Cluster-based training data preselection and classification for remote sensing images

In classical image classification approaches, it assumes that there are a number of labeled training data per class. In real applications, labeled data generally are difficult to obtain while unlabeled data are sufficient and helpful to improve the accuracy of classifier. Bipartition based clustering method is to generate better initial cluster centers and to preselect representative data samples from each cluster region with given area under clustering model. To attack the quantity and quality problems of training samples, we propose a Cluster-based Classification Algorithm (CCA) for remote sensing images, and different data samples selection methods are evaluated. Using this approach, the confident unlabeled data both cluster centroid and the ones nearest to the centroid are labeled as training data and extracted. SVM can subsequently be trained with the labeled dataset. The conducted experiments by clustering and classification on real remote sensing Images have validated the proposed approach.