Abstract: Landslide damming and subsequent flood breaches pose significant threats to the human lives and property of the downstream communities, and quantitative analysis of dam stability is essential for downstream risk assessment and disaster prevention. However, incomplete, imbalanced landslide dam databases severely impact the development of data-driven models. This study develops a stability assessment model for landslide dams that integrates five machine learning algorithms－Support Vector Machine (SVM), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and K Nearest Neighbor (KNN)－and accounts for geomorphological parameters, hydrological conditions, dam materials, and climate factors. We identify optimal methods for imputation of various factors, then apply three sampling techniques (oversampling, undersampling, and hybrid sampling) to balance the data and mitigate model bias while examining their combined effects on model performance. Five-fold cross-validation results indicate the oversampling combined with LightGBM achieves an average accuracy of 0.84, a bias of 0.24, and a comprehensive evaluation index (CEI) of 1.32, outperforming other algorithms. Validation in typical landslide dam cases shows that our new model outperforms previous models, offering a novel approach to emergency response planning.

Key words: landslide dam, stability, database imputation, model evaluation, sampling method