Bayesian lifetime analysis for landslide dams

Landslide dams are a common hazard which threaten downstream human settlement or infrastructure, as their collapse may result in a flash flood. The danger is compounded by the amount of water accumulated; therefore, estimation of the time to failure becomes crucial for assessing engineering risk mitigation procedures. Dam dimension indices, descriptive multivariate analysis and logistic regression have been used to produce a static image of the dam conditions, estimating the probability of failure, but they provide no information of when a failure might occur. We propose a Bayesian model to predict the time to failure of landslide dams, based on imputing missing dam and reservoir measurements via an analysis of their covariate structure. The resulting data is then used in a Bayesian survival model which links the (censored) failure time to dam and reservoir variables. A case study on heterogeneous Italian events is presented, on which our model is tested. Results show that it is possible to produce a probabilistic model of time to failure. The length and height of dams, and the catchment area behind them, are identified as the most important covariates controlling the time to failure. The addition of area-based intercepts confirms the robustness of the methodology. Examples of potential results (forecasting) and possible generalizations of the model are presented.