Research On Multiple Endmember Spectral Unmixing Of Hyperspectral Image

Hyperspectral image has been widely used for its high spectral resolution.Due to the limitation of spectroscopy spatial resolution,mixed pixels are ubiquitous present which affects the accuracy of subsequent image processing,and therefore,it is necessary to research on spectral unmixing.Currently,the unmixing model is mainly based on linear spectral mixture model,simple structure,clear physical meaning and widely used.However,hyperspectral image space range is great.So the variability within a class is usually large,leading to synonym spectrum.Traditional linear spectral mixture model uses a single fixed endmember to describe a type of feature and uses all the endmembers to unmix all of the pixels without distinguishing the endmembers number,causing an error.So,how to overcome the spectral variability and change of endmember number is an urgent problem.To address this problem,we have conducted intensive research.For multiple endmember unmixing efficiency,this paper have improved multiple endmember spectral mixture analysis and proposed a hierarchical algorithm;for complex practical application of multiple endmember spectral unmixing,a fully orthogonal matching pursuit multiple endmember spectral unmixing algorithm has been proposed.The main content and innovation include the following aspects:1.Classic MESMA(Multiple Endmember Spectral Mixture Analysis,MESMA)algorithm is the most widely used algorithms for solving spectral variability and change of endmember number,but there are shortcomings on computation intensive,cumbersome endmember preselected,over-fitting and so on.For this issue,an improved algorithm has been proposed,according to the characteristics of OSP(Orthogonal Subspace Projection,OSP)that can separate signals of interest or not interest.It projects the pixels onto the orthogonal subspace composed by the entire surface feature classes(each class selects only one spectral)endmembers and separate different pixels by projection value of zero(theoretical value).Then determine the optimal number of endmember combinations according to the reconstruction error variation,which avoids over-fitting.Experimental results show that OSPMESMA algorithm improves not only computational efficiency but also unmixing precision.2.We resolve complex issues into simple steps through layer to achieve the purpose of optimization problems.Based on the idea of decomposition,we have proposed a hierarchical multiple endmember hyperspectral unmixing algorithm.Under the guidance of multiple endmember spectral mixture model,the first layer is to determine the feature category by solving the maximum unmixing abundance error,and the second is stratified to find the optimal number of endmemers contained in the pixels by spectral angular distance on the basis of the first layer.Synthetic and real hyperspectral data experimental results show that the proposed algorithm can effectively overcome the unmixing errors causing by spectral variability and the change of endmember number.3.In recent years,sparse regression algorithms have applied increasingly to hyperspectral unmixing and have achieved certain results.Among them,Orthogonal matching pursuit algorithm(Orthogonal Matching Pursuit,OMP),a greedy optimization algorithm with relatively simple structure and high computational efficiency,do not consider the abundance non-negative and sum-to-one constraints,obtaining unsatisfactory unmixing results.In addition,many unmixing algorithms are difficult to handle complex scenes actually For this issue,we have proposed a full-constrained OMP multiple endmember hyperspectral unmixing algorithms.First,add non-negative and sum-to-one constraints to OMP algorithm,then unmix by the proposed algorithm.Synthetic and real hyperspectral data experimental results show that FOMP algorithm unmixing results are ideal and has a strong anti-noise performance.