A methodology for purification of training samples for the pixel-wise Maximum Likelihood Classification is proposed. In this method, pixels which show comparatively high local spectral variability as well as spectrally separable classes are removed from the preliminary designated training samples. An example using agricultural Thematic Mapper data shows that separability can be improved 3-78 times in terms of divergence between a specific class pair; goodness of fit to Gaussian can be improved 014 times in terms of chi-square; II’9 per cent improvement of the weighted mean percentage classification accuracy can be achieved; and, most importantly, a 20-6 per cent improvement of probability of correct classification can be achieved for a specific class.