Selecting statistical or machine learning techniques for regional landslide susceptibility modelling by evaluating spatial prediction

With so many techniques now available for landslide susceptibility modelling, it can be challenging to decide on which technique to apply. Generally speaking, the criteria for model selection should be tied closely to end users’ purpose, which could be spatial prediction, spatial analysis or both. In our research, we focus on comparing the spatial predictive abilities of landslide susceptibility models. We illustrate how spatial cross-validation, a statistical approach for assessing spatial prediction performance, can be applied with the area under the receiver operating characteristic curve (AUROC) as a prediction measure for model comparison. Several machine learning and statistical techniques are evaluated for prediction in Lower Austria: support vector machine, random forest, bundling with penalized linear discriminant analysis, logistic regression, weights of evidence, and the generalized additive model. In addition to predictive performance, the importance of predictor variables in each model was estimated using spatial cross-validation by calculating the change in AUROC performance when variables are randomly permuted. The susceptibility modelling techniques were tested in three areas of interest in Lower Austria, which have unique geologic conditions associated with landslide occurrence.