Extracting Road Networks from Natural Terrain for Visualization Edward Riseman, Xiaoguang Wang Brian Hanechak, Howard Schultz, Allen Hanson {riseman, xwang, hanechak, hschultz, hanson}@cs.umass.edu University of Massachusetts Computer Vision Laboratory Amherst MA 10003 USA Abstract The ability to effectively generate and assimilate terrain information is central to many computer vision (CV) and computer graphics (CG) applications. In particular, effective interactive visualization applications require techniques for generating a digital elevation map (DEM) and ortho-image, classifying ground cover, and extracting symbolic information. In this paper we describe a program underway at the University of Massachusetts for generating complex geospatial databases using photogrammetric techniques and context-sensitive strategies. We present a specific example in which the input are two monochromatic aerial images of a rural area of Fort Hood, Texas, and the results are a DEM, ortho-image, classified ground cover and a symbolic representation of a road network. 1. Introduction Effectively generating and applying terrain information is crucial for both civilian and military applications. Advances in computer vision and graphics, including sensors, sensor platforms, computing platforms, and storage capacity, can facilitate the automated classification of terrain and the extraction of features for both natural terrain and cultural sites. This paper focuses on a subset of work that forms part of a major DARPA (Defense Advanced Research Projects Agency) program encompassing a wide range of image understanding techniques. Our emphasis is on providing more accurate three-dimensional reconstruction, extracting symbolic information, and enhanced visualization techniques. To demonstrate our approach, we analyzed a rural scene containing trees, bare ground, a dry river bed, a partially occluded rural road network, and a bridge. The goals of this investigation were to determine the three-dimensional structure of the terrain, the type of ground cover, and to extract the road network. We obtained a semantic interpretation of aerial images to classify the terrain, and then applied context-sensitive strategies to extract the road