Seg-Road: A Segmentation Network for Road Extraction Based on Transformer and CNN with Connectivity Structures

Abstract

Acquiring road information is important for smart cities and sustainable urban development. In recent years, significant progress has been made in the extraction of urban road information from remote sensing images using deep learning (DL) algorithms. However, due to the complex shape, narrowness, and high span of roads in the images, the results are often unsatisfactory. This article proposes a Seg-Road model to improve road connectivity. The Seg-Road uses a transformer structure to extract the long-range dependency and global contextual information to improve the fragmentation of road segmentation and uses a convolutional neural network (CNN) structure to extract local contextual information to improve the segmentation of road details. Furthermore, a novel pixel connectivity structure (PCS) is proposed to improve the connectivity of road segmentation and the robustness of prediction results. To verify the effectiveness of Seg-Road for road segmentation, the DeepGlobe and Massachusetts datasets were used for training and testing. The experimental results show that Seg-Road achieves state-of-the-art (SOTA) performance, with an intersection over union (IoU) of 67.20%, mean intersection over union (MIoU) of 82.06%, F1 of 91.43%, precision of 90.05%, and recall of 92.85% in the DeepGlobe dataset, and achieves an IoU of 68.38%, MIoU of 83.89%, F1 of 90.01%, precision of 87.34%, and recall of 92.86% in the Massachusetts dataset, which is better than the values for CoANet. Further, it has higher application value for achieving sustainable urban development.