Comparison of GIS-based gullying susceptibility mapping using bivariate and multivariate statistics: Northern Calabria, South Italy

This paper presents GIS-aided procedures for the evaluation of gullying susceptibility on a statistical basis. Field surveys and air-photo interpretation allowed us to identify gullies, geology and land use, while some predisposing factors were derived from a 7.5 m cell-size DEM. In order to investigate the role of each instability factor in controlling the spatial distribution of gullies, we implemented four models that differ in the types of method (bivariate vs. multivariate), training set sampling (polygon or point-based) and variables (continuous or categorical). The models were applied to the Turbolo river catchment, a tributary of the Crati River, Northern Calabria, Italy. The susceptibility values were classified into five categories. Model performance was evaluated using prediction rate curves and by comparing the proportion of each susceptibility class with gully distribution in the validation set. Although all models showed relatively good performances, the bivariate method overestimated the extent of the high susceptibility classes. The multivariate logistic regression model using continuous variables was found to be the best model in predicting gullying susceptibility of the study area. The area classified as stable increased with DEM cell size, meaning that a reasonably fine DEM should be used.

► Statistical models for gullying susceptibility assessment for Calabria were compared.
► The produced maps differ in method, variables and the type of training set sampling.
► The spatial mismatch between bivariate and multivariate approaches is > 40%.
► The best method for gullying susceptibility evaluation is the logistic regression.
► The accuracy of map is dependent on DEM accuracy and scale of gully areas.