Nonlinear dimension reduction methods and segmentation of hyperspectral images

In this paper nonlinear dimension reduction methods are applied to hyperspectral images and the segmentation performance is investigated. The biggest disadvantage of nonlinear dimension reduction techniques is their long computational processing time. To overcome this problem, prototypes which represent the spectral distribution of the scene have been obtained with vector quantization and dimension reduction has been applied on these prototypes. Dimension reduction of all pixels in the scene has been accomplished using Radial Basis Function (RBF) neural networks and the developed `K-point mean interpolation' method. The positive effects of these methods on the segmentation of the scene have been presented in the experimental results section using objective evaluation criterion.