关键词: 路网模式, 空间认知, 地图分析, 图预训练, 图结构学习

Abstract:

Road network pattern recognition is a classical problem in the fields of cartography and geospatial information science, and is widely applied in tasks such as map generalization, spatial cognition, and vector data updating. The extraction and analysis of road network pattern features constitute a crucial step in reflecting urban landscape characteristics and enabling intelligent map operations. However, existing methods face two major challenges: first, a heavy reliance on large amounts of high-quality labeled data; second, an incomplete representation of the spatial-topological information of road networks. As typical unstructured graph data, road networks cannot be directly processed by pre-training methods designed for regular grid data. To address these issues, this paper proposes a self-supervised pre-training model that integrates Euclidean spatial proximity and topological adjacency information. Method ologically, on one hand, a graph convolutional kernel network is employed to encode local path structures, enhancing topological awareness; on the other hand, a spatial attention bias mechanism is designed to incorporate Euclidean distance information into graph node representations, achieving an effective fusion of dual spatial information and thus better capturing the characteristics of road networks. In terms of the learning paradigm, a pre-training strategy is adopted, learning universal representations using only unlabeled road network data, thereby reducing dependence on labeled data. Experimental results on public urban road network datasets demonstrate that the proposed model achieves significant advantages in the road network pattern classification task: an accuracy of 88.03%, surpassing the optimal baseline method (CRHD) by 2.29 percentage points; precision and recall reach 88.89% and 87.47%, respectively, with improvements of 3.19 and 1.57 percentage points; and anF₁ score of 0.878 0, an increase of 0.020 9. This paper validates the effectiveness of the fusion architecture and the pre-training strategy. The proposed model can learn more generalizable intrinsic representations of road networks, providing foundational model support for downstream tasks.

Key words: road network pattern, spatial cognition, map analysis, graph pre-training, graph structure learning