Bayesian belief network modelling for landslide hazard assessment using probabilistic estimates from experts

This study applies a belief-based probabilistic approach to find the relative influence of environmental conditions and human actions on the probability of occurrence of landslides in the Darjeeling Himalayas, India. Subjective (belief-based) probabilistic methods are an approach to model complex relationships in regions with low or no data. Here, we use two subjective probabilistic methods: expert elicitation and Bayesian belief network. Expert elicitation (EE) is a technique to quantify the knowledge of experts based on their theoretical or practical experience on a topic of interest. A Bayesian belief network (BBN) takes into consideration and represents the experts’ knowledge (and other data, if available) to give a probabilistic, rational outcome under the influence of uncertainty. BBNs are represented as a graph consisting of a set of random variables (called ‘nodes’) that are interconnected via edges (called ‘arcs’). BBN follow the Bayes’ theorem.

We first use expert judgement and secondary literature to determine 20 prominent variables that influence landsliding, divided into 2 triggering variables (earthquake, rainfall), 11 geomorphic variables (e.g., soil type, flood depth, elevation), and 6 anthropogenic variables (e.g., infrastructure development, machinery vibration). We then conduct EE with ten landslide experts, to find the prior and conditional probabilities of each of these 20 landslide-related variables. Prior probability is the probability of occurrence of the variables (A) that influence or trigger landslides (P(A)) and conditional probability is the probability of occurrence of landslide(s) given P(A)). Using BBN modelling, we then provide a comparison of answers across all ten experts per variable and across all variables per expert. Finally, we examine single and multiple combinations of variables and their relative influence on landsliding in the study area. We finally list suggestions, challenges faced, and limitations on designing and carrying out belief-based probabilistic procedures.