Kernel Principal Component Analysis for Feature Reduction in Hyperspectrale Images Analysis

Feature extraction of hyperspectral remote sensing data is investigated. Principal component analysis (PCA) has shown to be a good unsupervised feature extraction. On the other hand, this methods only focus on second orders statistics. By mapping the data onto another feature space and using nonlinear function, Kernel PCA (KPCA) can extract higher order statistics. Using kernel methods, all computation are done in the original space, thus saving computing time. In this paper, KPCA is used has a preprocessing step to extract relevant feature for classification and to prevent from the Hughes phenomenon. Then the classification was done with a back-propagation neural network on real hyperspectral ROSIS data from urban area. Results were positively compared to the linear version (PCA) and to a version of a algorithm specially designed to be use with neural network (DBFE)