Automatic Extraction of Main Road Centerlines from High Resolution Satellite Imagery Using Hierarchical Grouping

Automatic road centerline extraction from high-resolution satellite imagery has gained considerable interest recently due to the increasing availability of commercial high-resolution satellite images. In this paper, a hierarchical grouping strategy is proposed to automatically extract main road centerlines from high-resolution satellite imagery. Here hierarchical grouping means that, instead of grouping all segments at once, the selective segments are grouped gradually, and multiple clues are closely integrated into the procedure. By this means, the computational cost can be reduced significantly. Through the stepwise grouping, the detected fragmented line segments eventually form the long main road lines. The proposed method has been tested and validated using several Ikonos and QuickBird images both in open areas and build-up urban environments. The results demonstrate its robustness and viability on extracting salient main road centerlines. Introduction The demand for road database generation and updating from high-resolution satellite imagery is increasing dramatically due to its high spatial resolution (1 to 4 m), fast orbit repeatability, rich multi-spectrum information and stable, affordable acquisition cost. Such a promising end-user market drives the need for automating road information extraction from high-resolution satellite imagery. Various linear feature extraction methods have been developed by the photogrammetry, remote sensing, and computer vision communities in the past decades. Most of them were developed mainly for general linear feature extraction, while some are particularly designed for road extraction. In general, these methods can be classified into four categories: 1. Linear feature detection methods: Roads appear as ridges or valleys of grey value function of an image. Road finding can be considered as a process consisting of ridge finding, ridge points linking, and road segments forming. Image filtering or image convolution using edge or ridge templates is a classic approach (Nevatia and Babu, 1980). Based on the analysis of Automatic Extraction of Main Road Centerlines from High Resolution Satellite Imagery Using Hierarchical Grouping Xiangyun Hu and Vincent Tao the scale-space behavior of a line profile, Steger (1998) and Mayer and Steger (1998) presented an unbiased linear feature detector to detect road lines and their width. The road lines have different lateral contrast. Vosselman and de Knecht (1995) used a profile matching method derived from the road radiometric profile for road extraction. 2. Rule or knowledge-based methods: This is so called ‘highlevel’ processing compared to pixel based low-level and intermediate level processing. In order to handle the issues of linear feature alignment and fragmentation, heuristic or rule-based methods were developed by Trinder and Wang (1998). Tönjes and Growe (1998) employed a semantic-net technique to combine information from multiple sensors for road extraction. 3. Optimization-based methods: In this category, road extraction is treated as an optimization problem. Dynamic programming is a popular method. It is applied to formulate the linkage of candidate points (Gruen and Li, 1995). Snakes (active contour) and least square template matching have also been applied to automatic (Laptev et al., 2000) or semiautomatic road extraction from mono and stereo images (Gruen and Li 1997; Hu et al., 2004), and image sequences (Tao et al., 1998). 4. Content-supported and map guided methods: Using image features alone, a road network can hardly be extracted completely and correctly. Contextual information, such as buildings, trees, rivers, is obviously valuable for road identification and road segment linking. Baumgartner et al. (1999) developed a method for road extraction from multiscale images and discussed the role of grouping using contextual information. Map guided feature extraction is also considered in this category. Existing maps can provide important clues for road finding. Some algorithms have been developed to perform image-map matching for road identification. Fiset et al. (1998) used a multi-layer perceptron to extract road segments from SPOT-HRV panchromatic images. Stilla (1995) described a method to find new roads based on the assumption that new roads are always connected to the old ones. Agouris et al. (2001) proposed a method that detects the changed road segments using differential snakes for image-based GIS updating. It has been realized that robust feature extraction is largely dependent on the strategy developed and the constraints employed (Tao et al., 2001). Many factors such as the objective, image scale or resolution, image quality, and image complexity need to be taken into account. We have noticed that more and more methodologies are based on PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Sep t embe r 2007 1049 Xiangyun Hu is with Leica Geosystems Geospatial Imaging, LLC, USA, and was formerly with the Department of Earth and Space Science and Engineering, York University, 4700 Keele Street, Toronto, ON, Canada M3J 1P3 and the Department of Photogrammetry and Remote Sensing, Wuhan University, China. ([email protected]). Vincent Tao is with Microsoft, Corporation, USA, and formerly with the Department of Earth and Space Science and Engineering, York University. Photogrammetric Engineering & Remote Sensing Vol. 73, No. 9, September 2007, pp. 1049–1056. 0099-1112/07/7309–1049/$3.00/0 © 2007 American Society for Photogrammetry and Remote Sensing 05-080.qxd 8/14/07 9:00 AM Page 1049