m 
ree m ee 
REEL -- 
; and 
nts 
unctions 
. During. 
unction 
on ly 
s per 
M 
  
3.2 Determination of polynomial coefficients 
The polynomial coefficients for classification can bederived in different ways. One 
possibility would be to apply coefficients determined with other data sets to the new data 
set using appropriate signature extension methods. Another procedure estimates the 
direction dependance in the scene to be classified, e.g. with unsupervised methods or 
using training areas, the method applied here. Using this approach training areas should 
be regularly distributed over the strip without greater gaps to make the polynomials reliable. 
Furthermore the training fields should be marked as long (in flight direction) and narrow 
(perpendicular to the flight direction) areas. 
Mean values are calculated for each field and sorted according to scan angle. 
Trendcorrected covariance matrices in different scan angle ranges are also estimated. The 
polynomial coefficientsfor means and trendcorrected covariance matrices are then 
determined by least squares adjustment. 
3.3 Results 
To test this modified classification algorithm the same scanner image as analyzed in 
chapter 2 (Hartheim, 4000 m) was classified with three different statistical sets. These 
statistics differ in the mean vectors but use all the same constant, not trendcorrected 
covariance matrices. In all cases 8 classes (bare soil, grass, winterwheat, coniferous forest, 
deciduous forest, Rhine, channel, lake) were classified with three channels (3,6,9). 
Table 1 shows the differences in the number of training areas used in the three versions. 
191 
EVI Meo quum MD SPADA Er 
(& : sia :