Hyperspectral Remote Sensing Images Terrain Classification Based on PCA-KMFA

Abstract

Hyperspectral remote sensing images (HRSIs) have the problems of high dimensionality, strong linear correlation among dimensions, and poor data separability, which result in low terrain recognition rate. A new feature extraction algorithm based on principal component analysis (PCA) and kernel marginal Fisher analysis (KMFA), namely PCA-KMFA, is presented. Firstly, PCA is used for removing the linear correlation redundancy among the dimensions, and then KMFA is used for extracting the nonlinear separable features in the resulting PCA feature space, Bayesian classifier is used for performing classification in the resulting PCA-KMFA subspace. Based on the experimental results of two airborne visible-infrared imaging spectrometer (AVIRIS) HRSIs, we can see that the presented PCA-KMFA subspace method is superior to the linear discriminant analysis (LDA) subspace method and marginal Fisher analysis (MFA) subspace method.