Due to the influence of topography, landform, vegetation, and other factors, the morphology and spectral characteristics of landslides are different. Therefore, it is difficult to extract the landslide information accurately by classification and identification methods based on spectra and shapes. Deep learning technology can extract the features of objects by continuous training of samples, which avoids the subjectivity of artificially constructed features, and can form more abstract and stable features of the target object, which improves the recognition accuracy greatly. DeepLabV3+ is a semantic segmentation model based on deep learning method, which integrates image segmentation and target recognition, and can effectively segment the target object from the original image while recognizing the target. In this paper, we combine the landslide characters with DeepLabV3+ to identify landslides, and the network architecture in the original model is optimized by Xception, MobileNetV2, and ResNet. Through the Bijie landslide data set and the satellite landslide images of the Wenchuan Earthquake, it is verified that the optimized model can more accurately identify and segment the landslides appearing in the images. The maximum overall segmentation accuracy, the maximum recognition accuracy, and the maximum F1_Score have reached 95.3%, 97%, and 97.1% respectively. [ABSTRACT FROM AUTHOR]