Hyperspectral Imagery Unmixing Algorithm Based On Sparse Representation

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.Traditional unmixing algorithm for linear spectral mixture model uses the endmember extraction algorithm to obtain the pure pixels,and spectrum of pure pixels as the prior knowledge is used in abundance inversion algorithm.However real hyperspectral data may not have pure pixels,so the results of endmember extraction be used for abundance inversion algorithm will make serious error and lead to the final solution is not satisfied.To solve this problem,the theory of sparse is introduced to hyperspectral unmixing algorithm and the algorithm based on sparse representation is proposed.Linear spectral mixture model based on sparse representation,which can be equivalent to get the equation solution in a complete spectrum library,chooses matching tracking algorithms for sparse solution.OMP algorithm for sparse unmixing,can obtain good unmixing effect,but the unmixing accuracy needs to be improved under the condition of low signal-to-noise ratio.In order to further improve the effect of OMP sparse unmixing algorithm and make the algorithm more practical applications,this paper conducted a series of improvement,the main content include the following three aspects:1.OMP sparse unmixing algorithm uses inner product to measure the matching degree between the endmember spectrum in spectral library and the residual spectrum.In the low signal noise ratio(SNR),the matching effect is not ideal.To solve the problem,this paper puts forward using generalized Dice coefficient to measure spectrum matching degree instead of inner product.Due to the Dice coefficient considers more spectrum information,it has better ability to resist noise.In this paper,the hyperspectral unmixing algorithm with Dice coefficient as matching criterion is called DOMP sparse unmixing algorithm.Simulation and real data experiments show that under the condition of low signal noise ratio(SNR),DOMP algorithm can get the higher unmixing precision and show better anti-noise performance.2.OMP algorithm as a sparse solution algorithm,the sparse solution is unconstrained,and even negative.And the abundance coefficient of mixed pixels represent the proportion of endmember in each mixed pixels,should meet the "negative" and "sum-to-one".For OMP sparse unmixing algorithm to get the sparse solution meet abundance coefficient of nonnegative constraints and fully constraints,this paper adds constraint in DOMP algorithm to make the abundance coefficients satisfy physical meaning.By analyzing the simulation experiments with simulated and real data,the abundance constrained of sparse unmixing algorithm greatly enhance the unmixing precision and FDOMP algorithm is slightly better than NDOMP,also the algorithm complexity is relatively higher.NDOMP algorithm can meet the unmixing accuracy requirements under the condition that needs high efficiency.3.When matching tracking algorithms applied in the hyperspectral unmixing field,it needs set the sparse degree of mixed pixels and different sparse degree has a greater influence on the unmixing effect.OMP algorithm needs to set sparse degree which will lead to poor practical applications.In order to solve this problem,this paper proposes a adaptive sparse degree unmixing algorithm.The algorithm by means of restraining spectral residual to give each pixels adaptive sparse degree and on this basis with fully abundance coefficient constraints is called FAOMP sparse unmixing algorithm.The results of simulation and real data experiments show that FAOMP algorithm not only realizes the adaptive of sparse degrees but also obtains high unmixing accuracy and also has more practical application value.