A comparative study of regional landslide susceptibility mapping with multiple machine learning models

Abstract

The purpose of this study is to utilize three machine learning models—random forest, logistic regression and extreme gradient boosting—to assess the landslide susceptibility of Wushan County and compare the predictive performance of each algorithm. To achieve this, the database was randomly divided into training (80%) and testing (20%) datasets. We considered the impact of soil thickness, which has rarely been explored in previous research. A spatial database of 19 conditioning factors related to the occurrence of a landslide was constructed to develop the landslide susceptibility maps. Thereafter, three models were estimated and compared using metrics such as accuracy, recall, F1 score and area under the receiver operating characteristic curve (AUC). The results show that the random forest model with the testing dataset has higher accuracy (0.848), F1 score (0.740) and AUC (0.904) values. Soil thickness is found to play a significant role in the occurrence of a landslide. The quantitative analysis of landslide susceptibility maps indicates the superiority of the random forest model. Finally, the outcomes of the random forest model incorporating multicollinearity analysis and factor selection were thoroughly discussed. In conclusion, the random forest is the recommended algorithm for evaluating landslide susceptibility to aid in disaster management in Wushan County.