Segregating wood wastes by repetitive principal component analysis of near infrared spectra

Abstract

To improve the recycling rate of wooden materials, it is necessary to classify wood waste by disposal method and usage. In the industrial manufacture of these materials, rapid and accurate determination of their chemical and physical properties is critical for a stable supply of wood products with reliable quality. In this study, we investigated a discriminant analysis process for waste wood products using hyperspectral imaging with a newly developed repetitive principal component analysis. Hyperspectral images of four types of wood waste (plywood coated with resin, preservative-treated wood, hardwood and softwood) were acquired. The mean spectrum of each sample was extracted from a hypercube in order to build a classification model. A novel classification method based on principal component analysis, named repetitive principal component analysis, was developed. A total of three repetitions of principal component analysis were performed to classify the four types of wood waste. Cross-validated results of repetitive principal component analysis resulted in classifications greater than 85% for any of the four wood waste types. The discriminant model was then applied to single-pixel spectra of the hypercube to form a prediction map. Hyperspectral imaging, with the aid of the new repetitive principal component analysis discriminant analysis, is a powerful tool in wood recycling processes.

Keywords

Hyperspectral imaging wood waste classification repetitive principal component analysis recycling

Introduction

Wood-based materials are generally regarded as renewable and eco-friendly from the viewpoint of carbon storage and sustainable material. Once discarded, they are classified by material type and are then used as raw materials for new products such as particle boards, papers and fillers for wood–plastic composites. Some of wood waste which cannot be used in material recycling can be burnt to obtain thermal energy or landfilled. With the objective of improving the recycling ratio and reducing carbon dioxide emissions, it is necessary to improve the longevity of wood as a ‘wood-based material’. Wooden constructions comprise many types of wood-based materials: plywood for sheathing and walls, solid wood and glue-laminated timber for posts and beams or particle or fibre board for furniture. Since most wood waste is produced as demolition debris, it must be classified properly based on the purpose of recycling in order to improve the material recycling ratio.

When it comes to wood recycling mills, the shape and surface condition of wood waste collected from demolition sites vary widely. It is necessary to develop an analysis technique that can rapidly measure a large area without contact. Spectral imaging might be suitable for assessing a massive amount of wood waste at a time. From a material recycling standpoint, it is necessary to eliminate non-recyclable materials from the overall waste. Furthermore, solid wood must be separated from other wood waste.

In particle board production using recycled wood, wood waste is checked using a metal detector, in order to detect metal fastenings and fittings. It is then separated into chips and other non-wood components via sieving and centrifugation. However, these methods cannot grade wood waste by the type of wood-based material. Although visual inspection enables the classification of such materials or tree species, it is not realistic because it is slow and labour intensive, which results in inefficient processing.

Near infrared (NIR) spectroscopy has been widely used for fast, non-destructive quality control in numerous fields. Furthermore, in the field of wood science, the potential of NIR spectroscopy as a non-destructive tool has been demonstrated by many researchers.^1–8 The basic technique for assessing wood with NIR spectroscopy was summarized by Sandak et al.¹ In addition to the determination of chemical composition,² NIR spectroscopy has a potential for the quality evaluation of logs,³ identification of wood species⁴ and classification of wood and wood-based material.⁵ There are many absorption bands in wood-based materials due to the overtone and combination of the vibrations of functional groups such as CH, OH and NH.⁹ In previous research, the feasibility of near infrared imaging for the identification of wood-based materials was examined.^6–8 Recently, NIR imaging analysis has been used with hyperspectral imaging (HSI) for the determination of wood density,^10,11 moisture content,¹² assessment of the extractives content¹³ and identification of compression wood.¹⁴ Based on the concentrations of chemical components, NIR can be applied to both qualitative and quantitative analyses.

However, diffuse reflectance light is strongly influenced by scattering created at the surface of an object. Thus, the surface condition must be controlled if NIR spectroscopy is to be used for industrial applications.¹⁵

Since an NIR spectrum comprises hundreds of variables that have strong multicollinearity, multivariate analyses, such as the soft independent modelling of classification,^{6–8,16–18} partial least squares regression discriminant analysis^19,20 and artificial neural network or support vector machine^20,21 are often employed. These classifications are mostly based on the projection of a vector in linear space to enhance the differences between classes. However, projecting in linear space means compressing spectral information, which sometimes results in the loss of smaller spectral variation. When several types of classes are considered, those with small spectral differences might not be classified adequately.

The aim of this study was to develop a classification technique based on NIR spectral analysis. HSI was employed to obtain the reflectance spectra with high spatial resolution. Four types of wood waste (plywood coated with resin, preservative-treated wood, softwood and hardwood) collected from a recycling mill were analysed to build a discrimination model using the newly developed classification method named repetitive principal component analysis (rPCA). The classification ratios of the discrimination model were validated using a cross-validation method, and finally, models were applied to build a prediction image.

Materials and methods

Sample preparation

Wood waste was collected from a wood recycling mill in Nagoya city. Figure 1 shows the collection site. These materials originate from the demolition debris of wooden constructions. The collected samples were classified into four categories: solid wood (softwood), solid wood (hardwood), preservative-treated wood and plywood coated with resin (Figure 2). The preservative-treated wood was impregnated with creosote. The surface of the plywood coated with resin was coated with a polyurethane resin. All the samples were cut into 50 mm × 50 mm × 15 mm squares, and their surfaces were wiped with a dry paper towel to remove any mud. Twenty-five specimens were prepared for each category (a total of 100 specimens).

Figure 1.

View of a pile of wood at a recycling mill.

Figure 2.

Representative surfaces of four types of wooden waste: (a) plywood coated with resin, (b) preservative-treated wood, (c) hardwood and (d) softwood.

Image acquisition and spectra extraction

Image data acquisitions were carried out using a HSI system (Compovision, Sumitomo Electronic Industries, Ltd, Osaka, Japan). This is a push-broom line scanning camera covering the spectral range of 1046–2546 nm with a 6 nm wavelength interval. The standard pixel size was 0.44 mm × 0.44 mm with a 140 mm field width. The moving speed of the conveyer was set to 45 mms⁻¹, and the exposure time was set to 3 mms⁻¹ for each line. Finally, 534 (lines) × 237 (wavelength) × 320 (field width) pixels of hypercube data were acquired. Each pixel corresponded to a 16-bit reflectance value. Two halogen lamps were placed over the conveyer to illuminate the samples at an angle of 45° from both sides of the sample. Dark signals were acquired by covering the lens with the lens cap, and a white plate image was acquired as a reference before the acquisition of sample images. Images of four samples were taken simultaneously. Both were acquired before each sample image acquisition. The raw data were transformed into corrected reflectance as follows

Reflectance = \frac{Sample - Dark}{Reference - Dark}

(1)

where Sample is the spectra value of the sample image and Dark and Reference are the dark and white plate signals, respectively, at the same pixel position. The mean spectrum of each sample surface (50 × 50 pixels) was obtained from the hypercube and each of them was transformed into a second derivative spectra by Savitzky–Golay filter.

Data processing – rPCA

The aim of this study was to develop a novel classification model for waste wood products based on HSI. The authors developed the concept of a classification method based on rPCA.²² As with conventional classification methods, rPCA needs model building using calibration data that are set before the practical application. The details of rPCA are as follows: a first round of PCA is performed for all the calibration samples consisting of N classes. Then, one class is classified by the single threshold value for a particular score value. In this study, PC2 scores were used for classification. The first round of the second PC2 loading vector and the threshold value for the PC2 score will be used as the classification model for Class 1. Then, PCA is performed again without the Class 1 sample. Another class is classified by the single threshold value for the second round of the PC2 score value. PCA is repeated N−1 times until all classes were classified. The outline of model building and discrimination of unknown samples by rPCA is illustrated in Figure 3.

Figure 3.

The repetitive principal component analysis (rPCA) process.

As mentioned above, rPCA classifies each category one by one. In other words, each round of PCA discriminates the samples into two groups which are the target group and the remaining groups. Only a few PCs will be required for each discrimination. In this study, PCA was performed without any pre-treatment such as mean centring and auto scaling. In this case PC1 score will not be mean centred. Therefore, PC2 was selected to be the discriminating principal component. All samples are ultimately assigned a certain class. Furthermore, rPCA is performed repeatedly with different data sets. Thus, the loading vectors used for the discriminant model are not necessarily orthogonal to each other. These characteristic features will contribute to a robust classification.

Results and discussion

Band assignment of mean spectra

Figure 4 shows the example of second-derivative spectra of four kinds of the wood wastes. Several characteristic absorptions are summarized in Table 1.⁹ Strong absorptions around 1433 and 1924 nm were observed for all types of wood waste. The 1433 nm absorption was assigned to the first overtone of the OH symmetric stretching vibration due to cellulose. The 1924 nm absorption was due to a combination of the deformation and asymmetry stretching vibration of OH. There were also several characteristic absorption peaks for each wood waste type. Absorptions at 1205 and 1685 nm in the spectra of preservative-treated wood were assigned to the second and first overtones of CH stretch, respectively, in polycyclic aromatic hydrocarbons. In the case of plywood coated with resin, there were absorptions at 1685 nm (first overtone of the stretching vibration of CH) and 2156 nm (combination due to the amide bond) due to the surface polymer. The reflectance spectra of softwood and hardwood were similar. However, hardwood had a small absorption at 1717 nm (CH₃). This may due to the different structure of lignin: in softwood, only guaiacyl lignin exists, whereas the lignin in hardwood comprises both guaiacyl and syringyl.

Figure 4.

Representative second-derivative spectra of the wooden wastes and their significant absorption bands: 1 – 1205 nm, 2 – 1433 nm, 3 – 1685 nm, 4 – 1717 nm, 5 – 1924 nm and 6 – 2156 nm. Note that the raw spectra were obtained as reflectance, thus absorption peaks of second derivative appear as positive peak.

Table 1.

Summary of band assignments.

Numbers in Figure 4	Band location (nm)	Bond vibration	Component⁹
1	1205	CH str. 2nd ov	Polycyclic aromatic hydrocarbons
2	1433	OH str. 1st ov	Cellulose
3	1685	CH str. 1st ov	Polycyclic aromatic hydrocarbons/polymer in waterproof coating
4	1717	CH str. 1st ov	Lignin
5	1924	OH def. + OH str.	Water
6	2156	Combination of amide	Polymer in waterproof coating

Discriminant model building

In this study, rPCA was conducted on 100 mean spectra, and PC2 was applied for discriminating each type of wood waste. The PC2 score values for classification are shown in Figure 5. The threshold for each classification model was set to zero. Plywood coated with resin was classified by the first PCA. The second PCA was then performed on the remaining 75 spectra without plywood coated with resin. As shown in Figure 5, the second round of PCA classified preservative-treated wood from untreated wood. Finally, softwood was discriminated by the third PCA.

Figure 5.

Principal component analysis (PCA) score values for the discriminant model of (a) plywood coated with resin, (b) preservative-treated wood and (c) softwood.

Figure 6 shows the PC2 loading vector of first, second and third rounds of PCA in rPCA. As shown in Figure 6(a), loading was high at 1685 and 2156 nm, where the second-derivative spectra of plywood coated with resin had strong absorption. This indicates that the synthetic polymer applied to the surface of plywood has contributed to its classification. The PC2 loading vector of the second PCA, which was used for the classification of preservative-treated wood, had the highest value at around 1205 nm. This wavelength range exhibits absorption because of the polycyclic aromatic hydrocarbons of creosote. These two discriminant models (first and second PCA) projected spectral differences based on the characteristic absorption bands for each wood waste type. In contrast, the difference of the spectra between softwood and hardwood was rather smaller because they were solid, and no artificial chemical compositions were added. In contrast to the other loadings, the PC2 loading of the third PCA did not have any wavelength which dominated the loadings. However, in this study softwood and hardwood were correctly separated. As shown in Figure 6(c), various wavelengths are involved in discriminating softwood. In the PC2 loading of the third PCA, there was a negative peak at 1717 nm, which is assigned to the CH of lignin. Probably the difference between the lignin structures of softwood and hardwood contributed to the classification.

Figure 6.

Loading vectors for the discrimination of (a) plywood coated with resin, (b) preservative-treated wood and (c) softwood.

Cross-validation was then conducted to determine the correct classification ratio. The validation samples were selected based on the method called Venetian Blinds. Eighty of the 100 samples were used for calibration, and the remaining 20 samples (five samples each of the four types of wood waste) were subjected to validation. Validation samples were selected by the PC2 score value of the first round of PCA calibration. Each sample was sorted by PC2 score value within each category and every fifth sample was selected as a validation. Cross-validation was performed five times by changing the combination of validation samples. As summarized in Table 2, the average correct classification ratios of plywood coated with resin, preservative-treated wood, softwood and hardwood were 100, 90.7, 88.0 and 88.6%, respectively. What is most important for the recycling of wood waste is to separate the hazardous material (in the case of this study, preservative-treated wood) from the other recyclable materials. The correct classification ratio for preservative-treated wood (90.7%) is either equal or surpassing.⁸ When higher classification ratios are required, we need only to set the threshold at a lower value. However, when the threshold is lowered, the likelihood of softwood or hardwood being classified as preservative-treated wood increases. Throughout this study, all the thresholds were set to zero. Of course, another method such as the percentile or mode method could also be applied. In practical use, the threshold could be changed to match the required tolerance of the discrimination model.

Table 2.

Correct classification ratios from cross-validation.

N of cross-validation	Plywood coated with resin	Preservative-treated wood	Softwood	Hardwood
1	100	86.7	90.0	83.3
2	100	100	100	100
3	100	80.0	90.0	100
4	100	93.3	80.0	80
5	100	93.3	80.0	80
Average	100%	90.7%	88.0%	88.6%

Mapping of rPCA discriminant model

Figure 7 shows the RGB images of the four types of wood waste and the false colour image calculated by each classification model. For each false colour image, the range was scaled from zero to the maximum score. As shown in Figure 7(b), plywood coated with resin is clearly classified. In the case of classifying preservative-treated wood (Figure 7(c)), several pixels showed negative value, which means that these pixels are classified as other types of wood. However, the average score value of preservative treated wood was greater than zero. By contrast, the mean values of softwood and hardwood were less than zero, four-thirds of the pixels (76%) showed positive value. When building the classification model, the spectra of each sample were averaged. Thus, the spectral variation caused by the position of the sample was reduced. Furthermore, the S/N ratio of the spectra used for model building was higher than that used for mapping. With the aid of another type of image processing such as segmentation, this technique will be ready for use in recycling mills for wood-based materials.

Figure 7.

(a) Visible images of four types of wooden wastes, (b) score image for plywood coated with resin discrimination, (c) score image for preservative-treated wood discrimination and (d) score image for softwood discrimination.

Conclusion

In this study, a new discrimination method (rPCA) was developed to evaluate the feasibility of discriminating four types of wood waste: plywood coated with resin, preservative-treated wood, softwood and hardwood. The first round of PC2 scoring was used for discriminating plywood coated with resin, and the second round of PC2 scoring was then used for classifying preservative-treated wood. Finally, the third round of PCA was performed for softwood and discrimination. Cross-validation showed sufficient accuracy (over 80% correct classification) for classifying the four types of wood waste. In practice, the threshold could be changed to match the required tolerance of the discrimination model. With rPCA, classification will be performed for the class which has large spectral differences compared to other classes. In this study, the spectral differences between plywood coated with resin, preservative-treated wood and solid wood were large. By comparison, for solid wood, the spectral difference between softwood and hardwood was much smaller. Despite such small spectral differences, rPCA enabled the correct classification of softwood and hardwood. These results suggest that rPCA could be a powerful tool for wood recycling mills. Despite this, there are still a number of problems to be solved. For example, the HSI images were acquired under controlled conditions, with the same distance and angle between the measuring surface and focal plane. In a recycling mill, conditions including the angle, size and surface roughness of wood waste vary widely. Some of these problems can be solved by image processing. Others might become much easier to solve if the sample surface conditions (angle and size) can be physically controlled. Therefore, the development of an imaging system with a hardware that can obtain a flat, uniform image of samples is required in order to apply this technique in recycling mills.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

References

Sandak

Meder

. Tutorial: assessing trees, wood and derived products with near infrared spectroscopy: hints and tips. J Near Infrared Spectrosc 2016; 24: 485–505.

Ramadevi

Hegde

Varghese

. Evaluation of lignin syringyl/guaiacyl ratio in Eucalyptus camaldulensis across three diverse sites based on near infrared spectroscopic calibration modelling with five Eucalyptus species and its impact on Kraft pulp yield. J Near Infrared Spectrosc 2016; 24: 529–536.

Sandak

Böhm

. Near infrared spectroscopy as a tool for in-field determination of log/biomass quality index in mountain forests. J Near Infrared Spectrosc 2016; 24: 587–594.

Lang

Costa

FRC

Camargo

JLC

. Near infrared spectroscopy facilitates rapid identification of both young and mature Amazonian tree species. PLos One 2015; 10: e0134521.

de Groot

Postma

Melssen

. Influence of wavelength selection and data preprocessing on near-infrared-based classification of demolition waste. Appl Spectrosc 2001; 55: 173–181.

Tsuchikawa

Yamato

Inoue

. Discriminant analysis of wood-based materials using near-infrared spectroscopy. J Wood Sci 2003; 49: 275–280.

Tsuchikawa

Yamato

. Discriminant analysis of wood-based materials with weathering damage by near infrared spectroscopy. J Near Infrared Spectrosc 2003; 11: 391–399.

Kobori

Yonenobu

Noma

. Discriminant analyzing system for wood wastes using a visible-near-infrared chemometric imaging technique. Appl Spectrosc 2008; 62: 854–859.

Schwanninger

Rodrigues

Fackler

. A review of band assignments in near infrared spectra of wood and wood components. J Near Infrared Spectrosc 2011; 19: 287–308.

10.

Fernandes

Lousada

Morais

. Measurement of intra-ring wood density by means of imaging VIS/NIR spectroscopy (hyperspectral imaging). Holzforschung 2013; 67: 59–65.

11.

Mora

Schimleck

Yoon

. Determination of basic density and moisture content of loblolly pine wood disks using a near infrared hyperspectral imaging system. J Near Infrared Spectrosc 2011; 19: 401–409.

12.

Kobori

Gorretta

Rabatel

. Applicability of Vis-NIR hyperspectral imaging for monitoring wood moisture content (MC). Holzforschung 2013; 67: 307–314.

13.

Lestander

Geladi

Larsson

. Near infrared image analysis for online identification and separation of wood chips with elevated levels of extractives. J Near Infrared Spectrosc 2012; 20: 591–599.

14.

Meder

Meglen

. Near infrared spectroscopic and hyperspectral imaging of compression wood in Pinus radiata D. Don. J Near Infrared Spectrosc 2012; 20: 583–589.

15.

Thumm

Riddell

Nanayakkar

. Mapping within-stem variation of chemical composition by near infrared hyperspectral imaging. J Near Infrared Spectrosc 2016; 24: 605–616.

16.

Adedipe

Dawson-Andoh

Slahor

. Classification of red oak (Quercus rubra) and white oak (Quercus alba) wood using a near infrared spectrometer and soft independent modelling of class analogies. J Near Infrared Spectrosc 2008; 16: 49–57.

17.

Fujimoto

Tsuchikawa

. Identification of dead and sound knots by near infrared spectroscopy. J Near Infrared Spectrosc 2010; 18: 473–479.

18.

Watanabe

Hart

Mansfield

. Detection of wet-pockets on the surface of Tsuga heterophylla (Raf.) Sarg. by near infrared (NIR) spectroscopy. Holzforschung 2010; 64: 55–60.

19.

Batista Braga

Monteiro Pastore

Rauber Coradin

. The use of near infrared spectroscopy to identify solid wood specimens of Swietenia macrophylla. IAWA J 2011; 32: 285–296.

20.

Nascimbem

LBLR

Rubini

Poppi

. Determination of quality parameters in moist wood chips by near infrared spectroscopy combining PLS-DA and support vector machines. J Wood Chem Technol 2013; 33: 247–257.

21.

Mora

Schimleck

. Kernel regression methods for the prediction of wood properties of Pinus taeda using near infrared spectroscopy. Wood Sci Technol 2010; 44: 561–578.

22.

Tsuchikawa S, Kobori H and Higa S. Method of classifying and discerning wooden materials. Patent US9164029 B2, USA, 2015.