Application Of Low-rank Tensor Analysis In Hyperspectral Image Processing

Currently,hyperspectral image(HSI)processing has become a hot topic in the field of hyperspectral remote sensing.It has very high value both in application and theory research.However,there are still some deficiencies in the current HSI processing methods.First,the improvement of resolution in HSI brings about the growth of data and data dimension which have big chanllenges to the traditional image processing methods.Second,with the improvement of current camera technology,although the high resolution images may be obtained,the images are easily disturbed by all kinds of noise.This affects the subsequent image processing.Finally,the mismatch of spatial and spectral resolution results in the phenomenon of heterogenous spectrum and homogeneous spectrum.In order to solve the above deficiencies,four HSI processing methods based on tensor analysis are proposed in this dissertation,all these methods process HSI in both spatial and spectral dimensions.The dimesionality reduction,noise reduction and classification are achieved,respectively.The accuracy of classification is improved.The content of the dissertation is as follows:(1)In hyperspectral image processing,tensor decompositions have been successfully applied to joint noise reduction in spatial and spectral dimensions of hyperspectral images,such as parallel factor analysis(PARAFAC).However,the PARAFAC method does not reduce the dimension in the spectral dimension.To improve it,three new methods were proposed in this dissertation,that is,combine the classical dimensionality reduction methods and PARAFAC to reduce both the dimension in the spectral dimension and the noise in the spatial and spectral dimensions.The experimental results indicate that the new methods improve the classification compared with the PARAFAC method.(2)In hyperspectral image processing,Tucker decomposition has been successfully applied to noise reduction of HSI.However,it is not the only solution.To overcome this disadvantage,various constraints are introduced in Tucker decomposition.Although it can provide more interpretable components with physical meanings,it must suffer from slow convergence.To solve it,the multiway component analysis which is applied to noise reduction of HSI was proposed in this dissertation.The contribution of the new method is that multiway component analysis is applied to the noise reduction of HSI,which has the advantages of good flexibility,strong robustness,and high efficiency and simplicity.The experimental results show that compared with the constrained Tucker decomposition method,this method not only increases the signal-to-noise ratio,but also improves the classification accuracy.(3)To solve the phenomenon of heterogenous spectrum,an HSI is modled as a tensor in many methods.However,an HSI only is a data cube in these methods and these methods do not consider the root cause of this phenomenon.That is,the spectral signatures of ground objects are impacted by multiple factors,such as illumination,mixture,atmospheric scattering and radiation,and so on.In addition,these factors are very difficult to distinguish.Therefore,the spectral tensor synthesis analysis is proposed in this dissertation,in which these factors are synthesized as within-class factors.The within-class factor,the class factor and the pixels are selected as a mode respectively.The pixels in training set are modeled as a third-order tensor.The advantages of this new method are as follows:(a)It is the first in which factors of signatures of ground objects are modeled as a tensor;(b)Unlike in earlier methods wherein parameters have to be adjusted to achieve better classification results,the new method can be used without the adjustment of parameters after the model is created;(c)The most important advantage of the new method is that all the pixels of one class are mapped to the same coefficient vector.Thus,the influence of many factors is minimized.This results in an improved and stable classification.(4)The core tensor contains the variation information for all factors in spectral tensor synthesis analysis.To obtain further improvement of the classification accuracy,factor variation must be explored from the core tensor.The spectral tensor synthesis analysis based on joint eigenmode is proposed in this dissertation.The advantages of the new method are as follows: first,the proposed method defines a unified basis for projection.Second,the basis involves the eigenmodes of class and within-class,thereby truly exploiting the multilinear relations obtained from the multilinear decompositions.The experimental results indicate that compared with the spectral tensor synthesis analysis method,the new method improves the classification accuracy.