Towards a holistic assessment of landslide susceptibility models: insights from the Central Eastern Alps

Abstract

Statistical landslide susceptibility modelling is commonly used for identifying areas with an increased likelihood of landslide occurrence, given evidence of historic events and a potentially arbitrary number of explanatory features. Despite its widespread use, the actual utility and plausibility of the resulting models and maps is sometimes neglected at the expense of model performance. Here we present a landslide susceptibility map for the northern part of Carinthia, Austria, using random forest models within an extensive ensemble modelling and hyperparameter tuning framework. We discuss the importance and effects of the most relevant features retained after feature selection through a geomorphic lens. These results form the basis on a discussion of integrating considerations of geomorphic plausibility, model interpretability and reproducibility next to quantitative model performance metrics for assessing model utility. Including these aspects enhances the applicability of the results for decision-making in landslide risk management, thereby also increasing their reliability under scientific scrutiny.