A Study Of Hyperspectral Subpixel Target Detction

In recent years, detection of targets, especially targets in hyperspectral image, suchas vehicles, aerosols, minerals, man-made objects, has found an increasingly wideutilization in both military and civilian use. Hyperspectral imaging sensors acquire datain many contiguous and narrow bands which cover the range from visible light toinfrared ray. This makes hyperspectral image a three-dimensional data, two of whichare spatial dimensions and the remaining one is spectral dimension. Different bands ofone pixel form a continuous spectrum, and different ground materials often havedifferent spectral signatures, which makes it possible to remotely identify targets. Onthe other hand, in hyperspectral images the spatial resolution is limited. Sometimes apixel contains more than one materials, and target is mixed with background, which iscalled subpixel target. Subpixel target is difficult to detect, for the spectral spectrum ofthe mixed pixel is often quite different from the known target spectrum.Many algorithms have been proposed to reduce this problem, and most of themassume that the distribution of hyperspectral image is multinormal. However, thisassumption may not be an appropriate description of the distribution in hyperspectralimage.After carefullystudythe distribution of hyperspectral image, it is concluded thatthe gradient of noise should also be considered. In order in improve the performance ofhyperspectral target detection algorithm, it is necessary to make full use of theaforementioned a prior. The main contributions of this paper include (1) a new modelregarding the inner structure of mixed pixel is proposed, which assumes that gradientof the noise also follow Gaussian distribution;(2) based on the given model, twodetectors, mixture gradient structured detector (MGSD) and mixture gradientunstructured detector (MGUD) are proposed. Compared with the existing works, theproposed detectors has the following merits:(1) the proposed detectors are more robustwith noise. This is because these detectors take advantage of the new model, in whichthe distribution of noise is more accordant with the practical situation.(2)The proposedmodel provide a more accurate description of the structure within a mixed pixel, thus it can locate subpixel target better.