-' 333 =
APPLICATION OF FACTOR ANALYSIS
FOR THE DETERMINATION OF PHYSICAL
EFFECTS FROM CAUSAL INTERRELATIONS
OF MULTISPECTRAL BAND SIGNALS
Rudolf H. Dittel
Institut fiir Nachrichtentechnik
DFVLR
8031 Oberpfaffenhofen
1. Introduction
For the purpose of classification the spectral characteristics of different
surfaces are described by the mean value of the spectral intensities and the
covariance matrix of all spectral channels selected. Experience shows, that
increasing the number of spectral channels does not necessarily increase the
quality of the classification result. For the "visible part" of the electro-
magnetic spectrum and its vicinity into the UV and near IR region an optimum
number of four to five spectral channels assuming a bandwidth of 0.5 - 1.0 um,
has been obtained. The decrease of the classification quality by using an
extended number of spectral channels results from the partly high linear de-
pendence of the signals for a scene, combined with statistically occuring
systematic errors in the signal of either one of the highly correlated channels.
The statement should not be confused to the conditions of a case, where the
spectral intensities of only a certain number of different surfaces are highly
correlated. In these cases the high correlation coefficients do not have any
influence to the decrease of the classification quality because the different
surfaces can easily be discriminated by their average spectral radiation inten-
sities.
For those cases where the spectral intensities are highly correlated for a
certain surface either one of these channels does not contribute new informa-
tion for the explanation of physical and chemical effects that take place.
Statistically occuring systematic errors are then considered as significant
information. This is particularly evident if a linear transformation is main-
tained for the purpose of data reduction and data compression [1]. Thus the