Graph neural network reconstructs 80 years of global water storage data

Terrestrial water storage (TWS) — which includes snow, soil moisture, surface water, and groundwater — is a critical indicator of climate change and human water use. The GRACE and GRACE-FO satellite missions have provided the only direct global observations of TWS changes since 2002, but the record is too short for long-term climate analyses. In a new arXiv paper, researchers introduce a deep learning approach that extends GRACE-like TWS anomaly (TWSA) reconstructions back to 1940 by training a multi-variate time series graph neural network (MTGNN) on daily ERA5 meteorological forcing data.

The model adapts an architecture originally designed for urban traffic forecasting. It uses a hybrid adjacency matrix that combines geodesic proximity with lagged correlations of climatic time series, capturing both local hydrological coupling and large-scale teleconnections like El Niño. The reconstruction achieves a grid-cell Pearson correlation of 0.69, a basin-mean correlation of 0.94, and near-zero bias. It accurately reproduces the spatial fingerprints of the 2015/16 El Niño and 2020/21 La Niña events. Compared to existing reconstruction methods (GTWS-MLrec, RM-REC, GRAiCE), the graph-based model is statistically competitive at basin scale — within 0.025 correlation of the best baseline — while using roughly half to a tenth of the predictors. The full implementation is publicly available on GitHub.