Research On GRACE Terrestrial Water Storage Reconstruction Based On Deep Learning

Terrestrial water storage(TWS)is an essential component of the global hydrological cycle,and grasping the dynamic changes of terrestrial water storage is crucial for regional water resources management.Since the launch of the Gravity Recovery and Climate Experiment(GRACE)satellite in 2002,the satellite has been able to observe the Earth’s time-varying gravity field signals with high precision.Therefore,the GRACE satellite provides a new monitoring method for large-scale TWS,including groundwater storage,surface water storage,soil water storage,snow water equivalent and biological water storage.Given the far-reaching impact of the GRACE satellite,NASA successfully launched and operated GRACE Follow-On(GRACE-FO)in 2018.However,the discontinuity between GRACE and GRACE-FO severely restricts the application of this data in long-term hydrological studies.In order to effectively overcome this limitation,this study proposes a hybrid algorithm(STL-LSTM)that integrates Seasonal and Trend decomposition using Loess(STL)and Long short-term memory(LSTM)neural network to reconstruct changes in terrestrial water storage beyond the operational cycle of the GRACE and GRACE-FO satellites.In addition,this paper uses ERA-Interim data as the driving data source of the World-Wide Water Resources Assessment system(W3)to obtain long-term,high-resolution hydrological model results,which are used as STL-LSTM.The scheme reconstructs the input components of the terrestrial water storage model.The main results of this research are as follows:(1)The comparison between the hydrological data derived from the W3 hydrological model and the Global Land Data Assimilation System(GLDAS)data shows that the correlation coefficient,Nash coefficient and root mean square error between the W3 simulation results and GRACE observations are 0.81,0.59 and 25.32 mm,all of which are better than the three evaluation index results between GLDAS and GRACE.Therefore,the W3 model calculated in this study has a reliable performance in the Yangtze River Basin,and using the output data of the W3 model as the input component of the reconstructed model is beneficial to improving the model accuracy.(2)The reconstruction results based on STL-LSTM and LSTM in the Yangtze River Basin show the correlation coefficient,Nash coefficient,and root mean square error of STL-LSTM and GRACE-FO observation data are 0.87,0.38,and 1.97 mm,respectively.Better than LSTM in correlation coefficient(0.85),Nash coefficient(0.27),and root mean square error(2.13 mm)with GRACE-FO observation data.The STL-LSTM method proposed in this paper can reconstruct the time series of terrestrial water storage outside the operational cycle of satellites in regional and global hydrological studies.(3)According to the reconstructed TWS,a total of 12 drought events in the Yangtze River Basin were identified,of which the most severe event occurred from August 2006 to November 2006,with a duration of 4 months,a drought peak of-1.37,and the drought category D2(severe drought).In addition,the reconstructed TWS in this paper can also monitor drought events before April 2002,such as June 1988-July 1988,duration February,drought peak-0.62,drought category D0(mild drought).The above conclusions confirm that the reconstructed TWS data in this study can be effectively used for monitoring extreme events and provide a reliable data source for long-term hydrological research.There are 27 figures,8 tables,and 139 references in this paper.