Research On Sparse Unmixing Algorithm For Hyperspectral Images Based On Dictionary Extension

Hyperspectral unmixing has always been the key technology of hyperspectral remote sensing information processing,in the field of which the spectral library based unmixing is a hot topic.However,in the real application of unmixing,the spectral dictionary is extremely limited by complex conditions,resulting in precision decline and algorithms performance degradation.In view of the limitation of spectral dictionary,this paper mainly studies the spectral dictionary extension in the following three parts:1.Hyperspectral image unmixing based on kernel sparse representation is proposed.The spectral dictionary is extended through kernel mapping and observation hyperspectral data as well as spectral library are mapped into high-dimensional space.The optimization problem is solved by building a proper kernel function and utilizing the alternating direction method of multipliers for unmixing.The algorithm effectively overcomes the precision decrease caused by spectral dictionary’s high mutual coherence when retaining the endmembers of the spectral library and improves the abundance inversion accuracy.2.On the basis of experimental analysis of spectral mismatch,a spectral mismatch compensation model based on endmember and band is built.Spectral scaling factors belonging to endmember and band are introduced to extend the spectral dictionary.The objective function is proposed by adding the smooth constraints to scaling factors and solved by an alternating optimization algorithm for unmixing,which effectively overcomes the limitation to unmixing algorithms caused by the fixed spectral dictionary.High unmixing accuracy can be achieved,and the algorithm has strong robustness against noise.3.Combining the sparse unmixing property with the nonlinear model,sparse unmixing based on nonlinear model withl₁-l ₂ norm constraints is studied.The spectral dictionary is extended by introducing the secondary scattering term between endmembers and then the nonlinear relationship between the mixed spectrum is transformed into a linear issue.Thel₁-l ₂ norm is taken as the sparseness constraint and the multi-weight sparse unmixing model is proposed,which is finally solved by alternating direction multiplier method.The algorithm makes up for the lack of the description of endmember nonlinear mixing in the spectral dictionary and it effectively improves the unmixing accuracy.Besides,it is robustness to nonlinear models.Experimental results show that spectral dictionary extension can obtain proper spectral libraries to describe the ground distribution characteristics and effectively improve the unmixing performance.The research results enhance the applicability of the semi-supervised hyperspectral image unmixing technology in the real scene.