SWOT data assimilation in 2D hydrodnamic model for flood studies

To understand the dynamics of a flood, radar satellite imagery is a valuable tool because it can provide data during the event, even with cloud cover. The SWOT mission has the dual advantage of providing surface elevation readings and mapping flooded areas. While the revisit interval can be a drawback, coupling with a hydrodynamic model allows for real-time interpolation of water levels and velocities across the entire study area. SWOT data integration is achieved through data assimilation methods, which also improve hydrodynamic models, thus being useful for early warning and forecasting future events.

The work presented here demonstrates how to incorporate both water  surface elevations and flood front positions from SWOT into a data assimilation framework. Two types of methods are used to assimilate water masks. The first method has already been implemented in other studies; it is an ensemble Kalman filter where the domain is divided into homogeneous zones. Within each zone, the water level is modified at each analysis step to minimize the discrepancy between simulated and measured flooded area. The other method is an ensemble transform Kalman filter where the Chan Vese metric is used to minimize the distance between measured and simulated flood front positions. The novelty of this study therefore stems from the comparison of these two methods, each of which offers advantages for assimilating water masks from SWOT remote sensing data. Furthermore, it is necessary to verify that the two methods are also compatible for assimilating other data, whether in-situ water level measurements or direct water elevations in the river from SWOT products.

To carry out this study, a hydrodynamic model was built on an area around the Loire-Vienne confluence that experienced several flood events during 2023 and 2024. Calculations were performed using Telemac 2d software over a three-month period. Dual state-parameter data assimilation method estimates model parameters such as friction coefficients and inflow rates for each analysis cycle, as well as a uniform correction to the model state in subdomains of the flood plain. These parameters can vary depending on the method, and their interpretation provides information on the quality of the data used. Compared to previous studies, the dynamics of the floodplain can be modified by specific friction coefficients, in addition to the potential adjustment of water depth by zone (modification of the system state). These coefficients thus incorporate various phenomena responsible for flood propagation, such as hydraulic structures and land use.

 

Regardless of the assimilation method, SWOT data significantly reduces the estimation error for water depths and flooded areas with respect to SWOT water depths and flood extents. However, the Chan Vese method improves performance at the expense of computation time. When in-situ data is added, the simulation closely matches measurements in terms of water level at the observation stations. Nevertheless, thanks to the input of elevations measured by SWOT, it is possible to discuss the uncertainties of ground measurements and the best way to integrate data from different sources into the assimilation process.