Target Detection Approach for Hyperspectral Imagery Based on Independent Component Analysis and Local Singularity

In this paper, a target detection method for hyperspectral images is proposed. The main advantage of this algorithm is to combine independent component analysis (ICA) and local singularity (LS) to make full use of the high order statistics of hyperspectral images. Comparing the ICA model and the linear spectral mixing model (LSMM), the mixing matrix corresponding to statistically independent components (ICs) can be reinterpreted as the spectral signature matrix. Thus, the mixing matrix can be used to construct a subspace operator to suppress interfering signatures. Then by projecting the original data onto the operator, a new background-removed hyperspectral image can be obtained. To extract target information more effectively, LS based on high-order statistics is adopted to measure the singularity of principal components (PCs). It can be used to select effective PCs of interest. The experimental results show the performance of proposed algorithm is superior to that of conventional RX algorithm.