A Wavelet Transform Based Method for Road Centerline Extraction

This paper introduces a new wavelet transform based method of road centerline extraction from high resolution remote sensing images. In the one dimensional case, we characterize different kinds of sudden changes of signals by comparing the magnitudes of the local extreme values of the wavelet transforms under different dilation scales of the same wavelet. The platform-like signals, which come from cross sections of roads, can be characterized through the evolution of the wavelet transform across scales. A two-dimensional wavelet transform of an image consists of two components, each component is a one dimensional wavelet transform in one coordinate direction followed by a smoothing process in the perpendicular direction. The road edges can be characterized through the evolution of the two-dimensional wavelet transform across different scales. Edges of the main roads can then be extracted using these characterizations, and the road centerlines can be obtained by proper post-processing. Introduction Linear feature extraction for highways and road centerlines using remote sensing data is important in applications such as Geographic Information System (GIS) updating, image registration, and digital photogrammetry. With the increasing amount of source data, greater spatial detail, and increasing demand in GIS applications, automated spatial information extraction techniques such as improved road detection algorithms are urgently required to convert data into information (Wang and Zhang, 2000). In the past, efforts have been made to extract highway/ road information using lower resolution satellite images, such as Landsat TM, SPOT or IRS imagery. On these images, main roads appear as light lines or dark lines at one or two pixels in width. Therefore, the goal of the existing methods has been to detect lines or line following. Among the commonly used methods are dynamic programming and “snakes” (Merlet and Zerubia, 1996; Gruen and Li, 1997), an active testing model (Geman and Jedynak), neural network approach (Bhattacharya and Parui, 1997), and a sequence of algorithms (Karathanassi, et al., 1999). Satellite and airborne SAR images have been used to extract roads using various algorithms (Tupin, et al., 1998; Chanussot, et al., 1999; Toutin, 2001; Dell’Acqua and Gamba, 2001; Jeon, et al., 2002). A Wavelet Transform Based Method for Road Centerline Extraction Tieling Chen, Jinfei Wang, and Kaizhong Zhang A gradient direction profile analysis (GDPA) for line detection method was proposed by Wang and has been applied for road extraction from lower resolution remotely-sensed data (Wang, et al., 1992; Wang, 1993; Wang and Liu, 1994; Wang and Zhang, 2000). Although the GDPA algorithm performs relatively well for lower resolution satellite images (Landsat TM/ SPOT) in areas where there is sufficient contrast between roads and their background and the road network is not too dense to be detected from the lower resolution images, it is not applicable to very high resolution digitized air photos (approximately 1 m) because roads are no longer single pixel lines (Wang and Zhang, 2000). With the recent launch of the satellites which carry very high spatial resolution (such as 1–4 m Ikonos) sensors, much more detailed information, such as urban streets, can be obtained which would not have been possible with lower resolution data. These data have the potential to be used to maintain and update urban and regional GIS databases in an efficient and lower cost manner if the extraction can be automated with sufficient accuracy. This is especially useful in the hard to access places and rapidly changing areas. Some algorithms have been developed to recognize roads and other objects from high-resolution aerial imagery (Barzohar and Cooper, 1996; Gong and Wang, 1997; Baumgartner, et al., 1999; Laptev, et al., 2000; Auclair, et al., 2001; Katatzis, et al., 2001; Chen, et al., 2002). These algorithms may be modified and applied for feature extraction from the highresolution satellite data, since not much work has been reported for road extraction from very high-resolution satellite data and for the complex urban areas with dense road networks (Guindon, 2000; Couloigner and Ranchin, 2000). To date, automatic road detection techniques still cannot produce perfect road maps, especially in urban areas. Many road extraction tasks for database updating still have to be done manually. One of the problems with road detection from very high-resolution remote sensing images is that most edge/line detection methods extract both roads/streets and some extra or irrelevant features, such as driveways, sidewalks, and building boundaries. Another problem is that there are missing road segments because of the surrounding environment, such as tree-covered road segments. Wavelet transform has been used for edge detection (Mallat and Zhong, 1992; Merlet and Zerubia, 1996; Hsieh, et al., 1997; Aydin, et al., 1996), but not particularly for road extraction from satellite data. The basic requirement of a wavelet function is that the integral of the function over ( , ) is zero (Mallat and Zhong, 1992), although an admissibility condition is imposed in the multi-resolution analysis of wavelet transforms. The method of edge detection using P H OTO G R A M M E T R I C E N G I N E E R I N G & R E M OT E S E N S I N G December 2004 1 4 2 3 Tieling Chen is with the Mathematical Sciences Department, University of South Carolina Aiken, 471 University Parkway, Aiken, South Carolina 29801 ([email protected]). Jinfei Wang is with the Department of Geography, The University of Western Ontario, London, Ontario N6A 5B8, Canada ([email protected]). Kaizhong Zhang is with the Department of Computer Science, The University of Western Ontario, London, Ontario N6A 5B8, Canada ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 70, No. 12, December 2004, pp. 1423–1431. 0099-1112/04/7012–1423/$3.00/0 © 2004 American Society for Photogrammetry and Remote Sensing LFX-533.qxd 11/9/04 16:12 Page 1423