GRACE and GRACE-FO observations provide essential information on terrestrial water storage anomalies, but their coarse spatial resolution limits their direct use for subbasin-scale hydrological analysis. Machine-learning reconstruction offers a practical way to express the GRACE-derived storage signal in a spatial framework compatible with higher-resolution hydroclimatic datasets. However, reconstruction performance may depend strongly on the choice of predictor dataset. This study compares two GRACE reconstruction configurations in the Amazon using an identical XGBoost model structure but different hydroclimatic input sources.
The target variable is CSR GRACE Mascon liquid water equivalent thickness. Two predictor configurations are evaluated. The first combines CHIRPS precipitation with GLEAM variables representing soil moisture, evaporation, evapotranspiration components, and evaporative stress. The second uses CHIRPS precipitation and ERA5/Copernicus hydroclimatic variables, including soil moisture and evapotranspiration. Both configurations are harmonized to a common spatial and temporal grid, trained using the same model architecture, and evaluated over the same two Amazon subbasins. To reduce autocorrelation-driven performance, selected semesters are withheld from training, and model transferability is assessed through application beyond the training subbasin.
The comparison focuses on whether differences in predictor source lead to systematic differences in reconstructed terrestrial water storage patterns. Evaluation is performed using spatial maps and summary statistics, including absolute differences, RMSE, MAE, coefficient of determination, as well as Nash–Sutcliffe efficiency, and correlation with independent global total water storage anomaly data from the GRACE-SeDA (Self-supervised Data Assimilation) dataset. Feature-importance analysis is also used to examine whether the two configurations rely on similar hydrological controls or whether model behaviour changes depending on the predictor dataset.