Comparison of Hyperspectral Change Detection Algorithms

M. Piepera , D. Manolakisb , E. Truslowa , T. Cooleyc , M. Brueggemand , A. Weisnerd , and

J. Jacobsond
a Northeastern
University, 360 Huntington Ave, Boston, MA 02115
b MIT Lincoln Laboratory, 244 Wood Street, Lexington, MA 02420
c Space Vehicles Directorate Air Force Research Laboratory, 2251 Maxwell Avenue, Kirtland

AFB, NM 87117
d National Air and Space Intelligence Center, Wright-Patterson AFB, OH 45433
∗

ABSTRACT
There are a multitude of civilian and military applications for the detection of anomalous changes in hyperspectral
images. Anomalous changes occur when the material within a pixel is replaced. Environmental factors that
change over time, such as illumination, will affect the radiance of all the pixels in a scene, despite the materials
within remaining constant. The goal of an anomalous change detection algorithm is to suppress changes caused
by the environment, and detect pixels where the materials within have changed.
Anomalous change detection is a two step process. Two co-registered images of a scene are first transformed
to maximize the overall correlation between the images, then an anomalous change detector (ACD) is applied
to the transformed images. The transforms maximize the correlation between the two images to attenuate the
environmental differences that distract from the anomalous changes of importance.
Several categories of transforms with different optimization parameters are discussed and compared. One
of two types of ACDs are then applied to the transformed images. The first ACD uses the difference of the
two transformed images. The second concatenates the spectra of two images and uses an aggregated ACD. A
comparison of the two ACD methods and their effectiveness with the different transforms is done for the first
time.

Keywords: hyperspectral, anomalous change detection

1. INTRODUCTION
Remote change detection using hyperspectral imaging data is a broad area with many civilian and military
applications. In this paper we consider a special class of change detection applications concerned with “small”
and “anomalous” changes [1, 2]; a review of techniques for general image change detection is given by [3]. In
anomalous change detection (ACD) problems two hyperspectral images of the same scene, taken at different times
and under different conditions are available. The objective is to detect small objects that have been inserted,
deleted, or moved between the two observations. The data may be registered with a one-to-one correspondence
∗
This work is sponsored by the Air Force Research Laboratory under Air Force Contract #FA8721-05-C-0002. Opinions, in-
terpretations, conclusions, and recommendations are those of the author and are not necessarily endorsed by the United States
Government.

Imaging Spectrometry XX, edited by Thomas S. Pagano, John F. Silny, Proc. of SPIE Vol. 9611,
96110Z © 2015 SPIE · CCC code: 0277-786X/15/$18 · doi: 10.1117/12.2188316

Proc. of SPIE Vol. 9611 96110Z-1

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 between pixels of the same location [4]. Environmental changes and radiance changes due to illumination are of
no interest.
The basic assumptions underlying the operation of hyperspectral change detection algorithms are: (a) the
background clutter and environmental conditions in the two images are correlated; therefore, we can predict
one from the other; (b) “small” and “anomalous” changes in one image constitute “new” information which
cannot be predicted from the other image. Therefore, a reasonable approach to change detection shown in Fig. 1
involves two steps: (a) transform the two images to bring out their overall similarities; (b) use an ACD algorithm
to find differences between the two transformed images.
The paper is organized as follows, in Section 2 several image transforms are discussed, then the ACD al-
gorithms are discussed in Section 3. A comparison of the results using the different permutations of image
transforms and ACD algorithms is given in Section 4 for an AFRL change detection test. The conclusions are
discussed in Section 5.

2. IMAGE TRANSFORMS
Due to environmental changes, the pixel radiances of the two images taken of a scene at different times will differ.
However, the materials within most pixels remain consistent so that there is a correlation between the images.
Linear transforms are applied to each of the images to attenuate the subtle changes caused by time, and bring
out the correlation between the pixel materials of the two images. For unchanged pixels, the differences between
the transformed spectra from the two images should be small. When the material in a pixel is replaced by an
anomalous man-made material, the correlation between the pixel pair is lost, causing a large spectral change
allowing for detection.
The transformed spectra are
v = B T y and u = AT x (1)

where x is a p × 1 pixel spectrum in the first image and y is a corresponding q × 1 pixel spectrum in the second
image. The mean vectors mx = E(x), my = E(y) are assumed to both be 0 and the covariance matrices of the
images " #! " #
x Cx Cxy
C = cov = (2)
y Cyx Cy

are assumed known. The transforms A and B, are p × r and q × r matrices, where r ≤ min(p, q). Table 1 shows
the A and B transforms derived in [5], [6], and [7] from optimizing different criteria.

2.1 Linear Prediction

The Linear Prediction-X (LP-X) provides a minimum mean square estimate of X from the Y image space. The
Linear Prediction-X transform is defined by
x̂ = yCy−1 Cyx (3)

The X image is left unchanged [5]. The Linear Prediction-Y (LP-Y) is obtained by interchanging the roles of
x and y in the relevant formulas. Unlike the other transforms, the linear prediction transforms work in the
un-whitened full spectral spaces of the images. Accurate image registration is important due to the use of the
cross covariance in the transforms.

Proc. of SPIE Vol. 9611 96110Z-2

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 Image B Detected
Linear Residual
Changes
Transform e=v-u Anomaly
B Detector

Image A
Aggregated Detected
Linear u Anomaly Changes
Transform
A v Detector

Figure 1. Generalized framework for change detection.

Method A B Required quantities Optimization Criterion Spectral Space
Linear Prediction-X I Cy−1 Cyx Cx , Cy , Cxy MSE Original Full X
Linear Prediction-Y Cx−1 Cxy I Cx , Cy , Cxy MSE Original Full Y
−1/2 −1/2 T
CCA Cx P Cy Q C̃xy = P DQ Correlation Reduced Whitened
−1/2 −1/2 T
Modified CCA Cx P D Cy Q C̃xy = P DQ MSE Reduced Whitened
SVD P0 Q0 Cxy = P0 D0 QT0 Shared Variance Reduced Un-Whitened
PCA P1 Q2 Cx = P1 D1 QT1 Individual Variance Reduced Un-Whitened
Cy = P2 D2 QT2
−1/2 −1/2
CE P1 D 1 Q 2 D2 Cx = P1 D1 QT1 None Reduced Whitened
Cy = P2 D2 QT2
δ −1/2 −1/2
where C̃xy = Cx Cxy Cy and svd(C) = P DQT

Table 1. Summary of image transforms.

The LP-X and LP-Y are not equivalent transforms, and will differ in the quality of their anomalous change
detection results. Since all anomalies are only a small portion of the two images, and have a small role in the
overall image statistics, it is difficult to predict anomalies between the two image spaces. Due to difficulty of
predicting anomalous changes between the two image spaces, keeping the image with a new anomalous object
constant and predicting the other image to the image space containing the object will provide the highest quality
anomalous change detection results. Prediction in the other direction will cause some lose of the anomalous
change trying to be found, and hinder the anomalous change detection results. However, in a practical situation
one doesn’t know which image contains the anomalous object, and therefore doesn’t know which image transform
to use.

2.2 CCA and Modified-CCA

Canonical Correlation Analysis (CCA) is a method to find independent relationships of how Images X and Y
are most correlated. The CCA transforms A and B transform both X and Y into canonical variables defined
as

u = AT x, ui = aTi x, i = 1, . . . , r (4)
T
v = B y, vi = bTi y, i = 1, . . . , r (5)

Proc. of SPIE Vol. 9611 96110Z-3

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 The correlation coefficient between the canonical variables ui and vi are knows as the canonical correlations
and are defined as
δ cov(ui , vj ) aTi Cxy bj
ρi = ρui vj = p p =p q (6)
var(ui ) var(vj ) aTi Cx ai bTj Cy bj

CCA seeks to maximize the canonical correlations among all linear combinations that are uncorrelated with the
preceding variables {uk , vk , k = 1, . . . , i − 1}. The solution is determined by the left and right eigenvectors in the
SVD of the coherence matrix

δ
svd(C̃xy ) = svd(Cx−1/2 Cxy Cy−1/2 ) = P DQT (7)

where P T P = I, QT Q = I, and are the directions that maximize the correlation between the two image
spaces. The nonzero singular values of C̃xy in D = diag(ρ1 , . . . , ρr , 0, . . . , 0) are the canonical correlations and
are arranged as ρ1 ≥ ρ2 . . . ρr ≥ 0. The optimum transform matrices to the X and Y canonical variable spaces
are given by [6]
A = Cx−1/2 P , B = Cy−1/2 Q (8)

As shown by the A and B transforms, the CCA transform is implemented in a whitened space, and each
of the canonical variables will have unit variance. Due to the orthogonality of the canonical variables, the
dimensionality of the data can be reduced by utilizing only the first few canonical variables that show significant
correlation. To detect anomalous changes, one wants to find the anomalous pixels that show a large difference
for the highly correlated canonical variables.
The modified CCA transform provides a minimum MSE estimate of one canonical variable space to the other.
The cosine of the angle between the two canonical spaces with unit variance is the canonical correlation ρi . If the
unit variance constraint is only kept in one of the two canonical spaces, for example, the Y . The minimum MSE
can be reduced by projecting the X canonical space into the Y canonical space. This projection is accomplished
by simply multiplying the X canonical space by the canonical correlations ρi . The new modified CCA transfom
matrices A and B are then [8]
A = Cx−1/2 P D, B = Cy−1/2 Q (9)

2.3 Linear Prediction with Modified CCA

The CCA transform can be used as a reduced component Linear Predictor from one images space to another.
The CCA transform is derived as
svd(Cx−1/2 Cxy Cy−1/2 ) = P DQT (10)

If all the CCA components are used, the transform from the X image space to the Y image space is

ŷ T = xT Cx−1/2 (Cx−1/2 Cxy Cy−1/2 )Cy1/2 (11)

which simplifies to the Linear Prediction-Y

ŷ T = xT Cx−1 Cxy (12)

Proc. of SPIE Vol. 9611 96110Z-4

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 Using a subset of CCA components the Linear Prediction-Y can be reduced by

ŷ T = xT Cx−1/2 (P DQT )Cy1/2 (13)

similarly the Linear Prediction-X can be reduced by

x̂T = y T Cy−1/2 (QDP T )Cx1/2 (14)

2.4 Singular Value Decomposition

CCA links the two images by maximizing their correlations, however, the most correlated components might not
necessarily explain most of the variance of the images. Similar to CCA, if one wants to maximize the variance
in one image explained by the other image, the transform can be determined by the left and right eigenvectors
in the SVD of Cxy . Therefore, given the SVD

svd(Cxy ) = P0 D0 QT0 (15)

the following transformation matrices are obtained

A = P0 , B = Q0 (16)

where u = P0T x, v = QT0 y, and di are the singular values of Cxy . This approach is also known as canonical
covariance analysis [7].
The dimensionality of the data is reduced by using only the first few SVD components. The quality of the
transform components are determined by the magnitude of their variance, and the correlation of the component
pairs. Components with high variance will explain the majority of the image, and a high correlation allows for
detection of subtle differences.

2.5 Principal Component Analysis and Covariance Equalization

All of the previous change detection algorithms require the matrix Cxy in their implementation, whose compu-
tation is possible if the two data sets are registered with sufficient accuracy. If the images are not well registered,
the estimate of Cxy is practically unusable for change detection applications due to differences in edge loca-
tions being seen as changes. When Cxy is unavailable, we can use the principal component analysis (PCA) and
covariance equalization (CE) transforms.
Similarly to SVD, PCA finds orthogonal components which maximize the variance, but for each image
individually. The eigenvectors that maximize the variance for each individual image are determined by the
eigenvectors in the SVD of Cx and Cy . Therefor given the SVD

svd(Cx ) = P1 D1 QT1 , and svd(Cy ) = P2 D2 QT2 (17)

the following transformation matrices are obtained

A = P1 , B = Q2 (18)

Proc. of SPIE Vol. 9611 96110Z-5

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 where u = P1T x, v = QT2 y.
Covariance Equalization was introduced in [5] and whitens each individual image. Whitening does not have
an optimization parameter. It simply transforms the two images so that they have identical first and second order
statistics. One method of whitening an image is to normalize the variance along each principal component of the
image. This is done by normalizing each of the principal components by the square roots of their corresponding
singular values, and obtaining the transformation matrices.

−1/2 −1/2
A = P1 D 1 , B = Q2 D 2 (19)

The dimensionality of the data is reduced by using only the first few SVD components for each of the images.
The maximum variance components that explain the relatively consistent image materials should be highly
correlated, and similar to the SVD transform components. However, due to individual principal component
transforms not taking into account each others direction, it is possible to have component pairs with a high
negative correlation.

3. ANOMALOUS CHANGE DETECTION

Anomalous differences are detected by comparing the pixels of the two transformed images in one of several
ways. In this section the methods used to find four ACD scores d1 , du1 , d2 , and d3 are discussed. In the first
two methods, the difference between the two transformed images is taken where the residual image is

e=v−u (20)

An anomaly detector (AD) can be run on the residual image

d1 = eT Ce−1 e (21)

or an un-whitened anomaly detector that is essentially a sum or squares of the residual can be run

du1 = eT e (22)

The quality of the whitened and un-whitened anomalous change detectors is determined by the distribution of
the residuals for the individual components of each image transform. The variance of the residual of a component
is equal to
σei = σv2i + σu2 i − 2σvi σui ρi (23)

If a component has a high residual variance, and contains a majority of the anomalous change information,
whitening will lower the weight of the component and worsen change detection performance.
The second method uses the aggregated image AD which finds anomalies that are within one or both of the
images. The aggregated image AD is defined in [9]
" #T " #
u −1 u
d2 = G (24)
v v

Proc. of SPIE Vol. 9611 96110Z-6

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 where the aggregated covariance is
" #! " #
u Gu Guv
G = cov = (25)
v Gvu Gv

and its inverse is " #

−1
−G−1
u|v G−1 −1
u Guv Gv|u
G = (26)
G−1 −1
v|u Gvu Gu G−1
v|u

where
Gu|v = Gu − Guv G−1 −1
v Gvu and Gv|u = Gv − Gvu Gu Guv (27)

Due to the increased dimensionality of the aggregated image, the average response of the aggregated AD is
increased compared to ADs run on individual images.
The aggregated image AD can be simplified to

d2 = (u − û)T G−1 T −1
u|v u + (v − v̂) Gv|u v (28)

ûT = v T G−1 T T −1
v Gvu is a linear prediction of v to u and v̂ = u Gu Guv is a linear prediction of u to v.

The last method considered is the Hyperbolic Anomalous Change Detector (HACD), which takes the aggre-
gated image AD and subtracts each individual image AD [9]

d3 = d2 − uT G−1 T −1
u u − v Gv v (29)

The aggregated image AD finds anomalies that are within one or both of the images. By subtracting the two
individual image AD, anomalies in both images are suppressed and end up having a negative response. The
average response for the overall ACD image is zero.

4. RESULTS
4.1 Test Scene
The image transforms and anomalous change detection algorithms discussed in the previous sections were tested
on two hyperspectral images collected by AFRL for a change detection test. The images shown in Fig. 2a and
2b were collected on August 25 and October 14, 2005 between which noticeable environmental changes occurred.
Both images contain 4 anomalous panels in front of the trees. The October image has two black tarps inserted
in the grass below the panels. The two tarps are anomalous changes of interest that must be found with as few
false alarms as possible. The images were taken with a stationary hyperspectral imager allowing for optimal
image registration [10].

4.2 Reduced Dimensionality Image Transform Results

Following the framework in Fig. 1 the August and October images were transformed using one of the last six
image reduction transforms in Table 1. The first 10 components of each of the transformed image pairs were
then processed with each of the ACD algorithms.
The correlations between the X and Y components for the CCA, SVD, and PCA transforms are shown in
Fig. 3. CCA showed the maximum correlation for every component because the correlation was the optimization

Proc. of SPIE Vol. 9611 96110Z-7

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 a) August 25, 2005 b) October 14, 2005

Figure 2. AFRL change detection test scene

4) CCA, SVD, PCA

O-X A-Y
0.5
ó
U
o
iv 0

2 4 6 8 10 2 4 6 8 10
Component Number Component Number
Figure 3. Correlation between transform component pairs. Figure 4. Variance of X and Y transform components.

parameter. The SVD showed high correlation for the first component, however the correlation quickly tapered
off while still remaining positive for all the components. The first individual PCA components were nearly as
correlated as the first SVD components, but then switched between a positive and negative for several compo-
nents. The switches occurred because the individual PCA transforms do not take into account the directions of
each others components like the SVD. If the component directions were switched to be positively correlated, the
result would be the dashed green line in Fig. 3. For a majority of the components the PCA pairs have a lower
correlation than the SVD.
The variances of the X and Y components for the CCA, SVD, and PCA transforms are shown in Fig. 4.
It can be observed from the plot that the variances of the components are quite similar, and will explain the
variability of the images roughly the same. Since CCA is calculated in the whitened image space, and has a
variance of one, the CCA transform vectors were un-whitened then renormalized to unit length. The un-whitened
CCA transform vectors were no longer perpendicular, so gram-shmidt was run on them to create an orthogonal
bases. The variance of the images along the orthogonal basis vectors were then calculated.
Using ground truth for the black tarp pixel locations, receiver operating characteristic (ROC) curves were

Proc. of SPIE Vol. 9611 96110Z-8

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 1 1
o PCA
V) 0.8
SVD
a) a)
L2 0.6 CE L2 0.6
CCA
0.4 m -CCA 0.4

0.2 0.2
o o
^L
LL
^L
LL
o o
10 -6 10.4 10 -2 10° 10 -6 10.4 10 -2 10°
Probability of False Alarm Probability of False Alarm

Figure 5. ROC curves for the residual error AD. Figure 6. ROC curves for the un-whitened residual error
AD.
1 1
o o
PCA PCA
V) 0.8 V) 0.8
SVD SVD
a) a)
L2 0.6 CE L2 0.6 CE
CCA CCA
0.4 m -CCA 0.4 m -CCA

0.2 0.2
o o
^L
LL
^L
LL
o o
10 -6 10.4 10 -2 10° 10 -6 10.4 10 -2 10°
Probability of False Alarm Probability of False Alarm

Figure 7. ROC curves for the Aggregated AD. Figure 8. ROC curves for the HACD.

created for each of the anomalous change detection and transform pair results. The ROC curves for the residual
error whitened and un-whitened, aggregated image, and hyperbolic ACDs are shown in Figs. 5, 6, 7, and 8.
The mean ACD scores of the black tarps and the third panel from the left for each combination of image
transform and ACD algorithm are shown in Table 2. To compare the ACD scores of the black tarps and panel
to the background scores the mean and standard deviation of the whole ACD images are also shown in Table 2.
In order for an ACD to be effective, the scores of the black tarps should be higher than the anomalous panels
present in both of the images. For the whitened residual error ACD, the CCA and m-CCA maximized the
correlation between the two transformed images and were the only transforms that showed effective suppression
of the anomalous panel compared to the black change tarps. The ACD mean score of the black tarp was twice
that of the panel providing the best ROC curves in Fig. 5. The PCA, SVD, and CE transform components
provided similar ROC curves. For the un-whitened residual error ACD, the PCA and SVD showed improved
ROC curves, while the ROC curves for the rest of the transforms worsened. The PCA and SVD had large
residuals for their first components due to the large differences in their variances. The first components also
clearly explained the black tarps, which provided high quality ROC curves. When the residuals were whitened,
the residual for the first components were equally weighted worsening the ROC curves.
The aggregated ACD was ineffective at suppressing shared anomalies in the images. It is an anomaly detector
that detects anomalies in either image, and lacks the means of suppressing anomalies shared by both images.
The aggregated ACD image for CCA is shown in Fig. 9. It is easily visible that the two right most panels
and the brightness in the upper right hand corner are false alarms, and that the black tarp on the right is more

Proc. of SPIE Vol. 9611 96110Z-9

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 ACD PCA SVD CE CCA m-CCA
Black Residual Whitened 45.78 45.90 54.61 67.67 64.58
Tarp Residual Un-Whitened 5.82e3 5.80e3 35.3 23.6 18.9
Mean Aggregated 83.42 88.27 83.42 74.55 74.55
Hyperbolic 53.98 56.68 53.98 54.02 54.02
Panel Residual Whitened 158.77 153.60 159.43 16.64 15.49
Mean Residual Un-Whitened 5.20e3 5.45e3 237.9 14.3 9.9
Aggregated 264.4 286.5 264.5 95.69 95.69
Hyperbolic -18.73 -28.63 -18.73 -28.21 -28.21
Image Residual Whitened 10 .00 10.00 10.00 10.00 10.00
Mean Residual Un-Whitened 352.6 277.2 19.6 11.1 7.9
Aggregated 20 .00 20.00 20.00 20.00 20.00
Hyperbolic 0 0 0 0 0
Image Residual Whitened 12.50 14.60 12.57 8.62 8.44
STD Residual Un-Whitened 495.9 475.1 22.7 8.4 6.0
Aggregated 26.72 28.25 26.71 20.53 20.53
Hyperbolic 9.18 9.94 9.18 10.27 10.27

Table 2. Anomalous Change Detection Image Scores

Figure 9. Aggregated ACD on CCA Image Figure 10. Hyperbolic ACD on CCA Image

visible. The CCA and m-CCA showed more comparable scores for the black tarps and panel, which explained the
slightly better ROC curves in Fig. 7. The transform pairs PCA/CE and CCA/m-CCA have identical orthogonal
components that are either unweighted/weighted. After the whitening of the aggregated ACD, the transform
pairs give identical statistics and ROC curves. The aggregated ACD gives the highest ACD scores because the
aggregated image has double the spectral dimensionality of the residual image, which also doubles the ACD
image mean score shown in Table 2.
The HACD is an aggregated ACD with each image AD subtracted. The hyperbolic ACD image for CCA is
shown in Fig. 10. It can be observed that the false alarms in Fig. 9 have been subtracted out. The individual
anomaly detector images that were subtracted are shown in Figs. 11 and 12. The image ADs have half the
spectral dimensionality of the aggregated image and half the ACD image mean score. When the two image ADs
are subtracted from the aggregated ACD, the mean becomes centered at 0. The shared anomalies in both of the
images are subtracted twice, and are suppressed with negative values, as shown for the panel in Table 2. The
black tarp mean scores are comparable to the means for the residual error, and the suppression of the shared

Proc. of SPIE Vol. 9611 96110Z-10

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 Figure 11. X CCA Anomaly Detection Image Figure 12. Y CCA Anomaly Detection Image

1 1
o
V) 0.8
a) a)
L2 0.6 L2 0.6
LP-X-10
0.4 0.4 LP-Y-10
LP-X
o
0.2
o
0.2 r
^L ^L
LP-Y
LL LL
o o
10 -6 10.4 10 -2 10° 10 -6 10.4 10 -2 10°
Probability of False Alarm Probability of False Alarm

Figure 13. ROC curves for the residual error AD. Figure 14. ROC curves for the un-whitened residual error
AD.

anomalies provides the best ROC curves of all the ACD algorithms. The HACD ROC curves for the different
image transforms were nearly identical.

4.3 Linear Prediction and Linear Prediction with Modified CCA Results
The ACD algorithms were run on the full linear predictions of X and Y and the reduced linear predictions of
X and Y created with the first 10 CCA components. The ROC curves for the residual error whitened and
un-whitened, aggregated image, and hyperbolic ACDs are shown in Figs. 13, 14, 15, and 16.
The residuals were calculated between each linear prediction and its corresponding original images. The
whitened and un-whitened residual anomaly detectors where then run on each of the residuals. The corresponding
ROC curves are shown in Figs. 13 and 14. All of the linear predictors for the whitened anomaly detectors showed
a lack of suppression of anomalies in both images and had sub-par ROC curves. The anomalies in the light spots
of the trees and the anomalous panels provided most of the false alarms. Due to a lack of statistics on the image
anomalies, there will always be large residuals when predicting anomalies between image spaces.
For the un-whitened residual anomaly detector, the residuals of the black tarps were extremely high and
provided the high quality ROC curves shown in Fig. 14. It was observed that the residual errors of the Y linear
predictions were higher than the X linear predictions and provided better ROC curves. The reason why the Y

Proc. of SPIE Vol. 9611 96110Z-11

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 1 1
o LP-X-10
o LP-X-10
V) 0.8 V) 0.8
LP-Y-10 LP-Y-10
a)
L2 0.6 LP-X L2 0.6 LP-X
LP-Y LP-Y
0.4 0.4

0.2 0.2
o o
^L
LL
^L
LL
o o
10 -6 10.4 10 -2 10° 10 -6 10.4 10 -2 10°
Probability of False Alarm Probability of False Alarm

Figure 15. ROC curves for the Aggregated AD. Figure 16. ROC curves for the HACD.

linear predictions ROC curves were superior was the residual contained the Y image with the black tarps, and
none of the black tarp information was lost by prediction. The un-whitened AD showed better results because
the first principal components of the linear prediction residuals were high for the black tarps. When the residuals
were whitened, the residuals for the first components were equally weighted worsening the ROC curves.
The aggregated AD showed a lack of suppression of the anomalous panels compared to the black tarps and
provided the sub-par ROC curves in Fig. 15. As with the other transform methods, the HACD suppressed the
anomalous panels and provided comparable ROC curves in Fig. 16.

5. CONCLUSION
Of all the anomalous change detection algorithms, the hyperbolic ACD was the most effective overall at finding
the black tarps and suppressing illumination changes and the anomalous panels. The HACD provided nearly
identical results for all of the reduced component image transforms. The low PF A ROC curves were possible
because the the aggregated ACD was able to detect the black tarps, and the subtraction of the individual ADs
suppressed the shared image anomalies. The results were similar due to each of transforms explaining a similar
amount of variability in each of the images. It was shown that without subtracting the individual ADs, the
aggregated ACD was an ineffective ACD algorithm despite which image transform was used.
The differences between the whitened and un-whitened transforms were most evident for the residual ACDs.
The CCA transform residuals showed the greatest suppression of the anomalous panels and environmental
changes, due to the maximized correlation of the components. The black tarp changes were visible in several of
the first components, which had smaller variance residuals. Whitening of the residuals gave the first components
equal weights as the higher variance components, thus improving the ACD results and providing the best whitened
residual ACD ROC curves. The un-whitened transforms such as the linear predictors and variance transforms,
had the black tarps only visible in the first component, which had the highest variance. Whitening of the
components and giving equal weight to the first component with all the change information hindered the ACD
results. Using the un-whitened residual AD, the response consisted almost entirely of the the fist component with
the highest variance, and the black tarps were clearly visible providing low PF A ROC curves. The anomalous
panels were un-suppressed and also had a high response, but were coincidentally still lower than the black tarps.
However, if the prediction errors for the anomalous panels were higher, the ROC curves could have easily been
far worse.

Proc. of SPIE Vol. 9611 96110Z-12

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 References
[1] Michael T. Eismann, Joseph Meola, Alan D. Stocker, Scott G. Beaven, and Alan P. Schaum, “Airborne
hyperspectral detection of small changes,” Appl. Opt., vol. 47, no. 28, pp. F27–F45, 2008.

[2] James Theiler, “Quantitative comparison of quadratic covariance-based anomalous change detectors,” Appl.
Opt., vol. 47, no. 28, pp. F12–F26, 2008.

[3] Richard J Radke, Srinivas Andra, Omar Al-Kofahi, and Badrinath Roysam, “Image change detection
algorithms: a systematic survey,” Image Processing, IEEE Transactions on, vol. 14, no. 3, pp. 294–307,
2005.

[4] Barbara Zitova and Jan Flusser, “Image registration methods: a survey,” Image and vision computing, vol.
21, no. 11, pp. 977–1000, 2003.

[5] A Schaum and A Stocker, “Linear chromodynamics models for hyperspectral target detection,” in Aerospace
Conference, Proceedings. IEEE, 2003, vol. 4, pp. 4 1879–4 1885.

[6] A. Rencher, Multivariate Statistical Inference and Applications, Wiley Series in Probability and Statistics.
John Wiley & Sons, Inc., New York, 1998.

[7] Steve Cherry, “Singular value decomposition analysis and canonical correlation analysis,” Journal of
Climate, vol. 9, no. 9, pp. 2003–2009, 1996.

[8] C Radhakrishna Rao, “Separation theorems for singular values of matrices and their applications in multi-
variate analysis,” Journal of Multivariate Analysis, vol. 9, no. 3, pp. 362–377, 1979.

[9] J. Theiler, C. Scovel, B. Wohlberg, and B.R. Foy, “Elliptically contoured distributions for anomalous
change detection in hyperspectral imagery,” Geoscience and Remote Sensing Letters, IEEE, vol. 7, no. 2,
pp. 271–275, April 2010.

[10] Joseph Meola, Michael T. Eismann, Kenneth J. Barnard, and Russell C. Hardie, “Analysis of hyperspectral
change detection as affected by vegetation and illumination variations,” Proc. SPIE, 2007, vol. 6565.

Proc. of SPIE Vol. 9611 96110Z-13

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx