DeepRec: Global Terrestrial Water Storage Reconstruction Since 1941 Using Spatiotemporal‐Aware Deep Learning Model

Abstract Terrestrial water storage (TWS) plays an important role in describing the Earth system, as water availability is decisive for ecosystems and human development. Since 2002, the Gravity Recovery and Climate Experiment (GRACE) and its Follow‐On (GRACE‐FO) mission have measured TWS anomalies with unprecedented accuracy, enabling a leap in hydrological research. However, the use of the GRACE/‐FO data in climate research is restricted by the lack of measurements prior to 2002 and the 1‐yr gap between the missions. Here we present DeepRec, a deep learning approach for reconstructing GRACE‐like monthly TWS anomalies starting from 1941, covering the global land area except Greenland and Antarctica. DeepRec uses climate reanalysis variables and land use data sets as inputs to capture both natural and anthropogenic variations. Its deep learning architecture combines convolutional layers with a long short‐term memory layer to consider the spatiotemporal variations of the inputs. DeepRec quantifies the aleatoric (data) and epistemic (model) uncertainty in the TWSA estimates through a deep ensemble approach. The reconstruction achieved an area‐weighted mean basin‐scale root mean squared error (RMSE) of 17 mm against GRACE/‐FO (2002–2023) and showed improved accuracy compared to previous reconstructions when evaluated against solutions from satellite laser ranging and DORIS (weighted mean basin‐scale correlation of 0.68 for 1995–2001). Evaluations against the ERA5 water balance showed low and consistent closure errors across time periods, with weighted mean basin‐scale RMSEs of 12 mm for both 1980–2001 and 2002–2019. DeepRec achieved the lowest sea level budget closure error (RMSE of 7 mm) among all evaluated reconstructions for 1984–2001, outperforming others by 3–7 mm.