Jp 
rtography 
| road data from 
will be launched 
nd derivative of 
1 adjacent pixels 
matic grey scale 
e line segments, 
is considered as 
d by a simulated 
data. There still 
expectedly poor. 
nsional property 
n their previous 
xts (rural, forest 
iere selected and 
ch worked quite 
| images whereas 
ods were needed 
method standing 
t streets crossed 
quite fine and 
assumption may 
-automatic road 
solution satellite 
ulti spectral data 
racked based on 
tained. Then this 
he best direction 
1 urban area was 
. missed, as road 
irrounding roads. 
s were properly 
| automatic road 
ale images. This 
n, edge thinning, 
er simple image 
residence area it 
t some additional 
ceded. 
'd road extraction 
point to extract a 
int is calculated 
mplate matching 
ods using snake. 
1 the road cast by 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
Hu and Tao (2002) proposed template matching based main 
road extraction method for high-resolution satellite image. A 
two direction (horizontal and vertical) template matching is 
applied to the reduced resolution image. In line grouping 
process, they classified line segments and the segment 
connectivity matrix is constructed. Then collinear chains are 
extracted. Finally, precise position of road centre line is 
extracted on the original image by least square template 
matching. They applied their method on a IKONOS image. 
Main road is correctly extracted in the open rural area while the 
result in the complex building site is not so good because of the 
complicated road network. 
According to Vosselman & Knecht (1995), road characteristics 
can be classified in five groups; Photometric, Geometric, 
topological, functional, and contextual characteristics. As 
ground resolution of PRISM data is 2.5m and small marks on a 
road arc not visible, precise road model are not adequate for 
this study. Instead of a precise model, simple line detection 
methods, which use photometric and geometric characteristics, 
arc employed for road extraction in this study. Geometric, 
photometric and contextual characteristics are used for 
grouping line segments and building road network topology. 
In this paper, we propose a semi-automatic road extraction 
method based on three stages; line feature extraction stage, line 
segment classification stage, and line segment grouping stage. 
À centre line-detecting algorithm proposed by Steger (1998) is 
employed for feature extraction. It picks up pixels where 
second derivative of brightness becomes maximal. Those pixels 
are linked and chains of pixels are created if both probability of 
line and angle difference between adjacent pixels satisfy given 
conditions. In line segment classification stage, acquired line 
candidates are classified by its photometric property. We test 
both automatic grey scale threshold method and traditional 
unsupervised multi band classification method for this stage. 
After eliminating false line segments, a line linking method is 
applied on the line segment grouping stage. If all of angle 
difference, lateral offset, and net gap are less than threshold, the 
pair is considered as one long line. This process is applied 
iteratively while a connectable pair remains. 
In the following section, methods for cach stage are described. 
Then we present extraction results for each stage. Finally, we 
discuss the result and give direction for further work. 
2. METHODS 
2.1 Data used 
A simulated ALOS PRISM data set has been used for this 
study. It has been offered by NASDA/EORC for limited use. 
Data was acquired with an airborne three-line CCD sensor to 
simulate along-track PRISM sensors. Original data (25cm 
resolution) was thinned down to fit the targeting resolution 
(2.5m) by averaging surrounding 10x10 pixels. Ancillary data 
such as sensor position and attitude corresponding each image 
were also provided, though image data is only used for this 
study. 
As the image size is too large to manipulate, some target arca 
are selected. Figure 1 (a) and (b) are used for evaluation. Figure 
| (a) is 329 by 642 pixels and dense residential area. It contains 
some residential quarters and some single lane or dual lane 
roads divide them. Dual lane roads separate residential area and 
its adjacent area such as forest and agricultural area. Figure | 
(b) is 300 by 600 pixels. Most part is paddy field and dry field. 
À major road goes through the northward. Some habitations can 
be seen at the east area and minor roads go through them. All 
roads are clear to human eyes. Ground truth data is created 
from 1:25,000 vector map data and the simulated images by 
manual compilation. 
As the data set only include the simulated PRISM images but 
not contain simulated AVNIR-2 image, both PRISM and 
AVNIR-2 simulation data were created from IKONOS images 
to investigate effectiveness of multi band classification. 
Standard geometrically corrected IKONOS 4band image in 
Farnborough, Hampshire, the United Kingdom and IKONOS 
panchromatic image in the same area with the same process 
level were used. Both images are firstly reduced in resolution to 
fit the ALOS resolution. The resolution merge tool of ERDAS 
was used to create the pan-sharpen image. The merging method 
is the principal component method and the re-sampling 
technique is the cubic convolution method. Its image size is 400 
by 400 pixels. Main roads are clearly recognized as grey 
elongated area. 
ense residence area (a) and for 
rural area (b). The image size is 329 by 642 
pixels (a) and 300 by 600 pixels (b). Histogram 
equalisation was applied for both images. 
2.2 Line Feature Extraction Stage 
Many studies employ edge detector such as Canny, SUSAN for 
feature extraction method. A ridge extraction method proposed 
by Steger (1998) is used in this study. It follows centre of bright 
(or dark) blob where the second derivatives of profile crosses 
zero. The method gives centre (not edge) of road, needs no 
excess parallel edge process, and has sub-pixel accuracy. It 
consists of two stages, line point finding stage and line linking 
stage. It assumes that roads are almost homogeneous and have 
clear contrast to their adjacent areas. However, as real images 
have some noises that give us false ridge, it used Gaussian 
smoothing kernel for convolution for noise reduction in line 
point finding stage. Characteristic of this method is, smoothing 
and ridge extraction is integrated for avoiding loss of 
information. 
The employed method has three parameters to be determined. 
c, sth, and lth. o is called Gaussian parameter to determine 
degree of smoothness in smoothing process. Sth and Ith is 
called seeding threshold and linking threshold respectively. If 
absolute value of the second derivative of a pixel is more than 
sth, the pixel is considered as the seeding point and the line 
following process would begin. The extended direction of the