Relevant Vector Machine Classification of Hyperspectral Image Based on Wavelet Kernel Principal Component Analysis

Hyperspectral image classification by the Relevance Vector Machine (RVM) is a relatively new hyperspectral image classification method, however this method exists some shortcomings such as when the sample data is large and high dimension, the training time will be quit long and the classification accuracy is not so good. To solve these problems, this paper proposes a RVM classification method based on the new Kernel Principal Component Analysis (KPCA). This method uses the kernel function into the PCA and replaced the traditional kernel function with the wavelet kernel function. By using the feature of multiresolution analysis, the new method improves the nonlinear mapping capability of KPCA and the experiment completes the RVM hyperspectral image classification based on the wavelet kernel function PCA, And then the different effects of the hyperspectral image classification between the traditional PCA and the wavelet kernel PCA are analyzed and compared. The results show that by using the WKPCA method, the Euclidean distance of AVIRIS hyperspectral image data between the different categories and the same categories is lower 20% and the variance has been sharp rised. The classification accuracy, by using the RVM, improves the 3%~5%.