Improving the spatial resolution of global mass changes observed by GRACE(-FO) using deep learning — from terrestrial water to the ocean

Gravity field solutions from the Gravity Recovery and Climate Experiment (GRACE) and its follow-on (GRACE-FO) satellite mission provide an essential way to monitor mass changes in the climate system, comprising terrestrial water storage (TWS) anomalies and ocean bottom pressure (OBP) fluctuations. However, the coarse spatial resolution of the GRACE fields blurs important spatial detail, such as OBP gradients or mass fluxes in individual catchments. By contrast, classical hydrological or ocean models provide small-scale mass change information but of doubtful accuracy, especially for trends. To combine the strengths of both data sources, we develop a self-supervised data assimilation algorithm based on deep learning concepts and apply it successfully to global TWS and OBP anomalies. The specific design of the loss function allows the model to be optimised without requiring a high-resolution ground truth. Instead, the model parameters are optimised by weighing monthly GRACE(-FO) and numerical model inputs according to their respective advantages (i.e., spatial scales where their fidelity is highest). We obtain downscaled TWS and OBP anomaly maps with grid spacings of 0.5° and 0.25°, respectively, preserving reasonable large-scale agreement with monthly GRACE(-FO) fields. We validate our downscaled products on various time scales against satellite altimetry-measured water levels, tide gauge data, and in-situ bottom pressure measurements. The downscaled products facilitate analyses beyond the nominal GRACE(-FO) resolution, including closing the water balance equation in small basins or monitoring coastal ocean mass changes. Over the terrestrial water, our downscaled product provides an average improvement of 0.79 in terms of Nash–Sutcliffe efficiency (w.r.t. ERA5-Land water budget components) for the basins smaller than the effective resolution of GRACE(-FO) missions. Over the ocean, our downscaled product agrees better with coastal tide gauge measurements at more than 78% of studied stations. We anticipate our approach to be generally applicable to other GRACE(-FO) level-3 products and other gridded Earth observation data.