ist dann nur für Verbesserungen der Transformationen oder in Ausnahmefällen
notwendig. Der Artikel legt die Eigenwert-Theorie dar, eine Auswahl
strategie wird gegeben. Derzeit können noch keine experimentellen Ergebnisse
gezeigt werden.
INTRODUCTION
The need for data compression is evident when working with multi-spectral
data. A relatively simple pre-processing system is therefore being developed
at ITC, based on the use of mini computers and digital to film converters
or possibly optical processing. In the pre-processing stage, the illumination
effects can be eliminated by, for example, ratio processing, and the data
compression can be accomplished by means of Principal Component analysis,
both techniques being well known. In this article the idea is discussed of
optimising the Principal Component transformation for visual interpretation.
This transformation is widely used in data compression since it retains maximum
variation, though this will generally not imply maximum visual information.
A Principal Component transformation is fully determined by the covariance
matrix of its sample set. By using a favourable sampling strategy, the inter
preter can optimize the Principal Component transformation to give maximum
visual information in a few transformed pictures. This could also be a suitable
strategy in the composition of false colour pictures, involving the transform
ation and compression to three b & w pictures.
In order to make this system practicable, the user should have some know
ledge of how sampling influences the Principal Component transformation. A
summary of the essential theory required is therefore presented in this paper,
together with a few simple examples.
The user will have to optimise the Principal Component transformation
for his field of interest by experimenting with the method of selective-
limited sampling. It may be found that standard transformations are applicable
for certain disciplines, in certain regions or for particular seasons, in
which case there would be no necessity to define a new transformation by
sampling methods. Furthermore, it could be found that an original spectral
band has a rather low weight factor in all of the transformed pictures. That
band could then be omitted in the processing phase and the data reduction
thus being performed before transformation.
THEORY OF THE PRINCIPAL COMPONENT TRANSFORMATION
The PCT (Principal Component Transformation), also known as the Karhunen-
Loeve transformation is a linear (pointwise) transformation involving the
weighted summation of the original intensities of all bands. In MSS (Multi-
Spectral Sensing), N intensity records are obtained of a scene. Each picture
point is thus characterised by N number, X^ , X^ X„, which together
form the componejrts of an N-dimensional (intensity) vector X = (X^ , X^.^X^.).
The position of X in the picture is uniquely determined by an index k,
k varying from 1 Id M if the picture has M points. For each point k, the PCT
matrix Q" 1 transforms X into Y, of the same dimension N. (Equivalent to N
PC pictures)
