or 
ntrary 
the 
low. 
The 
m old 
o the 
ind 
ilua- 
recommended. A more sophisticated problem is the extraction of non-trivial 
factors. À statistical approach to determine the number of non-trivial 
factors exists for a small number of procedures only and may lead to the 
extraction of factors of trivial importance. À simple, but very relaible 
method for quick analysis is the determination of the per cent of variance 
extracted by a reduced number of factors. À cumulative per cent of 75 - 85 
of variance is considered as efficient. Usually the extraction of factor 
loadings is stopped, when the next factor to be extracted contributes a much 
smaller percentage of variance than the preceding one. This point can easily 
be recognized by plotting the eigenvalues of the principal component analysis 
or the variance of the factors vs. the factor sequence number. The basic idea 
behind this test is, that most of the variables measure a limited number of 
factors well and a large number of (error) factors less well. This test fits 
particularly to the requirements of the principal component analysis, where 
the factors are extracted in the sequence of their substantivity. The inter- 
pretation is becoming complicated, when several breaks occur in the factor 
variance plot. If this break occurs toward factors of higher order, the drop 
can be ignored, because extracting the same numbers of factors as there are 
variables, is of no interest and of no use. The problem is more serious, if 
(two) breaks occur in the first half of the number of factors. Here other 
criteria have to be applied for the extraction of the factors. For the use, 
when no obvious break exists, a solution can still be obtained by determining 
the point where the best variance contribution no larger forms or straight 
line. The exact number can, however, only be estimated. The correlation matrix 
can still be significant, as random eigenvalues do not necessarily have signi- 
ficant breaks. 
5. Rotation of Coordinate Systems 
  
As from the plots of the factor loadings in the coordinate system of the 
factor pattern space can be observed, there is a clustering of factors re- 
presenting items of uniform properties of the variables involved. The factor 
loadings are very often difficult to interprete, as the solution obtained is 
influenced by the special data properties. This discrepancy in the case of 
multispectral data evaluation has its origin in