Advancing Graph Convolutional Networks to Improve Road Network Selection for Small-Scale Maps

Legibility at smaller scales with preservation of the essential characteristics and structural patterns of an area is the most crucial task.Thus the process of cartographic generalization, relying on level of detail reduction, is fundamental to map design (Kraak et al., 2020).Road networks, which serve as the structural foundation of most general maps, pose a significant challenge for automated selection.This challenge is especially highlighted in the transition from medium scales (1:200,000-1:250,000) to small scales (1:500,000 and smaller), where feature omission rates often reach 60-70%.Such a high degree of reduction creates an unbalanced selection problem.This is also a particularly hard task to automate because it risks disrupting network connectivity or distorting the structural properties of the road system when using machine learning methods (Karsznia et al., 2024;Adolf & Karsznia 2025).While traditional methods such as stroke-based or mesh-based approaches have attempted to preserve network hierarchy, they can struggle with the complex relational structures required for small-scale representations.This research addresses these limitations by advancing a graph convolutional networks (GCN-based) approach that enables a model to learn selection patterns directly from the network's topology.