Hyperspectral image classification with sparse representation classifier and active learning

Abstract

Sparse representation classifiers have been widely studied for hyperspectral image classification. The success of sparse representation classifiers depends highly on the training dictionary. However, the definition of training samples, often in the form of field investigations, is time consuming and costly. To mitigate the problem, active learning tries to iteratively define the most informative training samples based on the outputs of the classifiers, thus reducing the quantities of samples to be labeled. For different classification models, several different active learning strategies have been proposed. In this paper, we studied one active learning strategy for sparse representation classifiers. The main idea of the proposed algorithm is to select the samples with most similar reconstruction errors for two different classes. The experiments are performed on two public hyperspectral data. The results show the effectiveness of the proposed algorithm.