5.1 Roads 
Segmentation by "original minus median" gave a noisy result 
with a lot of small pixel segments. The smallest segments 
were deleted by the noise filters. In addition, the method is 
extremely sensitive to the threshold level. Only level 131 
gave acceptable results. The template matching method was 
not sensitive to the threshold level and contained much less 
noise. However, the lines were wider using this method. We 
selected the template matching result for further processing. 
The result of map-guided detection is shown in fig. 4. While 
interpretating the image, the feature image and profile 
analysis was used 73% of the time, the binary image was 
used in the remaining 27%. 76% of the roads were 
recognized. From the 24% not recognized 16% can be 
eliminated as noise (small parts a few pixels long which were 
not detected due to poor contrast or too large map 
generalization). This noise could be removed by 
postprocessing. The remaining 8% should be controlled 
visually by the user. Two larger road sections were not 
recognized for reasons of poor contrast and map 
generalization. One of these sections followed the coastline 
and had therefore been moved on the map more than our 
maximum distance difference parameter of 40 m in the terrain 
(this section amounts to 3%). The other section that was not 
found was lying in a shaded area caused by a relatively steep 
hill. Neither of the sections could be found by visual 
inspection of the original image data while performing 
interactive contrast enhancement. 
Various steps for detection of new roads are shown in fig. 5. 
Fig. 5 a) shows the resulting image from template matching, 
described in section 2.1, of a lower middle section of our 
test scene, (band 1). In fig. 5 b) are the thinned line 
structures of the result from the template matching 
segmentation. Threshold value 13 has been used. Fig. 5 c) 
shows the result after removing the old, recognized roads and 
coastline, and the line structures present within old and 
proposed new urban areas. Lines within urban areas having 
an area of less than 100 pixels are, however, not removed to 
reduce the number of gaps generated in the line structures. 
Fig. 5 d) shows the result after the generation and testing of 
the hypothetical connection lines. For verified connection 
lines there have been generated lines in the binary image. 
Lines having a length less than 15 pixels are removed. 
The large proposed road structure is a new forest road, and 
the complete road has been well recognized as one connected 
line. The straight line in the upper right part of the image is a 
pier. Most of the remaining line structures are private roads. 
The private roads were not present in the old map data so 
they were not removed and therefore proposed as new roads. 
5.2 Water Bodies 
SPOT band 3 (near-infrared) was used for interpretation of 
hydrographic features. For interpretation of the sea shores the 
system used the binary segmented image 62% of the time 
and band 3 directly for profile analysis 38 % of the time. 
76% of the shores were recognized. From the 24% 
unrecognized 5% were due to real changes of the sea shore. 
A relatively large sea area had been filled in. The largest 
section of the 19% that were unchanged and undetected was 
again in a shaded area caused by a steep hill (the section 
amounts to about 5% of the unrecognized shores). 
2% of the sea and lake areas were not detected. The changed 
areas amounts to about 2/3 of this. The rest of the areas were 
mainly classified as land due to shallow water (sea bottom 
was clearly visible in the water). Surprisingly, 88% of the 
rivers were reconized (only rivers wide enough to be 
represented with two lines were present in our map data). 
The borders of the rivers were detected using a binary image 
52% of the time and a dark line feature image 48% of the 
time. 
5.3 Urban Areas 
The template matching algorithm was applied to extract line 
features. The thinned line feature image is shown in figure 6, 
while the resulting image from interference filtering is shown 
in figure 7. As expected, other line structures than roads have 
responded and created "noise". The airport in our test scene 
has given considerable response, as well as parts of the 
coastline, road intersections in rural areas and roads going 
into the urban areas. 
The interference image is smoothed and thresholded before 
the recognition. In map guided recognition, 93% of the urban 
areas marked in the last map revision are recognized. The 
unrecognized parts are mostly in the outer edges of the areas. 
The proposed urban areas from unguided recognition is 
shown in figure 8. The manually updated urban areas are are 
shown in figure 9 (air photo based revision). In general, the 
result is a coarse approximation to the new map data with 
some exceptions. Some small groups of urban areas are 
connected to larger structures, and the proposed new areas 
often tend to be somewhat larger than in the map. Some 
noisy areas are also present, and the airport has been detected 
as an urban area (correct?). 
Figure 6. The thresholded and thinned line image. To the 
right is a subsection of the image to the left showing an urban 
area (Værnes). 
554