In the image remote sensing domain, the current trend in classification consists in using nonparametric or data driven models. Thanks to the development of hardware, new techniques which could never have been implemented 20 years ago are now widely spread. For instance, multivariate density estimates are now available in some statistical packages (e.g. S.A.S.). In kernel density estimation, bandwidth selection has become a field of great interest, also from the theoretical point of view. A great number of papers published e.g. in J.A.S.A., Annals of Statistics, Biometrika, discuss the optimality and efficiency of the bandwidth. Most authors propose an adaptive bandwidth. Kernels have also been investigated, but it appears that the choice of a particular kernel is of less importance. 2-dimensional Markov random field models are also very popular. They are now becoming associated with spectral densities models, allowing contextual I.C.M. filters to take into account both the spatial and the spectral information. For incomplete data, E.M. and S.E.M. algorithms are being revisited. This paper proposes a behavioral approach combining the so-called Convex Hull and Non Parametric Classifications. Applications of the behavioral answer proposed for the classification of Remote Sensing data using the nonparametric intensities of the training sets have been performed in the aim to discuss the contribution of the authors’ research in Supervised Classification. The comparison between parametric and nonparametric methods, as well as with classical methods, shows a superiority of their method, sometimes very important in terms of C.P.U. time and rates of well classified pixels. Specific attention is devoted to the interdisciplinarity characteristic of the present research: for several years, statisticians and geographers are working together in the Geosatel laboratory in order to integrate, as far as possible, theoretical and practical constraints.
