Landslide susceptibility assessment including a set of novel explanatory variables: soil sealing, and multi-criteria geological parameterization.

Landslide susceptibility maps (LSMs) depict the probability of occurrence of a given type of landslide in a given area, based on the spatial distribution of a set of selected predisposing factors. Therefore, the susceptibility assessment is very sensitive to the parameters chosen and the identification of new parameters to be used as input data is a promising field of research in susceptibility studies as it may contribute to enhance the results.

In this work the machine learning algorithm called Random Forest (RF) has been applied, employing, in addition to the most common predisposing factors, a set of newly proposed parameters, with the aim of verifying their applicability in the landslide susceptibility analysis. The study area, 3100 km² wide, contains the provinces of Lucca, Prato and Pistoia, in northern Tuscany (Italy).

The first innovative parameter introduced is the soil sealing map, derived from the national map updated yearly by ISPRA (Italian Institute for Environmental Protection and Research). Soil sealing represents the degree of anthropization of the soil, which can radically alter the geotechnical equilibrium or the hydrological system of hillslopes. This may be directly or indirectly linked to an increased landslides hazard.

In addition, multi-parametric geological information has been included. Usually, LSMs exploit only the lithological information provided by geological maps, neglecting potentially relevant geological information (e.g. degree of weathering or tectonic stress history). We created a set of geologically-based explanatory variables by reclassifying a high resolution geological map (where 194 lithostratigraphic units were mapped at the 1:10,000 scale) using five different approaches: lithological, genetic, paleo-environmental, structural and chronological.

The model was run twice, with and without these innovative parameters, and the two resulting LSMs were compared with three approaches: (1) the area under receiver-operator characteristic curve (AUC) highlighted that the advanced parameterization increases the effectiveness of the model; (2) the Out-of-Bag Error (OOBE). OOBE was used to assess the relative importance of each predisposing factors, and the new parameters showed high predictive power; (3) the resulting maps were compared, and the main differences could be explained by local complex geological settings, which are better accounted for using the multi-criteria geological parameterization.