A Research On The Model And Algorithm Of Hyperspectral Image Unmixing

Hyperspectral image is a three-dimensional image formed by using spectral imager to shoot the same area with dozens or even hundreds of wavebands.However,due to the lack of spatial resolution of the spectral imager and the complexity of the surface environment,the spectral response of a single pixel may be a mixture of different material spectra.The existence of mixed pixel problem may make the recognition of objects in hyperspectral images inaccurate and lead to the problem of misclassification.The purpose of spectral unmixing of hyperspectral image is to find out the proportion of various pure substances in the mixed pixel.The mixed spectral signal is represented by the superposition of pure spectral components(endmembers),and its corresponding weight is called abundance vector.Specially,the total variation(TV)term is included into the classical sparse regression formulation to exploit the spatial information of hyperspectral data.The TV term assumes that there are similar mixed substances and similar abundance coefficients between adjacent pixels.However,the assumption of TV term is too strict.Therefore,TV term usually brings some staircase effects.To alleviate this drawback,we will introduce bilateral filter to relax the assumption of TV term for hyperspectral unmixing problems.Because bilateral filtering has the ability of smoothing image and preserving image edge information,we first apply bilateral filtering to every abundance map.It makes the smooth area of the abundance map smoother,and does not lose the edge information.Therefore,the abundance map preprocessed by bilateral filter is easier to satisfy the assumption of the piecewise constant transitions of the TV term.It is on this basis that we propose a bilateral filter based TV regularization term and present an unmixing model combining the new regularizer and the sparsity regularization term.To solve the proposed model,we design an algorithm called sparse unmixing via variable splitting augmented Lagrangian and bilateral filter based TV(SUnSAL-BF-TV),under the alternating direction method of multipliers(ADMM)framework.Then we apply the new algorithm to simulated and real data sets,and compare it with several classic hyperspectral unmixing algorithms.Experimental results show that compared with other algorithms,this algorithm has better or comparable performance for simulated and real data sets.