A reappraisal of unsupervised classification, II: optimal adjustment of the map legend and a neighbourhood approach for mapping legend units

Abstract

Abstract Unsupervised classification, as customarily implemented, assumes simple correspondence between spectral classes of pixels and the conceptual classes that are to be mapped. In practice rather more complex patterns of correspondence between spectral and conceptual classes will occur. This paper describes procedures for adjusting the proposed map legend to produce complex classes of land cover that are ‘mappable’ and for mapping conceptual classes that do not correspond to simple, non-overlapping subsets of the spectral classes. The results obtained suggest that this approach may be successful.