4. SEPARABILITY OF CLASSES 
As presented in chapter 3, microwave signatures of 
different classes vary within time, depending on the 
environmental conditions like changes in vegetation 
cover, vegetation and soil moisture, freezing and snow. 
The aim of this chapter is to study the separability of 
different classes under these conditions. 
Class separability was measured using Bhattacharyya 
distance. It is a probabilistic distance between two 
classes. There are also many other measures for class 
separability, but Bhattacharyya distance is important 
because it is related to upper and lower bounds of Bayes 
error E'. In our case, classes were supposed to be 
normally distributed, so Bhattacharyya distance is 
defined by using class means and covariances (Devivjer, 
1982). Then Bhattacharyya distance was transformed so 
that the range of distance would be between 0 and 2, 
latter meaning perfect separability. It should be noted 
that Bhattacharyya distance is not being used to predict 
actual classification accuracies, but to rank different 
alternatives with their ability to separate different 
classes. 
Figure 3 represents the mean class separabilities of 
different images. Solid line represents the mean class 
separabilities computed from the original images and 
dashed line from median the filtered images. It is 
evident that median filtering increases class 
separability, and this increase is larger for images taken 
in snow-free and unfrozen conditions. The mean class 
separabilities are largest for Jers, summer and late 
autumn ERS-1, snow-free and unfrozen Radarsat and 
frozen Radarsat images. The separabilities are smallest 
for winter and very rainy ERS-1 images, so ground 
freezing and heavy rain decrease the suitability of ERS-1 
images for land-use classification. The smaller 
separability of Radarsat image with unfrozen ground 
  
Mean class separabilites 
T 
  
o 
o 
Bhattacharrya distance 
‘ 
\ 
  
  
  
eap t] V jf can 
Ni uae bie V zm 
A k A 1 1 A n i 
2 4 6 8 10 12 14 16 18 20 
Number of image 
  
  
  
  
Figure 3: The temporal variation of the mean of 
the class separabilities. Solid line means that 
class separabilities are computed from original 
images and dashed line from median filtered 
images. 
and snow (image 18) compared to other Radarsat images 
is probably due to steeper incedence angle. 
The average separability of class water from other 
classes was good. The best separabilities were achieved 
in Jers, late autumn ERS-1 and Radarsat images and 
the worst in the winter or heavy rain ERS-1 images. The 
average separability of class agricultural field from other 
classes was moderate. The best separabilities were 
achieved in summer and heavy rain ERS-1, and frozen 
and old-snow ERS-1, Radarsat and Jers images. The 
worst separabilities were achieved in early winter ERS-1 
images. The average separability of class urban area 
from other classes was poor. The best separabilities were 
achieved in winter Jers, late autumn ERS-1 and 
Radarsat images and the worst in the winter or heavy 
rain ERS-1 images. The average separabilities of class 
mire and forest classes from other classes were also 
poor. The best separabilities were achieved in Jers, 
summer and late autumn ERS-1, snowfree and old-snow 
Radarsat images. The worst separabilities were achieved 
in the winter or heavy rain ERS-1 images. 
The best pairwise separabilities and corresponding 
image are summarized in Table 2, when the 
separabilities were computed from the original images 
(upper right triangle) and the median filtered images 
(lower left triangle). Also here it is evident that median 
filtering increases class separability, except the 
separability of urban area from forest classes. Pairwise 
separabilities are largest in Jers and Radarsat images. 
A very good or good separability can be achieved 
between water and other classes and between 
agricultural field and forest classes with high stem 
volume. Otherwise separabilities vary from moderate 
(between agricultural field and urban area) to very poor. 
Usually, the pairwise separabilities were largest in Jers 
(images 15 and 16), Radarsat (mainly in images 17 and 
20) and late autumn ERS-1 images (images 5 and 6). 
There were some exceptions. The pairwise separabilities 
between agricultural field and forest classes, and 
agricultural field and mire are larger in heavy rain or 
winter ERS-1 images. Also, the steeper incidence angle 
in Radarsat image (image 18) increases the pairwise 
separability between mire and some other classes like 
agricultural field, urban area and forest with low stem 
volume. 
5. IMAGE PROCESSING 
In order to decrease the effect of speckle images were 
filtered using median filter. Median filter filters image 
so that the center pixel of the filter window has been 
replaced by the median of the pixel values in window 
(Richards, 1993). In other words, this filter removes low- 
and high-valued pixels from image. 
Feature extraction was performed by using principal 
component analysis (PCA). It is a linear transformation 
which rotates coordinate axis according to variance of 
data. First principal component axis is placed according 
568 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
  
Tabl 
Jers- 
Nr 
  
  
ee lee | = [8 Le 1 | 
re 
E 
  
  
  
  
  
  
Table 
origin. 
Cla: 
1 
S |@ |O: |» [o2 | to