Analytical

Methods
View Article Online
PAPER View Journal | View Issue

Determination of pesticides in a flour substrate by

Published on 04 October 2018. Downloaded by Iowa State University on 1/21/2019 12:29:28 AM.

Cite this: Anal. Methods, 2018, 10,

chemometric methods using terahertz
5097 spectroscopy
Binghua Cao,a Hui Li, *a Mengbao Fan,b Wei Wanga and Mengyun Wanga

Terahertz time-domain spectroscopy (THz-TDS) is utilized as an effective tool for quantitative analysis of
imidacloprid and carbendazim in a flour substrate. The partial least squares (PLS), principal component
analysis (PCA), support vector machine (SVM) and PCA-SVM methods were used to construct linear and
nonlinear regression models to correlate absorption spectra and concentrations of 21 samples based on
the whole absorption spectra (0.2–1.4 THz). The multiple spectra baseline correction (MSBC) method
which is based on asymmetric least squares smoothing was adopted to correct the slope baselines. The
algorithm can eliminate scatter effects on the spectra and improve the signal-to-noise ratio of the THz
spectra. The models were optimized by cross-validation, and their performances were evaluated with
respect to root mean square error of prediction (RMSEP), correlation coefficient in the prediction set (Rp)
and correlation coefficient in the calibration set (Rc). The results show that PLS delivers the best
Received 3rd August 2018
Accepted 4th October 2018
performance with the lowest errors. The optimized PLS models for both imidacloprid and carbendazim
are obtained with RMSEP ¼ 0.5439%, Rp ¼ 0.9992 and Rc ¼ 0.9999. Our experimental results also
DOI: 10.1039/c8ay01728j
demonstrate that THz-TDS combined with chemometrics can be used for quantitative determination of
rsc.li/methods pesticides in agricultural products.

In general, chromatography is used for pesticide detection.

1. Introduction However, it is labor intensive and time consuming and cannot
Pesticides play an important role in modern agriculture due to offer efficient pesticide detection.3–5 Spectroscopy technique has
their ability to increase crop yields. However, pesticides have attracted increasing attention owing to its simplicity, rapidity,
a series of adverse effects on food safety, ecology and environ- and almost non-destruction to samples. In recent years,
ment.1 Environmental pollution and health problems caused by infrared spectroscopy and Raman spectroscopy for pesticide
pesticide residues have attracted a great deal of attention in detection have been reported.6–8 However, infrared spectroscopy
recent years. Imidacloprid is a kind of neonicotinoid insecticide and Raman spectroscopy are subject to the scattering effect and
whose chemical name is 1-[(6-chloro-3-pyridinyl) methyl]-N- thermal radiation, and then the accuracy tends to degrade.
nitro-2-imidazolidinimine. It can be used for crops such as Terahertz spectroscopy has attracted increasing attention as
wheat, rice, corn, cotton, vegetables and fruit trees. Neverthe- a new method developed in recent decades. Compared with
less, the toxicity of imidacloprid may be lethal to non-target infrared spectroscopy, the terahertz time-domain spectroscopy
animals. The chemical name of carbendazim is methyl N-(1H- (THz-TDS) technique is insensitive to ambient thermal radia-
benzimidazol-2-yl)carbamate, which is a kind of fungicide tion, and the light source is stable. Moreover, the scattering
widely used in China. Carbendazim can effectively prevent and effect is weaker due to its long wavelength. In addition, the
control a variety of crop diseases caused by fungi, but its residue THz-TDS technique uses coherent measurement technology
not only interferes with the biochemical and physiological and can provide the amplitude and phase information of the
reactions of plant pathogens, but also affects the population or electric eld simultaneously, which can give the absorption
activity of other non-target microorganisms in soil.2 Thus, a fast coefficient and refractive index spectra or complex permittivity
and reliable method for quantitative analysis of imidacloprid spectrum of a sample without using Kramers–Kronig relation-
and carbendazim is badly needed. ships. Therefore, THz-TDS has been widely applied in many
elds like chemistry, biology, medical science, industry and
security monitoring to the qualitative and quantitative analysis
a
of materials.9–15
School of Information and Control Engineering, China University of Mining and
In the eld of pesticide detection, Ma et al. used THz-TDS to
Technology, Xuzhou 221116, China. E-mail: lihui_swi@cumt.edu.cn
b
School of Mechatronic Engineering, China University of Mining and Technology,
identify different mixtures of thiabendazole and polyethylene
Xuzhou 221116, China with different mixing ratios.16 Four models, partial least squares

This journal is © The Royal Society of Chemistry 2018 Anal. Methods, 2018, 10, 5097–5104 | 5097
 View Article Online

Analytical Methods Paper

(PLS), interval PLS (iPLS), moving window PLS (mwPLS) and by Coherent. The femtosecond laser pulse has an output power
backward interval (biPLS), were employed to predict the of 960 mW, a center wavelength of 800 nm, a width of 56 fs and
concentration of thiabendazole and the performances were a repetition frequency of 80 MHz. Aer passing through the l/2
compared. Wang and Ma applied THz-TDS to detect mixtures of plate, the generated femtosecond laser pulse is divided into
nitrofen and polyethylene with different weight ratios.17 PLS, a pump beam and a probe beam by a beam splitter, and then
iPLS, mwPLS and biPLS algorithms were used to quantitatively the two beams are used for the generation and detection of the
analyze nitrofen. The results revealed that the terahertz spectra terahertz beam. The pump beam is focused on a GaAs photo-
of THz-TDS combined with biPLS algorithm gave the best conductor antenna via a delay system driven by a computer-
results with the minimum detection limit of 2.52%. Generally, controlled stepper motor to generate terahertz beam, which is
Published on 04 October 2018. Downloaded by Iowa State University on 1/21/2019 12:29:28 AM.

polyethylene (PE) is used as a matrix for sample preparation due focused on the sample by a pair of parabolic mirrors. The probe
to its low absorption in the THz spectrum. Taking practical beam, which passes through the ITO glass and the terahertz
measurement into account, other matrices like rice and our beam, which carries the sample information aer penetrating
serve as substrates instead. Hua and Zhang used THz-TDS the sample, are confocal on the ZnTe electro-optic crystal. The
combined with the PLS method to realize quantitative anal- terahertz beam makes the ZnTe crystal birefringent. The
ysis of imidacloprid in glutinous rice our mixtures.18 Chen polarization direction of the probe beam will be changed when
et al. reported THz-TDS combined with chemometrics for it passes through the ZnTe crystal. It will become elliptically
quantitative analysis of imidacloprid in a rice matrix.19 Qin et al. polarized aer the modulation of the l/4 plate, and will become
utilized PCA combined with clustering by fast search and nd of two polarization components of unequal size aer passing
density peaks (PCA-CFSFDP) in order to analyze mixtures of through a Wollaston prism. Therefore, the output current of
carbendazim and tomato powder.20 The results showed that a differential photodiode is nonzero and the magnitude of the
carbendazim was successfully identied using THz-TDS differential current is proportional to the intensity of a terahertz
combined with the PCA-CFSFDP method. However, these light. The amplied differential current by a lock-in amplier is
studies blended one pesticide ingredient with a food matrix. In led into a computer. In order to lessen the inuence of water
order to protect crops from the damage of fungi and insect vapor in air on the THz-TDS system, the optical path where the
pests, two or more kinds of pesticides are mixed and used at terahertz electromagnetic radiation passes through is enclosed
different stages of crop growth, which leads to various pesticide in a nitrogen gas box. The humidity is maintained at a constant
residues. Therefore, it is crucial to conduct research on the value of 4% during the experiment.
determination of mixtures of various pesticides and food
matrices.
In our work, simultaneous detection of two pesticides 2.2. Sample preparation
mixed with a our matrix is studied. First, imidacloprid and Imidacloprid and carbendazim with a purity of 97% were
carbendazim were mixed with a our substrate as the samples. purchased from the Jiangsu Changqing Agrochemical Company
Then, the absorption spectra of the mixture of imidacloprid Ltd and Jiangsu Liangmancang Agrochemical Company Ltd,
and carbendazim in the our matrix were obtained by THz- Jiangdu City, Jiangsu, China, respectively. All samples were
TDS. In addition, the slope baseline of the absorption spectra dried at 313 K for 1 h to remove water.
was corrected by utilizing a multiple spectra baseline correc- In this work, imidacloprid and carbendazim for spectral
tion (MSBC) method for elimination of scatter effects and analysis were mixed with polyethylene at a weight ratio of 1 : 1,
improvement of the signal-to-noise ratio. By using PLS, PCA, respectively. Moreover, the two pesticides were mixed with our
support vector machine (SVM) and PCA-SVM methods, linear instead of polyethylene. Twenty-one mixtures containing imi-
and nonlinear regression models were established to relate dacloprid, carbendazim and our were prepared. The mass
absorption spectra to the concentration of the two pesticides fraction of our was xed at 50%, and those of imidacloprid and
based on the whole absorption spectra (0.2–1.4 THz). carbendazim varied from 0% to 50%, with 2.5% interval. All the
Furthermore, the performances of the models were evaluated mixtures were ground into powders and then pressed using
by the root mean square error of prediction (RMSEP), correla- a hydraulic press at a pressure of 20 MPa for one minute. The
tion coefficient in the prediction set (Rp) and correlation pressed sheet was disc-shaped with a diameter of 13 mm and
coefficient in the calibration set (Rc). This paper reveals that thickness of 1–2 mm, which can be measured using an elec-
THz-TDS combined with chemometrics can be used for the tronic bench micrometer. The samples were then sealed in
determination of various pesticide components mixed with labelled bags.
food matrices.

2.3. Data acquisition

2. Experimental methods
All the measurements are performed using the THz-TDS system
2.1. Experimental setups illustrated in Fig. 1. Time domain signals of the reference Eref(t)
Fig. 1 shows the THz-TDS system used in the experiment. The and sample Esam(t) are recorded. The frequency spectroscopy
THz-TDS system consists of a Model Z-1 terahertz system can be obtained by using a FFT algorithm.
developed by Zomega Corporation of the United States and Based on the Fresnel formulae, the THz amplitude trans-
a Vitesse-800-5 titanium sapphire femtosecond laser produced mittance T for most low-loss materials can be expressed as

5098 | Anal. Methods, 2018, 10, 5097–5104 This journal is © The Royal Society of Chemistry 2018
 View Article Online

Paper Analytical Methods

Published on 04 October 2018. Downloaded by Iowa State University on 1/21/2019 12:29:28 AM.

Fig. 1 Schematic of the THz-TDS experimental system.


Esam ðuÞ 4n iuðN  1Þd 1 Xm
TðuÞ ¼ ¼ A expði4Þ z exp the weight parameter and gk ¼ ak(2  ak), q ¼ ðyv  zv Þ.
Eref ðuÞ ð1 þ nÞ2 c m v¼1
(1) The parameters l ¼ 800, mk ¼ 105 and p ¼ 0.01 are determined
based on the literature.21
where Eref(u) and Esam(u) are the incident and transmitted THz
amplitude spectra, respectively; A and 4 are the amplitude ratio
2.5. Modeling methods and evaluation
and phase difference of reference and sample signals, respec-
tively; N ¼ n + ik is the complex refractive index of a sample and Four methods (PCA, PLS, SVM and PCA-SVM) are used to
k is the extinction coefficient; d is the sample thickness; u is the analyze the concentrations of samples, in which PCA and PLS
angular frequency and c is the speed of light in vacuum. Then are linear methods and SVM and PCA-SVM are nonlinear
refractive index n(u) and absorption coefficient a(u) can be methods. It is noteworthy that the models for the two kinds of
obtained from eqn (1). pesticides are separately developed.
The THz signals were collected three times for each sample,
4ðuÞc
nðuÞ ¼ þ1 (2) and then were averaged to reduce noises. Twenty-one samples
ud
were divided into calibration and prediction sets randomly. The
2kðuÞu 2 4nðuÞ former (16 samples) was used to build a regression model, and
aðuÞ ¼ ¼ ln (3) the latter (5 samples) was used to evaluate the regression model.
c d AðuÞðnðuÞ þ 1Þ2
PCA is a multivariate statistical method.22,23 The essence of
PCA is to reduce the dimension of data and it is widely used in
2.4. Multiple spectra baseline correction method data analysis. The original variables are linearly combined into
new variables, where new variables retain as much information
The multiple spectra baseline correction method based on
as possible of the original variables and are independent.
asymmetric least squares smoothing can be used to correct the
PLS is one of the commonly used multiple linear regression
slope baseline.21 Assume spectrum yk is the column vector of
methods.24 PLS extracts orthogonal features from spectral data
length l, and zk is the corresponding baseline, where k ¼ 1, 2,
and establishes correlations between features and target vari-
3.m. The baselines zk, relaxation factors ak and weight
ables. In this paper, PLS is used to establish the relationship
matrices Qk are dened as
between absorption coefficient and concentration for quanti-
 1
zk ¼ ðm  gk ÞE þ mlQk þ mmk DT D tative analysis of pesticide samples.
" # SVM is a supervised learning method.25–27 Based on support
Xm
 ðm  gk Þyk  gk ðyi  zi Þ þ mlQk yk (4) vector regression theory, a support vector machine prediction
i¼1;isk model is established. Determining the penalty factor c and the
kernel parameter g is the key to the SVM model. Cross-
ak ¼ (qTq)1qT(yk  zk) (5) validation method is used to select c and g in the SVM model.
8 The ow chart about the SVM regression prediction model is
< p; if yk . zk ; shown in Fig. 2.
Qk ¼ 1  p; if yk # zk ; (6)
: PCA-SVM combines PCA and SVM to establish a model.25
0; otherwise:
First, PCA is used to extract features of the THz absorption
where E is the l  l identity matrix, D is the second order spectra. Then, the extracted features serve as the input of the
differential matrix, l and mk are regularization parameters, p is SVM model, and the relationship between the extracted features

This journal is © The Royal Society of Chemistry 2018 Anal. Methods, 2018, 10, 5097–5104 | 5099
 View Article Online

Analytical Methods Paper

Fig. 2 Flow chart of the SVM regression prediction model.

and the concentration is established to perform quantitative corresponding spectrum shows absorption peaks of imidaclo-
Published on 04 October 2018. Downloaded by Iowa State University on 1/21/2019 12:29:28 AM.

analysis. prid, which are about 0.89 THz, 1.13 THz, 1.22 THz and
The cross-validation method is used to select the best 1.30 THz corresponding to Fig. 3(a). When the mass fraction of
parameters for the model. The models were optimized with
respect to the root mean square error by cross-validation
(RMSECV) using the calibration set. The performances of the
developed models were evaluated from RMSEP, Rp and Rc.
RMSE and R are dened as follows:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uX n
 p 2
u
u yi  yri
t
i¼1
RMSE ¼ (7)
n

X
n
 
yri  yr ðypi  yp Þ
R ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
i¼1
ffisffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (8)
X n
 r 2 X n
2
yi  yr ðypi  yp Þ
i¼1 i¼1

Here, ypi is the predicted concentration of the i-th sample,

is the nominal concentration of the i-th sample, i ¼ 1, 2, 3/
yri
21. yr and yp represent the average of all sample references and
prediction values, respectively.

3. Results and discussion

3.1. THz spectra Fig. 3 Absorption spectra of (a) imidacloprid and (b) carbendazim.

The absorption coefficients of samples can be calculated using

eqn (2) and (3). Fig. 3 shows the absorption spectra of imida-
cloprid and carbendazim from 0.2 to 1.8 THz. It can be seen that
imidacloprid and carbendazim have different spectral charac-
teristics, including the number and position of absorption
peaks. The absorption peak positions of imidacloprid are 0.89
THz, 1.13 THz, 1.29 THz and 1.52 THz, and those of carben-
dazim are 1.16 THz and 1.35 THz in the region of 0.2–1.8 THz,
which can be observed in Fig. 3(a) and (b), respectively.
Fig. 4 shows the absorption spectra of the mixtures of
imidacloprid, carbendazim and our from 0.2 THz to 1.4 THz.
The arrows in Fig. 4 indicate the increase of the mass fraction of
imidacloprid, and the numbers represent the frequency posi-
tions of the absorption peaks. The second and fourth absorp-
tion peaks of the mixture are marked with two numbers, and the
number from the bottom to the top with a slanting arrow
indicates the gradual change of the absorption peak.
From Fig. 4, it can be seen that the mixture has four distinct
Fig. 4 Absorption spectra of several samples. (The mass fraction of
absorption peaks at 0.89 THz, 1.13 THz, 1.22 THz and 1.30 THz,
flour is fixed at 50%, and those of imidacloprid (MFIMI) and carbendazim
which are the main features to determine the concentration of vary from 0% to 50%, with 2.5% interval. The arrows indicate the
the samples. In addition, when the mass fraction of imidaclo- increase of the MFIMI, and the numbers represent the frequency
prid is 50% and the mass fraction of carbendazim is 0%, the positions of the absorption peaks.)

5100 | Anal. Methods, 2018, 10, 5097–5104 This journal is © The Royal Society of Chemistry 2018
 View Article Online

Paper Analytical Methods

imidacloprid is 0% and the mass fraction of carbendazim is imidacloprid and carbendazim. Thus, in order to eliminate the
50%, the absorption spectrum of the mixture exhibits the scattering effect of the spectrum and improve the signal-to-
absorption peaks of carbendazim at positions around 1.16 THz noise ratio of the terahertz spectrum, a multiple spectra base-
and 1.33 THz corresponding to Fig. 3(b). With the increase of line correction (MSBC) method is performed on the absorption
the mass fraction of imidacloprid, the positions of the second spectrum of the sample.
and fourth absorption peaks of the mixture gradually shi from Fig. 5 shows the absorption spectra aer MSBC. Comparing
the absorption peaks of carbendazim to those of imidacloprid. with Fig. 4, the slope baselines are removed.
Further analysis indicates that the baselines of the absorp-
tion spectra of the mixtures increase approximately linearly 3.2. Quantitative analysis results
Published on 04 October 2018. Downloaded by Iowa State University on 1/21/2019 12:29:28 AM.

with increasing frequency due to the effect of our absorption.

For the PCA model, the cumulative contribution rate of the rst
Furthermore, the absorption spectra in the high frequency
three principal components of the absorption spectra is 98%,
region are unstable. The scattering effect may result in a lower
and the PCA model delivers a smaller RMSECV using the rst
signal-to-noise ratio in the high frequency region. In general,
three principal components. Therefore, the rst three principal
the absorption spectra of pesticide mixtures vary with the mass
components are selected to determine the concentration.
fraction of the constituent pesticides.
To avoid under-tting and over-tting, the number of PLS
Flour is a kind of powder made from wheat. Different from
factors based on the principle of choosing a small RMSECV and
polyethylene, our absorbs terahertz radiation. The absorption
a small number of factors is selected. As shown in Fig. 6, we
of the our matrix is related to the frequency, causing sloped
selected 1 to 20 factors and calculated the RMSECV of the
baselines. Therefore, when our is used as the matrix, it has an
samples in turn. As a result, 16 factors were chosen to establish
effect on the analyte signal and the background. From Fig. 4, it
the PLS model of both imidacloprid and carbendazim.
can be seen that the absorption band of the our may change
The cross-validation method is used to select the optimal
the amplitude and shi the positions of the absorption peaks of
penalty factor c and the kernel parameter g in the SVM algo-
rithm. The SVM parameter optimization is shown in Fig. 7. The
performances of the SVM model of imidacloprid are optimal
with c ¼ 2.82843 and g ¼ 0.00390625, while those of carben-
dazim are optimal with c ¼ 2 and g ¼ 0.00390625.
The cross-validation method is also used to determine the
best parameters for the PCA-SVM model. The rst three prin-
cipal components were extracted from the absorption spectra as
the input to the PCA-SVM model. The performances of the PCA-
SVM model of imidacloprid are optimal with c ¼ 0.707107 and
g ¼ 1, while those of carbendazim are optimal with c ¼ 45.2548
and g ¼ 0.5.
The prediction results with the four models were obtained
aer parameter optimization. The scatter plots of the predicted
concentration versus the experimental concentration for the
four models are shown in Fig. 8.
As shown in Fig. 8, quantitative analysis of imidacloprid,
carbendazim and our mixtures was performed using PCA, PLS,
Fig. 5 Absorption spectra of several samples after MSBC. (The mass
SVM and PCA-SVM models. Compared with the other three
fraction of flour is fixed at 50%, and those of imidacloprid (MFIMI) and models, the concentration point distribution predicted by the
carbendazim vary from 0% to 50%, with 2.5% interval.) PLS model is closer to the reference line.

Fig. 6 The relationship between the number of factors and RMSECV of (a) imidacloprid and (b) carbendazim.

This journal is © The Royal Society of Chemistry 2018 Anal. Methods, 2018, 10, 5097–5104 | 5101
 View Article Online

Analytical Methods Paper

Published on 04 October 2018. Downloaded by Iowa State University on 1/21/2019 12:29:28 AM.

Fig. 7 3D charts of parameter selection results of (a) imidacloprid and (b) carbendazim.

Fig. 8 Scatter plots of predicted concentration versus experimental concentration for PCA, PLS, SVM and PCA-SVM models of (a) imidacloprid
and (b) carbendazim.

In order to compare the performances of the four models absorption spectra as input to the PCA-SVM model. The
developed, the predicted errors were calculated, as shown in Table performances of the PCA-SVM model of imidacloprid are
1. It can be seen that for the two pesticide components, PLS yields optimal with c ¼ 1.41421 and g ¼ 1, while those of carbendazim
the best prediction results with the highest Rp, Rc and the lowest are optimal with c ¼ 22.6274 and g ¼ 0.25.
RMSEP, while the other three models show unsatisfactory results. The THz spectra in the frequency range (0.2–1.4 THz) are
Aer baseline correction with MSBC, the PCA, PLS, SVM and selected for quantitative analysis. The scatter plots of the pre-
PCA-SVM models were re-established. For the PCA model, only dicted concentration versus the experimental concentration for
the rst principal component of the absorption spectra was the four models are shown in Fig. 9. Compared with Fig. 8, the
selected. Two factors were used to establish the PLS models of concentration point distributions predicted by PCA and SVM
both imidacloprid and carbendazim. models aer baseline correction are closer to the reference line.
The performances of the SVM models of both imidacloprid The RMSECV, Rc, RMSEP and Rp of the four models for the
and carbendazim are optimal with c ¼ 4 and g ¼ 0.00390625. absorption spectra with and without MSBC pretreatment are
The rst principal component was extracted from the listed in Table 1.

Table 1 The results obtained by PCA, PLS, SVM and PCA-SVM models

Imidacloprid Carbendazim
Baseline correction
Model method RMSECV (%) Rc RMSEP (%) Rp RMSECV (%) Rc RMSEP (%) Rp

PCA — 6.8431 0.9683 6.5698 0.9868 7.4124 0.9699 6.2500 0.9975

PLS — 2.3018 0.9999 0.5439 0.9992 2.3172 0.9999 0.5439 0.9992
SVM — 0.4628 0.9961 3.5806 0.9922 0.3197 0.9961 3.5804 0.9921
PCA-SVM — 0.5502 0.9774 1.7450 0.9901 0.4843 0.9756 1.7863 0.9907
PCA MSBC 7.5302 0.9865 5.9558 0.9981 7.9028 0.9865 5.9558 0.9981
PLS MSBC 1.8633 0.9999 1.0618 0.9972 1.7889 0.9999 1.0618 0.9972
SVM MSBC 0.3600 0.9993 1.6954 0.9978 0.5180 0.9993 1.6878 0.9978
PCA-SVM MSBC 0.4610 0.9755 3.1725 0.9771 0.5581 0.9740 3.3728 0.9850

5102 | Anal. Methods, 2018, 10, 5097–5104 This journal is © The Royal Society of Chemistry 2018
 View Article Online

Paper Analytical Methods

Published on 04 October 2018. Downloaded by Iowa State University on 1/21/2019 12:29:28 AM.

Fig. 9 Scatter plots of predicted concentration versus experimental concentration for PCA, PLS, SVM and PCA-SVM models of (a) imidacloprid
and (b) carbendazim.

Fig. 10 Error graphs of PLS of (a) imidacloprid and (b) carbendazim without MSBC.

From Table 1, it can be seen that the performances of the 4. Conclusions

PCA and SVM models are improved aer the baseline correc-
tion, whereas the performance of the PCA-SVM model degrades THz-TDS is used to quantitatively analyze imidacloprid and
slightly aer the baseline correction. In addition, PLS delivers carbendazim in a our substrate. The absorption spectra of
the best prediction results with the highest Rp, Rc and the lowest imidacloprid and carbendazim can be distinguished in the
RMSEP without the treatment of baseline correction. Moreover, frequency range of 0.2–1.4 THz. Multiple spectra baseline
the errors between the predicted concentration and experi- correction was used to eliminate the effects of the slope back-
mental concentration for PLS without baseline correction are ground and can improve the prediction accuracy in the ter-
presented in Fig. 10. Most of the errors fall within 2%. ahertz spectrum. Furthermore, the concentrations of the two
The detection errors are mainly due to experimental errors pesticides were predicted with the help of chemometric
and random errors. Above all, the experiment has uncertainty, methods. PLS, PCA, SVM and PCA-SVM were used for quanti-
which includes the system and background noises. Then, unlike tative analysis of sample concentration. The experimental
PE, our absorbs terahertz radiation, which will inuence the results show that PLS yields optimal results. The optimized PLS
analyte signal. Moreover, the absorption spectra of mixed models of both imidacloprid and carbendazim determination
pesticides are the combination of those of imidacloprid and are obtained with RMSEP ¼ 0.5439%, Rp ¼ 0.9992 and Rc ¼
carbendazim, contributing to the peak shi of each one. This 0.9999. In this work, the feasibility of quantitative determina-
may induce the change of prediction errors. In summary, this tion of pesticides in a our substrate is demonstrated by using
work indicates the feasibility of simultaneous measurement of THz-TDS combined with chemometrics, which suggests that
two or more pesticides in our. In practical applications, THz-TDS could work as a tool to determine the content of two or
however, the content of pesticides in our is very small. The more pesticides in mixtures. To go further, the reported
performance of the experimental system and the accuracy of the approaches can be extended to predict the concentrations of
models need to be improved, such as improving spectral reso- various components in mixtures of other substances.
lution and the sensitivity of the THz system, and searching for
the optimal frequency range for the established models. THz-
TDS is believed to be greatly enhanced with the development Conflicts of interest
of terahertz related technology including terahertz emitters and
detectors, thereby will be more suitable for practical analysis. There are no conicts to declare.

This journal is © The Royal Society of Chemistry 2018 Anal. Methods, 2018, 10, 5097–5104 | 5103
 View Article Online

Analytical Methods Paper

12 S. Nakajima, H. Hoshina, M. Yamashita, C. Otani and

Acknowledgements
N. Miyoshi, Appl. Phys. Lett., 2007, 90, 2549.
This work is supported by the Fundamental Research Funds for 13 B. H. Cao, M. B. Fan and S. Y. Jing, Spectrosc. Spectral Anal.,
the Central Universities under grant 2018QNA41. 2010, 30, 1748–1751.
14 M. Fan, B. Cao and G. Tian, J. Sens., 2017, 3520967.
15 D. Jiang and J. Shen, Conference on Terahertz Photonics,
References Beijing, China, November, 2007.
16 Y. Ma, Q. Wang and L. Li, J. Quant. Spectrosc. Radiat.
1 K. Müller, A. Tiktak, T. J. Dijkman, S. Green and B. Clothier, Transfer, 2013, 117, 7–14.
Published on 04 October 2018. Downloaded by Iowa State University on 1/21/2019 12:29:28 AM.

Encyclopedia of Agriculture & Food Systems, 2014, pp. 17–34. 17 Q. Wang and Y. H. Ma, Chemom. Intell. Lab. Syst., 2013, 127,
2 C. M. Tu, J. Environ. Sci. Health, Part B, 1993, 28, 67–80. 43–48.
3 R. C. Martinez, E. R. Gonzalo, M. A. Moran and J. H. Mendez, 18 Y. Hua and H. Zhang, IEEE Trans. Microwave Theory Tech.,
J. Chromatogr., 1992, 607, 37–45. 2010, 58, 2064–2070.
4 J. M. Bonmatin, I. Moineau, R. Charvet, C. Fleche, M. E. Colin 19 Z. Chen, Z. Zhang, R. Zhu, Y. Xiang, Y. Yang and
and E. R. Bengsch, Anal. Chem., 2003, 75, 2027–2033. P. B. Harrington, J. Quant. Spectrosc. Radiat. Transfer, 2015,
5 T. Nakajima, Y. Tsuruoka, M. Kanda, H. Hayashi, 167, 1–9.
T. Hashimoto, Y. Matsushima, S. Yoshikawa, C. Nagano, 20 B. Qin, Z. Li, Z. Luo, Y. Li and H. Zhang, Opt. Quantum
Y. Okutomi and I. Takano, J. Chromatogr. B: Anal. Technol. Electron., 2017, 49, 244.
Biomed. Life Sci., 2014, 32, 1099–1104. 21 J. Peng, S. Peng, A. Jiang, J. Wei, C. Li and J. Tan, Anal. Chim.
6 S. Armenta, G. Quintas, S. Garrigues and M. De la Guardia, Acta, 2010, 683, 63.
Trends Anal. Chem., 2005, 24, 772–781. 22 A. D. Burnett, W. Fan, P. C. Upadhya, J. E. Cunningham,
7 S. Armenta, S. Garrigues and M. de la Guardia, Anal. Bioanal. M. D. Hargreaves, T. Munshi, H. G. M. Edwards,
Electrochem., 2007, 387, 2887–2894. E. H. Lineld and A. G. Davies, Analyst, 2009, 134, 1658.
8 M. Khanmohammadi, S. Armenta, S. Garrigues and M. de la 23 S. Roweis, Adv. Neural Inf. Process Syst., 1997, 10, 626–632.
Guardia, Vib. Spectrosc., 2008, 46, 82–88. 24 H. Abdi, Encyclopedia of Measurement & Statistics, 2003, 6,
9 T. W. Crowe, T. Globus, D. L. Woolard and J. L. Hesler, Philos. 792–795.
Trans. R. Soc., A, 2004, 362, 365–374. 25 H. Ge, Y. Jiang, F. Lian, Y. Zhang and S. Xia, Food Chem.,
10 D. Dragoman and M. Dragoman, Prog. Quantum Electron., 2016, 209, 286–292.
2004, 28, 1–66. 26 M. E. Mavroforakis and S. Theodoridis, IEEE Transactions on
11 S. A. Zvyagin, M. Ozerov, E. Čižmá, D. Kamenskyi, Neural Networks, 2006, 17, 671–682.
S. Zherlitsyn, T. Herrmannsdörfer, J. Wosnitza, R. Wünsch 27 C. Cortes and V. Vapnik, Mach. Learn., 1995, 20, 273–297.
and W. Seidel, Rev. Sci. Instrum., 2009, 80, 300.

5104 | Anal. Methods, 2018, 10, 5097–5104 This journal is © The Royal Society of Chemistry 2018