Examining the role of class imbalance handling strategies in predicting earthquake-induced landslide-prone regions

This study was undertaken to propose a comprehensive prediction scheme containing the hybrid use of class imbalance handling strategies and machine learning methods to assess the earthquake-induced landslide susceptibility for the North Sikkim region. It is worth to mention that taking the class imbalance handling techniques into account is essential to mimic real-world conditions. To tackle this issue, this research for the first time focused on the comprehensive evaluation of nine scenarios comprising four oversampling, four undersampling, and a RAW data analysis techniques. The predictions were conducted with the stochastic gradient boosting (SGB) algorithm. Analysis results depicted that the SVM-SMOTE-SGB outperformed its counterparts (with an AUROC of 0.9878), followed by the models subjected to the pre-processing with BL-SMOTE (AUROC: 0.9876) and RUS (AUROC: 0.9859), respectively. Also, the major drawback of the black-box models, i.e., lack of interpretability, was overcome with a game-theoretical SHapley Additive explanation (SHAP) analysis. The SHAP application with respect to the best-performed model ensured the importance of distance to road, distance to stream, and elevation in the identification of earthquake-induced landslide prone regions.