773 
G i . = four-dimensional spectral vector corresponding to a single 
ground resolution element and defined by digital counts of 
pixel P. .in the four spectral bands: 
1 , J 
G. . = 
Any meaningful statistical description of the whole ERTS scene by a 
single matrix is impossible because of multitude and variability of classes 
present in typical satellite imagery. However, the 185 km x 185 km ground 
area of the ERTS scene can always be partitioned into classes exhibiting 
common features which can be described by measures of central tendency and 
dispersion. 
For example, Class C, representing coniferous forest, is described 
by mean spectral vector G^, and covariance matrix W^, if dispersion of spectral 
vectors G. . within Class C is due to random factors that is, if Class C 
i,l 
exhibits multivariate gaussian (normal) distribution of pixel values. This 
assumption, if true, greatly simplifies the scene classification (Swain, 
1972). 
Since the size and location of Class C within the ERTS scene is 
generally unknown (its delineation being our objective), the Class mean 
spectral vector and covariance matrix must be estimated from pixels located in 
training areas only. Hence, as was stated before, any error made in selection 
of training areas will directly effect the Class statistical definition and 
accuracy of its classification. 
The mean spectral vector G^. of Class C is defined by arithmetic 
means of digital counts of pixels located within the class training area in 
each spectral band: