About the Automatic Detection of Training Sets for Multispectral Images Classification

In images’ discriminant analysis, training sets are usually given by experts of the field of interest. A procedure is proposed in this paper to avoid such a resort to experts. Two steps are required to achieve this scheme. Firstly, thanks to a multivariate and nonparametric approach of supports comparison, some homogeneous areas are detected on the image. Secondly, building a similarity measure based on the same criterion, those so found areas are merged into a small number of classes. Afterwards, these classes can be used as training sets for any discriminant analysis procedure.