A NEW ALGORITHM FOR AUTOMATIC ROAD NETWORK EXTRACTION IN MULTISPECTRAL SATELLITE IMAGES E. Karaman, U. Çinar , E. Gedik , Y. Yardımcı, U. Halıcı METU, Ankara, Turkey – [ersin, ucinar, ekin, yardimy, halici]@ii.metu.edu.tr KEY WORDS: Road Extraction, Multispectral Satellite Images, Edge Detection, Structural Analysis ABSTRACT: The aim of this study is to develop automatic road extraction algorithm in satellite images. As roads have different width and surface material characteristics in urban and rural areas, a modular approach for road extraction algorithm is desired. In this study, edge detection, segmentation, clustering and vegetation and land cover analyses are used. In order to combine the results of different methods, a score map based on segmentation analysis is constructed. Quantitative and visual results show that this method is successful in road extraction from satellite images. 1. INTRODUCTION The importance of satellite systems has constantly been rising in response to the increasing demand on military decision making, urban planning, traffic regulation, emergency management, crop estimation and so on. Roads are important components for these applications. The purpose of this study is to use multi-spectral high resolution satellite images for automatic road extraction to get accurate and detailed road map. A large number of automatic and semi-automatic approaches have been developed for the analysis of transportation infrastructure in satellite images. Canny edge detection, Hough transform, morphological operations are some of the techniques used in road detection studies (Zhao et. al, 2002; Wang et. al, 2005; Vandana et. al.,2002; Zhang et. al., 1999). In the study by Zhao et. al, 2002, the road mask is extracted by using commercial remote sensing software. This binary mask shows possible road pixels. Then, Canny filter is applied on the image for detection of road edges. By tracing edge pixels, sudden and fast change points are determined and edge map is broken. In addition, edge lines which have similar direction and small gaps are merged. Some assumptions about the consistency of road features have been made. These are: road width, road direction and local average gray vary slowly, the contrast between road and background pixels is likely to be large, and roads are often long. This study uses 4-band (red, green, blue, and NIR) and 1 meter resolution satellite images obtained from IKONOS. In the study conducted by Wang et. al, 2005, in addition to Canny filter, some morphological operations and Hough transform have been used to enhance road extraction results. The employed road characteristics are that the width of a road and its curvature varies slowly; the texture enclosed by the road edges is rather homogeneous and roads create a connected network. IKONOS panchromatic 1 m resolution image was used in the study. Multi-resolution approach has been used to extract road sides and the centerline. Then, results obtained from different resolutions are combined. It is reported that this method is more suitable for main road detection. Moreover, there is manual post-processing to handle incomplete road segments. In another semi-automatic approach (Vandana et. al.,2002), edge detection is used at pre-processing step, then seed points provided by the user are used to complete the roads by using path following considering the variance from the mean of the segment, potential directions, length and width on 1-m resolution IKONOS and aerial images.In the study by (Zhang et. al., 1999), mathematical morphology was applied on 1-m resolution satellite images to find roads. It is assumed that roads are areas rather than lines because of high resolution, and road networks forms elongated areas. After classification (gray level analysis), segmentation and size distribution analysis, trivial opening, hole filling, removing small paths and closing methods have been applied respectively for road extraction. It is reported in the study that this method cannot handle occluded road areas by tree shadows, cars and other objects, and cannot remove houses connected to the roads and shorter road parts because of image frames. Landsat images with 25 m spatial resolution and 7 spectral channels have been used to extract the road network. Red band has been used to find the roads, and then watershed transform has been applied. In addition, curve adjacency graph has been constructed on watershed result. Lastly, Markov random fields have been used for road network extraction from this graph (Géraud & Mouret, 2004). On the other hand, knowledge base methods have been used to extract roads from satellite images. In the study of (Lee et. al., 2000), 1 meter, IKONOS images have been used. The method consists of two steps. Firstly, road primitives were extracted using an intensity based segmentation approach called hierarchical gradient watershed algorithm. This algorithm was modified to avoid over segmentation problem. With the assumption that roads are elongated and large objects with constant intensity and have high contrast with their surroundings, road segments obtained by intensity based segmentation were expected to be elongated and large. So, road segments can be selected by analysis of mean gray value, size (number of pixels in segment) and shape information (major/minor axis). Other studies generally focused on supervised techniques (Amini et. al., 2002), frequency based techniques (Hu et. al, 2007), and segmentation based methods (Géraud & Mouret, 2004; Lee et. al., 2000). Moreover, K-means clustering [Zhang et. al., 19990), Fuzzy c-means (Kim et. al., 2004), genetic algorithms, Markov Random Fields (MRF) (Lee et. al, 2000), Template Matching (Kim et. al., 2004) are some other techniques used in road extraction from satellite images studies. Proceedings of the 4th GEOBIA, May 7-9, 2012 - Rio de Janeiro - Brazil. p.455 455