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

Abstract

During the study, we investigate and compare spatial prediction result of landslide hazards with a relative less-used model BFT (Best-first Decision Tree) and its three integrated models RSBFT (RandomSubspace ensemble based BFTree), MBBFT (MultiBoost ensemble based BFT), BABFT (Bagging ensemble based BFT) in Meixian County, Baoji city, Shaanxi province, China. BFTree is a machine learning technique by optimizing split nodes of standard decision tree. Integrated learning is an excellent method by combining several weakly supervised models into a strong supervised model. For data preparation, 87 historical landslide events as landslide inven-tory map and 16 landslide conditioning factors as spatial database have been collected and organized in the study area. At last, the FR (frequency ratio) method was applied for the correlation analysis and CAE (correla-tion attribute evaluation) method was applied for analyzing contribution value of each factor. For the model studies, landslide susceptibility indexes would be possible to measure using BFT, BABFT, MBBFT, RSBFT models and prepared data. Then, four landslide susceptibility maps are generated. At last, randomly assigned 61 (70%) landslides locations has been used to build the landslide models. The other 26 (30%) landslide loca-tions were used to validate. The result of verification shows that three ensemble models have boosted the pre-dictive ability of the base model; MBBFT have better prediction ability than others; RSBFT model has no overfitting problems.