Application of random forest and multi-linear regression methods in downscaling GRACE derived groundwater storage changes

Abstract

The advent of Gravity Recovery and Climate Experiment (GRACE) has opened the doors for remote monitoring of gravitational changes and its derivatives across the globe, but received less attention due to poor spatial and temporal representation. Statistical models of varying complexity are commonly employed to downscale the GRACE datasets for use with local to regional applications. This study presents the application of two commonly employed machine learning models, multi-linear regression (MLR) and random forest (RF), in spatially downscaling (from 1° to 0.25°) the GRACE-derived terrestrial water storage anomalies (TWSA) by establishing a correlation with various land surface and hydroclimatic variables. The downscaled TWSA was further converted into groundwater storage anomalies. Applicability of the proposed methods was tested on four contrasting hydrogeological basins of India. For each basin, the significant predictor variables were considered to establish the relations. Seasonal groundwater levels observed in 236 wells during 2006–2015 were used for method validation and accuracy assessment. We observed a close match between GRACE-derived groundwater levels and the measurements for three of the four basins (r = 0.40–0.92, Root mean square error (RMSE) = 3.6–10.5 cm). Our results indicate that the predictor variables to downscale TWSA should be considered cautiously based on the hydrogeological, topographical, and meteorological characteristics of the basin.