In order to address the limitations of traditional high-resolution landslide image segmentation methods in handling details and blurred boundaries, an enhanced Deeplabv3+ model (SCPD-Deeplabv3+) was proposed, which integrated Swin Transformer network, convolutional block attention module (CBAM), position attention feature pyramid network (PA-FPN), and multi-layer convolutional decoder. Firstly, the baseline model Deeplabv3+ was improved by adopting Swin Transformer as the backbone network, introducing CBAM into the atrous spatial pyramid pooling (ASPP) module of Deeplabv3+, integrating PA-FPN into the decoder, and adding more convolutional layers during the upsampling process. Secondly, the improved Deeplabv3+ model was trained. Finally, the high-resolution landslide image test set was fed into the trained SCPD-Deeplabv3+ model for ablation experiments to analyze the role of each module, and comparisons with mainstream models such as UNet, proportional-integral-derivative network (PIDNet), and real-time transformer (RTFormer) for semantic segmentation were performed through quantitative evaluation and visualization. The results show that SCPD-Deeplabv3+ achieves an average intersection over union of 90.18%, precision of 93.57%, recall of 94.47%, and F1-score of 93.58%, respectively, which are improved by 3.39 percentage points, 1.45 percentage points, 3.90 percentage points, and 3.51 percentage points compared with the unmodified model. The proposed method effectively enhances the segmentation accuracy and detail restoration capability for landslide areas, providing a reliable technical means for remote sensing landslide monitoring and disaster assessment.