Machine learning downscaling of GRACE satellite data for local-scale water storage assessment in South-East Queensland, Australia

Monitoring terrestrial water storage (TWS) at local scales remains challenging due to the coarse spatial resolution of Gravity Recovery and Climate Experiment (GRACE) data. Although machine learning (ML)-based downscaling techniques offer a pathway to improve resolution, their reliability in capturing localised hydrological variability requires further evaluation. The main aim of this study is to evaluate five ML models - Support Vector Machine (SVM), Random Forest (RF), Gaussian Process Regression (GPR), Generalised Additive Model (GAM), and Artificial Neural Network (ANN) - to downscale GRACE-derived TWS anomalies from 1.0°×1.0° to 0.05°×0.05° across South-East Queensland, Australia, for the 2002–2022 period. High-resolution predictors of precipitation, evapotranspiration, and runoff from the Australian Water Outlook platform were used as inputs into the ML models. The downscaled outputs were validated using in-situ observations from 43 precipitation gauges, 98 groundwater monitoring wells, and 42 surface water level stations distributed across the study region. Among all the models, GPR consistently achieved the best performance (MAE = 1.4 mm, RMSE = 2.5 mm), followed in performance by RF (MAE = 4.6 mm, RMSE = 7.2 mm), GAM (MAE = 5.1 mm, RMSE = 8.1 mm), ANN, and SVM, respectively. Correlation coefficients between in situ groundwater levels (GWL) and downscaled outputs ranged from 0.77 to 0.79, while for surface water levels (SWL) they varied from 0.47 to 0.48, with RF, GPR and GAM providing the best fit for GWL estimates. This study provides an opportunity to enhance the use of GRACE in regional hydrological assessments where dense monitoring networks are absent or limited. • Evaluated five machine learning models to downscale GRACE TWS at catchment scale. • GPR emerges as the strongest single model across evaluation targets. • Random Forest and GAM show high accuracy and stable groundwater representation. • Downscaled TWS preserves GRACE mass signals with minimal error and high fidelity. • Enhanced TWS supports drought detection and water-resource planning in SEQ.