>st 
/as 
on 
ed 
re 
m 
he 
be 
a). 
ge 
he 
ine 
6, 
wn 
VE 
ne 
ng 
is 
are 
he 
ith 
are 
^as 
me 
led 
Figure 7. The resulting image from interference filtering 
before postprocessing. 
The method is fairly robust with regard to the parameter 
settings, except for the parameter Nfmin. 
The results from the line interference operator have been 
compared to 1) multispectral classification 2) classification 
using two texture bands in addition to the multispectral 
bands, and 3) Multivariate Image Analysis (MIA). 
In 1) and 2) a maximum-likelihood classifier was used. 
Several measures of texture are described in the literature. In 
Haralick 1973, 14 different measures are defined. The 11 
first of these have all been tested. The measures "sum of 
squares: variance" and "entropy" turned out to give the best 
results and these two measures have been used here. Four 
classes were used in the classification task. Class statistics 
were generated from training areas. The classification result 
for class urban was smoothed by a median filter of size 3x3. 
For both results 88% of the old urban areas are recognized, 
however, when texture was used fewer areas were wrongly 
proposed us urban. The unrecognized parts were mainly in 
the outer edges. The central urban area was recognized 
retaining the U-shape, as well as other smaller areas, but 
Figure 8. The interference image after postprocessing (the 
urban area proposal). 
555 
there were large areas which did not correspond to urban 
areas. Both results are inferior to the results from 
interference filtering. However, the classification results are 
sensitive to the class statistics. Defining more classes and 
combining the classification results could improve the overall 
result. 
The MIA system (Esbensen et al. 1989) is developed for 
multiband imagery and calculates principal components of the 
multiband image. These are visualized pairwise in "score- 
plots", allowing for tentative class delineations in the feature 
space. MIA was applied on SPOT band 1 and 2, in addition 
to four texture images made from band 1 with the texture 
features "contrast", "sum of squares: variance", "sum 
variance", and "entropy" (defined in Haralick 1973). MIA 
outlined noisy candidate urban areas which were further 
processed by 3x3 median filtering, three iterations of dilation 
and two iterations of erosion, followed by 9x9 median 
filtering to remove noise and create larger continuous areas. 
The candidate urban areas showed up to mostly cover real 
urban areas. However, a lot of urban areas are also missing. 
6. Conclusions 
Our experiments so far indicate that it is possible by semi- 
automatic methods to extract most of the necessary 
information to perform a coarse revision of topographic maps 
at a scale of 1:50000. Manual control and editing of the 
automatic interpretation result is clearly necessary, and the 
higher degree of uncertainty should be clearly shown by 
printing the changes in other colours. The satellite based map 
revisions will be more uncertain than the airphoto based 
photogrammetric revisions, but with a map revision cycle of 
10-20 years at the present time in Norway, semi-automatic 
revisions two to four times more often would make the maps 
much better approximations to reality. 
The proposed method of interference filtering seem to be 
superior to the other methods tested to recognize urban areas, 
even if textural features are introduced. Additional built-in 
logic in the high level post-processing could further refine the 
result. Our method for road recognition was able to detect a 
new forest road completely without any fragmentation. This 
also indicates that to achieve a very good result using satellite 
imagery of this resolution, it is important to perform the 
revision before too much vegetations grow up along side of 
the road. 
As far as we know, no other system for semi-automatic 
satellite based topographic map revision exist. However, 
Figure 9. Air photo based map for the same urban area 
generated manually.