Optimal landslide susceptibility zonation based on multiple forecasts

Environmental and multi-temporal landslide information for an area in Umbria, Italy, was exploited to produce four single and two combined landslide susceptibility zonations. The 78.9 km2 study area was partitioned in 894 slope units, and the single susceptibility zonations were obtained through linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), logistic regression (LR), and by training a neural network (NN). The presence or absence of landslides in the slope units in the period from pre-1941 to 1996 (training set) was used as the dependent variable for the terrain classification. Next, adopting a regression approach, two “optimal” combinations of the four single zonations were prepared. The single and the combined zonations were tested against landslides in the 9-year period from 1997 to 2005 (validation set). Different metrics were used to evaluate the quality of the susceptibility zonations, including degree of model fit, uncertainty in the probability estimates, and model prediction skills. These metrics showed that the degree of model fit was not a good indicator of the model forecasting skills. Zonations obtained through classical multivariate classification techniques (LDA, QDA and LR) produced superior predictions when compared to the NN model, that over fitted the landslide information in the training set. LDA and LR produced less uncertain zonations than QDA and NN. The combined models resulted in a reduced number of errors and in less uncertain predictions; an important result that suggests that the combination of landslide susceptibility zonations can provide “optimal” susceptibility assessments.