A robust band compression technique for hyperspectral image classification

Dimension reduction is the key step of hyperspectral image classification. Many techniques have been developed in the past years, but our classification experiments show that some of these techniques are not robust while others suffer from the accuracy and the effectiveness for the classification of hyperspectral data. In this paper, a novel band compression algorithm is proposed based on the fusion of segmented principle component analysis (SPCA) and linear discriminant analysis (LDA) for dimension reduction. We first select the bands independently via SPCA and LDA. Theoretical analysis shows that the selected bands have little correlation, and therefore, an iterative algorithm is adopt to adaptively co-optimizing both the parameter of merging SPCA bands and LDA bands, and the classification accuracy. Our extensive experiments on two real hyperspectral datasets (AVIRIS 1992 Indian pine image and HYDICE image of Washington DC Mall), proves that the proposed technique is not only robust but offers more classification accuracy than many conventional dimension reduction techniques over several well known classifiers.