Hyperspectral Remote Sensing Images Feature Extraction Based on Weighted Classwise Non-locality Preserving Projection

Abstract

In order to solve the high dimensionality and high spectral correlation problems of hyperspectral remote sensing images (HRSIs), a new feature extraction method, named weighted classwise non-locality preserving projection (WCNLPP), is proposed. WCNLPP introduces uncorrelation coefficient to express the dissimilarity degree between the samples of different classes and constructs a non-nearest neighbor graph, such that the non-locality manifold structure of the samples is preserved after feature extraction. Firstly, principal component analysis (PCA) is used to reduce dimensionality and remove the spectral correlation of HRSIs; then, WCNLPP is utilized to guide the procedure of feature extraction after PCA; finally, minimum distance (MD) classifier and discriminant analysis (DA) classifier are used to perform terrain classification in the final feature subspace. The experimental results based on two real HRSIs show that, comparing with PCA, linear discriminant analysis (LDA) and classwise non-locality preserving projection (CNLPP) methods, the presented WCNLPP method can improve the terrain recognition accuracy.