Developing a Long Short-Term Memory (LSTM)-Based Model for Reconstructing Terrestrial Water Storage Variations from 1982 to 2016 in the Tarim River Basin, Northwest China

Estimating Terrestrial Water Storage (TWS) not only helps to provide a comprehensive insight into water resource variability and the hydrological cycle but also for better management. In current research, Gravity Recovery And Climate Experiment (GRACE) data are combined with available reconstruct longer record of Anomalies (TWSA) prior 2003 Tarim River Basin (TRB), based on Long Short-Term Memory (LSTM) model. We found that TWSA generated by LSTM using soil moisture, evapotranspiration, precipitation, temperature best matches GRACE-derived TWSA, high correlation coefficient (r) 0.922 Normalized Root Mean Square Error (NRMSE) 0.107 during period 2003–2012. These results show model is an feasible method generate TWSA. Further, reveals significant fluctuating downward trend (p < 0.001), average decline rate 0.03 mm/month 1982–2016 in TRB. Moreover, amount north TRB was less than south basin. Overall, our findings unveiled GRACE can be effectively analyze long-term large-scale basins limited data.