PRINCIPAL COMPONENT TRANSFORMATION WITH SELECTIVE LIMITED SAMPLING 
In data compression, the PC transformation is usually applied in the 
transmission and storage of data. However, this need not he optimal for 
interpretation purposes. If, for example, an MS record is 1/3 agricultural 
area, 1/3 forest and 1/3 miscellaneous objects, the PC transformation based 
on samples of the whole image might contain very little information for an 
agricultural or a forestry interpreter as the covariance matrix could be mainly 
determined by the variation in the area with miscellaneous objects or by 
the difference between the agricultural and forest areas. 
In the development of a strategy, a first point would be to only sample 
relevant areas for each particular discipline, whereby the PC pictures will 
contain maximum variation between the classes of interest to the interpreter. 
A refinement on this procedure is possible if the idea of separable classes 
is elaborated. 
Interpretation is classification. After a first examination of PC 
images, an interpreter can often already discriminate between several classes. 
Of the remaining classes, the interpreter often knows that it is either class 
A or B (or C). The simplest way to enhance the visual difference between A 
and B is to take two samples, one from an area thought to be A and the other 
from an area thought to be B. The resulting PC transformation shows the 
maximum variance between areas A and B. The procedure is not economical, 
however, since only one eigen vector is determined with two points. 
If the MS record consists of N bands, at least N+1 independent samples 
are required to fully determine N eigen vectors and thus N PC’s. It may, 
however, be desirable to take more than N+1 samples, i.e. 4 samples of C 
and 4 samples of D could be taken in order to give more weight to the 
difference between two nearly equal classes C and D. This difference will 
then appear with a greater eigen value in a PC. 
This implies that the interpretation given to the eigen values in the 
previous section, i.e. a large eigen value corresponds to much of information 
in the PC, has to be reconsidered. The first PC often contains obvious 
information which could be obtained from an examination of one of the original 
pictures. Use can be made of this property by compressing illumination 
effects such as sun angle, cloud shadow etc. into the first PC. This can be 
achieved by sampling the classes cf interest both within and outside a shadow 
area. Depending on the variance between the sampled classes, the illumination 
effect can now be singled out in either the first or one of the next PC’s,. 
The remaining PC's will now be free of the linear part of illumination 
effects - an interesting alternative to ratio processing. Prom this it can 
be expected that the PC pictures in the middle range of eigen values contain 
or can be made to contain most of the information required by the interpreter 
in his specific discipline. Principal Component pictures should not be 
discarded just on account of their eigen values, but should rather be evaluated 
according to their usefulness for differentiating between separate classes.