ibul 2004 
  
SEMI-AUTOMATED MAP OBJECT EXTRACTION FROM IM RESOLUTION SPACE IMAGES 
Taejung Kim , Young-Jae Lim", Soo Jeong®, Kyung-Ok Kim” 
Department of Geoinformatic Engineering, Inha University, 253 Yonghyunn-Dong, Nam-Gu, Incheon, ROK, tezid@inha.ac.kr 
* Telemetics Research Division, Electronics and Telecommunications Research Institute, 161 Gajeong-Dong, Yuseong-Gu, Daejeon, 
ROK, (yjlim, soo, kokim)(etri.re.kr 
KEY WORDS : Building Detection, Road Extraction, Least Squares Matching, IKONOS 
ABSTRACT: 
Extraction of map objects such as roads, railroads, rivers and building boundaries from 1m resolution space images is one of the 
important research issue. Automation of this task is crucial for the success of the application but full and reliable automation is yet to 
be achieved. This paper describes the development of algorithms to extract two major map objects, roads and buildings. We adopt 
the *semi-automatic" approach for reliability and efficiency. For road extraction, we designed a new least squares template matching 
algorithm. For buildings, we combined line analysis and template matching for semi-automatic extraction. Our algorithms were 
tested with IKONOS images over a very dense urban scene. The algorithms developed showed promising results. The major 
contribution of this paper is the development of monoscopic algorithms little human intervention that produces a fair amount of 
information. 
1. INTRODUCTION 
Since the first commercial earth observing satellite with better 
than 1m spatial resolution has been launched in 1999, a number 
of follow-on satellites are already in orbit or in the process of 
development. For example, the Korea Aerospace Research 
Institute is developing the second satellite of its earth 
observation satellite series, KOMPSAT-2, with Im resolution 
imaging capability and the planned launch year of 2005. These 
satellites are offering spaceborne images with high quality for 
various applications on ground. 
Among many technologies needed to handle Im resolution 
images, automated extraction of map objects from images 
seems one of the most essential and urgent ones. This 
technology is required not only for mapping but also for urban 
planning, environmental research, logistics, and etc. Currently, 
the task of map object extraction is done manually through 
head-up digitization. This process consumes too many 
resources and causes the usage of spaceborne images not very 
cost effective in real applications. 
Therefore, research should be carried out to extract map objects 
such as roads, rivers, building boundaries from 1m resolution 
spaceborne images with an automated manner. This paper will 
report techniques developed for extracting two map objects, 
roads and building boundaries. While fully automated 
techniques are ideal, they often have accuracy and reliability 
problems. Instead, we focus on semi-automated approach. The 
work reported here is done as a part of operational SW 
development work. We design the extraction process such that 
users will guide and interact with automated algorithms for 
better results. 
The first part of this paper is about the extraction of roads. So 
far, various methods have been proposed for this theme, 
including perceptual grouping, (Trinder and Wang, 1998; 
Katartzis et al, 2001), scale-space approaches (Mayer and 
Steger, 1998), neural network and classification (Doucette et al., 
2001), “snakes” or energy minimization (Gruen and Li, 1997), 
and template matching (Vosselman and Knecht, 1995; Gruen ar 
al, 1995; Hu et al, 2000). Although many authors have 
focused on the development of fully automated algorithms, it 
  
* 7" . 
Corresponding Author 
seems that semi-automated algorithms such as “snakes” and 
template matching seemed to gain acknowledgement. 
In this paper we introduce a new semi-automatic road 
extraction algorithm based on template matching. Our work 
was motivated by the previous work of Gruen ef al. (1995) but 
our algorithm differs from the previous work in the following 
ways: we focused on tracking road centerlines from high 
resolution images while the previous work focused on tracking 
roads from mid or low resolution images; and we eliminated the 
need to have additional constraints for match guidance by 
designing a new least squares correlation matching scheme. The 
next section will describe this scheme step-by-step. 
The second part of this paper is about the extraction of building 
boundaries. Previous approaches for this task includes line 
analysis and perceptual grouping (Shufelt and McKeown, 1993; 
Kim and Muller, 1999), the use of shadow information and 
perspective geometry (Huertas and Nevatia, 1988). To improve 
the quality of building extraction, several approaches used 3D 
information. Shufelt and McKeown (1993) and Kim and Muller 
(1998) combined stereo matching and line analysis for building 
extraction. Cochran and Medioni (1992) and Kim and Muller 
(1996) tried to use building detection results to improve stereo 
matching process. Some authors assumed 3D information on 
buildings was available and used this information for building 
extraction (Baltsavias et al., 1996). 
In this paper, we will focus on extraction of relatively large and 
rectangular shaped buildings such as apartment or industrial 
buildings. We will propose a monoscopic algorithm, which 
extracts buildings from a single image without any additional 
information. For better results, we may assume other data 
sources such as 3D building heights obtained from a LIDAR 
sensor or building footprints from digital maps. However, we 
intent to maintain with this “image only” approach to find out 
the maximum amount of information we can retrieve from one 
single image only. We, however, decided to include manual 
interaction for the process. We realize the building extraction 
by line analysis and least squares template matching with a 
manually-given input point. Section 3 will describe this process 
step-by-step.