Binary logistic regression versus stochastic gradient boosted decision trees in assessing landslide susceptibility for multiple-occurring landslide events: application to the 2009 storm event in Messina (Sicily, southern Italy)

Abstract This study aims to compare binary logistic regression (BLR) and stochastic gradient treeboost (SGT) methods in assessing landslide susceptibility within the Mediterranean region for multiple-occurrence regional landslide events. A test area was selected in the north-eastern sector of Sicily (southern Italy) where thousands of debris flows and debris avalanches triggered on the first October 2009 due to an extreme storm. Exploiting the same set of predictors and the 2009 event landslide archive, BLR- and SGT-based susceptibility models have been obtained for the two catchments separately, adopting a random partition (RP) technique for validation. In addition, the models trained in one catchment have been tested in predicting the landslide distribution in the second, adopting a spatial partition (SP)-based validation. The models produced high predictive performances with a general consistency between BLR and SGT in the susceptibility maps, predictor importance and role. In particular, SGT models reached a higher prediction performance with respect to BLR models for RP-modelling, while for the SP-based models, the difference in predictive skills dropped, converging to equally excellent performances. However, analysing the precision of the probability estimates, BLR produced more robust models around the mean value for each pixel, indicating possible overfitting effects, which affect decision trees to a greater extent. The assessment of the predictor roles allowed identifying the activation mechanisms which are primarily controlled by steep south-facing open slopes located near the coastal area. These slopes are characterised by low/middle altitude downhill from mountain tops, having a medium-grade metamorphic bedrock, under grassland and cultivated (terraced) uses.