Semisupervised Classification of Hyperspectral Image Based on Clustering Kernel and LS-SVM

When classifying large scale hyperspectral image data, there are a lot of unlabeled samples. In order to enhance the classifier’s performance by using the information contained in the unlabeled data, this paper presents a semisupervised classification method. The proposed algorithm modifies the kernel function based on the clustering assumption. It assumes that the samples belonged to the same class will be assigned to the same cluster in the clustering at high probability. The algorithm clusters the unlabeled samples using K-means clustering algorithm. The K-means method uses spectral angle to measure the differences between the samples. The bagged kernel is constructed based on the multi clustering results of data set. Then the bagged kernel is combined with the RBF kernel using sum or product operation. So the information in the unlabeled samples is merged into the classification procedure. The proposed algorithm adopts the least squares SVM (LS-SVM). Instead of solving the quadratic problem of SVM, the LS-SVM changes it to linear equations. The proposed method is validated by the hyperspectral data. In the experiments the proposed method shows some superiority.