Landslide susceptibility mapping on the islands of Vulcano and Lipari (Aeolian Archipelago, Italy), using a multi-classification approach on conditioning factors and a modified GIS matrix method for areas lacking in a landslide inventory

Abstract

In areas prone to landslides, the identification of potentially unstable zones has a decisive impact on the risk assessment and development of mitigation plans. Active volcanic islands are particularly prone to instability phenomena as they are always in the early stage of dynamic unrest. A historical example of slope instability is the landslide which occurred in 1988 along the northwestern flank of La Fossa Cone on the island of Vulcano (Aeolian Archipelago). Based on this past activity, a susceptibility assessment using the bivariate technique of the GIS matrix method (GMM) was carried out on the islands of Lipari and Vulcano. Nevertheless, this case is congruent with those where a part of the surface was not assigned to stable or unstable areas, since a comprehensive inventory was only available for the island of Lipari. Some of the implemented steps of the susceptibility matrix method were modified to enable the model developed in the Lipari area to be applied to both islands. Considering the important role that the classification of conditioning factors plays in susceptibility analysis, the degree of association with landslide spatial distribution for the multiple classifications of each factor was assessed. Furthermore, an innovative clustering approach based on text and data mining techniques (self-organizing map neural network) was applied and compared with a heuristic classification of the categorical variable of lithology units. In addition to the extensive contingency analysis, up to 14 factor combinations were submitted to the GMM, validated and compared so as to select the one that best explains the susceptibility zoning. The effects of these incorporated processes in the previous phase of classification were discussed and preliminary susceptibility map was generated for both islands. After the validation of the susceptibility assessment, it is shown that the highest classes (High and Very High) matched 76.9% (relative accuracy) of the test inventory, while the lower susceptibility classes (Very Low and Low) resulted in a degree of fit of 14.39% (relative error).