To perform a more realistic estimation of the classification, it is necessary to have two groups of samples for each class:

• one of the groups is used to train the system to recognize the classes,
• the other group is used to test the classification achieved.

It is obvious that this method gives an estimation of the efficiency of the classifier closer to reality because tests are done on unknown areas.

In our application, each class is described by three samples (only two for class 3). It is a little number of samples. Moreover, these samples are not completely homogeneous following the radiometric feature. The small number of samples available for each class and the insufficient radiometric homogeneity of each class brought to divide each sample into two halves, instead of using two groups of samples for each class. A half is used to learn how to recognize the classes, other half is used to test the performances of classification.

Figure 14: Half-samples in spectral band MSS4.

Figure 14 presents, on the spectral band MSS4, the half-samples used to learn the classes, and the other half-samples which are used to test the result of classification.

In pages "Mono-band classification" to "Information fusion in belief theory", these two groups of half-samples will be used to learn to recognize each class and to test the classifications achieved. Of course, using the half-samples in the opposite order, or building them by a vertical cut, would provide results different from those presented. But the difference is rather weak.