A non-parametric method for feature selection

A non-parametric feature selection technique is proposed. It is hoped that a finite number of classes is represented by some finite number of unknown probability structures which are distributed in a finite discrete measurement space. No assumptions of statistical independence between pattern measurements will be made. The proposed non-parametric feature selection criterion is based on the direct estimation of the minimal expected error rates for a given data set of training samples and is independent from the classification technique used. The properties of the proposed feature section are demonstrated using data from agricultural remote sensing.