Spatially Distributed Identification of Debris Flow Source Areas by Credal Networks

Debris flows represent a very destructive natural hazard, affecting buildings, transport infrastructures, and, very often, causing human losses in mountain regions. That makes the identification of potential source areas of debris flows inside a watershed particularly important. In this paper we present a general identification procedure based on the credal network (that is an imprecise probabilistic graphical model generalizing Bayesian networks) originally introduced by Antonucci et al. [2004]. That model is significantly improved by a more refined description of the meteorological and hydrological processes contributing to the debris flow initiation. As a counterpart of such improvement, the model pays a slight increase in terms of computational time for identifications. That does not prevent its extensive, spatially distributed, application to whole basins, thanks to a preliminary deterministic analysis that rejects local areas where the triggering of a debris flow cannot take place. The overall procedure is tested for a debris flow prone watershed in Southern Switzerland. The model detects the areas in the basin more prone to debris flow initiation and also shows that different rainfall return periods produce different patterns of hazard in the basin. That makes it possible with this procedure to determine the return period of the critical rainfall that triggers debris flow as a result of channel-bed failure in a specific point along the drainage network.