A Multi-Scale Residential Areas Matching Method Using Relevance Vector Machine and Active Learning

Xinchang Zhang, Guowei Luo, Guangjing He, Liyan ChenISPRS Int. J. Geo Inf 2017 · DOI

TL;DR

This paper introduces a novel method for matching residential areas across different spatial scales, crucial for updating and integrating geographic data. By combining Relevance Vector Machines (RVM) with active learning, the approach automates the complex process of identifying corresponding features, significantly reducing manual effort and improving accuracy compared to existing methods.

What Problem Does This Paper Address?

Geographic Information Systems (GIS) rely on accurate and up-to-date spatial data for various applications, from urban planning to emergency relief. A significant challenge is integrating and updating data that exists at different scales. When data is generalized from a larger scale to a smaller scale, features can change in shape, size, and even disappear, making direct comparison difficult. This paper tackles the problem of matching residential areas across these multi-scale datasets. Traditional methods often struggle with the complexity of one-to-many or many-to-many relationships between features at different scales and require manual intervention for setting similarity thresholds and weights. This makes automated updating and fusion of spatial data a difficult task.

Methods

The core of this paper's contribution lies in its multi-scale residential area matching method, which consists of several key components:

  • Object Merging: To handle the complexities of one-to-many and many-to-many relationships, the authors propose a merging strategy that converts these into one-to-one relationships without simplification. This is achieved using Delaunay triangulation to identify and merge overlapping or adjacent residential areas from the larger scale data.
  • Feature Similarity Measurement: Five characteristics are used to quantify the similarity between residential areas:
  • Position Similarity: Based on the Euclidean distance between centroids.
  • Area Similarity: Calculated as the ratio of the minimum to maximum area of the two entities.
  • Shape Similarity: Uses a compactness measure (perimeter squared divided by area).
  • Orientation Similarity: An improved method using projective histograms of Smallest Minimum Bounding Rectangles (SMBR) to account for 180-degree rotations.
  • Surroundings Similarity: A grid-based approach that measures the similarity of neighboring features within a defined grid around the target residential area.
  • Relevance Vector Machine (RVM): This machine learning algorithm is employed for binary classification ("match" or "mismatch"). RVM is chosen for its suitability for small sample datasets and its ability to provide probabilistic outputs, indicating the confidence of a match.
  • Active Learning: To reduce the burden of manual sample labeling, an active learning strategy is integrated. This approach iteratively selects the most informative unlabeled samples for manual labeling, thereby improving the classifier's performance with fewer labeled examples.

Datasets

The study uses two datasets for experimental evaluation:

  • Residential areas data at 1:5000 scale.
  • Residential areas data at 1:25,000 scale.

These datasets are from the Tianhe District, Guangzhou, China. The paper does not introduce new benchmark datasets but uses existing multi-scale data to validate its method.

Key Insights and Ideas

  • Merging without Simplification: The decision to merge multi-scale features without simplification is crucial. It preserves more detailed geometric information, which is then leveraged by the similarity measures.
  • RVM for Probabilistic Matching: Using RVM provides a probabilistic output, allowing for a more nuanced assessment of matching reliability beyond a simple binary decision. This is particularly useful for handling ambiguous cases.
  • Active Learning for Efficiency: The integration of active learning is a significant practical contribution, directly addressing the time-consuming nature of manual sample labeling in machine learning for GIS. It demonstrates that high accuracy can be achieved with substantially less manual effort.
  • Improved Orientation Similarity: The novel orientation similarity measure using projective histograms is a key technical improvement, overcoming limitations of previous methods that struggled with rotated objects.

Key Results

The proposed method demonstrates strong performance in multi-scale residential area matching:

  • High Accuracy: The overall matching precision exceeds 90%.
  • Matching Type Performance:
    • 1:1 matching achieved the highest precision (95.5%), recall (96%), and F1 score (95.7%).
  • 1:m matching achieved precision of 91.9%, recall of 91.4%, and F1 of 91.6%.
  • m:n matching achieved precision of 82.2%, recall of 83.3%, and F1 of 82.7%.
  • Comparison with Other Methods: The RVM-based method with active learning outperformed other methods, including buffer overlap, weighted similarity, and SVM, in terms of precision, recall, and F1 scores.
  • Reduced Manual Intervention: The active learning approach significantly reduces the workload for manual sample labeling.

Limitations and Future Work

The authors acknowledge several areas for future improvement:

  • Shape Similarity Measurement: The current shape similarity measure needs further development to accurately handle area entities with very complex shapes, such as those with numerous cavities.
  • Processing Time for Large m:n Relations: When the number of entities involved in 1:m or m:n matching relations (m and n) becomes very large, the number of candidate matches increases, leading to longer processing times. Enhancing the efficiency of selecting matching groups is suggested.

Summary updated

BibTeX
@article{zhang2017multi,
  author = {Xinchang Zhang and Guowei Luo and Guangjing He and Liyan Chen},
  title = {{A Multi-Scale Residential Areas Matching Method Using Relevance Vector Machine and Active Learning}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2017},
  doi = {10.3390/ijgi6030070},
}
Figure