SDCnet: A Deep Learning Network Model for Road Segmentation Based on CNN and Transformer

Road segmentation is a key task in remote sensing semantic segmentation, and the existing deep learning methods still have the problems of insufficient fineness, difficulty in modeling long-distance dependency, and poor adaptability to complex scenes. To improve the segmentation accuracy and robustness, this article proposes an improved road segmentation model SDCnet, which combines the advantages of CNN and Transformer, adopts Swin Transformer as an encoder to enhance the global information extraction, utilizes the DenseASPP+SP module to realize the multiscale feature fusion, and introduces coordinate attention to improve the spatial relationship modeling capability. The experimental results show that on the Massachusetts Road dataset, the proposed method achieves an IoU and mIoU of 71.09% and 84.16%, respectively; on the DeepGlobe Roads dataset, it achieves 68.08% and 83.00%, respectively. In addition, we conducted comparative experiments and analyses of different methods in complex scenarios. The results show that the method proposed in this article outperforms other comparative methods in terms of recognition performance, demonstrating high segmentation accuracy and robustness for road objects.