Improving the forecast performance of landslide susceptibility mapping by using ensemble gradient boosting algorithms

Landslide is the most dangerous natural hazard in mountainous regions. Disasters due to landslides annually result in human casualties, destroyed property, and monetary damages. Landslide susceptibility maps, highlighting landslide-prone areas, can provide useful spatial information for risk management and mitigation. These maps are required to be updated continuously because of the complexity of the landslide formation and movement processes. This underlines the need to develop and use cutting-edge machine learning algorithms to produce more landslide predictive maps. The study aimed to compare the predictive performance of advanced gradient boosting algorithms for modeling landslide susceptibility, including Gradient Boosting (GB), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CB), and Natural Gradient Boosting (NGBoost). Fifteen landslide influencing factors were collected and selected based on the relationship between historical landslide locations and local geo-environmental characteristics. The statistical parameters were used to compare and verify the models’ predictive performance. All proposed models have excellent forecast performances, of which the CB model has the best forecast performance (AUC = 0.921), followed by the GB model (AUC = 0.915), the LightGBM model (AUC = 0.911), the NGBoost (AUC = 0.900), and the XGBoost model (AUC = 0.897). Landslide susceptibility maps created by the CB model are recommended for the Bac Kan province in Vietnam after being validated with current landslide events recorded by the Vietnam Disasters Monitoring System. There is potential for gradient boosting models and landslide susceptibility maps to improve disaster management activities in hilly regions.