Research On Hyperspectral Image Unmixing Based On Spectral-spatial Combination

Hyperspectral images are widely used in many research fields due to their rich spatial and spectral information.However,there are a large number of mixed pixels in the obtained hyperspectral images,which seriously affects the application and development of hyperspectral images.Therefore,hyperspectral image unmixing is an essential step in hyperspectral image application.Hyperspectral image unmixing refers to the process of extracting pure material features（endmembers）and determining their proportion（abundance）from mixed pixels in hyperspectral images.The hyperspectral image unmixing model based on spatial and spectral information can be established,according to he spatial and spectral information of hyperspectral images.The model makes full use of three-dimensional information of hyperspectral images and greatly improves the accuracy of unmixing.However,the unmixing model of hyperspectral images based on space spectrum information fails to fully consider other physical properties of hyperspectral images and is easily affected by noise.Therefore,in order to solve the above problems,on the basis of the hyperspectral image unmixing model based on space spectrum information,this paper fully considers the spatial correlation of hyperspectral image and sparsity of abundance matrix and noise,and proposes some new unmixing methods:1.A new hyperspectral image unmixing algorithm based on promoted abundance sparse（ATNMF）is proposed,which makes full use of spatial correlation of hyperspectral images and sparsity of abundance matrix.Firstly,the spectral difference of adjacent pixels in the original hyperspectral image is used as the weight,and the gradient model is used to describe the spatial correlation of hyperspectral image.Secondly,arctan function was used to promote the sparsity constraint of the abundance matrix.By changing the parameter values each iteration,lipschitz’s continuous arctan function values gradually approached the l₀ norm,which avoided the objective function falling into the local minimum and made the solving process more efficient and stable.2.Cauchy nonnegative tensor factorization（TV-CNTF）hyperspectral unmixing algorithm with total variation is proposed.The proposed method preserves the spatial structure information of hyperspectral images,makes full use of the piecewise smoothness of abundance tensor,and reduces the influence of noise on the unmixing performance of hyperspectral images.Nonnegative tensor factorization model can retain the spatial structure information of hyperspectral images well,under this framework,the Cauchy loss is used instead of the traditional least squares loss,by reducing the weight of noise point in unmixing model to reduce the noise impact on unmixing results,at the same time join the total variation（TV）operator in the model,the piecewise smooth structure of abundance tensor is guaranteed,the accuracy of abundance estimation is improved.3.Total variation nonnegative tensor factorization algorithm based on general loss function（TV-GLNTF）was proposed.Based on TV-CNTF method,the Cauchy loss function was replaced by the general loss function,which retained the spatial structure information of hyperspectral images and the anti-noise performance of unmixing.By adjusting the value of parameters in the general loss function,different loss functions can be obtained,which can effectively deal with hyperspectral image unmixing in a more complex and changeable noise environment and achieve better unmixing performance.The proposed method was tested on simulated and real data sets,and compared with the current better unmixing methods,it is found that the new method is superior to other methods both in visual effect and quantitative index.