The Performance of Bivariate Statistical Models in Landslide Susceptibility Mapping (Case Study: Cisangkuy Sub-Watershed, Bandung, Indonesia)

Abstract

Landslide occurrences are common in hilly and mountainous areas, especially in tropical countries with high rainfall and intensive weathering. Landslide susceptibility mapping (LSM) is an initial effort to mitigate landslide hazards. This research conducted a comparative study of four LSM maps, namely frequency ratio (FR), information value model (IVM), weight of evidence (WoE), and Shannon entropy (SE), for the Cisangkuy Sub-watershed, West Java. Those models determine the relationship between the landslide density and the causative factors. The model utilized 76 landslide pixels and 15 causative factors. 70% of the landslides were used as training data, and the remaining was used for validation. The 15 factors were selected from 27 causative factors. The highly correlated causative factors were removed to address multicollinearity. In addition, only causal factors related to landslide data are involved in the modelling. The receiver operating characteristics (ROC) curve and the landslide density index (LDI) method were used for model validation. All models indicate appropriate prediction rates for FR, IVM, WoE, and SE, which are 0.770, 0.790, 0.793, and 0.788, respectively. Based on the LDI analysis, the LDI values did not increase gradually from very low to very high susceptibility classes for each LSM map. However, the maps are still favorable because the classes that are most susceptible in all models have the highest LDI. The performance of the models may be influenced by the number of classes and classification methods used to categorize each continuous parameter, as well as the small quantity of landslide inventory data.