Automated Quantitative Spectroscopic Analysis Combining
Background Subtraction, Cosmic Ray Removal, and Peak Fitting
Timothy M. James,a Magnus Schlosser,b Richard J. Lewis,a Sebastian Fischer,b Beate Bornschein,b
Helmut H. Tellea,*
a
Department of Physics, College of Science, Swansea University, Singleton Park, Swansea, SA2 8PP United Kingdom
b
Institute for Technical Physics (ITeP), Tritium Laboratory Karlsruhe (TLK), Karlsruhe Institute of Technology (KIT), P.O. Box 3640, 76021
Karlsruhe, Germany
An integrated concept for post-acquisition spectrum analysis was
developed for in-line (real-time) and off-line applications that preserves
absolute spectral quantification; after the initializing parameter setup,
only minimal user intervention is required. This spectral evaluation suite
is composed of a sequence of tasks specifically addressing cosmic ray
removal, background subtraction, and peak analysis and fitting, together
with the treatment of two-dimensional charge-coupled device array data.
One may use any of the individual steps on their own, or may exclude
steps from the chain if so desired. For the background treatment, the
canonical rolling-circle filter (RCF) algorithm was adopted, but it was
coupled with a SavitzkyGolay filtering step on the locus-array generated
from a single RCF pass. This novel only-two-parameter procedure vastly
improves on the RCFs deficiency to overestimate the baseline level in
spectra with broad peak features. The peak analysis routine developed
here is an only-two-parameter (amplitude and position) fitting algorithm
that relies on numerical line shape profiles rather than on analytical
functions. The overall analysis chain was programmed in National
Instruments LabVIEW; this software allows for easy incorporation of
this spectrum analysis suite into any LabVIEW-managed instrument
control, data-acquisition environment, or both. The strength of the
individual tasks and the integrated program sequence are demonstrated
for the analysis of a wide range of (although not necessarily limited to)
Raman spectra of varying complexity and exhibiting nonanalytical line
profiles. In comparison to other analysis algorithms and functions, our
new approach for background subtraction, peak analysis, and fitting
returned vastly improved quantitative results, even for hidden details in
the spectra, in particular, for nonanalytical line profiles. All software is
available for download.
Index Headings: Automated baseline subtraction; Cosmic ray removal;
Peak analysis; Quantitative spectroscopic analysis; Raman spectroscopy;
Rolling-circle filter.
INTRODUCTION
Analytical (laser) spectroscopy techniques have become
commonplace in a wide range of applications in, e.g., analytical
chemistry, chemical processing, biomedical research, and
nanotechnology. Absolute and relative quantifiable results as
well as chemometric classification, based on the recorded
spectra, are of vital importance in all of these fields. Generally,
spectral responses are treated numerically during post-acquisi-
tion analysis, namely, to (i) remove background, (ii) eliminate
spurious cosmic ray events, (iii) fit spectral line profiles, or a
combination. If not applied with care, each of these measures
may affect the quantitative information contained in the spectra
Received 20 June 2012; accepted 18 March 2013
* Author to whom correspondence should be sent. E-mail: h.h.telle@
swansea.ac.uk.
DOI: 10.1366/12-06766
0003-7028/13/6708-0949/0
Volume 67, Number 8, 2013 2013 Society for Applied Spectroscopy
and also may influence detection limits. Note that spectra of
complex specimens are often evaluated using, e.g., principal
component analysis (PCA)/chemometrics instead of peak
fitting; the aim then is classification rather than precise
quantification.
Most modern data analysis software packages, either
commercial or user-written, incorporate some or all the
aforementioned post-acquisition functionality. But, normally
the full analysis chain, i.e., acquisition ! spectra correction !
data analysis, is only semiautomatic (i.e., some or all parts
require human input, and one often encounters strong model
dependence in the chosen correction routines). Furthermore,
they are often black-box implementations, and the sophis-
tication of individual procedures can vary significantly,
depending on which user community is targeted. This variation
becomes problematic in situations (i) where near real-time
(feedback) response is required to maintain stable operational
conditions, and (ii) when experiments run over extended
periods of time (weeks, months or even years), and thus
perpetual user attention is often out of the question.
Specific in-line experiments that may be seen as exemplary
for being affected by all the above-mentioned effects and
caveats are encountered in the International Thermonuclear
Experimental Reactor (ITER) project and in the Karlsruhe
Tritium Neutrino (KATRIN) experiment. For example, the
KATRIN experiment1 is set up to measure the neutrino mass
by means of analyzing the electron-energy spectra associated
with the b-decay of tritium, bound to run nearly continuously
over a period of three to five years. Thus automation of any
monitoring and control mechanisms during the gathering of
spectral information is paramount. Besides the recording of the
b-decay spectrum itself, knowledge of the temporal stability
and chemical composition of the tritium gas feed from the inner
tritium loop into the source tube of KATRIN is of ultimate
importance.2 Thus the amount and the purity of the tritium gas
injection have to be monitored continuously and can be
realized by using a laser Raman system.3 The Raman systems
requirements for data treatment and analysis procedures are
rather demanding:
(i) The operation and analysis procedure needs to be
automated, unstaffed, able to run for 60 days nonstop, and
provide real-time feedback of the gas composition (specifically
T2, DT, and HT) to the KATRIN run-control.
(ii) Acquisition and analysis time need to be minimized to
enable fast feedback of the monitored gas composition.
(iii) The extracted Raman line intensities have to be free
from systematic shifts to provide reliable and quantifiable
results, and the precision of the analysis output has to be of the
order 0.1% or better.
APPLIED SPECTROSCOPY 949
 These requirements lead to some specific problems encoun-
tered in the evaluation of recorded (Raman) spectra. For
example, the baseline in the spectra may change nonlinearly
with time, and transient changes in the gas composition can
occur. Both changes affect the sensitivity of KATRIN or, for
that matter, any other experiments experiencing such baseline
variations. The former change is mainly encountered in long-
term operations, caused by, e.g., temperature effects, the
generation of color centers in the Raman cell window, (trace)
generation of chemical reaction products,4 or a combination;
the latter change is due to system-specific injection of fresh
gases, and retraction of waste.
Any post-acquisition analysis still has to work fast and
reliably under the aforementioned circumstances, and it should
incorporate subtasks to deal with background subtraction,
cosmic ray removal, and quantitative spectral line evaluation.
This analysis should happen with as little user intervention as
possible and should be fast enough to provide the desired near
real-time response. The latter response should not be too
problematic for KATRIN Raman monitoring, where response
times of approximately 60100 s are specified. However, in the
context of the ITER project, the requirements are much more
challenging, since real-time process control with response
times of less than or equal to 1 s are required.
Furthermore, it should be possible to seamlessly link the
evaluation procedure to any data acquisition process, ideally to
run in parallel to an acquisition to provide near real-time
analysis; but which, conversely, is equally applicable to other
off-line evaluation of spectra. Also, it would be advanta-
geous if such an integrated routine were suitable for other
experimental situations in which spectra are generated in need
of similar data treatment.
For example, recently depolarization measurements for the
Q-branch Raman lines of the hydrogen isotopologues and other
molecules have been carried out by our group.5 These
measurements may require very long acquisition times of up
to the order of 1000 s, rather than the few seconds in the
aforementioned dynamic KATRIN and ITER response de-
mands, to gather the related spectra of huge intensity
differences with the necessary high signal-to-noise ratio. Such
long signal integration times lead to a large number of cosmic
ray events captured by the charge-coupled device (CCD)
detector during an acquisition, and the cosmic ray removal task
would need to be able to deal with this outcome. To provide
accurate depolarization ratios for individual rotational Raman
lineseven those with weak intensity or overlapping with each
otherthe precision in the peak evaluation would need to be
even higher than for the monitoring and control tasks in
KATRIN, ITER, or other experiments.
The integrated spectrum analysis procedure described here
evolves from cosmic ray removal via astigmatism correction
and background subtraction to peak intensity extraction. The
interplay of individual spectra treatment steps and the success
of the overall concept are presented for some selected
examples. It should be noted that parts of the analysis
procedure described in this publication have been successfully
applied to KATRIN-related measurements2,6 and other Raman
experiments.7,8 Once operating parameters are set, it constitutes
a fully automated procedure for post-acquisition data treatment
of (Raman) spectra and analysis of spectral line peaks. The
necessity for model input has been reduced to the lowest level
possible while maintaining full control on the methods.
950 Volume 67, Number 8, 2013
It is worth noting that by no means are we the first to attempt
to assemble a suite for the evaluation of Raman spectra. Since
Raman spectroscopy has become a widely accepted analytical
method, nearly all manufacturers of related instrumentation
provide software packages that incorporate data acquisition,
pre-evaluation data treatment, and final spectrum evaluation. In
addition, quite a few research groups have attempted to tailor
certain aspects to specific needs in their work (e.g., Reisner et
al.9 and Vicarra Rossel10). Three aspects of all approaches
seem to be common: (i) usually, the data evaluation is geared
toward chemometrics, i.e., sort-of pattern-recognition
against library data; exact quantification is often of lesser
interest; (ii) because the chemometric aspect is the key in most
of the published works, detailed information on the pre-
evaluation procedures is normally sparse (e.g., it is not always
clear how well the particular, selected procedure would suit
quantification, as we require in our research); and (iii) data
acquisition and evaluation are normally sequential.
DATA ANALYSIS METHODS
The overall analysis procedure described here is composed
of a sequence of individual steps, each associated with its own
LabVIEW subroutine (subVI); these subroutines also can be
used on their own, in principle. The schematic flow chart for
this concept is shown in Fig. 1; the individual routines are
described in the sections below, in the sequence as they are
executed in the overall program chain.
For on-line applications, all steps are fully incorporated
into a program flow and require only minimal user intervention
during the initial set-up; for off-line applications, the sequence
shown in Fig. 1 is overlaid with a graphical front end. By
setting option switches in the program flow, individual steps in
the sequence may be skipped, should they not be required for a
particular spectrum analysis. It should be noted that LabVIEW
stands out when seamless integration of instrument control,
data acquisition, and signal analysis is desired. A similar
LabVIEW approach, albeit for the analysis of biochemical
samples and chemometric evaluation, has been described by
Vicarra Rossel.10 Of course, all individual tasks may be
programmed differently; thus, the underlying generic algo-
rithms for each are summarized in a supplemental material
document wherein we also provide download options for the
documented programs.
Cosmic Ray Removal. In any spectrum recorded by photon
detectors, cosmic ray events are encountered on a frequent but
random basis. For CCD array detectors, they manifest
themselves as (mostly) single-pixel responses where the
particular pixel carries a far greater intensity in comparison
with that of neighboring pixels. For accurate analysis of such
affected spectra, the cosmic-ray events need to be removed.
Numerous techniques and algorithms exist that can be
implemented for cosmic ray removal from one-dimensional
(1D) or two-dimensional (2D) spectral recordings (e.g.,
Home,11 Kelson,12 Li and Dai,13 and Mozharov et al.14). Since
in the work described here we only deal with 1D spectrum
traces, we do not elaborate further on 2D methodology (for a
brief summary on the latter, see Section 1 of the supplemental
material).
Having sets of spectra recorded over time, like in our case
(large sets of spectra are recorded during KATRIN runs and
during off-line control measurements), the least complicated
and very efficient method for identifying and eliminating

FIG. 1. Flowchart for integrated, sequential evaluation procedure, indicating
the action of each individual step.
cosmic rays is the use of spectral difference comparison.
Comparing at least two consecutive spectra recordings and then
determining the signal difference pixel by pixel implement this
concept. If the absolute value of this difference is less than a
threshold, the two spectral data points are averaged, and this
average is taken as the cleaned output value; otherwise, the
smaller valued data point is taken as the output value. The
threshold value has to be set large enough not to be
compromised by noise fluctuations but small enough to
efficiently capture the majority of cosmic ray events that can
have a wide range of random amplitudes. We have dubbed this
process temporal cosmic ray removal (TCRR), and it is
repeated for the selected set of rows or binning segments of the
detector chip, and for the full set of spectra.
It should be noted that TCRR has certain limitations. TCRR
works assuming that cosmic ray events are random and sparse,
whereas the spectral features of interest are static for
consecutive spectra. This assumption is no longer valid for
very long acquisition times. As the acquisition time expands,
the probability that cosmic rays will strike the detector in the
same or in close-proximity places in consecutive acquisitions
increases. Hence, an acquisition time threshold will be
encountered, where the simple (two back-to-back acquisitions
only) temporal routine will no longer remove all cosmic rays.
Above this threshold, more than two consecutive acquisitions
will be needed to ensure the spectra are completely clean of
cosmic rays.
In principle, one could see this multiple-comparison
approach to be akin the differentiation method described by
Mozharov et al.,14 who for a multitude of repeat measurements,
differentiated the spectra along the time axis at each pixel
(wavelength) position. The cosmic rays show up as disconti-
nuities in the derivative spectrum, and they can be eliminated
via neighboring pixel replacement.
The routine TCRR has been tested on high- and low-
intensity spectra for various acquisition times; for each
measurement series, cosmic ray events left in the spectra were
counted. The number of cosmic rays present in a spectrum will
vary depending on the location, shielding, and time of day.
Based on the test measurements with our PIXIS:2KB
(Princeton Instruments) and Synapse FIOP (Horiba) detectors,
the TCRR routine removed all cosmic rays with only two
consecutive spectra for acquisition times for an individual
spectrum of up to approximately 600 s (normally, acquisition
times are much shorter than this value). Above this value, three
or more consecutive spectra were required. The numbers stated
here may change for other detectors and measurement
conditions. However, the results show that TCRR can be very
successful and that for short acquisition times almost always
only two consecutive spectra are needed to remove all cosmic
rays.
Astigmatism Correction. All optical systems suffer from a
certain degree of astigmatism, i.e., distortion of straight lines by
curved optical components. As a consequence, in a relatively
short focal length (small f-number) spectrograph with wide
dispersion, the shift of a line maximum within pixel rows near
the top and bottom edges of a CCD array detector can be quite
severe. The simplest approach to correcting astigmatism
aberration in a well-aligned CCD-array detector is to shift all
individual rows until full alignment with a selected line
maximum in a reference row is achieved. This per-row
approach consists of two steps: (i) characterization of the
astigmatism curvature; and (ii) interpolation for the (fractional-
value) pixel shift, an interpolation that also includes a fractional
shift correction associated with the slight change of astigma-
tism as a function of wavelength. The implementation of this
procedure is based on our earlier work on Raman spectroscopy
of hydrogen isotopologues15 and is described in more detail in
the supplemental material section.
Background Removal. The removal of background from
spectral data has been subject to many studies, and numerous
algorithms have been developed for a variety of methodolo-
gies. However, it has to be noted that because of the large
variation of individual needs for such post-acquisition data
APPLIED SPECTROSCOPY 951
treatment, in different spectroscopic techniques and individual
experiments, particular procedures may not be universally
successful. Over the past few years, some efficient procedures
have been described and cross-compared, normally based on
well-established theories, but incorporating novel algo-
rithms.16,17 Although these procedures may be judged as quite
successful, it is equally clear that some approaches are time-
consuming (e.g., Schulze et al.18 state for their method that it
requires further development for in-line or real-time applica-
tions), or they need extensive user intervention for their
overall success. However, for real-time evaluation of spectra
that may suffer nonlinear variation of background and noise,
and that are accumulated continuously over extended periods
of time, processing speed is of the essence and user
intervention should be reduced to a minimum. For example,
the planned Raman monitoring in the KATRIN experiment or
the future ITER fusion experiment will generate spectra every
few seconds to minutes, over periods of several months for
individual measurement campaigns, with the need for frequent
feedback for measurement or process control.
We have explored the various published methodologies and
found the rolling circle filter (RCF) concept to have high merits
for background removal in conjunction with such months-long
tasks and for dealing with odd-shaped and time-varying base
lines. RCF requires no further user intervention once
parameters have been set for the specific task, but it can
thereafter cope automatically with nearly any not-too-sudden
variation in background or noise during long measurement
periods. Next, we describe our particular approach and the
modifications we made to the basic methodology. Although
this new approach to the RCF methodology was specifically
developed with the KATRIN and ITER experiments in mind, it
has already been used successfully in other experiments.7,8
Rolling Circle Filter. The RCF routine was originally
proposed by Mikhailyuk et al.19 The associated canonical
RCF is described, e.g., in Brandt et al.20 The concept of the
action of the RCF is visualized in Fig. 2. As suggested by its
name, it is broadly equivalent to rolling a circle of radius r
below a measured array of points, treating the array as a rigid
surface. The vertical locus at the top of the circle thus rolled is
taken as an estimate of the varying background level. For this
work, the RCF was implemented in LabVIEW, as summa-
rized in the supplemental material accompanying this
publication.
FIG. 2. Action of the RCF on an example spectrum; the filter tends to roll into
the peak, overcompensating the background level. This deficiency is overcome
by using our SCARF filter routine.
952 Volume 67, Number 8, 2013
SavitzkyGolay Coupled Advanced Rolling Circle Filter
(SCARF). The canonical RCF algorithm has a serious
deficiency, namely, that the baseline level is overestimated if
a gap (e.g., a peak or broad unresolved feature) is encountered
whose baseline width, w, is larger than double the RCF-circle
radius, i.e., w  2r. Then, the propagating circle rolls into the
gap (see Fig. 2), thus subtracting more than the actual
background. But, simply increasing the RCF-circle to much
larger values does not necessarily fully solve the problem, as
will be shown in the examples below.
Our proposed SCARF algorithm can ameliorate this
deficiency. Simply put, it works by applying a SavitzkyGolay
filtering (SGF) step to the locus-array generated from a single
RCF pass. Two key parameters in the SGF algorithm have to
be set appropriately, namely, (i) the number of side points, s 
r, to be included (the number of side points corresponds to the
half-width of the SGF window), and (ii) the polynomial order n
of the smoothing process. Note that in this work the
polynomial order was always n = 3 (but this can be altered
if so desired). It also should be stressed that the SGF-step is
applied to the background estimate, and not to the spectral data
itself.
The procedure is surprisingly robust even if parameters were
not set to their optimum values. However, care should still be
taken to choose a suitable value of r in the RCF segment of the
routine and the number of side points, s. Specifically, r should
be as small as possible in order to capture details in the
variation of the baseline level but normally should not be
smaller than the widest spectral peak of interest (i.e., 2r 
wmax), whose shape one wishes to preserve, sitting upon the
moving baseline. For broad background features within the
moving baseline, the limit of r may be reduced to ensure the
baseline is followed correctly. But, as a consequence of
reducing 2r to values below wmax, the number of side points
needs to be increased. However, decreasing 2r too far may
have unwanted, rather adverse effects, as shown in the
examples below. Finally, for most efficient removal of odd-
shaped background from peaks of varying width multiple
passes of the routine may be required.
To demonstrate the action of the standard RCF and the
advanced SCARF routines, we applied them for a range of
filter parameters, specifically to (i) a synthetic line spectrum
and (ii) a real Raman spectrum.
In the first instance, a simple synthetic spectrum has been
generated, consisting of a pure Gaussian line of full width half-
maximum (FWHM) w = 10 pixel, sitting on an offset
background dressed with random shot noise fluctuations. The
amplitude of the spectral line was changed from values just
above the noise level to values at which the noise becomes
nearly negligible. The default starting parameter for the RCF
procedure was r = 2w. The results are summarized in Fig. 3
and Table I; note that the entries in the table are the amplitudes,
in units of counts, recuperated by peak-fitting the background-
corrected spectra after filter treatment.
Both the figure and table data clearly reveal that selecting the
filter parameters too small with respect to the base width of the
peak results in the filter rolling into the peak, and thus
subtracting an incorrect value. It is clear that peaks of equal
FWHM, but substantially different amplitude, exhibit different
behavior. One notices that it is indeed the width at the base of
the peak (here, the width points are taken at two times the noise
fluctuations above the baseline) that is responsible for how far
FIG. 3. Background removal for synthetic spectrum of a Gaussian peak, with
FWHM = 10 pixels, and different amplitude Apeak: (a) Apeak = 103 and (b)
Apeak = 105. The baseline is offset by 1000 counts, exhibiting random shot-
noise background. The background-fit traces a, b, d, and f are associated with
the various RCF and SCARF parameter entries in Table I; for clarity, they are
consecutively offset downward.
the filter rolls into the peak and hence would falsify the actual
baseline level. The application of the SCARF routine with the
same width parameter improves this but cannot fully
compensate. In fact, one may even complicate matters: due
to the sharp edges of the RCF (20) signal at the position of the
peak, the SCARF routine introduces a negative-going second-
derivative shape.
The implication is that the underlying RCF algorithm has to
start with a sensible value that needs to account for the width
at the base of any peak of interest for quantitative analysis. It
is also clear that in that case even RCF on its own results in
deviations of the order of noise median. However, it should be
noted that this only holds true for nearly flat background
levels. Finally, a nearly perfect background removal and
baseline correction function is achieved when applying the
SCARF routine twice, with staggered number of side points
included. Note that in this repetitive application the second
SCARF-passage acts on a modified data set, namely, the
original (raw) spectral data from which the background
estimate of the first SCARF-run has been subtracted. Note
also that in general the parameters for the first SCARF run are
set to remove an overall offset and slowly varying slope
features and that the in the second run (on the modified data
set) aims at dealing with rapidly varying background features
(see the example in the supplemental material for a
demonstration example); in general, this means for the related
SG parameter that srun#1 . srun#2.
It is also worth noting that repetitive application of RCF with
appropriately selected parameters may improve the overall
background subtraction. However, since RCF always acts on
the dataother than SCARF that incorporates coupled residual
smoothing before subtraction of the background estimate
there will always be a noticeable effect on the peak data, which
may affect exact quantification.
The same procedure as mentioned above was applied to a
real N2-Raman spectrum that was, in addition, superimposed
with a nonlinear background contribution (here, e.g., light from
a 605 nm light-emitting diode). In Fig. 4, the results from a
selected range of filter functions on this spectrum are shown.
The original spectral data traces are overlaid with the derived
background functions, namely, the circle loci for the normal
FIG. 4. Background removal test for a Raman spectrum of N2, overlaid with
shaped background light; traces are offset by 1500 units, consecutively from
top to bottom. The traces are annotated with the respective filter actions used,
RCF(r) and SCARF(r,s).
RCF-routine, RCF(r), with two different radii, r = 20 and r =
60; our SCARF-routine SCARF(r,s) with (r = 60, s = 60); and
the repetitive SCARF-routine with varied SavitzkyGolay
parameters (r = 60, s = 240) followed by (r = 60, s = 120).
Clearly, in the standard RCF routine, the filter has rolled into
the spectral lines for RCF with r = 20 pixels; but even for r =
60 pixels that is much wider than the narrow S1- and O1-branch
Raman lines, hints of roll-in are evident. This results in loss of
quantifiable spectral information because incorrect amounts of
background intensity are subtracted. Although this may not be
seen as critical for the large-intensity feature around 2330 cm1
(the capped Q1-branch of 14N2 in Fig. 4), where the lost
amount accounts for roughly 102, or less for the wider circle
radius, the low-intensity feature near 2285 cm1 (Q1-branch of
14 15
N N) suffers a loss of as much as 2030%. Therefore, for
the latter, one would have extracted incorrect values from the
spectra, and reliable quantification is most likely lost.
As in the synthetic spectrum case discussed above, applying
the SCARF routine has nearly eliminated the problem of roll-
in. However, with the larger radii parameters r = 60 and
s = 60, the background trace now does not follow the
background curvature correctly; consequently, the actual
background is not yet completely removed. Full background
compensation is finally achieved by applying SCARF a second
time on the adjusted spectrum from the first run (see the lowest
trace in Fig. 4). The actual quantitative data for the abundance
APPLIED SPECTROSCOPY 953
TABLE I. RCF and SCARF applied to a synthetic Gaussian peak of FWHM w = 10 (see Fig. 3), with different amplitude, and offset by Aoffset = 1000 6
random shot noise (in units of counts; width of noise-only distribution = 33). The parameters r and s of RCF(r) and SCARF(r,s) are in units of pixel. The
data entries correspond to the amplitudes obtained by peak fitting after application of the filter procedure to the spectra (in units of counts), with fitting
errors in parentheses.

Recovered peak amplitude

Nominal peak amplitude Apeak (counts) 100 1000 10 000 100 000
Peak width at base (pixel) 26 48 62 77
(a) RCF (20) 67 6 12 829 6 17 8590 6 30 86 413 6 35
(b) SCARF (20,20) 76 6 13 899 6 27 9350 6 67 94 218 6 98
(c) SCARF (20,80) (20,40) 82 6 13 916 6 26 9574 6 58 96 430 6 87
(d) RCF (60) 83 6 10 980 6 15 9963 6 22 99 940 6 25
(e) SCARF (60,60) 89 6 10 984 6 14 9985 6 20 99 968 6 22
(f) 23SCARF (60,240) (60,120) 98 6 8 993 6 12 9991 6 15 99 990 6 15

of N2 isotopologues, based on the data in Fig. 4, are included in
Table II.
Finally, for quantitative analysis of spectral lines via peak-fit
routines, a flat background with the noise oscillating around
zero is required. However, the application of RCF/SCARF for
background removal as is exhibits a minor deficiency. Since
the circle always rolls below all data points, this slightly offsets
any noise baseline to a minute value above zero, namely, about
the half-width of the peak-to-peak noise fluctuations. This
noise-median value for shifting the SCARF-treated spectrum to
a nominal zero-level can be obtained semiautomatically. For
this, one simply calculates the median of a (reasonably) flat,
noise-only region of the spectrum. Note that the flattest
background, most suitable for the determination of the noise
median, is achieved by tuning the radius of the circle in the
SCARF routine to rather small values; however, other than for
the (once or occasional) determination of the noise median, and
the SCARF radius and the side-point ratio needs to be run at
the task-optimum.
Analysis of Peak Intensities. The last step in the post-
acquisition data treatment is the extraction of peak intensities
(or areas). Two types of techniques have been considered for
the determination of intensities (or areas).
The first technique uses integration by a simple summing of
the intensity of all pixels within a peak profile, arriving at a
peak-area value. However, this method is unsuitable if one
wishes to separate overlapping lines. Fitting combinations of
Voigt profiles21,22 or other line shape functions may in general,
treat convoluted lines. This approach has some notable
drawbacks. First, the spectral line emission traverses through
an imaging system that normally contains line shape-
influencing components, like small-core optical fibers, the
spectrometer entrance slit, the grating and the CCD-detector
pixel structure. As a consequence, the original line shape
normally deviates from pure Voigt profile functions that are
frequently used in conjunction with quantum-dominated
spectral lines. Second, the number of free parameters in
minimization problems rises by four for each peak (position,
intensity, FWHM, and GaussLorentz fraction).
The method we have used throughout in this work is a
routine we named ShapeFit. The method can be applied
without restriction, if the line shape of all spectral lines of
interest is the same. This is the case if (i) the line width is
limited by the slit width of the spectrometer, optical fiber, or
both; or (ii) the spectral line width (i.e., natural line width plus
broadening) of all lines is the same. Note that if neither of these
conditions is met, the procedure becomes more complex, since
multiple peak shapes and their relative intensity weighting are
involved; this generalized case is not treated here, although we
have already programmed a suitable routine.
In the first step of the routine, a line is selected from the
spectrum that has sufficient intensity and stands isolated (no
convolution with other lines); for example, this could be a line
from a spectral calibration lamp. The shape of this peak is
stored pixel by pixel in an auxiliary data array.
In the second step, all lines within a spectrum are fitted,
using the previously determined digital peak-shape function
and multiplying it by the appropriate amplitude factor. The
process incorporates the LevenbergMarquart algorithm23 in
which amplitude and the center-position of each peak are
treated as fitting parameters. Subpixel translation of the peak
positions is enabled by interpolation of the numerical peak
shape. The baseline of the spectrum can be either fixed, or
added as a fitting parameter. Since the peak shape is stored
numerically, very complicated and odd shapes can be used,
whereas the number of free parameters per peak is still only
two (position and amplitude). The principal details of the
routine algorithm are summarized in the supplemental material
section.
Particular examples for the application of ShapeFit to
measured Raman spectra are discussed in the Results section.
There, it will be seen that despite strong convolution of some
peaks and the occasional use of intentionally bad line shapes,
the fitting could be applied successfully.

TABLE II. Isotopologue abundance for O2 and N2, extracted from Raman spectra, after application of RCF or SCARF in conjunction with ShapeFit.

Isotopologue ratio RCF (20) 3 103 RCF (60) 3 103 SCARF (60,240) (60,120) 3 103 Theoreticala 3 103
16
O17O/16O2 (b)
0.51 6 0.08 0.63 6 0.09 0.75 6 0.06 0.76 6 0.03
16
O18O/16O2 3.54 6 0.20 3.81 6 0.18 4.02 6 0.15 4.02 6 0.14
14 15 14
N N/ N2 5.70 6 0.22 6.93 6 0.21 7.27 6 0.16 7.35 6 0.20
a
Stochastic distribution data for isotopologues for atmospheric molecules27; errors estimated from isotope abundance data.26
b
Data for 16O17O corrected for hot-band contribution from 16O2.

954 Volume 67, Number 8, 2013

 For the sake of completeness, it should be mentioned that
line-fitting procedures based on data-driven multivariate
analysis could be implemented, such as partial least squares
(PLS). However, strictly speaking, this is not (individual) peak
fitting, but rather a comparison of the spectrum of a constituent,
as a whole, to reference data sets. For PLS in the context of
Raman spectroscopy, please refer to various publications for
further details.9,24

RESULTS AND DISCUSSION

In the previous section, the individual steps of the analysis
procedure were discussed. Now, we discuss the results from the
complete analysis chain for some selected examples that
specifically highlight the interplay of background removal
routines with quantitative peak analysis. Predominantly, we put
results obtained using our SCARF and ShapeFit techniques
into context with expected theoretical or experimental values.
The data for the molecules addressed in the examples below
were obtained using the same measurement setup as in the
systematic study of Raman depolarization ratios of the
hydrogen isotopologues (see James et al.5). In brief, the
orientation of the excitation laser (532 nm Nd : YAG laser, 25
W) is well controlled before entering the Raman cell, using a
combination of a k/2-plate and a Glan polarizer. The
polarization in the light collection path is predetermined by
using a high-quality linear (sheet) polarizer. The light is
coupled to the spectrometer (Acton SP500i) via a slit-shaped
fiber bundle for optimum light collection; the spectrometers
grating is selected to provide sufficient resolution to resolve the
relevant ro-vibrational branch lines and at the same time cover
the full spectral range of interest. The detector is a TE-cooled,
2D-array CCD detector.
Example 1: Fundamental Vibrational Raman Bands of
FIG. 5. Raman spectrum of O2 (at p = 970 mbar), recorded under 532 nm
O2 and N2. A high-resolution Raman spectrum of the laser excitation (Plaser = 2 W). Top panel. (a) Raw data spectrum (trace a), with
fundamental vibrational band of oxygen is shown in Fig. 5. the background determined using a RCF (20) filter function (trace b) and a
The individual lines of the ro-vibrational branches with DJ = SCARF (60,240) (60,120) filter function (trace c); the background traces are
6 2, i.e., O1(J) and S1(J), are well resolved while those of shifted downward by 200 and 400 units, respectively, for clarity. Bottom panel.
(b) The O1- and S1-branch lines (fitted using ShapeFit) have been subtracted,
the Q1-branch are not. Closer inspection of the low- leaving the Q1-branches of the oxygen isotopologues in isolation; note the
wavenumber part of the spectrum reveals that some additional different amplitude scale. The symbols indicate atomic lines from stray light of
features are underlying to the O1-branch. To interpret these the laboratory lighting.
features quantitatively, correct background treatment followed
by peak fit procedures is required. The top panel of Fig. 5
clearly shows different widths of the various spectral features combined SCARF/ShapeFit approach, and leaving the back-
of O2 and a curvature of the background contribution. Thus, ground plus subtraction-residual fluctuating at the dark current
care needs to be taken of how to correctly compensate the noise level. The remaining features are now easy to interpret; as
background so that accurate quantification of the actual annotated in the figure, they can be identified as (i) laboratory
spectrum is maintained. In the figure, the two background lighting having found its way into the enclosure (some Hg-
treatments using RCF and SCARF demonstrate this. lines from fluorescent tubes), and (ii) the Q1-branches of the
Subsequent to the removal of background, the rotational minor isotopologues of oxygen.
lines of the two resolved branches were fitted using our Since the (unresolved) Q1-branches of all oxygen isotopo-
ShapeFit routine. From the extracted line intensity data, the logues exhibit roughly the same envelope, we have fitted the
rotational temperature could be calculated and was determined minor contributors, i.e., 16O17O and 16O18O, with the shape of
as Trot = 297 6 2 K, which closely replicates the average the dominant 16O2 feature (using this secondary profile in
temperature in our Raman experiment enclosure (298 6 1 K). ShapeFit). The results from this evaluation are collected in
This information is then used to construct the O1(J = 3) Table II, where the data based on different background removal
line, using the theoretical thermal amplitude and the line shape options and the theoretical values are compared with each
associated with the above-mentioned ShapeFit routine. This other. It is evident that less-than-optimum background removal
line eluded accurate, direct fitting in the original spectrum can lead to substantial deviations from the expected isotopo-
because of its severe overlap with the Q1-branch. logue ratios, assuming stochastic isotope distribution25,26 (in
Finally, all S1(J) and O1(J) lines are subtracted from the the worst case, .30%).
original, background-corrected spectrum. The result is shown It is important to note that the Q1-branch of 16O17O overlaps
in the bottom panel of Fig. 5, demonstrating the strength of the with the hot-band v = 1 ! v = 2 of 16O2, with their band

APPLIED SPECTROSCOPY 955

origins only differing by approximately 0.8 cm1 (calculated
using data from Eiler,27 a value that is below the spectral
resolution of these measurements). Said hot-band can be
separated out by subtracting the band contribution deduced
from the vibrational population temperature (assumed to be
equal to the room temperature) and the band shape of the main
Q1-branch of 16O2.
In a similar manner, we analyzed the ro-vibrational branches
of nitrogen (N2) recorded under high resolution (see Fig. 4). As
for oxygen, the rotational O1(J) and S1(J) branch lines are
well resolved, whereas those of the Q1-branch are not. When
inspecting the residual of the background spectrum, after
subtraction of the O1- and S1-branch lines fitted using ShapeFit,
one is left with the weak Q1-branch of the isotopologue 14N15N
at 2290 cm1, accompanying the main 14N2 peak. In the
original spectrum in Fig. 4, the minor isotopologue can just be
identified as a small deviation from the O1-branch line pattern.
Only when appropriately using SCARF is no quantitative
information is lost. As for oxygen, the quantification of the
minor nitrogen isotopologue, relative to the dominant isotopo-
logue, is summarized in Table II. Those values are derived after
application of the background removal procedure and appro-
priate subtraction of the masked O1(J = 3) line, using its
theoretically derived intensity. This latter step becomes
possible, as for oxygen, by exploiting the knowledge of the
rotational temperature Trot = 299 6 2 for N2, based on all
(fitted) O1- and S1-branch lines.
This example of the O2 and N2 isotopologue data
demonstrates how vitally important correct background
treatment and line fitting are, if any exact, quantifiable data
are to be extracted from the spectra. Clearly, SCARF fits this
bill, so that by using ShapeFit one is able to extract accurate
peak intensity information so that selected features in the
spectra may be subtracted.
Example 2: Raman Spectrum of Q1-Branch of DT. In the
second example, we demonstrate the ability of the ShapeFit to
quantitatively retrieve semi-hidden lines from strongly
overlapping peaks, even if the line profiles are of odd shape
that cannot be modeled by standard line shape functions such
as Gauss, Lorentz, or Voigt profiles. The DT spectra shown in
Fig. 6 were recorded from a gas mixture of T2/DT/D2, of total
pressure 400 mbar; the DT fraction contributing to the
displayed spectra was approximately 46.7 % (determined via
gas chromatography analysis).
The Raman-scattered light for the spectrum displayed in Fig.
6a was imaged optimally onto the detector plane. This FIG. 6. Raman spectra of DT recorded for (a) an optimum fiber bundle image
constitutes a diffraction-limited case for which the peaks can and (b) a purposely misaligned fiber image on the entrance slit. The data in
be approximated reasonably well by a Gaussian profile. Thus, panel a fit well to the experimental data using a Gauss function, and a numerical
we attempted a Gaussian fit to the spectral lines; note that for line profile for our ShapeFit routine; the data in panel b only fit using a
numerical line profile for our ShapeFit routine. Note the different scales for the
the fit, the line width parameter was conditional, i.e., it was set residuals.
to be the same for all individual Q1(J) peaks. The residuals for
the Gaussian fit reveal that it is indeed a good approximation.
This fit is improved even further when applying our Clearly, the very asymmetric line profiles are no longer
numerical ShapeFit routine, whose digitized profile function represented by Gaussian-, Lorentzian- or Voigt-profile func-
accounts for minute deviations from a perfect Gaussian. In the tions. In attempting fits using any of the three profile functions,
latter case, the residuals become even smaller, approximating only the Gauss-function fit converged, providing a reasonable
the line profiles to better than 104, reaching the shot-noise sequencing of the relative line intensities, as the results in Table
limit in the measurements. III reveal. However, the fits do not quantitatively replicate the
The situation changes dramatically in the case of a peak experimental peak-heights and line-widths. Specifically, (i)
shape that is no longer describable by an analytical function. deconvolution of the J = 0 line from the dominant J = 1 line
Such an example is shown in Fig. 6b; here, the light was failed, and (ii) the Gaussian approximation of this odd line
focused asymmetrically onto the spectrometer entrance slit. shape would not allow one to carry out a quantitative line

956 Volume 67, Number 8, 2013

TABLE III. Relative rotational population in DT(v = 0), extracted from the Raman Q1-branch spectra displayed in Fig. 6, after application of peak-
fitting routines. Data for experimental enclosure temperature of T = 299 6 1 K.
Distorted line profile
Rotational level J Gauss fitb ShapeFit
0 c 0.3743 6 0.0040
1 1.0238 6 0.0139 1.0011 6 0.0050
2 1 1 1
3 0.6444 6 0.0101 0.7004 6 0.0044
4 0.3191 6 0.0092 0.3510 6 0.0039
5 0.1217 6 0.0057 0.1309 6 0.0026
6 0.0355 6 0.0032 0.0358 6 0.0017
7 0.0093 6 0.0020 0.0087 6 0.0010
8 c 0.0012 6 0.0005
9 d d e
a
Theoretical population data calculated for T = 299 K, based on the transition moment values calculated by LeRoy (personal communication).
b
The Gauss fit only replicates the relative peak relation, but neither the absolute amplitudes nor the line widths.
c
Did not converge.
d
Below noise level.
e
Overlap with emission of impurity molecule.
subtraction, as required for the extraction of the minor O2 and
N2 isotopologues.
In contrast, using our ShapeFit routine based on the
numerical representation of the profile, one is able to fit each
of the peaks and extract the exact convoluted intensities,
despite the very odd line shape. The residuals for the line
profile fits are of the order 103 to 104, being close to the shot-
noise limit in the measurements. The apparently lower
residuals are due to the fact that the overall signal amplitudes
are four times lower than for the optimum-imaging case,
depicted in Fig. 6a; note that the overall integrated line
intensities are nearly equal in both cases.
The example shown in Fig. 6 demonstrates the improvement
our ShapeFit approach can have over standard, analytical-
function fitting routines. The relative line amplitudes obtained
for the two line shape cases and the different fitting approaches
are collected in Table III, and the experimental values are
compared to theoretically derived amplitudes, based on the
Boltzmann distribution function with T = 299 6 1 K,
reflecting the enclosure temperature of Texp = 298 6 1 K in
the experimental setup.
All data were normalized to the Q1(J = 2) line, the first
(nearly) fully resolved line. Clearly, the ShapeFit results
reproduce the theoretical values best, being off by less than 1%
for most lines; larger deviations are only observed for the two
overlapping lines (deconvolution algorithms always introduce
some bias errors), and for the very weak lines J = 7 and
J = 8, for which background noise starts to become a
significant contributing uncertainty.
It should be noted that in all likelihood the deviations would
be even less were the spectral sensitivity variations of the
detection system taken into account. Within the spectral region
displayed in Fig. 6, the transmission function of the specific
532 nm Raman edge-filter (RazorEdge, Semrock), used to
suppress the laser excitation line, exhibits small oscillations in
its transmission function (of the order 0.20.3%).
CONCLUSIONS
As described in this work, we developed an integrated
concept for post-acquisition spectrum analysis that encom-
passes the all important data treatment procedures necessary to
Clean line profile
Gauss fit ShapeFit Theoreticala
0.4157 6 0.0012 0.4137 6 0.0013 0.3930
0.9717 6 0.0017 0.9756 6 0.0016 0.9834
1 1
0.6868 6 0.0015 0.6857 6 0.0014 0.6834
0.3387 6 0.0011 0.3391 6 0.0011 0.3400
0.1282 6 0.0008 0.1280 6 0.0006 0.1278
0.0368 6 0.0005 0.0369 6 0.0005 0.0371
0.0083 6 0.0004 0.0083 6 0.0004 0.0084
0.0016 6 0.0003 0.0015 6 0.0003 0.0015
e 0.0002
improve on the quality of relevant, quantitative results. These
procedures include, amongst other routines: cosmic ray
removal, specifically; background subtraction; and peak
analysis and fitting.
The individual spectra-treatment algorithms were selected to
only minimally, or not at all, affect the quantitative integrity of
any spectrum. This spectral evaluation suite is composed of a
sequence of the aforementioned tasks, and further steps.
However, individual steps may be easily used on their own if
so desired, or skipped if not required. Having all steps fully
incorporated into a single program flow that does require only
minimal user intervention after the initializing parameter setup
is not only paramount in on-line applications but also desirable
for the majority of (routine) off-line applications. In our
approach, we adapted and expanded some of the most elegant
ideas reported in the literature to arrive at solutions with low
coding effort and as little intervention as possible after
training the procedure for specific, repetitive spectrum types.
Within the overall program flow, the two subtasks central to
this work were those of background treatment (together with
cosmic ray removal) and peak analysis and fitting. Generally,
these are the most important subtasks when it comes to
quantification of spectrum content information.
For the former, we adopted the canonical RCF algorithm,
but coupled it with a SGF step on the locus-array generated
from a single RCF pass. This vastly improved on RCFs
deficiency to overestimate the baseline level in spectra with
broad peak features. This only two-parameter procedure,
dubbed SCARF, was surprisingly robust; it resulted in superb
background suppression, even if the parameters were not set to
their full optimum. Of course, the parameters have to be
sensibleeven the most intelligent routine normally cannot
cope with nonsensical requests. Also, in the majority of
complex background cases dual application of SCARF with
appropriate parameter settings is advisable.
For the latter task of peak analysis, we developed a routine
we named ShapeFit. As outlined above, this is an only two-
parameter (amplitude and position) fitting algorithm that relies
on numerical line shape profiles rather than analytical
functions, such as, e.g., a Gauss profile function. Of course,
the profile function has to be supplied by the user, and it is
normally based on an isolated spectral line feature. The great
APPLIED SPECTROSCOPY 957
advantage of this approach is that even for rather odd-shaped
spectral lines full quantitative information is returned. Note
that in applications that require spectral typing, encountered
by and large in the characterization of biomedical samples, the
peak-fitting routine will need to be replaced by a PLS/PCA
routine.
We also demonstrated the strength of our integrated
program sequence for the analysis of a range of Raman
spectra with nonideal line profiles and background contribu-
tions that may vary over time. As an example, it was shown
that in comparison to some other analysis algorithms and
functions, our approach succeeded to return improved
quantitative results for the relative abundance of minor
molecular isotopologues that were nearly hidden in the
spectra. In particular, ShapeFit excelled in the case of
nonanalytical line profiles. Furthermore, our integrated
analysis package clearly has the potential for real-time
analysis. Performed on a modern desktop computer with a
standard configuration (dual-core 2.7 GHz CPU, 4GB PC3
SDRAM memory and Windows 7 OS), the full spectrum
evaluation, including cosmic-ray removal using TCRR,
background evaluation using 23SCARF, and fitting up to
five peaks, is complete in 200350 ms, depending on the
number of peaks and the complexity of the spectrum. This
indicates that our goal of real-time analysis is indeed within
reach, since the overall evaluation times were less than
average Raman signal-recording and read-out times of 400
450 ms in process control applications and, therefore, can be
performed in parallel to the data acquisition. For some
additional details on the timing procedure, see Section 1 of the
supplemental material.
Finally, the overall analysis chain presented here was fully
programmed in National Instruments LabVIEW. Although
this may not be seen as the most obvious programming
methodology, this approach allows one to easily incorporate
this spectrum analysis suite, and any possible extensions, into
any LabVIEW-controlled instrument-control or data-acquisi-
tion environment, or both.
The umbrella program and the generic algorithms for all
tasks are described in the supplemental material associated with
this publication. The documented program suite can be
downloaded from http://sourceforge.net/projects/spectools/.
SUPPLEMENTAL MATERIAL
All supplemental material mentioned in the text, associated
with the description of the evaluation procedures and
algorithms, can be found in the online version of the journal
at http://www.s-a-s.org.
ACKNOWLEDGMENTS
This research was supported by the Helmholtz Association (HGF), the
German Ministry for Education and Research (BMBF), through grants
05A08VK2 and 05A11VK3; and the Deutsche Forschungs Gemeinschaft
(DFG) within the framework of the TransRegio project Neutrinos and
Beyond, grant SFB/TR 27. T.M.J. acknowledges the financial support for his
Ph.D. studentship by the UK Engineering and Physical Sciences Research
Council (EPSRC). The authors thank the Karlsruhe House of Young Scientists
(KHYS) for supporting part of this study via a visiting scientist stipend. Also,
we thank T.L. Le, K.H. Simon, and E. Fanghanel of the TLK-CAPER team for
the preparation of tritium containing gas samples.
1. J. Angrik. and the KATRIN Collaboration. The KATRIN Design Report
2004. FZK Scientific Report. 7090. 2005. http://bibliothek.fzk.de/zb/
berichte/FZKA7090.pdf [accessed Mar 18 2013].
958 Volume 67, Number 8, 2013
2. M. Babutzka, M. Bahr, J. Bonn, B. Bornschein, A. Dieter, G. Drexlin, K.
Eitel, S. Fischer, F. Gluck, S. Grohmann, M. Hotzel, T.M. James, W.
Kafer, M. Leber, B. Monreal, F. Priester, M. Rollig, M. Schlosser, U.
Schmitt, F. Sharipov, M. Steidl, M. Sturm, H.H. Telle, N. Titov.
Monitoring of the Properties of the KATRIN Windowless Gaseous
Tritium Source. New J. Phys. 2012. 14(10): 103046. doi: 10.1088/
1367-2630/14/10/103046.
3. S. Fischer, M. Sturm, M. Schlosser, B. Bornschein, G. Drexlin, F. Priester,
R.J. Lewis, H.H. Telle. Monitoring of Tritium Purity During Long-Term
Circulation in the KATRIN Test Experiment LOOPINO Using Laser
Raman Spectroscopy. Fusion Sci. Technol. 2011. 60(3): 925-930.
4. M. Sturm, M. Schlosser, R.J. Lewis, B. Bornschein, G. Drexlin. H H.
Telle; Monitoring of all Hydrogen Isotopologues at Tritium Laboratory
Karlsruhe Using Raman spectroscopy. Laser Phys. 2010. 20(2): 493-507.
doi: 10.1134/S1054660X10030163.
5. T.M. James, M. Schlosser, S. Fischer, M. Sturm, B. Bornschein, R.J.
Lewis, H.H. Telle. Accurate Depolarization Ratio Measurements for All
Diatomic Hydrogen Isotopologues. J. Raman Spectrosc. 2013. 44(6):857-
856. doi: 10.1002/jrs.4283.
6. M. Schlosser, S. Rupp, H. Seitz, S. Fischer, B. Bornschein, T.M. James,
H.H. Telle. Accurate Calibration of the Laser Raman System for the
Karlsruhe Tritium Neutrino Experiment. J. Mol. Struct. 2012. 1044: 61-
66. doi: 10.1016/j.molstruc.2012.11.022.
7. J.F.M. Almarashi, N. Kapel, T.S. Wilkinson, H.H. Telle. Raman
Spectroscopy of Bacterial Species and Strains Cultivated Under Repro-
ducible Conditions. Spectroscopy. 2012. 27(5-6): 361-365. doi:10.1155/
2012/540490.
8. A.G. Gonzalvez, A. Gonzalez Urena, R.J. Lewis, G. van der Zwan.
Spectroscopy and Kinetics of Tyrosinase Catalyzed Trans-Resveratrol
Oxidation. J. Phys. Chem. B. 2012. 116(8): 2553-2560. doi: 10.1021/
jp209753q.
9. L.A. Reisner, A. Cao, A.K. Pandya. An Integrated Software System for
Processing, Analysing, and Classifying Raman Spectra. Chemometr.
Intell. Lab. Syst. 2011. 105(1): 83-90. doi: 10.1016/j.chemolab.2010.09.
011.
10. R.A. Vicarra Rossel. ParLeS: Software for Chemometric Analysis of
Spectroscopic Data. Chemom. Intell. Lab. Syst. 2008. 90(1): 72-83. doi:
10.1016/j.chemolab.2007.06.006.
11. K. Horne. An Optimal Extraction Algorithm for CCD Spectroscopy.
Publ. Astron. Soc. Pac. 1986. 98(6): 609-617. doi: 10.1086/131801.
12. D.D. Kelson. Optimal Techniques in Two-Dimensional Spectroscopy:
Background Subtraction for the 21st Century. Publ. Astron. Soc. Pac.
2003. 115(808): 688-699. doi: 10.1086/375502.
13. S. Li, L.K. Dai. An Improved Algorithm to Remove Cosmic Spikes in
Raman Spectra for Online Monitoring. Appl. Spectrosc. 2011. 65(11):
1300-1306. doi: 10.1366/10-06169.
14. S. Mozharov, A. Nordon, D. Littlejohn, B. Marquardt. Automated
Cosmic Spike Filter Optimized for Process Raman Spectroscopy. Appl.
Spectrosc. 2012. 66(11): 1326-1333. doi: 10.1366/12-06660.
15. R.J. Lewis, H.H. Telle, B. Bornschein, O. Kazachenko, N. Kernert, M.
Sturm. Dynamic Raman Spectroscopy of Hydrogen Isotopomer Mixtures
In-Line at TILO. Laser Phys. Lett. 2008. 5(7): 522-531. doi: 10.1002/
lapl.200810026.
16. G. Schulze, A. Jirasek, M.M.L. Yu, A. Lim, R.F.B. Turner, M.W. Blades.
Investigation of Selected Baseline Removal Techniques as Candidates for
Automated Implementation. Appl. Spectrosc. 2005. 59(5): 545-574. doi:
10.1366/0003702053945985.
17. C. Rowlands, S. Elliott. Automated Algorithm for Baseline Subtraction in
Spectra. J. Raman Spectrosc. 2011. 42(3): 363-369. doi: 10.1002/jrs.
2691.
18. H.G. Schulze, R.B. Foist, K. Okuda, A. Ivanov, R.F.B. Turner. A Small-
Window Moving Average-Based Fully Automated Baseline Estimation
Method for Raman Spectra. Appl. Spectrosc. 2012. 66(7): 757-764. doi:
10.1366/11-06550.
19. I.K. Mikhailyuk, A.P. Razzhivin. Background subtraction in Experimen-
tal Data Arrays Illustrated by the Example of Raman Spectra and
Fluorescent Gel Electrophoresis Patterns. Instrum. Exp. Tech. 2003.
46(6): 765-769. doi: 10.1023/B:INET.0000008930.87741.b9.
20. N.N. Brandt, O.O. Brovko, A.Y. Chikisev, O.D. Paraschuk. Optimization
of the Rolling-Circle Filter for Raman Background Subtraction. Appl.
Spectrosc. 2006. 60(3): 288-293. doi: 10.1366/000370206776342553.
21. F. Alsmeyer, H.J. Koss, W. Marquardt. Indirect Spectral Hard Modeling
for the Analysis of Reactive and Interacting Mixtures. Appl. Spectrosc.
2004. 58(8): 975-985. doi: 10.1366/0003702041655368.
22. E. Kriesten, F. Alsmeyer, C.B. Minnich, L. Greiner, W. Marquardt.
Identification of Unknown Pure Component Spectra by Indirect Hard
 Modeling. Chemometr. Intell. Lab. 2008. 93(2): 108-119. doi: 10.1016/
j.chemolab.2008.05.002.
23. D. Marquardt. An Algorithm for Least-Squares Estimation of Nonlinear
Parameters. SIAM J. Appl. Math. 1963. 11(2): 431-441.
24. S.K. Oh, S.J. Yoo, J.M. Lee. Predicting Concentrations of a Mixture in
Bioreactor for On-Line Monitoring Using Raman Spectroscopy. Paper
presented at: 8th IFAC Symposium on Advanced Control of Chemical
Processes. Furama Riverfront, Singapore; July 10-13 2012.
25. K.J.R. Rosman, P.D.P. Taylor. Isotopic Compositions of the Elements,
1997. Pure Appl. Chem. 1998. 70(1): 217-235. doi: 10.1351/
pac199870010217.
26. J.K. Bohlke, J.R. de Laeter, P. De Bie`vre, H. Hidaka, H.S. Peiser, K.J.R.
Rosman, P.D.P. Taylor. Isotopic Compositions of the Elements, 2001. J.
Phys. Chem. Ref. Data. 2005. 34(1): 57-68. doi: 10.1063/1.1836764.
27. J.M. Eiler. Clumped-Isotope GeochemistryThe Study of Naturally-
Occurring, Multiply-Substituted Isotopologues. Earth Planet. Sci. Lett.
2007. 262(3-4): 309-327. doi: 10.1016/j.epsl.2007.08.020.
APPLIED SPECTROSCOPY 959