Displacement prediction for landslide with step-like behavior based on stacking ensemble learning strategy

Abstract

Predicting landslide displacement is crucial for the prevention and mitigation of landslide disasters. This study proposes a method based on a stacking ensemble learning strategy to predict landslide displacement, incorporating distinct yet effective individual models: the Voight model, the GM(1,1) grey model, and the backpropagation neural network (BPNN). These models are respectively emblematic of empirical, statistical, and nonlinear approaches to modeling. The stacking ensemble learning method marries creep theory, grey system theory, and nonlinear theory to accurately capture the statistical trends and step-like behavior characteristic of landslide displacement. A key feature of this approach is the tailored use of non-cross-validation, partial cross-validation, and 5-fold cross-validation for the Voight, GM(1,1), and BPNN models, respectively. This ensures that the conditions for model applicability are satisfied while fully leveraging their strengths, allowing the ensemble method to enhance prediction performance. The method is demonstrated through its application to the Xintan landslide in Zigui County, Hubei, China. Comparative analysis of the Voight, GM(1,1), BPNN, and the proposed stacking ensemble learning model reveals that the ensemble model achieves superior accuracy, underscoring its effectiveness in predicting landslide displacement. This promising method can effectively capture the landslide evolution process and be promoted to predict displacement in other landslide scenarios.