Hyperspectral Remote Sensing Image Anomaly Detection Based On Representation Learning

Hyperspectral anomaly detection aims at finding pixels or subpixels which have different spectral signatures compared with their neighbor background without any prior information.There are many challenges in hyperspectral anomaly detection,such as the lack of spectral prior about anomaly,complex background,and unbalanced distribution between the background and anomaly.To tackle these problems,the methods based on representation learning have been proposed.As the prior information is of great important in representation learning,we propose to introduce the prior information into representation model for hyperspectral anomaly detection.The main contributions of this thesis can be summarized as follows.(1)Hyperspectral anomaly detection based on low-rank representation with projection and dictionary construction.Low-rank representation-based methods have attracted extensive attentions and achieved promising performances in hyperspectral anomaly detection.These methods assume that the hyperspectral data can be decomposed into two parts: the low-rank component representing the background and the residual part indicating the anomaly.In order to improve the separability of the background and anomaly,we propose a novel hyperspectral anomaly detection based on low-rank representation with dictionary construction and projection.To construct a robust dictionary that contains all categories of the background objects whilst excluding the anomaly's influence,we adopt a superpixelbased tensor low-rank decomposition method to generate a comprehensive and pure background dictionary.Considering the spectral redundancy in the hyperspectral data,the projection matrix is introduced to the low-rank representation to project the original data into a low-dimensional feature space to better separate the anomaly and the background.Experimental results on four real hyperspectral datasets show that the proposed anomaly detection method outperforms the other anomaly detectors.(2)Spectral-difference low-rank dictionary learning for hyperspectral anomaly detection.Low-rank dictionary learning plays an important role in exploiting the low-rank prior of background for hyperspectral image(HSI)anomaly detection.Thus the low-rank dictionary learning is introduced to learn a dictionary which can reconstruct the background positively,whilst anomaly cannot.Considering the high correlation of data especially between the adjacent bands,we resort to spectral-difference low-rank dictionary learning for global background modeling which can fully exploit the low-rank prior of background.Then the residual matrix is used to distinguish anomaly.Different from existing anomaly detection methods based on dictionary construction which is constructed or learned in a separated step,our proposed model can simultaneously learn the dictionary and separate anomaly by iterative learning.The experimental results on five real hyperspectral datasets demonstrate the superior performance of the proposed method compared with other state-of-the-art methods.(3)Low-rank dictionary learning with structural constraint for hyperspectral anomaly detection.To detect anomaly,many methods for background representation have been proposed.However,the prior information about both background and anomaly is not fully explored in these methods.To tackle this issue,we propose to combine low-rank dictionary learning with total variation regularization for hyperspectral anomaly detection.To be specific,the low-rank dictionary learning is introduced for background representation to explore the low-rank prior of background.To exploit the smooth structural characteristic of background in spatial,we introduce the total variation regularization on coefficients matrix for better background representation learning for anomaly detection.The experiments on three real datasets demonstrate the effectiveness of the proposed method compared with state-of-the-art methods.(4)Non-local tensor low-rank decomposition for hyperspectral anomaly detection.We presents a novel method for hyperspectral anomaly detection considering the spectral redundancy and exploiting spectral-spatial information at the same time.We proposed a nonlocal tensor low-rank decomposition combing with dimensionality reduction framework.Firstly,k-means++ algorithm is implemented to spectral bands and centers of each group are selected to reduce the HSI dimensionality in spectral direction.To jointly utilize spectralspatial information,the cubic data(two spatial dimensions and one spectral dimension)is treated as a 3-order tensor.Then the non-local self-similarity is fully explored in our method.For the reason to reduce the ringing artifacts caused by over-lapped segmentation in reconstruction,we introduce the hyper-Laplacian constrained tensor low-rank decomposition and we get the separated background and residual parts.Finally,to eliminate the effect of Gaussian noise,we use local-RX basic detector to detect the residual matrix.The experimental results demonstrate that the proposed method can exploit the spatial and spectral characteristics of data to improve the detection performance.(5)Low-rank constrained autoencoder for hyperspectral anomaly detection.Recently,deep learning has attracted increasing interest as the latent features of the data can be exploited.Suppose that the background is the main component and with low-rank prior,a low-rank constrained autoencoder method is proposed in this paper.The low-rank constraint applied on the latent representation of autoencoder can explore the low-rank prior in deep feature space,and the background can be well reconstructed by the decoded data.Then the residual part between the input data and the decoded data is used to discriminate anomaly.The proposed method can fully explore the deep features of data,and the obtained detection results are superior to the other methods.