Exploring the effect of absence selection on landslide susceptibility models: A case study in Sicily, Italy

A statistical approach was employed to model the spatial distribution of rainfall-triggered landslides in two areas in Sicily (Italy) that occurred during the winter of 2004-2005. The investigated areas are located within the Belice River basin and extend for 38.5 and 10.3 km2, respectively. A landslide inventory was established for both areas using two Google Earth images taken on October 25th 2004 and on March 18th 2005, to map slope failures activated or reactivated during this interval. Geographic Information Systems (GIS) were used to prepare 5 m grids of the dependent variables (absence/presence of landslide) and independent variables (lithology and 13 DEM-derivatives). Multivariate Adaptive Regression Splines (MARS) were applied to model landslide susceptibility whereas receiver operating characteristic (ROC) curves and the area under the ROC curve (AUC) were used to evaluate model performance. To evaluate the robustness of the whole procedure, we prepared 10 different samples of positive (landslide presence) and negative (landslide absence) cases for each area. Absences were selected through two different methods: (i) extraction from randomly distributed circles with a diameter corresponding to the mean width of the landslide source areas; and (ii) selection as randomly distributed individual grid cells. A comparison was also made between the predictive performances of models including and not including the lithology parameter.The models trained and tested on the same area demonstrated excellent to outstanding fit (AUC > 0.8). On the other hand, predictive skill decreases when measured outside the calibration area, although most of the landslides occur where susceptibility is high and the overall model performance is acceptable (AUC > 0.7). The results also showed that the accuracy of the landslide susceptibility models is higher when lithology is included in the statistical analysis. Models whose absences were selected using random circles showed a significantly better performance when learning and validation samples were extracted from the same area; whereas, conversely, no significant difference was observed when testing the models outside the training area.