Research On Hyperspectral Anomaly Detection Based On Low Rank Representation

Hyperspectral remote sensing images have a high resolution.As a combination of high spectral resolution and two-dimensional image,its features enable it to acquire the geometric information of the image scene while also acquiring rich spectral information,which fully reflects the essential attributes of the materials and subtle difference between materials.Many objects that cannot be distinguished in two-dimensional images or multispectral images can be clearly identified by hyperspectral images.This provides the theoretical support for the application of target detection by hyperspectral images.As a kind of unsupervised target detection,anomaly detection mainly detects anomalous targets with large spectral differences of different ground objects without prior information.It has strong practical value and has become a research hotspot in the field of hyperspectral remote sensing.Traditional anomaly detection algorithms are mainly based on statistical models,that is,assuming that the image background obeys a specific probability distribution,but in practice,the background usually does not meet this assumption,which makes the traditional methods have big limitations.In addition,due to the limitation of the spatial resolution of the hyperspectral sensor,there are many mixed pixels in the image data.Mixed pixels will also have a great impact on detecting anomaly targets.Therefore,at present,how to avoid discussing the background distribution of the image when performing anomaly detection;how to combine the spatial and spectral information of the image more effectively;and how to reduce the influence of mixed pixels on the detection.These are all the key points to be solved to improve target detection.Based on the research of the development status at home and abroad,this thesis deeply explores the sparse representation and low rank representation methods from compressed sensing theory,and explores how to use the low rank representation method on the problem of anomaly detection.The knowledge of tensor decomposition,deep learning and graph theory in different fields are also combined so as to effectively solve the above problems,and has achieved good results.The main research work and structure of this paper are as follows:(1)This thesis introduces the definition of remote sensing technology and the specific differences from two-dimensional images,as well as the development status of hyperspectral remote sensing,including the development of hyperspectral imaging technology and the development of hyperspectral image processing technology.Then the definition of hyperspectral anomaly detection is described.The development status of anomaly detection technology and the shortcomings of classic algorithms are also expounded.Then it leads to the hyperspectral anomaly detection algorithms based on sparse representation.Sparse representation does not need to set background assumptions in advance,and the mathematical principles and differences of the two models of sparse representation and low-rank representation are elaborated in the thesis.At the same time,the paper also demonstrates the rationality and specific application of low-rank representation in anomaly detection.(2)For the mixed pixel problem caused by the limited resolution of the sensor,the paper proposes a hyperspectral anomaly detection algorithm based on endmember extraction and low rank decomposition.Firstly,each endmember and the corresponding abundance map of the original image are extracted by using the sequential maximum angle convex cone,and the hyperspectral pixels are mapped to a separate subspace,so as to increase the separation of background and anomalous components and improve the target detection rate.Next,a background dictionary construction method based on tensor decomposition and clustering is proposed.Tensor decomposition is used to separate the prominent components in the data,including anomalous components,high-frequency noise and background elements similar to the anomalies.The main and relatively uniform background is preserved.Then the background is classified unsupervised by clustering algorithm,and the most representative elements are selected as dictionary atoms in each cluster.The effective dictionary is used in the low rank decomposition model,and the residual matrix is calculated to obtain the detection results.This algorithm can effectively avoid the assumption of background probability distribution in traditional statistical model,and achieve the suppression of background and noise,highlighting the abnormal target.(3)The method of endmember extraction and abundance map construction based on sequential maximum angle convex cone described above is characterized by high speed and high reliability,but the disadvantage is that in abundance maps,the estimated probability of which feature each pixel belongs to is not very accurate.Regarding to this problem,we propose an anomaly detection algorithm based on convolutional neural network unmixing and low rank decomposition.Firstly,we design a convolutional neural network model to generate the abundance matrices of hyperspectral images.Next,a density-based clustering method is used to construct an effective dictionary to accurately estimate the background and anomalous components.This kind of unsupervised clustering based on density does not need to know the specific number of categories to be separated in advance,and can identify noise points meanwhile.By calculating the residuals between the abundance matrix and the estimated background,the anomalous components can be obtained and the noise interference can also be minimized.(4)In addition to extracting spectral features through endmember extraction,we are also committed to discovering the overall features of hyperspectral images,such as texture features,and apply the features to anomaly detection.In the thesis,a hyperspectral anomaly detection algorithm based on low rank representation of graph dictionary and texture feature extraction is proposed.We have discussed that when the low rank representation model is utilized to decompose the original hyperspectral data,the image data is considered to be distributed in multiple subspaces,and the l₂₁ norm should be used to constrain the objective function so that the function has an analytic solution.When the feature vectors are used as the input data,the input is considered to be located in a single subspace.In this case,the constraint of sparse matrix can be relaxed appropriately so that there can be more non-zero elements in the matrix.Therefore,l₁ norm is used to allow some non-zero values in the columns of sparse matrix,which makes it more sparse.We take the texture features extracted by the gray level co-occurrence matrix and the original image data as the input data of the low rank representation model respectively,and decompose them with different constraints.Then the decomposition results are fused to improve the detection rate of the whole detector.In addition,we also design a new strategy to construct dictionary based on graph theory.The dictionary does not need to predict the parameters in advance,and can retain the most important structure in the background,which is more universal.