Stability evaluation of landslide dams based on machine learning

A landslide dam is a natural hazard formed by the blockage of river channels due to landslides, mudslides, debris flows, or other slope failures. Such events create temporary dams that obstruct water flow and substantially increase the risk of catastrophic downstream flooding. Landslide dams pose significant threats to human life, property, and critical infrastructure. Early prediction of their stability is crucial for effective emergency response, disaster management, and post-disaster reconstruction, as it enables timely evacuation and mitigates further environmental and economic losses. This study aims to develop a predictive model employing machine learning algorithms to accurately assess the stability of landslide dams. A total of 380 real-world cases—both domestic and international—were collected as the initial dataset. From these, 46 complete cases were selected for in-depth analysis based on data quality and completeness. Prior to model development, correlation analysis was performed to identify key variable relationships influencing dam stability. Furthermore, a logarithmic transformation was applied to the dataset to normalize the data and enhance its suitability for modeling. Correlation analysis and normal transformation (specifically, logarithmic transformation) were conducted to enhance the dataset’s suitability for modeling and to identify relationships among factors influencing dam stability. Following transformation, the Shapiro–Wilk test was employed to assess the normality of each variable. Based on p-values, the results confirmed that all transformed variables satisfied the assumptions of normality. Subsequently, five machine learning algorithms were utilized to construct stability assessment models for landslide dams: Bayesian Network, Decision Tree, Bagging, Radial Basis Function (RBF) Neural Network, and Logistic Regression. These algorithms were selected for their capacity to handle complex, nonlinear data and to produce predictive models with strong generalization performance. Models were trained using historical data and evaluated for their ability to predict landslide dam stability under various environmental, geological, and hydrological conditions. Evaluation metrics included classification accuracy on both training and testing datasets, as well as comprehensive performance indicators. Among the five models, the Bayesian Network demonstrated the highest performance, achieving an accuracy of 89.19% and the best comprehensive performance score of 1.28. These findings underscore the multifactorial nature of landslide dam stability, which is influenced by geological conditions, triggering events, material characteristics, and hydrological factors. Despite this complexity, the models—particularly the Bayesian Network—proved effective for rapid stability assessment. In contrast, several limitations were noted in other models; for example, Decision Trees may misrepresent the distribution of small datasets, leading to biased accuracy outcomes. Bagging, due to its overly complex classification structure, tends to suffer from overfitting. Although the RBF neural network is stable, it is generally unsuitable for small datasets. The logistic regression model, which retains only three coefficients in its output expression, excludes numerous relevant parameters, rendering it inadequate for assessing the stability of landslide dams. In contrast, the Bayesian Network model demonstrates clear advantages in managing uncertainty and noisy data. Its inherent robustness minimizes the risk of overfitting and enables the construction of stable models even when data is limited—an essential feature for early-warning systems and disaster management. This capability facilitates informed decision-making and the implementation of precautionary measures before a potential failure occurs. In conclusion, the study demonstrates the potential of machine learning models in predicting landslide dam stability. By enabling early identification of instability risks, such models can enhance disaster preparedness and response efforts, offering valuable insights for risk management and mitigation in hazard-prone regions.