
266 Supervised Classification Part 1
the availability of a second dataset, which is usually referred to as the valida-
tion dataset. An optio n in the ENVI/IDL neural network program described
in Appendix C allows the training data to be split in two, so that half the data
are reserved for validation purposes. (If test data were held back, the training
dataset is then one third of its original size.) Both ha lves are still represen-
tative of the class distributions and are statistically independent. During the
training phase, which is carried out only with the training data, cost functions
calculated b oth with the training data as well as with the ...