Learning-based Reconstruction of GRACE Data Based on Changes in Total Water Storage and Its Accuracy Assessment

Since April 2002, the Gravity Recovery and Climate Experiment Satellite (GRACE) has provided monthly total water storage anomalies (TWSAs) on a global scale. However, these TWSAs are discontinuous because some GRACE observation data are missing. This study presents a combined machine learning-based modeling algorithm without hydrological model data. The TWSA time-series data for 11 large regions worldwide were divided into training and test sets. Autoregressive integrated moving average (ARIMA), long short-term memory (LSTM), and an ARIMA–LSTM combined model were used. The model predictions were compared with GRACE observations, and the model accuracy was evaluated using five metrics: the Nash–Sutcliffe efficiency coefficient (NSE), Pearson correlation coefficient (CC), root mean square error (RMSE), normalized RMSE (NRMSE), and mean absolute percentage error. The results show that at the basin scale, the mean CC, NSE, and NRMSE for the ARIMA–LSTM model were 0.93, 0.83, and 0.12, respectively. At the grid scale, this study compared the spatial distribution and cumulative distribution function curves of the metrics in the Amazon and Volga River basins. The ARIMA–LSTM model had mean CC and NSE values of 0.89 and 0.61 and 0.92 and 0.61 in the Amazon and Volga River basins, respectively, which are superior to those of the ARIMA model (0.86 and 0.48 and 0.88 and 0.46, respectively) and the LSTM model (0.80 and 0.41 and 0.89 and 0.31, respectively). In the ARIMA–LSTM model, the proportions of grid cells with NSE > 0.50 for the two basins were 63.3% and 80.8%, while they were 54.3% and 51.3% in the ARIMA model and 53.7% and 43.2% in the LSTM model. The ARIMA–LSTM model significantly improved the NSE values of the predictions while guaranteeing high CC values in the GRACE data reconstruction at both scales, which can aid in filling in discontinuous data in temporal gravity field models..