Network Cleaning

Raw street-network data is not ready for analysis. Sources such as OpenStreetMap, Ordnance Survey Open Roads, and US Census TIGER/Line represent the physical road surface, not the topological structure that centrality measures need. A single high street is often drawn as two parallel carriageways. Roundabouts become rings of segments, and complex junctions become clusters of slip roads and pedestrian crossings. These extra nodes and edges do not correspond to distinct decision points in the network.

Left uncleaned, this inflates node counts and distorts the metrics. Closeness and betweenness are pulled towards wherever the source data happens to be most finely subdivided. Angular centrality is affected most, because every spurious node introduces a spurious turn, and the measure is built from turn angles (see choosing shortest or angular). Cleaning is a required step before the metrics can be trusted.

The automated method

cityseer implements an automated network-cleaning and consolidation method that reduces a raw road surface to an analysis-ready topological graph: it merges parallel carriageways, consolidates complex junctions to single nodes, removes interstitial (“filler”) nodes that only subdivide a continuous street, drops short dangling stubs, and irons out geometric noise. This automated approach is evaluated by Abdeldayem et al. (2026), Environment and Planning B: Urban Analytics and City Science, which compares fully automated modelling against hybrid (manually refined) models across four cities. It finds that automated workflows capture large-scale integration patterns well, while local, flow-sensitive measures are more sensitive to how the network is segmented, which is exactly why the pipeline below is configurable rather than fixed.

On the high-level path, CityNetwork.from_osm applies a sensible default pipeline when simplify=True (the default), so most users get a well-cleaned network without extra steps. Every CityNetwork constructor also cleans input geometries during construction: short self-loops are removed, chains of segments meeting at filler (degree-2) endpoints are welded into single segments, short dead-end stubs are removed, and near-duplicate parallel edges are merged. Each step is controlled by a constructor parameter (remove_fillers, remove_danglers, merge_parallel_dist) and can be disabled to pass a network through as-is; directed networks skip the welding and merging steps automatically. This construction-time cleaning is deliberately conservative and does not replace the topological simplification described above: it does not consolidate complex junctions or merge separated dual carriageways.

It is configurable

The automated defaults suit most cases, but the right consolidation distance for a dense European core differs from a sprawling suburban grid, and some morphologies need more or less aggressive merging. The pipeline is fully exposed through the graphs module, so you can assemble and tune it step by step:

• nx_remove_filler_nodes removes degree-2 nodes that only subdivide a street.
• nx_consolidate_nodes merges nearby nodes (e.g. the several nodes of one junction) within a buffer distance.
• nx_merge_parallel_edges collapses dual carriageways and other parallel edges.
• nx_remove_dangling_nodes drops short dead-end stubs (the despine length) and disconnected fragments.
• nx_iron_edges straightens geometric noise below an angle threshold.

Tuning these parameters, and inspecting the result before computing centrality, is the reliable way to get a clean network for a specific dataset and morphology. The Network Simplification recipe works through the configurable pipeline step by step, and the Network Preparation recipes show the full range of construction routes.