Measuring Landslide Susceptibility of Phuentsholling, Bhutan Using Novel Ensemble Machine Learning Methods

Abstract

In this research study, the main aim is to create a landslide susceptibility map (LSM) for Phuentsholing of southern Bhutan. For the advancement of study, various individual and novel ensemble machine learning models, namely, Additive regression (ADR), Random subspace (RSS), M5P, RSS-M5P, and ADR-M5P were used. Out of all the locations of the landslide inventory data, 70% were used for the training purpose whereas the remaining 30% were used for the validation purpose. In this study, 20 landslide conditioning factors (LCFs) were considered and multicollinearity analysis by tolerance (TOL) as well as variance inflation factor (VIF) methods were done to check whether the conditioning factors are relevant or not. Those LSMs were then validated through receiver operating characteristic curve (ROC) as well as seed cell area index (SCAI) methods. Finally, it is seen that the RSS-M5P model proved to be the most effective for the prediction of landslide susceptibility. The research study found that the methods incorporated for the area under consideration could be applied to similar areas worldwide.KeywordsRandom subspaceM5PAdditive regressionEnsemble modelsLandslide susceptibility map