Rssearch On Hyperspectral Mixed Pixel Separation Algorithm

Hyperspectral image(HSI) is a kind of high dimensional data set, it contains abundant spectral bands information and spatial location information, we can use the rich spectral information to identify and recognize the materials, but due to the sensor resolution limit, mixed pixels are widespread in the hyperspectral images(HSI). It become an main obstacle to the in-depth development for quantification analysis of HSI. The study of hyperspectral mixed pixel separation has very vital significance, it can push through the resolution limit,obtain the property information of the mixed pixels on the subpixel precision and improve the classification precision. This paper studies several classical mixed pixels decomposition(MPD) algorithm, including the PPI, N-FINDR, VCA, SGA,NMF and ICA. These hyperspectral unmixing algorithm are on the base of linear spectral mixture model( LSMM).Conventional independent component analysis(ICA) is based on the assumption that all components are independent of each other, this assumption is not established when applied to hyperspectral unmixing. This paper introduce the constraints of abundance nonnegative and abundance sun-to-one to the objective function to solve this problem. The separation result is sensitive to the initial value on account of there are many local extreme points in objective function. Particle swarm optimization(PSO) was introduced and avoid trapping in local extreme.The performance of the algorithm is validated by simulation experiments, it is concluded the number of endmembers, the number of pixels, purity of endmembers on the mixing effect. Spectral Angle distance and root mean square error as the Performance evaluation index are used to evaluated the efficacy of the unmixing result. spectrum is used than the original components and algorithm to extract the endmember spectral similarity degree. Experimental results on both simulated and real hyperspectral data verified that the proposed algorithm has obvious superiority in enhancing accuracy, it performs well for noisy data, and can also apply to the unmixing of hyperspectral images in which pure pixels do not exist.