Study Of Unmixing Models And Algorithms For Hyperspectral Images

Hyperspectral imaging which incorporates in imaging and spectral technique is a rapid developing research filed in remote sensing.There are a wide range of applications such as military target disrimination,remote surveillance,biomedical,food safety and environmental monitoring.However,due to the limited spatial resolution of the imaging spectrometer's sensors,the observed pixels may consist of distinct materials,which results in mixed pixels existing in hyperspectral data extensively.Mixed pixels result in inaccurate classification of some applications in scientific research practice.Hence,the spectral unmixing is a crucial and serious problem for hyperspectral remote sensing.Firstly,this thesis introduces two spectral mixing models: the linear mixing models and the nonlinear mixing models.The linear mixing models assume that the observed signature is the linear combination of the pure spectral signatures.Being quite opposite,the nonlinear models consider the physical interactions between the light scattered by multiple materials in the scene.Secondly,this thesis presents some classical hyperspectral unmixing models.Among this models,this thesis introduces the sparse unmixing via variable splitting augmented Lagrangian and total variation(SUnSAL-TV)in detail.The SUnSAL-TV model makes use of spatial information of the hyperspectral images to establish the regularizer of the fractional abundances of endmembers which makes the experimental results improved in terms of SRE values and visual inspection.However,the disadvantage of the SUnSAL-TV method is that the observed results suffer from staircase effects in smooth regions of images.In this thesis,we consider the total variation with overlapping group sparsity as the regularization term for the proposed model.We develop an efficient alternating direction method of multipliers to solve the corresponding minimization problem.Instead of minimizing a difficult cost function directly,we solve a sequence of simple optimization problems to obtain the global optimal solution.In the process of applying the alternating direction method of multipliers,we adopt the majorization minimization method to solve the overlapping group sparsity subproblem.The experimental results with both simulated and real hyperspectral data demonstrate the effectiveness and efficiency of the proposed method in terms of SRE values and visual inspection.Besides,the staircase effects are efficiently allevaited and the estimated fractional abundances maps looks more smooth,besides,we can find the superiority of the visual inspection of the proposed method clearly.