Hyperspectral Unmixing Based On Nonnegative Autoencoders And Nonnegative Matrix Factorization

With the applications of remote sensing tend to be quantitative and accurate,hyperspectral unmixing has become one of the key technologies of hyperspectral remote sensing image processing,which has attracted more and more attention from domestic and foreign scholars.Hyperspectral unmixing is such a process in which hyperspectral images are used to unmix the mixed pixels into several basic types of materials spectral vector(endmember),and to obtain the proportion of these basic materials(abundance).The improvement of hyperspectral unmixing performance is not only beneficial to the development of hyperspectral applications,such as the classification and recognition of objects,image interpretation and visualization,image enhancement and compression,but also has important significance for geological exploration,agriculture monitoring,military investigation,and so on.This thesis summarized and analyzed the research status of hyperspectral unmixing,and proposed two new nonlinear unmixing algorithms.(1)Hyperspectral unmixing based on NNSAE-BPBy making full use of the advantages of nonnegative sparse autoencoders(NNSAE)in mining the internal structure of data and extracting features,a nonlinear hyperspectral unmixing model based on NNSAE-BP is proposed,the main procedure of which consists of two key stages: unmixing model learning in supervised way,and nonlinear unmixing of hyperspectral data.Therefore,the problem of hyperspectral unmixing is transformed into two core points: hyperspectral feature representation by NNSAE encoding model,and endemember abundance prediction based on BP model.In the research of hyperspectral feature representation by NNSAE encoding model,to optimaly get the hidden layer nodes,a new adaptive selection method based on Otsu is proposed,which can effectively reduce the computational complexity.The BP network uses the small batch gradient descent method to realize the regression analysis,and introduces the L2 weight decay term in the objective function to prevent over-fitting.Experiments on real hyperspectral databases show that this method can effectively improve the unmixing performance.(2)Hyperspectral unmixing based on SSC-rNMFFrom the perspective of blind source separation,using the nonnegative matrix factorization(NMF)theory as the research tool,and introducing the L2 regularizationconstraint of endmember and the L1/2 regularization constraint of abundance,a smoothness and sparseness constrained robust nonnegative matrix factorization(SSC-rNMF)method is proposed.The method based on SSC-rNMF is unsupervised,and the required priori information is only the number of endmembers.The low-rank decomposition of hyperspectral images is realized by the algorithm of block coordinate descent in multiplicative iteration.The regularizations which are consistent with hyperspectral unmixing characteristics reduce the solution space of matrix factorization.Different from NMF,the SSC-rNMF algorithm computes the nonnegative outlier term by multiplicative iteration,which describes the effects of the nonlinear factors,improving the generalization ability of the model and the anti-interference ability of the noise.Finally,experiments demonstrate the effectiveness of this proposed method.