Generating Long-Term Grace-Like Total Water Storage Change Products Using Conditional Generative Adversarial Networks

Since 2002, the Gravity Recovery And Climate Experiment (GRACE) and its Follow-On (GRACE-FO, hereafter GRACE) missions have offered global observations of total water storage (TWS). However, the relatively short record of GRACE data poses a significant challenge for researchers to investigate the full range and long-term variability in TWS. In this study, we present RecGAN, a novel Conditional Generative Adversarial Network (CGAN) comprising a RecNet generator and pixel discriminator. Our approach aims to generate long-term GRACE-like TWS observations by calibrating the WaterGAP Global Hydrology Model (WGHM). The generator is trained to produce observations conforming to the distribution of GRACE data, while the discriminator is trained to assess whether each generated pixel resembles GRACE data. Our results show that RecGAN effectively enhances the consistency between GRACE observations and WGHM-derived TWS changes, achieving improved correlation coefficients, Nash-Sutcliffe Efficiency, and Normalized Root-Mean-Square Error. In addition, RecGAN is robust to different GRACE mascon data, crop sizes used during the training period, and hydrological models targeted for calibration. This study illustrates a promising application of employing CGANs to fine-tune the WGHM output to match GRACE observations. This approach enables the generation of longterm TWS change datasets, which are invaluable for evaluating long-term water storage fluctuations, allocating water resources, and forecasting future hydrological extremes.