A Comparative Study of Landslide Susceptibility Evaluation Based on Three Different Machine Learning Algorithms

Accurate landslide susceptibility evaluation results are the key to landslide disaster prevention and control in mountainous areas, which can effectively avoid the risk caused by potential landslides. To obtain an accurate and reliable reference for landslide prevention, this paper selects nine evaluation factors, including elevation, stratigraphic lithology, average annual rainfall et al, and constructs a landslide susceptibility evaluation index system in the study area through multiple covariance analysis, taking Mangcheng, Yunnan Province as the research object. Subsequently, three typical machine learning models based on support vector machine (SVM), BP neural network and random forest (RF) were used for landslide susceptibility evaluation. Finally, the accuracy of the model evaluation results was compared and validated by using accuracy (ACC), area under the ROC curve (AUC), landslide ratio (S_ei) and field fieldwork. The results showed that the RF model had the highest S_eV values of 0.867, 0.94 and 9.21 for ACC, AUC, and very high susceptibility areas, respectively; the BP neural network model had the second highest values of 0.829, 0.90 and 9.14; the SVM had the lowest values of 0.794, 0.88 and 6.85; and the RF model results were more consistent with the field study. The results of experiments show that compared with the other two algorithms, the RF algorithm has higher accuracy and reliability in the Mangshi region and is more suitable for landslide susceptibility modeling in the region, and the evaluation results obtained by using the model can provide a theoretical basis and scientific reference for landslide control in the Mangshi region.