Modified Self-Learning with Clustering for the Classification of Remote Sensing Images

Acquiring labeled data for the training a classifier is very difficult, times consuming and expensive in the area of remote sensing. Many semi-supervised techniques have been developed and explored for the classification of remote sensing images with limited number of labeled samples. Self-learning is a semi-supervised technique in which a classifier is trained in iterative manner. Recently a clustering technique has been integrated in self-learning framework to improve the performance of semi-supervised classifications. This technique considers only the most confident samples to train a classifier. It is possible that most confident samples may not able to improve discriminative capability of the classifier. In this paper, an approach is proposed to select confident as well as informative samples. The experimental results with probabilistic support vector machine and semi-supervised fuzzy c-means on two publically available remote sensing images show that the proposed technique is effective and efficient.