Landslide susceptibility maps of Italy: lesson learnt from dealing with multiple landslide classes and the uneven spatial distribution of the national inventory

Landslide susceptibility corresponds to the probability of landslide occurrence across a given geographic space. This probability is usually estimated by using a binary classifier which is informed of landslide presence/absence data and associated landscape characteristics. Here, we consider the Italian national landslide inventory to repare slope-unit based landslide susceptibility maps. These maps are prepared for the eight types of mass movements existing in the inventory, (Complex, Deep Seated Gravitational Slope Deformation, Diffused Fall, Fall, Rapid Flow, Shallow, Slow Flow, Translational) and build one susceptibility map for each type. The analysis – carried out by using a Bayeian version of a Generalized Additive Model with a multiple intercept for each Italian region – revealed that the inventory may have been compiled with different levels of detail. This would be consistent with the datases being assembled from twenty sub– inventories, each prepared by different administrations of the Italian regions. As a result, this spatial inhonomegenity may lead to a biased national–scale susceptibility maps. On the basis of these considerations, we further analyzed the national database to confirm or reject the varying quality hypothesis suggested by the multiple intercepts results. For each landslide type, we then tried to build unbiased susceptibility models by removing regions with a poor landslide inventory from the calibration stage, and used them only as a prediction target of a simulation routine. We analyzed the resulting eight maps finding out a congruent dominant pattern in the Alpine and Apennine sectors. The whole procedure is implemented in R–INLA. This allowed to examine fixed(linear) and random (nonlinear) effects from an interpretative standpoint and produced a full prediction equipped with an estimated uncertainty. We propose this overall modeling pipeline for any landslide datasets where a significant mapping bias may influence the susceptibility pattern over space.