ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision‘, Graz, 2002 
  
Road lines are extracted in a semi-automatic way, where given a 
starting point, a road line is traced by iteratively matching a 
road template with both road mask and road seeds, control 
points indicating the right directions are assigned by operator if 
the road line got lost. 
  
1. Road Mask Extraction: 
Road pixels are identified using an 
existing remote sensing software. 
Y 
2. Road Seed Extraction: 
Road seeds are extracted by tracing 
edge pixels. 
  
  
  
v 
3. Road Line Extraction: 
Road is semi-automatic extracted 
using template matching. 
  
  
  
  
Figure 1. Flow of road extraction 
In the following chapters, we address each steps in detail. 
Experiments are conducted using an IKONOS image nearby 
KAWAGOE City, Japan. Experimental results and discussions 
are given subsequently, where road extraction and map 
generation are studied in detail at both dense building area and 
countryside. 
2. EXTRACTION OF ROAD MASK 
A pixel-based road mask is extracted using a commercial 
remote sensing software IDL/ENVI. Given the training values 
of vegetation, soil, building (concrete), water, express highway, 
main road and small road, image pixels of a multi-spectrum 
satellite image are classified into different groups using 
maximum likelihood method. As the multi-spectrum satellite 
image used in this research has four bands, i.e. red, green, blue 
and near infrared, classification result is much more reliable 
than those on three bands, i.e. red, green and blue, or less. 
Image pixels classified as express highway, main road and small 
road are exploited in this research to generate a “road mask”. A 
"road mask" is a binary image, where white pixel suggests that a 
road like object is most probably over there. The road mask is 
exploited in road line extraction. 
  
1. Edge Pixel Extraction: 
Edge pixels are extracted using 
Canny filter. 
2. Tracing Edge Line Patches: 
Continuing edge pixels are traced, 
and edge lines of sudden and/or 
fast change of direction are broken. 
Y 
3. Jointing Edge Line Patches: 
Edge line patches that are on a 
continuous line of a slow change of 
direction, and with only a limited 
blanks between them are jointed. 
  
  
  
  
  
  
  
Figure 2. Flow of extracting road seed 
3. EXTRACTION OF ROAD SEEDS 
Road seeds are extracted in three steps as shown in Figure 2. 
Edge pixels extracted by Canny filter reflect not only the 
boundaries of roads, buildings, rivers, or shadows, but also any 
other local rapid changes of photometric characteristics. 
Although especially in central town, edge pixels of buildings, 
trees and shadows are mixed with that of roads, so that road 
boundaries are not as clear as to be extracted, a long edge line 
with only a slow change in direction do strongly suggest the 
boundary of a road or a river. Thus, they are extracted in this 
research as "road seed" (see Figure 3), and exploited as the 
complimentary information of “road mask” in road line 
extraction. 
   
Edge line patch 
Road seed 
Figure 3. Definition of *Road Seed" 
Applying Canny filter on satellite image results in a binary 
image of edge pixels, where edges are thinned to a width of 
single pixel. Extraction of long edge lines from the binary 
image of edge pixels is conducted in two steps, tracking and 
jointing edge line patches. Road seeds are generated for any 
edge pixels on the edge line patches, where two directions of 
the road seed, indicating two ways of the road line, are 
calculated using the tangential line of the edge line patch at the 
edge pixel. 
3.1 Tracking Edge Line Patches 
The binary image of edge pixels is examined row by row from 
top-left to bottom-right. Whenever an edge pixel is met, if it is 
an isolated one (there is no other edge pixel in its 8-neighbors 
as shown in Figure 5), the edge pixel is removed form the 
binary image, and tracking procedure continues in looking for 
other edge pixels. Otherwise an edge line patch is tracked from 
the edge pixel as shown in Figure 4. Iteration continues until no 
edge pixel is found in the binary image. 
  
1. Linking Edge Pixels: 
Edge pixels that are continue in 
their 8-neighbors or with a few 
blanks in its local extensions are 
linked together. 
Y 
2. Breaking Edge Line Patches: 
Edge line patches are broken down 
as the pixel of sudden and/or fast 
change of direction. 
Y 
3. Post Process: 
Edge line patches shorter than a 
given threshold are discarded, 
while others are recorded for 
further process. 
  
  
  
  
  
  
  
Figure 4. Flow of tracking edge line patches 
A - 407