Landslide susceptibility prediction method based on HSOM and IABPA-CNN in Wenchuan earthquake disaster area

Landslide susceptibility prediction assesses the likelihood of landslides occurring in specific areas, providing crucial scientific support for mitigating the threat to people's lives and property posed by landslide disasters. To address the challenges in the existing landslide susceptibility prediction methods, such as insufficient representativeness of selected nonlandslide points, limited ability to capture nonlinear relationships, and a tendency to fall into local optima, the traditional Self-Organizing Maps(SOM) is improved in this paper by using a hierarchical approach to form Hierarchical Self-Organizing Maps(HSOM), and a model integrating the Information Value(IV), Adaptive Bat Precise Algorithm(ABPA), and Convolutional Neural Network(CNN) is proposed, termed IABPA–CNN. Evaluation factors such as topography, basic geology and hydrometeorology were selected with the 2008 Wenchuan earthquake-hit disaster area as the study area. The data were preprocessed, Pearson correlation coefficient, tolerance(TOL), and variance inflation factor(VIF) were employed to assess the correlation among all the factors. Subsequently, the information value for each evaluation factor's classification was calculated, thereby establishing a high-quality sample dataset, which was input into the IABPA-CNN model and IV-CNN model(contrast model), respectively. The Receiver Operating Characteristic(ROC) curve was used to compare and analyze the accuracy and performance. The Area Under Curve(AUC) values for the two models is 0.89 and 0.85, respectively, indicating that the IABPA–CNN model has higher predictive accuracy. Compared with the IV-CNN model, the proportion of landslides predicted by the IABPA-CNN model in the high susceptibility and very high susceptibility area increased to 28.18% and 30.76%, respectively. Although the area proportions of very low susceptibility and low susceptibility area increased, the proportion of landslides quantity decreased. Furthermore, Monte Carlo method was employed to analyze the uncertainty of IABPA-CNN model, and average variance of 0.0083, which indicates that the model has high reliability in landslide susceptibility prediction. Therefore, the research results in this study provide a reliable scientific basis for the work of landslide disaster prevention and mitigation in Wenchuan earthquake disaster area.