A Composite Kernel-based Classification With RVM For The Hyperspectral Image Classification

Hyperspectral imaging is a multi-dimensional information retrieval technology which combines imaging technology with subdivided spectroscopy technology. Hyperspectral image classification techniques have a great development gradually with the help of imaging spectroscopy, and it has been used widely in all kinds of applications. The hyperspectral data which have a high phenomenon and correlation-larger band is used in the traditional classification algorithms such as maximum likelihood estimation and neural networks will suffer the curse of dimensionality problem. Comparing with the SVM as a another kernel method, it need less relevant vectors and faster test time and can get a similar classification accurate rate.In this paper a composite kernel-based classification with RVM for the hyperspectral image classification algorithm is proposed, to make up the shortage of the lower classification accuracy of RVM, which study relevance vector machine as the basic model and combines the spatial information and spectral information together. The main work of this paper can be summarized as follows:(1)A variational RVM classification algorithm is proposed when the RVM theory is combined with hyperspectral image classification, and it performs simulate experiments compared with the traditional SVM algorithm on real hyperspectral data sets and make a sufficient comparison. (2) A composite kernel-based classification with RVM for the hyperspectral image classification algorithm is proposed. The algorithm combines the spatial characteristics with spectral characteristics from the viewpoint of composite kernel, considering both the spectral characteristics which is the main role for the image classification and the space diversity of hyperspectral data. To avoid the regulation of the balance between the spectral and spatial information specially in the weighted summation kernel, the generalized composite kernel framework is proposed, and for ensuring the reliability of the related experiment, meanwhile we discuss the influence of different window sizes on all kinds of classifiers.(3) The method based on composite kernel in this paper can make better use of the spatial information of the hyperspectral data, and the experiments on the real hyperspectral data illustrate the effectiveness of the proposed method compared with some other methods.