Calibration of a model for mapping shallow-landslide susceptibility: the study area of central Calabria (South Italy).

This paper is aimed to both calibrate a shallow landslide susceptibility method and examine the relative importance of conditioning factors controlling the spatial pattern of shallow landslides. The used case study is the Catanzaro strait, located in the central Calabria (South Italy), periodically affected by shallow landslides and the used method consisted of four different steps. Step I consisted in a spatial database creation where an inventory map of about 4000 shallow landslides and several conditioning factors were collected. Step II consisted in implementing the logistic regression model where shallow landslide inventory map was randomly partitioned into two groups: calibration and validation sets, consisting of the 70% and 30% of the total shallow landslides respectively. Step III consisted in as-sessing the accuracy and robustness of the spatial prediction model and the relat-ed shallow landslide susceptibility map. The logistic regression model has run it-eratively 10 times each one, randomly portioning 70/30 the shallow landslide da-taset, obtaining 10 different shallow landslide susceptibility maps. For each map, accuracy and robustness have been evaluated by means of Kappa index, Accura-cy, receiver operating characteristic curve and related area under the ROC curve. The shallow landslide susceptibility map was realized averaging the 10 prediction maps performed by the 10 replicates. Exploiting these 10 maps, standard devia-tion of each pixel was also computed and the error map was created. The obtained results showed that the logistic regression model had high values of accuracy (> 0.83), kappa index (> 0.66), and AUC (> 0.90). The findings revealed also that the capacity of the method for mapping shallow landslide susceptibility was quite stable when the calibration and validation sets were changed through the 10 repli-cates. Step IV, using the jackknife test, tested the relative importance of each shal-low landslide conditioning factor. The effectiveness of each conditioning factor revealed that slope gradient, elevation, stream power index, lithology, and soil texture, were the most important factors; whereas, aspect, drainage density and plan curvature had a least importance.