Improving spectral image classifications by incorporating context data using likelihood vectors

Statistical pattern recognition procedures, such as maximum likelihood classification, are applied to (multi-spectral) satellite images, in order to produce thematic maps (eg. land-use/land-cover maps) in most cases. Sometimes, the purpose is to obtain estimates of the sizes of the areas covered by the different classes. Area estimates that are "easily" created by counting the numbers of pixels per class label after a maximum likelihood classification (histogram) are not reliable, since classifiers tend to be biased in favour of some classes, at the expense of others. On the other hand, knowing areas per class and using them as input for the classifier in the form of prior probabilities, can improve the classification accuracy (but still not the resulting area estimates when making a histogram afterwards: they would be different from what you input). The purpose of this paper is to find a way out of this somewhat strange situation. It presents a slightly modified k-nearest-neighbour strategy to calculate feature probability densities. Also it reviews the method of using spatially distributed prior probabilities and see how it can be perfectly combined with the proposed method.