How GENREGION Revolutionizes City Region Segmentation with Vector Road Networks

City region segmentation underpins spatial analysis tasks such as urban analytics, trajectory mining, and traffic forecasting. Accurate, semantically meaningful partitions enable precise updates of mapping data, clearer transportation planning, and reduced manual interpretation errors.

Existing approaches fall into three categories, each with notable drawbacks:

Traditional grid method : divides the map into equal‑size rectangles, ignoring semantic structures like functional boundaries and road networks.

Road‑network‑driven raster model : simplifies the network with morphological operations and labels connected components, but relies on raster images, limiting topological analysis.

Road‑network‑driven vector model : merges nodes and applies Dijkstra‑based recursion; however, boundary logic is undocumented and computational complexity grows super‑linearly with network size.

To address these issues, the authors present GENREGION , the first open‑source, vector‑based, hierarchical road‑network segmentation system, released on PyPI and hosted at GitHub . The system treats the road network as a graph (edges = road segments, vertices = endpoints) and consists of three core modules.

PART1: Clustering‑Based Road Segment Simplification

The method reduces redundant details (e.g., interchange loops, curved or circular roads) via hierarchical clustering. The key innovation is a new distance metric and representative‑point design:

Instead of Euclidean distance, clusters are defined by a Minimum Bounding Rectangle (MBR); the MBR center becomes the representative point, eliminating density bias.

Cluster distance is infinite if MBRs exceed a threshold; otherwise it is computed from MBR coordinates. A grid index limits candidate comparisons to neighboring cells, merging clusters whose distance falls below the threshold.

PART2: Closed‑Polygon Generation

Using the simplified network, the algorithm performs “split – vectorize – recursive edge search” to build region boundaries:

Road segments are split at intersections so each segment connects exactly two endpoints.

Each segment is represented by two opposite vectors (invec/outvec). A “leftness” value, derived from trigonometric calculations, identifies the most left‑ward outward vector.

Starting from an unused vector, the process repeatedly finds the leftmost outward vector until the path closes, forming a polygon. Each vector is used once, ensuring efficient, non‑redundant polygon creation.

PART3: Polygon Refinement

After initial polygon generation, two types of invalid regions are removed and the results are standardized:

Sub‑polygons : polygons wholly contained within a larger one are discarded using MBR comparisons.

Micro / elongated polygons : tiny areas or long‑narrow shapes (e.g., highways) are identified via area and width thresholds and merged into adjacent regions.

The final output is a list of shapely Polygon objects ready for downstream spatial analysis.

Experimental Evaluation

Semantic effectiveness : Using ROI data from five Chinese cities (Beijing, Shanghai, Shenzhen, Wuhan, Chengdu) and Jaccard similarity as the metric, GENREGION achieved the highest scores (41.0%–42.7%) compared to shortpath, mapseg, and grid methods.

Computational performance : Runtime grows slowly with increasing road‑network nodes for GENREGION, while shortpath exhibits steep growth, demonstrating superior scalability.

Demonstration on Beijing road network : Using publicly available data from Figshare, the workflow (preprocessing, clustering, polygon generation, refinement) produces 12,813 polygons for the full network, illustrating practical applicability.

Technically, GENREGION’s MBR‑based clustering mitigates node‑density bias, while recursive left‑most vector search reduces algorithmic complexity, maintaining high performance on large networks. Ecologically, the open‑source release provides a valuable tool for the research community, fostering further advances in urban spatial analysis.