An ensemble learning algorithm for one-class classification of hyperspectral images

Since the spectral resolution of hyperspectral images has reached 10 nm or even higher, it is available to recognize many ground objects which can't be identified in other remote sensing images. But it also brings the problem of high-dimensional data processing in higher spectral resolution images. An ensemble learning algorithm for one-class classification of hyperspectral images is presented to take full advantages of high-dimensional information of hyperspectral image and improve the performance of one-class classifier. In this method, multiple low-dimensional stochastic subspace training sets are generated from the training samples. Then support vector data description (SVDD) on these subspace training sets is trained and subspace sets computing is reduced. These classifiers in average is combined to form an ensemble classifier. Experiment results demonstrate that the method proposed can achieve higher classification accuracy compared with spectral angle match, One class support vector machine (OC-SVM) and direct SVDD algorithm, the overall accuracy(OA) is not less than 90%. ©, 2014, Guangxue Xuebao/Acta Optica Sinica. All right reserved.