2eferential 
supervised 
Difference 
cies. 
  
The number of the p(x/w,) will be equal to the number 
of the ground cover classes. This means, for a pixel at 
aposition x in multispectral space a set of probabilities 
can be computed for each class. The required p(w/Xx) 
of the class and the available p(x/w) of training data 
are related by the Bay's Theorem as follows: 
p(wy/x) = p(x/iw) p(w)! p(x) 
where p(wj is the probability that class w; occurs in 
the image. The rule in classifying a pixel at a position 
x will be: 
x 5 wif pew) pw) > p/w) pw) 
forall j zi 
As we mentioned above this method is carried out in 
two stages, Figure 2, in the first stage and as a result to 
applying the maximum likelihood classification 
mentioned above, an index related to the measured 
spectral content is assigned to each pixel. Then in the 
second stage each pixel at a time is taken along with its 
coordinate and the spectral index which was derived in 
the first stage, then this index is led to the pixel data 
base and checked against the indices of a pixel with the 
same coordinates there. If the check result is true this 
means that the index derived in the first stage is the real 
spectral content of the classified pixel and classification 
decision will be confirmed. On the other hand if the 
check result is false this points to conflicting pixels and 
means one of two things: the characters of the 
conflicting pixel have been changed after constructing 
the data base, or the criteria established for the 
classification in the first stage is not as accurate as 
required. In both cases the final decision on classifying 
such pixels is given to the user where he can adopt the 
classification results though they are not in line with the 
information from the data base or he can change the 
classification criteria. 
The result of applying this method on classifying a test 
site image resulted in a good improvement in the image 
quality and its general appearance. Comparison of the 
accuracy results between the ordinary and referential 
classification (table 1) shows that there is 0%, 7%, 
13%, 16% and 25% when classifying poplar, water, 
chestnut, forest, grass, water bodies and sugar beet 
respectively, and the average accuracy in classifying all 
the studied classes has been improved by 17.2%. 
Further comparison between the accuracies of both 
classification procedures in figure 3 shows that the 
accuracy of the referential classification is always higher 
than that of ordinary classification (supervised), the 
worst case of the referential classification accuracy 
happens when the data base does not contain any pre- 
collected information about pixels required to be 
classified, even in this worst case the referential 
classification is just as accurate as ordinary classification 
and is never less as in the poplar class case. 
3. CONCLUSIONS 
Referential classification of image data is a vital step 
toward automating the classification process which is an 
important step in automating the whole image processing 
and analysis process. This can be achieved by cancelling 
the role of the user in classifying the conflicting pixels 
where the spectral data allocated to them in the ordinary 
classification stage can be adopted, changing their 
classification to ordinary and not referential or they can 
be rejected and reported as unknown pixels; a multifold 
process which involves three steps that reflect high level 
of expert and artificial intelligent behaviour. Reporting 
about rejected and unknown pixels could be a highly 
advantageous feature of this method especially when 
using multitemporal images in constructing the data base 
about one area and using another image of later time 
about the same area in the referential classification. 
Then all rejected pixels in the referential classification 
could denote possible change in the area between the 
time of constructing the data base and the present date 
of classification. However, the referential classification 
method has slightly disadvantageous features such as the 
fact that constructing a really indicative data base about 
individual pixels is not an easy task, though not 
impossible; also the algorithm for performing this 
method is more complicated and needs more time and 
more efficient hardware to be executed. 
REFERENCES 
Alhusain, O., 1992. Studies in the Design and 
Implementation of Microcomputer-based Satellite image 
Processing System. Ph.D. Thesis, Technical University 
of Budapest, Budapest, Hungary. 
Baxes, G.A., 1985. Vision and the Computer- An 
Overview. Robotics Age, March, pp. 12-19. 
Hunt, E., 1984. Digital Image Processing- Overview 
and Areas of Applications. Siemens Forsch.- u. 
Entwickl. - Ber. Bd. 12, pp. 250-257. 
Richards, J.A., 1986. Remote Sensing Digital Image 
Analysis. Springer Verlag, p. 281. Berlin [etc.], 
Germany. 
Swain, P.H., 1978. Fundamentals of Pattern 
Recognition in Remote Sensing. In: Swain and Davis 
(eds.). Remote Sensing: The Quantitative Approach, 
pp.136-187. McGraw-Hill, NY, USA. 
335