A new decision rule significantly improves the classification accuracy
of the nonparametric spectral classifier described by Skidmore and Tu
rner (1988). The new rule normalizes class histograms with the mean fr
equency, rather than with the number of training pixels in a class, an
d gives a classifier which consistently outperforms the maximum-likeli
hood classifier. Another type of classifier, which combines a two-dime
nsional and a one-dimensional Skidmore/Turner (S/T) classifier, has hi
gher classification accuracies than the S/T classifier with the improv
ed decision rule, in two of four study areas. Tree species were classi
fied for two study areas and land-cover classes for the other two.