Hyperspectral Image Unmixing Based On Deep Learning Network Modeland Non-negative Matrix Factorization

With the continuous improvement of the resolution of imaging spectrometers,hyperspectral remote sensing image data has become more and more widely used in various fields.However,due to the existence of mixed pixels in hyperspectral data,its application in various fields is very restricted,which seriously hinders the development of hyperspectral image data towards quantitative development.Therefore,research scholars in various places have proposed various unmixing models to study hyperspectral data.Based on the results published by previous scholars,this paper proposes to unmix the hyperspectral data based on the network model with deep learning as the framework.At the same time,according to the characteristics of hyperspectral data,strict feature constraints are to obtain higher-precision unmixing results.The main research contents of this article are as follows:(1)This paper proposes to integrate the generative adversarial network with the3 D convolutional neural network,extract the hyperspectral deep network structure features,and remove the noise interference during the feature learning process.The focus of this chapter on noise reduction is to use the characteristics of the game between the generating network and the discriminant network to design the generator as the main noise reduction network.The discriminating network discriminates the data on the feature domain.After the network reaches the Nash balance,the generator will produce enough noise to deceive the discriminator and achieve an unsupervised denoising process.(2)This paper constructs a stack-based non-negative sparse variational autoencoder network(SNN-SVAE)with highly mixed hyperspectral data to unmix the objects.The SNN-SVAE model consists of two parts.The first part of the network uses stacked autoencoders(SAEs)to generate the initial value of the end element value,thus generating a good initialization for the unmixing process and automatically detecting the outliers in the original data.In the second part of the network,non-negative sparse variational autoencoder(NN-SVAE)is to perform non-negative matrix decomposition on hyperspectral data.The nodes of the hidden layer are the abundance estimates,and the corresponding weight values are the endmember.At the same time,the unsupervised unmixing is finished.(3)This paper proposes to use the spectral unmixing model to achieve the spatial resolution improvement of hyperspectral data.The main content is to use joint non-negative matrix decomposition and the cascade network model proposed in this paper to achieve image fusion of hyperspectral data and multispectral data,and finally use a multiplicative iterative update algorithm to achieve abundance matching between the two to obtain reconstructed image with high spatial resolution.