Landslide hazard assessment based on improved Stacking model

Abstract

The early warning of landslides is crucial in mitigating the losses caused by frequent and abrupt landslide disasters along the railway. The scientific construction of an evaluation model is pivotal in conducting a comprehensive landslide hazard assessment. Using a railway section in Ya’an City as a case study, an improved Stacking model was developed to assess landslide hazard by selecting eight evaluation factors and employing support vector machines, random forests, K-neighborhood, and naive Bayesian learning. Logical regression was utilized as a meta learning tool to evaluate the model’s performance. To address the issue of a limited number of input samples for the meta learner, the proposed approach incorporates reduced dimensionality data from the original dataset as input for the meta learner. This is based on the output of the base learner, resulting in the establishment of an improved Stacking model. The ROC curve is used to verify the accuracy of the model, compare the accuracy of the Stacking model and the single model before and after the improvement, and generate the risk zoning map of the study area. The results show that the AUCs of support vector machines, random forests, and stacking models are 0.8068, 0.8203, and 0.8368 , respectively, with good performance, while the accuracy of the improved stacking model reaches 0.8806 . A reference for the prevention and management of geological catastrophes, the accuracy of the landslide hazard zoning map created using ArcGIS in the research area has reached 0.853 , which is essentially compatible with the real distribution.

Keywords

• landslide
• support vector machines
• random forest
• stacking model