The spectral separability of thirteen topical vegetation classes, including twelve forest types, was assessed. Although the thirteen classes could not be classified to a high accuracy the results of a set of supervised and unsupervised classifications revealed that three groups of classes were highly separable; a classification of the three groups by a discriminant analysis had an accuracy of 92·20 per cent. These three spectrally separable groups also corresponded closely to ecological groups identified from an ordination of data on tree species contained within a detailed ground data set. On the basis of the class separability analyses the three spectrally separable groups were mapped, with an accuracy of 94·84 per cent, from Landsat TM data by a maximum likelihood classification. It was apparent that some of the errors in this classification could be resolved through the use of contextual information and ancillary information, particularly on topography.