Research On Hypergraph Spatial Representation Based Autoencoder Blind Unmixing Methods For Hyperspectral Images

Hyperspectral images contain abundant spectral and spatial information,and have high spectral resolution.The characteristics of spatial-spectral integration makes hyperspectral images have high application values in research fields of earth remote sensing observation and so on.However,the problem of mixed pixels caused by complex ground object distribution and limited image spatial resolution often limits the accuracy of practical application of hyperspectral images.In order to extract more precise information from hyperspectral images and provide more accurate data references for subsequent research and application,spectral unmixing has become a key technology in hyperspectral image processing,which aims to extract the pure material endmember spectra and their corresponding abundances of each pixel in hyperspectral images.In recent years,deep neural networks have been widely used in the fields of text processing and computer vision.Among them,the autoencoder model with strong feature extraction capability makes the autoencoder-based hyperspectral image unmixing be more and more concerned.The application of deep network structures can generate more representative data features and better improve the unmixing accuracy,and a large number of existing mature deep learning frameworks also reduce the computing cost.However,most of the current autoencoder-based unmixing methods only use the data’s spectral information,and do not fully consider the beneficial functions of the spatial characteristics of hyperspectral images on the unmixing.To solve this problem,this paper combines the idea of hypergraph learning with the autoencoder model,makes full use of the similarity of adjacent pixels in local regions of hyperspectral images,and promotes the autoencoder to extract more accurate high-order hidden layer features,so as to improve the accuracy of unmixing.On this basis,two autoencoder unmixing algorithms based on hypergraph spatial representation for hyperspectral images are proposed.The experimental evaluation of hyperspectral simulation data sets and real data sets verifies the superiority of the proposed algorithm compared with the existing related algorithms.The proposed algorithm is integrated to further complete an application-oriented autoencoder unmixing system for hyperspectral images.The main contents of this paper are as follows:1.A hyperspectral image autoencoder unmixing algorithm based on hypergraph regularization is proposed.Each hyperedge of the hypergraph can contain multiple vertices,and different vertices can also be located in multiple hyperedges at the same time.This property can well represent the high-order complex relationships among vertices.The algorithm constructs hyperedges in the spatial window centered on each pixel of hyperspectral image,obtains the hypergraph Laplacian matrix that represents the abundance similarity between adjacent pixels in local space,which realizes innovatively the hypergraph regularization for the loss function of the traditional autoencoder.Furthermore,the L_1/2 norm constraint of abundance is introduced to enhance the sparse expression of abundance.The hidden layer features of the autoencoder corresponding to the abundances have the spatial characteristics of the actual ground object distribution due to the influence of the local spatial similarity and sparsity expressed by the hypergraph structures,which effectively improves the autoencoder unmixing accuracy.2.An improved method of hyperspectral image unmixing based on superpixel and hypergraph is proposed.Firstly,the hyperspectral image is divided into several irregular homogeneous local spatial sub-regions by using the superpixel segmentation method,which overcomes the influence of different pixels on the hyperedge construction under the fixed regular window settings of the first algorithm.Then,hyperedges are constructed in the superpixels of each pixel to improve the consistent expression of adjacent pixel abundances in the unmixing results.On this basis,in order to further improve effectively the learning of image spatial information in the autoencoder unmixing network,a dual autoencoders cooperative unmixing mode combining superpixel segmentation and hypergraph space representation is designed.The first autoencoder reconstructs each pixel and learns from the central pixel of its hyperedge constructed in the superpixel,and obtains the primary abundance features with high local similarity.The second autoencoder uses such primary abundance features to build an abundance regularization term in the loss function,which can enhance the abundance similarity among adjacent pixels and the accuracy of unmixing.3.According to the application requirements of spectral unmixing in the practical engineering fields using hyperspectral images,the applicability and effectiveness of the proposed algorithm in practical problems are further verified.In this paper,a hyperspectral unmixing application system based on autoencoders is designed and implemented.The system integrates the algorithms proposed in this paper.The completed functional modules mainly include:user login,user management,data set uploading and preprocessing,selection of disambiguation algorithm,calling of disambiguation algorithm,viewing and downloading of disambiguation results,etc.Finally,the real hyperspectral remote sensing images and textile fabric images are used to test and verify the application value of the system in the field of remote sensing and textile material analysis.