AUTOMATIC ROAD EXTRACTION FROM IRS SATELLITE IMAGES IN AGRICULTURAL AND DESERT AREAS Uwe BACHER and Helmut MAYER Institute for Photogrammetry and Cartography Bundeswehr University Munich D-85577 Neubiberg, Germany Email: {uwe.bacher, helmut.mayer}@unibw-muenchen.de Working Group III/4 KEY WORDS: Road Extraction, Fuzzy Logic, IRS, Vision Sciences, Automation ABSTRACT The appearance of roads in northern Africa differs from that of roads, e.g., in central Europe, which most of the approaches for automated road extraction in literature focus on. In this paper we propose a road model for areas with different road appearance in IRS satellite image data with a panchromatic resolution of 5 m and 20 m multispectral resolution. We model areas where water makes agriculture possible on one hand, and areas dominated by the desert and dry mountainous areas on the other hand. In the desert and mountainous areas paved roads appear as more or less distinct lines and the Steger line extraction algorithm can be used to extract roads in combination with global grouping. In mountainous areas detected, e.g., in a DEM, much larger curvatures are expected to occur than in the desert. In agricultural areas, on which we focus in this paper, roads often do not appear as distinct lines. Borders of the fields represented by edges in the image and the knowledge that these borders can be collinearly grouped, possibly together with lines, into longer linear structures are used to construct road sections. To close gaps, pairs of lines or edges are connected by ziplock snakes. To verify these road sections, the paths of the snakes are evaluated using the line strength and the gradient image. The verified road sections are finally globally grouped using the knowledge that roads construct a network between important points. Gaps which have a high impact on the network topology are closed if evidence supporting this is found in the image. Results show the validity of the approach. 1 INTRODUCTION For the road network in regions consisting in larger parts of desert or dry mountainous areas, e.g., in northern Africa, there is either no digital data available, or it is often very imprecise and not up to date, i.e., incomplete, or even wrong. Because of the large areas to be mapped, it is important to use highly automated means as well as cheap and readily available data. IRS-1C/D (Indian Remote Sensing Satellite) data with a ground resolution of about 5 m in the panchromatic and about 20 m in red, green, and NIR (near infrared) is a good choice for this. We use pan-sharpened images. The appearance of roads in these regions differs from that of roads, e.g., in central Europe, which most of the approaches for automated road extraction in literature focus on. In the following we give a short overview over related work, focusing on con- tributions which employ similar data or similar techniques, e.g., snakes, as our approach. One of the first approaches to automatic road extraction is (Fis- chler et al., 1981), where two types of operators are combined: the type I operator is very reliable but will not find all features of interest, whereas the type II operator extracts almost all fea- tures of interest, but with a large error rate. Starting with the reliable type I road parts, gaps are bridged based on the type II results employing a search algorithm termed F * . (Wiedemann et al., 1998) extract and evaluate road networks from MOMS- 2P satellite imagery with a resolution similar to IRS employing global grouping. The basis of this approach is the Steger line operator (Steger, 1998). The use of snakes for the detection of changes in road databases in SPOT and Landsat satellite imagery is demonstrated in (Klang, 1998). (P´ eteri and Ranchin, 2003) employ a multiresolution snake based on a wavelet transformed image to update urban roads based on given unprecise road data. In (Laptev et al., 2000) linear scale space and ziplock-snakes are used for the extraction of roads from high resolution aerial im- agery. (Dal Poz and do Vale, 2003) propose a semi-automated approach for the extraction of roads from medium and high reso- lution images based on dynamic programming. Active testing for the tracking of roads in satellite images is introduced by (Geman and Jedynak, 1996). A semi-automated system for road extrac- tion based on dynamic programming and least squares B-spline (LSB)-snakes is proposed by (Gr¨ un and Li, 1997). The automatic completion of road networks based on the generation and verifi- cation of link hypotheses given in (Wiedemann and Ebner, 2000). (Wallace et al., 2001) present an approach designed for a wide va- riety of imagery. It is based on an object-oriented database which allows the modeling and utilization of relations between roads as well as other objects. Road extraction using statistical model- ing in the form of point processes and Reversible Jump Markov Chain Monte Carlo is proposed by (Stoica et al., 2004). Our approach makes use of the 5 m panchromatic resolution as well as the multi spectral information of IRS. It is designed for the extraction of roads in mostly agricultural as well as in arid areas, the latter also comprising mountainous regions. Section 2 describes model and strategy. In Section 3 the individual steps of the extraction process, namely line / edge extraction, generation of connection hypotheses, verification of connection hypotheses, and global grouping are detailed. Section 4 presents experimen- tal results showing the validity of the approach. An outlook con- cludes the paper.