Predicting storm-triggered debris flow events: Application to the 2009 Ionian Peloritan disaster (Sicily, Italy)

The main assumption on which landslide susceptibility assessment by means of stochastic modelling lies is that the past is the key to the future. As a consequence, a stochastic model able to classify past known landslide events should be able to predict a future unknown scenario as well. However, storm-triggered multiple debris flow events in the Mediterranean region could pose some limits on the operative validity of such an expectation, as they are typically resultant of a randomness in time recurrence and magnitude and a great spatial variability, even at the scale of small catchments. This is the case for the 2007 and 2009 storm events, which recently hit north-eastern Sicily with different intensities, resulting in largely different disaster scenarios. The study area is the small catchment of the Itala torrent (10 km2), which drains from the southern Peloritani Mountains eastward to the Ionian Sea, in the territory of the Messina province (Sicily, Italy). Landslides have been mapped by integrating remote and field surveys, producing two event inventories which include 73 debris flows, activated in 2007, and 616 debris flows, triggered by the 2009 storm. Logistic regression was applied in order to obtain susceptibility models which utilize a set of predictors derived from a 2 m cell digital elevation model and a 1 : 50 000 scale geologic map. The research topic was explored by performing two types of validation procedures: self-validation, based on the random partition of each event inventory, and chrono-validation, based on the time partition of the landslide inventory. It was therefore possible to analyse and compare the performances both of the 2007 calibrated model in predicting the 2009 debris flows (forward chrono-validation), and vice versa of the 2009 calibrated model in predicting the 2007 debris flows (backward chrono-validation). Both of the two predictions resulted in largely acceptable performances in terms of fitting, skill and reliability. However, a loss of performance and differences in the selected predictors arose between the self-validated and the chrono-validated models. These are interpreted as effects of the non-linearity in the domain of the trigger intensity of the relationships between predictors and slope response, as well as in terms of the different spatial paths of the two triggering storms at the catchment scale.