Assimilation of future SWOT-based river elevations, surface extent observations and discharge estimations into uncertain global hydrological models

International audienceGlobal estimates of river dynamics are needed in order to manage water resources, mainly in developing countries where in-situ observation is limited. Remote sensors such as nadir altimeters can complement ground data. Current altimeters miss however a large number of continental surface water bodies. This issue will be largely resolved by the future Surface Water and Ocean Topography (SWOT) mission, thanks to its wide swath altimeter. SWOT will provide almost globally two-dimensional water elevation maps for rivers over 100 m wide and water bodies over 250 m × 250 m. During this research, we investigated the potential of SWOT to correct hydrological models on a global/continental scale, through data assimilation. For this purpose, an Observing System Simulation Experiment (OSSE), also known as “twin experiment”, has been implemented. Model forcings and parameters were perturbed to jointly achieve global hydrological models (GHMs) uncertainties, which is the expected scenario in which the SWOT community will mainly evaluate the future SWOT data. SWOT-like observations of water surface elevation (WSE), flooded water extent (FWE), and/or SWOT derived discharge (Q) were used to correct modelled Q, WSE and FWE from a large-scale hydrological and hydrodynamic model (MGB – portuguese acronym of “Modelo de Grandes Bacias”), using a Ensemble Kalman filter (EnKF). The results indicate that SWOT products could largely improve hydrological simulations on a global and continental scale. SWOT-like discharge can reduce ~40% of model errors in daily discharge. Furthermore, when anomalies of the WSE DA approach were implemented, the error reduction was even greater for all state variables compared to the absolute WSE DA, achieving average error reduction values of about ~30% compared to ~24%. Finally, the simultaneous DA of all the SWOT-like variables together reduces errors from ~14% to ~22% compared to the average of assimilating only one variable