AUTOMATIC LINEAR FEATURE EXTRACTION OF IRANIAN ROADS FROM 
HIGH RESOLUTION MULTI-SPECTRAL SATELLITE IMAGERY 
A. Mohammadzadeh**, A. Tavakoli”, M. J. Valadan Zoej* 
* Geodesy and Geomatics Engineering Faculty, K.N.Toosi University of Technology, No. 1346, Vali Asr St., Tehran, 
Iran, Postal Code: 1996715433 - ali_ mohammadzadeh2002@yahoo.com - valadanzouj@kntu.ac.ir 
° Dept. of Electrical Engineering, Amirkabir University of Technology - tavakoli@aut.ac.ir 
KEY WORDS: Extraction, Fuzzy Logic, IKONOS, GIS, High resolution, Change Detection 
ABSTRACT: 
Attaining geospatial information is a challenge for many scientific practitioners. Such information is a necessary tool for spatial 
decision making. Remote Sensing (RS) is the leading art/science providing the data for many global or local applications such as: 
green house effect, pollution, military, urban and land use. Graphical elements of geospatial information can be divided into: points, 
lines, and planimetric features. The most prominent linear topographic features are roads, rivers, railways and vegetation boundaries. 
Roads are important large-network man-made structures. All the elements can be derived from RS images. Many efforts have been 
performed to extract proper information efficiently. They can be classified into: manual and automatic feature extraction. Manual 
techniques are fading away as they are inefficient and inaccurate. Automatic extraction of geospatial phenomena has been the 
subject of extensive research for the past decade. Feature extraction approaches are diverse especially for linear features whose 
major methodologies are: fusion-based, fuzzy-based, mathematical morphology, model-based approach, dynamic programming and 
multi-scale grouping. In this paper, an approach based on fuzzy and mathematical morphology is introduced. In the developed fuzzy 
process, each pixel is transformed into a matrix of membership degrees representing the fuzzy inputs. A minimum-reasoning rule is, 
then, applied to infer the fuzzy outputs. Finally, a defuzzification step is applied to extract features. Advanced morphological 
concepts: “trivial opening’, ‘granulometry’ and ‘skeleton’ are applied to remove small objects, narrow paths and noises 
automatically. In addition, shadows of trees and buildings that cause partially covered roads are recovered. This object of this paper 
is to illustrate an applied method for automatic extraction of Iranian roads from pan-sharpened IKONOS images. The method is 
successfully executed on different regions including urban, suburban and rural areas. It is concluded that extracted road centrelines 
are so accurate and precise. The results are more promising at the crosses and curved segments. The extracted road centreline is 
easily inserted in a G/S. For future work, authors intend to introduce the fuzzy classification method into an artificial neural network 
program. In addition, the illustrated method can be used for the purpose of change detection in the road network system of a city 
from high resolution satellite images. 
1. INTRODUCTION 
For the development of a large number of countries topographic 
mapping from space must be regarded as a necessity (Konecny 
and Schiewe, 1996). Today many suitable and operational 2.1 Stage 1: Developed Fuzzy System 
sensors exhibiting various spatial, spectral and temporal 
2. METHODOLOGY 
resolutions and continuously delivering raw imagery are in In the first step, we have developed a fuzzy method used by 
orbit, and more are to come. Thus, the time and cost intensive Melgani et al. (2000). In this approach multispectral remote 
manual procedure necessary for turning these images into sensing images are segmented in two classes: road and non- 
useful geographic information constitutes the main bottleneck, road. By sampling from road surface, mean value of the road in 
Which needs to be overcome. The solution is an increase in each band will be obtained. We have defined 5 membership 
automation in order to improve the efficiency of satellite functions (MFs) with special means and standard deviations in 
topographic mapping. The major types of linear topographic each band. Then, a fuzzification step is applied for obtaining an 
objects are roads, rivers, railways and vegetation boundaries. estimation of the class contributions in each band assuming a 
The major linear features of interest right now are roads. Gaussian distribution of the classes. After a MIN and a MAX 
Automatic extraction of reads from digital images has been an operation on these fuzzy inputs results to the fuzzy 
active research subject for a decade. This field is quite young classification of the scene. Then a hard classification can be 
and the major approaches are not settled. Research in feature deduced in the defuzzification step in order to achieve to the 
extraction is still very diverse and object extraction is a segmented image. This method is successfully applied on 
fundamental computer vision operator. IKONOS  Pan-sharpened image. Figure 1 shows defined 
There are different methodologies for feature extraction, hypothetic MFs for bands one and two. There are 5*5=25 
especially for linear features such as image fusion for feature classes for two bands and 5*5*5=125 classes for three bands. 
extraction (Pigeon et al., 2001), fuzzy-based approach (Agouris Figure 2 shows implementation flowchart used in the fuzzy 
et al., 1998), mathematical morphology (Zhang, 1999), model- step. 
based approach (Buckner, 1998), dynamic programming (Gruen 
et al., 1995), multi-scale grouping and context (Mayer et al., 
1997), and kalman filtering (Vosselman et al., 1995). 
764 
I 
th 
cl 
ba 
ap 
Wi 
thi 
th: 
pb 
the