As customarily implemented, unsupervised classification assumes a simple pallern of correspondence between spectral classes and the conceptual classes that are to be mapped. This will not always apply. In this paper various possible patterns of correspondence between spectral and conceptual classes are considered. It is shown that land facets and land systems in a region of south-east Spain do not correspond simply to spectral classes derived by cluster analysis of SPOT data. Rather, these conceptual classes correspond to overlapping subsets of spectral classes that differ in their proportional composition. In such situations rather more sophisticated methods than conventional, unsupervised classification will be necessary in order to generate a map of conceptual classes from an image segmented into spectral classes.