Characterization of groundwater storage changes in the Amazon River Basin based on downscaling of GRACE/GRACE-FO data with machine learning models

Groundwater storage changes in the Amazon River Basin (ARB) play an important role in the hydrological behavior of the region, with significant influence on climate variability and rainforest ecosystems. The GRACE and GRACE-FO satellite missions provide gravity anomalies from which it is possible to monitor changes in terrestrial water storage, albeit at low spatial resolution. This study downscaled GRACE and GRACE-FO data from machine learning models from 1 degrees (110 km approx) to 0.25 degrees (27.5 km approx). It estimated the spatiotemporal variability of terrestrial and groundwater storage anomalies between 2002 and 2021 for the Amazon River Basin. In parallel, the Random Forest and AdaBoost algorithms were compared and analyzed. The results reflected a good fit of the models with a very low error and a slight superiority in the predictions obtained by AdaBoost. On the predictions at 0.25 degrees, spatial patterns associated with the strong influence on storage changes of some rivers and snow-capped mountains were identified, as well as an increase in the accuracy of the scaled data of the original ones. Positive long-term behavior was also obtained in terrestrial and groundwater storage of 14.26 +/- 1.18 km3/yr and + 22.24 +/- 1.18 km3/yr, respectively. Validation of the time series of groundwater anomalies to water levels in the monitoring wells obtained maximum correlation coefficients of 0.85 with confidence levels of 0.01. These results are promising for satellite information in water management, especially