Detection of Brominated Plastics from E-Waste by Short-Wave Infrared Spectroscopy

: In this work, the application of Short-Wave Infrared (SWIR: 1000–2500 nm) spectroscopy was evaluated to identify plastic waste containing brominated ﬂame retardants (BFRs) using two different technologies: a portable spectroradiometer, providing spectra of single spots, and a hyperspectral imaging (HSI) platform, acquiring spectral images. X-ray Fluorescence (XRF) analysis was preliminarily performed on plastic scraps to analyze their bromine content. Chemometric methods were then applied to identify brominated plastics and polymer types. Principal Component Analysis (PCA) was carried out to explore collected data and deﬁne the best preprocessing strate-gies, followed by Partial Least Squares—Discriminant Analysis (PLS-DA), used as a classiﬁcation method. Plastic fragments were classiﬁed into “High Br content” (Br > 2000 mg/kg) and “Low Br content” (Br < 2000 mg/kg). The identiﬁed polymers were acrylonitrile butadiene styrene (ABS) and polystyrene (PS). Correct recognition of 89–90%, independently from the applied technique, was achieved for brominated plastics, whereas a correct recognition ranging from 81 to 89% for polymer type was reached. The study demonstrated as a systematic utilization of both the approaches at the industrial level and/or at laboratory scale for quality control can be envisaged especially considering their ease of use and the short detection response.


Introduction
In recent years, the management of Waste from Electrical and Electronic Equipment (WEEE) is becoming more and more challenging due to its growing volume rate [1]. WEEE waste streams require special treatment and sometimes complex management activities due to their intrinsic potential toxicity for the environment and their harmfulness to human and animal health [2]. Their complexity in terms of composition, which may vary in time according to technological improvement and environmental regulations, lead to difficulties in recycling and recovery processes.
WEEE contain important metals (i.e., gold, silver, iron, steel, copper, and aluminum) and other valuable materials, such as plastics, glass, and wood [3]. Their correct handling, in terms of processing actions to obtain raw materials of secondary origins [4], thus produces a positive impact both in an economic (i.e., materials recovery) and in an environmental (i.e., reduced material disposed-off or incinerated) perspective.
In Europe, the plastic recycling rate ranges between 26 and 52% [5], showing that there is a need to increase such rates, developing innovative strategies. Considering that plastics from WEEE represent on average 25% of all WEEE annually generated by weight [6],  The first two fractions can be then further sorted using electrostatic separation methods, whereas the last fraction, with a density higher than 1.1 g/cm 3 , is typically disposed of by incineration, co-processing in cement kilns, or landfilled, and generally classified as hazardous waste.
However, the density method, despite its routine use in separating brominated plastics, is affected by errors, including plastics with high Br content in the recycled low Br plastic stream (floating fraction) and losing profitably recoverable plastics in the sink fraction, due to the presence of polymers with similar densities, such as ABS and PS [6,23]. Indeed, the sink-float method suffers from relatively poor selectivity. In more detail, plastics sorting from CRT screens using this method produces a waste fraction (Br-rich fraction) representing about 30% of the input [6].
The fast identification of polymers with certain flame-retardant grades in a mixed waste stream is not an easy task. The methods used at laboratory scale to detect flame retardant grades of polymers are commonly based on the utilization of different spectroscopic techniques [24], the most utilized include Raman scattering, mass pyrolysis, sliding spark (spark ablation), X-ray Fluorescence (XRF), Laser-Induced Plasma Spectroscopy (LIPS), Near Infrared (NIR), Short-Wave Infrared (SWIR), Mid Infrared (MIR) reflection, MIR pyrolysis and MIR Acousto-Optic Tunable Filter (MIR AOTF).
However, not all the above-mentioned methods are fast, sustainable, not destructive, and cost-effective. In a plant scenario, a method that is rapid, effective, and does not require sample preparation is needed. NIR-SWIR spectroscopy meets these requirements.
In recent years, Hyperspectral Imaging (HSI) has rapidly emerged and fast grown in many industrial fields, including the solid waste sector [25,26]. HSI is a technology allowing the collection of both spatial and spectral information at the same time, thus providing several physical and chemical characteristics of the analyzed material [27]. In literature, several studies related to NIR-his-based applications for WEEE material detection [28][29][30][31][32] and polymer identification and/or characterization [33][34][35][36][37][38][39][40] are reported.
In this work, two different spectroscopic techniques, working in the SWIR range (1000-2500 nm), were applied to identify brominated plastic scraps from CRT monitors and televisions, adopting a multi-analytical approach: a portable spectrophotoradiometer, working on "single spot base" and a HSI platform, acquiring spectral images. The first one was evaluated in order to perform a rapid test on batch samples collected in the feed and/or on some of the plastic waste flow streams generated by the recycling process, as schematically represented in Figure 1b. The second one was tested for the "on-line" assessment of sorted plastic quality, with reference to the BFRs presence, and for the development of sorting strategies, as the example reported in Figure 1c.

Materials
The analyzed plastic samples were collected from the recycling plant of Galloo Plastics (Halluin, France), in which density medium separation is used for concentrating the BFRs plastics from a WEEE stream. The samples consisted of shredded plastic scraps from cathode ray tube (CRT) monitors and televisions. The physical features (i.e., color, thickness, and weight) of the samples are reported in Table 1. In more detail, scraps were mainly made of acrylonitrile butadiene styrene (ABS) and polystyrene (PS) polymers with variable bromine content. Thirty-six scraps ( Figure 2) were randomly selected from the samples, and their polymer typology (ABS and PS) was verified using Raman spectroscopy. mainly made of acrylonitrile butadiene styrene (ABS) and polystyrene (PS) polymers with variable bromine content. Thirty-six scraps ( Figure 2) were randomly selected from the samples, and their polymer typology (ABS and PS) was verified using Raman spectroscopy.

X-ray Fluorescence Analysis
The scraps were preliminary analyzed to determine the total bromine content by a portable XRF spectrometer. This procedure was adopted, as it is commonly used in recycling plants, to perform controls on plastic scraps which could contain BFRs added polymers.
A hand-held XRF Niton™ XL2 (Thermo Fisher Scientific Inc. Waltham, MA, USA), with a silver anode X-ray Tube (6-45 kV, 1-200 uA, 2 W max) mounted on a benchtop stand aided by software dedicated for plastics (with automatic thickness correction) was used for Br measurements. Measurement time for each individual scrap was about 60 s. Preliminary tests have been carried out in order to verify the robustness of XRF determinations. To reach this goal, Br determinations were performed on a certified reference material during different days at a Br concentration of about 1000 mg/kg. A value of 0.03 was thus obtained for the Coefficient of Variation (CV), where CV is the ratio of Standard Deviation (SD) to the Mean (M) collected data.
According to the detected Br contents, individual scraps were then divided into two representative sample sets: one for training and another for validation.

SWIR Spectral Analysis
Portable spectroradiometer. Spectra acquisitions were performed in reflectance mode for each individual scrap, using the ASD FieldSpec ® 4 Standard-Res field portable spectroradiometer (ASD Inc., Boulder, CO, USA) with a contact probe. This portable instrument, working in VIS-SWIR regions (350-2500 nm), had a spectral resolution of 3 nm at 700 nm and 10 nm at 1400/2100 nm [41]. The spectroradiometer was composed of a detector unit and a fiber optics cable connected to a contact probe, controlled by a personal computer. The detector unit was realized by coupling different separate holographic diffraction gratings with three separate detectors. The detector architecture consisted of a VNIR detector (512 element silicon array: 350-1000 nm), a SWIR1 detector (Graded Index InGaAs. Photodiode, Two Stage TE Cooled; 1001-1800 nm), and a SWIR2 detector (Graded Index InGaAs. Photodiode, Two Stage TE Cooled; 1801-2500 nm). The ASD Contact Probe consisted of a halogen bulb light source with a color temperature 2901 +/− 10% • K. The contact probe spot size was 10 mm in diameter. Data acquisition and calibration procedures were performed using the ASD RS 3 software (ASD Inc., Boulder, CO, USA) [42]. The calibration of the ASD FieldSpec ® 4 Standard-Res spectroradiometer was performed by dark acquisition, calculated referencing the dark current calibration file, and by mean of the white reference measurement, acquiring a standardized white Spectralon ® ceramic material. After this calibration stage, the spectrum was acquired, and reflectance was then computed for each sample. For the purposes of this study, only the SWIR range of the collected spectra was investigated.
Hyperspectral imaging. Hyperspectral images were acquired in the SWIR range (1000-2500 nm) using the SisuCHEMA TM XL Chemical Imaging Workstation (Specim, Spectral Imaging Ltd., Oulu, Finland), equipped with the ImSpector TM N25E imaging spectrograph (Specim, Spectral Imaging Ltd., Oulu, Finland). A 31 mm lens with a field of view of 50 mm was adopted. The spectral resolution was 10 nm. This device combined NIR spectroscopy with high-resolution imaging. A push-broom hyperspectral camera acquired and built the hyperspectral image of sample line by line, simultaneously acquiring all wavelengths for each line. Image data were automatically calibrated in reflectance by measuring an internal standard reference target before each sample scan. A comparison between the technical characteristics of the two spectral devices is reported in Table 2.

Experimental Procedure
The flow chart reported in Figure 3 schematically explains the experimental procedure defined to perform the analysis using the different sensing devices and detection architectures. Two sample sets, belonging to different polymer families (ABS and PS) and characterized by different bromine content, "High Br content" (Br > 2000 mg/kg) and "Low Br content" (Br < 2000 mg/kg), were thus investigated.  The flow chart reported in Figure 3 schematically explains the experimental procedure defined to perform the analysis using the different sensing devices and detection architectures. Two sample sets, belonging to different polymer families (ABS and PS) and characterized by different bromine content, "High Br content" (Br > 2000 mg/kg) and "Low Br content" (Br < 2000 mg/kg), were thus investigated. Data acquisition and handling. The data collected by XRF analysis were used to set reference classes for SWIR investigation.  Data acquisition and handling. The data collected by XRF analysis were used to set reference classes for SWIR investigation.
ASD FieldSpec ® 4 Standard-Res spectra ".asd" data files were stacked into an ASCII text file using ViewSpec Pro (ver. 6.2.0; ASD Inc., Boulder, CO, USA). The ASCII text file was then imported into the MATLAB ® environment (MATLAB R2019a ver. 9.6.0; The Mathworks, Inc., Natick, MA, USA) using "fieldspec_import3.m", an ad hoc written script [43]. Imported data files were analyzed using Eigenvector Research, Inc. PLS_toolbox (ver. 8.6; Eigenvector Research, Inc, Wenatchee, WA, USA) within the MATLAB ® environment. Hyperspectral images were also imported into the MATLAB ® environment and subsequently analyzed with the aid of PLS_toolbox and MIA_toolbox (ver. 3.0; Eigenvector Research, Inc, Wenatchee, WA, USA) and finally concatenated together in order to create a hyperspectral image mosaic.
Acquired plastic scraps were divided into two classes according to the recommended Br concentration limits [22]: "High Br content" (>2000 mg/kg) and "Low Br content" (<2000 mg/kg), based on XRF analysis. For each plastic scrap, 5 "single spot based" spectra acquisitions in reflectance mode were carried out using the ASD FieldSpec ® device. On the same areas, the average reflectance spectra measured on circular regions of interest (ROIs) of about 10 mm diameter were extracted from the hyperspectral images ( Figure 4). The ROIs diameter was established in respect of the average area investigated by the spectroradiometer probe. The average reflectance spectrum of each ROI was compared with the corresponding spectrum measured by the single-point device. ASD FieldSpec ® 4 Standard-Res spectra ".asd" data files were stacked into an ASCII text file using ViewSpec Pro (ver. 6.2.0; ASD Inc., Boulder, CO, USA). The ASCII text file was then imported into the MATLAB ® environment (MATLAB R2019a ver. 9.6.0; The Mathworks, Inc., Natick, MA, USA) using "fieldspec_import3.m", an ad hoc written script [43]. Imported data files were analyzed using Eigenvector Research, Inc. PLS_toolbox (ver. 8.6; Eigenvector Research, Inc, Wenatchee, WA, USA) within the MATLAB ® environment. Hyperspectral images were also imported into the MATLAB ® environment and subsequently analyzed with the aid of PLS_toolbox and MIA_toolbox (ver. 3.0; Eigenvector Research, Inc, Wenatchee, WA, USA) and finally concatenated together in order to create a hyperspectral image mosaic.
Acquired plastic scraps were divided into two classes according to the recommended Br concentration limits [22]: "High Br content" (>2000 mg/kg) and "Low Br content" (<2000 mg/kg), based on XRF analysis. For each plastic scrap, 5 "single spot based" spectra acquisitions in reflectance mode were carried out using the ASD FieldSpec ® device. On the same areas, the average reflectance spectra measured on circular regions of interest (ROIs) of about 10 mm diameter were extracted from the hyperspectral images ( Figure 4). The ROIs diameter was established in respect of the average area investigated by the spectroradiometer probe. The average reflectance spectrum of each ROI was compared with the corresponding spectrum measured by the single-point device.  . Topological assessment strategy, adopted for each plastic scrap, to select the five spectra acquisition points by ASD FieldSpec ® 4 Standard-Res and the resulting raw reflectance spectra (1000-2500 nm) (a) and the corresponding five Regions of interest (ROIs) utilized to collect raw reflectance spectra (1000-2500 nm) by the SisuCHEMA XL™ HSI system (b). . Topological assessment strategy, adopted for each plastic scrap, to select the five spectra acquisition points by ASD FieldSpec ® 4 Standard-Res and the resulting raw reflectance spectra (1000-2500 nm) (a) and the corresponding five Regions of interest (ROIs) utilized to collect raw reflectance spectra (1000-2500 nm) by the SisuCHEMA XL™ HSI system (b).
Data pretreatments and exploratory analysis. In spectroscopic applications, spectral preprocessing was not only used to reduce instrument noise, scattering, and other physical phenomena linked to spectra acquisition but also to enhance sample differences and for better solving information in the spectral domains [44]. Scale differences in spectra were mainly due to scattering phenomena and changing environmental conditions affecting illuminants or detectors.
Different preprocessing algorithms were sequentially applied to the acquired reflectance spectra collected by both instruments to enhance differences between the two classes of products according to the bromine contents. In more detail, the Standard Normal Variate (SNV) algorithm was used for reducing scattering effects [45], Savitzky-Golay (S-G) first derivative was used to enhance signal differences [46], Detrend was applied to remove the mean offset from each spectrum [45] and, finally, Mean Center (MC) was used to center columns to have zero mean. Furthermore, spectra were also preprocessed, in order to highlight polymer spectral signatures (i.e., ABS or PS), using SNV, Detrend, and MC.
Principal Component Analysis (PCA) was then carried out in order to evaluate and detect outliers. PCA is a well-known chemometric technique used as a data exploratory tool, that is able to extract the dominants patterns of a spectral data matrix in terms of the product of two smaller matrices of scores and loadings [47,48].
PLS-DA classification models. Partial Least Square Discriminant Analysis (PLS-DA), a supervised technique for pattern recognition, was used to classify the sample data according to bromine content and polymer type. PLS-DA is essentially an inverse-least square approach to the Linear Discriminant Analysis (LDA). The Partial Least Square (PLS) regression was used in this technique to develop a model able to predict the class number for each investigated sample [49][50][51].
For each device, classification models were developed in order to perform discrimination based on bromine content first and then to identify the polymer type.
Classification models for bromine content. As shown in Figure 3, the PLS-DA method was used in order to discriminate "High Br content" and "Low Br content" classes in the 1000-2500 nm spectral range by using the portable spectrophotoradiometer and the HSI benchtop system.
Two classification models were set up: one for the spectra collected by the ASD FieldSpec ® 4 Standard-Res and the other for those acquired by the SisuCHEMA XL™. For these models, about 72% of the spectral data was used as the training set, whereas the remaining percentage (28%) was used as the validation set. In more detail, 26 samples were selected to be included in the training set and 10 individuals for the validation set. Venetian Blinds (VB) was used as a cross-validation method for assessing the optimal complexity of the models and choosing the number of Latent Variables (LVs). In this case, 3 LVs were chosen for both models.
The PLS-DA model built from FieldSpec data was applied to the reflectance spectra collected on the validation set samples (10 plastic scraps).
The built PLS-DA of the SisuCHEMA XL™ system spectral data was first applied to the average spectra of the ROIs selected on the validation image ("average spectra") and then to the validation image ("imaging"-validation set) and the global set ("imaging").
Classification models for polymer identification. A similar procedure, as described in the previous paragraph, was followed in order to set up the PLS-DA models for polymer classification (Figure 3). The classification models were built using reference spectra of ABS and PS for training and were validated on all samples (global set) for both the sensing devices. In more detail, the PLS-DA model was applied to the spectra collected from the global set by the portable spectrophotoradiometer and, concerning HSI, to the average spectra of the ROIs extracted from the validation set image and to the whole global set image ("imaging"). Four LVs were chosen for both models.
Classification performance assessment. The confusion matrix was considered to evaluate classifier performance coupled with the commonly used performance metrics, that is, Sensitivity, Specificity, Precision, Accuracy, and Class Error [33,52,53]. Sensitivity estimates the model's ability to avoid false negatives. On the contrary, Specificity is used to evaluate the model's ability to avoid false positives. Precision, also called positive predictive value, is defined as the ratio between the number of true positives and the number of positives calls. Accuracy is the proportion of correct predictions (both true positives and true negatives) among the total number of cases examined. Those parameters assume values between 0 and 1, where 1 is the ideal value for the model. Finally, Class Error is defined as the misclassified global fraction of particles.
In an object-based logic, the performance of each classification model was computed in terms of Recognition/Accuracy (i.e., the ratio of the well-recognized samples on the total number of samples) and Error (i.e., number of misrecognized samples in respect of the total number of samples) [53,54]. An object was deemed correctly classified when: (i) for punctual spectra acquisitions, more than 3 out of 5 spectra were correctly assigned to their belonging classes and (ii) adopting a hyperspectral approach, 50% or more of the pixels were correctly assigned to their respective classes.

X-ray Fluorescence Analysis
Bromine concentrations obtained by XRF of plastic scraps are reported in Table 3. The analyses showed 27 samples (75%) were characterized by a Br content higher than 2000 mg/kg and 9 (25%) by Br content lower than 2000 mg/kg.

SWIR Spectral Analysis
Exploratory analysis. Figure 4 shows an example of the strategy adopted to perform the five punctual acquisitions by the ASD FieldSpec ® 4 Standard-Res and the reflectance spectra acquired by the SisuCHEMA XL™ HSI system in the five corresponding ROIs. Figure 5 shows the raw reflectance spectra in the SWIR range (1000-2500 nm), included in the training and validation datasets, for the two devices. Spectral signature analysis for bromine content assessment. The raw reflectance spectra of the samples averaged according to the two bromine classes (high and low content), collected with the two devices, were preprocessed utilizing the same algorithms in order to correctly compare the further classification results ( Figure 6).
The raw reflectance spectra of samples characterized by a high Br content, differently from those characterized by a low Br content, show an extra absorption peak around 1469 nm. This peak can be linked to the presence of TBBPA (tetrabromobisphenol A), as reported in past studies [36,37,40]. Spectra preprocessing enhances the differences among the two classes. As it can be seen in Figure 6, the higher spectral resolution of ASD Field-Spec ® 4 Standard-Res produced a more enhanced preprocessed spectrum. Spectral signature analysis for bromine content assessment. The raw reflectance spectra of the samples averaged according to the two bromine classes (high and low content), collected with the two devices, were preprocessed utilizing the same algorithms in order to correctly compare the further classification results ( Figure 6).
The raw reflectance spectra of samples characterized by a high Br content, differently from those characterized by a low Br content, show an extra absorption peak around 1469 nm. This peak can be linked to the presence of TBBPA (tetrabromobisphenol A), as reported in past studies [36,37,40]. Spectra preprocessing enhances the differences among the two classes. As it can be seen in Figure 6, the higher spectral resolution of ASD FieldSpec ® 4 Standard-Res produced a more enhanced preprocessed spectrum.
Spectra signature analysis for polymers identification. Figure 7 shows the raw and preprocessed reflectance spectra of the reference polymers used to train the classification model for plastic-type classification (i.e., ABS and PS).  Figure 7 shows the raw and preprocessed reflectance spectra of the reference polymers used to train the classification model for plastic-type classification (i.e., ABS and PS).

Spectra signature analysis for polymers identification.
The SWIR range is very useful to identify different polymers because it provides information on the overtone bands of the fundamental groups containing O-H, C-H, N-H, and C-O bonds [55]. In this specific range, the identification of plastics is mainly based on the stretching vibration modes of CH, CH2, and CH3 groups between about 1100 and 1250 nm, corresponding to the 2nd overtone, the 1st overtone ranging from 1650 to 1700 nm, and the combination bands. More in detail, the spectra of PS samples were characterized by absorptions in the 3 rd region of harmonics (1043, 1151, 1214, and 1308 nm) due to absorption of C-H2 and C-H, in the second harmonic region (1352, 1415, and 1648 nm) of C-H2 and δC-H2, and in the 1 st combination region (1817, 1868, 1918, 2012, and 2074 nm) of C-H group [56]. ABS spectrum had a similar fingerprint to that of PS; the main differences are related to less intensity absorptions of ABS with respect to PS in the 3rd region of harmonics and a different absorption in the 1st combination region (1817, 1868, and 1918 nm) of C-H group. Classification models for bromine content. Both the built PLS-DA models based on the collected spectra carried out by "single spot" and ROIs based analysis showed good performance in discriminating the bromine content classes as reported in Tables 4 and 5, respectively, in calibration, cross-validation, and prediction. However, some spectra were incorrectly classified (please refer to Supplementary Materials for details). The SWIR range is very useful to identify different polymers because it provides information on the overtone bands of the fundamental groups containing O-H, C-H, N-H, and C-O bonds [55]. In this specific range, the identification of plastics is mainly based on the stretching vibration modes of CH, CH 2 , and CH 3 groups between about 1100 and 1250 nm, corresponding to the 2nd overtone, the 1st overtone ranging from 1650 to 1700 nm, and the combination bands. More in detail, the spectra of PS samples were characterized by absorptions in the 3 rd region of harmonics (1043, 1151, 1214, and 1308 nm) due to absorption of C-H 2 and C-H, in the second harmonic region (1352, 1415, and 1648 nm) of C-H 2 and δC-H 2 , and in the 1 st combination region (1817,1868,1918,2012, and 2074 nm) of C-H group [56]. ABS spectrum had a similar fingerprint to that of PS; the main differences are related to less intensity absorptions of ABS with respect to PS in the 3rd region of harmonics and a different absorption in the 1st combination region (1817, 1868, and 1918 nm) of C-H group.
Classification models for bromine content. Both the built PLS-DA models based on the collected spectra carried out by "single spot" and ROIs based analysis showed good performance in discriminating the bromine content classes as reported in Tables 4 and 5, respectively, in calibration, cross-validation, and prediction. However, some spectra were incorrectly classified (please refer to Supplementary Materials for details). In the case of ASD FieldSpec ® 4 Standard-Res data (Table 4), during the calibration phase, Sensitivity and Specificity values were 0.833 and 1.000 for "High Br content" and 1.000 and 0.833 for "Low Br content", respectively. As shown in Figures S1 and S2, three plastic scraps were not correctly assigned to their real classes (E4, E20, and E60). In the prediction phase, Sensitivity and Specificity values were 0.857 and 1.000 for "High Br content" and 1.000 and 0.857 for "Low Br content". In this case, sample E14 was misclassified as a "Low Br content" scrap in prediction. Concerning the performance parameters related to SisuCHEMA XL™ average spectral data (Table 5), during the calibration phase Sensitivity and Specificity values were 0.900 and 0.833 for "High Br content" and 0.833 and 0.900 for "Low Br content", respectively. Sample E20, E42, and E60 were not correctly assigned to their real class. The built classification model, in prediction, confirmed the good performance achieved in calibration (Sensitivity = 0.886, Specificity = 1.000 for "High Br content" and Sensitivity = 1.000, Specificity = 0.886 for "Low Br content"). In this case, the model did not recognize sample E14 as a "High Br content" one. Finally, the performance parameters related to the PLS-DA classification of the validation image set and global image set are reported in Table 5. For the validation set, the results showed that all the samples, except E14, were correctly classified (Sensitivity = 0.950, Specificity = 0.951 for "High Br content" and Sensitivity = 0.951, Specificity = 0.950 for "Low Br content"). For the global set, the performance parameters showed also in this case good values (Sensitivity = 0.881, Specificity = 0.932 for "High Br content" and Sensitivity = 0.932, Specificity = 0.881 for "Low Br content"), with four samples incorrectly classified: E42, E14, E60, and E20. With reference to the latter case, the classification results are also presented in terms of prediction maps in Figure 8, for both validation image set and global image set, together with the false-color image, indicating the true class membership, highlighting the few misclassified fragments. All the misclassified samples show flat spectra with low reflectance intensities, as can be seen in Figure 5 and Figure S3 (between 0.16-0.2).
The results (Tables 4 and 5) showed similar performances between the ASD FieldSpec ® 4 Standard Res-based approach compared to the SisuCHEMA XL™ one. In calibration and cross-validation phases, performance metrics obtained using the spectrophotoradiometer data were better than those obtained by the HSI system data. On the contrary, the performances obtained in prediction were slightly better for the classification performed by HSI system data than the portable system, since the Sensitivity and Specificity, as well as Precision and Accuracy, were higher for both classes.
Results obtained in prediction for the validation image set and global image confirm the reliability of the built model.
Classification models for polymer identification. The classification performance metrics were obtained following three different approaches: the first two based on the global set referred to ASD FieldSpec ® 4 Standard-Res and SisuCHEMA XL™ spectra and the third based on SisuCHEMA XL™ hyperspectral images mosaic ( Table 6).
The classification model applied to ASD FieldSpec ® 4 Standard-Res data was not able to correctly classify seven samples (E29, E31, E38, E39, E40, E46, and E55): six PS individuals were incorrectly classified as ABS and one ABS as PS. This fact was confirmed from Sensitivity and Specificity values (Sensitivity = 0.950, Specificity = 0.638 for ABS and Sensitivity = 0.638, Specificity = 0.950 for PS), reported in Table 6.
The classification model based on the analysis of SisuCHEMA XL™ global set spectra was not able to classify four PS samples (E31, E39, E46, and E57) that were incorrectly classified as ABS (Sensitivity = 1.000, Specificity = 0.713 for ABS and Sensitivity = 0.713, Specificity = 1.000 for PS).
A similar performance was achieved by applying the classification model based on hyperspectral images mosaic global set. The prediction map resulting from the PLS-DA is shown in Figure 9, together with the false-color image, indicating the true class membership (Figure 9a). As can be seen in Figure 9b, PS samples (E31, E39, E46, and E57) were incorrectly classified as ABS (Sensitivity = 0.985, Specificity = 0.682 for ABS and Sensitivity = 0.682, Specificity = 0.985 for PS).
The observed misclassification in discriminating ABS and PS was also observed by other authors. Wu et al. (2020) reported that only 72.8% of PS was correctly classified, and the remaining percentage was misclassified as ABS in PLS-DA classification using NIR spectroscopy (900-1700 nm). Similar results were also achieved by Amigo et al. (2015) using this approach. In this latter case, a large number of pixels were not correctly assigned to the real polymer class [33].
Overall classification models performances. Overall classification performances of the set-up models are reported in Table 7. The set-up models to discriminate Br classes with FieldSpec and SisuCHEMA spectral data achieved a Recognition of 88.46% and 90.00% for the training and validation set, respectively. The models performed to identify the type of polymer achieved a Recognition of 80.56% and 88.89% for FieldSpec data and SisuCHEMA (global set), respectively.   The classification model based on the analysis of SisuCHEMA XL™ global set spectra was not able to classify four PS samples (E31, E39, E46, and E57) that were incorrectly classified as ABS (Sensitivity = 1.000, Specificity = 0.713 for ABS and Sensitivity = 0.713, Specificity = 1.000 for PS).
A similar performance was achieved by applying the classification model based on hyperspectral images mosaic global set. The prediction map resulting from the PLS-DA is shown in Figure 9, together with the false-color image, indicating the true class membership ( Figure 9a). As can be seen in Figure 9b, PS samples (E31, E39, E46, and E57) were incorrectly classified as ABS (Sensitivity = 0.985, Specificity = 0.682 for ABS and Sensitivity = 0.682, Specificity = 0.985 for PS).
The observed misclassification in discriminating ABS and PS was also observed by other authors. Wu et al. (2020) reported that only 72.8% of PS was correctly classified, and the remaining percentage was misclassified as ABS in PLS-DA classification using NIR spectroscopy (900-1700 nm). Similar results were also achieved by Amigo et al. (2015) using this approach. In this latter case, a large number of pixels were not correctly assigned to the real polymer class [33]. Overall classification models performances. Overall classification performances of the set-up models are reported in Table 7. The set-up models to discriminate Br classes with FieldSpec and SisuCHEMA spectral data achieved a Recognition of 88.46% and 90.00% for the training and validation set, respectively. The models performed to identify the type of polymer achieved a Recognition of 80.56% and 88.89% for FieldSpec data and SisuCHEMA (global set), respectively.  Table 7. Overall classification model performances. Calculations were performed applying an object-based recognition logic that was assuming a discriminating threshold of more than three out of five spectra for a punctual analysis and of 50% or more pixels correctly assigned for the hyperspectral approach, respectively. Combining bromine content measured by XRF and polymer type of the single scrap ( Figure 10), it can be seen that all individual objects identified as ABS (n = 20; Br content = 76,378 ± 20,593 mg/kg) exceeded the bromine content regulation limit of 2000 mg/kg. On the contrary, only 47% of PS scraps exceeded the Br content regulation limit (n = 16; Br content = 19,001 ± 30,015 mg/kg). As seen from the HSI-based classification for bromine content, the misclassified samples were PS plastic scraps (Figures 8 and 9). The misclassification mainly occurred for PS with low Br content, which was incorrectly assigned to the "high Br class".

Classification
Recycling 2021, 6, x FOR PEER REVIEW 18 of 22 Table 7. Overall classification model performances. Calculations were performed applying an object-based recognition logic that was assuming a discriminating threshold of more than three out of five spectra for a punctual analysis and of 50% or more pixels correctly assigned for the hyperspectral approach, respectively. Combining bromine content measured by XRF and polymer type of the single scrap ( Figure 10), it can be seen that all individual objects identified as ABS (n = 20; Br content = 76,378 ± 20,593 mg/kg) exceeded the bromine content regulation limit of 2000 mg/kg. On the contrary, only 47% of PS scraps exceeded the Br content regulation limit (n = 16; Br content = 19,001 ± 30,015 mg/kg). As seen from the HSI-based classification for bromine content, the misclassified samples were PS plastic scraps (Figures 8 and 9). The misclassification mainly occurred for PS with low Br content, which was incorrectly assigned to the "high Br class".

Conclusions and Future Perspectives
This study was carried out in order to test the predictive ability of SWIR working instruments to discriminate plastic scraps with high bromine content from the low ones. This goal was achieved using two different equipment: a portable spectrophotoradiometer and a benchtop HSI system. The set-up models to discriminate bromine content classes with the portable spectrophotoradiometer and HSI data achieved equal Recognition (i.e., 88.46% and 90% for the training and validation set, respectively). The models performed to identify the type of polymer achieved a Recognition of 80.56% and 88.89% for FieldSpec and SisuCHEMA data (global sets), respectively. Both the instrumentations gave useful information about BFRs added polymer scraps, for example, to evaluate the performances of the currently adopted density separation process for brominated plastics in recycling plants, especially when dealing with different polymers characterized by similar densities, such as ABS and PS.
Both the instrumentations gave useful information about BFRs added polymer scraps, for example, to evaluate the performances of the currently adopted density separation process for brominated plastics in recycling plants, especially when dealing with different polymers characterized by similar densities, such as ABS and PS. The utilization of the portable instrument could be applied to a rapid test on samples collected at the entrance (i.e., feed) and/or on the process streams moving towards the exit and/or on the recovered product of a waste treatment plant. In addition, SWIR-HSI system can be utilized not only for the same above-mentioned purposes but also as a core engine to implement online sorting logics.
The proposed approach-if systematically applied at the recycling plant scale-could quickly, non-destructively, and efficiently carry out a complete control of brominated plastic contaminants. In fact, this method can both identify the bromine content of WEEE plastics and assess the polymer type.
Both methodologies can be seen as a tool to maximize the recyclable polymer fractions, removing those individuals containing toxic substances (i.e., Br), thus increasing the circular use of materials, reducing pollution, toxic emissions, landfill, and also contributing to mitigate climate change effects, specifically addressing the principles of circular economy and the Sustainable Development Goals (SDGs), as SDG 9 (Industry, Innovation, and Infrastructure), SDG 12 (Responsible Consumption and Production (SDG 12), SDG 13 (Climate Action), SDG 14 (Life Below Water), and SDG 15 (Life on Land).

Data Availability Statement:
The data presented in this study are available on request from the corresponding authors.