DRENet: A Dual-Branch Road Extraction Network for Enhanced Connectivity

Yao Yang, Pingfu Chao, Qiao Kun, Junhua Fang

Abstract: The road network is the foundation of the transportation system. However, the availability and the correctness of road networks always face challenges due to new road construction and frequent road changes. Instead of conducting labor-intensive ground surveys for map construction and update, automatic road network extraction via satellite images and/or trajectory data becomes the new trend. Nevertheless, although existing methods can extract road networks with the correct shape and decent coverage, few studies focus on road network connectivity, which is a crucial indicator of road network usability. In this paper, we propose a novel Dual-branch Network that improves connectivity and achieves accurate road extraction. Our model incorporates a Shape Reshaper Module for enriching the connectivity information and an Attention-based Fusion Module that dynamically captures the relationships between modalities, enabling effective fusion. Furthermore, we propose a connectivity measurement metric for road networks and a data augmentation method to decrease the impact of occlusions in satellite images. Extensive experiments on datasets from Beijing and Porto demonstrate that our approach achieves new state-of-the-art results. The source code can be found at https://github.com/sallwe/DRENet.

External IDs:dblp:conf/dasfaa/YangCKF25