Hazard assessment of slow slope movements

In recent years, landslide risk assessment has gained significant and ever increasing importance. In fact, soil and rock movements are natural threats that represent the major risk for both the population and infrastructure, particularly due to the anthropic influence on the continuous modifications of the territory. This is a typical scenery of the North Apennines region, in Italy. This area is in fact characterized by a high frequency of landslide events that often cause economic losses associated to human activities. From a geological point of view the North Apennines can be represented in a schematic way as a chain of stratums developed as a result of the collision of two Continental blockades. The formations show the following sequence, from bottom to top: sandstones-marls succession (Tuscan-Umbrian Domain), clays-marly clays (Subligurian and Ligurian Domains) and sedimentary material (clays-sandy clays, Epiligurian Domain). Most landslides occurring in this area consist of shallow movements, which involve fine, essentially clay material and the common movement is a translational or a roto-translational sliding. According to the Varnes classification, they can be identified as extremely slow or very slow movements, with velocities typically of few centimetres per year. The main triggering factor is hydrologic, since movements are usually strictly connected to ground water level fluctuations. The availability of a well established and highly reliable monitoring database of a few landslides located in the area – composed of inclinometer and piezometer records – has enabled the investigation of a new approach to predict soil movements. This paper discusses the case of a extensively monitored landslide. A well-defined dynamic-viscous model capable of returning a displacement prediction from a groundwater level input was considered. The deterministic solution of the inverse problem was performed by segmenting the historic data in start-end motions, allowing for the generation of empirical probability density functions of model initial condition parameters. By sampling these empirical functions using Monte-Carlo simulations the remaining model parameters were retrieved by the nonlinear least squares. In this way, all parameters were represented using a probability density function. Once the deterministic solution to the inverse problem is completed, it follows to solve the probabilistic inverse problem by the Bayesian approach. At this stage, both the prior and likelihood have been obtained, which permits the use of Markov-Chain Monte Carlo methods to sample the posterior, given in the form of probability density function for each model parameter conditioned on site specific data, including their corresponding correlation structure. Such approach represents a more rational tool for future risk management, as it enables to take into account the available information in a more effective way and to quantify the uncertainty related to the predictions.