• A prior-guided road extraction framework that fuses multidimensional priors to reduce background complexity. • An orientation-adaptive attention mechanism that aggregates features along estimated road directions to improve continuity. • An uncertainty-adaptive deep supervision strategy that balances multi-scale constraints to recover narrow roads. Accurate road extraction from remote sensing imagery is vital for urban planning and intelligent transportation. However, its accuracy is often limited by complex backgrounds, spectral ambiguity, and diverse road shapes. Recent deep learning methods have made strong progress, including topology-aware designs that explicitly encourage structural consistency; however, performance can still degrade under weak textures, occlusion, and rapidly changing orientations, where thin roads become fragmented and boundaries blur. To address these limitations, RoadAttNet is developed to enhance road network integrity and boundary delineation. The architecture integrates multidimensional physical priors through a learnable weighting mechanism. Furthermore, an Oriented Coordinate Attention (OCA) module is incorporated to reinforce directional continuity and refine boundary precision via oriented pooling operations. Feature representation and training stability are further optimized through an adaptive deep supervision strategy. Quantitative evaluations on the Massachusetts and Lanzhou datasets demonstrate the superiority of RoadAttNet over existing benchmarks. Compared to state-of-the-art methods such as MSMDFF-Net, the proposed architecture achieves improvements of up to 2.3% in IoU and 1.43% in F1-score. Additionally, the clDice connectivity metric increases by 2.1%, indicating a substantial reduction in network fragmentation and enhanced topological consistency. The relevant code and dataset are available at https://github.com/DLRS2025/RoadAttNet .
Zhang et al. (Mon,) studied this question.