Authors: Yasin Wahid Rabby*, University of Tennessee, Md Belal Hossain, Department of Sociology, Oklahoma State University, Joynal Abedin, University of Dhaka, Bangladesh
Topics: Hazards, Risks, and Disasters, Geographic Information Science and Systems, Spatial Analysis & Modeling
Keywords: Landslide Susceptibility, Rangamati, K-Nearest Neighbor, Random Forest, XGBoost
Session Type: Virtual Paper
Start / End Time: 4:40 PM / 5:55 PM
Room: Virtual 43
Presentation File: No File Uploaded
This study evaluates and compares three machine learning models: K-Nearest Neighbor (KNN), Random Forest, and Extreme Gradient Boosting (XGBoost) for landslide susceptibility mapping in Rangamati Sadar, Kaptai, and Kawkhali Upazilas in the Rangamati District. In this regard, 262 landslides in the study area and fifteen landslide causal factors such as slope, elevation, plan curvature, and profile curvature were used for modeling. Area Under the Success and Prediction Rate Curves (AUC) and statistical measures, including the Kappa index, were used to assess the models' performance. Moreover, Quality Sum (Qs) Index was employed to check the desirability and consistency of the models. From the three models, XGBoost has the highest AUCs for both success (95.27%) and prediction (90.63%), followed by the random forest (success rate: 89.26%; prediction rate: 84.74%) and KNN model (success rate: 85.54%; prediction rate: 81.02%). Kappa index shows both random forest and XGBoost has the same performance. The KNN model has the highest Qs index value (2.60), followed by random forest (2.17) and XGBoost (1.34). Spatial convergence analysis among the models showed that XGBoost and random forest are highly correlated (ρ=0.93), but all the models have a strong correlation (ρ>0.7) with each other. In terms of performance and accuracy, XGBoost out perm the other two models, but random forest balances performance, practical applicability, and desirability issues of landslide susceptibility map. Non-parametric Friedman's test reveals that the three models' performance is not significantly (P=0.23) different from each other.