A highorder statistical tensor based algorithm for anomaly detection in hyperspectral imagery.  
Jump to Full Text  
MedLine Citation:

PMID: 25366706 Owner: NLM Status: InDataReview 
Abstract/OtherAbstract:

Recently, highorder statistics have received more and more interest in the field of hyperspectral anomaly detection. However, most of the existing highorder statistics based anomaly detection methods require stepwise iterations since they are the direct applications of blind source separation. Moreover, these methods usually produce multiple detection maps rather than a single anomaly distribution image. In this study, we exploit the concept of coskewness tensor and propose a new anomaly detection method, which is called COSD (coskewness detector). COSD does not need iteration and can produce single detection map. The experiments based on both simulated and real hyperspectral data sets verify the effectiveness of our algorithm. 
Authors:

Xiurui Geng; Kang Sun; Luyan Ji; Yongchao Zhao 
Related Documents
:

17736826  Refuge theory and biological control. 22559876  Quantitatively integrating molecular structure and bioactivity profile evidence into dr... 20956326  Mental models and human reasoning. 17555376  The papm, diffusion theory, and violent death surveillance. 16754056  The philosophy, nature and nomenclature of tissue engineering and biomimetics. 17073326  Adaptive change detection in heart rate trend monitoring in anesthetized children. 
Publication Detail:

Type: Journal Article Date: 20141104 
Journal Detail:

Title: Scientific reports Volume: 4 ISSN: 20452322 ISO Abbreviation: Sci Rep Publication Date: 2014 
Date Detail:

Created Date: 20141104 Completed Date:  Revised Date:  
Medline Journal Info:

Nlm Unique ID: 101563288 Medline TA: Sci Rep Country: England 
Other Details:

Languages: eng Pagination: 6869 Citation Subset: IM 
Export Citation:

APA/MLA Format Download EndNote Download BibTex 
MeSH Terms  
Descriptor/Qualifier:

Full Text  
Journal Information Journal ID (nlmta): Sci Rep Journal ID (isoabbrev): Sci Rep ISSN: 20452322 Publisher: Nature Publishing Group 
Article Information Download PDF Copyright © 2014, Macmillan Publishers Limited. All rights reserved openaccess: Received Day: 24 Month: 07 Year: 2014 Accepted Day: 10 Month: 10 Year: 2014 Electronic publication date: Day: 04 Month: 11 Year: 2014 collection publication date: Year: 2014 Volume: 4Elocation ID: 6869 PubMed Id: 25366706 ID: 4219173 Publisher Item Identifier: srep06869 DOI: 10.1038/srep06869 
A HighOrder Statistical Tensor Based Algorithm for Anomaly Detection in Hyperspectral Imagery  
Xiurui Genga1  
Kang Sun1  
Luyan Ji2  
Yongchao Zhao1  
1Key Laboratory of Technology in Geospatial Information Processing and Application System, Institute of Electronics, Chinese Academy of Sciences, Beijing, China 

2Ministry of Education Key Laboratory for Earth System Modelling, Centre for Earth System Science, Tsinghua University, Beijing, China 

agengxr@sina.com.cn 
The research on the anomaly detection of hyperspectral data has drawn much attention recently in many fields^{1}^{, 2}^{, 3}. The socalled anomaly detection is basically to find out “abnormal” pixels from an image where the targets and their associated background are both unknown. Many anomaly detection methods have been proposed, among which RX Detector (RXD)^{4}^{, 5}^{, 6} is the most typical one. It has been applied to both multi and hyperspectral successfully in terms of anomaly detection. In fact, the expression of RXD is equivalent to the Mahalanobis distance. There are many anomaly detection operators derived from RXD, such as modified RX (MRX), normalized RX (NRX), weighted RX, Causal RX^{7} and adaptive causal anomaly detector algorithm (ACAD)^{8}. The low probability detector (LPD)^{9} is another anomaly detector used frequently. LPD determines whether a pixel is abnormal or not according to the relationship between any pixel of the image and the unity vector multiplied by the inversion of the sample autocorrelation matrix. The uniform target detector (UTD) is an evolved version of LPD which has a translational shift of the origin of the image to the mean vector. Kwon^{10} proposed a new anomaly detection method, dual windowbased eigen separation transform anomaly detector (DWEST). DWEST model involves two local windows, namely inner and outer windows, which are designed to maximize the separation between anomalies and background. The inner window is used to detect the anomalies presented in it, while the outer window is used to model the background of the anomalies. By moving these two local windows in an image, we can calculate the local mean and covariance matrix for each window and their differences. Consequently, anomalies can be extracted by projecting the differential mean between two windows onto the eigenvector associated with the largest positive eigenvalue of the differential covariance matrix. Similar to DWEST, nested spatial windowbased target detector (NSWTD) is presented in^{11}. NSWTD model involves three nested local windows, namely, inner, middle and outer windows. The first two windows are used to extract the smallest and largest anomalies respectively, while the outer window is used to model the local background. Moreover, the other key difference of this model from the DWEST and RXbased algorithms is to use the orthogonal projection divergence (OPD) instead of eigenvector projection or sample covariance matrix as a measurement. Based on a nonparametric model, the combined FTest anomaly detector (CFT) is presented by Rosario^{12}. The main assumption of this method has an asymptotic behavior of Fisher's F distribution for data sets which are examined by a common statistical test. Some other anomaly detection methods can be seen at^{13}^{, 14}^{, 15}.
The anomaly detectors mentioned above are all basically conducted from the statistical perspective. The statistics used include firstorder statistics (e.g., mean vector) and secondorder statistics (e.g., covariance matrix). It is not sufficient to use the firstorder and the secondorder statistics since the distributions of scatter points in feature space for most real images are not normal distributions.
In this study, the thirdorder statistical tensor is introduced to extract anomalies. In fact, there are many approaches in existing literature for anomaly detection using highorder statistics^{16}^{, 17}. However, all these approaches are the direct applications of the blind signal separation (BSS) methods (e.g., FastICA), which generally involve stepbystep iterations to reach the optimal solution. As a result, they are apt to be trapped in local minima. In order to address the convergence issue, Geng^{18} introduced the concept of coskewness tensor to hyperspectral data analysis and proposed a target detection method based on higher order singular value detection (HOSVD). Nevertheless, both the BSSSbased techniques and HOSVD are in the domain of feature extraction, aiming at the extraction of not only the anomalies but also the other independent components in the image.
In this paper, combining the concept of thirdorder statistical tensor and the idea of RXD, we present a new anomaly detection method termed coskewness tensor detector (COSD). The proposed method can directly get the distribution of the anomaly of a hyperspectral image without any iteration, which can therefore avoid problems of the BSSbased methods.
Although there are many anomaly detection algorithms based on the 2ndorder statistic, they are all generally derived from RXD. Similarly, we can also derive corresponding algorithms based on COSD. Therefore, the experiments in this study only focus on comparing the performances between COSD, FastICA and RXD. In order to facilitate the comparison between COSD and FastICA, we used skewness as a measure of nonGaussianity in FastICA. In addition, since FastICA can produce a lot of independent components, we just select the one with the greatest skewness as the detection result.
The simulated image of two bands with 50 × 50 pixels was first used in this experiment. Simulation data consists of two parts, abnormal targets (8*8) and their background. The background pixels fit a Gaussian distribution. The target was located in the upper left corner of the image and randomly scattered outside the background in the feature space of the image (see Fig. 1).
The comparisons between RXD, FastICA and COSD are given in Fig. 2. By a visual comparative analysis, the performance of COSD is superior to that of RXD. Fig. 3 shows the receiver operating characteristic (ROC) curve of detection rate versus false alarm rate for the three algorithms (see reference [15] for their definitions of detection rate and false alarm rate). Clearly, the detection performance of COSD is comparable to that of FastICA, and both are better than that of RXD.
The hyperspectral data of 100*100 pixels from OMISII (Operational Modular Imaging Spectrometer) is used to test these methods. The hyperspectral imaging system was developed by Shanghai Institute of Technical Physic, Chinese Academy of Sciences (SITP). The data, which was acquired by the Aerial Photogrammetry and Remote Sensing Bureau in Xi'an, China in 2003, includes 64 bands from visible to thermal infrared with 3.6 m spatial resolution and 10 nm spectral resolution in the visible and near infrared region (60 bands). There were small manmade targets simulated as common objects within the scene, which were distributed at two locations around the top right corner of the image (marked by the rectangles in Fig. 4 a), consisting of tens of pixels. From the true color (approximately) composition image (see Fig. 4 a), it is hard to find any information of the targets in the rectangles.
From Fig. 4b, we can see two small bright blocks at the top right corner, which are the manmade targets. It indicates that RXD can distinguish them as the abnormal pixels from the background. Fig. 4c is the result of COSD, where the two manmade targets are highlighted significantly as the abnormal pixels while the rest is greatly suppressed as the background. The result of NCOSD in Fig. 4d indicates that it is good enough to extract anomalies by only using the skewness information. Fig. 5 shows the ROC (receiver operating characteristic) curve of detection rate versus the false alarm rate for both RXD and COSD. It illustrates that the detection capability of COSD has a significant advantage over that of RXD.
Now we turn to the comparison between FastICA and COSD. Usually, FastICA gets all the independent components through iterations, and the one with the maximum skewness is chosen as the anomaly detection result. However, due to the local optimum, the first independent component of FastICA is not always corresponding to the global maximum skewness, thus different surface objects may be detected (see Fig. 6, which shows the inconsistency of anomaly detections. Besides, the skewnessbased FastICA does not have a global convergence. Nevertheless, our COSD does not have these problems.
In this study we proposed a new method of using highorder statistic tensor to detect the anomaly of a hyperspectral image and analyzed in detail the application of the skewness tensor in the anomaly detection. Compared to the traditional methods based on the secondorder statistics, COSD has a better capacity to extract the abnormal objects. Moreover, COSD can directly get the distribution of the abnormal objects by using a higherorder statistic tensor, compared to the traditional methods based on blind signal separation methods. Since COSD does not need iteration, it can avoid the shortcomings of the blind signal separation methods. By the experiment with simulated data, it shows that the detection performance of the COSD is better than that of RXD. In the experiment using real hyperspectral data, it is illustrated that the COSD can highlight the manmade targets as the anomalies out of the image successfully. It is noteworthy that, the obtained abnormal pixels might not be the ones of interest due to the uncertainty of the abnormal pixels in an image. However, the introduction of a higherorder statistic tensor will benefit a lot in the anomaly detection for hyperspectral images.
Although the introduction of coskewness tensor benefits anomaly detection a lot, COSD may suffer from larger computational complexity. Figure 7 shows the computational complexity (measured by the required float operations, flops) for RXD, and COSD. Assume the size of the hyperspectral data is N pixels and L bands. The flops required for RXD is while that for COSD is . We can see that from Fig. 7, COSD is more sensitive to the number of bands. When L is relatively large (for instance, >50), the computational complexity of COSD is about times of that of RXD.
It is noticeable that since the COSD method can be considered as the extension of the RXD method in formula expression from the 2ndorder statistics (covariance matrix) to the 3rdorder statistics (coskewness tensor), all the other 2ndorder statistics based anomaly detection methods (such as modified RX, weighted RX, causal RX, DWEST, NSWTD) can be simply extended to those 3rdorder statistics based ones or even higherorder statistics based ones. The advantage of our COSD algorithm in detecting anomaly of hyperspectral image ensures a rationality of this extension.
In conclusion, the anomalies generally show strong features in the highorder statistics. Thus, this paper presents a new anomaly detection method COSD based on thirdorder statistical tensor. Formally, the COSD is the natural extension of RXD from secondorder to thirdorder statistics. Essentially, the COSD take full advantage of angle information, which ensures the validity of COSD.
The RXD is a detector proposed by Reed and Yu^{4}. For each pixel vector r in an image, RXD can be implemented by a operator specified by
From Eq. (2) we can see that, δ_{RXD}(r) is actually the Euclidean distance of r and μ of the whitened image.
A real morder ndimensional tensor consists of n^{m} real entries^{19}, represented as where i_{j} = 1,…,n for j = 1,…, m. Fig. 8 shows an example of thirdorder 4dimensional tensor. The tensor is supersymmetric if its entries are invariant under any permutation of their indices^{20}, or mathematically, a_{ijk} = a_{ikj} = a_{jik} = a_{jki} = a_{kij} = a_{kji}.
The tensor defines an mthdegree homogeneous polynomial
For a hyperspectral image data set S = {r_{1},…, r_{N}}, its morder cumulant matrix (tensor) is defined as:
For a hyperspectral image data set , suppose that and , where I is L*L unit matrix. It means that each band of the image has a variance of 1, and the correlation coefficient between the bands is zero. That is to say, the hyperspectral data has been normalized. It is not difficult to reach these two conditions. If the mean vector of the image is not zero, it can be achieved by moving the origin of the image to the mean vector. Besides, the real hyperspectral image can meet the second condition by data whitening.
Here, we propose a new anomaly detector, named highorder statistic detector (HOSD), which is defined as follows:
Like RXD, the coskewness tensor based anomaly detector also uses all pixels of the image by eq. (7) to get a gray image, where the anomalies of the image will appear to be very bright or dark. The dark pixels in the gray image are caused by the negative values from Eq. (7). Therefore, we usually determine anomaly pixel by using the absolute value of the gray image.
In Eq. (7), if is a unit vector, then is the skewness of the image against the direction of . Eq.(7) can be transformed as follows
From Eq. (8), we can see that the anomaly extraction using Eq. (7) is mainly dependent on two indices: One is the skewness and the other is the cube of the 2norms. If we eliminate the item from Eq.(8), the COSD operator becomes a normalized COSD (NCOSD) operator
Practically, is the skewness of image data against the direction of .
Let us assume that hyperspectral image is composed of background and abnormal pixels,
In general, the number of abnormal pixels in an image is very small, thus μ_{b} ≈ μ = 0, where is the mean vector of the background image. Accordingly, can be approximately considered as the coskewness tensor of the background image. If we assume that the background image fits a Gaussian distribution, we can get
It means that all the elements of are close to zero. Thus the skewness of the image in the direction of can be expressed as
Since is an image after centering and whitening, the RX operator can be expressed as
From eq.(15–17), it can be seen that there are distinguished differences among RX, NCOSD and COSD. Specifically, RXD depends only on the distance between a pixel and the origin in the feature space of the whitened image. Only when abnormal pixels are far away from the origin and all the background pixels are relatively close to the origin, RXD can achieve a good anomaly detection result. NCOSD is based on the skewness of the image, and its detection performance is dependent on angles between an abnormal pixel and all the background pixels. When all the angles are large, NCOSD can get a good detection result. As for COSD, it does not only take the distance into account, but also the angle. So it can overcome shortcomings of both RXD and NCOSD, both of which focus only on one single index.
X.G. conceived the idea. X.G., K.S., L.J. and Y.Z. designed and performed the experiments and analyzed the data. X.G., K.S. and L.J. wrote the main manuscript text. All authors reviewed the manuscript.
We are very grateful to Dr. Suhong Liu from Beijing Normal University, China for her great help in language.
References
Tan K.,, Li E.,, Du Q., & Du P.,Hyperspectral Image Classification Using Band Selection and Morphological Profiles. IEEE J. Sel. Topics Appl. Earth Observ.7, 40–48 (Year: 2014).  
Chang C.I.,Hyperspectral Imaging: Techniques for Spectral Detection and Classification (Plenum Press, New York, Year: 2003).  
Manolakis D., & Shaw G.,Detection algorithoms for hyperspectral imaging applications. IEEE Signal Process Mag19, 29–43 (Year: 2002).  
Reed I. S., & Yu X.,Adaptive multipleband CFAR detection of an optical pattern with unknown spectral distribution. IEEE Trans. Acoust Speech Signal Process38, 1760–1770 (Year: 1990).  
Yu X.,, Reed I. S., & Stocker A. D.,Comparative performance analysis of adaptive multispectral detectors. IEEE Trans. Signal Process41, 2639–2656 (Year: 1993).  
Yu X.,, Hoff L. E.,, Reed I. S.,, Chen A. M., & Stotts L. B.,Automatic target detection and recognition in multiband imagery: A unified ML detection and estimation approach. IEEE Trans. Image Process6, 143–156 (Year: 1997).18282885  
Chang C. I., & Chiang S. S.,Anomaly detection and classification for hyperspectral imagery. IEEE Trans. Geosci. Remote Sens40, 1314–1325 (Year: 2002).  
Hsueh M., & Chang C. I.,Adaptive causal anomaly detection for hyperspectral imagery. In: IEEE International Geoscience & Remote Sensing Symposium5, 3222–3224 (Year: 2004).  
Harsanyi J. C.,Detection and classification of subpixel spectral signatures in hyperspectral image sequences Ph.D thesis, Baltimore, (Year: 1993).  
Kwon H.,Adaptive anomaly detection using subspace separation for hyperspectral imagery. Opt Eng42, 3342–3351 (Year: 2003).  
Liu W., & Chang C. I.,A nested spatial windowbased approach to target detection for hyperspectral imagery. In: IEEE International Geoscience and Remote Sensing Symposium20–24 (Alaska, Year: 2004).  
Rosario D.,A nonparametric Fdistribution anomaly detector for hyperspectral imagery. In: Aerospace Conference, IEEE2022–2029 (Year: 2005).  
Schwerizer S. M., & Moura J. M. F.,Efficient detection in hyperspectral imagery. IEEE Trans. Image Process10, 584–597 (Year: 2001).18249648  
Gu Y.,, Liu Y., & Zhang Y.,A selective KPCA algorithm based on highorder statistics for anomaly detection in hyperspectral imagery. IEEE Geosci. Remote Sens Lett5, 43–47 (Year: 2008).  
Chiang S.S.,, Chang C.I., & Ginsberg I. W.,Unsupervised target detection in hyperspectral images using projection pursuit. IEEE Trans. Geosci. Remote Sens39, 1380–1391 (Year: 2001).  
Xun L., & Fang Y.,Anomaly Detection Based on Highorder Statistics in Hyperspectral Imagery. In: The Sixth World Congress on Intelligent Control and Automation2, 10416–10419 (Year: 2006).  
Ren H., & Chang Y.L.,A Parallel Approach for Initialization of HighOrder Statistics Anomaly Detection in hyperspectral imagery. In: IEEE International Geoscience & Remote Sensing Symposium2, II1017II1020 (Year: 2008).  
Geng X.,, Ji L.,, Zhao Y., & Wang F.,A small target detection method for the hyperpectral image based on highoder singular value decomposition (HOSVD). IEEE Geosci. Remote Sens Lett10, 1305–1308 (Year: 2013).  
Qi L.,Eigenvalues of a real supersymmetric tensor. J. Symbolic Comput.40, 1302–1324 (Year: 2005).  
Kofidis E., & Regalia P.,On the best rank1 approximation of higherorder supersymmetric tensors. SIAM J Matrix Anal A23, 863–884 (Year: 2002). 
Article Categories:

Previous Document: The effect of Ce ion substituted OMS2 nanostructure in catalytic activity for benzene oxidation.
Next Document: The impact of radiologists' expertise on screen results decisions in a CT lung cancer screening tria...