MT-RoadNet: A heterogeneous network with local–global joint enhancement for road surface and centerline extraction

Road network information has extensive applications, such as urban planning and navigation. However, current road extraction methods mostly rely on single model architectures and single-source remote sensing images, neglecting the potential benefits of collaborative extraction with heterogeneous networks and multi-source data fusion. Moreover, existing methods often suffer from road fragmentation and poorly connected road graphs due to occlusions. To address these challenges, we propose MT-RoadNet, a local–global joint enhancement framework for extracting road surfaces and centerlines, which incorporates multi-network collaborative optimization. Specifically, MT-RoadNet adopts a dual-branch structure, which not only facilitates the collaborative extraction of local details and global semantics but also achieves the coupling of geometry and topology through a cross-task dynamic interaction mechanism. Besides, we propose the Local–Global Feature Fusion module (LGFF), which dynamically integrates local details and global semantics through multi-level feature interaction. Furthermore, to reduce interference features with high inter-class separability and low intra-class variation, we innovatively design the Visual State Space Module (VSSM) and the Spatial-Channel Mutual Attention (SCMA). The VSSM weighs features dynamically using multi-directional cross-scanning and global receptive fields, emphasizing prominent area information while improving computational efficiency. SCMA effectively guides the model to focus on semantically relevant regions. Finally, MT-RoadNet adopts a dual-path decoder to produce road surfaces and centerlines. Extensive experiments on three road datasets demonstrate that MT-RoadNet significantly outperforms existing state-of-the-art methods in terms of road completeness and recognition accuracy of topological structure. The code is available at https://github.com/508hz1207/MTNet . • MT-RoadNet is a scalable multi-network framework for road and centerline extraction. • LGFF performs bidirectional fusion between local and global features. • VSSM models features of directional linear structures. • SCMA performs parallel spatial-channel fusion to enhance feature discriminability. • Dual-branch decoder achieves differentiated modeling and cross-task interaction.