This paper discusses the pitfalls of classifying satellite images using automatic classifiers. The discussion is illustrated by application of quadratic and linear discrimination to high resolution data produced for the National Remote Sensing Centre campaign. Within the training set there is good discrimination between most land uses, but the lack of agreement between overall results of discrimination for the two discriminators shows that measures of accuracy based on accuracy within the training set can be misleading. Cross-validation of the quadratic discriminator and calculation of between and within site measures of spectral variability also illustrate how the usual calculations may give over-optimistic measures of discrimination accuracy and spectral variability.