° Y T 
[ Pixel number = 27362 
of Mean = 0.2195 d 
| pM Std 58 
td = 0.08 
S [ E oua 
i 4 M T THÉ, - 
mur Ek. 4 
d et ] 
a + I a 
r oF Mis ] 
BET al : A Hh 
0.00 0.10 0.20 0.30 0.40 
Pixel value 
14 T T T 
12 kk Pixel number - 4894 -J 
- ya + Mean = 0.0946 3 
+ Std = 0.0429 3 
ud 
* 
A 
| 
  
  
+ Pixel number = 142940 |] 
# Mean = 0.1507 3 
+ Std = 0.0582 
b 
  
  
  
Figure 4: Distributions of the AirSAR C-band 
HH intensity data for three uniform areas. 
processing algorithms (Lee, 1986). Fig. 5 shows 
the relationship between means and standard de- 
viations for the largest segments occupying more 
than 9596 of the total pixel numbers of the im- 
age. It can been seen that the dispersion of the 
RSTM is indeed very small, and the RSTM can 
be considered as constant as indicated by the 
slope of the straight line in the figure. 
  
  
  
  
  
0.10 P 
0.08 - 
à - 4 
2 7 
= L 4 
* 0.06r ei 
$ - 1 
a 
© 
be r 4 
5 0.04[- z 
a - 4 
3 
un L 4 
0.02 [- ; - 
L Average ratio of std to mean: 0.421 4 
0.00 L à à ; 
0.05 0.10 0.15 0.20 0.25 . 
Mean backscattering coefficient (m-2/m-2) 
Figure 5: Relationships between means and 
standard deviations for segmented uniform ar- 
eas. 
The texture of a segment can be measured by the 
difference matrix D describing the differences 
Table 1: Texture features of the image: the av- 
eraged difference matrix 
  
0.583 0.568 0.550 0.569 0.586 
0.563 0.504 0.390 0.504 0.564 
0.553 0.443 0.421 0.443 0.550 
0.567 0.505 0.389 0.504 0.560 
0.588 0.570 0.551 0.565 0.581 
  
between the current pixel and its surrounding 
neighbours. The element dj; of D is defined as, 
1.1 Le 
di; = mr 2 Imo) 
—m(io + i, jo + 3)1°] 7? (4) 
where M is the pixel number of the segment, 
m is the mean of the segment, m(io, jo) and 
m(io + 1, jo + J), (4, 7) # (0,0), are the measure- 
ments of the current pixel and its neighbouring 
pixels. Therefore, d;;, (i,j) # (0,0), is the av- 
eraged difference of the measurements between 
an arbitrary pixel and the pixel separated from 
it by 4 rows and j columns, and this difference 
is normalised by the mean of the segment. The 
central element, doo denotes the RSTM: 
1/2 
) ] 1 
doo = — | 37 3m - m(io, jo)? (5) 
M 
It has been found that the difference matri- 
ces for different segments are very similar. Ta- 
ble 1 shows the average texture measures, D, 
of the largest segments occupying more than 
95% of the total pixel numbers of the image. 
The neighbourhood size chosen is 5 x 5 (4,3 — 
—2,—1,0,1,2 and (7,7) — (0,0)). It is worth 
noting two features of the table. First the ta- 
ble shows approximate symmetry with respect 
to the center, and second the further the neigh- 
bouring pixel is away from the center, the larger 
the value of the difference matrix element, in- 
dicating that the greater the separation of the 
pixels, the less they are similar. 
The dissimilarities of the difference matrices of 
segments can be assessed by the standard devi- 
ation of the element {d;;}. Table 2 shows the 
normalised standard deviation, i.e., the value of 
element (i, j) is c;;/c;;, where cj; is the standard 
deviation of {d;;}. It can be seen that all values 
in Table 2 are less than 10%, indicating that the 
dissimilarities of difference matrices for different 
segments are small. Therefore, although Table 1 
520 Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
  
p rn put + 
«e et N EAN "S S Cho Us A er er Cho — $9 04 pF m FT pum ums pee pie Of) (0 — ou 08 m AY [f — C (TO em bdo pm 0M IP md 
po