Road Network Selection Improvement Using Graph Convolutional Networks for Small-Scale Mapping

This paper presents a method for automated road network selection in small-scale cartographic generalization based on graph convolutional networks (GCNs), with a focus on transitions from medium scales (1:200,000–1:250,000) to small scales (1:500,000 and smaller). Road networks are modelled using a line graph representation, enabling segment-level classification while preserving topological relationships. Each segment is enriched with geometric, centrality, functional, and proximity attributes, and higher-order subgraphs are constructed to capture the overall local and contextual network structure. The method incorporates a smoothness regularizer during training and a connectivity-based post-processing step to reduce fragmentation and enforce structural coherence. Experiments were conducted on datasets from Poland and Switzerland using national reference maps as ground truth, with country-specific training to reflect differences in data characteristics and road classification systems. Quantitative evaluation using accuracy, precision, recall, and F1-score, supported by detailed visual analysis, shows that the resulting networks retain substantial fragmentation, particularly in Poland, indicating that further improvements are needed to achieve network connectivity, although higher overall performance is observed in the Swiss case study. The results demonstrate the applicability of GCNs for small-scale road selection while highlighting the influence of geographic context, data quality, and the limitations of standard classification metrics in evaluating cartographic generalization.