It is of great significance to locate new landslide disaster quickly from remote sensing image for emergency rescue work. It is a hot research topic to segment landslides from a large number of remote sensing images by semantic segmentation method. This paper introduces Transformer series models for landslide segmentation for the first time and compares them with CNN models. At the same time, we set up two groups of experiments to discuss the influence of negative samples on the model recognition accuracy. The experiment results show that: (1)Transformer series models are better than CNN architecture models. Among them, Transunet has the highest identification accuracy for landslide(81.8%). Swin has the best comprehensive performance, and its overall accuracy reaches 98.5%, IoU reaches 73.13%(landslide), and MIoU reaches 85.79%. (2)The addition of negative samples improves the missed detection rate of the model, but greatly reduces the error detection rate of the model, which is of great significance for practical application.