International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
tl} ed Lt 
ised-ECHO performance and the feature-map 
   
  
E cud 
Figure 4. Ssuperv 
      
  
  
   
824 
Classification accuracy is dependent on both the classification 
algorithm and the training sample set, furthermore, it is slightly 
dependent on the window size. The performance of the object- 
feature is compared with the performance of the original pixel- 
features from the same scene, when the M.L. Bayes Gaussian 
decision rule is selected. 
Spectral information of surrounding pixels is correlated with the 
centred pixel under consideration. In object detection the 
spectral features of adjacent pixels are considered using 
neighbouring information; thus the object-feature which we 
represent them in this experiment only by (S <L) built upon 
both spectral and contextual information. Therefore, it is 
expected that the classification accuracy to be higher by using 
object-feature rather than the individual pixel-feature (notice 
that we did not consider effect of V in the classification of 
object-feature using M.L. decision rule). Tables | and 2 show, 
by using the object-feature, for example, the wheat field 
classified better than when the pixel-features are used for its 
classification. A test for robustness of the path hypothesis and 
accuracy of the unity relation shows that the functional based 
on path-hypothesis, can detect a single randomly selected pixel 
in a relatively large soybean field which is replaced by a pixel 
from some other ground cover types, see Figure 5. 
5. SUMMARY AND CONCLUSION 
In order to reduce data redundancy in multispectral imagery we 
have proposed a model, based on a scene object- description, 
for multispectral image representation. We have developed an 
on-line unsupervised object-feature extraction algorithm (called 
AMICA) which detects the objects by using the unity relation 
based on the path-hypothesis. The unity relation among the 
pixels of an object can be defined with regard to the: adjacency 
relation, spectral-feature and spatial-feature characteristics in an 
object. Based on the path-hypothesis the data read sequentially 
into the system. The unity relation between a current pixel and 
the path-segments (objects in the observation space) are 
examined, the current pixel may be merged into an appropriate 
object or it will initiate a new object. An object is represented 
by a relevant object-feature set. AMICA is implemented to real 
multispectral image data. The performance of the object- 
features is compared with the performance of the original pixel- 
feature. Three different evaluation strategies (overall 
misplacement error, feature classification performance and 
subjective object appearance) are selected for comparative 
feature evaluation using the pixel-features and the object- 
features. The experimental results indicate that data volume is 
reduced by a significant amount (the size of the feature-space 
for scene representation is reduced by a factor more than 25 
which is data dependent). In addition, the accuracy of 
information extracted from the object-features (as measured by 
classification accuracy) is greater than obtained when using the 
original pixel-features. 
The correlation among the adjacent pixels in the image data 
appears in the form of redundancy in the spectral-spatial 
features. Spectral information of surrounding pixels is 
correlated with the centred pixel under consideration. In object 
detection the spectral features of adjacent pixels are considered 
using neighbouring information. Therefore, it is expected that 
the classification accuracy to be higher by using object-feature 
rather than the individual pixel-feature. The improvement of the 
classification performance is consequence of incorporation of 
the spatial information in the object-feature extraction decision 
mt 
the 
the 
ob 
ob 
me 
Gl 
Ci