Predictive Performances of Ensemble Machine Learning Algorithms in Landslide Susceptibility Mapping Using Random Forest, Extreme Gradient Boosting (XGBoost) and Natural Gradient Boosting (NGBoost)

Abstract

Across the globe, landslides have been recognized as one of the most detrimental geological calamities, especially in hilly terrains. However, the correct determination of landslide-prone fields remained a challenging task due to the nonlinear, complex, and inconsistent nature of landslides. Therefore, it is essential to apply methods with superior capabilities in dealing with real-world problems for properly delineating potential landslide zones. Recently, ensemble learning techniques have been drawn intense interest in landslide susceptibility mapping studies due to their distinct advantages. This present work intended to propose natural gradient boosting (NGBoost), a novel member of the ensemble learning family, for modeling landslide susceptibility for Macka County of Trabzon province, Turkey. The predictive performance of NGBoost was compared to ensemble-based machine learning methods, namely random forest (RF) and XGBoost using five accuracy metrics including overall accuracy (OA), F1 score, Kappa coefficient, area under curve (AUC) value, and root-mean-square error to evaluate its competence and robustness. Besides, SHAP based on the game theory approach was implemented to interpret the influences of the predisposing factors on the produced model. Results indicated that the NGBoost method utilized for landslide susceptibility mapping problem for the first time had the greatest predictive ability (AUC = 0.898), followed by XGBoost (AUC = 0.871) and RF (AUC = 0.863), and outperformed the XGBoost and RF by approximately 6% in terms of OA. McNemar's statistical significance test results also confirmed the superiority of the proposed NGBoost method over the RF and XGBoost algorithms.