Research On Hyperspectral Nonlinear Unmixing Via Deep Autoencoder Networks

With the successful launch of the Gao Fen series,our country has maken great progress in the field of remote sensing,and a large amount of hyperspectral data needs to be processed.A piece of hyperspectral image contains hundreds of bands in the spectral dimension and tens of square kilometers in the spatial dimension.Faced with such a complex data,manual analysis is no longer sufficient,and computer image processing is necessary.One of the purposes of hyperspectral data processing is target detection,especially the detection of small targets,which is of great significance in the fields of early forest pest control and military investigation.However,affected by conditions such as sensors,communication capabilities,and atmospheric scattering,mixed pixels are widely present in hyperspectral image data,which is not conducive to the detection of small targets.Therefore,the study of unmixing algorithms for hyperspectral images has become an important work in the field of remote sensing.This paper is based on the current rapidly developing neural network,and introduces the autoencoder to hyperspectral unmixing.Using the advantages of neural network in processing large amounts of data and the powerful fitting ability to nonlinear problems,a new kind of autoencoder is proposed.The first is the unsupervised hyperspectral nonlinear unmixing network and the second is the spatialspectral unsupervised hyperspectral nonlinear unmixing network based on a 3-D convolutional autoencoder.The main contents are as follows:Research on prior estimation of nonlinear parameters based on Gaussian process.In the unmixing of hyperspectral image data,the nonlinear unmixing method has the characteristics of high accuracy and complex calculation,and the linear unmixing algorithm has the characteristics of simple calculation and low accuracy.It is unacceptable that nonlinear unmixing algorithms apply to a large hyperspectral image in terms of efficiency.Therefore,in order to combine the advantages of linear unmixing and nonlinear unmixing,we perform the preprocessing step of nonlinear region detection before unmixing.After Calculating the fitting error of the spectrum of each pixel based on statistical test,and marking it as a certain hybrid method through threshold processing,the hyperspectral image data is input into the unmixing network.Unsupervised autoencoder network for hyperspectral nonlinear unmixing.The autoencoder network can extract the hidden features of the input through the encoder,and reconstruct the input through the decoder.For hyperspectral data,the abundance corresponding to each endmember in each pixel is its intrinsic feature,which can be extracted by an encoder.The physical meaning of unmixing can be satisfied by setting the regular term on the endmembers and constraining the abundance value.The structure of the decoder in the autoencoder is optimized to enable it to perform linear and nonlinear reconstruction of hyperspectral data.and choose different reconstruction methods according to the results of nonlinear detection.When updating the network,the reconstruction error is used as the loss function.After unmixing is completed,the output of the encoder is the estimated abundance,and the weight connected to the next layer is the endmember matrix.The spatial-spectral unsupervised hyperspectral nonlinear unmixing network based on a 3-D convolutional autoencoder.Inspired by the application of autoencoders in unmixing,this paper uses convolutional autoencoder to introduce spatial information in hyperspectral data unmixing.Due to the particularity of hyperspectral data,we use a three-dimensional convolution kernel for convolution.Through convolution,the spatial relationship between the target pixel and the pixels in a certain area around is extracted,and the spatial information and spectral information are incorporated into the features extracted by the encoder.The unmixing algorithm that incorporates spatial information has advantages in the case of severe noise.