Landslide susceptibility and associated confidence by spatial analysis: Two applications in northern Italy's mountain areas

This contribution discusses how spatial prediction models of landslide susceptibility can be applied to datasets with combination of categorical and continuous data layers maintaining the variability of values: i.e., avoiding binary transformations that impoverish their significance. The data layers represent conditional factors, such as geology, land use, distance from thrusts, internal relief, slope, attitude and permeability, which are used to associate the distribution of mapping units with that of landslides, possibly of specific dynamic types and time intervals of occurrence. The landslide trigger areas, used as training sites, maintain in the modeling their spatial extension as polygonal patches. Predictions are made by mathematical models, e.g., empirical likelihood ratio, also using several occurrence subsets to obtain blind testing for empirical cross-validation of the spatial prediction results. Prediction maps are interpreted in relation with the corresponding prediction-rate curves expressing the goodness of the prediction. Datasets from two study areas in northern Italy, come from the Apennines and from the Alps. Different analytical strategies are followed in the two study areas due to the difference in density of landslide occurrence. The prediction-rate curves obtained provide continuous class values distributions that preserve variability thus avoiding prediction class grouping before interpretation. The analysis of the prediction-rate curves, tables and histograms leads to more articulate criteria for decisions aiming at hazard mitigation or avoidance. Maps of predicted susceptibility/hazard levels are generated that consist of relative values that need careful quantitative scrutiny to be interpreted: the only meaning of such relative values is their rank.