This paper discusses the problem of stabilizing image-discrimination statistics calculated from training samples. It is proposed to enhance the stability of the estimates by using the principle of ranking of the training information according to its reliability on the basis of an analysis of the actual information content of the measured results, characterized by the structure of the relationships between the measured scattering of the implementations in the sample and the measurement-error level. It is shown that the proposed algorithm is a generalization of algorithm for ridge estimates and discrimination in the subspace of the principal components of selective scattering matrices, and the region in which these algorithms are applicable is determined. The results are illustrated, using as an example the problem of the recognition of water and ice clouds from the spectra of radiation emitted in the spectral region 2-5.5 μm, measured from an airplane.
展开▼