Environmental Science and Pollution Research (2019) 26:7364–7374

https://doi.org/10.1007/s11356-018-2180-2

ADVANCEMENTS IN CHEMICAL METHODS FOR ENVIRONMENTAL RESEARCH

High-throughput NIR spectroscopic (NIRS) detection

of microplastics in soil
Andrea Paul 1 & Lukas Wander 1 & Roland Becker 1 & Caroline Goedecke 1 & Ulrike Braun 1

Received: 29 November 2017 / Accepted: 30 April 2018 / Published online: 12 May 2018
# Springer-Verlag GmbH Germany, part of Springer Nature 2018

Abstract
The increasing pollution of terrestrial and aquatic ecosystems with plastic debris leads to the accumulation of microscopic plastic
particles of still unknown amount. To monitor the degree of contamination, analytical methods are urgently needed, which help to
quantify microplastics (MP). Currently, time-costly purified materials enriched on filters are investigated both by micro-infrared
spectroscopy and/or micro-Raman. Although yielding precise results, these techniques are time consuming, and are restricted to the
analysis of a small part of the sample in the order of few micrograms. To overcome these problems, we tested a macroscopic
dimensioned near-infrared (NIR) process-spectroscopic method in combination with chemometrics. For calibration, artificial MP/
soil mixtures containing defined ratios of polyethylene, polyethylene terephthalate, polypropylene, and polystyrene with diameters
< 125 μm were prepared and measured by a process FT-NIR spectrometer equipped with a fiber-optic reflection probe. The resulting
spectra were processed by chemometric models including support vector machine regression (SVR), and partial least squares
discriminant analysis (PLS-DA). Validation of models by MP mixtures, MP-free soils, and real-world samples, e.g., fermenter
residue, suggests a reliable detection and a possible classification of MP at levels above 0.5 to 1.0 mass% depending on the polymer.
The benefit of the combined NIRS chemometric approach lies in the rapid assessment whether soil contains MP, without any
chemical pretreatment. The method can be used with larger sample volumes and even allows for an online prediction and thus meets
the demand of a high-throughput method.

Keywords Microplastics . Soil . Chemometrics . PLS-DA . Support vector machines . Near-infrared spectroscopy

Introduction materials originating from everyday products such as cosmetics

and scrubbing agents are referred to as primary MP, whereas
The increasing pollution of terrestrial and aquatic ecosystems particles formed as the outcome of material aging by a combina-
with plastic debris leads to the accumulation of microscopic plas- tion of UV radiation, mechanical, and oxidative degradation are
tic particles of still unknown amount and fate. Microplastic par- called secondary MP (Boucher and Friot 2017). While it is wide-
ticles or microplastics (MP) are defined as plastic particles with ly acknowledged that plastic fragments and MP in the ocean are a
diameters smaller than 5 mm and can be differentiated according serious issue with unclear effects on marine organisms (Kooi
to their sources into primary and secondary MP. Particles intro- et al. 2017; Herzke et al. 2016), information about MP in terres-
duced directly into the environment in form of small-size trial environments is fragmentary (Nizetto et al. 2016). The large
variability of soil organic matter and mineral constituents makes
the chemical identification of plastic residues in soil to a specific
Responsible editor: Philippe Garrigues
challenge (Blaesing and Amelung 2018). Fuller and Gautam
Electronic supplementary material The online version of this article (2016) recently detected 0.03 to 6.7% of plastic in soils of an
(https://doi.org/10.1007/s11356-018-2180-2) contains supplementary industrial area. Next to the introduction of particles by tyre abra-
material, which is available to authorized users.
sion, the application of compost, sewage sludge, or solid manure
could be a source of contamination. Furthermore, plastic
* Andrea Paul
[email protected] mulching, mostly by PE foils, irrigation of agricultural fields
by sometimes even untreated waste water, littering near roads
and trails, and atmospheric fallout may contribute to accumu-
1
Bundesanstalt für Materialforschung und –prüfung,
Richard-Willstätter-Str. 11, 12489 Berlin, Germany lation of MP in soil as well. A summary of known pathways of
Environ Sci Pollut Res (2019) 26:7364–7374 7365

introduction of MP into soils and estimated MP loads is pro- MP, quicker upstream technologies are still of great interest. One
vided by Blaesing and Amelung (2018). should be aware that the largest contribution to the uncertainties
Hence, suitable techniques are needed to get a more detailed of the obtained results is often caused by the process of sampling
overview on the source, the occurrence, and the fate of MP in and not the analysis itself (Szczendzina et al. 2009).
soils. At present, a number of common techniques for the iden- Consequently, it is important to analyze a representative number
tification and quantification of MP in solid environmental sam- of samples out of a larger stock to obtain reliable results.
ples is available (Hidalgo-Ruz et al. 2012; Vianello et al. 2013; In search for an approach, which may fulfill the high-
Imhof et al. 2016; Kaeppler et al. 2016). The actual analysis of throughput criterion together with an easy handling, we looked
material for MP includes primarily microscopic and micro-spec- for other technologies. High-throughput or even online measure-
troscopic, i.e., micro-Raman or micro-FTIR spectroscopic imag- ments are of central importance in process analytics, typically
ing methods on carefully prepared, cleaned, and up-concentrated using robust spectrometers which, nevertheless, producing reli-
materials. Infrared microscopy is one of the most powerful and able results over longer periods. Process spectrometers are easily
widely available techniques for the reliable detection and identi- operated and can be combined with a multitude of fiber optical
fication of MP in environmental samples, especially when probes that allow for the measurement of almost any type of
equipped with a focal plane array (FPA) detector. Both the sample and application. Beside the robustness of the measure-
micro-Raman and micro-FTIR techniques rely on the spectral ments, a major advantage of these methods is the short measure-
recording, mapping, and subsequent identification of vibrational ment time or even application as online analysis.
bands which are typical for synthetic polymers. However, a ma- The use of process Raman spectroscopy for the identification
jor precondition of these microscopic techniques, except for IR of MP in soil turned out as a promising approach for coastal
transmission, relates to the fact that MP can only be identified sands containing particles ≥ 2 mm (Paul et al. 2016), but failed
and studied in the absence of contaminating organic materials. In for lower particles sizes and non-coastal soils containing higher
the case of Raman spectroscopy, natural organic matter further- amounts of organic carbon. Extended studies applying a time-
more deteriorates spectra by a high fluorescing background gated Raman method that allows for the elimination of the fluo-
which exceeds the polymer signals by orders of magnitude. All rescence background (Kostamovaara et al. 2013) could detect
in all, despite delivering valuable information on size, number, polymer particles ≤ 125 μm at mass fractions ≥ 5 mass% but
and degree of aging, microscopic methods are time demanding, failed for lower concentration (Paul et al. 2017).
and provide information on small sample aliquots only (Elert In this study, we tested NIR spectroscopy (NIRS), which is
et al. 2017). Furthermore, typical pretreatments of soil are needed most common in process analytics. Whereas IR spectroscopy
including density fractionation, ultrasonic, acidic, alkaline or ox- comprises the spectral range 600–4000 cm−1 and mainly records
idizing treatment, electrostatic separation, and sequential extrac- the well-resolved fundamental molecular vibrations, the NIR fills
tion (Blaesing and Amelung 2018). the gap between the IR and the UV/Vis part of the electromagnetic
In addition to the microscopic techniques, thermal extraction spectrum, from 4000 to 15000 cm−1 (667 to 2500 nm). NIR
methods are an alternative (Fabbri 2001; Fries et al. 2013). spectra are characterized by the broad overlapping bands of over-
Although such pyrolysis techniques provide comprehensive in- tone and combination vibrations for a limited number of molecu-
formation about the plastic particles (type, additives, aging status), lar vibrations, mostly of the type X–H, e.g., C–H, O–H, N–H. The
these methods are hardly applicable for MP analysis, because the spectral evaluation in the sense of interpreting molecular structures
particles must be separated individually beforehand. A preceding from NIR spectra is in general poor. However, when it comes to
sample preparation reducing the natural organic fraction is possi- classification tasks or quantitative analyses, NIRS in combination
ble, but is laborious and assumes the risk of contamination or with chemometric methods is a powerful tool. Moreover, due to
partial MP particle degradation (Fischer and Scholz-Bottcher the low absorption coefficients of the higher overtones compared
2017). A compromise is the recently developed thermal extraction to fundamental vibrations, NIR can penetrate deeper than IR and
desorption gas chromatography mass spectrometry (TED-GC- thus handle larger sample volumes. These capacities, together
MS) (Duemichen et al. 2017) or chemical extraction with subse- with a much lower sensitivity to water and the possibility to work
quent analysis such as size exclusion chromatography (SEC) or with quartz materials for fibers and optical elements, render NIRS
FTIR has been demonstrated (Elert et al. 2017; Fuller and Gautam an ideal technology for our application. NIRS can be used for the
2016). TED-GC-MS, for example, allows the precise quantifica- classification of most common synthetic polymers and is for ex-
tion of polyethylene (PE), polyethylene terephthalate (PET), poly- ample applied for the automated sorting of plastic waste (van den
propylene (PP), and polystyrene (PS) in 20 to 100 mg samples Broek et al. 1998), often in combination with hyperspectral im-
within 3 h without any further preparation. SEC, in contrast, is aging (Moroni et al. 2015). Karlsson et al. (2016) demonstrated
restricted to the quantification of PS and PET in soil matrix, but the potential of three different hyperspectral imaging devices
can handle sample volumes in the order of 500 mg within 2 h. working in the NIR in combination with multivariate data analysis
Although some of the previously described thermal extraction for the mapping of MP on filters. However, in those experiments,
approaches mean a further step towards an efficient screening for the polymer particles were not embedded in the soil matrix.
7366 Environ Sci Pollut Res (2019) 26:7364–7374

Our approach aims at the development of an easy-to-use, To test for the effect of particle size on NIR spectra,
high-throughput method that allows a quick estimate on LUFA2.3 was investigated as obtained (LUFA2.3_raw),
whether or not a certain sample is contaminated with (micro) sieved, and after additional milling (LUFA2.3_milled) for
plastics. In this study, we will demonstrate the general feasi- 10 min using a cryo-mill (Retsch CryoMill) cooled with liquid
bility of a chemometric evaluation of NIR spectra of MP in nitrogen. If not stated otherwise, BLUFA2.3^ always refers to
soil which could be used as a screening method to identify the standard sieved material. A micrograph of LUFA2.3 is
batches that may be rendered a more detailed examination by provided in Fig. 1. The micrograph was obtained from a
alternative techniques which allows the determination of num- Zeiss AxioScope.A1 MAT microscope equipped with an
ber, size, and aging state of polymers. AxioCam MRc camera and × 5/NA = 0.13, × /100.25, × 20/
0.40, and × 50/0.90 objective lenses. The system was used in
reflected light darkfield illumination mode. The according in-
Materials and methods strument software (Zeiss ZEN) was employed to measure par-
ticle sizes. Micrographs of all fractions of LUFA2.3 are com-
Materials pared in Fig. S2 in the supplementary materials.

Preparation of microplastics Real-world samples

MP were prepared using industrial pellets (1–3 mm) of PE, PP, Fermenter residues were prepared as described by Duemichen
PS, and PET, kindly provided by plastic manufacturers with et al. (2017) containing PE, PP, and PS. An aliquot of this sample
the support of Plastics Europe Deutschland e.V. and contained was mildly H2O2 treated. Washing machine filter material was
no special stabilization additives. Pellets of each polymer were collected in an effluent filter containing both PP and polyamide
separately rinsed, first with methanol and then with water, and (PA) fibers next to hair, bristles, silk, cotton, wool, and cellulose
after drying and embrittlement with liquid nitrogen cryo- (pers. communication U. Braun, data not published). Domestic
milled using a ZM 200 centrifugal mill (Retsch, Haan, compost was collected by the author and used after drying and
Germany) equipped with a 2-mm ring sieve. After air-drying sieving (mesh size 100 μm). The polymer content in all real-
and sieving, the respective fractions below 125 μm were col- world samples was identified by use of TED-GC-MS, in the
lected and used for the further preparation of soil mixtures. latter the PS signal was close to the lower detection limit.
Micrographs of the obtained material are provided by Elert
et al. (2017). In addition, previously aged PP (PP_aged, Preparation of MP-soil mixtures
Bajer and Braun 2014) was milled for 10 min using a cryo-
mill (Retsch CryoMill) cooled with liquid nitrogen. Finally, Soil LUFA2.3 was added defined mass fractions of the cryo-
polyvinyl chloride (PVC) particles (averaged particles size milled polymers in the range 0.5–5 mass%, and was after-
~150 μm) were used as obtained from the manufacturer wards homogenized. These mixtures contained milled mate-
(Plastics Europe Deutschland e.V.). rial of (i) a single polymer, (ii) two polymers in equal amounts
of about 1 mass%, and (iii) four polymers, i.e., PE, PP, PET,
Soils and PS, each at 1 mass%. The latter is referred to as BAM
Microplastics Reference Material (BAM-RM) and has been
The air-dried standard soil type 2.3 (eutric fluvisol, organic characterized by Elert et al. (2017).
carbon content 0.67 mass%, LUFA Speyer, Speyer, Germany, All samples were kept open under ambient conditions if not
cf. Table 1) was sieved and the fraction below 125 μm was stated otherwise.
collected. This material is the basis for the soil-MP mixtures
used for calibration of chemometric models and will be re- Thermogravimetric analysis and TED-GC-MS
ferred to as LUFA2.3. A molic fluvic claysol containing a
higher humic fraction from the same supplier (LUFA2.2), To determine the amounts of organic and inorganic constituents
and an excavated urban medium sand, prepared by BAM of the selected soil samples, thermogravimetric analysis was per-
2009, completed the set of calibration samples. For validation, formed. Sixty milligrams of the soils were weighted in a 150-μl
agricultural MP-free soils from the BONARES project, which alumina pan and were placed in the oven of the TGA/SDTA 851
have been used in previous studies in different context (Barej single-arm thermobalance (Mettler-Toledo GmbH, Gießen,
et al. 2014; Heggemann et al. 2017), were included. A descrip- Germany). The sample was kept at 25 °C for 5 min and heated
tion of all soils including organic carbon contents as deter- up to 600 °C at 10 °C/min under a flow of nitrogen (30 mL
mined by thermogravimetric analysis (TGA) is provided in min−1). Subsequently, the valve was switched to synthetic air
Table 1. More detailed information on TGA is summarized and the sample was further heated up to 900 °C with the same
in the supplementary materials in Table S1 and Fig. S1. heating rate and gas flow rate.
Environ Sci Pollut Res (2019) 26:7364–7374 7367

Table 1 Soil materials used for calibration (C), production of MP mixtures, and as validation samples (V)

soil usage Soil description OC mass% Supplier Treatment

LUFA2.3 C Eutric fluvisol 0.67 LUFA Speyer Sieved and homogenized fraction
Germany below 125 μm, used for MP mixtures
LUFA2.3_raw C Eutric fluvisol 0.67 LUFA Speyer Raw material
Germany
LUFA2.3_milled C Eutric fluvisol 0.67 LUFA Speyer Milled (agglomerates of fine
Germany material ranging from 3 to 6.5 μm)
LUFA2.2_raw C Mollic fluvic vertisol 1.61 LUFA Speyer Raw material
Germany
Soil_B4 C Hortic-like anthrosol 0.5*) BAM 2009 Sieved and homogenized fraction
(medium sand) below 2 mm
Soil_B5-G1 V Cambisol (sandy) 1.7*) BONARES Sieved and homogenized fraction
0–30 cm below 2 mm
Soil_B6-G2 V Stagnosol (clayey) 4.2*) BONARES Sieved and homogenized fraction
0–30 cm below 2 mm
Soil_B7-G3 V Haplic luvisol (loamy) 2.0*) BONARES Sieved and homogenized fraction
(hypereutric, siltic), below 2 mm
AP 0–30 cm
Soil_B8-G4 V Haplic luvisol (loam) 1.3*) BONARES sieved and homogenized fraction
(hypereutric, siltic) below 2 mm
B 0–150 cm
Soil_B9-A1 V Dystric cambisol, 0.6*) BONARES Sieved and homogenized fraction
0–25 cm below 2 mm
*)
Estimated from TGA

The TED-GC-MS (combination of thermogravimetric 3 mm, Solvias) at room temperature (22–24 °C) using a
analysis and thermal desorption gas chromatography mass Spectralon standard for background measurements. Each
spectrometry) of selected samples was performed using spectrum was scanned over the spectral range 12000 to
40 mg of the sample substance. Further details about the meth- 4000 cm−1 with a resolution of 8 cm−1 and a zero fill of 2
od and the parameters are described in the publication of and transformed into absorbance values. Per run, 64 spectra
Duemichen et al. (2017). were accumulated and five or 10 repetitive measurements
were performed per sample with a period of 30 s between
NIR spectroscopy the measurements. For that purpose, the probe each time
was set on a different position of the powdered sample in an
NIR spectra were measured with a Matrix-F (Bruker) spec- aluminum cap of 30 mm diameter containing 1 g of material
trometer, equipped with a fiber-optic reflection probe (d = (cf. Fig. S3, supplementary materials). All samples were inde-
pendently investigated at least twice by different operators.

Chemometrics

NIR spectra were evaluated by using the software The

Unscrambler X, Vers. 10.5 (CAMO software). Data analysis
was in general performed with mean-centered spectra. Further
pretreatments included standard normal variate transformation
(SNV) or multiple scatter correction (MSC), smoothing by a
second-order polynomial (Savitzky-Golay), offset and/or lin-
ear baseline correction, first derivative transformation, and
always as a final step, unit vector normalization in a restricted
spectral range (5300 to 6100 cm−1).
Principal component analysis (PCA) was used both for ex-
plorative data analysis. PCA comprises the decomposition of a
Fig. 1 Light micrograph of sieved and homogenized fraction below data matrix X consisting of the spectral intensities of all included
125 μm of soil LUFA2.3 spectra (x variables), into a bilinear model of scores and loadings
7368 Environ Sci Pollut Res (2019) 26:7364–7374

and the unexplained rest, i.e., the residuals. Whereas the loadings unlike regression, the root mean square error of prediction
contain the information on the importance of the spectral vari- (RMSEP) is no figure of merit, but instead sensitivity and spec-
ables, the scores provide a description of the projection of the ificity may be applied. Using the labels Bpositive^ and
individual samples, i.e., spectra, into the new space generated by Bnegative,^ in a binary classification problem there are four pos-
the principal components. PCA can be considered as a lower sible outcomes: for correct classification, one obtains true-
dimensional approximation of X. positive and true-negative results, and for incorrect classification
Partial least squares regression (PLSR) is a bilinear procedure false-positive, and false-negative ones. Sensitivity is defined as
based on the reduction of dimension by the projection of both x the ratio of true positives and the sum of true positives and false
and y variables into a new space maximizing the covariance negatives, and specificity as the ratio of true negatives and the
between X and Y. Both the X and Y matrices are modeled sum of true negatives and false positives. In an ideal case, when
simultaneously to find the latent variables in X that will best all predictions are correct, both sensitivity and specificity reach
predict the latent variables in Y. The resulting PLS components the value B1^ and 100%, respectively.
are called factors. Partial least squares discriminant analysis
(PLS-DA) is a PLSR-based classification method. Here, we used
for MP-containing samples the response variable (y) of B1^; Results
whereas for MP-free soils, y was adjusted to B0.^ Model optimi-
zation was performed by cross validation over 20 randomly cho- NIR spectra of MP and soils
sen segments with mean centering inside the validation segments
as described by Eide et al. (2010). Typical NIR spectra of MP preparations in LUFA2.3 in the spec-
Support vector machines (SVM) is a binary linear classifier tral range 12000 to 4000 cm−1 are shown in Fig. 2a. After pre-
that optimizes a classification hyperplane between the surface treatment including SNV transformation, smoothing, and base-
data points of two clusters in the data space (Cortes and Vapnik line correction, the spectral features of the tested polymers can be
1995; De Boves Harrington 2015). In contrast to the previously clearly distinguished in the spectral range 5300 to 6100 cm−1 (cf.
described algorithms based on latent variables, in support vector Fig. 2b). The spectrum of the BAM-RM containing 1 mass% of
machines regression (SVR), dimensionality is increased using the polymers PE, PET, PP, and PS, summing up to 4 mass%, can
suitable Kernel functions. Support vector machine regression be understood as an envelope of the four components. The pri-
(SVR) models were established by cross validation with 20 ran- mary NIR active modes of PS belong to the first overtone vibra-
domly chosen segments using the same coding of samples as tions of aromatic C–H bonds with a maximum at about
described for PLS-DA. 5950 cm−1 (Workman and Weyer 2012). The spectrum of PET
Corresponding to the calibration strategy, prediction results of is characterized by a predominant absorption at approx.
B1^ and B0^ for Bunknown^ samples consequently assign spec- 6020 cm−1 (Workman and Weyer 2012). According to Miller
tra belonging to MP-containing and MP-free samples using the and Eichinger (1990), this is due to first overtone stretching of
PLS-DA and SVR models. Problematic is the question of how to aromatic C–H groups overlapped with first overtone vibrations
deal with intermediate values. According to the software recom- of methylene stretching at 5810 cm−1. In good agreement with
mendation, (i) samples with ŷi > 0.5 and a deviation that does not computational and experimental studies (Mizushima et al. 2012),
cross the 0.5 line are predicted members; (ii) samples with ŷi < the PE spectrum possesses two major absorption bands in the
0.5 and a deviation that does not cross the 0.5 line are predicted spectral range at 5790 and 5670 cm−1. In contrast to the domi-
non-members; and (iii) samples with a deviation that crosses the nating CH2 in PE, the first overtones of PP are more complex
0.5 line cannot be safely classified. resulting from the C–H vibrations of CH, CH2, and CH3 groups,
For PLS-DA and SVM model development, both the root with the two most prominent absorptions at 5890 cm1 and
mean square errors of calibration (RMSEC) and cross validation 5790 cm−1. In Fig. 2c, NIR spectra of the real-world samples
(RMSECV) were used. Best prediction performance is expected washing machine filter effluent, fermenter residue, and domestic
from models with small values of RMSEC and RMSECV. In (1), compost are provided weakly resembling the appearance of the
yi represents the known (e.g., from the gravimetric preparation BAM-RM. Besides polymer-specific differences, the spectra in
procedure) contents of the analyte in the calibration samples, ŷi Fig. 2a share a common spectral signature defined by the refer-
the predicted content values, and n the number of samples. ence soil used as matrix. Most noticeable are two NIR-bands at
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7090 and 5230 cm−1 flanking the region of interest highlighted in
∑ni¼1 ðyi −^yi Þ2 Fig. 2b–d, most likely caused by water residues in the soil.
RMSE ¼ ð1Þ However, other sources of similar absorption bands have been
n
discussed in the literature such as surface bound −OH groups and
To assess the prediction ability of chemometric models, vali- organic matter as summarized by Roberts and Cozzolino (2016).
dation by new samples having not been included in the calibra- Figure 2d summarizes the NIR spectra of soils used in the vali-
tion data set must be performed. However, in classification, dation set. The effect of grain size on the NIR spectra of the soil
Environ Sci Pollut Res (2019) 26:7364–7374 7369

0,35
a 0,5 b
0,30
0,4

SNV(absorbance)
absorbance

0,25 0,3

0,20 0,2

0,1
0,15

0,0
0,10
12000 11000 10000 9000 8000 7000 6000 5000 4000 6100 6000 5900 5800 5700 5600

0,5 c d
0,10
0,4
SNV(absorbance)

SNV(absorbance)
0,3

0,05
0,2

0,1

0,0 0,00

6100 6000 5900 5800 5700 5600 6100 6000 5900 5800 5700 5600

wavenumber / cm -1 wavenumber / cm -1
Fig. 2 NIR spectra: a raw spectra (stacked): 1 mass% mixtures of PE samples: fermenter residue (green), fermenter residue -ox (dark green),
(dark green), PET (red), PP (light-green) or PS (blue) in soil, and soil washing machine filter (blue), domestic_compost (red), compared to
LUFA2.3 (broken line); b processed spectra: First overtones (region of BAM-RM (broken line); d pure soils: ROI of soils LUFA2.3 (broken
interest, ROI) of 1 mass% mixtures of PE (dark green), PET (red), PP line), B5_G1 (red), B6_G2 (dark blue), B7_G3 (marine), B8_G4 (green),
(light-green), PS (blue), and BAM-RM (broken line), c ROI of real-world and B9_A1 (cyan)

LUFA2.3 and the natural variability of samples as obtained from of this material after mild H2O2 treatment, washing machine filter
three different batches of LUFA2.3 is illustrated in Fig. S4 in the material (containing PA and PP next to other organic fibers), and
supplemental materials. domestic compost, sieved and homogenized containing some PS
close to the lower detection limit of TED-GC-MS.
Identification of MP in soil by using chemometric At first, PCA was performed with cross validation over 20
models randomly chosen segments using the previously described cali-
bration spectra after mathematical pretreatment including SNV
Classification of microplastics transformation, smoothing (Savitzky-Golay using a second-order
polynomial over 13 spectral data points), linear baseline correc-
For the calibration of the chemometric models, we used 40 spec- tion, and unit vector normalization in the spectral range 5300 to
tra of soils without MP and 55 spectra of MP-containing mix- 6100 cm−1. Seven principal components (PCs) cover 98% of
tures providing a comparable number of samples in both groups. both calibration and validation variance. The loadings of the first
The calibration set comprised the spectra of MP-soil mixtures four principal components PC1–PC4 in Fig. 3a explain 46, 33, 7,
containing 1 mass% of single cryo-milled polymers prepared in and 5% of the variance. Whereas PC1 accounts mostly for a
LUFA2.3 and two out of three BAM-RM samples containing a separation between MP-free and MP-containing samples, PC2
mixture of 1 mass% of four different polymers in LUFA2.3. The to PC4 can be attributed to the spectral information of the poly-
soils LUFA2.3, LUFA2.3_raw, LUFA2.3_milled, mers. The scores plot of the first three PCs in Fig. 3b shows a
LUFA2.2_raw, and Soil_B4 (medium sand, cf. Table 1) were general trend of clustering on PC1, but no clear separation of
included as MP-free samples. MP-containing samples is obtained, which is here basically due
For the validation of the obtained chemometric models, we to the PE mixtures coinciding with the MP-free soils.
used three types of test samples: (i) gravimetrically prepared soil- PLS-DA of the same data set yielded a model which is based
MP mixtures, (ii) MP-free soils, and (iii) so called real-world on four factors explaining altogether 91 and 89% of calibration
samples, e.g., preparations from fermenter residue (containing and validation variance in random cross validation (cf. Fig. S5).
PE, PS, and PP next to other organic rest materials), an aliquot However, prediction results for the validation samples are not
7370 Environ Sci Pollut Res (2019) 26:7364–7374

0,3 a To assess the minimum MP concentration that could be iden-

tified by the combined NIR-SVR method, soil mixtures contain-
0,2 ing 0.5–5 mass%, of the cryo-milled polymers were tested. The
loading / a.u.

samples containing the mass fraction of 1% polymer were fresh-

0,1
ly prepared for that purpose and are not identical to those used
for calibration. As can be seen from the confusion matrix in
Table 3, except PS (0.5 and 1 mass%) and PET (0.5 mass%),
0,0
microplastics could be detected in all soil mixtures by the SVR
model. One should note that multivariate data analysis does not
-0,1
allow for extrapolation. However, here, we deliberately tested
the 0.5% samples and got a correct classification in four out of
-0,2
six cases. Next to the 0.5% sample, also the 1% mixture of PS in
6100 6000 5900 5800 5700 5600 LUFA2.3 yielded no distinct classification with a prediction of
-1 0.57. Moreover, an exceptional high standard deviation of 0.15
wavenumber / cm
was obtained. Inspection of the 10 individual spectra comprised
prediction values ranging from 0.36–0.78. We assume that the
observed variation is caused by an untypical poor mixing of soil
and MP within that sample highlighting the importance of prop-
er homogenization. Consequently, in case of missing classifica-
tion, together with high standard deviation, repeated measure-
ment and/ or inspection of the individual spectra is recommend-
ed. We further tested, how many spectra would be needed to
enable a detailed investigation of a nonhomogeneous sample,
and found about 20 spectra/ 100 mg would be necessary using
the reflection probe with a 3 mm spot size. Next to homogeneity,
sample conditioning includes drying of wet materials with sub-
sequent storage under dry conditions. For the prediction of MP
in the validation samples presented in Tables 2 and 3 by the SVR
model, a sensitivity of 92% together with a specificity of 100%
was obtained.

Fig. 3 a Loadings for the first four principal components (PCs): PC1 Classification of the polymers
(black), PC2 (red), PC3 (blue), PC4 (green), and b scores for PC1–PC3
(red – MP-containing samples, green - soils) as obtained from the PCA of To identify the polymeric material in soil, classification models
NIR spectra used for calibration of the PLS-DA and SVR model
using support vector-based models and PLS-DA were tested.
Here, SVM yielded no satisfactory results (not shown), but
satisfying as can be seen from Table 2. Whereas MP-free soils PLS-DA seemed promising. In that case, we used a PLS2 ap-
and samples prepared as mixtures of MP with LUFA2.3 could be proach, i.e., instead of using a vector y for the response variable,
correctly classified, classification of the real-world samples we now introduced the matrix Y containing the information on
failed. Samples studied in two independent campaigns in all the five polymers PE, PET, PP, PS, and PVC by the coding with
cases yielded comparable prediction results. Based on the 11 B1^ and B0^ indicating the presence of the respective polymer.
different samples, excluding the duplicates, the obtained sensi- Best classification results were obtained with spectra after a full
tivity was low with a true-positive rate of only 33%, whereas multiple scatter correction (MSC), followed by a first derivative
specificity amounted to 100%. Savitzky-Golay transformation over nine spectral data points
In a third approach, we tested classification of MP by SVR using a second-order polynomial, and unit vector normalization
using the same spectra as used in PCA and PLS-DA and coding in the spectral range 1530 to 1692 cm− 1. To enable at least a
for the response variable with B1^ and B0.^ Best results were decent number of training samples for each polymer, in addition
obtained by an epsilon-type SVR with a radial basis Kernel to the calibration spectra used before, binary mixtures of all
explaining 95% of the X variance in cross validation (parameters polymers PE, PET, PP, and PS at the 1% level together with
γ = 0.01, C = 12.92, ε = 0.1). The details of the SVR model are 1% soil-MP mixtures of PE, PET, PP, PS, and PVC in Soil_B4
summarized in Fig. S6 in the supplementary materials. Using (medium sand) were prepared for model development. The
this SVR model, all validation samples could be successfully resulting PLS-DA classification model is based on 185 spectra
classified (Table 2). from 26 samples representing all polymers except PVC at least
Environ Sci Pollut Res (2019) 26:7364–7374 7371

Table 2 MP in validation
samples as determined by using Sample name Target Sample type Polymer Predicted by Predicted by
PLS-DA or SVR. Target values value y contained PLS-DA ŷ ± s SVR ŷ ± s
(y) for MP-free and containing
samples are 0 and 1, predicted BAM-RM (10) 1 MP mixture PE, PET, PP, PS 1.09 ± 0.07 1.01 ± 0.04
values (ŷ) are provided with stan- 0.5%_PP-aged (10) 1 MP mixture PP 1.10 ± 0.05 1.99 ± 0.06
dard deviation (s) obtained from Soil_B5_G1 (5) 0 MP-free soil – − 0.11 ± 0.10 0.27 ± 0.04
averaging of 5 or 10 spectra (cf.
number in brackets provided with Soil_B5_G1 (10) 0 MP-free soil – 0.20 ± 0.04 0.04 ± 0.05
sample name) Soil_B6_G2 (5) 0 MP-free soil – − 0.13 ± 0.05 0.14 ± 0.03
Soil_B6_G2 (10) 0 MP-free soil – 0.17 ± 0.06 0.01 ± 0.03
Soil_B7_G3 (5) 0 MP-free soil – 0.02 ± 0.04 0.04 ± 0.05
Soil_B7_G3 (10) 0 MP-free soil – 0.34 ± 0.04 0.03 ± 0.06
Soil_B8_G4 (5) 0 MP-free soil – − 0.43 ± 0.04 0.10 ± 0.15
Soil_B8_G4 (10) 0 MP-free soil – 0.28 ± 0.03 − 0.14 ± 0.05
Soil_B9_A1 (5) 0 MP-free soil – 0.39 ± 0.07 0.31 ± 0.11
Soil_B9_A1 (10) 0 MP-free soil – 0.37 ± 0.06 0.19 ± 0.08
Fermenter residue (5) 1 Natural MP PE, PP, PS 0.40 ± 0.10 0.69 ± 0.16
Fermenter residue (10) 1 Natural MP PE, PP, PS 0.34 ± 0.11 0.82 ± 0.09
Fermenter residue-ox (5) 1 Natural MP PE, PP, PS 0.30 ± 0.05 0.69 ± 0.06
Fermenter residue-ox (10) 1 Natural MP PE, PP, PS 0.43 ± 0.02 0.78 ± 0.04
Washing machine filter (5) 1 Natural MP PP, PA 0.48 ± 0.03 0.96 ± 0.02
Washing machine filter (10) 1 Natural MP PP, PA 0.39 ± 0.02 0.95 ± 0.02
Domestic compost (10) 1 Natural MP PS 0.42 ± 0.02 0.67 ± 0.05

in four different samples. An illustration of the model together PP in domestic compost. For the calculation of sensitivity and
with regression coefficients, scores, and correlation loadings is specificity, these non-classified samples were added to the num-
provided in Fig. S7 in the supplementary material. The predic- ber of false positives, altogether yielding sensitivity and speci-
tion results for the validation samples using seven factors are ficity of 83 and 90%. In most cases, misclassification occurred
summarized in the confusion matrix in Table 4. For comparison, with PET which could be caused by a spectral overlap of that
in the last column, also the prediction results of the SVR model polymer and vibrational bands soils in the spectral region 6000–
indicating the presence of MP are added. With few exceptions, 6100 cm−1 (cf. Fig. 2).
polymers were correctly identified in all samples. In three cases,
polymers could not be detected, namely PET in the BAM-RM,
PVC in the 0.5% mixture, and PS in the oxidized fermenter
residue. One false-positive classification occurred for PE in the Discussion
washing machine filter, and in five cases no explicit classifica-
tion was obtained. Three of those cases refer to PET, e.g., in In contrast to the classical, microscope-based detection and
Soil_B5_G1 (cambisol, sandy), and the fermenter residue sam- classification of single MP particles, we aim at an overview
ples, one refers to PS in Soil_B5_G1, and the last one refers to on whether MP is contained or not in a certain sample volume.
In a first step of the decision tree, an unknown sample shall be
Table 3 Confusion matrix for the prediction of MP in mixtures of tested for the presence of MP, and in the second step, the type
LUFA2.3 or Soil_B4 (medium sand)*) of contained polymer(s) shall be identified. Originally, we also
aimed at a quantitative assessment of how much MP may be in
Mass fraction Polymer
a sample. However, trials to achieve a sensitive and specific
% PE PET PP PS PVC quantification were not successful so far, dose-response
0.5 1 0 1 0 1 curves were non-linear and standard errors high (not shown).
1 1 1 1 0.5 1 We assume that for a proper estimation of concentration, well-
1*) 1 1 1 1 1 defined particle size distributions will be a major precondition
2 1 1 1 1 1 demanding for extended calibration strategies.
3 1 1 1 1 1 Despite that point, the combined NIR spectroscopic chemo-
4 1 1 1 1 1 metric approach provides a complementary method with the
5 1 1 1 1 1 advantage of a comparable high throughput. For a well-homog-
enized, dried sample of 1 g screening by 10 spectra takes no
7372 Environ Sci Pollut Res (2019) 26:7364–7374

Table 4 Confusion matrix for the

prediction of polymers in Sample name Polymers PE PET PP PS PVC MP
validation samples obtained from
10 averaged spectra. In the right BAM-RM PE, PET, PP, PS 1 0 1 1 0 1
column, the prediction of MP by 0.5%_PP-aged PP 0 0 1 0 0 1
the SVR model is provided for 0.5% PE PE 1 0 0 0 0 1
comparison. False or missing
classification is indicted by 0.5% PET PET 0 1 0 0 0 0
italicized numbers 0.5% PP PP 0 0 1 0 0 1
0.5% PS PS 0 0 0 1 0 0
0.5% PVC PVC 0 0 0 0 0 1
Soil_B5_G1 – 0 0.5 0 0.5 0 0
Soil_B6_G2 – 0 0 0 0 0 0
Soil_B7_G3 – 0 0 0 0 0 0
Soil_B8_G4 – 0 0 0 0 0 0
Soil_B9_A1 – 0 0 0 0 0 0
Fermenter residue PE, PP, PS 1 0.5 1 1 0 1
Fermenter residue-ox PE, PP, PS 1 0.5 1 0 0 1
Washing machine filter PP, PA 1 0 1 0 0 1
Domestic compost PS 0 0 0.5 1 0 1

longer than 5 min including sample placement in a measure- mixture of soil and MP, of one or more polymers, represents a
ment well. In case of a nonhomogeneous sample, an aliquot of unique sample. Hence, the usage of single polymer spectra from
100 mg in a linear arrangement can be measured altogether with spectral libraries is not helpful. Our spectral library therefore was
20 spectra using a probe spot of 3 mm as applied here established by the measurement of well-defined mixtures of soils
amounting to 10 min of experimental time. and cryo-milled polymers. Such samples can only serve as
Following the recording of the NIR spectra, prediction of models for real-world samples which may behave differently
MP and polymer type is possible within seconds by using the due to aging, the presence of mixed materials, or surface cover-
previously calibrated chemometric models. No information, age, as may occur in contact with natural organic matter (Chen
however, is gained on the quantity of MP. To obtain reliable et al. 2018). However, by using a SVR model based on a radial
quantitative information, other techniques such as TED-GC- basis, Kernel function satisfying predictions were also obtained
MS or in case of PS and PET, LC-MS/MS or SEC may be for the real-world samples. This result indicates a high complex-
performed (Elert et al. 2017; Wang et al. 2017). These ity of the classification case finally demanding a non-linear ap-
methods produce results for sample aliquots in the order of proach. It should be noted that by inclusion of such samples into
50 to 500 mg within 1 to 2 hours. To gain detailed structural the calibration set, the performance of PLS-DA may be im-
information on MP such as number, size, and aging state, proved. Moreover, except the domestic_compost, which only
selected samples may be up-concentrated, purified, and small contained a minor fraction of PS even at the detection limit of
aliquots in the order of micrograms further analyzed by micro- TED-GC-MS, materials were studied in a difficult matrix con-
FTIR, or micro-Raman (Kaeppler et al. 2016). taining high amounts of non-plastic organic matter.
The time-demanding step in the combined NIR In contrast to the classification of MP-containing and MP-
spectroscopic-chemometric approach lies in the establishment free soils, PLS-DA yielded useful information on the type of
of the chemometric models. Depending on the matrix, suitable polymers. However, prediction fell off in quality when either
calibration samples must be selected and the resulting spectral MP levels were below 1 mass% or a combination of more than
information must be processed by mathematical filters two polymers was present. Furthermore, a trend of confusing
correcting for scattering effects, baseline offset and normali- predictions for PE and PP, and especially of false-positive
zation, together with the selection of relevant spectral ranges. detection of PET in MP-free soils, was observed. For the fu-
The validation material exclusively contained samples that ture, extended calibration strategies including more real-world
were not used in calibration. Whereas both the unknown soils samples are supposed to improve the prediction ability of the
and the MP-soil mixtures could be well classified by a PLS-DA classification model. In an ideal case, both the matrix material
model, for real-world samples, e.g., fermenter residue and com- and the contained polymers in calibration samples should best
post, PLS-DA failed. In contrast to the micro-spectroscopic ap- reflect the properties of the unknown samples.
proach, where spectra of each suspicious particle can be studied In the present work, we calibrated our models with samples
and compared to spectral libraries for identification, here each containing at least 1 mass% of cryo-milled polymers. This
Environ Sci Pollut Res (2019) 26:7364–7374 7373

mass fraction defines the minimum concentration that should from compost, fermenter residue, and washing machine filter
be recognized without any problems. Whereas the polymers effluent. Furthermore, we could demonstrate that even un-
PE and PP could be detected even at a mass fraction of about known blank soils of varying organic carbon content were
0.5% in MP-soil mixtures, for PS, PET, and PVC, a mass correctly classified, demonstrating robustness of the method
fraction of ≥ 1% was the necessary precondition. For certain towards changing soil composition. For the detection and clas-
scenarios, e.g., fermenter residue samples, no further condi- sification of MP, the applied SVR and PLS-DA models
tioning than homogenizing, drying, and sieving may be real- achieved a sensitivity of 92 and 83% and a specificity of
istic. However, as the average MP contamination of soil was 100 and 90%. The prediction ability of the chemometric
reported to be variable and often one to two orders of magni- models was found to be strongly dependent on the type and
tude lower (Blaesing and Amelung 2018), at least some kind number of calibration samples.
of up-concentration like electrostatic separation must be ap-
plied. Given the above summarized advantages regarding Acknowledgements The authors kindly thank Dr. Markus Ostermann
and Dr. Ute Kalbe for providing standard soils from their collections,
sample volumes and measuring times, the proposed NIR-
Andreas Sauer for the cryo-milling of polymers, Dr. Thomas Schmid
chemometric approach can be considered as a complementary for kindly providing microphotographs, Dr. Claus Gerhard Bannick
technique, especially when it comes to the assessment of large (Umweltbundesamt Berlin, Germany) for fermenter residue samples
sample volumes. and washing machine filter material, BONARES for soil samples, and
Yosri Hassanein for assistance with the preparation of samples. Many
Despite the well-known potential of multivariate spectra
thanks to Dr. Michael Maiwald (BAM), the Focus area BMicroplastics^
evaluation, there are only few chemometric approaches in at BAM, and the BMBF (MIWA, coordination Technische Universität
microplastics analysis. So far, only Karlsson et al. (2016) used Berlin, Prof. Martin Jekel) for the financial support of this project.
PCA and PLSR for the analysis of hyperspectral images of
purified MP samples emphasizing PCA as best tool for
classification. Renner et al. (2017) developed a combined References
curve-fitting and library searching algorithm for an automated
classification of plastics based on mid-infrared spectroscopic Bajer K, Braun U (2014) Different aspects of the accelerated oxidation of
data in surface reflection mode (standard FTIR-ATR) and re- polypropylene at increased pressure in an autoclave with regard to tem-
perature, pretreatment and exposure media. Polym Test 37:102–111
stricted to smaller sample volumes (few micrograms). An au-
Barej JAM, Pätzold S, Perkons U, Amelung W (2014) Phosphorus frac-
tomated approach based on micro-image FTIR analysis in tions in bulk subsoil and its biopore systems. Eur J Soil Sci 65:553–
transmission mode, library comparison, and benchmarks test 561. https://doi.org/10.1111/ejss.12124
for the identification of microplastics using FTIR has recently Blaesing M, Amelung W (2018) Plastics in soil: analytical methods and
been developed by Primpke et al. (2017). However, due to the possible sources. Sci Total Environ 612:422–435
Boucher J, Friot D (2017) Primary Microplastics in the Oceans: A Global
enzymatic sample preparation, this method is time consuming. Evaluation of Sources Imprint: Gland: IUCN, ISBN: 978-2-8317-
1827-9. https://portals.iucn.org/library/node/46622, Accessed 17
Nov 2017
Conclusions Chen W, Ouyang Z-Y, Qian C, Yu H-Q (2018) Induced structural changes
of humic acid by exposure of polystyrene microplastics: a spectro-
scopic insight. Environ Pollut 233:1–7
Microplastics of the materials PE, PP, PS, PET, and PVC can Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20:
be detected in soils at levels of about 1 mass% after minimal 273–297
conditioning, e.g., sieving and drying of the material, based on De Boves Harrington P (2015) Support vector machine classification
trees. Anal Chem 87:11065–11071
a combined NIR-chemometric approach. As soil samples can Duemichen E, Eisentraut P, Bannick CG, Barthel A-K, Senz R, Braun U
be examined without further purification or chemical treat- (2017) Fast identification of microplastics in complex environmen-
ment as described elsewhere (Blaesing and Amelung 2018), tal samples by a thermal degradation method. Chemosphere 174:
this seems to be an exceptionally easy procedure. Depending 572–584
Eide I, Neverdal G, Westad F (2010) Detection of 5 ppm fatty acid methyl
on the equipment used, e.g., the spectrometer type and optical ester (FAME) in jet fuel using electrospray ionization mass spec-
probe, larger sample volumes can be tested within short time. trometry and Chemometrics. Energy Fuel 24:3661–3664
Here, we used a process FT-NIR spectrometer coupled with a Elert AM, Becker R, Duemichen E, Eisentraut P, Falkenhagen J, Sturm H,
reflection probe to record on average 10 spectra on sample Braun U (2017) Comparison of different methods for MP detection:
what can we learn from them, and why asking the right question
aliquots of 1 g. Given the samples are well homogenized, before measurements matters? Environ Pollut 231:1256–1264
the measurement produced reliable results and took altogether Fabbri D (2001) Use of pyrolysis-gas chromatography/mass spectrometry to
5 min including both sample confection into the cap and mea- study environmental pollution caused by synthetic polymers. A case
surement. Compared to more common techniques like Raman study: the Ravenna lagoon. J Anal Appl Pyrolysis 58–59:361–370
Fischer M, Scholz-Bottcher BM (2017) Simultaneous trace identification
or FTIR spectroscopy, this is a significant improvement. By and quantification of common types of microplastics in environmen-
using a SVM model, recognition of MP was successful in- tal samples by pyrolysis-gas chromatography-mass spectrometry.
cluding both, polymer soil mixtures and real-world samples Environ Sci Technol 51:5052–5060
7374 Environ Sci Pollut Res (2019) 26:7364–7374

Fries E, Dekiff JH, Willmeyer J, Nuelle M-T, Ebert M, Remy D (2013) Moroni M, Mei A, Leonardi A, Lupo E, La Marca F (2015) PET and PVC
Identification of polymer types and additives in marine microplastic separation with hyperspectral imagery. Sensors 15:2205–2227.
particles using pyrolysis-GC/MS and scanning electron microscopy. https://doi.org/10.3390/s150102205
Environ Sci Process Impacts 15:1949–1956 Nizetto et al. (2016) Are agricultural soils dumps for microplastics of
Fuller S, Gautam A (2016) A procedure for measuring microplastics using urban origin? Environ Sci Technol 50:10777–10779
pressurized fluid extraction. Environ Sci Technol 50:5774–5780 Paul A, Breitfeld S, Sauer A, Becker R, Kalbe U, Bremser W, Braun U
Heggemann T, Welp G, Amelung W, Angst G, Franz SO, Koszinski S, (2016) A Raman spectroscopic method for the monitoring of
Schmidt S, Pätzold S (2017) Proximal gamma-ray spectrometry for microplastic in soil, 12. Kolloquium Prozessanalytik, Berlin
site-independent in situ prediction of soil texture on ten heteroge- Paul A, Becker R, Wander L, Breitfeld S, Koegler M, Sauer A, Braun U
neous fields in Germany using support vector machines. Soil Tillage (2017) An alternative spectroscopic approach for the monitoring of
Res 168:99–109 microplastics in environmental samples, 16 th International
Herzke D, Anker-Nilssen T, Nøst TH, Götsch A, Christensen-Dalsgaard Conference on chemistry and the environment, Oslo Norwegen,
S, Langset M, Fangel K, Koelmans AA (2016) Negligible impact of http://www.mn.uio.no/kjemi/english/research/projects/ICCE2017/
ingested microplastics on tissue concentrations of persistent organic monday-19.06/helga-eng-auditorium-1/Hr.-11%3A00/. Accessed
pollutants in northern fulmars off coastal Norway. Environ Sci 20 Nov 2017
Technol 50:1924–1933 Primpke S, Lorenz C, Rascher-Friesenhausen R, Gerdts G (2017) An
Hidalgo-Ruz V, Gutow L, Thompson RC, Thiel M (2012) Microplastics automated approach for microplastics analysis using focal plane
in the marine environment: a review of the methods used for iden- array (FPA) FTIR microscopy and image analyst. Anal Methods 9:
tification and quantification. Environ Sci Technol 46:3060–3075 1499–1511
Imhof HK, Laforsch C, Wiesheu AC, Schmid J, Anger PM, Niessner R, Renner G, Schmidt TC, Schram J (2017) A new chemometric approach
Ivleva NP (2016) Pigments and plastic in limnetic ecosystems: a for automatic identification of microplastics from environmental
qualitative and quantitative study on microparticles of different size compartments based on FT-IR spectroscopy. Anal Chem 89:
classes. Water Res 98:64–74 12045–12053. https://doi.org/10.1021/acs.analchem.7b02472
Kaeppler A, Fischer D, Oberbeckmann S, Schernewski G, Labrenz M, Roberts JJ, Cozzolino D (2016) Wet or dry? The effect of sample char-
Eichhorn K-J, Voit B (2016) Analysis of environmental acteristics on the determination of soil properties by near infrared
microplastics by vibrational microspectroscopy: FTIR, Raman or spectroscopy. Trends Anal Chem 83:25–30
both? Anal Bioanal Chem 408:8377–8391 Szczendzina G et al (2009) DCC-Ringversuche fester Brennstoffe,
Karlsson TM, Grahn H, van Bavel B, Geladi P (2016) Hyperspectral imaging Probenahme und Analytik. GIT Laborfachzeitschrift 3:163–165
and data analysis for detecting and determining plastic contamination in van den Broek WHAM, Wienke D, Melssen DJ, Buydens LMC (1998)
seawater filtrates. J Near Infrared Spectrosc 24:141–149 Plastic material identification with spectroscopic near infrared im-
Kooi M, EHv N, Scheffer M, Koelmans AA (2017) Ups and downs in the aging and articial neural networks. Anal Chim Acta 361:161–176
ocean: effects of biofouling on vertical transport of microplastics. Vianello A, Boldrin A, Guerriero P, Da Ros L (2013) Microplastic parti-
Environ Sci Technol 51:7963–7971 cles in sediments of lagoon of Venice, Italy: first observations on
Kostamovaara J, Tenhunen J, Kögler M, Nissinen I, Nissinen J, Keränen occurrence, spatial patterns and identification. Estuar Coast Shelf
P (2013) Fluorescence suppression in Raman spectroscopy using a Sci 130:54–61
time-gated CMOS SPAD. Opt Express 21:31632–31645 Wang L, Zhang J, Hou S, Sun H (2017) A simple method for quantifying
Miller CE, Eichinger BE (1990) Determination of crystallinity and morphol- polycarbonate and polyethylene terephthalate microplastics in envi-
ogy of fibrous and bulk poly(ethylene terephthalate) by near-infrared ronmental samples by liquid chromatography–tandem mass spec-
diffuse reflectance spectroscopy. Appl Spectrosc 44:496–504 trometry. Environ Sci Technol Lett 4:530–534
Mizushima M, Kawamura T, Takahashi K, Nitta KH (2012) In situ near- Workman J Jr, Weyer L (2012) Practical guide and spectral atlas for
infrared spectroscopic studies of the structural changes of polyeth- interpretive near-infrared spectroscopy. CRC Press Taylor &
ylene during melting. Polym J 44:162–166 Francis Group, Boca Raton