Transferability of Semi-Automatic Landslide Mapping Approach Using High-Resolution DTMs: a Case Study from the Swabian Alb, Germany

Landslides are among the natural hazards that significantly impact human life and infrastructure, making accurate landslide mapping essential for hazard assessment, risk reduction, and land use planning. However, mapping landslides, particularly in vegetated areas, remains challenging, as traditional field-based and manual mapping approaches are time-consuming and require substantial expert knowledge. Semi-automatic mapping methods based on high-resolution Digital Terrain Models (DTMs) have improved landslide inventory preparation; however, their transferability to larger and diverse environmental settings remains limited and require further assessment.  Therefore, this study aims to assess the transferability of a Geographic Object-Based Image Analysis (GEOBIA) landslide mapping approach using optimal moving window sizes, and to examine whether model performance varies across specific land use classes and improves with higher-quality DTM data.

A GEOBIA-based model, originally developed for forest covered landslides in the cuesta landscape of Jena region (Zangana et al., 2025), was transferred and applied to landslides at the Swabian Alb escarpment in south-western Germany, which are located not only in forests, but also in grasslands and settlements. The study area is characterized by Jurassic limestones overlying marls and clays. It is affected mainly by rotational slides, slump-earthflows, and translational landslides, some of which show repeated reactivation. The manually mapped landslide inventory was used for result validation and accuracy assessment. DTM derivatives (from the 2003 and 2023 data) were prepared using optimal moving window sizes following Zangana et al. (2025). The semi-automatic landslide detection workflow involved multi-resolution segmentation (MRS) and support vector machine (SVM) classification, followed by expert-based refinement and accuracy assessment against the reference map. Finally, transferability was further examined through land use class-based performance analysis and by evaluating the effect of higher-quality 2023 DTM data on model results.

The results indicate that the model developed for the Jena region is transferable to the Swabian Alb. When applied to the 2003 dataset, without differentiating between land use types, the model achieved a 75% detection rate for landslide body areas. Using the 2023 dataset increased detection accuracy to 86% compared to the 2003 data. The area-based detection accuracy in this study is approximately 30% higher than reported for the Jena region by Zangana et al. (2025). When considering only forested areas—for which the model was originally developed—the true positive rate increased by about 15%, while false positives decreased by a similar margin. Although the approach effectively identifies landslides, particularly in vegetated areas, it currently performs best for cuesta-related rotational slides. Further assessment and refinement are needed to extend its applicability to other landslide types. Nevertheless, the method shows strong potential for detecting landslides with distinct geomorphological signatures in high-resolution DTMs worldwide.

Reference: Zangana, I., Bell, R., Drăguţ, L., Sîrbu, F., and Schrott, L.: Mapping forest-covered landslides using Geographic Object-Based Image Analysis ( GEOBIA ), Jena region , Germany, Nat. Hazards Earth Syst. Sci., 25, 4787–4806, https://doi.org/10.5194/nhess-25-4787-2025, 2025.