TransRoadNet: a Transformer framework for multi-modal road network pattern recognition

Road network pattern recognition plays a vital role in traffic prediction, route planning, and the development of intelligent transportation systems. However, most existing approaches are limited to single-pattern recognition and often yield suboptimal accuracy. This study introduces TransRoadNet, a multimodal hybrid framework that combines geographic spatial relationship theory with a Transformer-based architecture. The method decomposes OpenStreetMap vector data into seven representative structural categories, including grid, radial, and irregular patterns. It constructs a ten-dimensional feature vector capturing road segment characteristics at the macro, meso, and micro levels. These features are embedded and processed through a multi-head self-attention mechanism to model cross-scale spatial dependencies. Experimental results on road data from Chengdu indicate that TransRoadNet achieves an accuracy of 0.97 ± 0.01, significantly outperforming conventional models, including CNNs, GCNs, SVMs, and random forests. Additional validation on Berlin and Shanghai road networks demonstrates the model’s strong generalization across different urban environments. Due to its modular design and compatibility with parallel computing, TransRoadNet exhibits strong scalability for large-scale deployments. Compared to pixel-based image segmentation methods, the model emphasizes topological and semantic structure modeling, offering enhanced adaptability for urban systems and supporting downstream applications such as traffic control, navigation, and infrastructure planning.