Hierarchical Classification of Multitemporal/Multispectral Scanner Data

Most techniques for computer classification of multispectral scanner data involve a "single-stage" approach in which every pixel in the data is classified in a single step, using a single set of training statistics and a single set of wave length bands. Hierarchical classifiers, on the other hand, involve a sequence of classification steps, each of which can involve a different wavelength band or combination of wavelength bands. In addition, at each step in the classification process only one spectral class or a specific group of spectral classes are separated from all other classes in the d~t~. Since a relatively small number of wave length bands are involved at each step, and (after the initial step) only a portion of the data is being classified at each step, such hierarchical classifiers are computationally very efficient. How ever, as compared to single stage classifiers, the effectiveness of hierarchic~l classifiers in terms of classification accuracy is not clear, especially when dealing with multi-temporal data sets. In this study, Landsat-1 MSS data sets ohtained in June1q7~ and February 1974 over the Monroe Reservoir and Hoosier National Forest in central Indiana were used. After digitally registering the two data sets, four classification procedures were compared. The first consisted of a standard single stage maximum likelihood classification using an eight channel training statistics deck (four wavelength bands from each two dates). The second utilized the 4 best channels of the 8 available. The third involved the Layered hierarchical classifier and the same eight channel training statistics. The fourth approach utilized the Layered classifier again, but the data from the two d~tes were treated independently for the purpose of developing training statistics. 062585 The results indicate that the Layered cla~sifier is a more effective and efficient approach for classification of multitemporal/multispectral scanner data. The classification accuracies we~e relatively high for all four classifications, but the Layered classifier required only one third of the CPU time used in the single stage classification.