Probabilistic Calibration of a Dynamic Model for Predicting Rainfall-Controlled Landslides

Abstract

Italy has a number of regions with mid to high vulnerable areas from a hydrogeological point of view. The causes are the result of both the fragility of territory and the anthropic influence on its continuous modifications. A quantitative landslide risk analysis is then necessary to avoid or reduce human life and property losses. In particular, the prediction of landslide occurrence should be estimated taking into account the uncertainties affecting the analysis process. In this paper, a specific type of landslide, triggered by rainfall and characterized by the viscous behavior of soil, is discussed and analyzed. The goal is to illustrate the applicability of a probabilistic approach, based on Bayesian theorem, which aims at developing an advanced analysis, and to predict slow-slope movements. The proposed methodology relies on the probabilistic calibration of a well-defined, viscoplastic-dynamic model that is able to predict the soil mass displacement evolution from groundwater level inputs and return a value of a mobilized friction angle. Making use of a well-established and highly reliable monitoring database of the Alverà landslide, the model is probabilistically calibrated by the Markov-chain Monte Carlo method. Starting from the prior and the likelihood, this numerical method allows sampling of the posterior, which represents the solution of probabilistic calibration given in the form of probability density functions for each model parameter, including the corresponding correlation structure. Furthermore, the uncertainty related to model predictions is fully described.