A novel landslide susceptibility optimization framework to assess landslide occurrence probability at the regional...

Landslide is a hazard that has drastic repercussions on population and the environment worldwide. Landslide susceptibility mapping (LSM) is vital for landslide disaster management and formulating mitigation strategies. In this study, with the support of geographic information system and remote sensing, a new LSM hybrid framework is developed based on random forest (RF) and cusp catastrophe model (CCM). Under the framework, 15 conditioning factors and 2082 historical landslides are selected to test and compare its performance in a landslide-prone area in Liangshan, Southwest China. The results depicted a better performance of the new LSM hybrid framework (RF-CCM) than those of RF or traditional application mode of catastrophe model (Catastrophe fuzzy membership functions, CFMFs) only. The RF-CCM achieved the highest accuracy (0.901), the narrowest confidence interval (0.895-0.907), and the smallest standard error (0.004) among all the models. Notably, RF-CCM successfully decreased the uncertainty of CFMFs in determining the relative importance of conditioning factors, overcame the dependence of the CFMFs on independence among the conditioning factors, and had a higher stability level than RF. Moreover, distance to human engineering activities and slope had the greatest impact on LSM in the modeling process. The study result can provide insights for developing reliable predictive models for other landslide-prone areas with similar geo-environmental conditions.