nbul 2004 
eee 
red result 
d one lake 
'er is not 
right blue) 
t context 
1Sses) and 
areas are 
ellow. 
ase from 
= 
ipted 
  
  
% 
  
).8 
  
of the rule 
re listed in 
transferred 
ion of Jijel 
' base an 
A and OK 
'e accuracy 
e first is the 
is from the 
from the 
Figure 4 is 
arts the OK 
he original, 
r example, 
increases to 
ion phases, 
: OK of the 
ile sets and 
spite of the 
it still falls 
ssifications. 
age OA for 
1ses to 82% 
the average 
n results for 
. Except for 
ve the best 
ed, adapted, 
e accuracies 
built-up and 
.appa values 
ly, they are 
delivers the 
0.39 for the 
e individual 
0.42. Roads 
ial geometry 
[ 5m is nol 
object-based 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
Analysing all of the above results, it can be noticed that the 
adaptation and adding of new rules to the base rule set do not 
improve the classification result significantly. in contrast to the 
transferability test where the adaptation of the rule base 
enhanced the accuracy considerably. The main difference 
between these two studies lies in the used object properties. The 
rule bases in the transferability test were optimised to one 
certain image and then transferred. In contrast, the base rule set 
comprises object properties which should be stable in several 
images of one specific geographical 
region. Therefore 
adaptation brings only small improvement. 
  
  
  
Overall Kappa values for the 4 coast test sites 
  
  
I= 
11| © Transferred base 
rule set 
  
  
||| Adapted base 
  
    
  
  
" pret le set 
L| rule se 
a | ese 
2 i 
x - || 0 Added rules to 
= || the base rule set 
i 
© : ; 
> E I il 
à i| D Original, 
|| individual 
  
[.i| eclassification 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
C 
c 
S Y 
  
  
  
  
Figure 4: Overall Kappa values for the 4 test sites Coast (four 
base classes) 
  
  
  
  
  
  
  
  
  
  
  
  
  
    
  
  
  
  
  
  
  
  
  
  
  
  
Overall Accuracy for the 4 coast test sites 
90% += | i 
BD% | | f 
% [eed [= | @ Transferred base 
|| rule set 
% + L—] EA 4 
70% +0 ; F] W Adapted base 
| rule set 
= i j 
> 60% + Lx] E F3 
$ {| O Added rules to 
8 E || the base rule set 
< 50% ++ pr Ft; Fi 
t | D Onginal, 
$ || individual 
> % 4 E I EL = A 
3 40% + | |L. classification | 
i 
30% T E fea] = 
{ 
20% + pd E wi 
7 | 
10% +4 L.T E] =) 
0% He A AL | Ly 
Algiers Mostaganem Jijel Mandoura Average 
  
  
  
Figure 5: Overall Accuracy for the 4 test sites Coast (four base 
classes) 
Transferring the base rule set to a new test site decreases the OK 
by 20% and the OA by 11% (four base classes). After adding 
new rules, the OK decreases by 15% and the OA by 894 in 
comparison to an individual classification which usually 
requires much more time. The investigations with the base rule 
Set show the feasibility of transferring knowledge-based 
Classification rules but the gain of automation and time for 
(Mage classification leads to a certain loss of accuracy. 
  
  
  
  
  
  
  
  
  
     
  
  
    
  
  
  
   
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Average Kappa values for individual classes 
2. | 
0,9 E i 
0,8 , 
| 
0,7 [i 
M | 
| 
: | | @ Transferred base 
eo V6 [i| rue cel 
= Í 
ë | 
> | 
s 0,5 E: 
& 
2 | 
S ay || D Added rules to 
04 = 3 |—| the base rule set 
i | D Onginal, 
0.3 4 he |-0| individual 
! P EM | classification 
02 JJ |i L- 
, i : i 1 
| 1 | 
0.1 +4 .—LI i 
n i 
0,0 | z a | 
Roads Water Non-Built up Built-Up 
  
  
  
Figure 6: Average Kappa values for the individual classes 
(coast, four base classes) 
5.3 Discussion 
When investigating the transferability of classification rules 
potential influencing factors have to be considered. The 
following items may influence the transfer of knowledge-based 
classification rules: 
- Date of image acquisition (season, sun elevation) 
- Relief (shadow) 
- . Atmospheric impacts 
- Geographic region (occurrence and appearance of 
objects) 
It is possible to decrease the impact of the listed factors. One 
option could be by limiting the proposed knowledge-base to a 
specific season and/or to a specific geographic region. In this 
work the rule base was limited to a specific geographic region. 
Nevertheless feature classes may have different appearances and 
new feature classes might appear. Atmospheric impacts and 
shadows due to the relief can be reduced by applying 
atmospheric and  topographic corrections. Because of 
unavailable meteorological data radiometrically uncorrected 
data was used here and it was a challenge to test the 
transferability of rule bases for this kind of data. 
The classification rules can use radiometric, texture, and shape 
properties of the image objects as well as relations between 
different image objects. Radiometric properties are strongly 
influenced by atmospheric impacts, season, and sun elevation. 
Texture is a more stable feature because it represents the 
structure of the pixels grey values and not absolute intensity 
values. Especially for man-made objects like roads shape 
features play an important role and the investigations with the 
base rule set show that they are transferable. Descriptions of 
context between image objects have to be universal to be 
transferable. 
6. SUMMARY AND CONCLUSION 
The transferability of classification rule sets for a specific 
geographic region is a challenging task. Topographic features 
appear in a broad variety and illumination effects additionally 
complicate finding stable and transferable properties for image 
classification. After the introduction of the object-based 
1063