Semisupervised classification method can improve classification accuracy using information of large non-labelled samples, whereas it bears a high cost to obtain labelled samples in hyperspectral image classification. Usual semisupervised classification methods need directly to estimate the classification of every sample in test group. A semisupervised classification method based on label mean is proposed in this paper. It first estimates label mean of samples in the test group. The optimal classification surface can be obtained by maximizing the classification space between label means. The least square support vector machine is adopted to transfer the quadratic planning problem into solving linear equations. The proposed method overcomes the faults of high time cost and much recall learning. It is also proved to be better on classification accuracy, complexity and sample scale by experiments.
展开▼
