In order to solve the challenges of occlusion and illumination changes in the automatic classification of urban landscape elements, Deeplabv3+ is used to achieve pixel-level semantic understanding and high-resolution boundary restoration. The ResNet-101 backbone network is integrated with the ASPP (Atrous Spatial Pyramid Pooling) module, and the multi-scale dilated convolution and global average pooling branches are combined to enhance the collaborative modeling of local details and global semantics. A lightweight decoder is designed to fuse low-level features with ASPP output, and boundary details are restored through bilinear interpolation. The edge-aware loss function and category weight balancing mechanism are introduced to optimize the detection accuracy of small targets and boundary pixels. Experiments based on the Cityscapes dataset show that the model has an average mIoU (Mean Intersection over Union) of 80.7% on 19 types of elements, outstanding performance in small target detection (76.4% for traffic lights, 79.8% for pedestrians) and occlusion scenes (mIoU decreases by only 3.6%), an inference speed of 25.4 FPS (Frames Per Second), and a boundary F1 score of 78.6%, verifying its robustness and real-time performance in complex environments. The model in this paper provides a high-precision solution for urban scene segmentation.