t 
i- 
r 
res 
the 
and 
nt 
st 
ate 
ight 
tral 
ced 
ines 
itly 
Thus, the test avlues resulting from S, can be directly compared with each 
other. Their position-dependent variation follows from the analysis of 
several test areas of the same size. 
The signal | emitted by each sensor element results from the superposition 
of the object-dependent signal ES and its distortion lg - 
zd. t 
o S 
For analysis of j^ it is suitable to take | as constant, i.e. to start 
from areas of identical land cover which caf be expected to have constant 
properties. 
As land areas of the required dimension and homogeneity are hard to find, 
the following investigations refer to the area of the North Sea. 
2. Test Areas 
The entire available data set encompasses 4 channels of LANDSAT-2 with 1100 
scan lines of 1100 pixels each. The overall scene is subdivided in sub- 
matrices (test areas) of equal size, seamlessly adjacent to each other in 
both the scan and flight direction. The following area sizes were selected 
for the 3 computing runs made: 
  
  
run no. of pixels in scan 
scan direction lines 
1 100 1080 
2 100 96 
3 50 48 
  
  
  
The following graphs are based solely on the results of the first strip in 
scan direction,which consists of up to 22 subareas of the size defined 
above (cf. fig. 1). 
3. Theoretical Basis 
The discrimination of different land cover types which are defined by dif- 
ferent n-spectral features is fundamentally dependent on the shape, the 
orientation, as well as the distance of their n-dimensional error ellip- 
soids (clusters ). 
Their form and orientation are defined by the length of the semimajor axes 
(eigenvalues of the Q-matrix) and the angles (resp. their cosines - eigen- 
vectors of the Q-matrix) which they form with the n coordinate axes [1]. 
The intersection set of two or more interpenetrating error ellipsoids will 
probably always be problematic as to the assignment of its elements to the 
correct!" class. As their n-dimensional extension, and thus the risk of 
misclassification, is essentially determined by the above-mentioned ellip- 
soid parameters, while it is still largely unknown whether, and to what 
degree, these are dependent on position and size of the test area, they 
will be studied in the following on a larger area of a single land cover 
173