Validation and evaluation of two multivariate statistical models for predictive shallow landslide susceptibility mapping of the Eastern Pyrenees (Spain)

This paper deals with the quality of two multivariate statistical models based on the Geographical Information System for shallow landslide susceptibility assessment in a test area at La Pobla de Lillet (Eastern Pyrenees, Spain). The quality, which was guaranteed by a rigorous methodology based on a suitable diagnosis, validation, and evaluation of the models, ensured a reliable contrast of the final susceptibility maps. This enables us to transfer the best results to the end user. Landslide susceptibility models were carried out by logistic regression and discriminant analysis of the significant conditioning factors related to the characteristics of the slope and the upslope contributing area captured from the digital elevation model and landslide distribution. The explanatory variables were tested (KS test, principal components and one-way and T-test) to select the most statistically significant ones before being introduced into the logistic and discriminant analyses. Accuracy statistics and the receiver operating characteristic curve used for diagnosis and validation showed similar prediction skills and a good fit to the data with more than 85% of unfailed cells properly classified for the two models. The evaluation of the study area and the correlation function (R ² = 0.83) between the models revealed that the discriminant model overestimated the susceptibility of the most stable zones with respect to the logistic model. Different methods of producing susceptibility maps showed marked differences in matching the models. Substantial spatial agreement (Kappa = 0.741) between binary maps produced by the standard cut-off value descended moderately (Kappa = 0.540) as a result of superimposing maps with five susceptibility levels defined by landslide percentage. Despite the fact that the two statistical models are similar in assessing susceptibility in the study area, the implications for hazard and risk management can be different because of the conservative nature of the discriminant model.