Content uploaded by David BL Bong
Author content
All content in this area was uploaded by David BL Bong on May 31, 2015
Content may be subject to copyright.
Automatic Road Network Recognition and
Extraction for Urban Planning
David B.L. Bong, Koon Chun Lai, and Annie Joseph
Abstract—The uses of road map in daily activities are numerous
but it is a hassle to construct and update a road map whenever there
are changes. In Universiti Malaysia Sarawak, research on Automatic
Road Extraction (ARE) was explored to solve the difficulties in
updating road map. The research started with using Satellite Image
(SI), or in short, the ARE-SI project. A Hybrid Simple Colour Space
Segmentation and Edge Detection (Hybrid SCSS-EDGE) algorithm
was developed to extract roads automatically from satellite-taken
images. In order to extract the road network accurately, the satel-
lite image must be analyzed prior to the extraction process. The
characteristics of these elements are analyzed and consequently the
relationships among them are determined. In this study, the road
regions are extracted based on colour space elements and edge details
of roads. Besides, edge detection method is applied to further filter
out the non-road regions. The extracted road regions are validated by
using a segmentation method. These results are valuable for building
road map and detecting the changes of the existing road database.
The proposed Hybrid Simple Colour Space Segmentation and Edge
Detection (Hybrid SCSS-EDGE) algorithm can perform the tasks
fully automatic, where the user only needs to input a high-resolution
satellite image and wait for the result. Moreover, this system can
work on complex road network and generate the extraction result in
seconds.
Keywords—Road Network Recognition, Colour Space, Edge De-
tection, Urban Planning.
I. INTRODUCTION
I
N recent years, the road network changes at a rapid rate be-
cause of urban development. Thus, it is hard to maintain the
accuracy and precision of the road network. In order to execute
massive applications such as city planning and management
system, automatic extraction of roads for updating city map
has recently come to a popular research topic [1]. Research
in Automatic Road Extraction using Satellite Images (ARE-
SI) has been carried out in Universiti Malaysia Sarawak. Road
extraction strategies can be divided into two categories namely
semi-automatic extraction and automatic extraction. In semi-
automatic road extraction, all the initial points or road seeds
need to be provided. However, in fully automatic extraction,
the road seeds can be detected automatically and linked to
complete the road network [2]. In generally, automatic way is
more preferable than the manual operations that acquire lots
of manpower and time in mapping the road network [3].
There are many researchers interested in this topic. Some
of them propose the use of Hough Transform to detect road
David B.L. Bong is with the Faculty of Engineering, Univer-
siti Malaysia Sarawak, 94300 Kota Samarahan, Malaysia. e-mail: bbl-
david@feng.unimas.my.
Koon Chun Lai is with Universiti Tunku Abdul Rahman, Perak, Malaysia.
Annie Joseph is with the Faculty of Engineering, Universiti Malaysia
Sarawak, 94300 Kota Samarahan, Malaysia. e-mail: jannie@feng.unimas.my.
Manuscript received April 21, 2009; revised June 19, 2009.
lines and implementing ”snake” method to reconnect the
broken road lines [4]. However, Hough Transform has low
efficiency in detecting curve road, and ”snake” requires user to
select initial seed points to start, turning its system into semi-
automatic which is more time-consuming. Furthermore, these
two methods are not applicable for complex road network
involving complicated curve lines.
There are alternative ways suggested by researchers. How-
ever, they only manage to extract simple road networks [5]. On
the other hand, there exists some methods that can work for
complex road network, but unfortunately, they always utilize
long and complicated procedures before the result is gained
[6].
In this project, a new algorithm based on hybrid sim-
ple colour space segmentation and edge detection (Hybrid
SCSS-EDGE) is proposed for the road extraction system to
solve the stated problems. It performs the extractions fully
automatic and properly works on complex road networks
with satisfactory results generated within seconds. The key
techniques used for the algorithm are segmentation using
edge detection together with colour space components (i.e.
luminance, saturation and hue values) of the satellite imagery.
Satellite images from Google Earth are considered because of
its wide availability and less deployment cost for this project.
II. H
YPOTHESES OF ROA D FEATURES
Several important hypotheses have been made for the road
extraction process, as mentioned in [7], [8], [9] and [10].
a) Road width variance is small and road width change is
likely to be slow.
b) Road direction changes are likely to be slow.
c) Road local average grey level is likely to vary only slowly.
d) Grey level variation between road and background is
likely to be large.
e) Roads are unlikely to be short and the curvature of roads
varies slowly.
f) Texture enclosed by the road edge is homogenous despite
of shadows.
g) There exists connectivity between roads which forms
road network.
Nevertheless, some hypotheses are not always true as road
images may vary with ground resolution, road type and density
of surrounding objects. Hence, in this study, the satellite
images from Google Earth are captured with aerial view
in a range of 1000 till 5000 feet for the sake of yielding
International Journal of Applied Science, Engineering and Technology 5:1 2009
54
more precise extraction results by applying these valuable
hypotheses.
III. C
OLOUR SPACE COMPONENTS
The satellite images in used were captured in full colour
forms. Thus, the basic RGB colour model can be utilized
to study the characteristics of the satellite image. In this
study, two more colour models, namely YCbCr and HSV were
selected for investigations. YCbCr model, which is commonly
used as digital video formats, consists of Y (represents lumi-
nance or a measurement unit of the energy amount that an
observer perceives from a light source), Cb (represents the
difference between the blue component and a reference value)
and Cr (represents the difference between the red component
and a reference value). Similarly, HSV also consists of Hue
(a colour attribute that describes a pure colour like yellow,
orange, or red), Saturation (a measurement of the degree to
which a pure colour is diluted by white light) and Value
(intensity or average value of R, G and B at specific location).
Each of these components can be obtained from RGB model
with the expressions as shown:
Y =65.481R + 128.553G +24.966B +16 (1)
C
b
= −37.797R − 74.203G + 112.000B + 128 (2)
C
r
= 112.000R − 93.786G − 18.214B + 128 (3)
H =
θ (B ≤ G)
360 − θ (otherwise)
with
θ = cos
−1
⎧
⎨
⎩
0.5[(R − G)+(R − B)]
(R − G)
2
+(R − B)(G − B)
⎫
⎬
⎭
(4)
S =1−
3
R + G + B
[MIN (R, G, B)] (5)
I =
R + G + B
3
(6)
IV. T
HE HYBRID SGSS-EDGE ALGORITHM
The overall flowchart of the hybrid SGSS-EDGE algorithm
is shown in Figure 1.
A. Image Acquisition
The Google Earth satellite images are used in this project
that can provide medium-high resolution imagery greater than
1-meter per pixel.
B. Road Regions Extraction
It seems that many elements can be found in a colour
satellite image, however, the authors had classified them in
five categories. These five general elements could be found
in a high-resolution satellite image, namely: bushes or trees,
roads, buildings, sandy region and water regions, as shown in
Figure 2. The characteristics of these elements are studied and
examined.
In Figure 2, humans can easily differentiate these five
elements with their naked eyes since all these elements are
Fig. 1. Flowchart of the hybrid SCSS-EDGE extraction
Fig. 2. Five elements in a satellite image
different from each other by their colour and luminance.
For the purpose to let the computers have the same ability,
they are trained to analyse the colour space components
(i.e. luminance, Saturation and Hue values) of the satellite
image. These components can be obtained by converting the
image into YC
b
C
r
and HSV colour spaces. The significant
relationships among them are noticed through examinations
over a number of imageries. Besides, detailed records are
tabulated in Table 1. Other components like C
b
, C
r
and
Value (or intensity) are not used due to their poor abilities
to provide useful information from the evaluations. The
relationships among the colour space components are listed
below:
a) Buildings have the smallest hue value which ranges from
0to0.1
b) The maximum saturation value for roads is 0.3
c) The minimum saturation value for sandy regions is 0.3
d) Bushes or trees are scattered and their luminance value
is below 100
International Journal of Applied Science, Engineering and Technology 5:1 2009
55
TABLE I
R
ELATIONSHIP OF THE 5GENERAL COMPONENTS
Bushes
or Trees
Roads Buildings Sandy
regions
Water
regions
Luminance 25-100 110-160 50-150 100-200 50-80
Hue 0.05-0.3 0.05-0.2 0-0.1 0.05-0.1 0.1-0.45
Saturation 0.05-0.4 0.1-0.25 0.1-0.5 0.3-0.5 0.05-0.2
Intensity 0.05-0.5 0.15-0.8 0.2-0.6 0.4-1 0.2-0.3
e) Sandy regions have similar luminance value with roads
f) Water regions are well separated from roads
g) Roads’ hue value is between 0.05 to 0.2
h) The range of urban road’s luminance is between 110
and 160 lumens.
Therefore, in this project, the road network can be extracted
out from the satellite images by applying these essential
adjustments:
a) Set luminance threshold value larger than 100 to eliminate
bushes or trees
b) Set Saturation threshold value smaller than 0.3 to elimi-
nate sandy region
c) Set Hue threshold value larger than or equal to 0.05 to
eliminate some buildings
d) Use segmentation to separate the water regions and other
small regions from road network
e) Use edge detection to recover the broken road segments
due to some undesired elimination to roads
Firstly, the satellite image is converted into YC
b
C
r
and
HSV modes using image processing software such as Matlab
program. The elimination process can be performed by setting
up a pixel-to-pixel examination loop to filter the unwanted
pixels. Since Y represents luminance in YC
b
C
r
mode whilst
H, S denote as Hue and Saturation correspondingly in HSV
mode, the examination loop can be thus written as
(100 <Y)AN D(S<0.3)AN D(H ≥ 0.05) (7)
Segmentation process is performed on the image as well.
Though the water regions can be eliminated by the loop
described above, the segmentation can further ensure its elim-
ination due to water regions such as lake or river will certainly
not attach to the roadside or any roads. Furthermore, by only
selecting the road regions, other small regions not belong to
the road network can be eliminated in this process. Note that
the authors always assume the road network appears as the
largest connected region in a satellite image that would remain
throughout the segmentation process, and only removes the
relatively small regions.
C. Edge Detection
Road extraction based only on the colour space components
may not be perfect. This is due to constant changes of the
road’s colour while being shaded by trees or buildings. Some
shaded road segments which luminance value is lower than a
predetermined threshold, i.e. 110 lumens, will be recognized
as bushes and thus eliminated by the examination loop. Hence,
edge detection is introduced in order to counter this problem.
Though the colour has been changed, the edge detectors are
still able to find the road edges to reconnect the undetected
road segments as long as the road segments are not entirely
detached or blurred.
Edges are the essential characteristic in road extraction
since the roads are laterally bounded by edges and can be
individuated by edge extractors [9][11]. The edges can be
extracted using various edge detection methods. The image
is converted into grayscale mode prior to applying the edge
detector. In this project, the Laplacian of Gaussian method, as
explained by equation (8), is used.
∇
2
f (x, y)=−
1
2πγ
4
2 −
x
2
+ y
2
γ
2
· e
−
x
2
+y
2
2γ
2
(8)
Equation (8) is a second-order derivative which can be
acquired when Gaussian method is defined as f(X), where
f (X)=
1
2πγ
2
· e
−
x
2
2γ
2
(9)
The Laplacian of Gaussian method finds edges in the input
image by looking for zero crossings after filtering the input
image. This method can detect the edges with different scale
by altering the value, which makes it superior to other edge
detection methods [9]. The threshold value for the Laplacian
of Gaussian filter can then be adjusted to give better results.
Figure 3 and 4 illustrates two set of results showing the
importance of edge detectors in reconnecting the road network
regions.
D. Determining the Road Network
Road extraction results obtained after elimination and seg-
mentation processes are combined with the edge detection
result to determine the actual road network. Based on the
assumption that connectivity exists between roads that form
a road network [10], the largest regions can be considered
as road regions. As a result, a simple calculation on the
total pixels is made to verify and select the largest connected
regions in the result images.
V. R
ESULTS AND ANA LY S I S
The proposed Hybrid SCSS-EDGE algorithm can work on
the complex road segments and large scale satellite image with
good finishing and accuracy. The satellite images captured
within area in Kuala Lumpur city, Malaysia are shown in
Figure 5 to Figure 9 in an increased complexity level. Figure
5 shows a straight road in rural area that surrounded by bushes
and tress. The road region is perfectly extracted without any
discontinuities.
On the other hand, Figure 6 shows another extraction
result on a curve road region (express highway) surrounded
by bushes and sandy regions. These results prove that the
algorithm used in this project has a very good finishing and
accurate performance on extracting the simple road regions.
International Journal of Applied Science, Engineering and Technology 5:1 2009
56
(a) Using edge detection (b) Without using edge detection
Fig. 3. Reconnect straight regions
(a) Using edge detection (b) Without using edge detection
Fig. 4. Reconnect curved regions
(a) (b)
Fig. 5. Extraction for a straight road
Figure 7 displays a moderate complicated road networks
in an urban area, involving multiples of curve road regions.
The algorithm proposed, in fact, is still able to eliminate
the unwanted regions such as buildings at top left corner
and the ground fields located between the curve regions. The
good performance of the algorithm is demonstrated obviously
by the corresponding extraction result image of Figure 7.
Discontinuity of roads at the bottom-middle part is caused
by the blockade of statement put by Google Company.
The road networks in the imageries shown in Figure 8 and
Figure 9 are at much more complicated level. Almost all road
regions in Figure 8 are successfully extracted and displayed in
the result image. Some roads failed to be extracted due to the
image is captured at a vey high aerial distance ( 4000 feet),
thus sometimes made the image blur at random spots. Figure 8
exhibits an image captured at 10000 feet. The extraction result
International Journal of Applied Science, Engineering and Technology 5:1 2009
57
(a) (b)
Fig. 6. Extraction for a curved road
(a) (b)
Fig. 7. Extraction for a small scale complex road
(a) (b)
Fig. 8. Extraction for a medium scale complex road
is able to give a good overview for the main road regions in
the satellite image.
It is noticeable from the results that the proposed extraction
method can perform at a high finishing and accuracy level.
Furthermore, this suggested methodology can generate the
results in a very short time. Each of these extraction outputs
(Figure 5 to Figure 9) is obtained within a minute using a
P4-1.8GHz PC with 256MB RAM, run by Matlab program.
Some objects, such as trees’ shadow, may change the road
luminance value and break the connectivity of the road regions.
Consequently, some road regions are unable to be extracted as
shown in Figure 8 and Figure 9. However, this problem only
affects the small road regions in which their connectivity is
totally broken while the main roads will not be affected. Other
small roads can still be extracted as long as their connectivity
to the main roads still exists.
International Journal of Applied Science, Engineering and Technology 5:1 2009
58
(a) (b)
Fig. 9. Extraction for a large scale complex road
Failure on extracting the road regions in Figure 8 and
Figure 9 may also due to the improper image acquisition
process. These images are probably captured under cloudy or
high humidity environment which caused the blurriness in the
image.
Some unwanted road-like regions are extracted together
with the road regions. This problem is due to the filter’s
disability to separate the road regions from the road-like
regions. For future works, the filtering capability needs to be
improved in order to yield better extraction results.
The main drawback of using Google Earth satellite image is
the non-standardized image resolution. This causes problems
such as luminance value of roads in different regions. Future
works is proposed to overcome this issue so that the road net-
work extraction process for urban planning and GIS database
updates can be performed successfully.
VI. C
ONCLUSION
The road extraction process can be divided into two groups,
namely automatic and semi-automatic (or manual). The auto-
matic way is getting popular attentions due to its short process-
ing time. The Hybrid SCSS-EDGE methodology mentioned
in this study is able to extract road regions automatically,
both in rural and urban areas, from high-resolution satellite
imageries in a very fast way. The results can be obtained from
the hybrid results from colour space elements (luminance,
saturation and hue) and the edge details of roads. Besides,
a number of adjustments are discussed to effectively extract
the road network. The advantages of this method are its fast,
accurate yet simple algorithms. In addition, it can provide
valuable references for organizations or companies which deal
with road maps and GPS (Global Position System).
A
CKNOWLEDGMENT
The authors would like to thank Universiti Malaysia
Sarawak for the facilities, resources and support given to this
project, and also members from Faculty of Engineering who
have contributed or helped in this project.
R
EFERENCES
[1] S. Hinz, A. Baumgartner and H. Ebner, ”Modelling Contextual Knowl-
edge for Controlling Road Extraction in Urban Areas.” IEEE/ISPRS Joint
Workshop on Remote Sensing and Dara Fusion over Urban Areas, pp.
40-44, 2001.
[2] X. Y. Jin and C. H. Davis, ”Multispectral IKONOS Imagery Automatic
Road Extraction from High-Resolution.” Proceedings of IEEE Interna-
tional Geoscience and Remote Sensing Symposium, Vol 3. pp. 1730-
1732, 2003.
[3] A. Chang S. K. Kyu and B. R. Sang, ”A Road Extraction Method
from Topographical Map Images.” IEEE Communications, Computers
and Signal Processing, Vol 2. pp.839-842, 1997.
[4] C. L. Jia, K. F. Ji, Y. M. Jiang and G. Y. Kuang, ”Road Extraction from
High-Resolution SAR Imagery Using Hough Transform.” Proceedings of
IEEE International Geoscience and Remote Sensing Symposium, Vol 1.
pp. 1-4, 2005.
[5] W. R. Chen, C. Wang and H. Zhang, ”Road Network Extraction in High
Resolution SAR Images.” Proceedings of IEEE International Geoscience
and Remote Sensing Symposium, Vol 6. pp. 3806-3809, 2004.
[6] P. Gamba, F. Dell’Acqua and G. Lisini, ”Improving Urban Road Extrac-
tion in High-Resolution Image Exploiting Directional Filtering, Percep-
tual Grouping, and Simple Topological Concepts.” IEEE Geoscience and
Remote Sensing Letters, Vol 3, No. 3, pp. 387-391, 2006.
[7] L. Wang, Q. Qin, S. Du, D. Chen and J. Tao, ”Road Extraction from
Remote Sensing Image Based on Multi-resolution Analysis.” International
Symposium on Remote Sensing of Environment (ISRSE),.
[8] G. Lisini, C. Tison, F. Tupin and P. Gamba, ”Feature Fusion to Improve
Road Network Extraction in High-Resolution SAR Images.” IEEE Geo-
science and Remote Sensing Letters, Vol 3, No. 2, pp. 217-221, 2006.
[9] X, Li, Y. Qiao, W. Yi and Z. Guo, ”The Research of Road Extraction for
High Resolution Satellite Image.” IEEE Geoscience and Remote Sensing
Symposium, Vol. 6. pp. 3949-3951, 2003.
[10] M. Barzohar and D. B. Cooper, ”Automatic finding of main roads
in aerial images by using geometric-stochastic models and estimation.”
IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 18,
No. 7, pp. 707-721, 1996.
[11] Y. G. Xiao, T. S. Tan, S. C. Tay, ”Utilizing Edge to Extract Roads in
High-Resolution Satellite Imagery”. IEEE International Conference on
Image Processing, Vol. 1. pp. 637-640, 2005.
David B.L. Bong David B.L. Bong is a lecturer with the Faculty of
Engineering, Universiti Malaysia Sarawak. His research interest is in the field
of image processing and artificial intelligence.
Koon Chun Lai Koon Chun Lai is currently a tutor with Universiti Tunku
Abdul Rahman, Malaysia. His research interest is in the area of image
processing and electrohydrodynamics.
International Journal of Applied Science, Engineering and Technology 5:1 2009
59
