INTEGRATION OF HYDROGEOLOGICAL VARIABLES IN RANDOM FOREST MODELS FOR DOWNSCALING OF TWS ANOMALIES IN CHILE

The coarse spatial resolution of Terrestrial Water Storage (TWS) data provided by the GRACE/GRACE-FO mission limits its applicability in regional water resource management studies. This work implements a Random Forest-based statistical downscaling model to increase the spatial resolution of TWS to 0.25° in the Chilean territory. The model is interconnected with hydrogeological, geospatial and temporal variables. The results show that the inclusion of spatiotemporal predictors improves significantly the model's predictive ability (R=0.74, RMSE=1.97 cm). The validation within situ observations of 485 wells shows positive correlations (R>0.5) in 60% of cases. This approach highlights the potential of machine learning to increase the regional applicability of GRACE data.