Skip to main content
Passa alla visualizzazione normale.

CHRISTIAN CONOSCENTI

Exploring relationships between grid cell size and accuracy for debris-flow susceptibility models: a test in the Giampilieri catchment (Sicily, Italy)

  • Authors: Cama, M.; Conoscenti, C.; Lombardo, L.; Rotigliano, E.
  • Publication year: 2016
  • Type: Articolo in rivista (Articolo in rivista)
  • OA Link: http://hdl.handle.net/10447/168555

Abstract

Debris flows are among the most hazardous phenomena in nature, requiring the preparation of suscep- tibility models in order to cope with this severe threat. The aim of this research was to verify whether a grid cell-based susceptibility model was capable of predicting the debris- flow initiation sites in the Giampilieri catchment (10 km2), which was hit by a storm on the 1st October 2009, resulting in more than one thousand landslides. This kind of event is to be considered as recurrent in the area as attested by historical data. Therefore, predictive models have been prepared by using forward stepwise binary logistic regression (BLR), a landslide inventory and a set of geo- environmental attributes as predictors. In particular, the effects produced in the quality of the predictive models by changing the grid cell size (2, 4, 16 and 32 m) have been explored in terms of predictive performance, robustness, importance and role of the selected predictors. The results generally attested for high predictive performances of the 2, 8 and 16 m model sets (AUROC [ 0.8), with the latter producing slightly better predictions and the 32 m showing the worst yet still acceptable performance and the lowest robustness. As regards the predictors, although all the 4 sets of models share a common group (topographic attri- butes, outcropping lithology and land use), the similarity resulted higher between the 8 and 16 m sets. The research demonstrates that no meaningful loss in the predictive performance arises by adopting a coarser cell size for the mapping unit. However, the largest adopted cell size resulted in marginally worse model performance, with AUROC slightly below 0.8 and error rates above 0.3.