Nonlinear contextual classification - a comparative study

We develop a framework for nonlinear contextual classification by
viewing classification as a regression problem and applying Markov
random field theory for neighboring pixels in an image. We compare
the effect of using a Gaussian classifier, neural nets, classification
trees and recent statistical approaches to nonlinear classification.
The effect of using different discriminant functions is compared to
the effect of a contextual model for three different data sets from
remote sensing applications. For all of the data sets, the use of
a contextual model increases the classification accuracy. For the first
data set, which fits a Gaussian model fairly well, there were no significant
differences in classifier performance for the various discriminant functions.
For the other two data sets, there were not very large differences between
the performance of the various noncontextual classifiers. However, for
the contextual model, some of the nonlinear classifiers performed
significantly better than the Gaussian classifier.