Abstract: Rapid identification and detection of landslides can both meet the requirement of timely responses to disasters and hold great significance for loss assessment and rescue post-disaster. This study proposed a deep learning-based automatic information extraction method for landslides to improve their detection accuracy. Specifically, the model input of this method includes the remote sensing images of the target areas, data from digital elevation models, and variation characteristics extracted using robust change vector analysis (RCVA). Furthermore, a U-Net model integrating dense upsampling and asymmetric convolution is designed to improve the identification accuracy. Taking Wenchuan, Sichuan Province as the study area, this study designed experiments to test the pixel-level image segmentation accuracy of landslides using different data combinations and methods. The results indicate that the improved U-Net model proposed in the study can produce the optimal image segmentation results of landslides.