Downscaling and validating GLDAS groundwater storage anomalies by integrating precipitation for recharge and actual evapotranspiration for discharge

Abstract

Study region: The Steenkoppies Catchment is located approximately 75 km southwest from Pretoria, South Africa (RSA). Study focus: This study tested a framework for downscaling Global Land Data Assimilation System (GLDAS-2.2) groundwater storage anomaly (GWSA) estimates from 0.25° to 0.05°. This was achieved in Google Earth Engine using the Random Forest algorithm with only precipitation and actual evapotranspiration (ETa) as input variables. Additionally, the study assessed whether accounting for temporal lags could minimise residuals and enhance model performance. New hydrological insights for the region: The greater range of downscaled GWSA values indicated that the product effectively captured local recharge (precipitation) and discharge (ETa) variations while maintaining conservation of mass. Optimising the temporal correlation (r) between input variables resulted in lower residuals and fewer outliers. Groundwater level measurements and downscaled estimates for the hard rock aquifer showed larger amplitudes and seasonality and yielded the highest r (0.6) and lowest RMSE (40 mm) and MAE (31 mm). Measurements near the spring and in the karst aquifer showed less evident amplitude and seasonality. The in situ derived and downscaled GWSA comparison demonstrated the effectiveness of the product for monitoring storage declines. When applied over aquifers experiencing significant land use change or below-average precipitation, the approach could monitor groundwater storage changes, even with limited in situ observations. The adaptable code is available for application in other study areas.

Keywords

• CHIRPS precipitation
• MOD16 ETa
• Remote and satellite sensing
• Machine learning
• South Africa