Author(s): Quan Liu, Brahim Boutkhamouine, Ayeley Tchangani, Helene Roux, François Peres
Linked Author(s): François Pérès
Keywords: Flash flood, flood risk, Bayesian techniques, object oriented approach, large scale models
Abstract: A flood forecasting approach based on Bayesian Networks (BN) is presented. In order to represent the behavior of an extended geographical zone like a basin for instance, two modelling levels are considered. In a first step, a Bayesian network later referred to as the elementary model is proposed to characterize the influence of the main variables controlling the physical processes (precipitations, runoff, infiltration, discharge) involved in a flood generation at local scale. The processes corresponding to forcing production (rainfall or snow), runoff production and runoff transfer are described through causal relations and the most relevant causal relations among variables are identified. In a second step, based on the observation that the use of BN techniques for risk assessment becomes increasingly complex when the size of the system increases, Object Oriented Bayesian Networks (OOBN) are introduced. Indeed when the system to be characterized exhibits a repetitive structure or a generic pattern representative of the various dimensions of a problem, OOBN appears to be a good alternative. With this in mind, individual elementary models are combined in a large network connecting a set of variables according to the catchment topology. The use of OOBN as modelling tools makes it possible not only as the representation of the spatial dimension of the flood phenomenon but enables also the introduction of dynamic variables characterizing the time evolution. These techniques are applied and give rise to a first tentative of modelling of natural disasters and in particular, flash flooding events
Year: 2017