Adaptive expansion of training samples for improving hyperspectral image classification performance

A relevant problem for supervised classification of hyperspectral image is the limited availability of labeled training samples, since their collection is generally expensive, difficult and time consuming. In this paper, we propose an adaptive method for improving classification of hyperspectral images through expansion of training samples size. The represented approach utilizes high-confidence labeled pixels as training samples to re-estimate classifier parameters. Semi-labeled samples are samples whose class labels are determined by ML classifier. Samples that their discriminator function values are large enough are selected in an adaptive process and considered as semi-labeled (pseudo-training) samples added to the training samples to train the classifier sequentially. The results of experiments show classification performance is improved and this method can solve the limitation of training samples in hyperspectral images.