Abstract
The aim of most data-driven landslide susceptibility assessments is to generate a map that depicts zones that are prone to future landsliding. The underlying algorithms are usually trained on the basis of a dichotomous response (i.e. landslide presences and absences) and a predictor set that includes derivatives of a digital terrain model (DTM). Literature indicates that numerous landslide susceptibility studies are built upon topographic variables, which were derived from the most recently available DTM. Thus, the available landslide information may mostly refer to events that were triggered before the topographic data was acquired. The resultant predictors are therefore likely to depict the archetypal morphometric signature of past landslides (i.e. the post-failure topography). However, landslide susceptible terrain, which did not yet fail, but may fail in the future may not exhibit such a distinct post-landslide topographic footprint. The delineation of landslide-prone terrain may therefore require an explicit consideration of the pre-landslide morphology (i.e. pre-failure topography).
This research aimed to explore differences among landslide susceptibility models that were based on post-failure topography and their counterparts associated with an approximated pre-failure topographic situation. The objective was to ascertain the situations in which pre-failure topography should explicitly be considered or can be ignored. In this context, a particular emphasis was set on elaborating the influence of spatial modelling resolution and landslide dimension. The research focused on a 12 km2 test site located in the Walgau region (Vorarlberg, Austria), where a comprehensive shallow landslide inventory was available.
The methodical framework consisted of pairwise confrontations of models trained on post-failure topographies with their counterparts based on approximated pre-failure situations. Mixed effects logistic regression modelling was applied separately for five raster resolutions (from 1 to 25 m) and two landslide size groups (smaller vs. larger landslides) to assess the associated effects of DTM generalization and landslide size. The results were evaluated by interpreting the discriminatory power of predictors, modelled associations, predictor importance, prediction skills and the spatial prediction pattern.
The findings showed that dissimilarities between post-failure models and their pre-failure equivalents were controlled by the selected raster resolution and the size of the underlying landslides. The highest degree of mismatch was detected among models that were based on a detailed topographic representation (i.e. high raster resolution) and landslides, which were associated with a larger morphometric footprint (i.e. larger landslides). However, these differences were observed to diminish systematically with a decreasing raster resolution.
The results indicate that a consideration of pre-failure topography is dispensable whenever the landslide data consists of smaller events whilst modelling is conducted at a comparatively coarse resolution. In contrast, landslide susceptibility modelling based on larger landslides may benefit from accounting for post-failure topography, even though obtained prediction skills might indicate a contrary tendency.