Application of Hierachical Bayeasian Modelling to Long Term Avalanche Forecasting
Nicolas Eckert
UR ETNA, Cemagref, Saint Martin d'Heres, France

In the context of rapid mass movements, evaluating extreme events is a crucial question for hazard zoning and the design of defense structures, i.e. long term forecasting. However, the direct use of standard extreme value theory is difficult because of the dependency of traveled distances on topography. The difficulty can be overcome by combining a mechanical model for flow propagation with a stochastic model describing the variability of the different inputs/outputs. Crucial problems are then model identifiability and finding a reasonable compromise between precision of the description of the flow and computation times.

In this talk, these points are illustrated with a depth-averaged set of equation describing the propagation of snow avalanches which is used within a hierarchical Bayesian framework. First, the joint posterior distribution of model unknowns is estimated using a sequential Metropolis-Hastings algorithm. Second, the point estimates are used to predict the joint distribution of different variables of interest for hazard mapping. Recent developments are employed to compute pressure distributions taking into account the rheology of snow. Third, the optimal design of a defense structure is performed by combining a loss function and the hazard model.The different steps of the method are illustrated with a real case study from the French Alps.

Keywords: Snow avalanches; Bayeasian modelling; Return periods; Risk evaluation

Biography: Nicolas Eckert is a young French researcher interested in developing statistical methods for quantifying natural hazards in mountainous environment, with special focus on long term forecasting of snow avalanches. Most of his work has for now been done within the Hierarchical Bayesian framework, with application to coupled statistical-dynamical simulations, spatio-temporal modelling and risk evaluation.