Subspace divided semi-supervised SVM classification for hyperspectral images

Abstract

Since the traditional SVM algorithm has high classification accuracy at the expense of huge training samples, A Subspace Divided Semi-Supervised Support Vector Machines (SDSS-SVM) classification method which need only one sample for each class is proposed. In the proposed method, a coarse classification result is obtained based on minimum distance clustering first. A General Sphere Criterion is introduced and applied to the coarse result, and the testing samples is divided into identified samples and unidentified samples. Then, the subspace division is accomplished according to the probable mixing. Samples which have the highest confidence in the subspace are selected as the training samples to subdivide the subspace of the unidentified samples to get the final classification. Classification experiment illustrates that the proposed approach can reach quite high classification accuracy. © 2012 Springer-Verlag.