Yandong Wang 
more detailed road segments in the grouping. Finally, knowledge of road networks is applied to remove non-road 
segments to form a complete road network. 
2 GENERATION OF LINE IMAGES 
2.1 Line Extraction 
In low-resolution images, a road is a smooth line feature which has distinct contrast against its background. Its width in 
the image depends on the image scale. A main road is usually one to three pixels wide when the ground resolution of 
the image is around 10 m. Thus, a road can be extracted by a line operator such as a morphological operator 
(Dougherty, 1992). The main advantage of the morphological operator is that it is unnecessary to define the width of the 
road to be extracted, and thus roads with different widths can be extracted at one time. However, due to the complexity 
of remotely sensed imagery, the extracted line features may not correspond to roads, but non-road features which have 
similar radiometric properties as roads. For road extraction, only roads should be retained while the detected non-road 
line features are noise, and therefore they should be removed. As an extracted line feature only possesses limited 
information about the road, such as pixel gray values of line and their gradients of gray values, it is usually difficult to 
distinguish roads and non-road objects by only using this type of information. It is recognized that a road is a smooth 
line feature while most non-road line features usually have irregular shapes. Based on this observation, a split-and- 
merge operation is used to generate smooth lines in this study. Lines with irregular shapes may be split into short 
segments in this operation, and thus can be removed by a thresholding operation. 
22 Generation of Smooth Lines 
A smooth line is defined as a linear feature in space which is continuous in its first and second derivatives. Thus, a 
smooth line can be generated by calculating its first and second derivatives and checking their continuity. In this study, 
smooth lines are created by using a split-and-merge operation which combines splitting and merging operations 
together. There are some existing algorithms for line splitting using different criteria (Ramer, 1972; Grimson, 1990). 
The method described in Grimson (1990) is very similar to the algorithm presented in Ramer (1972). It splits a line into 
straight segments in four steps: 
1) determine the straight line connection between two end points of a line and measure the maximum deviation of the 
line from that connection (Figure 1); 
2) if the maximum deviation exceeds a given threshold, the line is split into two segments at the point with the 
maximum deviation and the original connection line is replaced by two straight lines which connect the end points 
and the new splitting point; 
3) repeat the above steps until the deviation is below the threshold; and 
4) check consecutive segments to see if they can be merged into a single straight segment without exceeding the 
threshold. 
To generate smooth line segments, the above algorithm was modified in this study. The modified algorithm works in 
the same way in the splitting, but the merging process includes a polynomial approximation of the split segments and 
merging of consecutive segments into smooth lines. Once a line is split into two segments, they will be approximated 
by a polynomial, and their spatial directions at the splitting point will be calculated. Two split line segments will be 
merged together if they have similar spatial orientations at the splitting point. In this way, smooth lines are maintained, 
while lines with irregular shapes are split into short line segments in the split-and-merge operation. The threshold for 
the maximum deviation used in this study is set as three pixels. The reason for choosing a constant for the threshold, 
instead of a function dependent on the length of the connection between two end points, is that the point with the 
maximum deviation does not always appear around the middle of the line. 
C 
  
  
Figure 1. Split-and-merge operation of line segments 
  
944 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
3.1 
Af 
ro 
sel 
fe: 
se: 
mé 
res 
im 
of 
bo 
ea 
3. 
H: 
otl 
ch 
TO: 
TO: 
of 
rei 
frc 
ro 
gu 
re: 
gr 
eft 
In 
cri 
yk 
as. 
gr 
pr 
sir 
co