A deep Learning-based Framework that Uses Only Geometry Data to Predict Road Class, Leveraging Principles of Persistent Homology

Accurate road class information is essential in transport application, yet it remains inconsistent and partially incomplete across many regions of OpenStreetMap. Although recent advances in AI-based road extraction have substantially improved the availability of road geometries, these methods do not provide attributes such as road class. Existing approaches to road class prediction typically rely on contextual data, including max-speed and buildings, which are often unevenly distributed and incomplete, particularly in under-mapped regions. In this study, we introduce TopoNET, a deep learning–based framework that uses road geometry data for road class prediction while leveraging multi-scale topological features derived from principles of persistent homology. By encoding intrinsic topological patterns of road networks, TopoNET avoids reliance on external contextual data and focuses exclusively on network structure and geometric properties. We evaluated the approach using road networks from fifteen German administrative areas and compared its performance against both context-dependent and geometry-only state-of-the-art methods. Our results show that TopoNET consistently outperforms existing geometry-only models and substantially narrows the performance gap between geometry-only and context-dependent approaches. The model recorded accuracy of over 70% across all study areas, although classes with weak network integration, such as living streets and cycleways remain more challenging to predict. A comparative transferability analysis between Berlin (85%), and Frankfurt (Main) (79%) highlights differences in model generalization, with Frankfurt-trained model performing better with an average of 5% more and maximum of 7% more across all regions. Overall, this study demonstrates that persistent homology provides a foundation for geometry-only road classification.