to 
"th 
and 
] as 
nat, 
[he 
tor 
ine 
irst 
the 
ure 
of 
age 
s a 
out 
eas 
ize, 
the 
old 
> is 
sly 
rea 
res, 
and 
for 
the 
|, is 
ked 
together in a pointer list during the scanning. Each candidate 
area description object points to its outer border description 
object, and to the list of inner hole description objects. For 
the outer border the area is also a measure of border length. 
3. The list of area objects is traversed in order to decide 
whether candidate area objects should be accepted or 
rejected. The rejection criteria are: 
- area size < Nmin 
- Nf « Nfmin 
4. Holes in the accepted area objects are filled on these 
conditions: 
- Absolute size: Number of pixels in the hole is 
less 
Nhmin 
- Relative size: Size of hole is less than Ph percent 
of total area. 
Holes having more pixels than the absolute size are never 
removed. 
Figure 5. a) (upper left) shows the result from template 
matching ("the feature image"). The thresholded image is to 
the upper left (b), and the thresholded image after removing 
line segments in urban areas and the costline is to the lower 
553 
5. The resulting image containing the accepted area 
structures is optionally smoothed and thresholded. The 
smoothing is required if the binary image was the non- 
smoothed thresholded interference image. A standard 
median filter with kernel size 5x5 is used. 
The following parameter values have been used in our tests: 
NI =3 
Nfmin = 4 
Nmin =2 
Nhmin = 40 
Ph zm 
These values are set partly as a result of experiments and 
partly due to knowledge about the response values in known 
urban areas in the interference image. More test scenes are 
necessary to determine the optimum parameter value. 
5. Results 
The interpretation system has been tested on a SPOT scene of 
Varnes near Trondheim in Norway. A 512x512 subimage 
from band 1 has been used (the green band, 20m resolution, 
see fig. 4) for detection of roads and urban areas, and band 3 
(near-infrared) for detection of water bodies. 
left (c). To the lower right is the result from the road 
detection algorithm. A larger forest road is completely 
recognized (d).