Hyperspectral Unmixing Based On Deep Learning

Hyperspectral images have high spectral resolution.Usually,the spectral dimensions of each pixel can reach tens or even hundreds.However,the spatial resolution of hyperspectral images is low.At the same time,due to the factors such as microscopic mixing of ground materials,multiple reflections,atmospheric scattering and the instrument itself,hyperspectral images generally have problems with mixed pixels.That is to say,the spectral signals measured by imaging spectrometer are actually the mixing of reflectivity of various substances in the scene.Therefore,spectral unmixing has become a challenging topic in the exploration of hyperspectral image data.As we all know,deep learning has made remarkable achievements recently in text classification,speech recognition,machine translation,computer vision and other fields,which is mainly due to the strong fitting and feature extraction ability of deep learning.Although deep learning has been widely used in hyperspectral image classification,it is seldom used in hyperspectral unmixing.In this paper,the network structure matching the spectral unmixing is designed by using the related model of deep learning to achieve the hyperspectral image unmixing.Firstly,this paper describes the concepts of hyperspectral image unmixing,and expatiates the current research status of spectral unmixing methods.In addition,this paper also introduces the related knowledge of deep learning,sums up the advantages of deep learning and its application in related fields.Subsequently,for the unmixing of hyperspectral images,this paper constructs two network architectures under the framework of deep learning: 1)The spectral based network.The architecture uses a fully connected neural network and the spectral vector is used as an input for unmixing.This network architecture designs specific network layers,which considers not only the abundance non-negative constraints and the sum-to-one constraints,but also the sparsity of abundance.Through the continuous updating of network model parameters,the endmember matrix and the abundance matrix can be obtained simultaneously;2)The spatial-spectral based network.The architecture further combines the convolutional neural networks to fuse the spatial information and spectral information of the hyperspectral image for unmixing.By considering the spatial neighborhood information of each pixel,the finally obtained abundance maps can show a certain smoothness and can more accurately reflect the characteristics of the ground distribution.Finally,experiments are carried out on simulated datasets and real datasets.Compared with other popular unmixing methods,the results show that the proposed network structures have better performance in hyperspectral image unmixing.