Landslide susceptibility modeling using bivariate statistical-based logistic regression, naieve Bayes, and alternating decision tree models

摘要

The main aim of this study is to use weights of evidence (WoE), logistic regression (LR), naieve Bayes (NB), and alternatingdecision tree (ADTree) models to draw a landslide susceptibility map in Yanchuan County, China. First, 311 landslide pointswere identified through historical data, aerial interpretation, and field investigation to generate landslide inventory maps.Second, the landslide points were randomly divided into two groups (70%/30%) for training and validation. Then, 16 landslideconditioning factors were selected, namely slope aspect, slope angle, elevation, topographic roughness index (TRI), slopelength (SL), convergence index (CI), terrain positioning index (TPI), profile curvature, plan curvature, distance to rivers,distance to roads, lithology, soil, rainfall, land use, and normalized difference vegetation index (NDVI). Variance inflationfactors (VIF), tolerance (TOL), and Pearson correlation coefficient (PCC) were used to detect potential multicollinearityproblems between these factors. The performance of the model was evaluated using receiver operating characteristic (ROC)curves and area under curve (AUC) methods. The areas under the curve obtained through WoE, LR, NB, and ADTree methodsare 0.822, 0.833, 0.821, and 0.847 for the training dataset, and 0.888, 0.897, 0.898, and 0.823 for the validation dataset,respectively. The results show that the ADTree model has an overfitting state, so LR has the best balance performance. Thisalso proves that advanced machine learning models do not necessarily perform better than traditional models. The resultsobtained will assist in the future identification of landslide areas to better manage and reduce the negative environmentalimpact of landslides.