Research On Spectral-spatial Features-based Anomaly Detection Algorithms For Hyperspectral Images

Hyperspectral remote sensing is a technique that uses dozens or even hundreds of narrow and continuous spectral channels to image features continuously,which allows hyperspectral images to contain rich spatial and spectral information.Hyperspectral image anomaly detection has been widely concerned because it does not need any prior information of the scene.However,the complexity of hyperspectral image data and the variability of anomalies also bring great challenges to the current anomaly detection technology.The main problems are as follows:(1)The composition of background in hyperspectral image is complex,so it is difficult to model the background accurately by traditional statistical methods;(2)The shape and size of the anomaly is changeable,so it is difficult to detect and identify anomaly by a uniform pattern;(3)Hyperspectral image data has high dimension,strong correlation between bands,and some information redundancy,which leads to the detection effect of high-dimensional data is not ideal;(4)Most hyperspectral anomaly detection algorithms have high false alarm rate with high detection accuracy.Therefore,according to the characteristics of hyperspectral data,making full use of the spatial and spectral information in the image,designing an effective hyperspectral anomaly detection algorithm has important theoretical significance and practical value.The main contents and innovations of this paper are as follows:Aiming at the problem that it is difficult to model the complex background accurately by traditional statistical methods in hyperspectral images,and to detect anomaly with various shape and size by a unified pattern,this paper proposes a discriminative reconstruction for hyperspectral anomaly detection with spectral learning.According to the characteristics of low probability small targets and Gaussian background in hyperspectral data,a discriminative reconstruction network based on spectral constraints is constructed.During the network training,the encoder is constrained to produce latent features that obey the unit Gaussian distribution,which helps the decoder to reconstruct images that resemble the input background.At the same time,the spectral angular distance term is introduced into the loss function of the network to constrain the network to generate the reconstructed image with greater spectral similarity to the input.Through verification,it is found that the spectral difference on the anomalous pixel between the input and reconstruction is larger than that of the background pixel.Therefore,the spectral angular distance map between the network input and reconstruction is used as the input of the subsequent anomaly detection,and the background is suppressed and the anomaly is retained by attribute filtering and self-suppression processing.By reconstructing the background and using the spectral difference between input and reconstruction for anomaly detection,this method solves the problems that is difficult to establish an accurate background representation model,and to extract a uniform pattern for anomaly.And this method realizes the combination of spectral and spatial dimensions,and effectively avoids the detection deviation caused by the single use of spectral or spatial information.The experimental results show that the average detection accuracy of the proposed method is improved by about 3% compared to other comparative algorithms.Aiming at the problems of hyperspectral image with high spectral dimension,strong correlation between bands and large data redundancy,this paper proposes a dual feature extraction network for hyperspectral anomaly detection.In order to extract deep features in the image,a dual feature extraction network is constructed.The network is composed of two autoencoders cascaded.The two networks are made to learn potential features of the original data and the background by jointly applying end-to-end discriminative learning losses,including adversarial learning and Gaussian constraint learning,on the two networks,respectively,while achieving the goal of peace reduction and data redundancy reduction.Then the spatial differences and spectral differences between the two feature images are exploited for anomaly extraction to obtain the initial spatial and spectral dimensional detection results.Aiming at the problem that most hyperspectral anomaly detection algorithms have a high false alarm rate even with high detection accuracy,the final detection result is obtained by combining the two initial detection results through Hadamard product operation.Experiments on real hyperspectral datasets show that the extraction of deep features of hyperspectral images can improve the detection effect,and the combination of spatial dimension and spectral dimension can effectively suppress the background.The average false alarm rate is nearly an order of magnitude smaller than other algorithms with higher detection accuracy than all other comparison algorithms.