58 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B6b. Beijing 2008 
during the imaging stage; the second is produced in re-sampling 
or compressing process; the last is determinant, referring to 
noises that are caused by the difference of extracted features to 
non-extracted features. So, we should smooth the image by 
smoothing filter, which could reduce the former two kinds of 
noises and make different physical feature in accordance with 
different gray distribution region, thus facilitate the feature 
discrimination and extraction. In this paper, we utilize median 
filter to achieve the goal. Through repetitive tests, 3X3 
template proves to be effective in eliminating small vehicles on 
roadsides in the imagery. At the meantime, this method is 
simple, quick and automatic, hence the advantage of applying 
to high spatial resolution images of huge volumes. 
Then, clustering algorithm is used to identify road classes. The 
key parameter is the number of spectral clusters K into which a 
scene is classified. Many researches have done on the selection 
of K value, while in this paper, we adopt automatic K-mean 
Clustering algorithm to classify the image simply into two 
classes, one of which contains road information. Advantages of 
such clustering method are obvious: the method is higher in 
automation, less sensitive to initial conditions, and costs less 
computational expenses. Most importantly, as shadows on roads 
are classified into the class containing road information, the 
method ensures the road continuity. For non-road speckles or 
patches within the road class, they can be eliminated 
commendably in the next stage. 
2.3 Road Connection Algorithm 
In order to extract roadside in spite of heavy shadows on roads, 
we bring forward a new algorithm named “road connection 
algorithm”. This is the key procedure of the whole frame work, 
in which spectral and geometric features of roads are 
represented in two rules: (1) spectral feature: for each pixel, 
only when its three radiometric properties are similar to those of 
roads, it can be identified as road candidates. (2) geometric 
feature: on high resolution image, roads are usually smooth 
with no small wiggles. Thus in an appropriate distance, road 
segments, including smooth curves, can be recognized as 
straight line segments. 
According to road geometrical properties mentioned above, for 
each pixel on the image, search in specific direction in a straight 
line at fixed step. If the proportion of road candidate pixels 
(refers to pixels whose spectral property satisfies the road) to all 
pixels on the line exceeds a threshold value S that we give in 
advance, all pixels on the line L could be identified as road 
pixels. After that, we convert original non-road pixels to road 
pixels. 
Assume the image after rough classification as I, and the result 
after road connection algorithm as RI, and the value of number 
1 marks road. Then the pseudo codes are as follows: 
Pro Road Extract (I, L, R) 
W = image width, H = image height 
For i = 1 to H do begin 
For i = 1 to W do begin 
For 9 = 0 to 180 by 15 do begin 
DL -> pixels on the line segments starting from (ij) in 0 
direction with length L 
RN-> number of pixels in DL in image RI whose value 
equal 1 
If RN ^ 1 then continue 
End if 
In this paper, we take Beijing city as a study case and the data 
we utilize is multi-band QUICKBIRD image. In order to 
evaluate the result, we define an assessment system according 
to information in the imagery and achieved a satisfactory result. 
The approach has proved to be simple, quick, automatic and 
efficient. 
The remainder of this paper is organized as follows. The 
automatic road extraction framework which includes the rough 
classification, road connection and result grooming is 
introduced in section 2. In section 3, we utilize this method to 
extract road from QUICKBIRD image and give the accuracy 
assessment. At last, conclusions are presented in section 4. 
2. METHODOLOGICAL FRAMEWORK 
The approach includes both spectral and geometric constraints 
about roads network in urban areas. The methodological 
framework contains three steps: rough classification, road 
connection, and result grooming. 
2.1 Road Feature Analysis 
In high spatial resolution remotely sensed images, urban roads 
have following properties: 
Stability of spectral property: The spectral properties of 
uncovered roads are stable to a certain degree. Because urban 
roads are mainly constructed by asphalt or cement, especially 
asphalt dominates a large part; spectral properties of roads are 
limited to a fixed range which corresponds to the spectral range 
of road materials. However, in the imagery, objects on 
roadsides like zebra crossings, cars and people cause noises due 
to the huge spectral difference to roads. 
Continuity of roads: Normally roads in reality are continuous 
and regular in geometry, while in the imagery, trees and 
shadows of high buildings by roads interrupt the continuity of 
roads to a large degree. But on the whole, roads in the imagery 
still have impressive connection and regularity. 
Straightness: On high spatial resolution images, urban roads 
are straight and smooth with no small wiggles thus can be 
recognized as combinations of straight road segments. 
Topological property: Road segments are always connected 
with each other constituting road networks, and impossible to 
be broken suddenly. 
2.2 Rough Classification 
The first step is rough classification, which enhances the 
exploitation potential of spectral content for automated road 
extraction. 
Firstly, after atmospheric and geometric correction, we need to 
pre-process images to eliminate noises in the imagery. For high 
spatial resolution satellite images, there are primarily three 
types of noises: the first is called white noise that is formed