A novel generative adversarial network and downscaling scheme for GRACE/GRACE-FO products: Exemplified by the Yangtze and Nile River Basins.

The coarse spatial resolution of about 300 km in Total Water Storage Anomalies (TWSA) data from the Gravity Recovery And Climate Experiment (GRACE) and its follow-on (GRACE-FO, hereafter GRACE) missions presents significant challenges for local water resource management. Previous approaches to addressing this issue through statistical downscaling have been limited by the reliance on the scale-invariance assumption, residual correction, hydrological models, and a lack of consideration for spatial correlations among the TWSA grids. This study introduces the DownGAN generative adversarial network, which downscales GRACE TWSA to 25 km, as exemplified in the Yangtze River Basin (YRB) and the Nile River Basin (NRB). Additionally, we propose a novel downscaling scheme to address the above limitations. DownGAN receives static and dynamic variables as inputs while considering their potential time-delay effects. The downscaled TWSA is validated using a synthetic example, in-situ runoff, groundwater levels, and two hydrological models. The potential benefits of the downscaled TWSA in closing the water balance budget and monitoring hydrological droughts in the YRB and NRB are explored. The synthetic example indicates that DownGAN trained using the proposed downscaling scheme can downscale the YRB and NRB's TWSA from 1° to 0.5° and 0.25°, respectively. DownGAN outperforms RecNet, a fully convolutional neural network, producing continuous, consistent, and realistic downscaled TWSA. The downscaled TWSA exhibits high correlations with the runoff and groundwater levels in the YRB and NRB, respectively. In addition, DownGAN demonstrates better performance in closing the water balance budget and monitoring drought events in both the YRB and NRB than HR GRACE mascon products, as evidenced by its higher correlations with the total water storage changes derived from the water balance equation and two drought indices, respectively. DownGAN is adaptable to other downscaling tasks and regions, offering a flexible downscaling factor, minimal assumptions, cost-effectiveness, and realistic predictions.