Longevity prediction and influencing factor analysis of landslide dams

Abstract

The longevity prediction and analysis of landslide dams are significant for evaluating the failure risks and formulating emergency plans. However, owing to the large range of landslide dam longevity (from “minutes” to “millennia”), both the analysis of influencing factors within different longevity ranges and a reliable longevity prediction model are still challenging. Based on 1045 landslide dam cases, the statistical analysis of the available data and longevity records is conducted. The landslide dam longevity is classified into seven categories (L1: <1 h, L2: 1 h to 1 day, L3: 1 day to 1 week, L4: 1 week to 1 month, L5: 1 month to 6 months, L6: 6 months to 1 year, and L7: >1 year). The k-Nearest Neighbor algorithm is used to impute the missing data for each longevity category. Based on geometric parameters (dam height, length, width, and volume), hydrological parameters (dammed lake volume and upstream catchment area), triggers, and dam materials, a longevity category classification model and a longevity numeric regression model are developed. The former is established for estimating the longevity range of landslide dams with an accuracy of 91%, and the latter is constructed with an R2 value of 0.95 for predicting precise longevity. Furthermore, the application of longevity prediction models proposed in this paper is verified by three landslide dam cases with reliable recorded data in the recent few years, while their predictive performance is also compared with that of conventional statistical models. Based on the SHapley Additive exPlanations attribution theory, the importance of geometric/hydrological parameters, triggers, and dam materials in determining the longevity of landslide dams in each longevity category is demonstrated using the complete database. Additionally, the influence level of each input variable on the prediction results is analyzed by examining the specific cases. The results indicate that the models proposed in this paper are superior to conventional statistical models in applicability and prediction accuracy. Notably, geometric parameters serve as the most important factor in each longevity category. Generally, the hydrological parameters can be regarded as the second most important factor in terms of influencing the longevity of landslide dams, while dam material demonstrates the lowest impact. The impact of triggers is more significant for landslide dams with a longevity of < 1 day. For landslide dams with a longevity exceeding one month, the importance of hydrological parameters is more prominent. Analyzing the impact of input variables on the prediction results of the specific landslide dams can facilitate the identification of key factors and provide some preliminary guidance for emergency rescue efforts.