Abstract: Machine learning faces two difficulties in the evaluation of landslide susceptibility. One is the objective quantification of evaluation index, and the other is the selection of training sample-0.5pts. For that reason, the frequency ratio method is used to achieve the objective quantification of evaluation index, and the k-means clustering algorithm is used to achieve the selection of non-landslide sample data. The results show that based on the premise that the k-means clustering algorithm selects non-landslides, the training accuracy of the neural network has increased from 73% to 97%, and the training accuracy of the support vector machine has increased from 75% to 96%. Based on the GIS platform, the susceptibility index calculated by the neural network and support vector machine model is divided into five regions according to the natural break point method. The statistical results of the overlay analysis of the zoning map and the historical disaster points show that the evaluation result of the neural network is better than the support vector machine model in the global scope, and the global accuracy is 76% and 74%, respectively. The research results can provide reference for disaster prevention and mitigation in Nanjiang County of China.