ABSTRACT
Continuing in the present chapter on the subject of supervised classification, we will begin with a discussion of postclassification processing methods to improve classification results on the basis of contextual information. Then we turn our attention to statistical procedures for evaluating classification accuracy and for making quantitative comparisons between different classifiers. In this context, the computationally expensive n-fold cross-validation procedure will provide a good excuse to illustrate how to take advantage of modern cloud-computing services and perform several tasks in parallel. As an example of so-called ensembles or committees of classifiers, we then examine the adaptive boosting technique, applying it in particular to improve the generalization accuracy of neural network classifiers. This is followed by the derivation and implementation of a maximum likelihood classifier for polarimetric SAR imagery. The chapter concludes with a discussion of the problems posed by the classification of images with high spectral resolution, including an introduction to linear spectral un-mixing and the derivation of algorithms for linear and kernel anomaly detection.
