Image Analysis, Classification and Change Detection in Remote Sensing, 3rd Edition

Maximum likelihood classification 241

6.3.2 A modified classifier for E NVI and a Python script

In order to car ry out an unbiased assessment of the accuracy of s

upervised

classification methods, ENVI’s built-in evaluation procedures allow compari-

son with so -called “g round truth” ROIs or images containing areas of labeled

data not used during the training phase. We will prefer a somewhat different

evaluation philosophy, arguing that, if o ther representative training areas are

indeed available for evaluation, then they should also be used to train the

classifier. For evaluation purposes, some portion o f the pixels in all o f the

training areas can b