Multiscale Feature Modeling and Interpretability Analysis of the SHAP Method for Predicting the Lifespan of Landslide Dams

Abstract

Landslide dams, formed by natural disasters or human activities, pose significant challenges for lifespan prediction, which is crucial for effective water conservancy management and disaster prevention. This study proposes a hybrid CNN–Transformer model optimized using the Improved Black-Winged Kite Algorithm (IBKA) aimed at improving the accuracy of landslide dam lifespan prediction by combining local feature extraction with global dependency modeling. The model integrates CNN’s local feature extraction with Transformer’s global modeling capabilities, effectively capturing the nonlinear dynamics of key parameters affecting landslide dam lifespan. The IBKA ensures optimal parameter tuning, which enhances the model’s adaptability and generalization, especially when dealing with small-sample datasets. Experiments utilizing multi-source heterogeneous datasets compare the proposed model with traditional machine learning and deep-learning approaches, including LightGBM, MLP, SVR, CNN–Transformer, and BKA–CNN–Transformer. The results show that the IBKA–CNN–Transformer achieves R2 values of 0.99 on training data and 0.98 on testing data, surpassing the baseline methods. Moreover, SHapley Additive exPlanations analysis quantifies the influence of critical features such as dam length, reservoir capacity, and upstream catchment area on lifespan prediction, improving model interpretability. This approach not only provides scientific insights for risk assessment and decision making in landslide dam management but also demonstrates the potential of deep learning and optimization algorithms in broader geological disaster management applications.