The case when data for pure classes are not available.

If there are no sets of data that represent pure class regions, the statistics of the pure classes have to be inferred from training data, i.e. sites for which the mixing proportions are known from ground inspection. This can be done either by the least square error method or by a robust method. Once the statistics of the pure classes are known, one can create artificial sets of data to represent each pure class with the help of a random number generator. The sets of data points generated can then be treated the same way as real points that represent the pure classes.