Density-based unsupervised classification for remote sensing

Most image classification methods are supervised and use a parametric model of the classes that have to be detected. The models of the different classes are trained by means of a set of training regions that usually have to be marked and classified by a human interpreter. Unsupervised classification methods are data-driven methods that do not use such a set of training samples. Instead, these methods look for (repeated) structures in the data. In this paper we describe a non-parametric unsupervised classification method. The method uses biased sampling to obtain a learning sample with little noise. Next, density estimation based clustering is used to find the structure in the learning data. The method generates a non-parametric model for each of the classes and uses these models to classify the pixels in the image.