sensors

Letter
An Automatic Baseline Correction Method Based on
the Penalized Least Squares Method
Feng Zhang , Xiaojun Tang *, Angxin Tong, Bin Wang and Jingwei Wang
State Key Laboratory of Electrical Insulation & Power Equipment, Xi’an Jiaotong University, Xi’an 710049, China;
[email protected] (F.Z.); [email protected] (A.T.); [email protected] (B.W.);
[email protected] (J.W.)
* Correspondence: [email protected]; Tel.: +86-134-8846-0799




Received: 10 March 2020; Accepted: 1 April 2020; Published: 3 April 2020


Abstract: Baseline drift spectra are used for quantitative and qualitative analysis, which can easily
lead to inaccurate or even wrong results. Although there are several baseline correction methods
based on penalized least squares, they all have one or more parameters that must be optimized by
users. For this purpose, an automatic baseline correction method based on penalized least squares
is proposed in this paper. The algorithm first linearly expands the ends of the spectrum signal,
and a Gaussian peak is added to the expanded range. Then, the whole spectrum is corrected by
the adaptive smoothness parameter penalized least squares (asPLS) method, that is, by turning the
smoothing parameter λ of asPLS to obtain a different root-mean-square error (RMSE) in the extended
range, the optimal λ is selected with minimal RMSE. Finally, the baseline of the original signal is
well estimated by asPLS with the optimal λ. The paper concludes with the experimental results on
the simulated spectra and measured infrared spectra, demonstrating that the proposed method can
automatically deal with different types of baseline drift.

Keywords: automated baseline correction; infrared spectra; penalized least squares

1. Introduction
Spectroscopic analysis technology such as Raman and infrared spectroscopy has been widely used
in many fields because of its high sensitivity, rapid analysis speed, and nondestructive features [1–5].
However, any spectra measured by a spectrometer usually have a wandering baseline due to
environmental factors, such as light source, temperature, humidity, and vibration during long-time
continuous work [6]. When the baseline drift spectrum is used for quantitative and qualitative analysis,
the prediction accuracy of the model is reduced. Hence, it is necessary to correct the baseline of spectra
before subsequent analysis [7,8].
Recently, researchers have paid much attention to baseline correction and proposed several
methods [9–25]. These baseline correction methods are mainly iterative polynomial fitting [9–11],
derivative method [12,13], wavelet transform method [14–16], iterative averaging [17–19], and penalized
least squares [20–25]. Iterative polynomial fitting is performed by an iterative strategy where, in each
iteration, a polynomial is used to obtain the fitted baseline. The minimum value between baseline
and signal is taken as the new signal, and when the fitted baseline no longer changes or the changes
are minimal, a satisfactory baseline is obtained. However, an iterative polynomial tends to obtain a
boosted baseline in the peak regions and cannot be used in situations where the background changes
drastically. The derivative method amplifies the noise signal during baseline correction. When the
spectrum has high frequency noise, the corrected signal is distorted. Therefore, the spectrum should
be smoothed before applying the derivative method. When the wavelet transform method is used to
correct the baseline of the spectra, the spectra signals are decomposed into low and high frequency

Sensors 2020, 20, 2015; doi:10.3390/s20072015 www.mdpi.com/journal/sensors

Sensors 2020, 20, 2015 2 of 12

signals. Therefore, the baseline signal with low frequency is removed by the wavelet transform and
reconstruction. However, it is difficult to select the optimum wavelet basis, decomposition level,
and threshold of the wavelet coefficients. Iterative averaging tends to overestimate the baseline in
peak regions when the width of the spectral peak area is large. In addition, it is not suitable to fit the
baseline when the spectra are overlapped.
Baseline correction methods based on penalized least squares are widely used in various types of
spectral preprocessing because of its fast speed and its ability to avoid peak detection. The main idea is
to balance the fidelity and smoothness of the fitted baseline. The application of penalized least squares
in the field of baseline correction was first proposed by Eilers, who designed the asymmetric least
squares (AsLS) method [20] to correct the baseline of various spectra in 2003. For the AsLS method,
the weight vector updates iteratively, and if a signal is above the fitted baseline, then a small weight
is assigned to achieve automatic interpolation fitting. On the other hand, a large weight is given
when a signal is below a fitted baseline. Subsequently, researchers have developed several improved
algorithms based on the AsLS algorithm. Zhang et al. considered that the AsLS algorithm has two
turning parameters, λ and p. Moreover, the weight can only be set to p and 1-p, which results in the
same weight of the peak regions as that of the noise regions, so the estimated baseline obtained by
this method is not optimal. In this respect, an adaptive iteratively reweighted penalized least squares
(airPLS) method [21] was proposed that further improves the performance of the baseline correction.
The core idea of airPLS is to assign different weights according to the difference between the signal and
the fitted baseline; in addition, only one parameter needs to be optimized. He and Zhang proposed
an improved asymmetric least squares (IAsLS) [22] baseline correction method. Compared with the
AsLS and airPLS methods, the performance of the fitted baseline is further improved. However, when
a spectrum is corrupted with additive noise, both the airPLS and IAsLS methods tend to obtain an
underestimated baseline. In order to perform baseline correction in noisy signals, an asymmetrically
reweighted penalized least squares smoothing (arPLS) [23] method was proposed by Park, which
uses the iterative method to adaptively obtain weights according to the generalized logic function.
Although the arPLS can obtain a satisfactory baseline in the no-peak regions, a boosted baseline is
obtained in the peak regions. In order to obtain a satisfactory baseline in the whole spectrum region, we
previously proposed an approach called the adaptive smoothness parameter penalized least squares
(asPLS) [24]. However, the above baseline correction methods have one or more tuning parameters,
which limits the application of real-time online data analysis.
According to our experiments, the smoothing parameter λ has a great influence on the performance
of baseline correction, but in all of above methods λ is set randomly by the users. Therefore, a baseline
correction method of extended range based on penalized least squares (erPLS) is proposed in this
paper by selecting the optimal parameter λ of asPLS. The proposed method is compared with the
improved modified multi-polynomial fitting (I-ModPoly), morphological weighted penalized least
squares (MPLS) [25], and iterative averaging (IA) methods in the simulated spectrum and the real
infrared spectrum, and the comparison results confirm that the proposed method can better handle
diverse baseline types automatically.

2. Materials and Methods

2.1. Simulated Data Generation

For the presentation and validation of the proposed method, several mathematical spectra were
generated to simulate real infrared spectra in different practical applications. The simulated spectra
data consisted of an analytical signal and different types of baseline and random noise, which is
described as the following equation:

y(v) = s(v) + b(v) + n(v) (1)

 Sensors 2020, 20, 2015 3 of 12

where y(v) is the simulated data, s(v) denotes the analytical spectral data, b(v) represents the baseline,
and n(v) displays the random noise. The spectra scanned by FTIR are actually a convolution of the
real spectra and the apodization (window) function, so the shape of the spectra is similar to that of
the Gaussian
Sensors function. Therefore, the analytical signal is generated by several Gaussian peaks
2020, 20, x FOR PEER REVIEW
and
3 of 12
mathematically presented as follows:

v−100 2 −( v−100 2 v−200 v−400

2 v−500
2 2 v−800 v−1000
v−1000 22 v−1100 2 v−1200 2
)2 )−( )2−( v−400
−() )2−( v−500
−( ) )2 −(−v−800 ) )2 )2 ) )2 20 )
s(v) = 2es−((v)20= 2
) −( v−200 (
e+ e20 20+e +202e +240e 40+ e +e 30 30 ++4e4e 5050 ++0.5 0.5ee−−(( 1515 )) ++−( −( v−1100
e e 20 20 +1.5
−( −( v−1200
+e1.5e
20 (2)
(2)

whereEquation
where Equation(2) (2)isisthe
the generated
generated analytical
analytical signal,
signal, andand
the the numbers
numbers in thein equation
the equation and even
and even the
the number
number of Gaussian
of Gaussian peakspeaks
can be can be modified
modified according
according to thetoapplication
the application in different
in different situations.
situations. The
The spectra
spectra baseline
baseline is prone
is prone to linear
to linear or curved
or curved drift.drift.
Thus,Thus, we simulated
we simulated linearlinear
and and curved
curved baselines
baselines to
to evaluate the performance of the proposed baseline correction method: (1) a first-order
evaluate the performance of the proposed baseline correction method: (1) a first-order linear baseline linear baseline
and(2)
and (2)aasine
sinebaseline.
baseline.The Thetwotwotypes
typesofofbaselines
baselinesare
aregiven
givenbybythe
thefollowing
followingformula:
formula:

b=1 =−0.01
b1 −0.01+ + 22××10 xx
10−3
−3
(
b = sin(πx/1200) (3)
(3)
 2b2 = sin(π x /1200)
Random noise
Random noise was
wasgenerated
generatedusing MATLAB’s
using MATLAB’sawgn awgn
function. In orderIn
function. to verify
order theto performance
verify the
of the proposed
performance of themethod in different
proposed method noises, two different
in different levelsdifferent
noises, two of noiselevels
were of
generated—the
noise were
signal-to-noise signal-to-noise
generated—the ratio (SNR) of ratio
the high
(SNR)noise spectrum
of the wasspectrum
high noise set to 25 was
dB and that
set to 25ofdBthe low
and noise
that of
spectrum was set to 30 dB. The simulated spectra were obtained by adding the baseline signal,
the low noise spectrum was set to 30 dB. The simulated spectra were obtained by adding the baseline the noise
signal,the
signal, andnoise
the analytical
signal, andsignal, which aresignal,
the analytical shownwhich
in Figure 1. To distinguish
are shown theTosimulated
in Figure 1. distinguishspectra
the
easily, thespectra
simulated high noise spectra
easily, wasnoise
the high moved up bywas
spectra two.moved up by two.
Absorbance

Absorbance

(a) (b)
Simulatedspectra
Figure1.1.Simulated
Figure spectrawith
withlinear
linear and
and sine
sine baseline
baseline types
types inin high
high noise
noise and
and low
low noise:(a)
noise: (a)linear
linear
baseline;(b)
baseline; (b)sine
sinebaseline.
baseline.ToTodistinguish
distinguishthethesimulated
simulatedspectra
spectraeasily,
easily,the
thehigh
highnoise
noisespectra
spectrawas was
movedup
moved upby bytwo.
two.

2.2. Infrared Spectra Acquisition

2.2. Infrared Spectra Acquisition
The mid-infrared spectra were scanned using a Spectrum Two Fourier transform (FTIR)
The mid-infrared spectra were scanned using a Spectrum Two Fourier transform (FTIR)
spectrometer, produced by PerkinElmer (Waltham, MA, USA). The optical path was 10 cm. The spectrum
spectrometer, produced by PerkinElmer (Waltham, MA, USA). The optical path was 10 cm. The
resolution was set to 1 cm−1 and 8 scans were used for each spectrum. The spectral range and apodization
spectrum resolution was set to 1 cm –1 and 8 scans were used for each spectrum. The spectral range
function were set to 400~4000 cm−1 and Norton–Beer medium, respectively. Thus, there were 3601
and apodization function were set to 400~4000 cm–1 and Norton–Beer medium, respectively. Thus,
variables for each spectrum. Methane and n-butane were selected as target analytes, and the absorption
there were 3601 variables for each spectrum. Methane and n-butane were selected as target analytes,
spectra are shown in Figure 2.
and the absorption spectra are shown in Figure 2.
It is obvious from Figure 2 that the baseline drifted more seriously as concentration decreased.
It is obvious from Figure 2 that the baseline drifted more seriously as concentration decreased. This
This is because the absorption signal became smaller when the concentration decreased. Thus, baseline
is because the absorption signal became smaller when the concentration decreased. Thus, baseline drift
drift is prone to occur with a slight change of environmental factors.
is prone to occur with a slight change of environmental factors.
Sensors 2020, 20, 2015 4 of 12
Sensors 2020, 20, x FOR PEER REVIEW 4 of 12

(a)

(b)

Figure 2. Six
Figure 2. different concentrations
Six different concentrationsofofmethane and n-butane
methane and n-butanespectra
spectra scanned
scanned by Fourier
by Fourier transform
transform
infrared spectrometer
infrared spectrometer(FTIR): (a)(a)
(FTIR): methane;
methane;(b)
(b)n-butane.
n-butane.

2.3. The
2.3. Proposed Method:
The Proposed erPLS
Method: erPLS
The The methods
methods basedbased onon penalizedleast
penalized least squares
squares obtain
obtainaasatisfactory
satisfactory baseline
baselineby turning
by turning the the
parameters. For the AsLS method, baseline drift is eliminated by adjusting
parameters. For the AsLS method, baseline drift is eliminated by adjusting the asymmetry p and the asymmetry p and
smoothness
smoothness λ. λ. Forthe
For the airPLS,
airPLS,arPLS,
arPLS,andand
asPLS methods,
asPLS a satisfactory
methods, baseline is obtained
a satisfactory baselinebyischanging
obtained by
the smoothness parameter λ. However, these methods must be optimized by one or two parameters;
changing the smoothness parameter λ. However, these methods must be optimized by one or two
more importantly, these parameters have a great impact on the results, which limits the application of
parameters; more importantly, these parameters have a great impact on the results, which limits the
these methods in the field of real-time online spectra analysis. In order to overcome the shortcomings
application of these
of the above methods
methods, in thecorrection
a baseline field of real-time
method called online spectrarange
extended analysis.
based Inon order to overcome
penalized least
the shortcomings of the above methods, a baseline correction method called
squares (erPLS) is proposed. The core idea of erPLS is to select the optimal parameter λ of the asPLSextended range based on
penalized
method.least squares
Before (erPLS)the
introducing is proposed. The core
proposed method, idea
three of erPLS are
parameters is todefined
select asthefollows:
optimal parameter
selected
λ of the
wavenumber range Ω,Before
asPLS method. introducing
and added Gaussian thepeakproposed
width W method, threeH.parameters
and height We selectedare Ω asdefined
the as
follows: selected wavenumber range Ω, and added Gaussian peak width W and height H. We selected
wavenumber range at the end of the spectral signal used for linear fitting and we used W and H to
generate a Gaussian peak. In general, the length of Ω
Ω as the wavenumber range at the end of the spectral signal used for linear fitting and we used W and
is recommended to be set to one-twentieth of the
spectral length
H to generate N, whereas
a Gaussian peak.WIn is general,
usually assigned
the lengthone-fifth
of Ω is of recommended
N, and H is set equalto be to
setthe
tomaximum
one-twentiethof of
the ordinate values of the spectra signal. A detailed description of the erPLS algorithm is as follows:
the spectral length N, whereas W is usually assigned one-fifth of N, and H is set equal to the maximum
Step 1:values
of the ordinate Linear fit. A first-order
of the polynomial
spectra signal. is used to
A detailed perform a linear
description of thefiterPLS
on thealgorithm
spectral data in the
is as follows:
selected
Step wavenumber
1: Linear fit. A range to obtain
first-order a linear regression
polynomial is used tocoefficient.
perform The fitting
a linear fitequation is given data
on the spectral as in
follows:
the selected wavenumber range to obtain a linear regression coefficient. The fitting equation is given
as follows:
ŷi = ayi + b (i ∈ Ω) (4)
Sensors 2020, 20, x FOR PEER REVIEW 5 of 12
Sensors 2020, 20, x FOR PEER REVIEW 5 of 12
Sensors 2020, 20, 2015 5 of 12
yˆ iˆ=ayi +b (i ∈Ω)
yi =ayi +b (i ∈Ω) (4)
(4)
where
whereaaaand andb are first-order regression coefficients and yˆ and
is theŷi fitted
thevalue
fittedofofthe spectral
of the signal in the
where and bbare
are first-order
first-order regression
regression coefficients
coefficients and yˆ is the i
is
fitted
i
value value
the spectral spectral
signal insignal
the in
the selected wavenumber.
selected wavenumber.
selected wavenumber.
Step2:2:Linear
Step Linear expansion.
expansion. Extended
Extended signalsignal is obtained
ye isyobtained by linear
by linear expansion
expansion of the of theofend
end the of the
Step 2: Linear expansion. Extended signal ye is eobtained by linear expansion of the end of the
spectra
spectra signal
signal according
according to the
to regression
the regressioncoefficient, and the
coefficient, and length
the of ye isofW.
length y e is W.
spectra signal according to the regression coefficient, and the length of ye is W.
Step
Step3:3:3:Signal
Signal addition.
addition. AddAdd a Gaussian
a Gaussian peakpeak
signal yg toythe
signal toextended
the extended signal,signal,
the width of
the width
Step Signal addition. Add a Gaussian peak signal yg togthe extended signal, the width of of
In = ye +yg .
Gaussian is equal
Gaussianisisequal equalto W/2, and the height is H. The new extended signal is marked y , y = y +y . In this
toto W/2, andthethe height
is H.isTheH. new
The extended
new extended signal is marked , gy.eg
yeg, yeg= yyeeg+y
eg eg e g
Gaussian W/2, and height signal is marked this
way,
In theway,
this new thespectral
new signal
spectralynewsignal
is composed
y is of two
composedparts—one
of two is the original
parts—one isspectra
the signal
original y and
spectra thesignal y
way, the new spectral signal ynew is composed new of two parts—one is the original spectra signal y and the
other is the extended signal yeg, ynew= [y, yeg]. The progress of steps 1–3 is shown in Figure 3.
and the
other other
is the is the extended
extended signal yeg,signal
ynew= [y, The=progress
yegy,egy].new [y, yeg ].ofThe
steps progress of steps
1–3 is shown in 1–3
Figureis shown
3. in Figure 3.

Figure 3. The original

The spectrum
original withwith
spectrum linear extension
linear and an
extension added Gaussian peak. peak.
Figure
Figure 3.3.The original spectrum with linear extension andand an added
an added Gaussian
Gaussian peak.

Step
Step4:4:4:Calculate
Calculate thethe
RMSE
RMSE values for for
values the extended range. The extended rangerange
RMSERMSE is denoted
Step Calculate the RMSE values for thethe extended
extended range. range. TheThe extended
extended range RMSE is denoted
is denoted
asasRMSE
RMSE e. The
. Thewhole
whole spectrum
spectrum y y
new is corrected
is corrected by the
by asPLS
the asPLSmethod. RMSE
method. e is calculated
RMSE is by
calculated theby the
as RMSEee. The whole spectrum ynewnew is corrected by the asPLS method. RMSEe is calculated e by the
following equation:
followingequation:
equation:
following v
u
W
Wu
u
2
 [ yˆ ˆ(i ) −[ŷbˆ(eˆ(ii))]−2 2b̂e (i)]
tP
W


i =1
[ y
i =(1i ) − b e ( i )] (5)
e =e== i =1
RMSE
RMSE (5) (5)
RMSE e W W
W
wherebˆb̂eˆe is
where is the
thefitted
fittedbaseline
baselineofof the
the extended
extended range range obtainedobtained by the by asPLS
the asPLS method,
method, ŷi is defined
yˆ is defined in in
where be is the fitted baseline of the extended range obtained by the asPLS method, yˆ is defined i
i
in
step 1, and W is the length of the extended range. For this case,
step 1, and W is the length of the extended range. For this case, different RMSEe valueseare obtained by different RMSE values are obtained
step 1, and W is the length of the extended range. For this case, different RMSEe values are obtained by
by turning
turning λ in
λ in the the asPLS
asPLS method, method,
and are and
shown areinshown
Figure in Figure
4. From the4.figure,
Fromonethecan
figure, onethe
see that can see that the
optimal
turning λ in the 9.6 asPLS method, and are shown in Figure 4. From the figure, one can see that the optimal
λ is 10 9.6because this value obtained the lowest RMSEe.
optimal
9.6 λ is 10 because this value obtained the lowest RMSE e .
λ is 10 because this value obtained the lowest RMSEe.
1.5
1.5 0.02
0.02

0.015
0.015
1 0.01
1 0.01

0.005
0.005

0.5 0
0.5 07 8 9 10
7 8 9 10

0
0 4 8 6 10 12
4 8 6 10 12
log10( )
log10( )
Figure 4. 4.
The relation
The between
relation extended
between range
extended root-mean-square
range error error
root-mean-square (RMSE e) and λ.
(RMSE
Figure
Figure 4. The relation between extended range root-mean-square error (RMSE e )λ.
e) and and λ.

Step 5: Select the optimal λ. The λ with the lowest RMSEe is selected as the optimal
smoothing parameter.
Sensors 2020, 20, x FOR PEER REVIEW 6 of 12
Sensors 2020, 20, 2015 6 of 12
Step 5: Select the optimal λ. The λ with the lowest RMSEe is selected as the optimal smoothing
parameter.
Step 6: Output the final fitted baseline of the original signal y. The baseline of the original signal y
Step 6: Output the final fitted baseline of the original signal y. The baseline of the original signal y
is corrected by the asPLS method with optimal λ.
is corrected by the asPLS method with optimal λ.
2.4. Parameter Setting
2.4. Parameter Setting
In this research, all experiments were performed on MATLAB (R2016b, The Math Works,
In thisMA,
Natick, research,
USA).all
Inexperiments were performed
order to evaluate on MATLAB
the performance of the (R2016b, The Math
above baseline Works, methods,
correction Natick,
MA,
the USA). In order to evaluate
root-mean-square the performance
error (RMSE) was adopted:of the above baseline correction methods, the root-
mean-square error (RMSE) was adopted:
v
u
N
u
2
u
tNP
i = 1
ˆ(ib̂)](2i)]
[b[(bi )(i−) b−
(6)
RMSE== i =1
RMSE (6)
NN

whereb bisisthe
where thesimulated
simulatedbaseline
baselinethat
thatisisdenoted
denotedasasthe
thereal baseline, b̂ is the
realbaseline, the estimated
estimated baseline, and N
baseline, and
Nis
is the length
lengthofofthe theoriginal
original spectra
spectra signal.
signal. As mentioned
As mentioned above,above, the critical
the critical step of step of the proposed
the proposed erPLS
erPLS method
method is to find is to
thefind
optimal λ for λthe
the optimal forasPLS
the asPLS method.
method. In order
In order to observe
to observe howhowthethe smoothness
smoothness
parameter
parameter λλ affectsthe
affects theperformance
performance ofofthe
thearPLS
arPLSand
andasPLS methods,λ λwas
asPLSmethods, wasturned
turnedfrom
from10103 3 12 12
toto1010
with
with varying
varying inin a log
a log scale.
scale. The The experiments
experiments were
were tested
tested inin a simulated
a simulated spectrum
spectrum with
with a sine
a sine baseline
baseline
and
and with
with lowlow noise.
noise. The The RMSE
RMSE values
values obtained
obtained by by
twotwo methods
methods are are
shownshown in Figure
in Figure 5. 5.

1.2

1 arPLS
asPLS

0.8

0.6

0.4

0.2

0
4 6 8 10 12
log10( )
Figure
Figure5.
5.The
TheRMSE
RMSEof
ofthe
theasymmetrically
asymmetrically reweighted
reweighted penalized least squares
squares smoothing
smoothing (arPLS)
(arPLS)and
andadaptive
the the adaptive smoothness
smoothness parameter
parameter penalized
penalized leastleast squares
squares (asPLS)
(asPLS) methods
methods λ. λ.
for different
for different

From
From Figure
Figure 5, 5,one
one cansee
can seethat
thatthetheparameter
parameter λ λhad
hada agreat
greatimpact
impactononthe thearPLS
arPLSand andasPLS
asPLS
6.6 8.3 , while for the arPLS
methods. For the arPLS method, RMSE gets small value when 10 ≤λ≤10 , while for the arPLS
methods. For the arPLS method, RMSE gets a small value when 10 6.6 ≤ λ ≤ 10
8.3
8.4 ≤ λ ≤ 9.9
8.4 ≤ λ
9.9 . For convenience, this paper refers to8.4 8.4 9.9
method,
method, RMSEkeeps
RMSE keepslow lowwhen
when1010 ≤ 10 10. For convenience, this paper refers to 10 10≤ λ ≤ ≤ 10
λ≤ 9.9 10
as
asproper
the the proper
range. range. A careful
A careful looklook at Figure
at Figure 5 reveals
5 reveals thatthat
the the
RMSERMSE valuevalue is greater
is greater than
than 1 when
1 when λ is
λ is
4 , and less than 0.05 when λ is in the proper range. Therefore, we developed a method to
less than 10 , and less than 0.05 when λ is in the proper range. Therefore, we developed a method to
less than 10
4

select
select parameter
parameter λ of
λ ofthethe asPLS
asPLS method.
method.
In the erPLS algorithm, three
In the erPLS algorithm, three parameters—the parameters—the selected
selected wavenumber
wavenumber range
range ΩΩ and
and thetheadded
added
Gaussian peak width W and height H—are important and should
Gaussian peak width W and height H—are important and should first be optimized. Experiments of first be optimized. Experiments
of baseline
sine sine baseline simulated
simulated spectraspectra
with low withnoise
low were
noiseconducted
were conducted
to observe to observe the influence
the influence of the
of the above
above
parameters.parameters.
InIn order
order to to research
research howhowthethe Gaussian
Gaussian peakpeakwidthwidth
W andW and height
height H affectH affect the performance
the performance of theof
the erPLS
erPLS method, method,
the erPLS the erPLS was calculated
was calculated using different
using different values ofvalues
W and of H Wmade
and Hwith made withΩ
a fixed of theΩ
a fixed
of the wavenumber range 1151~1200 cm −1 . The parameter W was changed from 200 to 400 with an
wavenumber range 1151∼1200 cm . The parameter W was changed from 200 to 400 with an interval of
–1

10,interval of parameter
while the 10, while the parameter
H was turned from H was turned
3 to 6 withfrom 3 to 6 with
an interval of 0.2.anTheinterval λ were
of 0.2.
different Thecomputed
different λ
 Sensors 2020, 20, 2015 7 of 12
Sensors 2020, 20, x FOR PEER REVIEW 7 of 12

with
werethecomputed
various pairs
withofthe
W various
and H, and the
pairs ofthree-dimensional
W and H, and thefigure of λ vs. W and
three-dimensional H is given
figure inW
of λ vs. Figure
and H
6. is given in Figure 6.

11
Sensors 2020, 20, x FOR PEER REVIEW 7 of 12
10.5

with the various pairs of W10and H, and the three-dimensional figure of λ vs. W and H is given in Figure
log10( )

6. 9.5

8.5
6

5 400
350
H 4 300
250
log10( )

W
3 200

Figure
Figure 6. The
6. The selection
selection λ between
λ between parameters
parameters WW and
and H the
H in in the extended
extended range
range based
based on on penalized
penalized least
least
squares (erPLS) algorithm.
squares (erPLS) algorithm.

It is obvious from Figure 4 that λ is in the proper range (108.4 ≤ λ ≤ 109.9 ) when W is varied from
It is obvious from Figure 4 that λ is in the proper range (108.4 ≤ λ ≤ 109.9) when W is varied from 200
200 to 400 and H from 3 to 6, which confirms that the erPLS is robust to the parameters W and H.
to 400 and H from 3 to 6, which confirms that the erPLS is robust to the parameters W and H.
To investigate the relationship between the parameter Ω and the erPLS algorithm’s performance,
the following settings were used:5 Ω = 1141~1200 cm−1 , Ω = 1151~1200 cm−1 , and Ω = 1161~1200 cm−1 ,
Simulated spectra
where
Figure W6.and H are set
The selection to 200 and
λ between 5,=1141–1200
respectively.
parameters W
-
The relation between Ω and erPLS algorithm’s
cm and H in the extended range based on penalized least
1
4 -
performance is shown
squares (erPLS) algorithm.in Figure 7. It is clear
=1151–1200 cm from
-
1
Figure 7 that the fitted baselines almost overlap,
which also shows that the3 erPLS is robust to the selected wavenumber range Ω. Through these
1
=1161–1200 cm

It is obviouswe
experiments, from
canFigure
reach4the λ is in the proper
thatconclusion that therange above 8.4 ≤ λparameters
(10three ≤ 109.9) whencan
W is
bevaried from 200
set automatically
Absorbance

to 400 and H from

according to the3spectrum
to 6, which confirms
signal
2 when that the
applying erPLS the is robust
erPLS to
method.the parameters W and H.

-1
Absorbance

0 200 400 600 800 1000 1200

-
1
Wavenumber/cm

Figure 7. Comparison of the fitted baseline using the erPLS method between different selected
ranges.

To investigate the relationship between the parameter Ω and the erPLS algorithm’s performance,
the following settings were used: Ω = 1141∼1200 cm–1, Ω = 1151∼1200 cm–1, and Ω = 1161∼1200 cm–1,
where W and H are set to 200 and 5, respectively. The relation between Ω and erPLS algorithm’s
performance is shown in Figure 7. It is clear from Figure 7 that the fitted baselines almost overlap, which
also Figure
shows 7.that
Figure the erPLSofisthe
7. Comparison
Comparison ofrobust
the
fitted to
fitted the
baselineselected
baseline
using wavenumber
using
thethe erPLS
erPLS range
method
method Ω. Through
between
between these
different
different experiments,
selected
selected ranges.
we3.can reach the conclusion that the above three parameters can be set automatically according to the
ranges.
Results
spectrum signal when applying the erPLS method.
To investigate the relationship between the parameter Ω and the erPLS algorithm’s performance,
3.1. Experimental Simulated Data
the following settings were used: Ω = 1141∼1200 cm–1, Ω = 1151∼1200 cm–1, and Ω = 1161∼1200 cm–1,
3. Results
where W To and
evaluate
H arethe setperformance
to 200 and 5,ofrespectively.
the erPLS method, somebetween
The relation Ω and erPLS
lately proposed baseline correction
algorithm’s
methods,
3.1. namely
Experimental
performance I-ModPoly,
Simulated
is shown in Figure MPLS,
Data
7. and from
It is clear IA, were used
Figure to simulate
7 that thebaselines
the fitted infraredalmost
spectraoverlap,
for a comparison
which
also shows that the erPLS is robust to the selected wavenumber range Ω. Through these experiments,
with erPLS. All of the above four methods have one or more parameters that must be set before
To evaluate
performing the the performance
baseline correction. of the
The erPLS method,
parameters of some are
erPLS lately proposed baseline
automatically set correction
according to the
we can reach the conclusion that the above three parameters can be set automatically according to the
methods,
spectrum namely
signal. I-ModPoly, MPLS, and IA, were used to simulate the infrared spectra for a
spectrum signal whenFor the I-ModPoly
applying the erPLSmethod,
method. the polynomial order was turned from 1 to 10 with
comparison with erPLS. All of the above four methods have one or more parameters that must be set
before
3. performing the baseline correction. The parameters of erPLS are automatically set according to
Results

3.1. Experimental Simulated Data

To evaluate the performance of the erPLS method, some lately proposed baseline correction
Sensors 2020, 20, 2015 8 of 12

increments of 1. The optimal order was determined after a few loops. For the MPLS, the critical
smoothness parameter λ was turned from 103 to 109 with varying in a log scale, the order of difference
penalties was set to 2, and the half window w size of the morphological structuring elements was set
from 10 to 200 with increments of 10. The optimal parameters λ and w were carried out by finding
the minimum value of the RMSE in the MPLS method. For the IA method, we chose the optimum
parameter threshold T according to reference [19]. The RMSE values obtained by the four baseline
correction methods under optimal parameters are listed in Table 1.

Table 1. The RMSE values obtained by the four baseline correction methods.

Baseline Type (Low Noise) Baseline Type (High Noise)

Linear Sine Linear Sine
Methods
Optimal Optimal Optimal Optimal
RMSE RMSE RMSE RMSE
Parameters Parameters Parameters Parameters
I-ModPoly 1 order 5 = 6 0.1043 order = 6 0.0790 order = 5 0.1315 order = 6 0.1164
2 λ 6 = 108 λ = 107.6 λ = 108.8 λ = 107.6
MPLS 0.0131 0.0148 0.0764 0.0312
w 7 = 70 w = 80 w = 60 w = 70
IA 3 T 8 = 10−2.2 0.1128 T = 10−2.2 0.0831 T = 10−2 0.1907 T = 10−2 0.1641
erPLS 4 - 0.0061 - 0.0036 - 0.0098 - 0.0109
1.I-ModPoly—improved modified multi-polynomial fitting method; 2. MPLS—morphological weighted penalized
least squares method; 3. IA—iterative averaging method; 4. erPLS—the proposed method; 5. order—polynomial
order of the I-ModPoly method; 6. λ—smoothness parameter of the MPLS method; 7. w—half window parameter of
the MPLS method; 8. T—threshold parameter of the IA method.

From Table 1, among the four baseline correction methods, the erPLS method obtained the smallest
RMSE values, the MPLS and I-ModPoly methods obtained relatively better RMSE values, while IA
obtained the worst RMSE values. Hence, we can conclude that the erPLS method can obtain an
improved result for linear or curved baselines compared to the I-ModPoly, MPLS, or IA methods,
especially when the spectra are corrupted with high noise. It can also be seen from Table 1 that the
RMSE values obtained in high noise were larger than those obtained in low noise. Therefore, we can
confirm that noise has a great influence on the above baseline correction methods.
Detailed results of the baseline corrected by the four methods are shown in Figure 8. It is obvious
from Figure 8 that the estimated baseline obtained by the erPLS method is more accurate than others.
On the other hand, the I-ModPoly and MPLS methods are relatively worse, while IA obtained the
worst result. The estimated baseline obtained by the erPLS almost overlapped the real baseline in both
the linear and curved baselines with different noise levels. The I-ModPoly method obtained a boosted
baseline, and the reason is that the calculated approximation of the noise level was too high. The fitted
baselines handled by MPLS almost overlapped with the real baseline when the noise was low, while in
high noise, the fitted baseline was lower than the real baseline. The IA method tended to pass through
the lowest point of the spectrum, resulting in different degrees of underfitting in the whole spectrum.
Therefore, it is necessary to perform a smoothing operation before applying the MPLS and IA methods.
From the above experiments, we can conclude that the proposed baseline correction method shows
better performance compared with the other methods.
Sensors 2020, 20, 2015 9 of 12
Sensors 2020,20
Sensors2020, 20, ,xxFOR
FORPEER
PEERREVIEW
REVIEW 99of
of12
12

66

Simulated
Simulatedspectra
spectra
55
I-ModPoly
I-ModPoly
MPLS
MPLS
44 IA
IA
erPLS
erPLS
33 real
realbaseline
baseline

22

11

00

00 200
200 400
400 600
600 800
800 1000
1000 1200
1200
-1
Wavenumber/cm-1
Wavenumber/cm
(a)
(a) (b)
(b)

(c)
(c) (d)
(d)
Figure
Figure8.8.
Figure 8.The
Thefitted
The fitted
fitted baselines
baselines
baselines with
with
with four
four
four methods
methods
methods in
inboth
in high
both
both and
high
high low
andand noise.
lowlow (a)
(a)Linear
noise.
noise. baseline
(a) Linear
Linear in
inlow
baseline
baseline lowin
noise; (b)
(b)linear
low noise;
noise; baseline
(b) linear
linear in
inhigh
baseline
baseline innoise;
high high (c)
noise; (c)sine
noise; baseline
sine(c) in
inlow
lownoise;
sine baseline
baseline in low(d)
noise; sine
sinebaseline
noise;
(d) (d) sine in
baseline inhigh
highnoise.
baseline ItIt
in high
noise.
can
can be
be seen
noise. canthat
Itseen the
be seen
that erPLS
thethat themethod
erPLS isis more
erPLS method
method moreisaccurate compared
more accurate
accurate with
with the
compared
compared the I-ModPoly,
with MPLS,
the I-ModPoly,
I-ModPoly, and
MPLS,
MPLS, IA
andandIA
methods.
IA methods.
methods.

3.2. Experimental Infrared

3.2. Spectra
3.2.Experimental
ExperimentalInfrared
InfraredSpectra
Spectra
To verifythe
To the capability of the proposed erPLS method, different
six different concentrations of methane
Toverify
verify thecapability
capabilityof
ofthe
theproposed
proposederPLS
erPLSmethod,
method,six six differentconcentrations
concentrationsofofmethane
methaneandand
and n-butane spectra
n-butane were used for the experiment. In order to clearly distinguish the performance of
n-butanespectra
spectrawere
wereused
usedfor
forthe
theexperiment.
experiment.InInorder
orderto toclearly
clearlydistinguish
distinguishthe
theperformance
performanceof ofthe
the
the
aboveabove methods, the baseline-corrected spectra of n-butane with a concentration of 50 ppm and
abovemethods,
methods,thethebaseline-corrected
baseline-correctedspectra
spectraof
ofn-butane
n-butanewith withaaconcentration
concentrationof
of50
50ppm
ppmand
and100
100ppm
ppm
100 ppm
obtained obtained by the four methods are shown in Figure 9.
obtainedby
bythe
thefour
fourmethods
methodsareareshown
shownin inFigure
Figure9.9.
0.005
0.005 0.010
0.010

0.004
uncorrected
uncorrected uncorrected
uncorrected
0.004 0.008
I-ModPoly
I-ModPoly 0.008 I-ModPoly
I-ModPoly
MPLS
MPLS MPLS
MPLS
0.003
0.003 IA IA
IA IA
0.006
0.006
erPLS
erPLS erPLS
erPLS
Absorbance
Absorbance

Absorbance
Absorbance

0.002
0.002
0.004
0.004
0.001
0.001

0.002
0.002
0.000
0.000

0.000
0.000
-0.001
-0.001

-0.002
-0.002 -0.002
-0.002
4000
4000 3500
3500 3000
3000 2500
2500 2000
2000 1500
1500 1000
1000 500
500 4000
4000 3500
3500 3000
3000 2500
2500 2000
2000 1500
1500 1000
1000 500
500
-1-1 -1
Wavenumber/cm
Wavenumber/cm Wavenumber/cm
Wavenumber/cm
-1

(a)
(a) (b)
(b)
Figure9.9.
Figure
Figure Baseline-corrected
9.Baseline-corrected
Baseline-corrected n-butane
n-butane
n-butane spectra
spectra
spectra byby
by four
four
four methods:
methods:
methods: (a)(a)
(a) thethe
the concentration
concentration
concentrationof of n-butane
ofn-butane
n-butaneisis50is
50
50
ppm;ppm;
ppm;(b)
(b)the(b) the concentration
theconcentration
concentrationof of n-butane
ofn-butane is
n-butaneisis100 100
100ppm.
ppm.ppm.
the infrared spectrum of n-butane, the performance of these methods can be evaluated by the degree of
deviation from zero in the non-absorption region. As seen in Figure 9a, the MPLS and IA methods
obtained a boosted corrected spectrum in the middle spectrum region (1600–2200 cm–1), while the I-
ModPoly method overestimated the baseline in the left half of the non-absorption region (3000–3500
cm –1), resulting in a declining corrected spectrum in the left spectrum region. From Figure 9b, it can be
Sensors 2020, 20, 2015 10 of 12
seen that the MPLS and IA methods obtained a boosted corrected spectrum in the right spectrum region
(600–1000 cm–1), while the I-ModPoly method obtained an underestimated baseline-corrected spectrum
According
in the left half of theto Lambert’s
non-absorptionlaw, the ideal
region absorption
(3000–3500 cmis–1);zero
this in non-absorption
phenomenon bands.
is more obviousTherefore,
when
for the infrared spectrum of n-butane, the performance of these
the concentration of n-butane is 50 ppm. The corrected spectrum performed by the erPLS method methods can be evaluated by the
degreeoverlapped
almost of deviation from
zero zeronon-absorption
in the in the non-absorption region. As
region, whether theseen in Figure 9a,
concentration the MPLSwas
of n-butane and50IA
methods −1 ),
ppm or 100obtained
ppm, which a boosted corrected
proves that spectrum
the baselines in the by
corrected middle
the erPLSspectrum
methodsregion (1600–2200
are more precise cm
than
while
those ofthe
theI-ModPoly
others. method overestimated the baseline in the left half of the non-absorption region
(3000–3500 −1
cm ),concentrations
resulting in a declining corrected
Six different of methane and spectrum
n-butane in the leftusing
spectra spectrum
the region. From Figure
erPLS method were9b,
it can be seen
performed that the
to show theMPLS and IA
capability ofmethods obtained
the proposed a boosted
method, and corrected spectrum
the corrected in the
spectra are right spectrum
displayed in
region10.(600–1000 −1
cmwith ), while
Figure Compared Figurethe
2, itI-ModPoly
is evident method obtained anbaseline-corrected
that the absorption underestimated baseline-corrected
spectra are close
tospectrum
zero in thein non-absorption
the left half of the non-absorption
region, and look quite region (3000–3500
similar cm−1 ); this
in the absorption phenomenon
regions, is more
which further
obvious when the concentration of n-butane is 50 ppm.
proves that the proposed method can successfully eliminate baseline drift. The corrected spectrum performed by the
erPLS method
Finally, it is almost overlappedthat
worth mentioning zerotheinoptimal
the non-absorption
parameters cannot region,bewhether
selected the concentration
by carrying out theof
n-butane
minimal was 50
RMSE ppmbecause
value, or 100 ppm,
we dowhich proves
not know thethat
realthe baselines
baseline. corrected we
Fortunately, by the
knowerPLS
the methods
absorption are
more precise than those of the others.
in the non-absorption region. Therefore, for the I-ModPoly, MPLS, and IA methods, we chose the
Six different
appropriate parametersconcentrations
according oftomethane and n-butaneFor
many experiments. spectra using the erPLS
the proposed erPLS method,
method were the
performedΩ,
parameters toW,show
andthe capability
H were proven of tothehave
proposed method,
little effect on theand the corrected
results, which canspectra
be setare displayed in
automatically
Figure 10.
according Compared
to the spectrum with Figure
signal. 2, it isin
As shown evident
Figure that thedrifting
10, the absorptionspectrabaseline-corrected
were successfullyspectra
corrected are
byclose to zeromethod;
the erPLS in the non-absorption
thus, we can conclude region,that
andthe look quite
erPLS can similar in the absorption
automatically regions,
handle various which
types of
furtherdrift.
baseline proves that the proposed method can successfully eliminate baseline drift.

(a) (b)
Figure
Figure10. Baseline-corrected
10.Baseline-corrected methane
methane and
and n-butane
n-butane spectra
spectra bybythe
theerPLS
erPLSmethod:
method:(a)(a)the
thecorrected
corrected
spectraof
spectra of methane; (b)
(b)the
thecorrected spectra
corrected of n-butane.
spectra It is shown
of n-butane. that the erPLS
It is shown method
that the erPLSsuccessfully
method
eliminates baseline
successfully drift.
eliminates baseline drift.

Finally, it is worth mentioning that the optimal parameters cannot be selected by carrying out
4. Conclusions
the minimal RMSE value, because we do not know the real baseline. Fortunately, we know the
In this research, an automated baseline correction method based on penalized least squares was
absorption in the non-absorption region. Therefore, for the I-ModPoly, MPLS, and IA methods,
proposed and the performance was evaluated by simulation and infrared spectra. The core idea of
we chose the appropriate parameters according to many experiments. For the proposed erPLS
erPLS was to add a Gaussian peak in the extended range combined with the asPLS method to perform
method, the parameters Ω, W, and H were proven to have little effect on the results, which can be
baseline correction automatically.
set automatically according to the spectrum signal. As shown in Figure 10, the drifting spectra were
For the simulated spectra analysis, the RMSE of the erPLS was smallest among the four baseline
successfully corrected by the erPLS method; thus, we can conclude that the erPLS can automatically
correction methods, and the estimated baseline almost overlapped the real baseline, whereas the other
handle various types of baseline drift.
methods tended to deviate from the real baseline. The experimental results of the two different
4. Conclusions
In this research, an automated baseline correction method based on penalized least squares was
proposed and the performance was evaluated by simulation and infrared spectra. The core idea of
erPLS was to add a Gaussian peak in the extended range combined with the asPLS method to perform
baseline correction automatically.
Sensors 2020, 20, 2015 11 of 12

For the simulated spectra analysis, the RMSE of the erPLS was smallest among the four baseline
correction methods, and the estimated baseline almost overlapped the real baseline, whereas the
other methods tended to deviate from the real baseline. The experimental results of the two different
concentrations of n-butane infrared spectra proved that the erPLS method is more precise than the
I-ModPoly, MPLS, and IA methods. Moreover, the six different concentrations of methane and n-butane
spectra performed using the erPLS method also confirmed that the proposed method is more effective
in dealing with different types of baselines. More importantly, the method can automatically correct
the baseline without users having to set the parameters. These consistent results show that the erPLS
method can correct various types of baselines automatically and can improve the accuracy of baseline
estimation. Therefore, we hope that erPLS can be applied to unsupervised online analysis of FTIR
systems in the future.

Author Contributions: Conceptualization, F.Z. and X.T.; methodology, F.Z.; software, F.Z.; validation, A.X., X.T.,
and B.W.; formal analysis, F.Z.; investigation, A.T.; resources, J.W.; data curation, X.T.; writing—original draft
preparation, F.Z.; writing—review and editing, X.T.; visualization, F.Z.; supervision, X.T.; project administration,
B.W. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by the National Key Research and Development Program of China, grant
number 2016YFF0102805.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Feng, L.; Zhu, S.; Chen, S.; Bao, Y.; He, Y. Combining Fourier Transform Mid-Infrared Spectroscopy with
Chemometric Methods to Detect Adulterations in Milk Powder. Sensors 2019, 19, 2934. [CrossRef]
2. Wen, S.F.; Xu, Y.Z. Fourier Transform Infrared Spectroscopy, 3rd ed.; Chemical Industry Press: Beijing, China,
2016; pp. 451–489.
3. Stuart, B. Infrared Spectroscopy: Fundamentals and Applications, 1st ed.; John Wiley and Sons Ltd.: Chichester,
UK, 2004; pp. 137–185.
4. Tang, X.J.; Li, Y.J.; Zhu, L.J.; Zhao, A.X.; Liu, J.H. On-line multi-component alkane mixture quantitative
analysis using Fourier transform infrared spectrometer. Chemom. Intell. Lab. Syst. 2015, 146, 371–377.
[CrossRef]
5. Usseglio, S.; Thorshaug, K.; Karlsson, A.; Dahl, I.M.; Nielsen, C.J.; Jens, K.J.; Tangstad, E. In situ infrared
emission spectroscopy for quantitative gas-phase measurement under high temperature reaction conditions:
An analytical method for methane by means of an innovative small-volume flowing cell. Appl. Spectrosc.
2010, 64, 141–148. [CrossRef] [PubMed]
6. Zhao, A.X.; Tang, X.J.; Li, W.D.; Zhang, Z.H.; Liu, J.H. The Piecewise Two Points Auto-Linear Correlated
Correction Method for Fourier Transform Infrared Baseline Wander. Spectrosc. Lett. 2015, 48, 274–279.
[CrossRef]
7. Liu, J.; Sun, J.; Huang, X.; Li, G.; Liu, B. Goldindec: A Novel Algorithm for Raman Spectrum Baseline
Correction. Appl. Spectrosc. 2015, 69, 834–842. [CrossRef]
8. Xu, D.; Liu, S.; Cai, Y.; Yang, C. Baseline correction method based on doubly reweighted penalized least
squares. Appl. Opt. 2019, 58, 1225–1232. [CrossRef]
9. Zhao, J.; Lui, H.; Mclean, D.L.; Zeng, H. Automated Autofluorescence Background Subtraction Algorithm
for Biomedical Raman Spectroscopy. Appl. Spectrosc. 2007, 61, 834–842. [CrossRef]
10. Krishna, H.; Majumder, S.K.; Gupta, P.K. Range-independent background subtraction algorithm for recovery
of Raman spectra of biological tissue. Appl. Spectrosc. 2012, 43, 1884–1894. [CrossRef]
11. Wang, T.; Dai, L. Background Subtraction of Raman Spectra Based on Iterative Polynomial Smoothing.
Appl. Spectrosc. 2016, 71, 1169–1179. [CrossRef]
12. Leger, M.N.; Ryder, A.G. Comparison of Derivative Preprocessing and automated polynomial baseline
correction method for classification and quantification of narcotics in solid mixtures. Appl. Spectrosc. 2006,
60, 182–193. [CrossRef]
13. Baek, S.; Park, A.; Kim, J.; Shen, A.; Hu, J. A simple background elimination method for Raman spectra.
Chemom. Intell. Lab. Syst. 2009, 98, 24–30. [CrossRef]
Sensors 2020, 20, 2015 12 of 12

14. Feng, J.; Tian, F.; Jia, Y.; He, Q.; Yue, S.; Pan, L. A background elimination method based on wavelet transform
in wound infection detection by electronic nose. Sens. Actuators B Chem. 2011, 157, 395–400. [CrossRef]
15. Bertinetto, C.G.; Vuorinen, T. Automatic Baseline Recognition for the Correction of Large Sets of Spectra
Using Continuous Wavelet Transform and Iterative Fitting. Appl. Spectrosc. 2014, 68, 155–164. [CrossRef]
[PubMed]
16. Qian, F.; Wu, Y.; Hao, P. A fully automated algorithm of baseline correction based on wavelet feature points
and segment interpolation. Opt. Laser Technol. 2017, 96, 202–207. [CrossRef]
17. Prakash, B.D.; Wei, Y.C. A fully automated iterative moving averaging (AIMA) technique for baseline
correction. Analyst 2011, 136, 3130–3135. [CrossRef]
18. Shen, X.C.; Ye, S.B.; Xu, L.; Hu, R.; Jin, L.; Xu, H.Y.; Liu, J.G.; Liu, W.Q. Study on baseline correction methods
for the Fourier transform infrared spectra with different signal-to-noise ratios. Appl. Opt. 2018, 57, 5794–5799.
[CrossRef] [PubMed]
19. Shen, X.; Xu, L.; Ye, S.; Hu, R.; Jin, L.; Xu, H.; Liu, W. Automatic baseline correction method for the open-path
Fourier transform infrared spectra by using simple iterative averaging. Opt. Express 2018, 26, 609–614.
[CrossRef]
20. Eilers, P.H.C. A Perfect Smoother. Anal. Chem. 2003, 75, 3631–3636. [CrossRef]
21. Zhang, Z.M.; Chen, S.; Liang, Y.Z. Baseline correction using adaptive iteratively reweighted penalized least
squares. Analyst 2010, 135, 1138–1146. [CrossRef]
22. He, S.; Zhang, W.; Liu, L.; Huang, Y.; He, J.; Xie, W.; Wu, P.; Du, C. Baseline correction for Raman spectra
using an improved asymmetric least squares method. Anal. Methods 2014, 6, 4402–4407. [CrossRef]
23. Baek, S.J.; Park, A.; Ahn, Y.J.; Choo, J. Baseline correction using asymmetrically reweighted penalized least
squares smoothing. Analyst 2015, 140, 250–257. [CrossRef]
24. Zhang, F.; Tang, X.; Tong, A.; Wang, B.; Wang, J.; Lv, Y.; Tang, C.; Wang, J. Baseline correction for infrared
spectra using adaptive smoothness parameter penalized least squares method. Spectrosc. Lett. 2020, 53,
222–233. [CrossRef]
25. Li, Z.; Zhan, D.J.; Wang, J.J.; Huang, J.; Xu, Q.S.; Zhang, Z.M.; Zheng, Y.B.; Liang, Y.Z.; Wang, H. Morphological
weighted penalized least squares for background correction. Analyst 2013, 138, 4483–4492. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).