Hyperspectral Image Unmixing Based On Auto-encoder Network

Hyperspectral remote sensing images not only record the spatial information on a twodimensional plane,but also contain hundreds of channels of spectral data.Due to the limited spatial resolution,It is a common fact that hyperspectral pixels are the mixture of multiple substances.It is a key hyperspectral remote sensing technology to extract the spectra of mixed substances and their corresponding proportions from hyperspectral pixels,which is the hyperspectral unmixing.With the increasingly mature hyperspectral imaging technology,unmixing methods have made considerable progress and development.Among them,unsupervised blind source unmixing is a common and challenging problem.The application of autoencoder networks to unsupervised unmixing is a popular research in recent years.However,due to the limitation and low rank of training data,unmixing networks based on autoencoders cannot train too many parameters and complex networks.Most of their structures are single-layer fully connected forms,which not only fail to make full use of image spectral information.also leads to the training of the network can only be input in the form of a single pixel,thus ignoring the spatial information in the original hyperspectral image.In view of this,this article improves the unmixing autoencoder network from the following three aspects:(1)A multi-to-one pixel reconstruction strategy based on heat kenerl descriptors is proposed to extract endmembers.First,select the neighboring pixels of the center pixel in the form of a square sliding window,and input them into the fully connected encoding layer one by one to find their abundance,and then use a heat kernel function to fuse them into an abundance from the perspective of spatial weighting,and finally input the it to reconstruct the center pixel.This way of enhancing the input samples utilizes spatial information,improves the coding ability of the network,and improves the endmember extraction effect of the network.(2)A deep spectral convolution network is proposed to estimate the abundance.First,design several small convolutional layers to extract spectral features,then design a two-layer fully connected layer to fuse the extracted features into abundance,and finally reconstruct the original pixels through the extracted endmembers.Convolutional networks have good feature extraction capabilities for highly correlated spectral channel data,effectively utilize spectral information,and ultimately improve the abundance estimation capabilities of the network.(3)A super-pixel segmentation method based on abundance is proposed to extract spatial information from hyperspectral images.First,the rougher pixel abundance is estimated by traditional methods,and then the spatial distance between it and the pixel is combined into a feature space.The hyperspectral image is divided into pixel clusters,namely super-pixels,using the SLIC algorithm,and finally As the neighborhood of the center pixel,enter the endmember extraction framework in(1)for endmember extraction.The super-pixel segmentation adaptively selects the homogeneous and uniform neighborhood of the center pixel,which overcomes the problem of over-smooth boundary caused by the fixed sliding window,and further explores the potential ability of spatial information to extract endmembers.The method proposed in this paper is verified on two real hyperspectral data sets.Experiments show that the method in this paper not only strengthens the endmember extraction ability of the autoencoder,but also improves the estimation accuracy of abundance and achieves better unmixing.effect.