Hyperspectral Unmixing Via Deep Neural Networks And Spatial-spectral Attention Mechanisms

The hyperspectral image data obtained by the hyperspectral imager has the characteristics of high resolution,large amount of data and multiple bands.Therefore,hyperspectral remote sensing has a relatively widespread application and broad prospects in various fields of today’s society,but hyperspectral remote sensing data also has the problem of high spectral resolution but low spatial resolution.Because of the low spatial resolution of the data obtained,mixed pixels are common in hyperspectral remote sensing images,and it is not a pure pixel,but a combination of various endmember according to a certain ratio,so the emergence of mixed pixels not only has it hindered us in the classification of direct pixels,it has also caused problems in improving the classification accuracy of hyperspectral remote sensing data.Hyperspectral unmixing technology refers to the process of separating hyperspectral images into spectral features of endmember and each pixel has a set of abundance scores.In this paper,the concepts of deep learning are introduced to the problem of hyperspectral image unmixing and the convolutional network framework is used to extract and fuse the spectral and spatial information of hyperspectral images to obtain a better unmixing effect.The main work of this article is as follows:(1)We propose a supervised hyperspectral image unmixing method based on convolutional neural networks.First,a supervised hyperspectral image unmixing based on a one-dimensional convolutional neural network is proposed.The method consists of two stages: the first stage uses the convolution to extract features from the input pixels,and the second stage the abundance value is obtained by fuzzy unmixing of the extracted features.The end-to-end network structure and richer features extracted by this method improve the accuracy of unmixing.And on the basis of one-dimensional convolution network,a supervised hyperspectral image unmixing method based on three-dimensional convolutional neural network is proposed.This method simultaneously extracts information in spatial and spectral dimensions through the operation of three-dimensional convolution.This method can fuse multi-dimensional information and effectively use the input pixel neighborhood information.This method has high accuracy.(2)We propose an unsupervised hyperspectral image unmixing method based on asymmetric convolutional autoencoder network.The deep learning unmixing method has changed from supervised to unsupervised,which makes the application of unmixing more extensive.This method uses an asymmetric convolutional autoencoder network to simulate a linear unmixing model of hyperspectral images for unsupervised unmixing,add unmixing constraints to the network,and finally can obtain the endmember and abundance values of the hyperspectral image at the same time.And it can achieve efficient high-precision unmixing without the need for class labels.(3)We propose an unsupervised hyperspectral image unmixing method,which uses an asymmetric convolutional autoencoder network based on spatial-spectrum joint self-supervision mechanism.A spatial-spectrum joint self-supervision mechanism is added to the asymmetric convolutional network.The spatial-spectrum joint self-supervision mechanism can extract long-distance dependencies and feature information in two dimensions of space and spectral,fuse the information extracted in the two dimensions to add spatial information,and make full use of the global information of the input pixel patch to improve the unmixing accuracy.