ROAD EXTRACTION FROM HIGH RESOLUTION SATELLITE IMAGE BY 
USING CIRCLE AREA 
Tsukasa HOSOMURA 
Department of Information and Arts, School of Science and Engineering, Tokyo Denki University 
- hosomura@ia.dendai.ac.jp 
Commission III, WG III/5 
KEY WORDS: Urban, Quickbird, Extraction, Algorithms, Feature, Accuracy, Experimental 
ABSTRACT: 
Many researchers conducted efforts for improving the accuracy of road extraction from satellite images. Many studies have used 
optical spectrum information of each pixel for road extraction. By applying this method for high resolution satellite image, some 
problems occurred. This situation is remarkable for urban area. Most of the case, we need the information about land use. In this 
study, we propose the method for detecting the road intersection by using a double circle. After detecting the road intersection 
position, we can find out the road network easy. This algorithm was applied for QuickBird image. 
1. INTRODUCTION 
Many researchers conducted efforts for improving the road 
extraction accuracy of satellite images. Most of the study has 
used optical spectrum information of each pixel for road 
extraction. By applying this method for high resolution satellite 
images, some cars or center lines on road are also extracted. 
This situation is remarkable in the urban area. Even if the land 
cove classification is carried out for many classes, color 
information of each car is not important in the classified result. 
Most of the case, we need the information that the object is road 
or not. 
In this study, we propose the method for detecting the roads by 
using the size and the shape. Double circle is introduced for 
detecting the size and the shape of road. 
2. CONCEPT OF DOUBLE CIRCLE 
The circle is designed according to the size of target roads. We 
will prepare a circle whose size is almost same as the size of 
target road intersection. If target road intersection is almost 
entered in the window, we can recognize the size of target road 
intersection. The problem in this case is how to decide whether 
the size of the target road intersection is almost same as that of 
the circle. This decision becomes easy if the situation of the 
target road intersection expansion to the outside of the circle. 
Therefore, large circle was set around the center circle. We call 
double circle for such structure. If this decision is carried out by 
using only the size of target area, other classes for example bare 
ground or grassland are also recognized as road intersection. In 
order to avoid such situation, training data is taken for most of 
the land cover except road. Supervised land cover classification 
is carried out for the center pixel of each circle. Inside circle and 
outside circle are called as core circle and peripheral circle 
respectively. 
This circle is designed according to the size of target road 
intersection. We will prepare a circle whose size is almost same 
as the size of target road intersection. If target road intersection 
is almost entered in the circle, we can recognize the size of 
target road intersection. The problem in this case is how to 
decide whether the size of the target road intersection is almost 
same as that of the circle. This decision becomes easy if the 
situation of the target road intersection expansion to the outside 
of the circle. Therefore, large circle was set around the circle. 
We call double circle for such structure. If this decision is 
carried out by using only the size of target area, other classes 
for example bare ground or grassland are also recognized as 
road intersection. In order to avoid such situation, training data 
is taken for most of the land cover. Supervised land cover 
classification is carried out for the center pixel of each circle. 
Inside circle and outside circle are called as core circle and 
peripheral circle respectively. 
Road intersection detection algorithm is as follows. 
(1) Supervised land cover classification is carried out for 
pixels in the core circle. 
(2) If classified results of some pixels in core circle are not 
road, the circle is moved. 
(3) If classified results of all pixels in core circle are road, the 
pixel whose value is almost same as that of central pixel 
and connected with central pixel is allocated “1”. Other 
pixels are allocated “0”. 
(4) Directional lines are drawn in all directions from the 
central pixel to the peripheral circle. 
(5) Run length of “1” on the directional line is calculated. 
(6) If many lines that run length is the same size of the radius 
of the peripheral circle are gathered in one group, this 
direction correspond to road. 
(7) If above line group exist at least 3, this central pixel is 
recognized as road intersection. 
Such procedure is carried out for each pixel by moving double 
circle. Central pixel of road intersection area satisfy above (7) 
condition. We can detect the central position of the road 
intersetion by using this algorithm. 
3. ROAD EXTRACTION ALGORITHM 
Double circle is introduced for detecting the road intersection. 
505