Change Detection for Building Footprints with Different Levels of Detail Using Combined Shape and Pattern Analysis

Xiang Zhang (张翔), Xiaodong Zhou, Zhe Chen, Xiang Zhang, Tinghua AiIsprs Int. J. geo inf 2018 · DOI

TL;DR

This paper tackles the challenge of accurately detecting changes in building footprints between datasets with varying levels of detail (LOD), like OpenStreetMap and official topographic data. They introduce a novel rule-based system that combines geometric analysis, shape similarity, and crucially, building pattern recognition to overcome common issues like boundary mismatch and generalization artifacts, achieving a high accuracy of 0.90.

What Problem Does This Paper Address?

Keeping spatial databases up-to-date is crucial for urban planning, geoinformatics, and real estate. While crowd-sourced data like OpenStreetMap (OSM) offers timely updates, it often has different quality and levels of detail (LOD) compared to professional datasets from national mapping agencies. This discrepancy leads to challenges in automated change detection, including:

  • Boundary Mismatch: Differences in how building outlines are captured.
  • Level of Detail (LOD) Differences: One dataset might include annexes while another only shows roofs, or one might be more generalized.
  • Many-to-Many Correspondences: Multiple objects in one dataset might correspond to a single object in another.
  • Non-Systematic Offsets: Small positional shifts that aren't easily corrected by geo-referencing.

These issues can lead to false positives (identifying non-changes as changes) and false negatives (missing actual changes), making automated change detection unreliable, especially when integrating heterogeneous data sources.

Methods

The authors propose a multi-stage, rule-based approach to address these challenges:

  1. Object Matching: Initially, adjacent buildings of the same type in OSM are merged to improve comparability. Then, an overlap-based analysis (using a CommonAreaRatio threshold) identifies corresponding objects between the two datasets. This step is designed to handle both 1-to-1 and many-to-many correspondences.
  2. Aggregation for Many-to-Many: To simplify analysis, many-to-many relationships are converted to 1-to-1 relationships through a "pair-wise aggregation" process. This method is guided by distance and includes a "post-evaluation" step to detect potential changes if aggregation creates too much empty space.
  3. Controlled Alignment: To mitigate boundary mismatch and non-systematic offsets, a controlled alignment technique is employed. Small buildings with simple shapes are aligned by their centroids. For more complex shapes, alignment is only performed if the difference parts are comparable in size, suggesting a positional shift rather than a shape change.
  4. Shape Similarity (Turning Function): A turning function is used to measure the global shape similarity between building footprints. This helps in situations where similar shapes might have shifted.
  5. Morphological Analysis of Difference Parts: The differences between overlapping polygons (A-B, B-A) are analyzed. This goes beyond simple size comparison to examine the shape, fragmentation, and distribution of these difference parts. For large buildings, criteria like width, length, and elongation ratio are used to distinguish true changes from non-changes.
  6. Pattern and Contextual Information: This is a key novel contribution. Building alignments (groups of buildings with similar forms, sizes, and layouts) are recognized. If a building identified as "changed" is part of a well-maintained alignment, and the change is minor, it can be re-classified as "unchanged" (inconsistency correction). This leverages the idea that entire neighborhoods are often modified or replaced, rather than individual buildings within them.

Datasets

  • OpenStreetMap (OSM): Used as the "more current" dataset.
  • Dutch Topographic Dataset (TOP10NL): Used as the "to-be-updated" dataset.

The study specifically used data from Geldermalsen, NL, which includes diverse land-use characteristics (city centers, urban, suburban, rural, and industrial/commercial). The paper does not introduce new benchmark datasets but rather uses existing ones to evaluate their method.

Key Insights and Ideas

  • Distinguishing "Real" Changes: The core insight is that simply comparing geometric properties (like size) is insufficient. True change detection requires accounting for data acquisition differences, generalization, and varying levels of detail.
  • Building Patterns as Heuristics: Recognizing that buildings often exist in contextual groups (alignments) is crucial. Changes within these groups are not always individual events; sometimes, entire alignments are updated or replaced. This pattern-based correction significantly improves accuracy by reducing false positives.
  • Small vs. Large Buildings Matter: Different rules and parameters are more effective for small buildings (where shape similarity is less reliable due to generalization) versus large buildings (where morphology of differences is key).
  • Controlled Alignment is Key: Addressing boundary mismatch and offsets through intelligent alignment, rather than brute-force geo-referencing, is vital for accurate shape comparison.
  • Morphology Over Size: For large buildings, the shape and distribution of the difference parts are more indicative of change than just their absolute or relative size.

Key Results

  • The proposed approach achieved an accuracy of 0.90 and a kappa coefficient of 0.77.
  • This significantly outperformed a basic geometric method, which achieved an accuracy of 0.77 and a kappa of 0.47.
  • The "Advanced + Pattern" version of their method, incorporating building pattern correction, yielded the highest performance with an accuracy of 0.92 and kappa of 0.81.
  • The controlled alignment technique was effective in handling non-systematic offsets, with a median offset of 3.38 m and a maximum of ~9 m.
  • The method successfully corrected 94 cases through the conservative pattern-based inconsistency correction.

Limitations and Future Work

  • Uncertainty in Labeling: The authors acknowledge that even human experts can disagree on change detection, especially for ambiguous cases. This uncertainty, partly due to differing specifications between OSM and TOP10NL in the study region, is a limitation.
  • User Parameters: While the approach offers flexibility through user-defined parameters (like Twidth and Telong_ratio), this can lead to fitting or overfitting issues.
  • Machine Learning Potential: The authors suggest that machine learning could be a future direction to handle uncertainty and potentially automate parameter tuning. However, they note challenges in translating their complex, rule-based, and context-dependent analysis into a standard machine learning framework, especially regarding feature extraction for morphological analysis and the dynamic nature of rules.
  • Generalization to Other Datasets: Parameter values were empirically determined and might need adjustment for different datasets or regions.
  • Handling Scale Differences: While the study focused on similar scales, the paper notes that significant scale differences would further complicate change detection. They recommend applying change detection to datasets with the closest possible scales.

Summary updated

BibTeX
@article{2018change,
  author = {Xiang Zhang (张翔) and Xiaodong Zhou and Zhe Chen and Xiang Zhang and Tinghua Ai},
  title = {{Change Detection for Building Footprints with Different Levels of Detail Using Combined Shape and Pattern Analysis}},
  journal = {Isprs Int. J. geo inf},
  year = {2018},
  doi = {10.3390/ijgi7100406},
}
Figure