= 
- ut 
  
r of 
by 
tand 
»utes 
  
  
  
i 
1 
1 
1 
| 
  
  
ms“ up SV 
i.e- PM uu co — - "MIS 
International Archives of the Photogrammetry, R 
        
In addition, it was investigated that adding artificial processed 
Bands facilitate improving the results in both approaches. Use 
of suitable processing bands such as those were used in the 
pixel-based classification together with main ETM+ bands 
could slightly improve the results (3%). These additional bands 
contained useful information, which have increased the 
seprabilty of forest types in the feature Space as well as have 
reduced images errors. In addition, in the Object oriented 
method, use of artificial bands in segmentation as well as in 
classification as suitable descriptions could improve the results 
(by 2.5 to 5.5 94), 
  
      
   
Pixel 
Object oriented techniques 
based 
    
  
     
  
  
  
    
  
   
method Membership Integration 
function technigue 
  
   
  
Overall 
accuracy 
(76) 
Kappa 
(76) 
58.5 
    
    
  
  
  
Table 4: comparison of classification results of pixel-based and 
objects oriented methods by main ETM+ bands 
  
   
   
   
  
  
Pixel Object oriented techniques 
based 
Nearest 
method | Membership 
function Neighbour 
i 
Table 5: comparison of classification results of pixel-based and 
objects oriented methods by main ETM- and the best bands 
   
   
  
  
   
  
    
  
Integration 
technique 
   
   
Overall 
accuracy 
(76) 
Kappa 
(70) 
60.7 
    
  
44.4 
  
  
In the object-oriented approach, image is segmented to different 
objects. Before segmentation at satellite images, the image 
objects are heterogeneous due to diversity of spectral 
information in pixels. After segmentation, heterogeneity of 
image objects will reduce. It led to easier seperabilty of objects, 
due to increasing of contrast between them. See figure 6B. 
  
  
  
  
  
  
  
  
Figure 6: The pixel-based classification in a masked area (6A) 
and the object-oriented classification in the same area (6B). See 
Salt-peppery image at the pixel-based method. 
Among three objects oriented techniques, which were used to 
extract the forest types, integration of the nearest neighbor and 
membership function method showed high capability to stratify 
forest types. 
1109 
emote Sensing and Spatial Information Sciences, Vol XXXV ; 
Part B7. Istanbul 2004 
  
In the membership function, determining of the suitable feature 
that exactly separates types was very difficult. Optical 
interpretation of images and their attributes could help to find 
the best descriptions. However, its result quality was low among 
other object-oriented methods (48.8 and 51.3 % in the both data 
set respectively). 
As the results of pixel based classification and separabilty 
assessment showed, the forest types could not be completely 
separated by a few features. In the other hand, the nearest 
neighbor method classified the forest types in the multi-feature 
Spaces. It caused better result than the membership function 
method (57.2 %). 
By using of both the nearest neighbor and the membership 
function in the classification integrately, the overall accuracy of 
result increased to 60.7 94. Whereas, use of obtained 
information from training objects to define suitable descriptions 
of classes and use of membership function to re-correct classes 
Were reasons of improvement. 
Increasing of the kappa by 255 y, in the pixel based 
classification to 44.4 % in the object oriented approaches 
showed the capability of multi-resolution segmentation of data 
to provide other useful attributes in addition to Spectral 
information as well as reducing of heterogeneity in image 
objects. 
The results show that since forest type’s signatures have high 
overlaps on every band individually, they can not be separated 
by a few parameters (feature spaces). As the result, it js 
recommended to use the nearest neighbor method at first step 
and then membership to refine classification of images. 
However, there are some reasons for low accuracy in both 
methods contrary to other image classifications such as land 
cover or land use. First, it refers to significant overlaps in the 
spectral attributes between the most of the mixed type and the 
mixed Fagus as well as the mixed Carpinus in some places in 
the study area. In addition, similarity Of spectral attributes in the 
pure Fagus with the mixed Fagus types caused that they could 
not be completely separated well (see table 3). Second, the 
effect of topography and different illumination at the different 
aspects in the study caused that a similar type reflected different 
Spectral attributes. 
Although the considerable result have got by the object oriented 
methods in compare with the pixel-based classification method, 
but recognition of heterogeneous objects in forest area 
especially in the hardwood forests was difficult because of 
mixed species and also contrast of objects borders was low. 
Use of different data in terms of resolution and Spectral 
information can be examined to extract the forest types and 
certificate the results by the object-oriented approaches in 
future. Integrating of ancillary data in corporation with satellite 
data is expected to improve results 
ACKNOWLEDGEMENTS 
This study was performed at the remote sensing laboratory of 
department of geography, Zurich University and has been 
funded by Tehran University and Zurich University. We would 
like to thank the head of group professor Claus Itten for his 
supports in this study.