The conjunctive fusion of the four spectral bands MSS4 to MSS7 leads to the rates of classification and the image presented table 18.

Unclassified pixels: 1.21%

 Rate of correct classification

This fusion improves the rate of correct classification, compared to the isolated classification carried out on each spectral bands, but also compared to the scores obtained by applying Bayes' rule. In effect, the rate passes from 40.90% (better rate obtained by spectral band) to 53.35%, i.e. a profit of 12.45 points, and from 45.52% obtained by Bayes' rule to 53.35%.

Class 3 benefits particularly from fusion since it passes from 27.78% (better rate given by MSS6) to 39.54%, i.e. a profit of 11.76 points.

Class 4 also, passing from 40.66% (rate of MSS5) to 52.17%, i.e. a profit of 11.51 points.

Only classes 1 and 9 decrease, passing from 56.21% (MSS4) to 32.90% for class 1, and from 46.41% (MSS5) to 43.14%. However, for these two classes, the rates obtained by fusion are better than those of the spectral bands taken separately.

 Inter-classes confusion rates

The improvement of classification is also measured by the decreasing of the inter-classes confusion rates. Table 19 shows the rates of confusion obtained after the fusion of the spectral bands. It is noted that they are generally weaker than those obtained during the classifications carried out separately for each spectral band.

The fusion of information shows its utility here, since sharing the information produced by each band makes it possible to obtain a better result than the classifications carried out independently for each class. The various information available overlap, and are reinforced mutually because the sources agree towards a same result. They are in agreement the ones with the others.

 Unclassified Pixels

The number of unclassified pixels increases. On the whole image, 3 174 pixel have not been classified, i.e. 1.21% of the total number of pixels of the image (for the spectral band, the higher rate of unclassified pixels is of 0.24% for MSS4). This is due to the effect of the severe behaviour of conjunctive fusion. To classify a pixel, all the sources must support together at least one class, assumed plausible for this pixel. The operator min attributes a membership degree of zero to each class and the pixel is then not classified.

It is also due to the difference of the data handled between

• the definition given for the conjunctive fusion which fuses possibility distributions, and
• the fusion of classification information which is a fusion of scalar numerical values.

It is of course more difficult to obtain a nonempty intersection with simple numerical values (if only one of the values is null, the intersection is regarded as being empty), than for distributions.

Table 19: Matrix of inter-classes confusion in percentages relatively to the classification after conjunctive fusion of spectral bands MSS4 to MSS7.

	Class observed
Class expected	1	2	3	4	5	6	7	8	9	Unclassified
1	32.90	7.41	49.02	6.10	2.18	1.53	-	-	-	0.87
2	5.66	50.11	15.03	13.73	14.38	1.09	-	-	-	-
3	11.76	6.54	39.54	7.52	19.61	0.98	9.48	-	2.29	2.29
4	4.09	3.58	25.83	52.17	11.76	2.56	-	-	-	-
5	0.22	2.40	2.83	4.79	58.39	16.56	9.80	-	5.01	-
6	-	-	9.59	5.45	17.43	64.05	0.87	-	2.61	-
7	-	-	1.09	-	6.97	0.44	70.37	1.31	16.78	3.05
8	-	-	-	-	-	-	4.36	64.71	26.58	4.36
9	-	-	-	-	14.81	0.22	34.86	6.97	43.14	-