Artificial Neural Networks applied to landslide susceptibility assessment

In the present application, two different ANNs, used in classification problems, were set up and applied: one belonging to the category of Multi-Layered Perceptron (MLP) and the other to the Probabilistic Neural Network (PNN) family. The hillslope factors that have been taken into account in the analysis were the following: (a) lithology, (b) slope angle, (c), profile curvature, (d) land cover and (e) upslope contributing area. These factors have been classified on nominal scales, and their intersection allowed 3342 homogeneous domains (Unique Condition Unit, UCU) to be singled out, which correspond to the terrain units utilized in this analysis. The model vector used to train the ANNs is a subset of that derived from the production of Unique Condition Units and consists of 3342 records organized in input and output variable vectors. In particular, the hillslope factors, once classified on nominal scales as binary numbers, represent the 19 input variables, while the presence/absence of a landslide in a given terrain unit is assumed to be the output variable. The comparison between the most up-to-date landslide inventory of the Riomaggiore catchment and the hazardous areas, as predicted by the ANNs, showed satisfactory results (with a slight preference for the MLP). For this reason, this is an encouraging preliminary approach towards a systematic introduction of ANN-based statistical methods in landslide hazard assessment and mapping.