An Improved Algorithm to Remove Cosmic Spikes in Raman
Spectra for Online Monitoring
SHENG LI and LIANKUI DAI*
State Key Lab of Industrial Control Technology, Zhejiang University, Hangzhou, China 310027
Raman spectral analysis integrated with multivariate calibration is a fast
and effective solution to monitor chemical product properties. However,
Raman instruments utilizing charge-coupled device (CCD) detectors
suffer from occasional spikes caused by cosmic rays. Cosmic spikes can
disturb or even destroy the meaningful chemical information expressed by
normal Raman spectra. In online monitoring, some cosmic spikes have
intensity and bandwidth similar to normal Raman peaks of chemical
components when a low resolution and cost-effective Raman instrument is
used. Moreover, the online Raman spectra always contain variations of
strong Raman peaks and fluorescence. Current spike-removal methods
seem to have difficulty detecting and recovering cosmic spikes in these
online Raman spectra. Therefore, an improved algorithm is proposed. In
this algorithm, a new scheme composed of intensity identification and
local moving window correlation analysis is introduced for cosmic spike
detection; intensity identification based on derivative spectra and local
linear fitting approximation are used for the recovery of cosmic spikes.
The algorithm is proved to be simple and effective and has been applied in
an online Raman instrument installed at a continuous catalytic reforming
unit in a refinery.
Index Headings: Raman spectroscopy; Online monitoring; Cosmic spikes;
Charge-coupled device; CCD detector; Data processing.
INTRODUCTION
Raman spectral analysis is commonly applied in the
chemical industry,1–5 as vibration information represented by
Raman spectral peaks is specific to the chemical bonds and
symmetry of molecules. It has become a suitable tool for real-
time monitoring of chemical product quality because of its high
sampling rate and reliability. Charge-coupled device (CCD)
detectors are widely used in Raman spectrometers because of
their high sensitivity and low noise, but they are vulnerable to
cosmic rays, which lead to visible spikes in Raman spectra.
Typically, these spikes have narrow bandwidth, positive
unidirectional peaks, and random positions. Cosmic spikes
pollute or even distort the effective Raman signals of chemical
products.
The aim of our research team is to monitor the properties of
reformed gasoline in a continuous catalytic reforming unit
using an online Raman instrument. Unfortunately, the cosmic
spikes result in unexpected abrupt fluctuations of gasoline
properties, so they must be detected and removed. In our
application, a low-resolution and cost-effective Raman instru-
ment is used, so some cosmic spikes have similar bandwidth to
some normal Raman peaks; the spectra are obtained over a
long integration time because the Raman signal of reforming
gasoline is not strong enough in our spectral apparatus; the
existing time of cosmic spikes is less than the integration time,
Received 22 October 2010; accepted 17 August 2011.
* Author to whom correspondence should be sent. E-mail: lkdai@iipc.zju.
edu.cn.
DOI: 10.1366/10-06169
0003-7028/11/6511-1300$2.00/0
1300 Volume 65, Number 11, 2011 Ó 2011 Society for Applied Spectroscopy
which results in the intensity of the cosmic spikes being smaller
than those in other literature and more difficult to detect.
Spectral processing should be quick so that the information can
be utilized in process optimization of the gasoline properties
online.
A wealth of approaches have been proposed in the literature
for the detection and removal of cosmic spikes. Some of these
approaches can only detect whether a spectrum contains spikes
and cannot find their exact positions in the spectrum.6,7 Others
can both detect and remove spikes; these approaches fall into
four categories.
The first category can be defined as single-scan methods,
including robust smoothing filter,8 weighted moving window
filter,9 missing point polynomial filter,10,11 and wavelet
recognition.12 These approaches are based on the hypothesis
that the bandwidth of cosmic spikes is much smaller than the
bandwidth of the substance of interest. However, the resolution
of our Raman instrument is low and gasoline is a relatively
complex mixture, so the bandwidth of some Raman peaks is
similar to that of some cosmic spikes. Some normal peaks will
be mistaken for cosmic spikes if these methods are used.
Moreover, filter-based approaches may distort gasoline Raman
spectra.
The second category, including the robust summation
method13 and the upper-bound spectrum method14 and its
improved versions,15,16 is based on the fact that the probability
of cosmic spikes appearing at the same CCD pixel in two
sequential spectra is very low and at least one of the same-
position pixels in two spectra is normal. These methods cost
more spectral acquisition time, which limits their application in
online monitoring.
The third category tries to improve optical hardware
performance to avoid spikes, such as image curvature
correction,17 which will increase instrumental complexity and
cost.
A fourth category of approaches is proposed to detect and
remove the cosmic spikes in Raman spectral images.18–20
These approaches are based on the fact that the probability of a
cosmic spike appearing at the same CCD pixel in two spatially
adjacent spectra is low and that the difference between the
spectrum of a spatial pixel and its neighbors in the spatial
domain is very small. The currently detected spectrum is
approximated by the most similar neighboring spectrum. This
spectrum can also be called the nearest neighbor of the
currently detected spectrum. The spikes are detected by
comparing the currently detected spectrum and its approximat-
ed spectrum. Therefore, these methods are also called the
nearest neighbor methods.
In online reformed gasoline monitoring, the change of
Raman spectra is slow and smooth under normal operation. A
series of spectra obtained just before the current spectrum in a
time domain can be taken as the neighboring spectra, and the
APPLIED SPECTROSCOPY
neighboring spectral information can also be utilized just as the
nearest neighbor methods. However, a more effective spike
detection and removal algorithm is needed because changes in
strong Raman peaks and fluorescence for online spectra result
in visible difference between the currently detected spectrum
and its approximated spectrum. The simple intensity identifi-
cation used in previous nearest neighbor methods is not
suitable.
An improved cosmic spike removal algorithm based on a
new spike identification scheme is proposed for online
monitoring of a low-resolution Raman system. In this new
scheme, both intensity identification and local moving window
correlation analysis are used to distinguish spectral variations
of strong Raman peaks from real cosmic spikes and confirm
whether a spike is just a strong Raman peak. Intensity
identification based on derivative spectra and local linear
fitting approximation are used for the recovery of cosmic
spikes. The algorithm is conceptually simple and effective for
online cosmic spike removal. Furthermore, it is very easy to
implement and costs little calculation time.
THEORY
Strong Raman Peak Area Definition. In order to
distinguish spikes on strong Raman peaks, the strong Raman
peak (SRP) area is defined in advance. First, automated
subtraction of fluorescence based on the iterative polyfit
method21 is carried out on the current spectra. Second, we
obtain the maximum of the fluorescence subtracted spectrum
im. Third, the partition ratio rp is introduced and the area in
which the intensity of the fluorescence subtracted spectrum is
more than im 3 rp is defined as the SRP area.
Most Similar Neighboring Spectrum Definition. In our
algorithm, the spectrum just before the current spectrum at the
last sampling time can be defined as the most similar
neighboring (MSN) spectrum since in process monitoring the
sample at the last sampling time has the most similar chemical
composition with the current sample.
Linear Approximation of Current Spectrum. Following
the spatial nearest neighbor method by Zhang and Henson,14
linear approximation of the current spectrum is implemented
for neighboring comparison as follows:
x ¼ axm þ b1 þ e ð1Þ
xa ¼ axm þ b1 ð2Þ
where xm denotes the MSN spectrum; x is the current raw
spectrum; xa is the linear fit approximated spectrum; a and b
are the linear fitting coefficients; and e is the fitting error.
Cosmic Spike Detection. For a current spectrum, a scaled
absolute residual vector r is given by:
jx  xa j
r¼ ð3Þ
s
where s is the standard deviation of the linear fit.
In our new scheme, an intensity threshold t and a correlation
coefficient threshold Rt are introduced.
First, the pixels with intensity greater than t in r are
identified and the index corresponding to each set of pixels
above t is defined as a detection zone z.
Then, the detection zones are checked if they are in the SRP
areas. The detection zone around an SRP area is defined as zs,
and the remaining detection zones are defined as zw. Each
detection zone zw is confirmed to contain a spike and the
detection zone zs will be checked in the third step. (If the pixel
number of zs is less than 3, neighboring normal pixels are
introduced to insure that the correlation analysis can be
realized.)
Third, the moving window correlation analysis between x
and xa at zs is used to detect spikes on the strong Raman peaks.
The window size is denoted as w and the analysis result as Rc.
Rc is a vector composed of correlation coefficients of each w-
pixel window. zs is determined to contain a spike if the
minimum of Rc is less than Rt. Conversely, zs is judged as the
normal zone at which the intensity of a strong Raman peak
varies. The moving window correlation analysis is based on the
fact that the cosmic spikes have more narrow width than the
strong Raman peaks.
Cosmic Spike Removal. According to the detection of
cosmic spikes, the removal procedure is specific to each kind of
spike zone. For spike zone zw , a local linear fitting
approximation between x and xa in zw is implemented. The
neighboring nr normal pixels at two sides of zw are used to
estimate the fitting coefficients. The locally approximated
spectrum in zw is used to replace the contaminated pixels.
For a spike zone zs, zs should be resized since zs also
contains the variation of strong Raman peaks and the size of zs
is larger than the real spike zone. Because variation of the
gradient intensity for strong Raman peaks is much smaller than
that for cosmic spikes, derivative spectra of x and xa are first
calculated and are denoted as xd and xda. Then, the absolute
residue of xd and xda is obtained and its standard deviation is
calculated and defined as sxd. A scaled absolute residual vector
rd is given as follows:
jxd  xda j
rd ¼ ð4Þ
sxd
Third, an intensity threshold tr is introduced and the resized
spike zone zsr in which the intensity of rd is greater than tr is
finally obtained in zs.
After resizing zs into zsr, the same local linear fitting
approximation is used to recover x at zsr.
Discontinuity of the Detection Zone. The detection zones
may be discontinuous in two cases. Case 1: a spike is on a
strong Raman peak that is decreasing, in which case some
spike pixels may have lower intensity than t in r. Case 2: the
absolute residual vector rd is used and the intensity at the pixel
which is the center of the spike is always close to zero. When
discontinuity happens, pixels having lower intensity are
ignored in the detection zones.
Algorithm Procedure. The procedure of the improved
algorithm is as follows:
(1) Use the current spectrum after fluorescence subtraction and
the partition ratio rp to acquire the SRP areas.
(2) Obtain the approximated spectrum xa according to Eq. 2.
(3) Calculate the scaled residual vector r using Eq. 3. Obtain
detection zones z according to the intensity threshold t and
divide them into zs and zw. Each zw can be taken as a spike
zone. Examine each zs using the moving window
correlation analysis to confirm whether zs is a spike zone.
(4) For a spike zone zs, zs is resized to zsr by the scaled
absolute residual vector rd and threshold tr. Finally, a local
APPLIED SPECTROSCOPY 1301
 linear fitting approximation is applied to recover the
contaminated pixels in x at zw and zsr.

SIMULATION EXPERIMENT
Sample Generation. A series of synthesized online Raman
spectra S(k) consist of the following three parts:
SðkÞ ¼ RðkÞ þ BðkÞ þ NðkÞ ð5Þ

where R(k) denotes the pure Raman signal, B(k) is the curved
baselines to represent the fluorescence, N(k) is random noise,
and k is the time index.
The pure Raman signal can be described by R(k) = fRi(k)ji
= 1, . . ., ng:
X
M
2Aj ðkÞcL; j
Ri ðkÞ ¼ ð6Þ
2
j¼1 p½cL; j þ 4ði  m0; j Þ2 

where n represents the number of pixels; M is the number of FIG. 1. Two simulated online Raman spectra without spikes.
Lorentzian peaks; and m0,j, c0,j, and Aj(k) denote the central
position, the width, and the area of the jth peak, respectively. 8
In this experiment, we set M = 5, n = 2000, k = 1 and 2, >
> Si ð2Þ þ 15000 i ¼ 353
>
>
cL,1 = 30, m0,1 = 200; cL,2 = 30, m0,2 = 500; cL,3 = 15, m0,3 = < Si ð2Þ þ 2000ð5  ji  500jÞ i 2 ½498; 502
1000; cL,4 = 40, m0,4 = 1100; and cL,5 = 10, m0,5 = 1700. For Ŝi ð2Þ ¼ Si ð2Þ þ 2000ð3  ji  1003jÞ i 2 ½1002; 1004
>
>
real online Raman spectra, Raman peak intensity varies >
> S ð2Þ þ 2000ð3  ji  1126jÞ i 2 ½1125; 1127
: i
according to operating conditions, which can be realized by Si ð2Þ else
adjusting the following peak areas in our simulation: ð14Þ
5
A1 ðkÞ ¼ 10 3½1 þ ðk  1Þ 3 0:02 ð7Þ
INDUSTRIAL ONLINE APPLICATION
A2 ðkÞ ¼ 106 3½1 þ ðk  1Þ 3 0:15 ð8Þ
Instrumental Setup. An online Raman instrument is
installed at a catalytic reforming unit in a refinery. As shown
A3 ðkÞ ¼ 106 3½1  ðk  1Þ 3 0:1 ð9Þ
in Fig. 3, the reformed gasoline is sampled at two sides of a
regulating valve V1, which is located at the bottom outlet of a
A4 ðkÞ ¼ 106 ð10Þ distillation column after a heat exchanger. The instrument
includes reformed gasoline sampling equipment, the Raman
A5 ðkÞ ¼ 106 ð11Þ system, and a central control system. In the Raman system, a

The curved baselines B(k) can be approximated by a

polynomial. In the gasoline refining application, it is found
that the fluorescence varies minimally, so we represent B(k)
with the following fifth polynomial and set the change of
baseline as follows:

Bi ð1Þ ¼ 2:7 3 1013 i5 þ 2:7 3 109 i4  1:3 3 105 i3

þ 0:03i2  40i þ 1:9 3 104 ð12Þ

Bi ð2Þ ¼ Bi ð1Þ 3½1 þ ðk  1Þ 3 0:03 þ ðk  1Þ 3 7500 ð13Þ

where i = 1, . . ., 2000. The random noise N(k) can be

represented by white noise, and its standard deviation is set to
be 35. Finally, the synthesized online Raman spectra are shown
in Fig. 1.
The width of spikes in real online spectra varies from one
pixel to several pixels and the intensity of spikes also varies,
so four spikes with different width and different intensity
are added into the synthesized spectrum S(2) in this
experiment. The synthesized spectrum with spikes is denoted
by Ŝ(2) = fŜi(2)ji = 1, . . ., 2000g, which is shown in Fig. 2, and FIG. 2. Simulated online Raman spectrum Ŝ(2) with four spikes.

1302 Volume 65, Number 11, 2011


FIG. 3. The installation position of the online Raman instrument.
785 nm near-infrared diode laser is used; the Raman scattering
spectrum is collected by a process Raman probe (InPhotonics,
Inc., Norwood, MA) and measured with a scientific-grade
spectrometer QE65000 (Ocean Optics, Inc., Dunedin, FL).
Sample Source. One thousand (1000) real gasoline Raman
spectra were obtained from 2008/12/5 1:51 to 2008/12/6 23:20
at a sampling interval of 2 minutes. The temperature of
reformed gasoline fluctuated between 8.9 8C and 22.3 8C and
the environment temperature fluctuated between 10.1 8C and
20.2 8C. The partial spectra from the 261st pixel to the 750th
pixel, which contain most gasoline Raman features, are chosen
as the input for the spike removal algorithm. As shown in Fig.
4, some of the sequential spectra were overlapped, indicating
that the operating conditions of the catalytic reforming process
were almost invariant. On the other hand, some of the
sequential spectra changed due to fluctuations of the operating
conditions. Moreover, the spectra will also change when the
intensity of the diode laser varies, but the probability of this is
quite low because the industrial diode laser is running under
long-time stable conditions.
RESULTS AND DISCUSSION
Simulation Experiment. Set Ŝ(2) to be the current spectrum
and S(1) to be its MSN spectrum. Spike removal algorithm
parameters are set as given in Table I. Figure 5 illustrates the
spike detection procedure. Three SRP areas are obtained, the first
lies in the 491st;509th pixels, the second lies in the
FIG. 4. Real online gasoline spectra.
TABLE I. Algorithm parameters for the simulation experiment.
Parameter Parameter specification Value
rp Partition ratio for confirming the SRP areas 0.3
t Intensity threshold for spike detection 1.7
w Window size for the moving window correlation analysis 3
Rt Correlation coefficient threshold for spike detection in 0.8
SRP areas
tr Intensity threshold for resizing spike zones in SRP areas 6
nr Number of neighboring normal pixels at two sides of a 2
spike zone
988th;1012th pixels, and the third lies in the 1687th;1713th
pixels, which is shown in Fig. 5a. Figure 5b illustrates that the
current spectrum x has a good correspondence with its
approximated spectrum xa at the areas with no Raman peaks
or weak Raman peaks. However, due to the intensity variation of
the strong Raman peaks, x is in visible contrast with xa in the
whole SRP area. Figure 5c illustrates that there are five detection
zones with the intensity threshold t. The first and fourth detection
zone belong to zw, the second, third, and fifth detection zone
belong to zs. The specific distribution of the detection zones can
be shown in Table II. The discontinuity of the third detection
zone is ignored. Each zw is confirmed to be a spike zone. For
each zs, the result of moving window correlation analysis is
shown in Fig. 6. The result indicates that the minimum of Rc is
smaller than the correlation coefficient threshold Rt in the second
and third detection zone, which is confirmed to be a spike zone.
In the fifth detection zone, Rc is always greater than Rt and the
FIG. 5. The spike detection procedure. (a) The SRP area on the current
spectrum; (b) the current spectrum x and its approximated spectrum xa; (c) the
absolute residual vector r.
APPLIED SPECTROSCOPY 1303
TABLE II. The distribution of the detection zones.

Detection zones Zone position Zone class

1 353rd pixel zw
2 494th;504th pixels zs
3 979th;1023rd pixels zs
4 1125th;1127th pixels zw
5 1686th;1713rd pixels zs

FIG. 8. The result of the resizing procedure. (a) The absolute residual vector
rd at detection zone zs (the 494th;504th pixels); (b) the absolute residual
vector rd at detection zone zs (the 979th;1023rd pixel).

the spike removal procedure is shown in Fig. 9. A local linear

FIG. 6. The result of the moving window correlation analysis. (a) Correlation
coefficient vector Rc for detection zone zs (the 494th;504th pixels); (b)
correlation coefficient vector Rc for detection zone zs (the 979th;1023rd
pixels); (c) correlation coefficient vector Rc for detection zone zs (the
1686th;1713th pixels).
fitting approximation is applied. Cosmic spikes at zw and zrs are
all well recovered and there is no distortion in the rest of the
current spectrum.
For the four well-recovered simulated cosmic spikes by our
spike removal algorithm, one spike is at an area with no Raman
signal, another is at a weak Raman signal area, and the other
two are on strong Raman peaks. The spikes also have different
widths. This illustrates that there is no limitation on the spike
position and widths in our algorithm. Furthermore, if the
intensity of a spike decreases, the intensity threshold t can also
be decreased and the correlation coefficient threshold Rt can be
increased to ensure that the spike can be detected.

fifth detection zone contains no spike but the normal variation of

a strong Raman peak.
In the spike removal procedure, for a detection zone zs, a
resizing procedure is applied. The derivative spectra xd and
xda are obtained and their parts at zs are shown in Fig. 7. As
shown in Fig. 8, the second detection zone is resized to the
498th;502nd pixels and the third is resized to the
1002nd;1004th pixels. We can also ignore that discontinuity
happens at the pixel which is the center of a spike. The result of

FIG. 7. The derivative spectrum of the current spectrum and its approximated
spectrum. (a) The derivative spectra xd and xda at detection zone zs (the FIG. 9. The result of the spike removal procedure. (a) The current spectrum x;
494th;504th pixels); (b) the derivative spectra xd and xda at detection zone zs (b) the recovered current spectrum; (c) the residue of the current spectrum and
(the 979th;1023rd pixels). the recovered current spectrum.

1304 Volume 65, Number 11, 2011

 TABLE IV. Spike distribution in 999 online Raman spectra.

Spectrum Spike pixel Spike pixel Spectrum Spike pixel Spike pixel
index index number index index number

3 299;300 2 633 742 1

62 558 1 645 268 1
65 479 1 649 532 1
99 739 1 662 566 1
145 692 1 665 583 1
162 368 1 720 395 1
218 563 1 732 654 1
231 564;565 2 749 514 1
285 379;381 3 787 441 1
349 525;526 2 815 471;472 2
364 274 1 866 675 1
389 715 1 874 664 1
476 338;340 3 943 629;632 4
589 738;739 2 945 529 1
599 695 1

FIG. 10. The spike detection procedure by the method of Zhang and Henson. (a)
The current spectrum x and its approximated spectrum xa; (b) the residual vector r1.
difference at two sides of the fourth spike zone, which will lead
to unusual discontinuity in the recovered x.
Industrial Online Application. Spike removal algorithm
parameters are given in Table III. The first spectrum is set to be
the initial MSN spectrum of the second spectrum and it is
confirmed that there is no cosmic spike on it, so 999 online
Raman spectra will be processed. Twenty-nine (29) cosmic
spikes are found in 29 online spectra, respectively. All of the
contaminated spectra are recovered; the spike distribution in
detail can be seen in Table IV. There are no spikes on the
strong Raman peaks. Two of the recovered spectra (the 476th
and 599th spectra) are shown in Fig. 11.

In the method of Zhang and Henson, a dual threshold

scheme is proposed to confirm the spike zone. In the scaled
residual value r1, the separate detection zones are divided by a
low intensity threshold t1, and the spike is confirmed if at least
one pixel in the zone has greater intensity than t2. However, for
our simulation spectra, t2 is invalid because of the variation of
strong Raman peaks. In Fig. 10b, there are five fluctuations in
r1. The third fluctuation is negative because the strong Raman
peak in this area is decreasing and the spike on this peak cannot
be detected. For the other four positive fluctuations, it is very
difficult to set t2 because the fifth fluctuation caused by
variation of a strong Raman peak is even greater than the
fourth, which is caused by a real cosmic spike. Moreover, the
spike removal strategy of Zhang and Henson is not fit for our
simulation. Take the fourth fluctuation, for example; the
contaminated pixels in x will be replaced by the corresponding
pixels from xa, but it is shown in Fig. 10a that there is still some

TABLE III. Algorithm parameters for online application.

Parameter Parameter specification Value

rp Partition ratio for confirming the SRP areas 0.33

t Intensity threshold for spike detection 5
w Window size for the moving window correlation analysis 3
Rt Correlation coefficient threshold for spike detection in 0.6
SRP areas
tr Intensity threshold for resizing spike zones in SRP areas 4
nr Number of neighboring normal pixels at two sides of a 3
spike zone
FIG. 11. Two contaminated spectra and their recovered spectra.

APPLIED SPECTROSCOPY 1305


FIG. 12. The confirmation of detection zone for the 162nd spectrum. (a) The
current spectrum (the 162nd spectrum) and its MSN spectrum (the 161st
spectrum); (b) the absolute residual vector r.
Figure 12 shows the detection zone for the 162nd spectrum.
Figure 12a illustrates the visible baseline change of the whole
spectrum and the variation of the strong Raman peak at the
410th pixel between the 162nd spectrum and the 161st
spectrum, which indicates that the operating condition of the
catalytic reforming process fluctuated at that moment. Figure
12b illustrates that there is strong fluctuation both at the 368th
pixel and the 410th pixel in the scaled absolute residual vector
r. Finally, two detection zones are obtained, which are the
368th pixel and the 406th;413th pixel. The SRP areas are also
obtained in advance, which are the 312th;313th pixel and the
404th;413th pixel. Therefore, the first detection zone is zw
and is confirmed to be a spike zone. The second detection zone
is zs and the result of the moving window correlation analysis
for zs can be seen in Fig. 13, which shows that the second
detection zone contains no spike but a varying strong Raman
peak.
CONCLUSION
An improved algorithm is proposed in this paper for cosmic
spike removal of Raman spectra obtained by a low-resolution
and cost-effective Raman system in online monitoring. The
change of strong Raman peaks in the time domain can be
distinguished from real cosmic spikes by the algorithm and the
false detection rate is low. Spikes on both the strong Raman
peaks and at weak Raman signal areas can all be well
1306 Volume 65, Number 11, 2011
FIG. 13. The result of the moving window correlation analysis for zs (the
406th;413th pixels).
recovered. The algorithm is simple, easy to implement, and
costs little calculation time, which satisfies the requirements of
real-time online monitoring.
ACKNOWLEDGEMENT
This research work was supported by the National High Technology
Research and Development Program (‘‘863’’Program) of China Grant No.
2009AA04Z123.
1. P. Marteau, N. Zanier-Szydlowski, A. Aoufi, G. Hotier, and F. Cansell,
Vib. Spectrosc. 9, 101 (1995).
2. M. Digne, K. Marchand, and P. Bourges, Oil Gas Sci. Technol. 62, 91
(2007).
3. S. J. Choquette, S. N. Chesler, and D. L. Duewer, Anal. Chem. 68, 3525
(1996).
4. P. E. Flecher, W. T. Welch, S. Albin, and J B. Cooper, Spectrochim. Acta,
Part A 53, 199 (1997).
5. J. B. Cooper, Chemom. Intell. Lab. Syst. 46, 231 (1999).
6. W. J. Egan and S. L. Morgan, Anal. Chem. 70, 2372 (1998).
7. P. J. de Groot, G. J. Postma, W. J. Melssen, L. M. C. Buydens, V. Deckert,
and R. Zenobi, Anal. Chim. Acta 446, 71 (2001).
8. B. M. Bussian and W. Harale, Appl. Spectrosc. 38, 309 (1984).
9. Y. Katsumoto and Y. Ozaki, Appl. Spectrosc. 57, 317 (2003).
10. G. R. Phillips and J. M. Harris, Anal. Chem. 62, 2351 (1990).
11. W. Hill and D. Rogalla, Anal. Chem. 64, 2575 (1992).
12. F. Ehrentreich and L. Summchen, Anal. Chem. 73, 4364 (2001).
13. H. Takeuchi, S. Hashimoto, and I. Harada, Appl. Spectrosc. 47, 129
(1993).
14. D. M. Zhang, K. N. Jallad, and D. Ben-Amotz, Appl. Spectrosc. 55, 1523
(2001).
15. D. M. Zhang and D. Ben-Amotz, Appl. Spectrosc. 56, 91 (2002).
16. D. M. Zhang, J. D. Hanna, and D. Ben-Amotz, Appl. Spectrosc. 57, 1303
(2003).
17. J. Zhao, Appl. Spectrosc. 57, 1368 (2003).
18. C. J. Behrend, C. P. Tarnowski, and M. D. Morris, Appl. Spectrosc. 56,
1458 (2002).
19. L. Zhang and M. J. Henson, Appl. Spectrosc. 61, 1015 (2007).
20. U. B. Cappel, I. M. Bell, and L. K. Pickard, Appl. Spectrosc. 64, 195
(2010).
21. C. A. Lieber and A. Mahadevan-Jansen, Appl. Spectrosc. 57, 1363 (2003).