 |
 |
   |
|
|
|
Tables 24 and 25 show the result of the disjunctive fusion applied to the spectral bands and to those of the nine out-image data. The rates of classification obtained are very poor.
For the four spectral bands MSS4 to MSS7, the rate of classification reaches only 28.95%, that is to say about the half of the score of the conjunctive fusion of the same images (53.35%).
Adding the out-image data to the spectral bands falls disjunctive fusion to a score of 16.96%, against 22.89% for conjunctive fusion. In particular, classes 6, 7, 8 and 9 are not recognized, and class 3 is hardly recognized (1.96%)!
The difficulty is to take the final decision when the number of sources increases. By taking the maximum of the opinions expressed by each class, several classes obtain often the same degree of maximum possibility after fusion.
How shall we choose a class, when several classes are presented plausible for a same pixel? Unless using additional information to decide which is the ``good'' class, the decision is impossible. However, disjunctive fusion is implemented in such a way that, if several classes are considered to be possible for a same pixel, it is the first class among the candidates which is retained.
Thus, if a pixel is likely the most probable to belong to classes 3 and 5, with the same degree of possibility for these two classes, then it will be arbitrarily affected to class 3. This is why classes 6 to 9 are not recognized at all. Another solution to deal with this problem can consist in not classifying the pixel and in assigning it to a class of rejection.
Disjunctive fusion gives the priority to the source which attributes the strongest degree of membership of a pixel to a class. This method is very sensitive to the noise due to the sensors or the construction defects of the out-image data. Consequently, the choice of the class to affect the pixel is often modified by the noise of the sensors and the incomplete definition of the classes.
