Académique Documents
Professionnel Documents
Culture Documents
To cite this article: Qiandong Guo, Ruiliang Pu, Lianru Gao & Bing Zhang (2016) A novel
anomaly detection method incorporating target information derived from hyperspectral
imagery, Remote Sensing Letters, 7:1, 11-20, DOI: 10.1080/2150704X.2015.1101177
Article views: 32
1. Introduction
Anomaly detection is one of the important topics in hyperspectral image processing
and hyperspectral remote sensing applications. Anomaly can be dened as a target
which appears with a low probability in a given scene (Eismann 2012). Although no
targets or background information is available in the process of detection, anomalies
still have two characteristics that make them as outliers: (1) their spectral signatures
are dierent from their neighbour pixels and (2) anomalies occur with a low prob-
ability (Stein et al. 2002). Based on these two characteristics, statistical models have
been developed and used to calculate the probability of a pixel under test (PUT) being
a target. One main assumption is that its background follows a multivariate normal
distribution. According to this assumption, ReedXiaoli detector (RXD) (Reed and Yu
1990) was developed and widely used for anomaly detection. It applies the probability
(Gorelnik, Yehudai, and Rotman 2010), the blocked adaptive computationally ecient
outlier nominator (BACON) (Billor, Hadi, and Velleman 2000), the random-selection-
based anomaly detector (RSAD) (Du and Zhang 2011), the weighted-RXD (Guo et al.
2014) and a discriminative metric learning-based anomaly detection method (Du and
Zhang 2014). In recent years, several other approaches such as a feature extraction
algorithm named sparse transfer manifold embedding and a supervised matric learning
algorithm are implemented for subpixel target detection (Zhang, Zhang, Tao, and Huang
2014; Zhang, Zhang, Tao, Huang, et al. 2014). BACON and RSAD are two novel algo-
rithms among these detectors. The two algorithms can prevent the contamination by
removing anomalous signatures when estimating background information. They are
very powerful to spot anomalies as outliers. However, it is noticeable that the two
methods can be used to estimate information from the background only. As mentioned
earlier, lacking target information could be a drawback that might cause a high FAR.
Usually, a scene of hyperspectral imagery can also contain several targets which can be
spotted by the RXD algorithm. For this case, we propose a new method that can be used
not only to estimate background information, but also to calculate target statistics from
the image. The additional anomaly information can be used to improve the ability of
detectors to spot anomalies.
H1 : x s b (2)
where x is the sample pixel vector; s is the target signal and b is the background signal
vector which is assumed as a multivariate normal distribution with mean vector and
REMOTE SENSING LETTERS 13
1 T 1
e2x S x
1
Px j H0 K=2 1=2
(3)
2 jSj
pixel. Based on this observation, RXD uses the following expression to detect anomalies:
2.2. BACON
BACON is an algorithm to spot outliers in multivariate and regression data. It uses two
main strategies to achieve the goal: (1) thresholding RXD values of pixels to rene the
background and (2) selecting a small group of background samples to keep the time
eciency. The BACON algorithm includes the following steps:
Step 1. Compute the RXD score for each PUT. Select m cK smallest values of RXD
scores as the initial background subset, where K is the number of bands; cis a small
integer chosen by the data analyst, and c should be greater than 1 such that the
condition of m > K can be met.
Step 2. Obtain the square root of RXD scores based on the current background subset.
Step 3. Those pixels whose square root values of RXD scores are smaller than cn;K;r K;
will be selected as new background samples, where K; is the square root of the
1 percentile of the x2 distribution with K degrees of freedom, and cn;K;r is
computed as follows:
h n K 1=2 (7)
where n is the total number of pixels and r is the number of pixels in current back-
ground subset.
Step 4. Iterate steps 2 and 3 until the size of the background subset no longer change.
Step 5. Map anomalies to the image space.
14 Q. GUO ET AL.
2.3. RSAD
Dierent from BACON, the RSAD algorithm randomly selects representative background
samples from the image to estimate background statistical information, identies
anomalies by statistical dierences, and nally, fuses all the detection results together.
The steps of RSAD are described as follows:
Step 4. Iterate steps 2 and 3 until the background subset no longer change.
Step 5. Label the pixels excluded by the nal background subset as anomalies.
Nm
PHm (10)
N0 N1
Therefore,
PHm j x Px j Hm PHm =Px (11)
REMOTE SENSING LETTERS 15
where Px is the probability that vector x occurs in the image, Sm is the covariance
matrix of Vm; m is the mean vector of Vm and Nm is the number of samples in Vm. In this
case, we can use the log-likelihood ratio to simplify the equation:
1 K
lnPHm j x x m T S1
m x m ln2
2 2 (12)
1
lnj Sm j lnPHm lnPx
2
Then subtracting lnPH0 j x from lnPH1 j x, we obtain the quantity d10 :
1 1
d10 x 0 T S1 T 1
0 x 0 x 1 S1 x 1
2 2 (13)
1 1
lnj S0 j lnPH1 lnj S1 j lnPH0
2 2
Downloaded by [Wuhan University] at 06:46 10 November 2015
d10 x 0 T S1 T 1
0 x 0 x 1 S1 x 1 RXD0 x RXD1 x (14)
Resulting from the previous equations, RXD0 x represents the distance from the
PUTx to the background set, and RXD1 x is the distance from x to the target set.
Although the algorithm of Equation (14) is not a constant FAR detector (Gao et al. 2014),
it still has three advantages: (1) this algorithm distinguishes the covariance matrix of the
background from the covariance matrix of targets. The additional target information can
benet for the target detection; (2) this algorithm highlights an anomaly twice by the
distance from the background set and by the distance from the target set; and (3)
background pixels are signicantly dierent from targets in spectral domain. Thus, the
RXD1 x of the background pixels will be very large to withhold them as outliers. As the
number of target samples chosen by the proposed method is determined by the
number of bands, it is necessary to control the size of target sample, otherwise more
background information may be involved in the target set. Since principal component
analysis (PCA) is a commonly used approach to reduce data dimensionality and increase
the detection eciency (Ma, Crawford, and Tian 2010; Taitano, Geier, and Bauer 2010; Du
and Zhang 2011), our method uses it to control the number of target samples while
preserving major information of the image. The procedures of our method can be
described as follows:
Step 1. Conduct PCA transformation for an image and keep the rst t principal
component (PC) images.
Step 2. Implement BACON or RSAD with the rst t component images to create
background subset.
Step 3. Select l ct pixels with the largest values of BACON scores or RSAD scores
into target subset, where c is the small integer for determining background
samples in BACON or RSAD.
Step 4. Using the background subset and target subset created by steps 2 and 3,
based on Equation (14) to produce a grey image as nal output.
16 Q. GUO ET AL.
locations of the hot spots provided by the United States Geological Survey. The
ground truth image was used to evaluate the performances of dierent algorithms.
Figure 1. (a) AVIRIS image over the World Trade Center in New York City on 16 September 2001; (b)
ground truth map indicating spatial locations of the hot spot res, available from the United States
Geological Survey.
REMOTE SENSING LETTERS 17
usually more than one pixel, the LRXD in this experiment had been implemented using a
dual window technique (Kwon, Der, and Nasrabadi 2003). We used dierent inner and
outer window sizes to process the image. 15 15as an outer window size and 11 11as
an inner window size were selected for this experiment as they led to the best detection
performance on this data set. The grey images obtained by the mentioned detectors are
presented in Figure 2. We could identify the re spots by their high brightness values.
The AUCs of dierent detectors for various PCs are shown in Figure 3. It is noticeable
that BACON and RSAD would produce the best detection results when using 15 PCs.
Thus, BACON-Target (15 PCs) and RSAD-Target (15 PCs) would be implemented in the
following experiments. The spectral information was inadequate if fewer components
were used. On the other hand, more PCs could probably contain more noise, which
could negatively impact the capability of detectors. Thus, either using too few or too
Figure 3. AUCs created by dierent algorithms by using dierent numbers of PCs on WTC data.
18 Q. GUO ET AL.
many components would lead to the reduction of accuracy. BACON-Target and RSAD-
Target outperformed BACON and RSAD on the rst 5 PCs to rst 17 PCs, which demon-
strated that target information could improve the capability of these anomaly detectors.
GRXD created a better result than LRXD, probably because some re spots in the study
area were large in size, and thus the local windows always included some anomaly pixels
when estimating information of background, which withheld the anomalies to be
detected.
All ROCs of the discussed methods are presented in Figure 4, and their correspond-
ing AUCs are shown in Table 1. From the gure, RSAD-Target (15 PCs) had the best
detection performance among all these detectors. BACON and RSAD performed with
almost no dierence when full data or 15 PCs were used. However, the computation
time of BACON (full) was 11.82 times higher than that by BACON (15 PCs). Thus, it is
wise to use several PCs instead of full bands of the image when implementing
BACON or RSAD. GRXD was better than LRXD in spite of using full bands or 15 PCs.
Downloaded by [Wuhan University] at 06:46 10 November 2015
In terms of time eciency, GRXD was the fastest algorithm as it was the simplest.
However, it created more false alarms than BACON, RSAD and the proposed methods.
BACON-Target (15 PCs) and RSAD-Target (15 PCs) outperformed BACON and RSAD in
spite of using 15 PCs or full bands of the image, which demonstrated that the target
information could help detectors spot more targets with less false alarms indeed.
Table 1. AUCs and processing times (seconds) produced by dierent detectors; also see Figures 2
and 4.
Algorithms AUC Time (s) Algorithms AUC Time (s)
GRXD (full) 0.9689 9.11 LRXD (full) 0.9642 36.10
GRXD (15 PCs) 0.9583 0.29 LRXD (15 PCs) 0.9082 13.08
BACON (full) 0.9947 97.03 RSAD (full) 0.9946 49.55
BACON (15 PCs) 0.9942 8.21 RSAD (15 PCs) 0.9944 8.92
BACON-Target (15 PCs) 0.9978 8.34 RSAD-Target (15 PCs) 0.9987 9.04
REMOTE SENSING LETTERS 19
Several recent papers mentioned using kernel methods to detect anomalies (e.g.
Kernel RX algorithm and Kernel PCA). These algorithms could mine the high-order
correlation between spectral bands and selects the non-linear features from the
hyperspectral image (Kwon and Nasrabadi 2005; Gu, Liu, and Zhang 2008).
However, kernel methods are greatly time consuming and need a lot of memory
space. Compared with these methods, our proposed methods are much easier to
implement.
5. Conclusion
Anomaly detection is a very active topic of research in hyperspectral imagery processing.
Usually, it is with a condition without any prior background or target information.
Currently existing algorithms (e.g. RXD, BACON and RSAD) only estimate the information
Downloaded by [Wuhan University] at 06:46 10 November 2015
of background to make decisions for anomaly detection. This study proposed new
methods of incorporating target information from the image to assist detection. The
new methods were combined with PCA technique in order to control the number of
target samples as well as increase the time eciency. In this study, we used real
hyperspectral data (AVIRIS) to test the performances of our new methods. The experi-
mental results demonstrate that the new methods have outperformed other traditional
algorithms which only consider background information (e.g. BACON, RSAD and RXD).
Moreover, the proposed methods were of time eciency as they could obtain satised
detection results with only a few PC images.
Acknowledgement
The authors would like to express appreciation to the NASAs Jet Propulsion Laboratory for
providing the WTC data set.
Disclosure statement
No potential conict of interest was reported by the authors.
References
Billor, N., A. S. Hadi, and P. F. Velleman. 2000. BACON: Blocked Adaptive Computationally Ecient
Outlier Nominators. Computational Statistics & Data Analysis 34 (3): 279298. doi:10.1016/
S0167-9473(99)00101-2.
Du, B., and L. Zhang. 2011. Random-Selection-Based Anomaly Detector for Hyperspectral
Imagery. IEEE Transactions on Geoscience and Remote Sensing 49 (5): 15781589. doi:10.1109/
TGRS.2010.2081677.
Du, B., and L. Zhang. 2014. A Discriminative Metric Learning Based Anomaly Detection Method.
IEEE Transactions on Geoscience and Remote Sensing 52 (11): 68446857. doi:10.1109/
TGRS.2014.2303895.
Eismann, M. 2012. Hyperspectral Remote Sensing. Bellingham, WA: SPIE Press. doi:10.1117/3.899758.
Gao, L., Q. Guo, A. Plaza, J. Li, and B. Zhang. 2014. Probabilistic Anomaly Detector for Remotely
Sensed Hyperspectral Data. Journal of Applied Remote Sensing 8 (1): 083538083538.
doi:10.1117/1.JRS.8.083538.
20 Q. GUO ET AL.
Gaucel, J.-M., M. Guillaume, and S. Bourennane. 2008. Adaptive Linear Minimum Mean Square
Error Restoration: Inuence on Hyperspectral Detection Strategy. International Journal of
Remote Sensing 29 (10): 29432961. doi:10.1080/01431160701408428.
Gorelnik, N., H. Yehudai, and S. R. Rotman. 2010. Anomaly Detection in Non-Stationary
Backgrounds. 2nd Workshop on Hyperspectral Image and Signal Processing: Evolution in
Remote Sensing, Reykjavik, Iceland: WHISPERS 2010, IEEE Computer Society, June 1416.
doi:10.1109/WHISPERS.2010.5594914.
Gu, Y., Y. Liu, and Y. Zhang. 2008. A Selective KPCA Algorithm Based on High-Order Statistics for
Anomaly Detection in Hyperspectral Imagery. IEEE Geoscience and Remote Sensing Letters 5 (1):
4347. doi:10.1109/LGRS.2007.907304.
Guo, Q., B. Zhang, Q. Ran, L. Gao, J. Li, and A. Plaza. 2014. Weighted-RXD and Linear Filter-Based
RXD: Improving Background Statistics Estimation for Anomaly Detection in Hyperspectral
Imagery. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7
(6): 23512366. doi:10.1109/JSTARS.2014.2302446.
Huck, A., and M. Guillaume. 2010. Asymptotically CFAR-Unsupervised Target Detection and
Discrimination in Hyperspectral Images With Anomalous-Component Pursuit. IEEE
Downloaded by [Wuhan University] at 06:46 10 November 2015