Researches On Hyperspectral Image Unmixing Algorithm Based On Non-negative Independent Component Analysis

Hyperspectral Images are simultaneously containing information of space and spectrum obtained from hyperspectral spectrometer.Due to the limit resolution of hyperspectral image spectrometer and complex diversity of nature objects,one single pixel often contains different ground-objects.Mixed pixels always exist,seriously impeding the further development of hyperspectral image processing towards quantitative direction.Therefore,how to effectively solve the problem of mixed pixels and realize hyperspectral unmixing is of great importance.In many actual situations,source signals mixed are non-negative.In consideration of non-negativity as well as independence conditions,nonnegative independent component analysis(NICA)algorithm can realize the nonnegative blind source separation based on statistical theory.As hyperspectral images are typical nonnegative signal and meet abundance nonnegative constraint(ANC)under linear mixture model(LMM),NICA can be used to solve the problem of hyperspectral unmixing.However,abundance sum-to-one constraint(ASC)under LMM is in contradiction to independence requirement of the nonnegative sources.If applying NICA directly to hyperspectral unmixing,its performance will be weakened.The main work of the thesis is as follows:(1)As gradient algorithm is quite easy to be affected by initialization and iteration step length and converge to local minima when optimizing the objective function of NICA algorithm,we adopt cuckoo search(CS)algorithm of better global convergence performance.Thus,we put forward to NICA algorithm based on CS(NICA-CS)to implement nonnegative blind source separation.(2)With respect to the demands of ANC and ASC for hyperspectral images under LMM,we introduce ASC into the objective function of the NICA algorithm to construct a new objective function and propose hyperspectral images unmixing algorithm based NICA.Then we use CS algorithm to optimize the newly-built objective function to accomplish unmixing for hyperspectral images.Experimental results on synthetic data and real hyperspectral data indicate that the proposed algorithm show very strong robustness to pixel purity,image size and noise.And under condition with noise without pure,it still exhibits good performance.(3)In the process of unmixing,in order to reduce parameter dimension and narrow CS search range,we employ the QR decomposition theory of matrix to convert the search process of unmixing matrix to the recognition process of a series of Gives matrixes,and thus decrease the computational complexity.