U3-Road: A CNN and Transformer Multi-Level Nested Model for Road Extraction from Remote Sensing Images

As a crucial prerequisite for intelligent transportation and urban digitalization, the importance of road extraction is evident. In recent years, the rapid development of remote sensing technology and deep learning has made road extraction from remote sensing images a research focus in both academia and industry. However, the topological complexity of road networks, interference from ground-object occlusion, and the spectral and spatial limitations of remote sensing images have presented numerous challenges to the design and optimization of related algorithms. To address these issues, this paper proposes a road extraction algorithm, U3-Road, which combines the advantages of CNN and Transformer architectures. This algorithm adopts an encoder–decoder structure and uses multi-path nesting technology in the encoder and bridge modules, enabling the extraction of multi-scale contextual features and semantic information, which are fully integrated with the decoder. Additionally, it employs a composite loss function to focus on difficult-to-identify sections and constrain problems such as road breaks and deformations, thereby achieving better accuracy and connectivity. Tests on the DeepGlobe, Massachusetts, and CHN6-CUG datasets show that the Intersection over Union (IoU) of U3-Road reaches 70.47%, 69.55%, and 70.87%, demonstrating leading performance. It effectively alleviates the problems of road breaks and misidentification, and compared with other representative methods, U3-Road performs better in the completeness and accuracy of road extraction results, achieving the best overall extraction effect while maintaining a lower parameter count and computational cost. This model provides a new solution for road extraction from remote sensing images.