GIS-based landslide susceptibility mapping using logistic regression, random forest and decision and regression tree models in Chattogram District, Bangladesh

Abstract

The frequency of landslides and related economic and environmental damage has increased in recent decades across the hilly areas of the world, no exception is Bangladesh. Considering the first step in landslide disaster management, different methods have been applied but no methods found as best one. As a result, landslide assessment using different methods in different geographical regions has significant importance. The research aims to prepare and evaluate landslide susceptibility maps (LSMs) of the Chattogram district using three machine learning algorithms of Logistic Regression (LR), Random forest (RF) and Decision and Regression Tree (DRT). Sixteen landslide conditioning factors were determined considering topographic, hydro-climatic, geologic and anthropogenic influence. The landslide inventory database (255 locations) was randomly divided into training (80 %) and testing (20 %) sets. The LSMs showed that almost 9–12 % of areas of the Chattogram district are highly susceptible to landslides. The highly susceptible zones cover the Chattogram district's hill ranges where active morphological processes (erosion and denudation) are dominant. The ROC values for training data were 0.943, 0.917 and 0.947 and testing data were 0.963, 0.934 and 0.905 for LR, RF and DRT models, respectively. The accuracy is higher than the previous research in comparison to the extent of the study area and the size of the inventory. Among the models, LR showed the highest prediction rate and DRT showed the highest success rate. According to susceptibility zones, DRT is the more realistic model followed by LR. The maps can be applied at the local scale for landslide hazard management.

Keywords

• Geographic information systems
• Logistic regression
• Random forest
• Decision and regression tree
• AUC of ROC
• Landslide susceptibility map