An automatic target detection algorithm for hyperspectral imagery based on feature-level fusion

Detecting unkown man-made targets in an unknown background is a great challenge in hyperspectral imagery analysis since all of the prior knowledge about targets, backgrouds and noise is not available. In this paper, we present an automatic spectral detection algorithm to deal with the problem. Unlike some hyperspectral target detection algorithm which take advantage of the prior spectral signature, the proposed algorithm is to estimate the spectral signaure completely from the observation and removing undesired signature using linear spectral mixture model and subspace projection before feature-level fusion. It consists of several successive processes: (1)estimating the spectral signatures of background and targets using eigenvalue analysis and automatic target detection and classification algorithm(ATDCA); (2)decomposing the observation space into a noise space and a signature space spaned by target and background spectral signatures; (S)projecting hyperspectral datum onto the signature subspace in order to reduce the noise effects; (4)projecting residual datum onto orthogonal complement subspace of background space spaned by backgroud spectral signatures and onto subspace spaned by targets spectral signatures, thus suppressing the residual undesired spectral signatures; and (5)a generalized likelihood ratio test(GLRT) which, as a fusion center, is used to achieve detection output from component images at feature level. The algorithm is tested with a HYDICE hyperspectral imagery in which simulated targets have been implanted. The results of experiment and theoretic analysis verify the effectiveness of the algorithm.