data within the framework of an on-going project at the Canada 
Centre for Remote Sensing (CCRS) known as the Crop Information 
System (Manore et al. (1989)). Water bodies were chosen for the 
comparison because they are simple to recognize visually. The 
data stratification scheme performed was a simple maximum 
likelihood classification on the original imagery. The result of the 
classification was an 11 class land cover theme image. In 
particular, the water-body theme was used for processing into a 
resultant vector coverage. 
A manually digitized coverage of water bodies for the study area 
was available. This vector coverage was the control vector 
coverage against which the output from the feature extraction were 
compared. 
3.0 METHODOLOGY 
The following is a generic methodology for feature extraction from 
raster imagery. The assumption made is that the imagery contains 
some useful geographic information, but that this information is not 
in the proper format, which in turn provides the impetus for the 
operator to extract these features from the imagery. Feature 
extraction gives a user a vector product that can easily be 
integrated into the GIS environment with as little operator 
interaction as possible. It is recognized, however, that given the 
current state of software development, this fully-automated 
methodology, while promising, is not yet feasible for operational 
use. The approach above was followed in this paper as much as 
possible with exceptions discussed. 
3.1 Raster to Vector Conversion 
Classification of Raster Image 
It is understood that it is possible to extract features from an 
unclassified image, likewise it is also understood that the job of 
feature extraction would proceed much more simply if the data 
were stratified. There are procedures that can be used to stratify 
the data (e.g., density slicing or supervised/unsupervised 
classification). The purpose of stratifying the data is to make the 
analysis procedures more practical, in terms of processing time and 
disk storage. 
Feature Extraction 
The purpose of the feature extraction procedure is to identify 
homogeneous clusters of pixels. In this case, only a single theme 
class (water bodies) was used from the original 11 theme 
classification. Sub-pixel elements were not considered to be 
significant in this comparison. 
Boundary Extraction 
Once the homogeneous clusters of pixels have been identified by 
the feature extraction process the next step is to delineate the edge 
of these clusters. This is the process of boundary detection, which 
is also known as image segmentation. The output from the 
boundary extraction procedure is, presumably, a vector 
representation of the original polygonal structure or feature. 
Generalization of extracted vectors 
If the extracted boundaries were examined at this point, they would 
appear to be ’step-like’. That is, they would follow the exact 
contours of the pixel edges. In order to smooth out these ’steps’ 
and create a more realistic representation of the feature, a 
smoothing-filter needs to be passed over the edge. The larger the 
filter size, the greater the effects of the smoothing. Thus, while a 
3 x 3 filter might smooth the step-edge slightly, a 9 x 9 filter might 
distort the edge and even larger filters might shift the X and Y 
coordinates. 
By generalizing the data, the data volume is also decreased. The 
benefit of this is decreased data storage requirements and 
increased processing speed. The potential deficiency of this is that 
the data may become too generalized, and not very well registered. 
Export of vectors to GIS 
Up to this point, the work done has been entirely in the image 
processing domain. The extracted features are now ready to be 
exported to the GIS. This procedure is a straightforward translation 
of the extracted vectors to a format compatible with the GIS, such 
as the Digital Line graph (DLG) format. 
366