Application of novel ensemble models to improve landslide susceptibility mapping reliability

Most landslides in the Eastern Golestan province in Iran occur in the Doji watershed. Their number, however, lies at the lower limit for reliable statistical analyses. By selecting a statistical sample in an area with rather homogeneous conditions (thereby reducing the number of meaningful covariates), significant insights can nevertheless be obtained. We relied on an inventory of 145 landslides which discerns between types of movement and implemented six machine learning algorithms (Decorate, DE-REPTree, Random Subspace, RS-REPTree, Dagging, and DA-REPTree) to produce landslide susceptibility maps. This allowed us to evaluate the relative importance and the effect of covariates in the models and identify factors that are consistently associated with the presence of landslides. Our results demonstrate that, even for a small landslide inventory, reliable susceptibility maps can be produced for homogeneous landscapes. We discuss that our approach could be used to assess the reliability of statistical approaches at small scales, where a distinctive trigger is lacking.