Study On Hyperspectral Unmixing Based On Deep Nonnegative Matrix Factorization And Cluster Spatial Processing

Hyperspectral image,which contains abundant spectral and spatial information of the scene,is widely used in the field of quantitative retrieval,environmental monitoring and target detection etc.Due to low spatial resolution of the imaging spectrometers,pixels are recognized to be mixtures of a few materials in the scene.Recently,Nonnegative matrix factorization(NMF)has been widely utilized in hyperspectral unmixing owing to its simplicity and effectiveness,whereas traditional NMF lacks of ability to obtain the information of the hidden layer with its single layer structure and thus can only achieve limited performance.On the other hand,spatial preprocessing methods can get better results by incorporating both spatial correlation and the spectral similarity,while such algorithms still confront poor efficiency and susceptible to outliers.In this regard,a novel deep NMF algorithm is proposed in this thesis by integrating the total variation and reweighted sparsity with deep layers.Moreover,a new spatial preprocessing algorithm is proposed using cellular clustering strategy.The main research contents are summarized as follows:(1)A deep NMF has been obtained by extending the traditional single-layer NMF to the multilayer with pretraining and fine-tuning stages.The former stage pretrains all factors layer by layer,and the latter reduces the decomposition error.Moreover,a weighted sparse regularizer is integrated into the deep NMF model for the sparsification of abundance matrix,and the weights are adaptively updated in view of the abundance matrix.Finally,the total variation is introduced to improve the piecewise smoothness of abundance maps.In this thesis,gradient descent method is implemented for the multiplicative update.Experimental results demonstrate that the proposed algorithm achieves improved performance with strong robustness and denoising ability.(2)A novel spatial preprocessing method has been proposed.The main idea is similar to Simple Linear Iterative Clustering(SLIC)method,which limits the global search scope of clustering to the neighbor area.After that,the cellular cluster strategy is used to initialize clusters.During the segmentation process,spectral information divergence and spectral correlation coefficient are adopted for the representation of spectral similarity.Also,we utilized GPU(graphics processing unit)to speed up the segmentation process for the practical application in big scenes.After the cluster phase that using both spectral similarity and spatial correlation,the principal component transform is carried out to select the candidate endmember sets in each clustered partition and confidence index is introduced to dilute the impacts of outliers.Finally,the final endmembers are extracted from the candidate endmember sets using spectralbased methods.Experimental results show that the proposed algorithm achieves a high cost-effectiveness ratio,and has a potential of handling very large images,which is more practical in the hyperspectral application.(3)The widely-used remote sensing processing software lacks the support for hyperspectral unmixing.To address such problem,we developed a new hyperspectral unmixing prototype software,which is implemented as standalone and complementary program to tackle the shortage of common remote sensing software.And it has certain practical application value,providing an alternative solution for the researchers.There are 35 figures,10 tables,and 123 references in this thesis.