Machine Learning-Based Identification of Plastic Types Using Handheld Spectrometers

Abstract

Plastic waste and pollution is growing rapidly worldwide and most plastics end up in landfill or are incinerated because high-quality recycling is not possible. Plastic-type identification with a low-cost, handheld spectral approach could help in parts of the world where high-end spectral imaging systems on conveyor belts cannot be implemented. Here, we investigate how two fundamentally different handheld infrared spectral devices can identify plastic types by benchmarking the same analysis against a high-resolution bench-top spectral approach. We used the handheld Plastic Scanner, which measures a discrete infrared spectrum using LED illumination at different wavelengths, and the SpectraPod, which has an integrated photonics chip which has varying responsivity in different channels in the near-infrared. We employ machine learning using SVM, XGBoost, Random Forest and Gaussian Naïve Bayes models on a full dataset of plastic samples of PET, HDPE, PVC, LDPE, PP and PS, with samples of varying shape, color and opacity, as measured with three different experimental approaches. The high-resolution spectral approach can obtain an accuracy (mean ± standard deviation) of (0.97 ± 0.01), whereas we obtain (0.93 ± 0.01) for the SpectraPod and (0.70 ± 0.03) for the Plastic Scanner. Differences of reflectance at subsequent wavelengths prove to be the most important features in the plastic-type classification model when using high-resolution spectroscopy, which is not possible with the other two devices. Lower accuracy for the handheld devices is caused by their limitations, as the spectral range of both devices is limited-up to 1600 nm for the SpectraPod, while the Plastic Scanner has limited sensitivity to reflectance at wavelengths of 1100 and 1350 nm, where certain plastic types show characteristic absorbance bands. We suggest that combining selective sensitivity channels (as in the SpectraPod) and illuminating the sample with varying LEDs (as with the Plastic Scanner) could increase the accuracy in plastic-type identification with a handheld device.

Keywords: benchmarking; handheld; machine learning; plastic identification; spectroscopy.

Figure 1

The three different spectroscopic methods compared in this research to identify plastic types 1–6. We measured plastic samples with (a) an integrating sphere, broadband source and high-resolution near infrared spectrometer (NIRS) from 1050–1950 nm, (b) the discrete illumination-based Plastic Scanner with LEDs ranging from 900 to 1800 nm (PScanner) and (c) an integrated photonics approach using the SpectraPod (SP) using broadband light providing spectral responsivity from 850 to 1700 nm. NIRS gives a high-resolution spectrum with 237 datapoints, the Plastic Scanner, a discrete spectrum with 8 points, while the SP gives 16 channels that each have a different responsivity, as given in [9]. Below: Resin identification codes for plastic types used in this research.

Figure 2

Spectral reflectance measurements obtained with (a,b) NIR Spectrometer, (c) Plastic Scanner and (d) SpectraPod (Note: the SpectraPod integrates reflectance over multiple infrared bands as described in [9]). All measured spectra for a variety of plastic samples of different types (e.g., varying colors and opacity) are shown dimly in the background, and the mean of all measurements is given in a solid color corresponding to the legend. (a) Displays all NIRS data in one graph, and (b) shows the normalized spectra (mean = 1.0) with an offset to more clearly show the spectra.

Figure 3

Plastic Scanner (PScanner) accuracy with (a) the confusion matrix showing misclassification of LDPE, PS and PVC affecting the overall accuracy, (b,c) the emission spectra of the PScanner LEDs and the sensitivity of the InGaAs detector, respectively. LED spectra were measured separately and the detector responsivity is from the manufacturer’s specifications. Misclassification is due to a lack of spectral information around 1080–1160 nm, 1320–1380 nm and the broad spectral emission and overlap above 1580 nm, combined with low responsivity of the detector.

Figure 4

(a) Feature importance scores for the 5 most important features attributed to the Random Forest model approach on the NIRS data. This shows that the difference (numerical derivative) at certain wavelengths is most important in the classification. (b,c) Show the zoom of the average reflectance spectrum for the different plastic types around the three most important features used in the Random Forest model.