PAG-Road:Pseudo-Node Attention Guided Adversarial Generation for Hierarchical Road Network Topology

Road network generation plays a pivotal role in urban planning. Classifying and structuring the generated road networks not only mirrors the actual road distribution more accurately but also provides crucial data for autonomous driving simulation tests. Consequently, this paper focuses on employing a graph-based generative approach to create hierarchical road network topologies. Initially, the paper introduces a topological representation model for hierarchical road networks. Addressing the current lack of publicly available hierarchical road network topology datasets, we have developed a comprehensive dataset that includes data from 71 medium to large cities in China, thus filling this significant gap. Lastly, this paper proposes a novel adversarial generative model for hierarchical road network topologies that facilitates feature sharing and achieves task decoupling and feature coupling in generating both node attributes and links. Notably, we introduce an innovative pseudo-node attention mechanism that models global associations between connected and unconnected nodes in a semi-supervised manner. Comparative experiments with other methods demonstrate that our approach exhibits superior precision and robustness in generating node attributes and links, with performance metrics exceeding 98.5 %.