Investigating Limits in Exploiting Assembled Landslide Inventories for Calibrating Regional Susceptibility Models: A Test in Volcanic Areas of El Salvador

Abstract

This research is focused on the evaluation of the reliability of regional landslide susceptibility models obtained by exploiting inhomogeneous (for quality, resolution and/or triggering related type and intensity) collected inventories for calibration. At a large-scale glance, merging more inventories can result in well-performing models hiding potential strong predictive deficiencies. An example of the limits that such kinds of models can display is given by a landslide susceptibility study, which was carried out for a large sector of the coastal area of El Salvador, where an apparently well-performing regional model (AUC = 0.87) was obtained by regressing a dataset through multivariate adaptive regression splines (MARS), including five landslide inventories from volcanic areas (Ilopango and Coatepeque caldera; San Salvador, San Miguel, and San Vicente Volcanoes). A multiscale validation strategy was applied to verify its actual predictive skill on a local base, bringing to light the loss in the predictive power of the regional model, with a lowering of AUC (20% on average) and strong effects in terms of sensitivity and specificity.

Predicting Earthquake-Induced Landslides by Using a Stochastic Modelling Approach Which Combines Preparatory and Triggering Factors: A Case Study of the Coseismic Landslides Occurred on January and February 2001 in El Salvador, CA

“…In stochastic and empirical assessment of landslide susceptibility, the presence or absence of a gravitational phenomenon within a selected mapping unit is employed as dependent variable. Thus, the accuracy and degree of completeness of the inventory largely affect the predictive performance of the models [46][47][48][49][50].…”

Investigating the Effects of Cell Size in Statistical Landslide Susceptibility Modelling for Different Landslide Typologies: A Test in Central–Northern Sicily

“…Elevation (ELE) is considered a proxy of the mean annual which typically reflects the altitude [29]; steepness (SLO) is linked to the inclination of potential underlying rupture surfaces, and it controls the speed of water [29,38]; aspects, expressed here in terms of northernness (N) and easternness (E), can be used to determine the seasonal wet/dry cycles of soils [40]; topographic curvatures, expressed in terms of plan (PLC) and profile curvature (PRF), are assumed to express the direction of flow and shallow mechanical stresses [29,38,41]; the topographic wetness index (TWI) is a good proxy for estimating the potential infiltration [31] or saturated soil thickness; the stream power index (SPI) expresses the intensity of surface water erosion [29,42]; landform classification (LCL) directly differentiates the morphological setting of the study area [29]; the outcropping lithology (LITO) is assumed to be a good proxy for the physical-mechanical properties of rocks at the basin scale [29]; finally, soil use, obtained here via CORINE 2018, is directly related to the potential hydrological and surface water disturbances induced by erosion [29,42].…”

“…where y is the dependent variable (the outcome) predicted by the function f(x), α is the model intercept, and β i is the coefficient of the h i basis functions, given the N number of basis functions. The MARS method is largely employed in several fields of science (e.g., [44,45]), but in the last ten years, MARS has been successfully employed in landslide susceptibility evaluation (e.g., [29,30,36,40,42,[46][47][48][49][50]). In this research, MARS analysis was performed by exploiting the "earth" R-package [51].…”

“…In this research, MARS analysis was performed by exploiting the "earth" R-package [51]. Based on previous MARS applications [26,29,30,37,42,50], in order to avoid overfitted models and obtain results with geomorphological adequacy in terms of evaluation, we accepted the default setting of the "earth" R-package (see [51] for details), limiting the number of terms in the range of 20-200 and setting it with no interaction.…”

“…For this reason, the random extraction of the same number of negative and positive cases was carried out. To test the independence of the results (resolution, precision and robustness) from the choice of the selected cases [29,30,42], the extraction of the balanced datasets was repeated fifty times. Moreover, a balanced random partition of each dataset into calibration (75%) and validation (25%) subsets was replicated one hundred times.…”