4.0 DISCUSSION 
The results profile the implications of importing raster data into the 
vector domain of a GIS and importing vector data into the raster 
environment of an IAS. = The effects of the various data 
transformations with respect to geographic accuracy is addressed. 
There is a obstacle with some GISs, in that, there are limitations in 
the software. Most GISs have a practical limit in terms of the 
number of elements that can be addressed in a single coverage. 
While this limit is generally large, it is still a limit. This leads one to 
consider the complexity of an image. That is, is the coverage 
going to exceed the limits of the software ? More and more, this 
is becoming a bottleneck for analysis. Many researchers must 
devise creative solutions to deal with these inherent software 
limitations. Occasionally, these limits are practical, rather than 
physical. That is, they reflect the hardware limits more than the 
software restrictions. Hardware limitations include disk storage 
capacity and processing speed. By increasing either of these, the 
user is faced with increasingly cost-ineffective solutions to their 
problems. 
Since feature extraction is concerned with the simplification of 
highly complex information, it follows that the actual process 
involved is likewise highly complex. The simplest solution to this 
problem is to first stratify the complex data and then perform the 
feature extraction procedure. Image classification is just such a 
stratification scheme, albeit a complex one. In a simple case, we 
Observe a standard 8 bit image channel to have 256 possible digital 
values. Imagine the increased complexity by adding further 
channels. Conversely, consider the case where the original image 
can be stratified, through a supervised classification technique, to 
a mere 11 classes. The task of feature extraction becomes 
considerably easier. 
Because we have stratified the data into a number of desired 
classes, we have some control over how the individual pixels 
become classified. This implies that the number of potential 
artifacts that could result from the feature extraction process is 
minimized. In a homogeneous field, the number of misclassified 
pixels is minimal. However, in a heterogeneous field, the potential 
368 
number of misclassified pixels increases. Thus, obviously, one 
could expect quite a few artifacts from a heterogeneous field, and 
little, if any, from a homogeneous field. 
The procedure of extracting vector features from a raster theme 
image is currently still in the research phase (O'Brien (1991), Taylor 
et al. (1991). Experiments are going on that are scene and 
situation specific, and thus can not be applied to a general case. 
Progress is being made in this field, but results must be taken with 
a grain of salt (Taylor (1991)). There are semi-automated 
procedures for feature extraction that are being used, but they 
require considerable operator input (Zelek (1990), Van 
Cleynenbreugel et al. (1990)). For example, a feature is identified 
by an operator and at a certain point the extraction or recognition 
algorithm takes over. This sort of procedure generally produces 
more reliable results than the fully automated procedures, but at 
the cost of greater operator interaction. Currently rule-based 
feature extraction techniques use operator expertise and knowledge 
of a specific site to aid in the extraction process (e.g. Van 
Cleynenbreugel, 1990). Although one would expect this approach 
to yield the most robust results, it is not always possible to have 
such in-depth knowledge of a study site. In cases where 
knowledge of a specific site is not known, the extraction algorithm 
must proceed without the benefit of any additional information. 
A common feature extraction application is image segmentation. 
Image segmentation refers to the selection of linear features from 
an image. Typically, segmentation is used to select road features 
from an image. |t is understood that pixel resolution has a 
profound effect on the ability of a feature extraction algorithm to 
pick out specific elements (Van Cleynenbreugel et al. (1990)). By 
increasing the resolution of the pixel, the feature being sensed is 
more truly represented, and is, therefore, more easily recognized. 
Feature extraction algorithms seek out regions of homogeneity. 
There is far more information in a digital image than can be seen 
with the naked eye. Image segmentation algorithms are designed 
to seek out a specific element and identify it as such. The end 
result is that, often, image artifacts or noise are extracted in 
addition to the desired elements. This noise can be dealt with 
through spatial filtering techniques or by selecting elements that 
meet a certain criteria and subsequently deleting them. In images 
where regions of homogeneity are fuzzy, a data stratification 
approach must be adopted. 
Abstracting vector data to a raster representation is a different 
matter, and is generally more straightforward. The procedure in