An Efficient Deep Learning Framework for Robust Estimation of Terrestrial Water Storage Anomalies From GRACE/GRACE-FO

The Gravity Recovery and Climate Experiment (GRACE) and its successor GRACE Follow-On (GRACE-FO) provide critical insights into global Terrestrial Water Storage Anomalies (TWSA) by precisely measuring changes in inter-satellite distances. However, various uncertainty sources cause geographically correlated noise in the GRACE/GRACE-FO Level-2 data, which is manifested as north-south striping noise in TWSA grids and complicates hydrological investigations. To address these challenges, we introduce a novel deep learning framework, the Bayesian Optimization Convolutional Neural Network with Spatial Attention Mechanism (BO-CNN-SAM), which directly mitigates systematic errors and recovers TWSA from Level-2 data. The results are as follows: (1) The BO-CNN-SAM model architecture demonstrates superior predictive accuracy and robustness, with the Root Mean Squared Error (RMSE) reducing by 142%, 120%, and 93% during training, and 136%, 89%, and 74% during validation, compared to other deep learning structures such as BO-CNN, CNN-SAM, and CNN, respectively. (2) In terms of recovered TWSA, BO-CNN-SAM outperforms other CNN-based models, achieving the lowest RMSE (4.25 cm, 4.40 cm, and 5.50 cm) and the highest Nash–Sutcliffe Efficiency (NSE) coefficients of 0.78, 0.64, and 0.60 during the training, validation, and testing phase, respectively. (3) The BO-CNN-SAM framework is successfully used to estimate ice sheet mass balance in Greenland and to accurately quantify prolonged drought in the Yangtze River Basin during 2022. Looking ahead, BO-CNN-SAM has the potential to directly recover TWSA from raw GRACE/GRACE-FO Level-1 data to gridded Level-3 TWSA products. This capability opens new opportunities for producing TWSA fields with low latency and enhancing our understanding of hydrological processes.