Figure 1. Main design of the prototype system. 
required. When making these changes, the operator uses 
editing functions similar to those found on most digital 
mapping systems. The final revised map is then used to 
update the cartographic database. 
In Norway, the map features for which there is the greatest 
requirement for revision are: roads, built-up areas, forested 
areas, and power lines. In this project, it has been decided to 
concentrate on roads and built-up areas. 
3. Map-guided Detection of Old Objects 
The old unrevised map is here used to guide the interpretation 
process. The relevant map features are converted to an 
internal raster format. Raster format is a more suitable 
representation of the map than vector for our applications, 
especially because the spatial relations are represented in a 
direct way as in the satellite imagery. Some information is 
built into the raster points, e.g. pointers to the neighbour 
points for each road raster point. 
3.1 Roads 
Roads are typically line structures with higher reflectance in 
the visible part of the electromagnetic spectrum than the 
surrounding environment, at least when the surrounding 
environment is covered by vegetation. We have found two 
suitable algorithms for the extraction of brighter lines on a 
darker background: "Original Minus Median" (Graga 1987) 
and "Template Matching" (Wang & Howart 1987). 
The idea behind the first method is that the median filter 
suppresses singular and one dimensional structures, i.e. 
points and very thin lines. Accordingly, an image made by 
subtracting the smoothed median filtered image from the 
original image should contain line features and points. The 
points are regarded as noise here. 
In the second algorithm, the image is filtered by a set of 
templates representing local line structure. The filter response 
values are measurements of how well the corresponding 
templates match the local line geometry represented by the 
templates. Fourteen templates are used, each with a pattern of 
weight numbers representing a smooth line of three pixels. 
550 
The pixel values in the resulting image are the maximum 
response values from the filtes. 
The interpretation algorithm described below uses the 
segmentation result directly as a "feature image" and a 
thresholded version of it as a binary image. The threshold is 
set by the user. The binary image is partly "cleaned up" by 
noise reduction consisting of two straightforward filter 
operators removing isolated single pixels and pixel pairs in 
two passes. In the first pass one pixel from each pixel pair is 
removed, in the second pass all isolated single pixels are 
removed. 
Interpretation of the segmented image is performed as a line 
tracing process. Each road network in the raster map is 
located by a scan search. When a network is located, it is 
traced pixel by pixel. A dynamic list of junctions is kept to 
assure that each segment connected to the junctions are 
traced. Tracing is performed in parallel in the map and in the 
image. The map tracing is the active part guiding the position 
and direction of search for the road counterpart in the image 
data. The satellite imagery has in advance been geometrically 
transformed to the map coordinate system, however, 
corresponding objects can have slightly different positions 
due to small geometric errors in the image and map 
generalization. The system has therefore been designed to 
tolerate positional errors less than a user-defined constant. 
However, the topological relations have to be equivalent in 
the map and image. 
The interpretation process dynamically selects between two 
interpretation algorithms. The least CPU-consuming is just to 
check if a line point is present in the binary image. However, 
a line point will only be present in the binary image if the 
contrast to the surrounding environment is relatively high 
(and the threshold has to be set relatively high to suppress as 
much noise as possible). When a line point cannot be found 
in the binary image, a profile analysing algorithm is 
performed on the feature image data. Intensity profiles are 
extracted in the different possible continuations for the road. 
The first derivative profiles are computed and analysed to 
locate a possible intensity top that can be the line 
continuation. The algorithm includes procedures to 
distinguish between small local intensity tops and less local 
tops.