Hyperspectral image classification by combination of active learning and extended multi-attribute profile

The main difficulty in hyperspectral image classification is the few labelled samples versus high dimensional features. Moreover, spatial information of hyperspectral image plays an important role. Therefore, we focus on combination of the active learning and extended multi-attribute profile spectral-spatial classification of hyperspectral image. We adopt active learning (AL) based on best versus second-best (BvSB) in order to iteratively select the most informative unlabeled samples and enlarge the training set. The spatial information is obtained by extended multi-attribute profile. To evaluate and compare the proposed approach with others, experiments were conducted on two hyperspectral data sets. Results demonstrated the effectiveness of the proposed method.