Landslide susceptibility modeling with frequency ratio, logistic regression, artificial neural network and the combination method

Several landslide models have been proposed to produce landslide susceptibility map, but no particular model has been considered as an optimal one in global scale due to the dominated varieties of landslide are different from regions. In this study, we try to integrate the advantage of each model for a better approach in landslide susceptibility mapping. Three commonly used models, namely frequency ratio (FR), logistic regression (LR) and artificial neural network (ANN), are examined to generate a combined susceptibility map in Thu Lum basin located in the mountainous area of Lai Chau Province, Viet Nam. For training and testing the models, landslide samples were selected from a landslide inventory map which was prepared by applying the change detection method based on Normalized Difference Vegetation Index (NDVI) images derived from Sentinel-2. Landslide susceptibility maps were constructed with 13 environmental factors. The performance of proposed model was assessed by using area under the receiver operation characteristic curve (AUC) and kappa coefficient (Kappa). The combined model outperforms the best results (AUC=0.953 and Kappa=0.79) when compared to the single models (AUC: 0.944, 0.929 and 0.91, and Kappa: 0.73, 0.72 and 0.65 for ANN, LR and FR, respectively) equipped with the high potential in mapping landslide susceptibility.