AN ACTIVE LEARNING METHOD BASED ON MARKOV RANDOM FIELDS FOR HYPERSPECTRAL IMAGES CLASSIFICATION

Spectral-spatial classification of hyperspectral images using Markov random fields (MRFs) has been demonstrated effective to improve the classification accuracy. However, it still suffers from the little availability of labeled samples. In order to collect the most informative samples and train an effective classifier, active learning (AL) methods are often used. In this paper, we present a new AL approach which takes advantage of MRF and iteratively selects the most inconsistent and conflicting instances from the unlabeled set. Experimental results on the AVIRIS Indian Pines data set demonstrated the superior performance of the proposed AL approach with respect to the state-of-the-art AL techniques.