively reflect the long-term cumulative effects of drought. The reconstruction of GRACE WSDI successfully detected the extreme drought event in the southwestern region of China in 2022, further confirming the effectiveness of GRACE WSDI in identifying and assessing drought severity.