Study On The Hyperspectral Anomaly Detection Based On Spectral Reconstruction And Adversarial Learning

Anomaly detection can be utilize to detect in hyperspectral image by accurate background modeling without any spectral prior information of target and background,which is widely used in precision agriculture,environmental monitoring,national defense security and other aspects etc..However,with the rapid development of hyperspectral remote sensing technology,the higher-dimensional spectra,more redundant features and finer background information of hyperspectral images bring new challenges to the accurate detection of hyperspectral anomaly targets.The spectral feature validity,hyperspectral background complexity,spectral similarity between background and anomaly pixel are still the barrier of the application and development of hyperspectral anomaly detection.In view of the above issues,the interrelationships between spectral feature selection and anomaly detection are modelled in this thesis with the theory of sparse representation model and deep generative learning.Moreover,discriminative spectral feature selection,accurate modeling of complex backgrounds,and separability of background and anomaly targets are explored.The contributions of this work are summarized as follows:(1)To address the issues of spectral high correlation and redundancy,the spectral feature extraction method based on structure tensor and the dual adversarial autoencoder network has been constructed.The correlation between spectral feature selection and anomaly detection is established.In order to select spectral features which can represent fine spatial structure,the structure tensor is modified by noise removal methods to obtain more discriminative spectral features.Moreover,the dual adversarial autoencoder network is constructed by including adversarial learning method to extract deep spectral features.The proposed model can accurately learn the statistical distribution of background samples,by which the effective features can be extract and the separability of background and anomalous targets can be improved.(2)A novel anomaly detection algorithm based on unsupervised nearest regularized subspace and spectral space reconstruction has been proposed,which can solve the spectral similarity between the anomaly target and the background pixels.Firstly,the different relative positions between the center pixel and spatial neighborhood pixels are analyzed,and the original spectrum is reconstructed by combining the spatial information and the non-negative parameters of the absolute distance,which can improve the separability of background and anomaly.In addition,by adding spectral distance weight and spatial distance weight,the adaptive weight expression is designed to improve the anomaly detection performance of the sparse representation model.Experimental results show that the proposed algorithm is superior to other anomaly detection algorithms on the four hyperspectral datasets.(3)A dual adversarial autoencoder and spectral distance adaptive weight anomaly detection algorithm has been proposed.Firstly,the effectiveness of spectral depth features and the accuracy of spectral reconstruction is improved by adding a latent space discriminator and a sample discriminator.Combined with the tight coupling training of encoder,decoder,latent space discriminator and sample discriminator,the mean error loss and spectral angular distance loss are included to ensure the high consistency between the original spectrum and the reconstructed spectrum of the background sample.In addition,the anomaly fraction is optimized for the anomaly detection combining the reconstruction error and spectral depth characteristics.The experimental results show that the generated spectra of the background samples using the dual adversarial network are highly consistent with the original spectra,while the generated spectra of the abnormal samples are different.Moreover,the proposed method outperforms other counterpart algorithms,which demonstrates the effectiveness of the proposed method.There are 42 figures,16 tables,and 133 references in this thesis.