Optimisation and application of a high-resolution shallow landslide model at regional scale

Abstract Shallow landslides pose a risk to infrastructure and inhabited areas. Regional landslide hazard mapping and modelling are important tools for indicating potential hazard areas. We developed a methodology to assess the absolute shallow landslide occurrence probabilities at regional scale in the canton of Bern, Switzerland. Relative failure probabilities from slope stability simulations, using the probabilistic model SlideforMAP and rainfall scenarios with different return periods, are combined. These values are then normalised by the number of historical landslides observed in an area to obtain an absolute landslide occurrence probability. We parametrised the simulations using soil property classes derived from machine learning (ML)-based soil property maps, ML-based landslide thickness, forest structure data based on individual tree detection, and modelled extreme point precipitation scenarios. A global sensitivity analysis, using Sobol indices for 608 hectare plots with observed landslides distributed throughout the canton, revealed that soil cohesion, angle of internal friction, and forest effect were the most influential parameters. We then optimised the three most important soil parameters, using the Levenberg-Marquardt algorithm, to maximise the Area Under the Curve ( AUC ) for 1216 hectare plots with and without observed landslides. We tested our methodology in four catchments of about 40 $$\hbox {km}^{2}$$ in different regions of Bern, resulting in AUC values ranging from 0.59 to 0.82. The best parametrisations based on optimisation performed similarly to those based on expert input. With plausible spatial distributions and return periods, our methodology can produce information suitable for practical applications, such as hazard indication mapping.