SEMI-AUTOMATIC EXTRACTION OF DIFFERENT-SHAPED ROAD CENTERLINES 
FROM MS AND PAN-SHARPED IKONOS IMAGES 
F. Ameri, A. M. Mobaraki, M. J. Valadan Zoej 
Faculty of Geodesy and Geomatics Engineering, K.N.Toosi.University of Technology, Vali-e-asr St., Mirdamad 
Cross, Tehran, Iran, P.C. 1996715433 - (ameri fa, mobaraki_am)@yahoo.com, valadanzouj@kntu.ac.ir 
KEY WORDS: Image processing, Road extraction, Space photogrammetry, Clustering, Updating 
ABSTRACT: 
In this work we develop semi-automatic road extraction system for updating and storage road network data bases. Combination of 
some of the existing road extraction techniques such as spectral and spatial data clustering, morphological functions and graph 
theory is used in this proposed system. 
Input data of the proposed road extraction system are multi-spectral and pan-sharpened IKONOS images of Lavasan city in Iran 
(with respectively 4 and 1 meters spatial resolution). The proposed system investigates capability and amount of system success in 
extraction of different shaped roads such as straight, spiral, junction and square. 
In the proposed method, primarily the input image is spectrally classified by use of Fuzzy C-Means (FCM) clustering technique and 
road class binary image is obtained by definition of threshold value. Afterwards, quality of detected road features is improved using 
morphological operators like dilation, erosion, opening, closing, bridge and etc. Our approach proceeds by performing spatial cluster 
analysis using C-Means technique and hence, road centerline nodes are attained. Finally by use of graph theory and minimum 
spanning tree (MST) and defining an appropriate cost function, these key points are connected and vector road centerline is 
obtained. 
The only drawback of this system is limitation in completely extraction of road centerline in place of squares and closed loops. 
Attaining mean overall accuracy (OA) of 98.2% and Kappa coefficient of 86.26% in classification of image to road and non-road 
classes, and also mean RMS error of 0.64 pixel in comparing automatic extracted road centerline with manual extracted one, are a 
good criterion of proposed system success in semi-automatic extraction of roads. 
1. INTRODUCTION 
Attaining geospatial information is a challenging issue for many 
scientists and experts in the field of spatial decision making. 
Feature extraction from aerial or satellite images is a 
supplemental technology to achieve these kinds of information 
and is able to facilitate image analysis and interpretation and 
updating existing databases which are one of the urgent 
requirements of each organization. At present, information 
extraction from images is performed mostly manually, and thus 
time and cost intensive. To overcome this drawback, automatic 
methods for extraction of information and analysis of image 
content are needed. 
On the other hand, the advent of high resolution optical satellite 
imagery such as IKONOS or Quickbird, creates new 
possibilities for the extraction of linear features such as roads. 
The advantages of this digital data compared to aerial imagery 
are the worldwide availability and higher radiometric resolution 
in fewer spectral bands. The worldwide availability of the 
satellite digital data makes it possible to produce topographic 
databases for nearly any region of the earth and for extensive 
applications such as military purposes, disaster managements, 
crop and weather predictions, traffic managements and etc. 
Although the geometric resolution of HR satellite images are 
often worse than that of the aerial images, but for the purpose of 
road extraction and updating spatial data bases is generally 
sufficient. 
The most prominent linear topographic features to be extracted 
from satellite images are roads. Roads are important large 
networks that seem to be the most vital transportation arteries of 
each country which extraction of them has been one of popular 
research areas in computer vision, remote sensing, 
photogrammetry and GIS communities. Therefore many efforts 
have been performed to extract them from digital images. 
An inclusive review on the proposed methods of road extraction 
could be found in (Mena, 2003). In (Long and Zhao, 2005) a 
new integrated system for automatic extraction of main roads 
from high resolution satellite images was presented which used 
multi-scale gray level morphological cleaning algorithm to 
reduce road gray level deviation and the binary image was 
analyzed by morphological operators and convex hull 
algorithms. 
The research presented in (Wang et al., 2005), showed that by 
utilizing the texture, edge, shape and size properties, non-road 
pixels could be removed in order to improve road detection 
results. 
The idea of geometrical and topological analysis of high 
resolution binary images for automatic vectorization of 
segmented road networks was presented in (Mena, 2006). 
(Mohammadzadeh et al., 2006) proposed a new fuzzy 
segmentation method for road detection in high resolution 
satellite images that needed only a few number of road samples. 
Then using an advanced mathematical morphological operator, 
road centerlines were extracted. Also, in (Zhang and 
Couloigner, 2006), the effectiveness of angular texture 
signature was evaluated to discriminate between parking lots 
and roads using high resolution satellite images. 
According to necessity and importance of using accurate and 
update spatial information of road network for traffic 
management, automatic vehicle navigation, natural hazard 
assessment and critical decision making, using automatic, 
621