Hyperspectral Data Unmixing Method Based On Sparse Representation And Deep Learnin

Hyperspectral images（HSIs）are three-dimensional data that contain both spatial information and spectral information.In recent years,it has been widely used in many fields such as environmental monitoring,land recognition,and precision agriculture.However,limited by low spatial resolution of imaging spectrometer,a pixel usually contains multiple features.The existence of such mixed pixels hampers the accurate analysis and application of HSIs.Therefore,the unmixing of hyperspectral mixed pixels（unmixing for short）has become a key issue in the analysis of HSIs.Its purpose is to extract pure spectra（called endmembers）from the mixed pixels and calculate their corresponding proportions（called abundance）.Its difficulty lies in the morbidity of the problem caused by insufficient information.At present,sparse unmixing based on the spectral library is a representative method,but in actual situations,HSIs usually contain noise such as Gaussian,impulse and dead line,and the intensity of the noise in each band is often different,so the commonly used sparse unmixing method is not robust enough,and the accuracy of unmixing needs to be improved.It is urgent to study hyperspectral optimization model and efficient robust algorithm.On the other hand,the unmixing algorithm based on deep learning is a new research trend.The mainstream method is the deep learning unmixing method based on autoencoder,which has the characteristics of unsupervised and blind unmixing.When the endmembers and abundance are not known,it is difficult to obtain a unique solution by commonly used autoencoder networks.In view of the above problems,this thesis further studies hyperspectral sparse unmixing and deep learning unmixing methods,and optimizes the model and algorithm.The main contribution is as follows:（1）Taking into account the mixed noise in HSIs,this thesis proposes a robust unmixing model based on collaborative sparse and low-rank representation（RUn CSLRR）.This algorithm models sparse structure noise and uses ₁ norm to promote the sparsity of noise,_2,0 norm to promote the row sparsity of the abundance,and then uses the global low-rank of the abundance to mine its low-dimensional structural features,and uses the alternating direction method of multipliers method to design an efficient iterative algorithm.Finally,experiments are carried out on simulated data sets and real data sets,and prove that this algorithm performs better than the stat-of-the-art algorithms.（2）Considering the mixed noise of real hyperspectral data and the statistical characteristics of different noise intensity in each band,this thesis establishes a non-negative sparse component analysis model based on the maximum posterior probability framework;then uses ₁ norm to describe the sparsity of noise,_2,0 norm to describe the row sparsity of abundance,total variation to describe the local homogeneity and segmental smoothness of the pixels.A hyperspectral robust unmixing optimization model（RUn SCA）based on non-negative sparse component analysis is established.Experiments on simulated data and real data show that the proposed algorithm is robust to mixed noise,and the accuracy of unmixing is higher than the stat-of-the-art algorithms.（3）Aiming at the problem that data annotation of hyperspectral unmixing is very scarce and difficult,this thesis proposes a blind unmixing algorithm based on an asymmetric sparse convolutional autoencoder（ASCAE）under the unsupervised learning framework.The network consists of a multi-convolutional layer encoder and a single-layer decoder.The loss function is composed of spectral angular distance and sparse regularization terms.By imposing _1/2 norm and ₂ norm on abundance and neurons,respectively,a better solution for blind source separation is obtained.Experiments show that the algorithm can perform better blind unmixing of HSIs without sample labeling.