Model training data: GRACE/GRACE-FO TES

    Input for machine learning model: 11 meteorological elements from the 5th Generation European Mid Range Weather Center; Two hydrological variables of ERA5; Land use and cover data; Vegetation indicators, namely LAI and solar induced fluorescence; 11 meteorological variables, etc.

    As a comparison, two of the most widely used global TWS reconstruction datasets (0.5 ° resolution) were used, namely GRACE Humphrey and Gudmundsson's REC dataset, as well as the recent GRACE like reconstructed TWS dataset.

    Reconstruction is achieved through a set of machine learning models that utilize a wide range of input drivers, including climate and hydrological variables, land use/land cover data, and vegetation indicators. The machine learning model is trained on GRACE/GRACE-FO measurement data

    The TWS estimation value reconstructed by machine learning (i.e. GTWS MLrec) is highly consistent with the GRACE/GRACE-FO measurement results, showing high correlation and low bias in the GRACE era. GTWS MLrec was also evaluated using other independent datasets, such as land ocean quality budget, large-scale water balance of 341 major river basins, and stream measurement data from 10168 stations. We found that the proposed method performs better or more reliably overall compared to the previous TWS dataset.