through subsampling of the classified data. 
Analysis techniques -- Signature extension 
One of the least well understood aspects of the image analysis pro- 
cedure, be it manual or automated, is that of signature extension. When 
choosing representative examples of a particular resource type to be 
used as a training set for a classifier, it is necessary to know over 
how large an area the spectral signature of the training set can be ex- 
tended and still yield acceptable classification results. 
The problem is not a simple one. Both the inherent spatial var- 
iability of the environemnt as well as the characteristics of the imagery 
or data being analyzed will affect the extent to which signatures can be 
extended. There certainly is no general answer, as there are many var- e o 
iables affecting the spectral signature of a given resource type. Never- 
theless, a much better understanding of the problem is necessary before 
it can be said that automatic classification can be applied in an oper- 
ational sense to a variety of resource mapping tasks. 
Analysis techniques -- masking 
The synoptic coverage of LANDSAT provides data with considerable 
variability in spectral data inherent to the data. A large number of 
training sets is often needed to account for the spectral variability 
when performing image classification. Variations in landform, land use, 
and natural ecosystems are factors contributing to the variability of 
spectral signatures. Studies have shown that stratification of a LAND- 
SAT scene into relatively homogeneous strata with respect to ecological 
parameters can substantially improve the accuracy of classification and 
reduce the amount of time required to complete the processing (Nichols 
et al, 1974)... Stratification prior to classification is intended to 
reduce the spectral variability within strata, reduce the number of 
training sets required within strata, reduce the number of pixels being Qo 
classified at one time, and increase classification accuracy.  Imple- o 
mentation of masking techniques requires digitizing strata boundaries 
and merging these strata boundaries with the LANDSAT data. Although soft- 
ware has been developed to perform this registration function, research 
is needed to define criteria for establishing the number of strata and 
size of strata required to optimize the classification procedure. 
Analysis techniques -- supervised vs. unsupervised training 
In recent years, a great amount of research has focused on the sub- 
ject of supervised and unsupervised classification. There is consider- 
able confusion, misconception, and misunderstanding of these two types 
of analysis and how they should be applied to image analysis problems. 
Supervised classification refers to the process of selecting homogeneous 
training sets which represent all of the spectral variability present 
in each cover type. The spectral signatures for a given training set 
are calculated from all pixels within the boundaries specified by the 
analyst. In many cases, the distribution of data within training sets 
selected for supervised classification violates the assumptions of gaus- 
sian statistics inherent in most maximum likelihood decision rules or © o 
does not totally represent the spectral variability in the desired re-