A test of transferability for landslides susceptibility models under extreme climatic events: application to the Messina 2009 disaster

A model building strategy is tested to assess the susceptibility for extreme climatic events driven shallow landslides. In fact, extreme climatic inputs such as storms typically are very local phenomena in the Mediterranean areas, so that with the exception of recently stricken areas, the landslide inventories which are required to train any stochastic model are actually unavailable. A solution is here proposed, consisting in training a susceptibility model in a source catchment, which was implemented by applying the binary logistic regression technique, and exporting its predicting function (selected predictors regressed coefficients) in a target catchment to predict its landslide distribution. To test the method, we exploit the disaster that occurred in the Messina area (southern Italy) on 1 October 2009 where, following a 250-mm/8-h storm, approximately two thousand debris flow/debris avalanches landslides in an area of 21 km2 triggered, killing 37 people and injuring more than 100, and causing 0.5 M € worth of structural damage. The debris flows and debris avalanches phenomena involved the thin weathered mantle of the Varisican low to high-grade metamorphic rocks that outcrop in the eastern slopes of the Peloritani Mounts. Two 10-km2-wide stream catchments, which are located inside the storm core area, were exploited: susceptibility models trained in the Briga catchment were tested when exported to predict the landslides distribution in the Giampilieri catchment. The prediction performance (based on goodness of fit, prediction skill, accuracy and precision assessment) of the exported model was then compared with that of a model prepared in the Giampilieri catchment exploiting its landslide inventory. The results demonstrate that the landslide scenario observed in the Giampilieri catchment can be predicted with the same high performance without knowing its landslide distribution: we obtained, in fact, a very poor decrease in predictive performance when comparing the exported model to the native random partition-based model.