SP-GEM: Spatial Pattern-Aware Graph Embedding for Matching Multisource Road Networks

Jian Yang, Chenghao Zheng, Yunfei Qiu, Jian Yang 0034, Bianying Zhang, Zeyuan Li, Zhangxiang Lin, Xianglin Zhang, Yang Hou, Li FangIsprs Int. J. geo inf 2025 · DOI

TL;DR

This paper introduces SP-GEM, a novel graph embedding method for matching road networks from different sources. By explicitly modeling spatial patterns within road network structures, SP-GEM significantly improves matching accuracy and recall, outperforming existing methods, especially in complex urban environments.

What Problem Does This Paper Address?

Matching road networks from different sources (e.g., OpenStreetMap vs. commercial datasets) is crucial for data integration, quality assessment, and urban planning. Traditional methods struggle with the inherent differences in data acquisition, resolution, and topological complexity, leading to poor performance, especially in dense and irregular urban road structures. Existing deep learning approaches, while promising, often fail to fully leverage the rich spatial context of road networks, limiting their matching capabilities.

Methods

The core innovation of SP-GEM lies in its "Spatial Pattern-Aware Neighbor Aggregation" within a Graph Neural Network (GNN) framework, specifically using GraphSAGE. The process involves:

  • Road Graph Construction: Road networks are converted into graphs where road segments are nodes and edges represent spatial proximity.
  • Feature Extraction: Geometric (e.g., MBR aspect ratio, direction) and topological (e.g., node degree) features are extracted for each road segment. Semantic features are intentionally excluded to reduce reliance on potentially missing data.
  • Spatial Pattern Identification: Four distinct spatial patterns are identified in road networks: Grid, High-branching degree, Irregular grid, and Circuitous. These patterns are characterized by features like average node degree, log circuity, and edge ratios.
  • Spatial Pattern-Aware Embedding: This is the key novelty. Instead of standard neighbor aggregation, SP-GEM divides a node's neighborhood into eight directional regions. Within each region, it aggregates features based on the identified spatial patterns of neighboring nodes. This explicitly models the spatial arrangement and patterns of roads, enhancing the expressiveness of the node embeddings.
  • Similarity Measure: Road segment embeddings are then used to calculate similarity scores (combining Hausdorff, Cosine, and Manhattan distances) to identify correspondences between road networks.

Datasets

The study utilizes two real-world road network datasets for evaluation:

  • OpenStreetMap (OSM): A crowdsourced, open-source global map dataset.
  • Tele Atlas: A commercial map provider known for its professional data collection.

The experiments are conducted on road network data from the city of Ansbach, Germany, using both OSM and Tele Atlas sources. The paper does not introduce new benchmark datasets but uses existing ones to validate its approach against established methods.

Key Insights and Ideas

  • Spatial Patterns are Crucial: The paper strongly emphasizes that explicitly modeling spatial patterns in road networks is key to improving matching performance, especially in complex and heterogeneous environments.
  • Beyond Simple Proximity: Standard GNNs often rely on basic spatial proximity. SP-GEM demonstrates that understanding how roads are spatially arranged (their patterns) provides more discriminative information than just their geometric or topological features alone.
  • Robustness to Granularity: SP-GEM shows improved performance in handling inconsistencies in road segment granularity between different datasets, a common challenge in real-world road network matching.
  • Complementarity of Features: The combination of geometric, topological, and spatial pattern features yields the best results, highlighting the benefit of multi-dimensional information fusion.

Key Results

  • Improved Accuracy and Recall: SP-GEM significantly outperforms baseline methods. It achieves at least 6.7% higher accuracy and 10.2% higher recall compared to a baseline method without spatial pattern modeling. Compared to a spatially explicit model (GS-SP), SP-GEM improves accuracy by 2.68% and recall by 1.29%.
  • High Matching Success Rate: The method achieves a high matching success rate of over 70%.
  • Superior Performance on Complex Scenarios: SP-GEM demonstrates notable advantages in scenarios with complex road structures, such as dense intersections or parallel roads, where other methods struggle. For instance, it achieved zero matching error in a case with many parallel highways.
  • Effective Embedding Visualization: t-SNE visualizations of road embeddings show that SP-GEM effectively clusters road patterns, clearly distinguishing between different types (Grid, Irregular Grid, etc.), unlike baseline models.

Limitations and Future Work

  • Time-Consuming Training: The current implementation requires significant computational resources and time for training (e.g., ~10 hours for the full Ansbach dataset on an RTX A6000 GPU).
  • Parameter Selection: The paper notes that parameter selection for constructing spatial proximity graphs can be a challenge.
  • Limited Geographic Scope: The study primarily focuses on European city regions.
  • Future Work Suggestions:
    • Efficient Graph Learning: Explore graph partitioning and sampling techniques to speed up training.
  • Broader Validation: Extend validation to underdeveloped cities and diverse global regions.
  • GeoEmbedding Service: Develop a standardized GeoEmbedding service for road networks to facilitate machine learning applications and data discovery.

Summary updated

BibTeX
@article{yang2025sp,
  author = {Jian Yang and Chenghao Zheng and Yunfei Qiu and Jian Yang and Bianying Zhang and Zeyuan Li and Zhangxiang Lin and Xianglin Zhang and Yang Hou and Li Fang},
  title = {{SP-GEM: Spatial Pattern-Aware Graph Embedding for Matching Multisource Road Networks}},
  journal = {Isprs Int. J. geo inf},
  year = {2025},
  doi = {10.3390/ijgi14070275},
}
Figure