i 
| 
i 
I 
| 
| 
| 
| 
i 
" 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
of general classes, the class hierarchy was extended to parent 
and children classes. First, all objects were classified to parent 
based on their existence in supper object classes. Then in 
corrector classes they were belonged to their correct classes. In 
the lowest level, those objects, which were reclassified to their 
correct classes, were belonged to main types by logical" or" 
term. In order to extract special types from main types, those 
were subdivided into special type's classes and were classified 
by suitable descriptions. In addition to spectral information, the 
aspect and elevation attributes were used to separate pure Fagus 
from mixed Fagus. 
subdivided into their special types and were classified to each 
class by aspect and elevation as well as spectral attributes. 
  
  
  
  
  
  
    
    
        
  
  
  
Coarse segmentation of 
channels by scale 
parameter and suitable 
homogeneity criterion on 
ETM+ data and creating 
Middle segmentation of 
channels by scale 
parameter and suite 
homogeneity criterion on 
ETM+ data and creating 
Fine segmentation of 
channels by scale 
parameter and 
homogeneity criterion 
on ETM+ data and 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
     
   
  
  
  
class heirarchy class heirarchy creating class heirarchy 
at l- AL 
Classification of Rectification of Creating general classes 
general classes of — || Ur d — through belonging of 
fa Class- | fagus,carpinus and || Class 
gus, carpinus, related shadow with related correct classes to each 
replantation, mixed Kay? : 
p hy nix ; featur | corrector dasses and | feature class by “or”, Extraction 
road and shadow with rJ reclassification with ]  ofpurefagus pure 
suitable expr essions for : : carpinus and mixed 
each lass suit expressions 
  
  
  
  
  
  
  
  
  
  
Fine segmentation to Coarse Medium Fine segmentation 
extract the road, segmentation on segmentation to on channels 
shadow and others the channels and rectification of — t 
by creating class es generaltypes 
Membership function hierarchy for Reconstruction of 
And classification gemeraltvmes —— general types by 
c- re b : Mun 
gicalterm “or 
EU Nearest Diiding of general = Cr 
Chssificatio n- neighbour types to corrector Chssificati 
based classification on chsses and S On ot 
segmenta on and generaltypes and chssification b y Sechl types by 
chssification useofrehtional | | suitable description | | P. 73b aspectand 
elevation attrib utes 
a = = 
E A A 
£ A n 
^ it 1^. 
»- 3 ' 
e 2 a Gt 
rl LP Ex 
a a . 
Ub, d Pid 
Leva? ^ CMM | Leal 
  
  
  
  
  
  
TT IT ST 
  
  
  
d 7 
^ - ku 
#- : ^x 
4 "s ; X ya 
4 EPL vs PE 
T, , * $ f: 4 * # 
s zy HM ; id : 
  
  
  
  
  
  
  
  
  
  
  
  
Figure 4: Flowchart of hierarchical forest types classification by 
membership function method on ETM+ bands and their results 
in each level 
4.2.3 Integration of Nearest Neighbour and 
Membership Function 
The last experiences were delineated that each of the nearest 
neighbour or membership function methods has advantages and 
disadvantages. In the nearest neighbour method, the spectral 
attributes can be only participated to classification. In the 
membership function method to get the best result, user should 
be find the best description for each class as well as should use 
more descriptions to classify all objects in different levels. In 
order to overcome these disadvantages, integration of both 
methods is offered for classification. Four multiresolution 
segmentations have been done for hierarchical classification of 
types (figure 5). 
First, the roads and shadows were extracted by suitable 
descriptions on a fine segmentation and the segmentation based 
on classification was used to merge objects of each class. In 
next step, high-resolution segmentation (level 3) was only done 
on the “others “merged section of image. They were classified 
by training samples on main classes (Fagus, Carpinus, mixed 
and replantation) in the nearest neighbour method. The fine 
objects, which belonged to road and shadows classes were 
classified by membership function. In the next level (level 2), 
these main classes were subdivided into corrector classes and 
objects in each class were belonged to their correct classes. 
Also, the shadow objects were reclassified to class of their 
neighbour, which had more relation border. The all corrected 
classes were belonged to their correct classes by use of logical 
“or” term in the lowest level. Also the main forest types were 
Figure 5: Flowchart of hierarchical forest types classification by 
integration nearest Neighbour & Membership function and 
results in each level. 
4.3 Accuracy assessment 
The results of the pixel-based and the object-oriented 
classification of the ETM+ images were compared, using a 
sample ground truth map that has already been generated in 
other project. The accuracy assessment of results was done in 
PCI programme. A confusion matrix was built to obtain all 
accuracy indices (see table 3) for assessing of quality of whole 
classification and each class. 
  
  
  
  
  
  
  
  
  
  
  
User Producer In class 
Forest type 
accuracy accuracy accuracy 
Pure Fagus 0.2647 0.2920 0.1922 
Mixed Fagus 0.6939 0.6341 0.9821 
Pure Carpinus 0.2059 0.4200 0.1909 
Mixed Carpinus 0.6162 0.7085 0.9669 
Mixed Alnus 1.0000 0.7200 0.5714 
dide 5 0.6550 0.4986 0.6525 
Replantation 0.8873 0.4300 0.6885 
Overall accuracy: 60.65 Kappa accuracy: 44.40 
  
  
  
  
Table 3: An accuracy assessment table obtained by confusion 
matrix for result of integration of both methods 
5 DISCUSION AND CONCLUSION 
The results of accuracy assessment showed that the object- 
oriented techniques could classify forest types better than the 
pixel based classification method. The kappa index was about 
20 % more in the object based method (see tables 4, 5 and 
figure 6). 
As expected, the results of pixel-based classification had less 
accuracy in comparison with the object oriented classification 
(overall accuracy was 43.7%). In the pixel-based classification 
due to incorporating of only pixels spectral attributes in 
classification of images, the results looks like salt-peppery 
picture (see figure 6A). In other hand, in the homogenous area 
as forest type some pixels had different reflectance with their 
niehbour pixels so that they classified to other classes. 
However, in the results of the object-oriented classification, 
those are assumed as a homogenous area and an object. 
1108 
  
  
Fig 
and 
salt 
Am 
extr 
men 
fore