The fusion of the four spectral bands MSS4 to MSS7 shows that Bayes' rule has rather a severe behaviour (of conjunctive type).

The final rate of classification of 45.52% is weak . That comes from the great confusion between the classes. It is noticed that classes 3, 4 and 9 are very badly recognized. Histograms of these classes are largely overlapped by those of the other classes.

However Bayes' rule selects, for each pixel, the classes achieving the strongest degree of probability. This method consisting in giving all the weight to the most probable classes is unfavourable for the recognition of the classes the least well defined.

But this rate is higher than the best rate of 40.90% obtained by the classification of band MSS4 only.

Classes 2, 6 and 8 largely benefit from the fusion of information:

• from 56.21% (better rate obtained by MSS4) to 87.15% for class 2,
• from 56,21% (MSS4) to 77.56% for class 6, and
• from 61.22% (MSS5) to 83.66% for class 8.

But it is not the case for all the classes. Classes 3, 4 and 9 have their rates recognition which fall down:

• from 27.78% (MSS6) to 3.92% for class 3,
• from 40.66% (MSS5) to 27.11% for class 4.

Those results are due to the choice of allotting all the probability of a source to the class which got the highest membership probability for a given pixel. As the number of sources (4) used is lower than the number of classes (9), classes which often get a rather high probability degree bring to them all the probability, to the detriment of classes which are less easy to recognize.

Because of the confusion which exists in the description of classes, classes with a more narrow histogram and with a well marked peak are chosen more frequently. In effect, this histogram often produces high measurements. It is the case for classes 1, 2, 6 and 8 .

On the other hand, classes with a more spread out histogram and with lower peak are difficult to recognize. It is the case for

• class 3 and 4 merged with class 1 and 2, and
• class 9 merged with class 7 and 8.

But these mitigated results are essentially due to the small number of fused sources compared to the classes to recognize. We will add out-image data to the spectral bands to test the behaviour of Bayes' rule when fusing a great number of sources (thirteen).

 

Unclassified pixels: 1.21%