The Automatic Extraction of Road Information by the Technology of Object-oriented

Ping Wang, Zhai Zhaokun, Hui Fu, Zhao ZhilingThe Open Automation and Control Systems Journal 2015 · DOI

TL;DR

This paper demonstrates that object-oriented classification significantly outperforms traditional pixel-based methods for extracting road networks from high-resolution satellite imagery. By grouping pixels into "objects" and analyzing their shape and texture rather than just their color, the researchers achieved a 92.94% accuracy rate—a 10% improvement over standard techniques.

What Problem Does This Paper Address?

As satellite imagery reaches higher spatial resolutions, you might expect road mapping to become easier. In reality, it often becomes harder for traditional algorithms. Traditional "pixel-based" classification looks at each pixel in isolation based on its spectral signature (its color).

This approach faces two major hurdles in complex environments like mining areas and urban districts:

  1. The "Same Spectrum, Different Object" Problem: A concrete road might have the exact same spectral signature as a building's roof or a paved parking lot, leading to "fragmented" results.
  2. The "Salt and Pepper" Effect: High-res images capture tiny details like shadows or individual trees overhanging a road. A pixel-based algorithm might classify these as "forest" or "shadow," breaking a continuous road into a series of disconnected dots.

The authors aim to bridge this gap by moving from pixel-level analysis to Object-Oriented Technology, which treats groups of related pixels as single entities with geometric properties.

Methods

The researchers employed a multi-stage workflow centered on the eCognition software platform. The process involves:

  • Image Fusion: They used Gram-Schmidt Pan sharpening to fuse multi-spectral data with high-resolution panchromatic (black and white) data, ensuring the final images had both rich color and sharp detail.
  • Multi-scale Segmentation: This is the core of the object-oriented approach. Instead of analyzing pixels, the algorithm groups neighboring pixels into "objects" based on spectral and shape heterogeneity.
  • Feature Space Construction: Unlike traditional methods that only use color (spectral) data, this method builds a "feature space" including:
  • Spectral Features: Mean intensity and standard deviation of bands.
  • Geometric Features: Length-to-width ratio (crucial for long, thin roads), compactness, and density.
  • Index Features: Vegetation and water indices to filter out non-road objects.
  • Classification Strategies: The authors tested two specific object-oriented paths: a Decision Tree approach (using specific thresholds for features like "Border Index") and a Multi-Feature approach that optimizes the feature space to distinguish roads from buildings.

Datasets

The study utilized two primary high-resolution satellite datasets:

  1. SPOT-5 Imagery: Focused on the Feicheng mining area in Shandong Province (June 2011). This area was chosen for its mix of wide and narrow roads in a complex industrial landscape.
  2. GaoFen-1 Imagery: Focused on an urban district in Chengde City, Hebei Province (September 2013), where road information is often obscured by dense buildings.

Key Insights and Ideas

  • Geometry is King: The most effective way to identify a road isn't by its color, but by its "long-ness." By using the length-width ratio and compactness as primary filters, the algorithm can easily distinguish a gray road from a gray square building.
  • Context Matters: The "Border Index" (measuring how jagged an object is) helps the system understand that roads generally have smooth, continuous edges compared to the more irregular shapes of natural features or "salt and pepper" noise.
  • Overcoming Heterogeneity: In mining areas, roads are often "dirty" or made of different materials. Because object-oriented methods look at the average properties of a segment, they are much more resilient to these local variations than pixel-by-pixel methods.

Key Results

The object-oriented approach proved superior across all metrics compared to traditional supervised pixel-based classification:

  • Accuracy Boost: The overall accuracy reached 92.94%, representing a 10% improvement over traditional methods.
  • Reliability: The Kappa coefficient (a measure of how much better the classification is than random chance) was 0.9056, indicating excellent performance.
  • Qualitative Improvement: The extracted road networks were significantly more continuous and consistent with actual ground conditions, successfully identifying narrow roads that pixel-based methods typically missed or fragmented.

Limitations and Future Work

Despite the success, the authors highlight a significant bottleneck: Segmentation Scale. The quality of the results depends heavily on choosing the right "scale" for grouping pixels into objects. Currently, this requires repeated manual experiments and trial-and-error to find the "reasonable" scale for a specific landscape.

Future research is needed to develop more automated, algorithmic ways to determine the optimal segmentation scale and to further refine the feature parameters to handle even more complex urban "canyons" where shadows and high-rise buildings create extreme occlusions.

Summary updated

Abstract

Road information is rich in urban districts and mining areas. With the increase in spatial resolution and spectral resolution of remote sensing data, it is possible to extract information of narrow roads. However, the traditional manual extraction method using high-spatial-resolution data has shortcomings of low accuracy and low efficiency. Considering the features of SPOT-5 and Gaofen-1 data and the actual road situation of the study areas, the object-oriented method is used in this paper. The main advantage of this method is overcoming the limit of road extraction only using spectral information. A comparison with the results of the traditional supervised classification based on pixels proves that the objectoriented method improves the accuracy by 10% and provides better road information results. In addition, it is a more effective method to extract information for geographical condition monitoring.
BibTeX
@article{wang2015automatic,
  author = {Ping Wang and Zhai Zhaokun and Hui Fu and Zhao Zhiling},
  title = {{The Automatic Extraction of Road Information by the Technology of Object-oriented}},
  journal = {The Open Automation and Control Systems Journal},
  year = {2015},
  doi = {10.2174/1874444301507011842},
}
Figure