Graph Neural Networks Reconstruct Global Water Storage from Satellite Data

Terrestrial water storage (TWS) is a critical indicator of the global water cycle, influenced by climate and human activity. While GRACE satellite missions provide direct observations, their record only extends back to 2002, limiting long-term climate analyses. To address this, a new deep learning approach reconstructs GRACE-like TWS anomalies (TWSA) monthly, extending the record back to 1940. This method leverages a multi-variate time series graph neural network (MTGNN) architecture, originally designed for urban sensor networks, adapted for satellite geodesy. It learns the complex relationship between daily meteorological forcing data (like precipitation and evapotranspiration from ERA5) and monthly GRACE observations. Spatial dependencies are modeled using a hybrid adjacency matrix that combines geodesic proximity with lagged climatic correlations, capturing both local hydrological processes and large-scale teleconnections. The reconstruction achieves high accuracy, with a grid-cell Pearson correlation of 0.69 and a basin-mean correlation of 0.94, accurately reproducing major climate events. Compared to established reconstruction methods, the graph-based model is statistically competitive at basin scale, often requiring significantly fewer predictors. This advancement provides a valuable, extended dataset for climate-scale analyses of water resources.