Unsupervised Hyperspectral Unmixing Theory And Techniques

Hyperspectral imagery(HSI)is a three-dimensional imagery generated by imaging spectrometer simultaneously to the same surface scenery at dozens even hundreds bands.One of the main applications of HSI is to use the rich spec-tral information to identify and recognize the materials,which is also the basic reason for the widely used of HSI to both military and civilian fields.However, owing to the low spatial resolution of the sensor,mixed pixels are widespread in HSI.The mixed pixel problem not only influences the precision of object recog-nition and classification,but also becomes an obstacle to quantification analysis of remote sensing images.Hence,how to effectively interpret the mixed pixels is a critical problem for hyperspectral remote sensing applications.Hyperspectral unmixing is the procedure by which the measured spectrum of a mixed pixel is decomposed into a collection of constituent spectra,or end-members,and their corresponding fractional abundances that indicate the pro-portion of each endmember present in the pixel.Conventional hyperspectral unmixing generally includes two steps:endmember selection and mixed pixel decomposition.Nevertheless,the non-completeness of ground object spectral library and the existence assumption of pure pixel of each endmember lead to the unsatisfactory unmixing results.On the contrary,based on the spectral mixing model and constrained conditions of mixed pixel,unsupervised hyper-spectral unmixing utilizes unsupervised signal processing techniques to obtain the endmember spectra and their corresponding fractional abundances directly from the HSI without knowing the knowledge of endmembers.The unsuper-vised approaches overcome the disadvantages of conventional methods,which provide a new idea for hyperspectral unmixing,and become a research hotspot these days.This thesis addresses the unsupervised hyperspectral unmixing.Ac-cording to the nonnegativity and continuity of both endmember spectra and abundances,and also the sparseness constraint of only abundances,we use the blind source separation(BSS)and nonnegative matrix factorization(NMF)to ob-tain the spectral information and corresponding abundances about mixed pix-els.The major works and contribution of this dissertation are as follows:1)Markov random field(MRF)model is adopted to incorporate the spatial information into independent component analysis(ICA),so the MRF-ICA mix-ture model is proposed.Because the ground objects are continuous in the spatial domain,which makes the probability that the neighborhood of any material is the same is high,the pixels are spatial correlated.MRF model is thought of as a powerful tool to model the joint probability distribution of the image pixels based on local spatial interactions.The experimental results demonstrate that the proposed MRF-ICA mixture model can produce better HSI unmixing results than FastICA.2)For ICA model,the sources are assumed to be independent and sta-tionary.However,in hyperspectral unmixing,this assumption is not valid. Hence,complexity based BSS algorithm is introduced to unmix hyperspectral data,which studies the complexity of sources instead of the independence.First, we extend the one-dimensional temporal complexity,called complexity pursuit, to the spatial domain to describe the autocorrelation of each endmember abun-dance,named spatial complexity BSS(SCBSS).Second,the temporal complex-ity of spectrum is combined into SCBSS to account for the spectral smoothness, which is termed spectral and spatial complexity BSS(SSCBSS).The complexity based BSS algorithms offer three main advantages over other unmixing tech-niques:●The complexity based BSS algorithms are based on a more reasonable assumption that the statistics of endmember spectra and the correspond-ing fractional abundances vary smoothly,due to the high spectral and low spatial resolutions of HSI.●For hyperspectral data,because the number of bands(L)is generally much larger than that of endmembers(P),general ICA methods usu-ally employ principal component analysis(PCA)to prewhiten the data for reducing the dimension,and retain the principal components with the largest eigenvalues.But some important features may reside within the subspace discarded by PCA.We find a very important property of complexity based BSS methods that the larger L is,the more precise the obtained sources are,which shows L>>P is an advantage rather than a limitation.●The huge amount of HSI makes the time complexity of data analysis be-ing an important index to measure the effectiveness of algorithms.Since the measure of signal(either temporal or spatial)predictability is easy to implement,the complexity based methods have a feasible computational cost.Experimental results show that complexity based BSS algorithms provide promising approaches for hyperspectral unmixing.3)Further,the complexity based BSS algorithm is extended by incorporat-ing the nonnegativity and sparseness constraints to unmix hyperspectral data. The measurement of sparseness is implemented by nonsmooth NMF(nsNMF) and NMF with sparseness constraints(NMFSC)algorithms respectively.The monotonous convergence of the algorithm is proved with the help of an aux-iliary function.Several experiments with synthetic data lead to the conclusion that the algorithm is more robust to the noise,signature and illumination vari-ability,applicable to the images containing larger number of pixels and end-members.It is also applied to real hyperspectral data.The results achieved show that the algorithm has the potential of yielding accurate estimates of both endmember spectra and abundance maps.