A hybrid classifier for remote sensing applications.

Abstract

This paper presents a hybrid-unsupervised and supervised-classifier for land use classification of remote sensing images. The entire satellite image is quantized by an unsupervised Neural Gas process and the resulting codebook is labeled by a supervised majority voting process using the ground truth. The performance of the classifier is similar to that of Maximum Likelihood and is only a little worse than Multilayer Perceptions while training and classifying requires no expert knowledge after collecting the ground truth. The hybrid classifier is much better suited to classifications with complex non-normally distributed classes than Maximum Likelihood. The main advantage of the Neural Gas classifier, however, is that it requires much less user interaction than other classifiers, especially Maximum Likelihood.