Machine learning and sensitivity analysis approach to quantify uncertainty in landslide susceptibility mapping

Mitigating the impacts of landslides requires quantifying the susceptibility of different infrastructures to this hazard through landslide susceptibility mapping. The mapping requires overlaying the spatial effects of multiple factors that contribute to the occurrence of landslide events (rain- fall, land cover, distance to roads, lithology, and slope) and this process requires assigning weights to the different factors contributing to landslides. This study introduces a new statistical approach for quantifying the weights used in landslide susceptibility mapping and their associated uncertainty. The proposed approach combines machine learning (random forest classification) with large-scale sen- sitivity analysis to derive the uncertainty ranges of weights used in landslide susceptibility mapping. The study demon- strates the approach for a case study of the Chittagong Hill Tracts and Sylhet divisions of Bangladesh to understand the implications of weight uncertainty for road susceptibility to landslides. The case study results show that distance to roads is the most influential factor to determine the like- lihood of the occurrence of landslide events, followed by the land cover type. Given weight uncertainty, the percent- age of road lengths in the study area under extremely high susceptibility to landslides ranges from around 20 to 38 per- cent. The tolerance level to weight uncertainty is a crucial determinant of investment costs and is ultimately a critical element for decision making to relevant institutions and affected stakeholders. A conservative selection of weights from within the uncertainty range (a weight combination that results in the highest susceptibility) means that the risk is minimized but with a high investment cost.