Independent component analysis-based band selection for hyperspectral imagery

An independent component analysis (ICA)-based band selection method suitable for target detection in hyperspectral imagery was proposed. Firstly, the number of important independent components were obtained by estimating virtual dimensionality(VD) value while prioritizing the independent components generated by FastICA. After all the bands were ranked according to their average contributions to the important independent components, a spectral similarity measure was used for band decorrelation to remove the redundant bands and make the final selected bands contain the most target information. Two real AVIRIS hyperspectral data sets were tested for target detection. As shown in the experimental results, the proposed method outperforms the other two existing second-order statistics-based band selection methods, and it selects less than 12% and 3% of the full bands respectively, with which the detection probability of adaptive cosine estimator (ACE) and adaptive matched filter (AMF) has been improved by 30% and 15%.