|Paper title||Assimilation of SAR-Derived Flood Extent Maps for Improving Fluvial Flood Forecast|
|Form of presentation||Poster|
As the severity and occurence of flood events tend to intensify worldwide with climate change, the need for high fidelity flood forecasting capability increases. However, this capability remains limited due to a large number of uncertainties in models and observed data. In this regard, the Flood Detection, Alert and rapid Mapping (FloodDAM) project, funded by Space Climate Observatory (SCO) initiatives, was set out to develop pre-operational tools dedicated to enabling quick responses in selected flood-prone areas, as well as improving the resolution, reactivity and predictive capability of existing decision support systems.
Hydraulic numerical models are used in hindcast mode to improve knowledge on flood dynamics, assess flood-related damage and design flood protection infrastructures. They are also used in forecast mode by civil security agencies in charge of decision support systems, for flood monitoring, alert, and management. These numerical models are developed to simulate and predict water surface elevation (WSE) and velocity with lead times ranging from a couple of hours to several days. For instance, Telemac2D (www.opentelemac.org) solves the Shallow Water Equations with an explicit first-order time integration scheme, a finite element scheme and an iterative conjugate gradient method. However, such models remain imperfect because of the uncertainties in their inputs that translate into uncertainties in the model outputs. These uncertainties are related, for instance, to the simplified equations, the numerical solver, the forcing and boundary conditions or to the model parameters resulting from batch calibration, such as friction and boundary conditions.
Data Assimilation (DA) allows to reduce these uncertainties by sequentially combining the numerical model outputs with observations, as they become available, and taking into account their respective uncertainties. These techniques are widely used in geosciences and have proven to be effective in river hydrodynamics and flood forecasting. The Ensemble Kalman Filter (EnKF) is implemented here to reduce uncertainties in upstream time-varying inflow discharge to the river catchment as well as in spatially distributed friction coefficients, with the assimilation of in-situ WSE data at observing stations. The optimality of the EnKF depends on the ensemble size over which covariances are stochastically estimated and on the observing network, especially in terms of its spatial and temporal density. The use of remote-sensing (RS) data allows to overcome the limits due to the lack and decline of in-situ river gauge stations, especially in flood plains. In recent years, Synthetic Aperture Radar (SAR) data has been widely used for operational flood management due to the ability to map flood extents over large areas in near real time, and its all-weather day-and-night image acquisition capabilities. Water bodies and flooded areas typically exhibit low backscatter intensity on SAR images, since most of the radar pulses are, upon arrival at the water surfaces, specularly reflected away. Therefore, these areas can be detected relatively straightforward from SAR images, with exceptions in built-up environments and vegetated areas. In the present work, RS-derived flood extents are obtained by a Random Forest (RF) algorithm applied on Sentinel-1 images. The RF was trained on a database that gathers manually delineated flood maps from Copernicus Emergency Management Service Rapid mapping products from past flood events. It also takes into account the MERIT DEM to improve the flood detection precision and recall.
This work highlights the merits of assimilating RS-derived flood extents along with in-situ data that are usually confined in the river bed, in order to improve the representation of the flood plain dynamics. Here, the RS-derived flood extents are post-processed to express the number of wet and dry pixels over selected regions-of-interest (ROI) in the floodplain. These pixel-count observations are assimilated along with in-situ WSE observations to account for errors in friction and upstream forcing. They provide spatially distributed information of the river and flood plain but with a limited temporal resolution that depends on the satellite overpass times; for instance Sentinel-1 has a revisit frequency of several days (maximum six days) while in-situ observations are available every 5 to 15 minutes for observing stations in the VigiCrue network (https://www.vigicrues.gouv.fr/).
The study area is the Garonne Marmandaise catchment (south-west of France) which extends over a 50-km reach of the river Garonne, between Tonneins and La Réole. The control vector for the EnKF-DA is composed of seven friction coefficient values (six on the main channel and one for the floodplain) and three corrective parameters to the inflow discharge. Results are shown for a flood event that occurred in January-February 2021, with a forecast lead time up to +24 hours. It was shown that the assimilation of both RS and in-situ data outperforms the assimilation of in-situ data only, especially in terms of 2D dynamics in the flood plains. Quantitative performance assessments have been carried out by comparing the simulated and observed water level time-series at in-situ observing stations and by computing 2D metrics computed between the simulated flood extent maps and the SAR-derived maps (i.e. Critical Success Index and F1-score based on the expression of the contingency table). This work paves the way toward a cost-effective and reliable solution for flood forecasting and flood risk assessment over poorly gauged catchments or even ungauged catchments. Once generalized, such developments could potentially lead to hydrology-related disaster risk mitigation in other regions. Future progresses built upon this work will extend to other catchments and the assimilation of other flood observations.