Research On Hyperspectral Unmixing Model And Algorithm Based On Similarity Measure

Hyperspectral remote sensing images harbor a profusion of spectral and spatial information.Highly mixed pixels exist in hyperspectral images owing to the limitation of imaging instrument resolution and the complex ground coverage in the objective environment.In order to improve the practicality of hyperspectral remote sensing,mixed pixel factorization has become a topical issue in the field of hyperspectral image analysis.Hyperspectral unmixing can identify and extract end element information by analyzing the differences in the spectral characteristics of various substances in the image elements,as well as calculating the proportion of each end element in the image elements,so as to extract the abundance information.Nonnegative Matrix Factorization（NMF）,as one of the typical blind source separation methods,is widely used in hyperspectral unmixing because its own nonnegative constraints well explain the inherent characteristics of hyperspectral image data.However,the blind hyperspectral unmixing based on NMF is a discomfort problem,which will lead to local extreme value problems and the unmixing results are vulnerable to the initial values and noise.In addition,considering that hyperspectral images contain rich correlation information between each dimension,while most of the existing NMF unmixing methods use Euclidean distance as a similarity measure to measure reconstruction error,which cannot effectively model the relationship between dimensions,making the relationship between the internal features of the data not for can be fully considered.To tackle the above problems,while considering the inherent data structure of hyperspectral images,this thesis optimizes and improves the traditional non-negative matrix unmixing algorithm.The following are the main research and innovation points of this thesis.（1）A graph-constrained non-negative matrix unmixing algorithm based on Sinkhorn distance is proposed to cope with the problem that the available NMF similarity measures cannot take into account correlation information in hyperspectral images well.Firstly,the Sinkhorn distance is used instead of the Euclidean distance to model the association tween aspects of dimensionality,capitalizing on the correlation between features.Secondly,in view of the spatial structure learning capability of the manifold approach,the graph regularity constraint is imposed for the abundance matrix based on the use of the Sinkhorn distance metric error to further capture valid features between the data and enhance the effectiveness of the graph regularity constraint in describing the geometric manifold structure.Finally,experiments are carried out on simulated and real data to analyse the effect of abundance estimation and the unmixing effect under different noise levels and different numbers of endelements.Furthermore,ablation experiments are carried out for the Sinkhorn distance part of the algorithm.The experimental results verify the effectiveness of the Sinkhorn distance in constructing the unmixing model,and with the graph regularity,the accuracy of the proposed algorithm is effectively improved by 44% compared to other methods.（2）For the problems that NMF tends to fall into local extremes and is greatly influenced by initial values and noise,a Sinkhorn distance-based smoothed initialized non-negative matrix unmixing algorithm is proposed.The algorithm uses smoothed vertex component analysis（SVCA）instead of traditional VCA as the initialisation method,which smoothes the original data set so as to ensure that a better subset of endmember spectral and initial abundance are recovered in the presence of noise,thus reducing errors due to the influence of initial values.Experiments are carried out on simulated and real datasets and the unmixing results are compared with other algorithms in terms of noise robustness,sensitivity to different endmembers and abundance estimation.Experimental results show that the initialization of the unmixing algorithm using SVCA can further improve the accuracy by 41% and the results are more robust to noise.