Dynamic integration of GRACE-based water storage into global hydrological models using machine learning

Abstract

Groundwater storage (GWS) is a critical component of the hydrologic cycle. Accurate representation of GWS dynamics is crucial for water availability estimates in global hydrological models (GHMs). This study explores the enhancement of grid-level water availability estimates by dynamically coupling machine learning (ML) models with GHMs. By integrating daily GRACE-based total water storage (TWS) data into the CWatM GHM using long short-term memory models, this study looks into updating daily GWS in a GHM, providing an observation-based estimate of an often-underrepresented process in hydrologic modeling. The results show an average 16% improvement in daily streamflow estimates. Kling Gupta Efficiency (KGE) scores increased at 61% of USGS stations (64 stations) considered in this study. Monthly streamflow estimates showed even greater gains—10% more stations reported KGE values above 0, and 58% of stations (61 stations) showed improvements, with an average monthly KGE increase of 13%. Additionally, the ML-based Data Integration (DI) approach, although it does not account for dynamic uncertainty, provides a computationally efficient alternative, reducing processing time from 2.33 min day−1 (using Ensemble Kalman Filters) to 4 s d–1. Furthermore, dynamically coupling ML with GHM to integrate GRACE-based daily TWS data yields a GRACE-informed estimate of GWS dynamics. This integration led to a 17% shift in CWatM-derived water-stress estimates across the CONUS, reflecting sensitivity to the GRACE-based storage baseline rather than a direct correction of true water stress.