Content uploaded by Yue Huang
Author content
All content in this area was uploaded by Yue Huang on Aug 24, 2021
Content may be subject to copyright.
Vol.:(0123456789)
1 3
Environmental Chemistry Letters
https://doi.org/10.1007/s10311-021-01240-9
ORIGINAL PAPER
Convolutional neural network withnear‑infrared spectroscopy
forplastic discrimination
JingjingXia1· YueHuang2· QianqianLi3· YanmeiXiong1· ShungengMin1
Received: 15 February 2021 / Accepted: 13 April 2021
© The Author(s), under exclusive licence to Springer Nature Switzerland AG 2021
Abstract
Plastic pollution is a global issue of increasing health concern, thus requiring innovative waste management. In particular,
there is a need for advanced methods to identify and classify the different types of plastics. Near-infrared spectroscopy is
currently operational in some waste-sorting facilities, yet remains challenging to discriminate different black plastics because
black targets have low reflectance in some spectral regions. Here we used partial least squares discrimination analysis, soft
independent modeling of class analogy, linear discriminant analysis and convolutional neural network to classify the plastics.
We analyzed 159 plastic samples, including 84 black plastics, made of high impact polystyrene, acrylonitrile butadiene sty-
rene, high-density polyethylene, polyethylene terephthalate, polyamide 66, polycarbonate and polypropylene. Results show
that the convolutional neural network model yielded an accuracy up to 98%, whereas other models displayed accuracy of
57–70%. Overall, convolutional neural network analysis of infrared plastic data is promising to solve the bottleneck problem
of black plastic discrimination.
Keywords Plastic solid waste· Identification model· CNN· NIR
Abbreviations
ABS Acrylonitrile butadiene styrene
CNN Convolutional neural network
HDPE High-density polyethylene
HIPS High-impact polystyrene
LDA Linear discriminant analysis
NIR Near-infrared spectroscopy
PA66 Polyamide 66
PC Polycarbonate
PET Polyethylene terephthalate
PLS-DA Partial least squares discrimination analysis
PP Polypropylene
SIMCA Soft independent modeling of class analogy
Introduction
Plastics are widely used in manufacturing and daily applica-
tions; global plastic production increased from 254 million
tons to 359 million tons between 2008 and 2018, with an
expected threefold increase by 2050 (Othman etal. 2021).
One of the main concerns associated with the extensive
applications of plastics comes from the large amount of
generated waste. Plastic waste embodies 11% of the total
waste, where about 31% was disposed in landfills (Alassali
etal. 2018). Much of this plastic waste is mismanaged, lead-
ing to its dispersal in the environment (Wang etal. 2019;
Zhang etal. 2014). In order to contribute to the optimiza-
tion of plastic waste management, an integrated strategy of
waste minimization, reuse and recovery should be consid-
ered, since the processes of recycling and recovery enable
waste to become a resource (Worrell and Reuter 2014). As
we all know, the prior step in waste removal is always waste
identification (Padervand etal. 2021; Padervand etal. 2020a,
b). Plastic mixing generally leads to decreased mechanical
properties as most of them are incompatible. For example,
* Yanmei Xiong
xiongym@cau.edu.cn
* Shungeng Min
minsg@263.net
1 College ofScience, China Agricultural University,
Beijing100193, People’sRepublicofChina
2 College ofFood Science andNutritional Engineering,
China Agricultural University, Beijing100193,
People’sRepublicofChina
3 School ofChinese Material Medica, Beijing
University ofChinese Medicine, Beijing100029,
People’sRepublicofChina
Environmental Chemistry Letters
1 3
almost each plastic has personal melting point, heating sev-
eral plastics with various melting points will cause unsure
changes, or even degrade, especially between crystalline
and amorphous plastics (Nanda and Berruti 2021). Thus,
the main technological obstacle to recycling is an efficient
and accurate plastic sorting scheme (Signoret etal. 2019).
There are several techniques have been used for the fast
separation and identification of plastic wastes. The study of
Mohsen Padervand reviewed the microplastic occurrence,
transport, raw polymers and additives, toxicity and methods
of removal (Padervand etal. 2020). Zhang proposed a novel
magnetic separation method, which is based on particles of
different densities that are driven by the magnetic force and
finally landed in different collection regions (Zhang etal.
2019). In the study of La Marca, the fluid dynamic con-
ditions were investigated by an image analysis technique,
separating plastic waste under different geometric configura-
tion (Marca etal. 2012). Shen investigated the floatability
of seven plastics in the presence of nonionic surfactant, but
was likely to distinguish acrylonitrile butadiene styrene from
polystyrene (Shen etal. 2002). For most cases, the processes
were slow (Marca etal. 2012; Li etal. 2015; Zhang etal.
2019), needed to smash samples before examination (Li
etal. 2015), did not require high-purity polymer streams
for varieties of plastics (Shen etal. 2002), even existed the
danger posed by inefficient chemical waste management
(Zulkifley etal. 2014).
Infrared spectroscopy is an essential method for the auto-
matic sorting of polymers as it is a non-destructive and fast
detection technique (Zhu etal. 2019). Especially NIR has
been used extensively to sort plastic automatically because
of its non-destructive remote high-speed measurements, high
penetration depth of the NIR radiation and high signal-to-
noise ratio (Macho and Larrechi 2002).
The middle infrared spectrum (MIR) using photon up-
conversion technique was applied to sort black polymer parts
in the research of Becker (Becker etal. 2017). Kassouf tested
the potential of combing MIR with independent component
analysis to rapidly discriminate the plastics packaging mate-
rials (Kassouf etal. 2014). Rani demonstrated that a minia-
turized handheld NIR can be used to successful fingerprint
and classify the different plastic polymers (Rani etal. 2019).
NIR, MIR and differential scanning calorimetry (DSC), as
quantitative methods, were used to obtained polymers from
recycled mixed plastic waste (Camacho and Karlsson 2001).
Zhu presented an identification system of plastic solid waste
based on NIR using support vector machine (SVM), and the
model’s accuracy could reach to 97.50% (Zhu etal. 2019).
Zheng proposed a discrimination model using NIR hyper-
spectral imaging system and yielded the accuracy of 100%
(Zheng etal. 2018). Regretfully, NIR spectroscopy is lim-
ited to samples with dark color that are difficult to identify
(Rozenstein etal. 2017). Carbon black strongly absorbs NIR
rays, consequently making NIR inappropriate for sorting
these plastics (Froelich etal. 2007). Hence in the most stud-
ies, samples were usually transparent or in a bright color but
excluded black (Zheng etal. 2018; Zhu etal. 2019). Gener-
ally, one solution is to improve the instrument performance
as the development of technology, the other is to build a
more reliable and robust model.
Convolutional neural network (CNN) has been the popu-
lar solution for different machine learning tasks, including
object detection, image classification and natural language
processing. It is common to employ the CNN to classify
the waste. In order to analyze the possibilities for automatic
waste sorting, the CNN was used for image classification,
and the best validation accuracy was 87.80% (Gyawali etal.
2020). The study of Zheng and Gu (2021) proposed a novel
ensemble learning model to classify the household solid
waste via waste images. The results demonstrate the effec-
tiveness of the proposed strategy in terms of its accuracy and
F1-scores. The research used a CNN-based algorithm which
deployed a high-resolution camera to capture waste image
and sensors to detect feature information. After training and
validating, the classification accuracy could achieve above
90.00% (Chu etal. 2018). All the above researches focused
on the images; however, it is still available to explore the
possibility which classify the plastic waste using the CNN
combine the spectral data.
In recent years, a few examples on implementation of
CNN for spectroscopic analysis started to appear, which
attracted a lot of attention due to CNN excellent advantages.
(1) compared with classical methods, CNN is less dependent
on preprocessing than the standard modeling methods for
vibrational spectroscopy data and achieves excellent per-
formance (Cui and Fearn 2018). (2) CNN exploits spatially
local correlation by enforcing a local connectivity pattern
between neurons of adjacent layers and in most cases, has
fewer parameters than traditional neural network (Acquarelli
etal. 2017). (3) The study of Ng (Ng etal. 2019) compared
with PLSR and Cubist models; CNN could not only give a
higher accuracy, but also maintain output correlation.
Therefore, in this study, a total of 159 plastic samples
were selected, among which the black plastics accounted for
half more (84/159), including seven kinds of plastic: high-
impact polystyrene (HIPS), acrylonitrile butadiene styrene
(ABS), high-density polyethylene (HDPE), polyethylene
terephthalate (PET), polyamide 66 (PA66), polycarbonate
(PC) and polypropylene (PP). We used of NIR spectra as
one-dimensional data to feed CNN and then built model to
identify the plastic after optimizing network parameters.
Meanwhile, in order to verify the performance of CNN
model, traditional algorithms were involved as a comparison.
Environmental Chemistry Letters
1 3
Experimental
Samples andanalytical methods
Sample set
The plastic samples were provided by China Nanchang Cus-
toms. All of samples were detected by standard methods and
confirmed the compounds. The number of plastic samples
with more than 10 were chosen in terms of the statistical
theory, including high-impact polystyrene (HIPS), acryloni-
trile butadiene styrene (ABS), high-density polyethylene
(HDPE), polyethylene terephthalate (PET), polyamide 66
(PA66), polycarbonate (PC) and polypropylene (PP). Out
of 159 plastic samples, 84 were black, 48 were transparent
or white, and 27 were other colors of plastic (TableS1).
The shapes for all samples included are cylindrical (about
1mm in diameter, about 4mm in height), oblate (about
3mm in diameter) and sphere (about 3mm diameter).
For each sample, the size of plastic particles is same and
well-proportioned.
Each of the sample was scanned five times after being
reshuffled and compacted. Thus, there were a total of 795
spectra. The data set was split into a training set and a test set
by systematic sampling, and the distance was four. Hence,
596 spectra as training set were used to build the model. The
other 199 spectra (test set) were used as unknown samples
to verify the accuracy of the model.
Near‑infrared spectroscopy analysis
Both background and samples were measured by MicroNIR
Pro ES 1700 (VIAVI Solutions Inc., USA), at a data interval
of 7nm. The spectral range covered the NIR window, from
950nm till 1650nm, and the measurements were obtained
in absorbance units. The NIR was controlled by a computer
equipped with MicroNIR v1.5.7 software.
The free classification_toolbox_4.0 was applied with
Matlab R2019a to develop principal component analysis
(PCA), partial least squares discrimination analysis (PLS-
DA), soft independent modeling of class analogy (SIMCA)
and linear discriminant analysis (LDA) models. Convolu-
tional neural network (CNN) model was built on Tensorflow
(Intel (R) Core (TM) i5-8265U, version 2.0.0).
Algorithms
Principal component analysis
PCA can be used to reveal the hidden structure within
large data sets. It provides a visual representation of the
relationships between samples and variables and provides
insights into how measured variables cause some samples
to be similar or how they differ from each other. PCA is
also known as a projection method, because it takes infor-
mation carried by the original variables and projects them
onto a smaller number of latent variables called princi-
pal components (PCs) (Fuentes-García etal. 2018). Each
principal component (PC) explains a certain amount of
the total information contained in the original data, and
the first principal component contains the greatest source
of information in the data set. Each subsequent principal
component contains, in order, less information than the
previous one. In representation, the PCA model with a
given number of components has the following Eq.(1):
where X is predictors matrix, T is the scores matrix, P the
loadings matrix, and E the error matrix.
Soft independent modeling ofclass analogy
SIMCA is known as a supervised pattern recognition
method, identifying different classes of samples based on
the value that provided by known samples (training set).
Unknown samples (test set) are then compared to the class
models and assigned to classes according to their prox-
imity to the training set. SIMCA requires building one
PCA model for each class which describes the structure
of that class as well as possible. The optimal number of
principal components should be chosen for each model
separately, according to a suitable validation procedure
(Vanden Branden and Hubert 2005). In this study, based
on cross-validation, the optimal number of principal com-
ponents is determined for each class.
Partial least squares‑discriminate analysis
PLS-DA is a classification method derived from PLS
regression, incorporating dimension reduction by combin-
ing predictors to generate latent variables which maximally
correlate with the targeted outcomes. PLS-DA implements
dependent variable Y by utilizing independent variables X,
providing the ability to construct a multidimensional model
for the prediction of the features. The basis of PLS-DA is
to reduce the size of original data X and replace it with the
matrix of scores and loadings maximizing the covariance
between X and Y. The level of reduction is described by the
number of significant latent variable; the optimal number
of latent variables is also decided by cross-validation in the
data analysis (Bevilacqua and Marini 2014).
(1)
X=TPT+E
Environmental Chemistry Letters
1 3
Linear discriminant analysis
LDA is the simplest of all possible classification methods
that are based on Bayes’ formula, providing a linear transfor-
mation of n-dimensional feature vectors (or samples) into an
m-dimensional space (m < n), so that samples belonging to
the same class are close together, but samples from different
classes are far apart from each other (Gerhardt etal. 2019).
LDA is a supervised classification method, as the catego-
ries to which objects are to be classified is known before
the model is created. The objective of LDA is to determine
the best fit parameters for classification of samples by a
developed model. The model can then be used to classify
unknown samples.
Convolutional neural network
A typical CNN model would divide layers into 3 sets:
input, hidden and output layers. Among them, hidden lay-
ers include convolution, pooling and fully connected (FC)
layers. The representation of the CNN architecture using
spectral input is included in Figure S1. The input layer is
the first layer, and it generally has a liner activation function.
The output layer is the last layer, and it usually has the num-
ber of neurons, which is equal to the number of class. In our
case, the initial input was spectral variables, and final output
was the prediction of the target classes. Rectified linear unit
(ReLU) was employed in the all layers because the network
was able to converge faster compared with sigmoid or tanh
functions (Ng etal. 2019).
Model evaluation
To evaluate the predictive ability of the models, the accuracy
and loss function were used.
Accuracy is the ratio of correctly assigned samples:
where
ngg
is the number of samples belonging to class g
and correctly assigned to class g. And n represents the total
number of samples.
The loss function adopted in this research is the Cross-
entropy loss function.
where ̂
yn
=𝜑
(
w⋅x
n)
,
𝜑(
⋅
)
is the active function,
xn
is the
n-th sample, w is output weights, and
yn
is the target label.
Parameter optimization
Principal component analysis
We used principal component analysis (PCA) to observe
the distribution of samples. Figure1a shows a two-dimen-
sional score diagram. The abscissa was PC1 which contained
69.00% of data information, and the ordinate was PC2 which
included 18.00%. Figure1b gives the three-dimensional
score diagram with PC3 contained 5.00% of information.
Different kinds of plastic samples were grouped together,
(2)
Accuracy
=
∑
G
g=1ngg
n
(3)
Loss
=−
1
N
N
∑
n=1
[ynlog ̂yn+(1−yn)log(1−̂yn
)]
Fig. 1 Two-dimensional diagram of scores (a); three-dimensional
diagram of scores (b). This figure shows the score plots of principal
component analysis (PCA) using seven kinds of plastic samples. The
PC, PP, HIPS, ABS, HDPE, PET and PA66 in the legend represented
the polycarbonate (PC), polypropylene (PP), high-impact polystyrene
(HIPS), acrylonitrile butadiene styrene (ABS), high-density polyeth-
ylene (HDPE), polyethylene terephthalate (PET) and polyamide 66
(PA66), respectively
Environmental Chemistry Letters
1 3
it and was impossible to classify all kinds of samples only
by PCA.
Spectral pre‑processing
Spectral pre-processing is used to correct sample variations
and noisy spectra. For CNN model, the required pre-pro-
cessing is normalization (Acquarelli etal. 2017; Cui and
Fearn 2018). For example, each variable has a mean of zero
and standard deviation of 1 along the column axis (Cui and
Fearn 2018). However, for PLS-DA, SIMCA and LDA, there
is no standard rule on choosing proper preprocessing meth-
ods (Gerretzen etal. 2015). Savitzky–Golay (S–G) smooth-
ing and standard normal variate (SNV) transformation are
commonly applied in preprocessing methods. Thus, in this
study, we obtained the best preprocessing combinations (the
one achieving the highest tenfold cross-validation accuracy
on the test set was selected) of the methods from the above-
mentioned. Illustrations for the pre-processing spectra were
included TableS2.
Optimization ofmodels
For PLS-DA, SIMCA and LDA, usually, the optimal princi-
pal component (PC) needs to be confirmed by tenfold cross-
validation. According to Figure S2 (a) (b), the abscissa was
the number of principal component (PC), and the ordinate
was error rate. The optimal principal component (PC) for
PLS-DA, LDA was 10, 8, and the corresponding error rate
was 0.41, 0.45, respectively. PCA model for each class was
constructed independently for SIMCA calculation. Training
set was used to determine the optimal principal component
(PC) by cross-validation; Figure S2 (c) shows the result in
SIMCA, and Figure S2 (d) presents the optimal principal
components (PCs) as 3, 6, 3, 7, 3, 6 and 3 for polycarbonate
(PC), polypropylene (PP), high-impact polystyrene (HIPS),
acrylonitrile butadiene styrene (ABS), high-density poly-
ethylene (HDPE), polyethylene terephthalate (PET) and
polyamide 66 (PA66).
In most cases, CNN has fewer parameters than traditional
neural network and with embedded regularization and drop-
out techniques, it is more robust to overfitting (Acquarelli
etal. 2017). For each CNN model, there are several param-
eters need to be considered.
Figure2a: The structure of model (Layers). Different net-
work structures directly impact the accuracy and reliability
of result and the speed of model. Therefore, different-layer
convolutions were trained and compared. Figure2a shows
two-layer, three-layer and four-layer convolutions. When epoch
was 3500, the accuracy was 80.00%, and the loss was 0.50
of test set in two-layer convolution. However, the accuracy
was 90.00%, and the loss near 0.25 in three-layer convolu-
tion, demonstrating that the structure of three layers not only
could converge faster, but also the result was more reliable.
Compared with three-layer convolution, although the accu-
racy could reach 0.90 and the loss was below 0.25 at 2000,
more glitches were showed in four-layer convolution. Glitch
is normal for loss if the model includes shuffle. However, we
tend a steady model when the results are near. In summary, the
model with three-layer convolution was chosen in this study.
Figure2b: Learning rate (lr). Learning rate determines
when the objective function can converge the minimum.
Learning rate is too small, and the update speed is slow,
learning rate is too large, the speed is quick; however, it
may skip the optimal value. In this research, learning rates
(0.0005, 0.0001 and 0.00005) were considered when the
model structure was same. As can be seen from Fig.2b,
when lr was 0.0005, the speed of converge was very fast (the
accuracy was close to 100.00% and the loss was near 0.00),
but the glitch was obvious. When lr was 0.0001, the loss
began to rise after epoch 4000. For lr was 0.00005, although
the speed was slow, the result was acceptable. The accuracy
could reach 98.00% and the loss was 0.12. Thus, 0.00005
was choose as learning rate value.
Figure2c: Kernel size. Kernel is a weight matrix used
for features detection. When epoch was 5000, we could see
from Fig.2c, kernel size set 2 and the model was steady, and
the highest accuracy of model reached 85.00%. But when
kernel size set 3 and 4, the accuracy began to increase as
the epoch increased. And the best accuracy could acquire
98.00%. Compared with kernel size 4, it was obviously seen
that the loss value was more stable when the kernel size was
3. Thus, 3 was choose as a value of kernel size.
Figure2d: Strides. Stride is the number of kernels shifts
over the input data. Figure2d gave the comparison chart of
different strides. The accuracies were approximate values
whatever the stride set be 1, 2 or 3. However, compared
with stride 2 and 3, when the stride was 1, the loss was more
stable. So, stride set 1 was chosen in this study.
CNN was implemented when used the spectra as one-
dimensional input. This network consisted of three convo-
lutional layers. The output of the convolutional layer was
passed to the pooling layer. Aimed on the structure of model,
the convolution layer contained 32 filters with a kernel size
of 3, stride as 1 and zero paddings (‘Same’ way). The num-
ber of filters at each subsequent convolutional layer was
increased by a factor of two, while the other parameters
were kept the same. The detailed architecture of the model
is summarized in TableS3.
Results anddiscussion
The main contribution of this work was the identification
of the black plastics by a novel model, thus overcoming
some of the detection limitation currently experienced in
Environmental Chemistry Letters
1 3
Fig. 2 Adjusting the network parameters: a with different layers convolutions; b under different learning rates; c with different kernel sizes; d under different strides
Environmental Chemistry Letters
1 3
the near-infrared (NIR) region to identify the plastics. NIR
spectra of seven kinds of plastics, including high-impact
polystyrene (HIPS), acrylonitrile butadiene styrene (ABS),
high-density polyethylene (HDPE), polyethylene terephtha-
late (PET), polyamide 66 (PA66), polycarbonate (PC) and
polypropylene (PP), was chosen as one-dimensional data
to feed convolutional neural network (CNN). Additionally,
several traditional algorithms including partial least squares-
discriminate analysis (PLS-DA), soft independent modeling
of class analogy (SIMCA) and linear discriminant analysis
(LDA) were also employed as a comparison.
Accuracies of prediction of seven kinds of plastic using
conventional PLS-DA, SIMCA and LDA techniques were
evaluated, respectively. The results obtained using the vari-
ous preprocessing were summarized in TableS2. The result-
ing preprocessing strategy for PLS-DA was Savitzky–Golay
(S-G) smoothing with the second-order polynomial, fol-
lowed by standard normal variate (SNV). SIMCA with S-G
smoothing with a first-order polynomial, followed by stand-
ard normal variate (SNV), could bring in good result. And
for LDA, standard normal variate (SNV) was chosen as pre-
processing. Aimed on CNN, only normalization was enough.
The spectra (199) in test set were not participate in
modeling, but as unknown samples to predict the perfor-
mance of model. The confusion matrixes of test set were
used to show the results of PLS-DA, SIMCA, LDA and
CNN models. In confusion matrix, the larger the number,
the darker of the grid, and the correctly classified samples
were on the diagonal while the misclassified samples fell
into the other areas. Figure3a gives the confusion matrix
of PLS-DA; the result was poor, and almost a half samples
were mistaken. Obviously, the SIMCA result was a lit-
tle bit better than PLS-DA from the comparison between
Fig.3a, b. LDA concluded the similar result (Fig.3c) with
PLS-DA. When all the samples were involved to build
the CNN model, only four samples misclassified based
on the confusion matrix (Fig.3d). The results obtained
using the various methods are summarized in Table1.
The accuracies of PLS-DA and LDA were close, about
57.00%. SIMCA gave a high accuracy relatively, which
was 69.98%. For CNN model, the accuracies both training
set and test set were 98.00%. The result of CNN yielded
far superior to traditional techniques for distinguishing the
types of plastic.
Fig. 3 Classification results of models. a the confusion matrix of par-
tial least squares-discriminate analysis (PLS-DA); b the confusion
matrix of soft independent modeling of class analogy (SIMCA); c the
confusion matrix of linear discriminant analysis (LDA); d the confu-
sion matrix of convolutional neural network (CNN)
Environmental Chemistry Letters
1 3
Table1 gives the results of various algorithms combined
the different preprocessing. The algorithms included partial
least squares-discriminate analysis (PLS-DA), soft inde-
pendent modeling of class (SIMCA), linear discriminant
analysis (LDA) and convolutional neural network (CNN).
Preprocessing obtained Savitzky–Golay (S-G) smoothing,
standard normal variate (SNV) and the combination of
Savitzky–Golay (S-G) smoothing, standard normal variate
(SNV).
NIR is currently operational in some waste-sorting facili-
ties. However, NIR remains challenging to discriminate dif-
ferent black plastics. The reason for this limitation is that
black targets have low reflectance in the NIR spectral region.
In this research, black plastic accounted for half of the total
samples (84/159). CNN still could give an outstanding result
(98.00% accuracy). From the view of algorithm, it made
up for the shortcomings of NIR. To some extent, we may
explain the result from the followed theories. On the one
hand, focusing on black plastic from different categories, the
compositions of carbon black in plastics are slightly different
(Becker etal. 2017; Rozenstein etal. 2017). Hence, the NIR
spectra is similar, but the absorption intensity is not same.
On the other hand, CNN can learn the characteristics of plas-
tic. Just like the study of Ng (Ng etal. 2019), CNN model
also could produce an excellent result for soil properties.
After analyzing of correlation between soil properties, CNN
may learn inherent structure from data to improve prediction
(Ng etal. 2019; Xu etal. 2017).
Conclusion
This study focused on the problem of black plastic classifi-
cation. Several traditional algorithms including partial least
squares-discriminate analysis (PLS-DA), soft independent
modeling of class analogy (SIMCA) and linear discriminant
analysis (LDA) were employed to classify plastics. After
preprocessing, the accuracies of PLS-DA and LDA were
similar, both reaching to 57.00%. The accuracy of SIMCA
outputted 69.98%, which was higher than those of PLS-DA
and LDA. Convolutional neural network (CNN), which had
strong learning ability, after data normalization, provided an
excellent accuracy of 98.00%. This paper offered a potential
approach for industrial sorting of plastic waste using NIR
technology.
This study mainly focused on seven kinds of plastic. We
have updated the model with more categories of plastics.
Additionally, the plastics used were basically clear and
purity samples. However, the environment plastics in soil,
water and air may reduce the accuracy, given the excellent
effect of the CNN model. In the next study, we will move
to the actual samples from the daily life and build a more
broaden and robust model. Meanwhile, although CNN had
little concern on preprocessing and time-saving, the param-
eters in network were manually adjusted. The following
novelty is to find a solution to optimize parameters auto-
matically, further widening the application of CNN in waste
management.
Supplementary Information The online version contains supplemen-
tary material available at https:// doi. org/ 10. 1007/ s10311- 021- 01240-9.
Acknowledgements Thanks to the China Nanchang Customs provid-
ing the plastic samples and the corresponding chemistry information.
And the authors would like to thank Yue Huang and Qianqian Li for
the English language review. This research did not receive any specific
grant from funding agencies in the public, commercial or not-for-profit
sectors.
References
Acquarelli J, van Laarhoven T, Gerretzen J, Tran TN, Buydens LMC,
Marchiori E (2017) Convolutional neural networks for vibrational
spectroscopic data analysis. Anal Chim Acta 954:22–31. https://
doi. org/ 10. 1016/j. aca. 2016. 12. 010
Alassali A, Fiore S, Kuchta K (2018) Assessment of plastic waste
materials degradation through near infrared spectroscopy. Waste
Manag 82:71–81. https:// doi. org/ 10. 1016/j. wasman. 2018. 10. 010
Becker W, Sachsenheimer K, Klemenz M (2017) Detection of black
plastics in the middle infrared spectrum (MIR) using photon up-
conversion technique for polymer recycling purposes. Polymers
9:9. https:// doi. org/ 10. 3390/ polym 90904 35
Bevilacqua M, Marini F (2014) Local classification: locally weighted–
partial least squares-discriminant analysis (LW–PLS-DA). Anal
Chim Acta 838:20–30. https:// doi. org/ 10. 1016/j. aca. 2014. 05. 057
Camacho W, Karlsson S (2001) NIR, DSC, and FTIR as quantita-
tive methods for compositional analysis of blends of polymers
obtained from recycled mixed plastic waste. Polym Eng Sci
41:1626–35. https:// doi. org/ 10. 1002/ pen. 10860
Chu Y, Huang C, Xie X, Tan B, Kamal S, Xiong X (2018) Multilayer
hybrid deep-learning method for waste classification and recy-
cling. Comput Intell Neurosci 2018:1–9. https:// doi. org/ 10. 1155/
2018/ 50608 57
Cui C, Fearn T (2018) Modern practical convolutional neural networks
for multivariate regression: applications to NIR calibration. Che-
mom Intell Lab Syst 182:9–20. https:// doi. org/ 10. 1016/j. chemo
lab. 2018. 07. 008
Froelich D, Maris E, Haoues N, Chemineau L, Renard H, Abraham
F, Lassartesses R (2007) State of the art of plastic sorting and
Table 1 Results of various algorithms combining the different pre-
processing
PLS-DA partial least squares discrimination analysis, SIMCA soft
independent modeling of class analogy, LDA linear discriminant anal-
ysis, CNN convolutional neural network
PLS-DA SIMCA LDA CNN
Accuracy (%)
Preprocessing S-G → SNV S-G → SNV SNV Normalization
Training set 59.73 76.28 56.47 98.00
Test set 57.79 69.98 54.36 98.00
Environmental Chemistry Letters
1 3
recycling: feedback to vehicle design. Miner Eng 20:902–912.
https:// doi. org/ 10. 1016/j. mineng. 2007. 04. 020
Fuentes-García M, Maciá-Fernández G, Camacho J (2018) Evalua-
tion of diagnosis methods in PCA-based Multivariate Statistical
Process Control. Chemom Intell Lab Syst 172:194–210. https://
doi. org/ 10. 1016/j. chemo lab. 2017. 12. 008
Gerhardt N, Schwolow S, Rohn S, Pérez-Cacho PR, Galán-Soldevilla
H, Arce L, Weller P (2019) Quality assessment of olive oils based
on temperature-ramped HS-GC-IMS and sensory evaluation:
comparison of different processing approaches by LDA, kNN,
and SVM. Food Chem 278:720–728. https:// doi. org/ 10. 1016/j.
foodc hem. 2018. 11. 095
Gerretzen J, Szymanska E, Jansen JJ, Bart J, van Manen HJ, van den
Heuvel ER, Buydens LM (2015) Simple and Effective Way for
Data Preprocessing Selection Based on Design of Experiments.
Anal Chem 87:12096–12103. https:// doi. org/ 10. 1021/ acs. analc
hem. 5b028 32
Gyawali D, Regmi A, Shakya A, Gautam A, Shrestha S (2020) Com-
parative analysis of multiple deep CNN models for waste clas-
sification. https:// arxiv. org/ abs/ 2004. 02168
Kassouf A, Maalouly J, Rutledge DN, Chebib H, Ducruet V (2014)
Rapid discrimination of plastic packaging materials using MIR
spectroscopy coupled with independent components analysis
(ICA). Waste Manage 34:2131–2138. https:// doi. org/ 10. 1016/j.
wasman. 2014. 06. 015
Li J, Wu GQ, Xu ZM (2015) Tribo-charging properties of waste plastic
granules in process of tribo-electrostatic separation. Waste Man-
age 35:36–41. https:// doi. org/ 10. 1016/j. wasman. 2014. 10. 001
Macho S, Larrechi MS (2002) Near-infrared spectroscopy and mul-
tivariate calibration for the quantitative determination of certain
properties in the petrochemical industry. Trac-Trends Anal Chem
21:799–806. https:// doi. org/ 10. 1016/ s0165- 9936(02) 01202-5
Marca FL, Moroni M, Cherubini L, Lupo E, Cenedese A (2012) Sepa-
ration of plastic waste via the hydraulic separator Multidune under
different geometric configurations. Waste Manag 32:1306–1315.
https:// doi. org/ 10. 1016/j. wasman. 2012. 02. 021
Nanda S, Berruti F (2021) Thermochemical conversion of plastic waste
to fuels: a review. Environ Chem Lett 19:123–148. https:// doi. org/
10. 1007/ s10311- 020- 01094-7
Ng W, Minasny B, Montazerolghaem M, Padarian J, Ferguson R, Bai-
ley S, McBratney AB (2019) Convolutional neural network for
simultaneous prediction of several soil properties using visible/
near-infrared, mid-infrared, and their combined spectra. Geo-
derma 352:251–267. https:// doi. org/ 10. 1016/j. geode rma. 2019.
06. 016
Othman AR, Hasan HA, Muhamad MH, Ismail N, Abdullah SRS
(2021) Microbial degradation of microplastics by enzymatic
processes: a review. Environ Chem Lett. https:// doi. org/ 10. 1007/
s10311- 021- 01197-9
Padervand M, Rhimi B, Wang C (2020) One-pot synthesis of novel
ternary Fe3N/Fe2O3/C3N4 photocatalyst for efficient removal of
rhodamine B and CO2 reduction. J Alloy Compd 852:156955.
https:// doi. org/ 10. 1016/j. jallc om. 2020. 156955
Padervand M, Lichtfouse E, Robert D, Wang C (2020) Removal of
microplastics from the environment: a review. Environ Chem Lett
18:807–28. https:// doi. org/ 10. 1007/ s10311- 020- 00983-1
Padervand M, Ghasemi S, Hajiahmadi S, Wang C (2021) K4Nb6O17/
Fe3N/α-Fe2O3/C3N4 as an enhanced visible light-driven quater-
nary photocatalyst for acetamiprid photodegradation, CO2 reduc-
tieon, and cancer cells treatment. Appl Surf Sci 544:148939.
https:// doi. org/ 10. 1016/j. apsusc. 2021. 148939
Rani M, Marchesi C, Federici S, Rovelli G, Depero LE (2019) Min-
iaturized near-infrared (MicroNIR) spectrometer in plastic waste
sorting. Materials 12:2740. https:// doi. org/ 10. 3390/ ma121 72740
Rozenstein O, Puckrin E, Adamowski J (2017) Development of a new
approach based on midwave infrared spectroscopy for post-con-
sumer black plastic waste sorting in the recycling industry. Waste
Manag 68:38–44. https:// doi. org/ 10. 1016/j. wasman. 2017. 07. 023
Shen HT, Pugh RJ, Forssberg E (2002) Floatability, selectivity and
flotation separation of plastics by using a surfactant. Colloids
Surf-Physicochem Eng Aspects 196:63–70. https:// doi. org/ 10.
1016/ s0927- 7757(01) 00706-3
Signoret C, Caro-Bretelle AS, Lopez-Cuesta JM, Ienny P, Perrin D
(2019) MIR spectral characterization of plastic to enable dis-
crimination in an industrial recycling context: I. Specific case of
styrenic polymers. Waste Manag 95:513–25. https:// doi. org/ 10.
1016/j. wasman. 2019. 05. 050
Vanden Branden K, Hubert M (2005) Robust classification in high
dimensions based on the SIMCA Method. Chemom Intell Lab
Syst 79:10–21. https:// doi. org/ 10. 1016/j. chemo lab. 2005. 03. 002
Wang Z, Qin Y, Li W, Yang W, Meng Q, Yang J (2019) Microplastic
contamination in freshwater: first observation in Lake Ulansuhai,
Yellow River Basin, China. Environ Chem Lett 17:1821–1830.
https:// doi. org/ 10. 1007/ s10311- 019- 00888-8
Worrell E, Reuter MA (2014) Chapter2-definitions and terminology:
9–16. https://doi.org/https:// doi. org/ 10. 1016/ B978-0- 12- 396459-
5. 00002-7
Xu Y, Ma J, Liaw A, Sheridan RP, Svetnik V (2017) Demystifying
Multi-Task Deep Neural Networks for Quantitative Structure-
Activity Relationships. J Chem Inf Model 57:2490–504. https://
doi. org/ 10. 1021/ acs. jcim. 7b000 87
Zhang C-C, Sahajwalla V, Zhang F-S (2014) Tetrabromobisphenol A
recovery from computer housing plastic by a new solvothermal
process. Environ Chem Lett 12:347–352. https:// doi. org/ 10. 1007/
s10311- 014- 0452-8
Zhang XC, Gu F, Xie J, Zhang CQ, Fu JZ, Zhao P (2019) Magnetic
projection: a novel separation method and its first application on
separating mixed plastics. Waste Manage 87:805–813. https:// doi.
org/ 10. 1016/j. wasman. 2019. 03. 008
Zheng H, Gu Y (2021) EnCNN-UPMWS: waste classification by a
CNN ensemble using the UPM weighting strategy. Electronics
10:427. https:// doi. org/ 10. 3390/ elect ronic s1004 0427
Zheng Y, Bai J, Xu J, Li X, Zhang Y (2018) A discrimination model in
waste plastics sorting using NIR hyperspectral imaging system.
Waste Manag 72:87–98. https:// doi. org/ 10. 1016/j. wasman. 2017.
10. 015
Zhu S, Chen H, Wang M, Guo X, Lei Y, Jin G (2019) Plastic solid
waste identification system based on near infrared spectroscopy
in combination with support vector machine. Adv Ind Eng Polym
Res 2:77–81. https:// doi. org/ 10. 1016/j. aiepr. 2019. 04. 001
Zulkifley MA, Mustafa MM, Hussain A, Mustapha A, Ramli S (2014)
Robust identification of polyethylene terephthalate (PET) plastics
through bayesian decision. PLoS ONE 9:20. https:// doi. org/ 10.
1371/ journ al. pone. 01145 18
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.