Using self-training and graph laplacian in semi-supervised band selection for hyperspectral image classification

A semi-supervised band selection based on self-training and graph laplacian is proposed for the classification of hyperspectral data when small-sized labeled samples presented. In the method, a feature ranking criterion using both labeled and unlabeled samples is first defined to select the initial feature subset. Next, the self-training scheme is used to expand the unlabeled samples into the labeled ones and thus the initial subset can be modified according to the ranking criterion. This correction procedure is repeated to obtain the final band subset. The band selection and classification experiments on hyperspectral datasets show the effectiveness of the proposed method.