The sources used so far (spectral bands and out-image data on the geography of the area) are globally of identical reliability. Nevertheless the rate of confusion between the classes remains very high. Some classes are better recognized by certain sources than by others. However, it is difficult to determine if a source is clearly more reliable than another.

We thus used another source of information, highly reliable, to be able to implement and to test fusion by priority. This additional source is the result of the classification carried out by a neural network.

Mascarilla [Mascarilla, 1996] [Mascarilla and Desachy, 1996] defined a neural method to determine in an automatic way the most favorable context of each class. The phase of training of the neural network consists in applying rules, describing the geographical context of each class, to a set of representative samples.

The network provides the ``ideal'' card of distribution of the classes in the image, according to the geography of the area.

Table 37: Theoretical image of the ``ideal'' distribution of the classes.

Rates of classification obtained
Class	Number of pixels   correctly classified (A)	Number of pixels in samples      (B)	Rate of pixels     correctly classified (A/B)
1	458	459	99.78%
2	441	459	96.08%
3	296	306	96.73%
4	377	391	96.42%
5	365	459	79.52%
6	420	459	91.50%
7	459	459	100.00%
8	459	459	100.00%
9	459	459	100.00%
TOTAL	3734	3910	95.50%

Table 37 presents the ``ideal'' card of distribution of the classes in the image, according to the rules calculated by Mascarilla. The rates of correct classification are very good for each class, however this classification is too ``artificial'' as it does not take account of the real distributions of each class in the image. The information of the spectral bands makes it possible to add real information to this theoretical card.