A Physics-Informed Dual-Stream Architecture for Spatiotemporal Downscaling of GRACE/GRACE-FO to Daily Resolution
TL;DR
A physics-informed Dual-Stream Deep Learning architecture that synchronously trains a monthly physics-constrained estimator by enforcing strict mass conservation via a global mass budget projection and a daily probabilistic estimator that models conditional predictive distributions using a Student-t likelihood to accommodate heavy-tailed GRACE/-FO TWSA residuals is developed.
Abstract
We introduce a daily and 25-km spatiotemporally downscaled GRACE/-FO Terrestrial Water Storage Anomaly (TWSA) dataset for the Contiguous United States (CONUS) spanning 2002–2024. To address the resolution gap between coarse monthly satellite observations and daily hydrometeorological dynamics, we developed a physics-informed Dual-Stream Deep Learning (DL) architecture. The framework synchronously trains a monthly physics-constrained estimator by enforcing strict mass conservation via a global mass budget projection and a daily probabilistic estimator that models conditional predictive distributions using a Student-t likelihood to accommodate heavy-tailed GRACE/-FO TWSA residuals. Internal validation demonstrates that the downscaled product preserves the long-term trends and spatial variance of the native GRACE/-FO observations while outperforming standard land surface models that are used as input of DL framework. Crucially, the daily temporal resolution enables the quantification of high-frequency hydrologic extremes that are typically obscured in monthly gravity fields. We extensively validated the daily TWSA dataset against a 22-year catalog of Atmospheric River (AR) events in California and 104 different Tropical Cyclone (TC) events across the eastern CONUS. The downscaled TWSA successfully resolves rapid mass loading with a near-zero phase lag relative to independent in-situ streamflow and precipitation networks. The resulting simulation yields a mass-conserving, physically interpretable daily TWSA product suitable for high-frequency hydrological and climate applications.