materials

Article
Miniaturized Near-Infrared (MicroNIR) Spectrometer
in Plastic Waste Sorting
Monika Rani 1,2,† , Claudio Marchesi 1,2,† , Stefania Federici 1,2, * , Gianluca Rovelli 3 ,
Ivano Alessandri 2,4,5 , Irene Vassalini 2,4,5 , Serena Ducoli 2,4 , Laura Borgese 1,2 ,
Annalisa Zacco 1,2 , Fabjola Bilo 1,2 , Elza Bontempi 1,2 and Laura E. Depero 1,2
1 Department of Mechanical and Industrial Engineering, University of Brescia, via Branze,
38-25123 Brescia, Italy
2 Consorzio Interuniversitario Nazionale per la Scienza e Tecnologia dei Materiali (INSTM),
50121 Firenze, Italy
3 ABCS Srl-Scientific Instruments & Materials Science, viale Regina Giovanna, 37-20129 Milan, Italy
4 Department of Information Engineering, University of Brescia, via Branze, 38-25123 Brescia, Italy
5 CNR-INO, Unit of Brescia, 25123 Brescia, Italy
* Correspondence: stefania.federici@unibs.it
† These authors are equally contributed.




Received: 24 July 2019; Accepted: 23 August 2019; Published: 27 August 2019


Abstract: Valorisation of the urban plastic waste in high-quality recyclates is an imperative challenge
in the new paradigm of the circular economy. In this scenario, a key role in the improvement of
the recycling process is exerted by the optimization of waste sorting. In spite of the enormous
developments achieved in the field of automated sorting systems, the quest for the reduction
of cross-contamination of incompatible polymers as well as a rapid and punctual sorting of the
unmatched polymers has not been sufficiently developed. In this paper, we demonstrate that a
miniaturized handheld near-infrared (NIR) spectrometer can be used to successfully fingerprint and
classify different plastic polymers. The investigated urban plastic waste comprised polyethylene (PE),
polypropylene (PP), poly(vinyl chloride) (PVC), poly(ethylene terephthalate) (PET), and poly(styrene)
(PS), collected directly in a recycling plastic waste plant, without any kind of sample washing or
treatment. The application of unsupervised and supervised chemometric tools such as principal
component analysis (PCA) and partial least squares-discriminant analysis (PLS-DA) on the NIR
dataset resulted in a complete classification of the polymer classes. In addition, several kinds of PET
(clear, blue, coloured, opaque, and boxes) were correctly classified as PET class, and PE samples with
different branching degrees were properly separated.

Keywords: urban plastic waste; near-infrared (NIR) spectroscopy; chemometrics; principal

component analysis (PCA); partial least squares-discriminant analysis (PLS-DA); circular economy

1. Introduction
The huge amount of urban plastic waste and the continuous growth in human plastic consumption
require a high valorisation of the collected waste in the direction of a whole-system economic
sustainability. By the numbers, European plastic production reached almost 60 million tonnes in 2016,
and 335 million tonnes globally (Source: PlasticsEurope Market Research Group (PEMRG)/Conversio
Market & Strategy GmbH). In this framework, over 8.4 million tonnes of plastic waste were collected
in order to be recycled inside and outside the EU [1]. The recycling of polymer waste has valuable
environmental benefits due to the substitution of primary production, and a key role in the improvement
of the recycling process is exerted by the optimization of waste sorting [2,3].

Materials 2019, 12, 2740; doi:10.3390/ma12172740 www.mdpi.com/journal/materials

Materials 2019, 12, 2740 2 of 13

Generally, automated sorting processes are widespread in plastic recycling facilities [4].
These systems are usually based on vibrational spectroscopy techniques for polymer identification [5–9]
and camera systems for optical recognition of clear and coloured products [10,11]. Other sorting
technologies include UV–Vis spectroscopy [12] or mass spectroscopy (Py-GC/MS) for identification [13],
hyper-spectral imaging methods [14], X-ray detection for the separation of specific containers [15],
fluorescence spectroscopy for identifying halogens and heavy metals [16], and laser-induced plasma
spectroscopy to detect additives [17]. This approach has increased the purity of the output plastic
over the years, reaching a high percentage of recyclates in the production of secondary materials.
Nevertheless, there are severe limitations of these techniques in handling mixed plastic, which requires
additional sorting elsewhere and can affect the quality of the recyclate, if not properly assigned.
The positive cost/benefit balance can take place only if the separated fractions of polymers match a
high purity grade, fulfilling the market requirement of high recyclate quality. Therefore, a key step in
post-consumer recycling is reducing the cross-contamination of incompatible polymers [10].
Besides the automated sorting of the plastic stream, manual sorting is still necessary for the full
separation of the collected waste. However, manual sorting based only on plastic code recognition is
inefficient and time consuming. In some cases, plastic goods can be badly crushed or have missing or
unreadable codes. Manufacturers can also use different polymers for the same product, depending on
marketing choices. In addition, the high turnover rate can be affected by an improperly trained or
unskilled workforce.
In this scenario, the possibility of exploiting a well-established technique for polymer identification
combined with a miniaturized, portable, low-cost, and real-time spectrometer for local and punctual
semi-automated sorting is highly desirable. Urban plastic waste normally contains plastics such as
polyethylene (PE), polypropylene (PP), poly(vinyl chloride) (PVC), poly(ethylene terephthalate) (PET),
and poly(styrene) (PS). These polymers have characteristic near-infrared (NIR) spectral fingerprints that
can be used for accurate sorting and separation [5]. The progress in miniaturization technologies has
generated specific instrumental features that have been reported in polymer identification applications,
but results have been reported only for a selection of polymers from commercial libraries, and are not
fully representative of real plastic waste [18].
In the present work, we challenged a compacted and miniaturized NIR spectrometer for a rapid
identification of urban plastic waste collected directly from a selection and recycling plant, without
any sample treatment. We collected a large amount of plastic samples directly from the plant and we
created a robust database of NIR spectra. Different chemometric tools were applied for data evaluation
of all the recorded samples. Principal component analysis (PCA) for preliminary data exploration and
partial least squares-discriminant analysis (PLS-DA) as supervised pattern recognition were applied to
fingerprint and classify the different plastic polymers.

2. Materials and Methods

2.1. Polymer Samples Collection

Plastic samples were collected at the selection division of the recovery and recycling plant Montello
SpA (Bergamo, Italy), which receives post-consumer plastic in the form of urban waste for recycling.
In total, 250 samples belonging to various classes of polymer were used (see Supplementary Materials
Figure S1). The polymers considered were: several kinds of PET (clear, blue, coloured, opaque,
and boxes), PE with different branching degrees (low-density polyethylene, LDPE, and high-density
polyethylene, HDPE), PVC, PP, and PS. The collection included bottles, containers, and packets of
different shapes, sizes, and colours. The samples were flattened or partially flattened at the sampling
site, and were not washed prior to testing. Black samples were not considered, due to the very low
reflectance in the NIR spectral region and the low signal-to-noise ratio of NIR sensors, which hamper
successful black polymer spectral acquisition [19].
Materials 2019, 12, 2740 3 of 13

Materials
2.2. NIR2019, 12, x FOR PEER REVIEW
Measurements 3 of 13

Spectra
Spectraacquisition
acquisitionwaswasrealized using
realized MicroNIR
using MicroNIROn-site (Viavi
On-site Solutions
(Viavi Inc., CA,
Solutions Inc.,United States
CA, United
in reflectance mode without any sample preparation. The instrument is a
States in reflectance mode without any sample preparation. The instrument is a miniaturized miniaturized palm-sized
portable
palm-sizedspectrometer of about 250ofg about
portable spectrometer in weight,
250 gless than 200
in weight, mm
less in 200
than length,
mmand 50 mmand
in length, in diameter.
50 mm in
This
diameter. This handheld spectrometer employs a linear variable filter (LVF) as dispersing contrast
handheld spectrometer employs a linear variable filter (LVF) as dispersing element, in element,
to
in traditional
contrast to diffraction-based spectrometers
traditional diffraction-based (see Figure(see
spectrometers 1 for an instrument
Figure operating operating
1 for an instrument scheme).
The controlling
scheme). parametersparameters
The controlling for spectral for
dataspectral
collection
datawere set at 10were
collection ms integration
set at 10 ms time and 50 scans,
integration time
resulting in a short measurement time of 0.25 s. Not less than five replicates in different
and 50 scans, resulting in a short measurement time of 0.25 s. Not less than five replicates in different positions were
acquired
positionsby a point-and-shoot
were acquired by a technique, in order
point-and-shoot to minimize
technique, effectstodriven
in order by the
minimize non-uniformity
effects driven by theof
samples. A totalof
non-uniformity ofsamples.
1303 NIRAspectra
total ofwere
1303 collected.
NIR spectra A dark
were and a reference
collected. A dark scan
andwere carried scan
a reference out
approximately every 10 min. Data acquisition was realized through MicroNIR TM Pro v3.0 software
were carried out approximately every 10 min. Data acquisition was realized through MicroNIRTM
(Viavi
Pro v3.0Solutions
software Inc., CA, Solutions
(Viavi United States).
Inc., CA, United States).

Figure 1. (a) Miniaturized near-infrared (microNIR) device and (b) instrument operating scheme.
Figure 1. (a) Miniaturized near-infrared (microNIR) device and (b) instrument operating scheme.
2.3. Signal Pretreatment
2.3. Signal
SignalPretreatment
correction methods are mathematical tools employed to remove additive and/or
multiplicative effects onmethods
Signal correction spectroscopic data, which often
are mathematical tools affect the performance
employed to remove of chemometric
additive and/or
analysis [20]. These phenomena are typical of light scattering effects, which induce
multiplicative effects on spectroscopic data, which often affect the performance of chemometric a photon loss
(addictive effect) and an increased path length (multiplicative effect), among others.
analysis [20]. These phenomena are typical of light scattering effects, which induce a photon loss These methods
are “row-wise”,
(addictive effect)and
andtherefore the correction
an increased is made
path length sample-by-sample.
(multiplicative This others.
effect), among is in contrast
These to mean
methods
centring and autoscaling,
are “row-wise”, whichthe
and therefore arecorrection
“column-wise”
is made treatments. In most cases,
sample-by-sample. Thisthe combination
is in contrast toof the
mean
two types of method is required in order to make a satisfactory data analysis [21]. In this
centring and autoscaling, which are “column-wise” treatments. In most cases, the combination of the study, spectra
were
two collected
types of into
methoda single 1307 × 125
is required in matrix
order to make a×satisfactory
(samples scans), and several pre-processing
data analysis toolsstudy,
[21]. In this were
applied to the raw data. The best result was obtained by applying Savitzky-Golay second-derivative
spectra were collected into a single 1307 × 125 matrix (samples × scans), and several pre-processing
procedure
tools wereofapplied
seven data points
to the rawand a second-order
data. The best resultpolynomial followedby
was obtained byapplying
standard Savitzky-Golay
normal variate
(SNV). Moreover, after
second-derivative the preprocessing,
procedure spectra
of seven data wereand
points normalized by meanpolynomial
a second-order centring. followed by
standard normal variate (SNV). Moreover, after the preprocessing, spectra were normalized by
2.4. Data Analysis
mean centring.
Different chemometric tools were applied in order to have a correct data evaluation of all the
2.4. Data samples.
analysed Analysis The first step consisted of an exploratory analysis by PCA aiming to explore the
data structure. PCA was carried out with a total of 1303 NIR spectra for all five polymer classes.
Different chemometric tools were applied in order to have a correct data evaluation of all the
Subsequently, PLS-DA was used as supervised pattern recognition with the aim of separating the
analysed samples. The first step consisted of an exploratory analysis by PCA aiming to explore the
data structure. PCA was carried out with a total of 1303 NIR spectra for all five polymer classes.
Subsequently, PLS-DA was used as supervised pattern recognition with the aim of separating the
different commodities, and it was performed on a total of 1287 NIR spectra (16 spectra were
Materials 2019, 12, 2740 4 of 13
Materials 2019, 12, x FOR PEER REVIEW 4 of 13

different
removed commodities,
as outliers).and
All it was performed
calculations wereon a total of using
performed 1287 NIR spectra(R2018a)
MATLAB (16 spectra were
with removed as
chemometric
outliers). All calculations
toolboxes [22,23]. were performed using MATLAB (R2018a) with chemometric toolboxes [22,23].

3. 3.Results
Resultsand
andDiscussion
Discussion
NIR
NIRspectroscopy
spectroscopyhas hasthethekey
keyadvantage
advantage of of being
being aa rapid-response analyticaltool,
rapid-response analytical tool,recording
recording
spectra
spectra with no prior manipulation and predicting physical and chemical parameters from aasingle
with no prior manipulation and predicting physical and chemical parameters from single
spectrum
spectrum[24].
[24].InInthe
theNIR
NIRregion,
region,the
theabsorption
absorption bands
bands occur due to
occur due to overtones
overtonesor
orcombination
combinationbands
bands
ofof
mainly
mainly carbon–hydrogen vibrations and oxygen-hydrogen vibrations. A correct band assignmentis
carbon–hydrogen vibrations and oxygen-hydrogen vibrations. A correct band assignment
challenging because
is challenging becauseit may
it maybebe
due
duetoto
different
differentcombinations
combinationsof of fundamental vibrations.Moreover,
fundamental vibrations. Moreover,
overtone
overtonevibrations
vibrationsare arehighly
highlyoverlapped
overlapped[25].
[25].
The
TheNIR
NIRreflectance
reflectancespectra
spectraof
of the
the five
five polymers (i.e., PET,
polymers (i.e., PET, PP,PP, PVC,
PVC, PE,
PE, and
and PS)
PS)in
inthe
therange
range
900–1700nm,
900–1700 nm,obtained
obtainedby bythe
theMicroNIR
MicroNIROn-site,
On-site,are
areshown
shown in in Figure
Figure 2.

Figure
Figure2. 2.Near-infrared
Near-infrared(NIR) spectra
(NIR) of of
spectra five classes
five of plastics:
classes (a)(a)
of plastics: representative raw
representative spectra
raw of the
spectra five
of the
classes; (b) corresponding spectra after pre-treatment by second derivative and standard normal
five classes; (b) corresponding spectra after pre-treatment by second derivative and standard normal variate
(SNV). PE:(SNV).
variate polyethylene; PET: poly(ethylene
PE: polyethylene; PET: terephthalate);
poly(ethylene PP: polypropylene;
terephthalate); PP:PS: poly(styrene); PVC:
polypropylene; PS:
poly(vinyl chloride).
poly(styrene); PVC: poly(vinyl chloride).

InIn
the
theabsorbance
absorbance spectra
spectra(Figure
(Figure2a), thethe
2a), principal
principalabsorbance
absorbance band forfor
band PET was
PET wasfound
found at 1660 nm
at 1660
due to the first overtone of C–H stretching [26], with two other
nm due to the first overtone of C–H stretching [26], with two other small peaks small peaks around 1130 nm
1130 nm and and
1415
1415nm.nm.InInthethecase
caseofofPVC,
PVC,thisthismain
mainpeak
peak waswas shifted
shifted and was not not recorded
recorded in in the
theinstrumental
instrumental
operating
operatingwindow,
window,and and only the the two
twosmall
smallpeaks
peaks around
around 1190
1190 nmnm andand
14201420 nm were
nm were seenThe
seen [26]. [26].
The same
same behaviour
behaviour appeared
appeared forforPEPE and
and PP,PP, assigningtwo
assigning twobands
bandsaround
around12101210nm nmandand1430
1430nm nmforforPEPE
and
andaround
around1190 1190nmnmandand1400
1400nm nmforforPP.
PP.PS PSexhibited
exhibited aa dominant
dominant band band around
around 16751675nm nmandandtwotwo
small
small peaksaround
peaks around1140
1140nm nmandand1205
1205nmnm[5].[5]. Figure 2b refers to to the
the corresponding
correspondingspectra spectrashown
showninin
Figure2a2aafter
Figure afterpre-treatment
pre-treatmentby bysecond
secondderivative
derivative and SNV. In general, by by comparing
comparingthe thespectra
spectraofof
samples made of one type of polymer, a variation in the absorbance values was noted andthis
samples made of one type of polymer, a variation in the absorbance values was noted and thiswas
was
ascribed
ascribed to to variations
variations in sample
in sample thickness
thickness [5]. However,
[5]. However, the overall
the overall form ofform of thewas
the spectra spectra was
preserved,
preserved,that
concluding concluding that dust
dust or liquid or liquid contaminations
contaminations randomly characterizing
randomly characterizing each sample each did notsample
affectdid
the
not affect the quality of NIR spectra (see Appendix A, Figure A1) [27]. In addition,
quality of NIR spectra (see Appendix A, Figure A1) [27]. In addition, NIR spectra of samples of different NIR spectra of
samples
colours didofnot
different
display colours did not display
considerable differencesconsiderable
(Figure A1), differences
meaning (Figure
that theA1), meaningpresent
colorants that thein
colorants present in the plastic resin matrix did not significantly influence
the plastic resin matrix did not significantly influence the NIR spectra. Moreover, no differences the NIR spectra.in
Moreover, no differences in band positions were found when studying the
band positions were found when studying the spectra of several kinds of PET (clear, coloured, opaque, spectra of several kinds
of PET (clear, coloured, opaque, box, and blue) (see Appendix A, Figure
box, and blue) (see Appendix A, Figure A2), implying that the spectra can be successfully exploited A2), implying that theto
spectra can PET
discriminate be successfully exploited
samples, despite to discriminate
their differences. For PETthis
samples,
reason, despite their differences.
hereinafter, only one PET For class
this
reason, hereinafter,
is considered. only one PET class is considered.

3.1. Principal Component Analysis

Principal component analysis is a very powerful chemometric tool for analysing data structure
[28]. The aim of PCA is to extract the information encoded in a certain number of variables into a
Materials 2019, 12, 2740 5 of 13

3.1. Principal Component Analysis

Principal component analysis is a very powerful chemometric tool for analysing data structure [28].
Materials 2019, 12, x FOR PEER REVIEW
The aim of PCA is to extract the information encoded in a certain number of variables into5aofsmaller 13

set ofsmaller
new orthogonal
set of new variables
orthogonalcalled principal
variables components.
called PCA calculation
principal components. PCA was performed
calculation was over
the entire
performed over the entire spectral region and using all collected spectra. The first three components for
spectral region and using all collected spectra. The first three components accounted
87.11% of the total
accounted variability
for 87.11% (PC1
of the total56.75%, PC2(PC1
variability 15.63%, andPC2
56.75%, PC315.63%,
14.73%).
andFigure 3 displays
PC3 14.73%). the3score
Figure
plots displays
of PC1 vs. thePC2
score (Figure 3a)
plots of and
PC1 vs.PC1
PC2vs. PC3 (Figure
(Figure 3a) and 3b).
PC1 In
vs.the
PC3score plot3b).
(Figure displayed in Figure
In the score plot 3a,
displayed
a clear separation in Figure
between3a, athe
clear separation
five between
commodities canthebefive commodities
seen. Indeed, in can benew
the seen.space
Indeed, in the
determined
by thenew
firstspace
two determined
components, by polymers
the first two components,
formed polymers
very tight formed very tight
and homogeneous and homogeneous
clusters. Along the first
component, a clear differentiation between PET, PVC, PP, and PE was highlighted, whilePE
clusters. Along the first component, a clear differentiation between PET, PVC, PP, and thewas
second
highlighted, while the second component allowed the differentiation of PS from the other polymers,
component allowed the differentiation of PS from the other polymers, and PE from PP. On the other
and PE from PP. On the other hand, Figure 3b shows the score plot of PC1 vs. PC3. In this case PET
hand, Figure 3b shows the score plot of PC1 vs. PC3. In this case PET was differentiated from the other
was differentiated from the other polymers along PC1, while PC3 distinguished between the
polymers alongsets
remaining PC1, of while PC3 distinguished between the remaining sets of polymers.
polymers.

Figure
Figure 3. Results
3. Results of principal
of principal components analysis
components analysis carried
carriedout outwith
withspectral data
spectral of the
data different
of the different
commodities. (a) The score plot of the first two components is shown, as well as (b) the
commodities. (a) The score plot of the first two components is shown, as well as (b) the score plot score plot of of
PC1 vs. PC3 EV: explained variance.
PC1 vs. PC3 EV: explained variance.
Materials 2019, 12, 2740 6 of 13

Materials 2019, 12, x FOR PEER REVIEW 6 of 13

Furthermore, the results of PCA applied to the restricted dataset of PET samples showed no
Furthermore, the results of PCA applied to the restricted dataset of PET samples showed no
separation among the five kinds of PET samples (Figure 4), confirming the interpretation of the band
separation among the five kinds of PET samples (Figure 4), confirming the interpretation of the band
assignment of NIR spectra.
assignment of NIR spectra.

Figure 4.4.Principal component

Principal componentanalysis (PCA) (PCA)
analysis analysisanalysis
of severalofkinds of poly(ethylene
several terephthalate)
kinds of poly(ethylene
(PET) samples for a total of 659 spectra considered. EV: explained variance.
terephthalate) (PET) samples for a total of 659 spectra considered. EV: explained variance.

3.2. HDPE versus

3.2. HDPE versus LDPE
LDPE
The
The discrimination
discriminationbetween
betweenHDPE HDPEandandLDPELDPEisisananimportant
important issue
issuein in
terms
termsof plastic recycling
of plastic [3].
recycling
PE
[3]. PE is largely used in the food packaging sector, thanks to its remarkable properties in terms of
is largely used in the food packaging sector, thanks to its remarkable properties in terms of
mechanical
mechanical and andoptical performances,
optical performances,water-vapor-resistance,
water-vapor-resistance,and heat-sealing attitude. The
and heat-sealing amount
attitude. of
The
side
amountbranches that are
of side attached
branches to the
that aremain PE polymer
attached to the chain
maindetermines
PE polymer the morphological
chain determines properties,
the
which in turn affects optical, physical, and thermal properties [29,30]. NIR spectroscopy
morphological properties, which in turn affects optical, physical, and thermal properties [29,30]. NIR successfully
discriminates between HDPE
spectroscopy successfully and LDPE [31,32],
discriminates betweeneven HDPEif severe limitations
and LDPE [31,32],can occur
even depending
if severe on the
limitations
characteristics
can occur dependingof the measuring instruments [27].
on the characteristics of In themeasuring
the sampling campaign,
instruments we collected
[27]. In the25 samples
sampling of
HDPE and 25 samples of LDPE. Raw NIR spectra did not show any significant
campaign, we collected 25 samples of HDPE and 25 samples of LDPE. Raw NIR spectra did not feature to discriminate
the
show twoanyplastics, nor did
significant a clear
feature toseparation
discriminatein the
the score plots resulting
two plastics, nor didfrom the separation
a clear previous PCA in theanalysis
score
(Figure 3). However, by limiting the PCA analysis to the PE samples, a separation
plots resulting from the previous PCA analysis (Figure 3). However, by limiting the PCA analysis between HDPE and
to
LDPE was obtained; the first two components explicated 61.22% of the
the PE samples, a separation between HDPE and LDPE was obtained; the first two components total variability (PC1 43.22%,
PC2 18.00%),
explicated confirming
61.22% of thethe successful
total application
variability of NIR PC2
(PC1 43.22%, spectroscopy
18.00%), inconfirming
discriminating these two
the successful
polymers. Figure 5 displays the score plots of PC1 vs. PC2.
application of NIR spectroscopy in discriminating these two polymers. Figure 5 displays the score
plots of PC1 vs. PC2.
Materials 2019, 12, 2740 7 of 13
Materials 2019, 12, x FOR PEER REVIEW 7 of 13

Figure 5.
Figure 5. Results
Results of
of principal
principal components
components analysis
analysiscarried
carriedout
outwith
withspectral
spectraldata
dataof
ofpolyethylene
polyethylene(PE)
(PE)
class polymers.
class polymers. Score
Score plot
plot of
of the
the first
first two
two components
components is is shown.
shown.

3.3.
3.3. Partial
Partial Least
Least Squares-Discriminant
Squares-Discriminant Analysis
Analysis
After
After thethe exploratory
exploratory PCA PCA analysis,
analysis, aa supervised
supervised classification
classification tool tool was
was applied
applied in in order
order to to
distinguish
distinguish the the different
different plastic
plastic classes.
classes. In In PLS-DA,
PLS-DA, the the well-known
well-known PLS PLS regression
regression algorithm
algorithm is is
modified with a classification goal. The response variable is categorical,
modified with a classification goal. The response variable is categorical, reflecting the belonging reflecting the belonging class
of the of
class statistical units. units.
the statistical PLS-DA returnsreturns
PLS-DA the prediction in a vector
the prediction in a of size equal
vector of sizetoequal
the number of classes
to the number of
in the predictor variables, with values ranging from
classes in the predictor variables, with values ranging from 0 to 1 [33].0 to 1 [33].
Validation
Validationtools toolsareare
very useful
very methods
useful methodsin chemometrics,
in chemometrics, and they andarethey
usedare
to verify
used the capability
to verify the
of the model
capability prediction.
of the Before the
model prediction. model
Before thecalibration, data were
model calibration, datasplit
wereintosplittraining and test
into training andsets
test
containing 901 and
sets containing 901386andsamples, respectively.
386 samples, The training
respectively. set wasset
The training used
wastoused
calibrate the model
to calibrate theboth
modelin
fitting and in cross-validation, while the test set was employed only
both in fitting and in cross-validation, while the test set was employed only at the end of the at the end of the procedure in
order to evaluate
procedure in order thetotrue predictive
evaluate capability
the true predictive of the calibrated
capability model
of the [22]. model [22].
calibrated
During
During the calibration, the selection of the optimal number of latent variables
the calibration, the selection of the optimal number of latent variables (LVs)
(LVs) isis aa crucial
crucial
point,
point, which
which is is performed
performed basedbased on on the
the cross-validation
cross-validation procedureprocedure [22]. [22]. This
This should
should allowallow thethe
optimization of the complexity of the multivariate model according
optimization of the complexity of the multivariate model according to the predictive capacity of the to the predictive capacity of
the model itself. Cross-validation is generally carried out by dividing
model itself. Cross-validation is generally carried out by dividing the training set into different the training set into different
cross-validation
cross-validationgroups, groups,andandduring
duringeacheachround
round oneonegroup
group is removed
is removed from the the
from training set. set.
training OverOver
the
rounds, the model is calibrated on the remaining training samples and
the rounds, the model is calibrated on the remaining training samples and then used to predict then used to predict samples of
the cross-validation group. The cross-validation procedure was based
samples of the cross-validation group. The cross-validation procedure was based on a Venetian on a Venetian blinds approach
with
blinds10approach
groups. In the10Venetian
with groups.blinds
In the procedure,
Venetian blinds the calibration
procedure,group and the cross-validation
the calibration group and the
group are selected by choosing every n-th sample from the dataset
cross-validation group are selected by choosing every n-th sample from the dataset starting starting from the first one
from[22].
the
The
firstselection
one [22].ofThe theselection
number of of cross-validation groups was set to
the number of cross-validation 10, as was
groups a reasonable
set to 10,number in order
as a reasonable
to avoid the
number so-called
in order to “overfitting” (i.e., the“overfitting”
avoid the so-called overestimation of the
(i.e., thepredictive capability
overestimation of theof the model).
predictive
Figure 6 shows the error rate (Figure 6a) and the degree of not assigned
capability of the model). Figure 6 shows the error rate (Figure 6a) and the degree of not assigned samples as a function of the
number
samplesof aslatent variables
a function (Figure
of the number6b).ofThe optimal
latent number
variables of LVs
(Figure was
6b). Thesetoptimal
to 6 (explaining
number of 94% LVsofwas
the
total variability and 0.0099% of not assigned samples), which was associated
set to 6 (explaining 94% of the total variability and 0.0099% of not assigned samples), which was with the minimum error
rate and, simultaneously,
associated with the minimum with the
errorminimum
rate and,percentage
simultaneously,of not with
assigned samples. Four
the minimum LVs reached
percentage of not
the same error
assigned samples. rate,Four
but 0.019% of not the
LVs reached assigned
same errorsamples.
rate, but 0.019% of not assigned samples.
 Materials 2019, 12, 2740 8 of 13
Materials 2019, 12, x FOR PEER REVIEW 8 of 13

Figure 6. Error rate (a) and not assigned samples (b) as a function of latent variables calculated in the
partial
partial least
leastsquares-discriminant
squares-discriminantanalysis (PLS-DA)
analysis model.
(PLS-DA) Six latent
model. variables
Six latent (LVs) was
variables the was
(LVs) optimal
the
number, marked in red, with 0.0099% of not assigned samples.
optimal number, marked in red, with 0.0099% of not assigned samples.

After the determination

After the determinationofofthetheoptimal
optimal number
number of of
LVs,LVs,
thethe calculation
calculation of the
of the PLS-DA
PLS-DA model
model was
was performed
performed selecting
selecting six latent
six latent variables
variables and and 10 cross-validation
10 cross-validation groups
groups for for internal
internal validation.
validation. In
In Table 1, the two confusion matrices achieved in fitting and in cross-validation were
Table 1, the two confusion matrices achieved in fitting and in cross-validation were presented. Both presented.
Both in fitting
in fitting and and in cross-validation,
in cross-validation, the the performance
performance of of
thethe modelwas
model wasvery
veryhigh,
high,proved
proved by
by the
the
correlation between observed and predicted classes. All the investigated samples
correlation between observed and predicted classes. All the investigated samples were correctly were correctly
designated,
designated, andand only
only four
four samples
samples inin fitting
fitting and
and five
five in
in cross-validation
cross-validationwere
werenotnotassigned.
assigned.

Table 1. Confusion
Table 1. Confusion matrices
matrices obtained
obtained from
from the
the PLS-DA
PLS-DA model,
model, both
both in
in fitting
fitting and
and in
in cross-validation
cross-validation
(based on Venetian blinds with 10 groups). The “Not Assigned” column contains samples
(based on Venetian blinds with 10 groups). The “Not Assigned” column contains samples whichwhich
were
not assigned to any of the considered classes.
were not assigned to any of the considered classes.
CalculatedClass
Calculated Class
Experimental Experimental Class
PET PE PP PVC PS Not Assigned Not
Class PET PE PP PVC PS
Fitting - - - - - - Assigned
Fitting PET
- 459
- 0 -0 0 -0 2 - -
PE 0 181 0 0 0 1
PET 459 0 0 0 0 2
PP 0 0 82 0 0 0
PE 0
PVC 1810 0 0 0 84 0 0 1 0 1
PP 0
PS 00 0 820 0 091 0 0 0
PVC 0
Cross-Validation 0- - 0- - 84- - 0 1
PS PET
0 460
0 0 00 0 00 1 91 0
PE 0 181 0 0 0 1
Cross-Validation - - - - - -
PP 0 0 82 0 0 0
PET 460
PVC 00 0 0 0 82 0 0 3 0 1
PE 0
PS 1810 0 00 0 091 0 0 1
PP 0 0 82 0 0 0
PVC 0 0 0 82 0 3
The performance of the classification model could be also evaluated by considering the classification
PS 0 0 0 0 91 0
parameters derived from the confusion matrix: sensitivity (Sn), specificity (Sp), and non-error rate
(NER). These parameters represent the ability of the model to correctly identify the samples of the
The performance of the classification model could be also evaluated by considering the
classification parameters derived from the confusion matrix: sensitivity (Sn), specificity (Sp), and
Materials 2019, 12, 2740 9 of 13

given class, the ability of a classifier to reject the samples of other classes, and the average of the class
sensitivities, respectively [34]. Table 2 collects the parameters of the PLS-DA model. In both fitting
and cross-validation, NER, Sn, and Sp were equal to 1, meaning that 100% of the polymers were
correctly classified.

Table 2. Classification parameters (non-error rate (NER), class sensitivity (Sn), and specificity (Sp))
calculated in fitting and in cross-validation.

- - PET PE PP PVC PS
NER Sn Sp Sn Sp Sn Sp Sn Sp Sn Sp
Fitting 1 1 1 1 1 1 1 1 1 1 1
Cross-validation 1 1 1 1 1 1 1 1 1 1 1

Furthermore, we analysed the confusion matrix and the classification parameters for the test set,
used as an additional trial to validate the model. Table 3 shows the confusion matrix derived from
the test set; again, 100% of the samples were classified, and only four samples were not assigned.
Table 4 summarizes the classification parameters obtained from the confusion matrix, with the NER,
Sn, and Sp of each class equal to 1.

Table 3. Confusion matrix obtained from the PLS-DA model fitted on the validation set. The “Not
Assigned” column contains samples which were not assigned to any of the considered classes.

Calculated Class
Experimental Class
PET PE PP PVC PS Not Assigned
PET 197 0 0 0 0 1
PE 0 78 0 0 0 1
PP 0 0 33 0 0 1
PVC 0 0 0 35 0 1
PS 0 0 0 0 39 0

Table 4. Classification parameters (non-error rate (NER), class sensitivity (Sn), and specificity (Sp))
calculated on the validation set.

- PET PE PP PVC PS
-
NER Sn Sp Sn Sp Sn Sp Sn Sp Sn Sp
Test 1 1 1 1 1 1 1 1 1 1 1

In the case of the restricted data set composed of HDPE and LDPE spectra, the optimal number
of LVs was set to six, assigning 90% of total variability (see Supplementary Materials, Figure S2).
With respect to the analysis of the polymer classes, the degree of not assigned samples as a function
of the number of latent variables was equal to zero, meaning that all the samples were assigned to a
class. After the determination of the optimal number of LVs, the calculation of the PLS-DA model
was performed, selecting six latent variables and 10 cross-validation groups for internal validation.
Both in fitting and in cross-validation, all the investigated samples were correctly designated (see
Table S1 in the Supplementary Materials for the two confusion matrices achieved in fitting and in
cross-validation).

4. Conclusions
In the present study, the use of a handheld NIR spectrometer combined with robust chemometrics
tools to fingerprint different urban plastic waste collected directly from a recycling plant was presented.
Plastic samples included clear, blue, coloured, opaque, and boxes of PET, PE with different branching
Materials 2019, 12, 2740 10 of 13

degrees (HDPE and LDPE), PP, PVC, and PS. In a first step, PCA was applied to NIR spectra of all
polymer classes, and polymers were correctly separated in clusters. Moreover, by limiting the analysis
on PE samples, HDPE and LDPE were correctly discriminated.
In addition, by comparing samples of different colours, no differences in band positions were
found, suggesting that colorants present in the plastic resin matrix do not significantly influence NIR
spectra. Moreover, even if samples were not washed or treated prior to the analysis, NIR spectra were
not affected by dust contamination or liquid residuals, meaning a successful application directly to the
urban plastic waste.
After the exploratory PCA analysis, PLS-DA was applied. Before the model calibration, data were
split into a training and test set, using the training set to calibrate the model. In both fitting and
cross-validation, the performance of the model was very high, proved by the correlation between
observed and predicted classes. All the analysed samples were also correctly designated in the case of
the restricted data set composed of HDPE and LDPE spectra.
The results show promising outcomes in order to improve the reliability and efficiency of the
manual sorting of plastic waste, increasing the volume of recycled plastic and the management of
multi-component polymer types.

Supplementary Materials: The following are available online at http://www.mdpi.com/1996-1944/12/17/2740/s1.

Figure S1: Systematic organization of the collected urban plastic waste for laboratory measurements, Figure S2:
Error rate as a function of latent variables calculated in the PLS-DA model for HDPE and LDPE classification. Six
LVs was the optimal number, Table S1: Confusion matrices obtained from PLS-DA model, both in fitting and in
cross-validation (based on Venetian blinds with 10 groups) restricted to the PE class.
Author Contributions: Conceptualization, S.F. and L.E.D.; Funding acquisition, G.R., E.B., and L.E.D.;
Investigation, M.R., C.M., I.V., S.D., A.Z., and F.B.; Resources, E.B. and L.E.D.; Supervision, S.F. and L.E.D.;
Writing—original draft, M.R., C.M., and S.F.; Writing—review and editing, S.F., I.A., L.B., and L.E.D.
Funding: This research was funded by PON “R&I” 2014-2020: SIRIMAP—SIstemi di Rilevamento
dell’Inquinamento MArino da Plastiche e successivo recupero-riciclo (No. ARS01_01183) CUP D86C18000520008.
Conflicts of Interest: The authors declare no conflicts of interest.

Appendix A

Appendix A.1 Evaluation of Sample Contaminations and Colour Influence

Figure A1 shows a representative collection of NIR spectra of coloured PET samples. The coloured
PET batch was chosen as a demonstrator set to compare samples with random contaminations and
to evaluate the presence of colorants. A total of 136 spectra were collected (25 samples with at least
5 replicates in different positions for each sample). The samples obtained from the recycling plant
were flattened or partially flattened directly on site and were not washed prior the testing. Liquid
residuals, pieces of labels, and environmental dust were largely present on the surface and inside the
samples. Nevertheless, NIR spectra showed only a variation in the total absorbance values, attributable
to variations in sample thickness, according to Lambert–Beer law [35]. Moreover, no differences in
band positions were found despite the different colours of the commodities.
Materials 2019, 12,
Materials 2019, 12, 2740
x FOR PEER REVIEW 11 of
11 of 13
13
Materials 2019, 12, x FOR PEER REVIEW 11 of 13

Figure A1. Representative NIR spectra of coloured PET samples.

Figure A1.
Figure A1. Representative
Representative NIR
NIR spectra
spectra of
of coloured
coloured PET
PET samples.
samples.
Appendix A.2. PET Samples
Appendix
Appendix A.2A.2.PET
PETSamples
Samples
Several kinds of PET were considered. In particular, clear, coloured, opaque, blue, and box
Several
Several kinds ofof
kindswerePET
PETwere considered.
were In particular,
considered. clear, coloured,
In particular, opaque,opaque,
blue, and box and
samples
samples of PET collected at the recycling plant (25clear, coloured,
samples for each kind). blue,
In orderbox to
of PET were
samples of the collected
PET were at the recycling
collected of plant
at the (25 samples
thedifferent
recyclingkinds, for
plantwe each kind).
(25acquired In
samplesat fororder to investigate the NIR
investigate NIR contribution leasteach
five kind). In order
replicates for eachto
contribution
investigate of the
the NIRdifferent kinds, of
contribution wethe
acquired at least
different fivewe
kinds, replicates
acquired forateach
leastsample
five (for a total
replicates forofeach
659
sample (for a total of 659 spectra) to make spectra comparison and multivariate analysis. Figure A2
spectra) to make
samplerepresentative
(for spectra
a total of 659 comparison and multivariate analysis. and
Figure A2 shows analysis.
representative NIR
shows NIRspectra)
spectra to
of make spectraof
a restriction comparison
two random multivariate
samples for each kind, in Figure
orderA2 to
spectra of a restriction
shows visualize
representative of two random
NIR spectra samples for each kind, in order to better visualize the spectra in
better the spectra in theofplot.
a restriction of two
The overall random
form of thesamples
spectraforwaseach kind, in order
preserved and no to
the plot. The
better visualizeoverall form
the spectraof the spectra was preserved and no significant differences
in the plot. The overall form of the spectra was preserved and no are noted.
significant differences are noted.
significant differences are noted.

A2. Representative NIR spectra of two random

Figure A2. random samples
samples for
for each
each kind
kind of
of PET
PET commodity.
commodity.
Figure A2. Representative NIR spectra of two random samples for each kind of PET commodity.
References
References
1. Plastics Europe. Available online: http://www.plasticseurope.org (accessed on 24 August 2019).
1.
2. Plastics Europe.
Al-Salem, S.M.; Available online:
Lettieri, P.; http://www.plasticseurope.org
Baeyens, J. Recycling and recovery (accessed
routes of on 24 August
plastic solid 2019).
waste (PSW): A
2. Al-Salem, S.M.;
review. Waste Lettieri,
Manag. P.;29,
2009, Baeyens, J. Recycling and recovery routes of plastic solid waste (PSW): A
2625–2643.
review. Waste Manag. 2009, 29, 2625–2643.
Materials 2019, 12, 2740 12 of 13

References
1. Plastics Europe. Available online: http://www.plasticseurope.org (accessed on 19 June 2019).
2. Al-Salem, S.M.; Lettieri, P.; Baeyens, J. Recycling and recovery routes of plastic solid waste (PSW): A review.
Waste Manag. 2009, 29, 2625–2643. [CrossRef] [PubMed]
3. Siddique, R.; Khatib, J.; Kaur, I. Use of recycled plastic in concrete: A review. Waste Manag. 2008, 28,
1835–1852. [CrossRef] [PubMed]
4. Gundupalli, S.P.; Hait, S.; Thakur, A. A review on automated sorting of source-separated municipal solid
waste for recycling. Waste Manag. 2017, 60, 56–74. [CrossRef] [PubMed]
5. Masoumi, H.; Safavi, S.M.; Khani, Z. Identification and classification of plastic resins using near infrared
reflectance spectroscopy. Int. J. Mech. Mechatron. Eng. 2012, 6, 877–884.
6. Masson, J.; Pelletier, L.; Collins, P. Rapid FTIR method for quantification of styrene-butadiene type copolymers
in bitumen. J. Appl. Polym. Sci. 2001, 79, 1034–1041. [CrossRef]
7. Allen, V.; Kalivas, J.H.; Rodriguez, R.G. Post-consumer plastic identification using raman spectroscopy. Appl.
Spectrosc. 1999, 53, 672–681. [CrossRef]
8. Chalmers, J.M.; Everall, N.J.; Measurement, I.C.I. Handbook of Vibrational Spectroscopy: Qualitative and
Quantitative Analysis of Plastics, Polymers and Rubbers by Vibrational Spectroscopy; John Wiley & Sons, Ltd.:
Hoboken, NJ, USA, 2007; ISBN 9780470027325.
9. Florestan, J.; Lachambre, A.; Mermilliod, N.; Boulou, J.C.; Marfisi, C. Recycling of plastics: Automatic
identification of polymers by spectroscopic methods. Resour. Conserv. Recycl. 1994, 10, 67–74. [CrossRef]
10. Hopewell, J.; Dvorak, R.; Kosior, E. Plastics recycling: Challenges and opportunities. Philos. Trans. R. Soc. B
Biol. Sci. 2009, 364, 2115–2126. [CrossRef]
11. Tachwali, Y.; Al-Assaf, Y.; Al-Ali, A.R. Automatic multistage classification system for plastic bottles recycling.
Resour. Conserv. Recycl. 2007, 52, 266–285. [CrossRef]
12. Safavi, S.M.; Masoumi, H.; Mirian, S.S.; Tabrizchi, M. Sorting of polypropylene resins by color in MSW using
visible reflectance spectroscopy. Waste Manag. 2010, 30, 2216–2222. [CrossRef]
13. Herrera, M.; Matuschek, G.; Kettrup, A. Fast identification of polymer additives by pyrolysis-gas
chromatography/mass spectrometry. J. Anal. Appl. Pyrol. 2003, 70, 35–42. [CrossRef]
14. Ulrici, A.; Serranti, S.; Ferrari, C.; Cesare, D.; Foca, G.; Bonifazi, G. Efficient chemometric strategies for
PET-PLA discrimination in recycling plants using hyperspectral imaging. Chemometr. Intell. Lab. 2013, 122,
31–39. [CrossRef]
15. Brunner, S.; Fomin, P.; Kargel, C. Automated sorting of polymer flakes: Fluorescence labeling and development
of a measurement system prototype. Waste Manag. 2015, 38, 49–60. [CrossRef]
16. Ernst, T.; Popp, R.; Van Eldik, R. Quantification of heavy metals for the recycling of waste plastics from
electrotechnical applications. Talanta 2000, 53, 347–357. [CrossRef]
17. Anzano, J.M.; Gornushkin, I.B.; Smith, B.W.; Winefofdner, J.D. Laser-Induced Plasma Spectroscopy for Plastic
Identification. Polym. Eng. Sci. 2000, 40, 2423–2429. [CrossRef]
18. Yan, H.; Siesler, H.W. Identification performance of different types of handheld near-infrared (NIR)
spectrometers for the recycling of polymer commodities. Appl. Spectrosc. 2018, 72, 1362–1370. [CrossRef]
19. Rozenstein, O.; Puckrin, E.; Adamowski, J. Development of a new approach based on midwave infrared
spectroscopy for post-consumer black plastic waste sorting in the recycling industry. Waste Manag. 2017, 68,
38–44. [CrossRef]
20. Jansen, J.J.; Downey, G.; Blanchet, L.; Engel, J.; Gerretzen, J.; Szyman, E. Breaking with trends in pre-processing?
Trend. Anal. Chem. 2013, 50, 96–106.
21. Rinnan, Å.; van den Berg, F.; Engelsen, S.B. Review of the most common pre-processing techniques for
near-infrared spectra. TrAC-Trend. Anal. Chem. 2009, 28, 1201–1222. [CrossRef]
22. Ballabio, D.; Consonni, V. Classification tools in chemistry. Part 1: Linear models. PLS-DA. Anal. Methods
2013, 5, 3790. [CrossRef]
23. Ballabio, D. A Matlab toolbox for principal component analysis and unsupervised exploration of data
structure. Chemometr. Intell. Lab. 2015, 149, 1–9. [CrossRef]
24. Blanco, M.; Villarroya, I. NIR spectroscopy: A rapid-response analytical tool. TrAC-Trend. Anal. Chem. 2002,
21, 240–250. [CrossRef]
Materials 2019, 12, 2740 13 of 13

25. Weyer, L.; Workman, J., Jr. Practical Guide and Spectral Atlas for Interpretive Near-Infrared Spectroscopy;
Workman, J., Jr., Weyer, L., Eds.; CRC Press: Boca Raton, FL, USA, 2012.
26. Crandall, E.W.; Jagtap, A.N. The near-infrared spectra of polymers. J. Appl. Polym. Sci. 1977, 21, 449–454.
[CrossRef]
27. Feldhoff, R.; Wienke, D.; Cammann, K.; Fuchs, H. On-line post consumer package identification by NIR
spectroscopy combined with a FuzzyARTMAP classifier in an industrial environment. Appl. Spectrosc. 1997,
51, 362–368. [CrossRef]
28. Bro, R.; Smilde, A.K. Principal component analysis. Anal. Methods 2014, 6, 2812–2831. [CrossRef]
29. Martinez-Salazar, J.; Sánchez Cuesta, M.; Baltá Calleja, F.J. Equilibrium and non-equilibrium melting of
branched polyethylene relating to defect incorporation within crystals. Colloid Polym. Sci. 1987, 265, 239–245.
[CrossRef]
30. Martínez-Salazar, J.; Cuesta, M.S.; Plans, J. On phase separation in high- and low-density polyethylene
blends: 1. Melting-point depression analysis. Polymer (Guildf) 1991, 32, 2984–2988. [CrossRef]
31. Miller, C.E. Use of near-infrared spectroscopy to determine the composition of high-density/low-density
polyethylene blend films. Appl. Spectrosc. 1993, 47, 222–228. [CrossRef]
32. De Biasio, M.; Arnold, T.; McGunnigle, G.; Leitner, R.; Balthasar, D.; Rehrmann, V. Detecting and
discriminating PE and PP polymers for plastics recycling using NIR imaging spectroscopy. In Proceedings of
the Thermosense XXXII, Orlando, FL, USA, 3 May 2010; Volume 7661, p. 76610V.
33. Barker, M.; Rayens, W. Partial least squares for discrimination. J. Chemom. 2003, 17, 166–173. [CrossRef]
34. Ballabio, D.; Grisoni, F.; Todeschini, R. Multivariate comparison of classification performance measures.
Chemometr. Intell. Lab. 2018, 174, 33–44. [CrossRef]
35. Willard, H.H.; Merritt, L.L.J.; Dean, J.A.; Settle, F.A.J. Instrumental Methods of Analysis, 7th ed.; Wadsworth
Pub Co.: Belmont, CA, USA, 1988.

© 2019 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/).