An unlabeled samples labeling method of TSVM for remote sensing image

Transductive SVM is a semi-supervised method, which can capture the intrinsic properties of each class' structure in feature space with the help of large number of unlabeled data. It can optimize the classification effect with little and poor representative labeled samples. A weakness of this method is one need determine the number of unlabeled samples which belongs to a specific class before iteration, and the labeling efficiency is very low. We proposed an unlabeled samples labeling method of TSVM for remote sensing image. With this method, we need not know the ratio of the unlabeled samples among classes any more. The first step of our method is clustering, and the number of the clusters must be more than 5 times as the number of classes to be classified. After clustering we get the mean value and the standard deviation of every cluster. Then we labeled the unlabeled samples which contained in the hyper-ball with the mean value as ball-center and the standard deviation as radius a time, instead of labeled one pair of unlabeled samples a time. The classification experiments results prove that the proposed method is not only effective but also can improve the classification accuracy to some extent.