Abstract: In view of the problem that the selection of characteristic factors in landslide susceptibility evaluation is highly subjective and the selection of landslide negative samples is highly random, resulting in low prediction accuracy, this paper proposes a landslide susceptibility evaluation method that uses multi-method integration to select characteristic factors and combines the information volume method to optimize the extraction of negative samples. Taking Bazhong city, Sichuan province as an example, the results of feature selection by five methods, namely maximum relevance minimum redundancy (mRMR), gradient boosting trees (GBT), extreme gradient boosting (XGBoost), ordinary least squares (OLS), and information gain (IG) are first normalized and accumulated to obtain a comprehensive score. Secondly, negative samples are selected by the information volume method to construct a sample data set. Then, the support vector machine (SVM) model is used to analyze landslide susceptibility, and a comparative experiment is conducted with the logistic regression (LR) model. Finally, the accuracy of the prediction results is verified from three aspects: landslide susceptibility zoning map, point density statistics, and ROC curve. The experimental results show that the multi-method integrated feature selection proposed in this paper and the application of information volume method for negative sample optimization can effectively improve the prediction accuracy of the model, and the susceptibility evaluation results are more accurate and reliable.

Key words: landslide disaster, feature selection, negative sample optimization, machine learning