Research On Multi-endmember Unmixing Methods Of Hyperspectral Images

Hyperspectral images are widely used in vegetation monitoring,mineral exploration,urban survey and military reconnaissance because of their high spectral resolution.However,due to the limited spatial resolution of imaging sensors and the spectral mixing effect caused by natural factors,there are generally mixed pixels composed of multiple ground object spectra in hyperspectral images,which limits the in-depth mining and accurate application of hyperspectral image information.In order to analyze the mixed pixel deeply,hyperspectral image unmixing methods emerges.The main objective of this method is to extract the spectrum of pure material（endmember）from hyperspectral image and estimate its proportion（abundance）in the mixed pixel spectrum.The development of hyperspectral image unmixing methods has greatly promoted the performance of other technical means such as image fusion,sub-pixel positioning and super-resolution restoration.The hyperspectral image unmixing methods can be divided into single-endmember unmixing methods and multi-endmember unmixing methods.The traditional single-endmember unmixing methods extract a single fixed endmember representing a class of spectral features of ground objects,and the accuracy of unmixing is low.This is because the geographical span of the image and the complex reflection and scattering and other factors cause the endmember changes of the same kind,that is,the endmember spectral variability exists.The multi-endmember unmixing methods represent the spectral characteristics of a class of ground objects by using multiple changing endmembers,which better reflect the endmember spectral variability.In this paper,the multi-endmember unmixing methods of hyperspectral image are taken as the research object,and the multi-endmember unmixing process of hyperspectral image is taken as the thread,and the problems of multi-endmember extraction,multi-endmember abundance estimation and multi-endmember unmixing uncertainty analysis are mainly studied.The specific research content is described as follows:1.In view of the fact that the traditional multi-endmember extraction methods fail to make full use of the spectral information and spatial information of hyperspectral images at the same time,and the method of extracting endmembers first and then clustering is easy to cause misclassification problems,this paper proposes an automatic endmember bundle extraction method based on probabilistic output support vector machine.This method first randomly selects a certain proportion of pixels from the original hyperspectral image to form a random image subset,and then extracts the endmembers using the vertex component analysis algorithm.At the same time,the neighborhood pixels of the extracted endmembers are used as candidate endmembers,and the probabilistic output support vector machine algorithm is used to determine the probability estimation value of the candidate endmembers,and the candidate endmembers whose probability estimation value is greater than the threshold value are extracted as endmembers.Finally,repeat the above two processes several times to obtain the multi-endmember set.The multi-endmember extraction method proposed in this paper makes full use of the spectral feature information and spatial neighborhood information of pixels,and classifies the endmembers at the same time,solving the problem of misclassification.2.In view of the problems of unclear meanings of multi-endmember models,sinsufficient utilization of spatial information and limited sparsity of estimated abundances in traditional sparse unmixing algorithms,this paper proposes a group sparse multi-endmember unmixing algorithm based on fractional mixed norm and spatial constraints.Because the endmembers extracted by the multi-endmember extraction methods have a group structure,the abundances estimated by the sparse unmixing methods also have a group structure.In this paper,the _G,1,q fractional mixed norm is used to induce the sparsity of the abundance matrix with group structure,in which the ₁ norm induces the sparsity of the coefficients in the abundance groups,and the _q norm（0<q<1）induces the sparsity of the abundance groups.At the same time,considering the smoothness of abundances distribution,the total variation（TV）regularization constraint is introduced into the _G,1,q unmixing framework.Compared with the traditional sparse unmixing algorithms,_G,1,q sparse enhancement constraint and TV space regular constraint improve the accuracy of abundance estimation.In addition,we designed an efficient ADMM algorithm for the proposed algorithm.3.In view of the difficulty in obtaining training samples and the low computational efficiency of traditional machine learning unmixing algorithm,this paper proposes a multi-endmember unmixing algorithm based on least squares twin support vector machines.The algorithm uses the endmembers extracted by the multi-endmember extraction methods as training samples to obtain two non-parallel classification hyperplane equations,and further obtains the abundances by calculating the distance from the mixed pixel to the classification hyperplanes.The algorithm proposed in this paper belongs to the small sample machine learning method.By solving the quadratic optimization problem in the form of least squares,the closed solution relationship between training samples（input）and estimated abundances（output）is established,taking into account the accuracy and efficiency of unmixing.4.The unmixing uncertainty widely exists in various multi-endmember unmixing models,but this problem is often ignored.This paper takes the least squares twin support vector machine unmixing algorithm as an example to study the uncertainty in multi-endmember unmixing.In this paper,the concept of within-class endmembers discrete angles（WEDA）is proposed to quantify the spectral variability of endmember.At the same time,two uncertainty models of abundance overlap and model overlap are proposed,and two quantitative indicators of abundance overlap angle（AOA）and abundance variability scale（AVS）are defined to measure the above two uncertainty models.In this paper,the specific algorithms of WEDA,AOA and AVS are given,and the positive correlation between the endmember spectral variability and the unmixing uncertainty is further revealed.This paper also presents methods to reduce uncertainty,and proposes the mean AVS algorithm（m AVS）to obtain the best compromise for abundance estimation when the unmixing uncertainty is high.