distribution, and that the statistical parameters (e.g. mean, variance,
covariance matrix) of the training samples are representative.
However, the assumption of a normal spectral distribution could
potentially lead to some errors if the data is not normally distributed.
We selected 5 to 7 training samples per class that are spectrally
different for the classification. We also attempted several different
sets of training samples and qualitatively evaluated the outputs. We
merged those classes generated by different training samples under
the same land-use and land-cover category. The resulting land-use
and land-cover categories were the same as those identified with the
object-based approach. The output map produced by the traditional
classifier is presented in Fig. 14.