Model for downscaling and validating GLDAS groundwater storage anomalies

This dataset underpins a framework for statistically downscaling GLDAS‑2.2 groundwater storage anomalies (GWSA) from a coarse 0.25° spatial resolution to a finer 0.05° resolution using a Random Forest machine‑learning algorithm. The approach integrates only two freely available remotely sensed inputs: CHIRPS precipitation to represent groundwater recharge processes and MODIS actual evapotranspiration (ETₐ) to represent discharge processes.

The study area is the Steenkoppies Catchment (quaternary catchment A21F), located approximately 75 km south‑west of Pretoria, South Africa, and includes both karst (dolomitic) and intergranular/fractured aquifers. Monthly GLDAS‑2.2 GWSA data (2003–2021) were processed in Google Earth Engine, with precipitation and ETₐ standardised to anomalies relative to a 2004–2009 baseline. Temporal lag analysis identified optimal lags of three months for precipitation and two months for ETₐ, which significantly improved model performance and reduced residuals.

Downscaled GWSA estimates were validated against in situ groundwater level measurements from the Department of Water and Sanitation monitoring network, converted to storage anomalies using aquifer‑specific specific yield values. The downscaled product preserved mass conservation, captured enhanced seasonal amplitudes, and achieved improved agreement with observations (up to r = 0.6, RMSE ≈ 40 mm). The dataset demonstrates strong potential for monitoring groundwater storage dynamics in data‑scarce regions, complementing rather than replacing in situ observations. The adaptable Google Earth Engine code is openly available for application in other regions.