Gaussian Pyramid Based Multiscale Feature Fusion for Hyperspectral Image Classification

Abstract

In this paper, we propose a segmented principal component analysis (SPCA) and Gaussian pyramid decomposition based multiscale feature fusion method for the classification of hyperspectral images. First, considering the band-to-band cross correlations of objects, the SPCA method is utilized for the spectral dimension reduction of the hyperspectral image. Then, the dimension-reduced image is decomposed into several Gaussian pyramids to extract the multiscale features. Next, the SPCA method is performed again to compute the fused SPCA based Gaussian pyramid features (SPCA-GPs). Finally, the performance of the SPCA-GPs is evaluated using the support vector machine classifier. Experiments performed on three widely used hyperspectral images show that the proposed SPCA-GPs method outperforms several compared classification methods in terms of classification accuracies and computational cost.