GLD-Road: A global–local decoding road network extraction model for remote sensing images

Road networks are essential information for map updates, autonomous driving, and disaster response. However, manual annotation of road networks from remote sensing imagery is time-consuming and costly, whereas deep learning methods have gained attention for their efficiency and precision in road extraction. Current deep learning approaches for road network extraction fall into three main categories: postprocessing methods based on semantic segmentation results, global parallel methods and local iterative methods. Postprocessing methods introduce quantization errors, leading to higher overall road network inaccuracies; global parallel methods achieve high extraction efficiency but risk road node omissions; local iterative methods excel in node detection but have relatively lower extraction efficiency. To address the above limitations, We propose a two-stage road extraction model with global–local decoding, named GLD-Road, which possesses the high efficiency of global parallel methods and the strong node perception capability of local iterative methods, enabling a significant reduction in inference time while maintaining high-precision road network extraction. In the first stage, GLD-Road extracts the coordinates and direction descriptors of road nodes using global information from the entire input image. Subsequently, it connects adjacent nodes using a self-designed graph network module (Connect Module) to form the initial road network. In the second stage, based on the road endpoints contained in the initial road network, GLD-Road iteratively searches local images and the local grid map of the primary network to repair broken roads, ultimately producing a complete road network. Since the second stage only requires limited supplementary detection of locally missing nodes, GLD-Road significantly reduces the global iterative search range over the entire image, leading to a substantial reduction in retrieval time compared to local iterative methods. Finally, experimental results revealed that GLD-Road outperformed current state-of-the-art methods, achieving improvements of 1.9% and 0.67% in average path length similarity (APLS) on the City-Scale and SpaceNet3 datasets, respectively. Moreover, compared with those of a global parallel method (Sat2Graph) and a local iterative method (RNGDet++), the retrieval time of GLD-Road exhibited reductions of 40% and 92%, respectively, suggesting that GLD-Road achieves a pronounced improvement in road network extraction efficiency compared to existing methods. The experimental results are available at https://github.com/ucas-dlg/GLD-Road.