Classification of Hyperspectral Data Based on Semi-Supervised Tri-Training Learning Framework
A semi-supervised learning framework based on the tri-training scheme is proposed for the classification of hyper spectral data. The framework involves two stages: multiple classifier learning by the improved tri-training and integrating the outputs of classifiers to the final hypothesis by decision fusion. To settle the ill-posed classification problem, in the stage of classifier learning, the label confidence of each learner is measured by the improved estimation of classifier error, and self-training is introduced to expand the labeled set using unlabeled samples with confident labels assigned by classifiers. Hyper spectral data classification experiments show the effectiveness of the proposed framework.
R. Huang and L. N. Zhou, "Classification of Hyperspectral Data Based on Semi-Supervised Tri-Training Learning Framework", Key Engineering Materials, Vol. 500, pp. 374-382, 2012