Abstract
Landslides of the slide-type movement represent a potential threat to people and infrastructure in mountain areas all over the world. At regional scales, data-driven models are typically used to assess landslide susceptibility, i.e., to map where landslides are more or less likely to occur. Such assessments frequently serve as basic input for landslide risk assessment, applications in spatial planning, and landslide early warning. Data-driven landslide susceptibility models strongly rely on the quality of the landslide inventory data, and therefore, their explanatory power depends on various errors associated with the underlying landslide data collection process. Previous research has highlighted how ignoring systematic errors inherent in the landslide data can lead to erroneous model inference and landslide susceptibility maps with limited practical applicability. In this context, this study aims to counteract the challenge of spatial landslide-inventory biases (e.g., incomplete landslide mapping far from infrastructure) by testing several possibilities to consider information on Effectively Surveyed Areas (ESA). The concept of ESA was introduced in previous research as the areas which were explicitly surveyed during the preparation of the landslide inventory, and hence, landslides occurring outside these areas are not usually reflected in the inventory data. Consequently, accounting for ESA may lead to landslide susceptibility maps more suitable for practical applications.
In this contribution, we carried out a comparative analysis of three different landslide susceptibility models that focus on different strategies to handle information on ESA. The analyses focused on the Italian province of South Tyrol (7,400 km²) and built upon landslide data from 2000 to 2020. This data relates to damage-causing and infrastructure-threatening landslide events and therefore ignores slope instability far from the elements at risk. The first tested strategy focused on accounting for the ESA during model fitting while averaging out its effect for the prediction into space. The second strategy builds upon the first strategy but additionally excludes easy-to-classify “trivial” terrain (e.g., alluvial plains, rock faces, water bodies). The third strategy does not account for the ESA during model fitting but considers its information during the preceding sampling process (i.e., the sampling of stable and unstable mapping units is based on an ESA mask).
The workflow comprised the delineation of slope units, the derivation of the ESA, and the spatial aggregation of the predisposing factors (e.g., lithology, land cover, topographic indices). The subsequent exploratory data analysis was conducted to explore the association of each factor with landslide occurrence data. Generalized Additive Models were implemented to derive linear or non-linear relationships between shallow landslide occurrence and the predisposing factors and to assess relative variable importance. The results were validated through random cross-validation, spatial cross-validation, and geomorphic plausibility checks.