Applications of hyperspectral imaging and machine learning methods for
real-time classification of waste stream components
(7573)Martin Sevcik
1, Jan Skvaril
1, Elena Tomás-Aparicio
121. School of Business, Society and Engineering, Future Energy Center,
Mälardalen University, Västerås, Västmanland, Sweden
2. Mälarenergi AB, Västerås, Västmanland, Sweden
ABSTRACT
Near-infrared (NIR) hyperspectral imaging (HSI) was applied together with machine learning methods to enable classification of typical municipal solid waste (MSW) components such as paper, biomass, food residues, plastics, textile and incombustibles. Classification models were developed using partial least square discriminant analysis (PLS-DA), support vector machine (SVM), and radial-basis neural network (RBNN). The overall accuracy of SVM model calculated from
classification sensitivity was 85% in prediction pixel by pixel for external sample set. The model outperformed other models in identifying incombustible material but it had higher computational time requirements. The accuracy of RBNN model reached 85% in prediction while being approx. 10 times faster. Minimum computational time was required by PLS-DA model reaching lower accuracy of 81% in prediction. The result indicate that developed models can be successfully used for real-time MSW component classification. NIR hyperspectral imaging coupled with machine learning methods has a large potential to be used for on-line material identification at waste sorting facilities or for pre-sorting at waste-to-energy powerplants.
INTRODUCTION
Modern industrialized world is characterized by ever-growing energy and material demands resulting in production of large amounts of wastes in various forms. Sustainable waste
management approach is required to address this environmental threat. According to European waste management hierarchy [1] waste re-use, recycle and energy recovery is strongly preferred over waste disposal. Most common waste type that covers household waste consisting of everyday items is a municipal solid waste (MSW). The MSW originates from residential areas, schools and hospitals and commercial sources such as small business and restaurants [2]. Composition of MSW varies greatly depending on its origin and typically contains: cardboard, food residues, glass, metals and ceramics, newsprint, plastics (i.e. organic polymers such as high density/low density polyethylene HDPE/LDPE, polyethylene terephthalate PET, polypropylene PP, polystyrene PS and polyvinyl chloride PVC), textiles, paper and wood.
Components of the waste such as plastics can be sorted out and reused or recycled in order to prolong their lifecycle. In waste energy recovery, some waste components e.g. substances that contains large amount of chlorines or flame retardants need to be removed from the stream as they may cause formation of harmful emissions or damage combustion equipment. To assist waste recycling and energy recovery, technologies that classify waste components for automated on-line sorting, preferably in real-time, are required.
Fast and non-destructive on-line measurement classifying organic materials can be achieved employing near-infrared hyperspectral imaging (NIR-HSI) technique [3]. Several studies have demonstrated its ability to identify: synthetic organic polymers [4,5] and specifically LDPE/HDPE in mixed plastic waste [6], plastic packaging [7, 8], compost samples, [9] bottom ash [10], wood-plastic composites and contaminated wood [11], and types of cardboard and paper [12]. The state-of-the-practice methods used for classification based on NIR-HSI are employing mostly linear algorithms such as partial least square discriminant analysis (PLS-DA). However, the great
potential of machine learning methods for given application has been identified based on literature review [13-16].
Therefore, the aim of this is study is to evaluate the application of NIR-HSI coupled with selected machine learning methods i.e. support vector machine (SVM), and radial-basis neural network (RBNN) and as reference the PLS-DA algorithm for classification of MSW components.
Materials:Samples included in this study are typical components of MSW and are listed in Table 1.
For each of 10 distinct categories several different sample objects were collected to simulate large variability of MSW. Composition of the organic polymer samples was identified by Nicolet iS50 FT-IR Spectrometer in ATR mode (ThermoFisher Scientific, USA) followed by the infrared spectral library search. Larger pieces of the samples were reduced to size of 10 to 30 mm when applicable. Sample objects were then randomly mixed together in each category and then randomly divided to form two different mixtures per category. Data acquired on the first mixture were used as training data set and internal validation set and data acquired on the second one as external validation set. Table 1. Samples included in the study
Spectroscopy: Spectral data were acquired on a push-broom line scanning HSI camera FX17e
(SPECIM, Spectral Imaging Ltd., Finland), with FOV = 38°, gathering approx. 1200 frames per image. The camera is equipped with an InGaAs based NIR detector with spectral range of 900-1700 nm, 224 spectral bands and 640 spatial pixels. Samples illuminated by two linear halogen light sources (were moving on a laboratory scanning table with tray dimensions 20 cm x 40 cm at velocity approx. 90 mm/s. Acquisition was done at framerate 300 fps and exposure time of 5 ms, parameters were synchronized in order to acquire images with correct aspect ratio. The
background material of the tray was a carbon black styrene-butadiene rubber a typical material of a conveyor belt. The temperature of the ambient air during the acquisition was 20±1 °C.
Analysis:
Calculations, decomposition of the hyperspectral cube, pre-processing and modeling were done using scripts programmed in MATLAB (MathWorks Inc., USA). The number of spectral bands was reduced to 192 by removing data points on the begging and end of the recorded spectral range due to low signal-to-noise ratio defined by a threshold of 30 dB. The training data set and internal validation set included different randomly selected data points (i.e. pixels), 1000 per material category. Prediction was then performed on external validation set predicting on all pixels of the image acquired on a different mixture. Spectral data were converted to relative absorbance based on raw reflectance image from the sample, dark reference image and white reference image. Data pre-processing techniques such as mean centering (MC), standard normal variate (SNV), Savitzky-Golay smoothening (SGS, 2nd order polynomial, 11 smoothing points) Savitzky-Golay 1st derivative
(SG1, 2nd order polynomial, 11 smoothing points) and Savitzky-Golay 2nd derivative (SG2, 2nd order
polynomial, 11 smoothing points) and combinations thereof were applied to the spectra and evaluated. Furthermore, machine learning classification algorithms such as partial least-squares discriminant analysis (PLS-DA), support vector machine (SVM) and radial-basis neural network (RBNN), i.e. basic single-layer artificial neural network (ANN), were employed in classification model development and validation. It is noted that no attempt to remove outliers was done. Model performance were evaluated by sensitivity (i.e. true positive rate, Eq.1) and specificity (i.e. 1 – specificity, Eq.2, defined as false positive rate) which were calculated for each class separately. Parameters are based on whether prediction is a true positive (TP), true negative (TN), false positive (FP), or false negative (FN). Overall classification accuracy of the model is then expressed as a mean value for all classes.
Equation 2.
RESULTS
Average NIR spectra of materials included in this study acquired by HSI camera and extracted from the hyperspectral cube are shown in Figure 1.
Figure 1. Relative absorbance spectra of materials included in the study
Results obtained from the identification of organic polymers (i.e. plastics) by FT-IR spectrometer reveals that samples in each group contained various amounts of plasticizers and other additives making the classification challenging. Results of subsequent classification modelling shows that data pre-processing has significant impact on resulting accuracy of classification models. Table 2 shows which of the pre-processing techniques or combinations being the most (in green) and the least (in red) successful in classification. The accuracy is expressed as average of sensitivities calculated for all of the classes according to Eq. 1. PLS-DA reaches overall the lowest accuracy. SVM reaches highest accuracy of 94 % for combined pre-processing by Savitzky-Golay
1st derivative and mean centering.
Table 3 shows results with respect to sensitivity, specificity and classification confusion, derived from the confusion matrix, for three developed classification models. The results are achieved using the best pre-processing method of highest overall accuracy, based on previous evaluation
presented in Table 2. Generally, the lowest sensitivity is achieved for incombustibles and PS. Most common misclassifications are done for incombustibles being confused with the background and for PS being confused with incombustible materials and background.
Table 3. Detailed classification results – internal validation set
Models were further externally validated, when used for prediction on hyperspectral imaging data acquired from separate validation mixtures (Figure 2).
The prediction was done by classifying each pixel of the acquired hyperspectral image. The accuracy of classification during external validation is evaluated for four different pre-processing methods presented in Table 4. Furthermore, demand on computational time is measured as the time needed for each model to classify approx. 1.5 million pixels (i.e. data points) of acquired hyperspectral images. The least demanding model on computational time is PLS-DA, followed by RBNN, requiring approx. 10 times more time, and the model most demanding on computational time is SVM, requiring approx. 100 times more time than PLS-DA.
Table 4. Prediction results – external validation set
Results of the prediction are shown in figures where different reference color of each pixel
represents a predicted class. Results shown in following figures are based on prediction using best performing pre-processing technique i.e. SG1 + MC. Figure 3 presents prediction of classes by RBNN model, Figure 4 prediction of classes by PLS-DA model and Figure 5 prediction of classes by SVM model. It is shown that consistency of classification by PLS-DA throughout the pile is lower compare to RBNN. Furthermore, when classifying by SVM a large part of the background was classified as polystyrene.
Figure 4. Prediction of classes by PLS-DA model – external validation set
Table 5 presents results for each material class. Generally, the best sensitivity was achieved for food, PS, and wood. Nevertheless, models often struggled with correct classification of
incombustible materials. Overall, best results were reached for SVM, even though the model had a very low sensitivity in background identification. Confusions between background and PS can be largely explained by their very flat spectral profiles in the given spectral range.
Table 5. Detailed prediction results – external validation set
To clarify spectroscopy background, the assignment of spectral bands to specific vibrational transitions of molecular functional groups was then performed based on regression coefficients from PLS-DA for each of the classes using spectral atlas [17] (not shown). The further from zero the regression coefficient is, the more influence on classification the given spectral band has.
CONCLUSIONS
The results demonstrated that it is possible to develop a classification model that can recognize most common components of MSW based on NIR-HSI data. The models developed based on machine learning methods can successfully classify pixels belonging to textile, wood, food, paper, and most common types of plastics with reasonable accuracy. Nevertheless, classification of
materials with no specific absorbance spectrum in the NIR region such as metals, glass, or ceramics is more challenging.
Three classification models were developed employing PLS-DA, SVM and RBNN. Models varied in complexity and achieved accuracies. The SVM model achieved accuracies of up to 94% for internal validation, 85% for external validation and was the best in identifying incombustible materials. However, due to high demands on computational time, limitations may arise in real-time application. Results of the RBNN reached 91% for internal validation and 85% for external validation. The model also predicted approx. 10 times faster than the SVM. The minimum computational time was required for model based on PLS-DA, however reaching accuracies only 88% for internal validation and 81% for external validation.
The developed classification models have a large potential be used on-line for real-time material identification at MSW sorting facilities or for pre-sorting at waste-to-energy plants. Applicability of the developed models can be further enhanced by including new samples of varying composition and new classes driven by a demand of which components need to be sorted.
ACKNOWLEDGEMENT
The presented work is based on Master's thesis authored by Martin Sevcik and supervised by Jan Skvaril: Sevcik, M. (2019) Near-infrared spectroscopy for refuse derived fuel: Classification of waste material components using hyperspectral imaging and feasibility study of inorganic chlorine content quantification. Mälardalen University, Västerås. 68 pages.