Landslide susceptibility assessment using hybrid method of best-first decision tree and machine learning ensembles

During the study, we investigate and compare spatial prediction of landslide susceptibility using Best-first Decision Tree (BFT) and three hybrid models RSBFT (RandomSubspace ensemble with BFT), MBBFT (MultiBoost ensemble with BFT), BABFT (Bagging ensemble with BFT) in Meixian County, Shaanxi Province, China. For data preparation, 87 historical landslide events as landslide inventory map and 16 landslide conditioning factors as spatial database have been collected and organized in the study area. Firstly, the frequency ratio (FR) method was applied for the correlation analysis. Secondly, 61 (70 %) landslides locations were randomly selected to build the landslide models, and the other 26 (30 %) landslide locations were used to validate models. Finally, landslide susceptibility indexes were measured using BFT, BABFT, MBBFT, RSBFT models, and four landslide susceptibility maps were generated. The susceptibility classes of the MBBFT model with the highest AUC values (0.883, 0.729) included very low (59.54 %), low (4.58 %), moderate (8.58 %), high (4.07 %) and very high (23.22 %). The result of validation shows three ensemble models have the better predictive ability than that of the benchmark model. The study contributes to managing and mitigating landslide hazards in the study area and similar regions worldwide.