Assessing Debris Flow Hazard by Credal Nets

Debris flows are among the most dangerous and destructive natural hazards that affect human life, buildings, and infrastructures. Starting from the ’70s, significant scientific and engineering advances in the understanding of the processes have been achieved [4, 7]. Yet, human expertise is still fundamental for hazard identification, as many aspects of the whole process are poorly understood; and the acquisition of evidence about the areas under consideration can only be done vaguely in practice, making it difficult to apply models. This paper presents a credal network model of debris flow hazard for the Ticino canton, southern Switzerland. Credal networks [5] are impreciseprobability models based on the extension of Bayesian networks [9] to sets of probability mass functions (see Sec. 2.2). Imprecise probability [11] is a very general theory of uncertainty that measures chance and uncertainty without sharp probabilities. The model represents expert’s causal knowledge by a directed graph, connecting the triggering factors for debris flows (Sec. 3.1). Probability intervals are used to quantify uncertainty on the basis of historical data, expert knowledge, and physical theories. They are also used to carefully model the vague acquisition of evidence, which is a basis to draw credible conclusions. The model presented here aims at supporting experts in the prediction of dangerous events of debris flow. We have made preliminary experiments