The road network similarity calculation based on autoencoders and its application in map generalization quality evaluation

This dataset comprises multi-scale road network data extracted from OpenStreetMap (OSM) for five Chinese cities—Chengdu (radial-ring pattern), Wuhan, Kunming, Shenzhen, and Hefei (grid-irregular mixed pattern)—supporting advanced spatial similarity analysis and cartographic generalization quality assessment. The data were manually extracted from OSM (accessed June 2025) at an original scale of 1:10,000 and generalized to five additional scales (1:25,000; 1:50,000; 1:100,000; 1:250,000; 1:500,000) by expert cartographers following national topographic mapping specifications. Each city's dataset includes 6 road network files (original + 5 generalized versions) in GeoJSON format, representing dual graph structures (nodes for road segment attributes, edges for connectivity) with extracted features: connectivity (node degree), global distribution (angle α and distance difference Δl relative to network centroid), and local geometry (length, tortuosity, MBR major axis direction, mean direction). Total: 30 files (~5-10 MB per city), preprocessed to remove pseudo-nodes and isolated roads, normalized for model input. These datasets enable replication of a Graph Convolutional Autoencoder (GCAE) model for road network similarity computation via cosine distance in latent space, construction of a logarithmic quantitative similarity-scale model (S_sim = -0.1261 ln(x) + 1.0147, R²=0.989), and quality evaluation via deviation (E = |S - R|) and rating sets (Excellent/Good/Fair/Poor). Derived from public OSM sources under ODbL, processed files are released under CC BY 4.0 for reuse in spatial cognition, multi-scale mapping, and deep learning applications. Related publication: Song et al. (2025), "The Road Network Similarity Calculation Based on Autoencoders and Its Application in Map Generalization Quality Evaluation," Cartography and Geographic Information Science (DOI: [pending]). For replication code and full methods, refer to the paper's supplementary materials.