Hyperspectral Image Unmixing Based On Deep Learning And Structured Matrix Factorization

With the development of spectral imaging technology and spectral analysis technology,hyperspectral images are widely used.From mineral exploration to pesticide residue detection in fruits and vegetable,hyperspectral images are gradually entering people's lives,and people are no longer unfamiliar with remote sensing images.Hyperspectral image contains rich data information,which not only carries the spatial image features,but also carries the spectral domain information.Hyperspectral images are divided into more detailed spectral dimensions and have high spectral resolution,which is the advantage of hyperspectral images.However,due to the limitation of the device and other environmental factors,mixed pixels generally exist in hyperspectral images.Mixed pixels seriously restrict the development of hyperspectral images from qualitative description to quantitative analysis.Therefore,the unmixing of hyperspectral image is very urgent and necessary.Although deep learning is not a new kind of technology,in today's big data era and with the support of computing equipment which have strong computing power,deep learning gradually shows great potential and advantages.From image recognition to speech processing to intelligent robot,deep learning shows its strong learning ability and feature extraction ability everywhere.Facing such opportunities and challenges,this paper mainly carried out the research of deep learning in hyperspectral image unmixing,and also considered the application of some statistical methods in hyperspectral image unmixing.The main work of this paper is as follows:1.An endmember extraction network based on fully connected neural network is designed.In addition,the output of the network can be improved by adding some utility layers such as sum-to-one layer and sparsity enhancing layer.During the network performance test,we consider the influence of noise and the number of endmembers in mixed pixels on the accuracy of endmember extraction.2.Combining the advantages of model driven method and data-driven method,a model driven deep neural network CLSUnSAL-Net is designed by using the idea of deep unfolding.CLSUnSAL-Net not only avoids the requirement of precise modeling in model driven method,but also overcomes the difficulty of selecting network structure in deep learning.CLSUnSAL-Net can achieve endmember extraction and abundance estimation at the same time,and it has good performance in both simulation and real data sets.3.A stochastic maximum likelihood structured matrix factorization algorithm using a sampling strategy optimized by MISSO is proposed.In the linear mixed model,the hyperspectral image unmixing problem can be abstracted into a simplex structured matrix factorization problem.With the framework of statistical methods,the prior information of noise and abundance can be integrated into the structured matrix factorization problem,and then solved by stochastic maximum likelihood method.In the process of sample average approximation,we used the sampling strategy optimized by MISSO,and then obtained better results.