What is wrong with landslide susceptibility mapping: Insights from data-driven analysis and field investigations

Shallow landslides pose a significant threat to people and infrastructure in mountainous regions and can occur abruptly on steep soil slopes. To assess their hazard potential, data-driven landslide susceptibility mapping aims to predict the spatial likelihood of such events. In recent decades, machine learning approaches and high-quality spatial information have continuously improved landslide susceptibility assessment. Nevertheless, discrepancies between predicted susceptibility and observed landslide occurrence seem to remain unavoidable. In simple terms, this mismatch between predicted and observed patterns can have two causes: 1) the information on covariates controlling landslide triggering is limiting (the predicted susceptibility is ‘wrong’) or 2) the observed time scale is too short to capture the failure of more areas with similar (correctly predicted) susceptibilities. To explore these two options, we first developed a landslide susceptibility map for Switzerland based on a wide range of spatial datasets and machine-learning methods. Next, we evaluated its performance against an independent inventory which contains detailed field information of 763 landslides. Information from soil profiles collected at the head scarps of these landslides allowed us to assess the specific conditions that lead to slope instabilities which large-scale spatial models are not capable of addressing. In a third step, we performed field investigations at selected past landslide sites and compared their subsurface structure (deduced from electrical resistivity tomography) with nearby locations that had not yet failed but exhibited similar predicted susceptibility values. These measurements revealed significant differences in the subsurface. Our approach highlights the critical role of subsurface complexity in controlling hydrological flow paths that ultimately govern slope failure. In particular, variations in soil texture, soil development, soil type and soil depth strongly influence the mechanical and hydrological conditions affecting slope stability. These findings provide new insights into the limitations of large-scale susceptibility mapping and emphasize the importance of subsurface hydrology in understanding shallow landslide initiation.