nto 
lent 
orm 
xels 
the 
f all 
ter, 
‚bil- 
um 
del 
fol- 
the 
the 
nent 
fea- 
; en- 
f (or 
pixel 
1not 
s its 
s of 
shed 
nta- 
nta- 
sure 
one 
pro- 
ging 
nax- 
1erg- 
nent 
tion, 
ered, 
nerg- 
mber 
  
  
of segments or a threshold of the merging ratio 
can be used to terminate the merging iteration. 
A 512x512 NASA /JPL AirSAR C-band HH po- 
larisation intensity image, acquired from Kakadu 
National Park, Northern Territory, Australia, is 
shown in Fig. 1. The image shows different types 
of vegetation cover typically seen in the region. 
Fig. 2 and Fig. 3 show the initial and final seg- 
mentation, respectively. 
  
Figure 1: A 512 x 512 pixel NASA/JPL Air- 
SAR C-band HH polarisation intensity image, 
acquired in 1993. 
  
  
  
  
  
Figure 2: The initial segmentation. 
3. CLASSIFICATION 
Apart from bare and/or little vegetated soil, 
which appears as the darkest areas in the image, 
three types of vegetation cover also show in the 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
  
  
  
  
  
  
  
Figure 3: The final segmentation. 
image. They are wet Melaleuca, dry Melaleuca 
and mixed woodland, marked as I, II and III, 
respectively, which are indicated in the classified 
image shown in Fig. 6. The measured C-band 
HH data distributions for these areas are shown 
in Fig. 4. It can been seen that those distribu- 
tions are highly overlapped (especially between 
I and III). The means of these areas are close to 
each other while their standard deviations are 
very high. Obviously the classification based on 
such pixel-based measurements would be very 
poor, as thresholds to separate these clusters 
hardly exist when dealing with only pixel-based 
measurements. 
. After the image is segmented, uniform areas have 
been circled, and the task of classification is to 
classify segments rather than pixels. In the case 
of the single-channel image, the useful informa- 
tion in terms of classification for each segment 
may include the mean, standard deviation and 
texture characteristics. Details are discussed in 
the following. 
The mean is the most important and useful in- 
formation, because radar backscatter models in- 
dicate that the mean (the average of backscat- 
tering coefficient) are highly correlated to physi- 
cal parameters of the area being imaged, such as 
above ground biomass, surface roughness, mois- 
ture and so on. 
The standard deviation of SAR image data is 
generally highly correlated to the mean for uni- 
form areas. The ratio of the standard deviation 
to the mean (RSTM) for homogeneous areas, in 
a given image, is approximately constant, which 
depends only on the number of looks and the 
519