In order to solve the problems of insufficient landslide catalog data in the process of landslide susceptibility evaluation, and low model accuracy due to subjective or random selection of non-landslide raster cells, 12 evaluation factors were selected from the aspects of topography and geomorphology, geological conditions, hydrometeorological conditions and human engineering activities to construct a landslide evaluation system for Xianrendong National Nature Reserve in southern Liaoning Province in this paper; Furthermore, the imbalance of sample categories between landslide and non-landslide was solved based on SMOTETomek comprehensive sampling method, and then a dataset of landslide susceptibility evaluation was established; Finally, for the nonlinear landslide data in the east and west sides of the study area (zones A and B), the convolutional neural networks (CNN) model was constructed to evaluate the landslide susceptibility, and the distribution map of landslide susceptibility in the study area was accurately drawn. The results show that the CNN model had good adaptability, and the zoning map of landslide susceptibility shows a reasonable spatial distribution. The AUC area of the test set in part A and part B of the study area was 91.2% and 94.3%, respectively.70% of landslides were distributed in higher and above grade prone areas, and 68.7% of non-landslides were distributed in lower and below grade prone areas. The high prone area of landslide is mainly located in the area of Maoling Beigou Mountain in the northeast of the study area, the northern mountainous area of the Bingyugou Scenic Area, and the coastal area of the Biliuhe Reservoir. The research results can provide an important scientific basis for the planning of geological disaster prevention and control, and the formulation of emergency plans in Xianrendong National Nature Reserve in southern Liaoning Province.