Integration of Satellite Data, Physically-based Model, and Deep Neural Networks for Historical Terrestrial Water Storage Reconstruction

Abstract

Terrestrial water storage (TWS) is an essential part of the global water cycle. Long-term monitoring of observed and modeled TWS is fundamental to analyze droughts, floods, and other meteorological extreme events caused by the effects of climate change on the hydrological cycle. Over the past several decades, hydrologists have been applying physically-based global hydrological model (GHM) and land surface model (LSM) to simulate TWS and the water components (e.g., groundwater storage) composing TWS. However, the reliability of these physically-based models is often affected by uncertainties in climatic forcing data, model parameters, model structure, and mechanisms for physical process representations. Launched in March 2002, the Gravity Recovery and Climate Experiment (GRACE) satellite mission exclusively applies remote sensing techniques to measure the variations in TWS on a global scale. The mission length of GRACE, however, is too short to meet the requirements for analyzing long-term TWS. Therefore, lots of effort has been devoted to the reconstruction of GRACE-like TWS data during the pre-GRACE era. Data-driven methods, such as multilinear regression and machine learning, exhibit a great potential to improve TWS assessments by integrating GRACE observations and physically-based simulations. The advances in artificial intelligence enable adaptive learning of correlations between variables in complex spatiotemporal systems. As for GRACE reconstruction, the applicability of various deep learning techniques has not been well studied previously. Thus, in this study, three deep learning-based models are developed based on the LSM-simulated TWS, to reconstruct the historical TWS in the Canadian landmass from 1979 to 2002. The performance of the models is evaluated against the GRACE-observed TWS anomalies from 2002 to 2004, and 2014 to 2016. The trained models achieve a mean correlation coefficient of 0.96, with a mean RMSE of 53 mm. The results show that the LSM-based deep learning models significantly improve the match between original LSM simulations and GRACE observations.