Research On Unmixing Methods Based On Nonnegative Matrix Factorization For Hyperspectral Remote Sensing Images

Owing to the complex distribution of materials and the limitation of spatial resolution,mixed pixels commonly exist in hyperspectral remote sensing images?HSIs?and prevent the in-depth development and application of hyperspectral remote sensing towards quantification analysis.As a key technology for solving the problem of mixed pixels,hyperspectral unmixing has become a research hotspot in the field of hyperspectral remote sensing technologies.According to different implementations,hyperspectral unmixing can be categorized as supervised unmixing and unsupervised unmixing.Due to limitation of conditions in practical applications,it is of great significance to obtain endmembers and abundances from hyperspectral data in an unsupervised way.As a popular unsupervised method,the applications of nonnegative matrix factorization?NMF?in hyperspectral unmixing have received broad attention and research,but are still confronted with a variety of difficulties and challenges,such as easily falling into local minima,interfered unmixing accuracy by various forms of noise and unsatisfactory efficiency of processing.Regarding to above problems,the thesis focuses on the research of unmixing methods based on NMF for HSIs.The main research works are listed as follows?1?The sparsity levels are various in hyperspectral images and imposing a sparseness constraint over the entire image may not contribute to the unmixing accuracy To solve this problem,a linear unmmixing method based on nonnegative matrix factorization with data-guided constraints for hyperspectral unmixing is proposed.In this method,the sparsity levels of pixels are effectively evaluated though an unmixing process based on NMF with no constraint.Then both l_1/2 regularizer and l₂ regularizer are adopted to combine their effect of regularization and impose adaptive constraints on pixels with various sparsity levels,and NMF with data-guided constraints which accommodates the distribution of sparsity is further derived for linear unmixing.Synthetic and real data experiments demonstrate that the proposed method achieves high accuracy compared to the existing sparsity-regularized unmixing methods?2?Aiming at solving the problem that HSIs are confronted with the interferences from both noisy bands and noisy pixels,a spectral-spatial robust nonnegative matrix factorization for hyperspectral linear unmixing is proposed.Considering the influence of noise corruption exists in both spectral dimension and spatial dimension,l_1/2 norm and l_2,1 norm are adopted to achieve robustness to noisy bands and noisy pixels respectively.And Huber M-estimation is introduced to achieve better assignment of weights.According to the residues of bands and pixels,hyperspectral data are divided in both spectral dimension and spatial dimension and split into four parts for robust estimation.The multiplicative update rules for endmembers and abundances are derived and provided.Experimental results validate the good robustness of the proposed method in both spectral dimension and spatial dimension.When applied to noisy occasions,the proposed method shows better unmixing performance than other unmixing algorithms?3?Aiming at further achieving sparsity constraint adapting to the sparsity distuibution of data on the basis of achieving robustness in both spectral dimension and spatial dimension,a correntropy-based spatial-spectral robust sparsity-regularized unmixing method is proposed.Based on the band weights learnt by correntropy-based NMF,we propose a correntropy-based spatial-spectral robust unmixing model and a correntropy-based spatial-spectral robust sparsity-regularized unmixing model,which are optimized though half-quadratic technique.Experimental results demonstrate that the proposed method simultaneously achieves good robustness to both noisy bands and noisy pixels,as well as adaptive weighted sparsity constraint for abundances,which improve the accuracy of unmixing effectively.In addtion,a general robust sparsity-regularized unmixing framework based on half-quadratic optimization is proposed Under this framework,novel robust sparsity-regularized unmixing methods can be designed and developed through various forms of potential functions to accommodate the needs of different applications?4?Aiming at solving the bottleneck problem of calculating and memory resource when applying kernel NMF?KNMF?to HSIs,an incremental KNMF?IKNMF?method is proposed.In this method,hyperspectral images are evenly divided and added to the unmixing process to correct the unmixing results in an incremental way.The proposed method improves KNMF by effectively reducing the usage of memory and computation time when applied on large scale and dynamic hyperspectral data.Aiming at solving the problem that the accuracy of abundances is unsatisfactory,an improved IKNMF?IIKNMF?is further proposed.IIKNMF improves the accuracy of abundances significantly by a recalculating process of abundances and retains the advantages of IKNMF in saving memory and computing resource.The efficiency of the derived algorithms is validated by experiments conducted on synthetic data and real hyperspectral images.