Data Assimilation and Hydrological Distributed Flash Flood Modelling
Bessiere, HB1; Roux, H.2; Dartus, D.2
1IMFT (Institute of Fluid Mecanics); 2IMFT
The MARINE model (Model of Anticipation of Runoff and Inundations for Extreme events) is a flash flood forecast model developed for real time exploitation of small watersheds. Inside this physically based model, the infiltration capacity is evaluated by the Green and Ampt equation and the surface runoff calculation is divided in two parts: the land surface flow and the flow in the drainage network both based on the hypothesis of the kinematic wave. In order to better represent the heterogeneities of the rainfall as well as the various behaviours of the land surface, the model is spatially distributed, which also helps to understand the surface processes. MARINE can integrate remote sensing data with spatial resolution adapted to hydrological scales. The model requires a minimum numbers of data to be run: the Digital Elevation Model for the topography of the catchment and its location; the rainfall data which come from meteorological radar; the land cover map and the location and description of the rivers. However some parameters are not directly measurable and others need to be calibrated. Consequently, they contain various uncertainties and errors.
In order to improve their identification as well as the outlet flow, data assimilation techniques were integrated to hydrological models. In hydrology, difficulties result from the non-linearities which make the model very dependant on initial conditions. Assimilation of observations in hydrology is the process by which observations are combined together with numerical models to produce a description of the state of the catchment and a prediction of the outlet flow as accurate as possible. A variational data assimilation technique called the adjoint state method was used in the MARINE model.
The study is applied to the Gardon d’Anduze catchment located in southern France. A weighted mean-squared error similar to the Nash criterion is the cost function to be minimized. The objectives are first to improve the understanding of land surface hydrology and mechanisms of modelling process and secondly to reduce uncertainties linked to hydrological system characterisation during flash flood generation. To achieve these objectives, an estimation process of some relevant parameters is implemented. The study shows that data assimilation techniques provide interesting contributions to data fitting in hydrological models. The estimated set of parameters allows simulating a hydrograph very similar to the observations. Moreover, the evolution of the parameters during the cost function convergence reveals the importance of each parameter with respect to the modelling process and correlations between them.
Finally, the study brings interesting conclusions about the validation of the physical hypothesis on which the model is based. From an operational point of view, data assimilation should contribute to the construction of a hydro-meteorological prediction chain for the identification of an imminent flood.