From the supervised to the unsupervised behavioral classification

The authors recall the complete classification procedure they have developed. This is based on a nonstationary Poisson process for the points distribution in the radiometric space. The intensity functions are estimated nonparametrically by using the uniform kernel. For the a priori probability estimates the authors use an EM-like algorithm, which keeps the densities constant. Then to the classified image they apply a special filtering: a contextual ICM classifier. The points distribution on the 2-geometrical lattice is modelized by a Markov random field. The authors´ ICM classifier takes also into account the radiometric densities. Until last year, the authors worked with training sets given by experts. Now they try to adapt this supervised classification to an unsupervised one. Their philosophy is to try to find automatically homogeneous zones which could be used as training data in the supervised classification. In view of the results, the authors have fixed their choice on a multivariate and nonparametric approach of supports comparison. The zones so found are then merged according to the same criterium and as training sets