Hyperspectral data reduction using pure pixel extraction and ICA
结合纯像元提取和ICA的高光谱降维方法*

Abstract: By comparing ICA with hyperspectral linear model, a new approach to hyperspectral data reduction based on combination of pure pixel extraction and ICA is proposed, in order to retain significant spectral characteristics in reduction process. In this method, virtual dimensionality estimation determined the number of spectral characteristics, and automatic target generation process extracted pure pixel vectors which could be applied as initialization vectors for ICA, and then negentropy iteration and higher order statistics were used for independent components generation and final components selection respectively. Classification results show that the approach not only solves the stochastic scheduling problem of traditional ICA, but also achieves a classification accuracy increment by 6.83 percent comparing with the classical dimensionality reduction method PCA. The proposed approach can protect characteristics of hyperspectral data very well in the case of significantly data reduction and makes it favorable for the following analysis and application.