Estimation of the earlywood and latewood ratio with different particle size of dry and wet milled Japanese cedar wood flour by using hyperspectral imaging

Abstract

Mechanical properties of wood–plastic composites are influenced by a particle size and surface morphology of wood flour. Generally various sizes of wood flour are produced from single solid wood even if the single process is used. If the different particle size of wood flour is produced from different wood tissue such as earlywood (EW) and latewood (LW), not only particle size but also density and chemical composition of wood flour might influence the mechanical property of final products. This study aims to investigate the relationship between particle size and their origin; EW and LW. EW and LW were separately milled to produce the EW and LW flour by dry and wet milling. Hyperspectral images ranging 400–1000 nm for each wood flour were used as training data. Discriminant model of EW and LW flour developed by PLS-DA showed over 0.77 of accuracy. Then the EW and LW were dry and wet milled together and screened by three different sieve openings to obtain different particle size wood flour. Discriminant model was applied for the hyperspectral images of each size of wood flour to estimate the EW and LW ratio. The result showed that increasing sieve opening resulted in the increasing ratio of LW for dry milled wood flours. The results suggest that the EW was easily pulverized than LW.

Introduction

Wood–plastic composite (WPC) is a product made from wood flour and thermoplastic resin. Because of higher water resistance than solid wood products and some wood-like texture, recently WPC is widely used in exterior. It has been reported that the particle size of wood flour affects the mechanical properties of WPC [1]. The various factors such as types of milling machine and condition result in the variety of size and morphology of wood flour. The effects of ball milling and drying condition of wood flour on the mechanical property of WPC was investigated by Murayama et al. [2]. The highest tensile strength and modulus of rupture (MOR) were observed when 30 min of ball milled wood flour was used. The dominant particle size of wood flour produced by 30 min of ball milling was 100–300 µm. Particle size of wood flour is one of the most considered factors for manufacturing WPC, since related characteristics such as aspect ratio or specific surface area also influence the mechanical properties of WPC [3].

By the way, both physical and chemical properties of wood vary within the position of log, i.e., juvenile and mature wood, heart and sapwood, or earlywood (EW) and latewood (LW), etc. In case of softwood, the density of LW is over 2 times higher than that of EW [4]. There is a strong relationship between density and strength. There is a possibility that the grindability of EW and LW is different, which may result in the different particle size of EW and LW flour, even the same milling process.

Nowadays visible–near infrared spectroscopy has been widely used for non-destructive evaluation into various fields [5], including wood science [6]. Recently hyperspectral imaging has become commonly used for the evaluation of biological materials [7]. Hyperspectral imaging can acquire a huge number of spectra at once, thus it is suitable to visualize the spatial distribution of moisture content [8], and density [9] within wood.

This study aimed to investigate the relationship between the particle size of wood flour and origin of wood tissue. Wood flour was produced from a heart wood of Cryptomeria japonica by two types of milling, and then separated by the particle size. EW and LW flour discriminant models for two different milled wood flours were developed by hyperspectral imaging with aid of partial least squares discriminant analysis (PLS-DA). Finally, EW and LW ratio for 3 different particle sizes of wood flour were estimated.

Experimental

Sample preparation

Experimental procedure is summarized in Fig. 1. A heartwood lamina of Cryptomeria japonica (longitudinal (L): 500 × radial (R): 50 × tangential (T): 10 mm) was used as a raw material. As a training sample, EW and LW regions were separated by a tabletop band saw, small hand saw and belt sanding machine. As a test sample, several small blocks with the size of L:5 × R:50 × T:10 mm were cut from the same lamina by a tabletop bandsaw. Then EW, LW and small blocks for test were firstly dry milled by a cutting mill (Pulverisette 15, Fritsch Japan Co., Ltd.) to produce coarse wood flour (In this study, these wood flours are called dry milled wood flour.)

Fig. 1

Schematic diagram of the experimental procedure (D: dry milled W: wet milled)

Then, 13.5 g of dry milled wood flours were wet milled by a ball milling machine (Pulverisette 5, Fritsch Japan Co., Ltd.) with 200 ml of distilled water for 30 min at 200 rpm. After wet milling, vacuum filtration, tert-butyl alcohol substitution and centrifugation were performed to the wood flours. Centrifuged wood flours were once cooled in a fridge, then freeze-dried (FDU-1200, Tokyo Rikakikai Co., Ltd.) under − 45 °C with 20 Pa for 1 week. To reduce the aggregation, freeze-dried wood flours were defibrated by a food mixer (IFM-800DG, Iwatani Co.). In this study, those wood flours are called wet milled wood flour.

Eight grams of dry and wet milled wood flour for the test were finally classified by particle size using a sieve shaker. Sieve openings for dry milled and wet milled wood flour were 710, 500, 300 and 425, 300, 212 µm, respectively. Note that all wood flours were not assigned to those 3 mesh sizes, several amounts of wood flour passed the smallest sieve opening.

Hyperspectral image acquisition

Hyperspectral images of EW, LW and 3 different sizes of both dry and wet milled wood flours for test were acquired (NH1-CK, Eba Japan Co., Ltd.). Acquired wavelength range, interval and were 400–1000 nm and 5 nm, respectively. A lens of 16 mm focal length with F1.8 was mounted. Distance between lens of the camera and sample was fixed as 390 mm, which resulted in the spatial resolution of 0.136 mm/pixel. Each of the wood flours was filled in a glass frame of 25 × 75 × 1 mm. Pixel values of hyperspectral data were converted into reflectance by a dark image and barium sulfate images as a reference. All images were trimmed by 100 × 400 pixels. Then the wood flour was removed and filled again to acquire the hyperspectral image. Finally, 5 images per wood flour were acquired to evaluate the variation according to sampling conditions.

Discriminant model development

A classification model of EW and LW flour for dry and wet milled wood flour were developed by reflectance spectra obtained from hyperspectral images. Generally, particle size affects the baseline drift of the spectra [10]. Therefore, standard normal variate (SNV) was performed to eliminate the baseline [11] for the spectra ranging from 450–950 nm. For both dry and wet milled wood flour, a total of 20,000 spectra (10,000 EW spectra and 10,000 LW spectra) were randomly selected as a training set. PLS-DA was applied to develop a discriminant model. Maximum latent variables were fixed as 5, and leave-one-out full cross-validation was used to evaluate the prediction accuracy. Ratio of true class within predicted class (precision) for EW and LW flour (p_e and p_l) were calculated by a result of a cross-validation. All data processing was performed by Matlab r2020b (Mathworks. Inc.). All procedures were repeatedly performed for 5 subset of data set.

Estimation of EW and LW ratio for different particle size wood flour

The developed discriminant models were applied to the hyperspectral data of different particle sizes of wood flours to estimate the EW and LW ratio. All spectra were transformed by SNV, and then all pixels (40,000 spectra per particle size) were assigned to early or LW pixels. The ratio of EW and LW pixels in the images (r_e and r_l) were calculated for different sizes of wood flour images. Finally, the EW ratio (R_e) were calculated as follows:

$${R}_{\text{e}}={p}_{\text{e}}{r}_{\text{e}}+\left(1-{p}_{\text{l}}\right){r}_{\text{l}}.$$

(1)

All procedures were repeatedly performed for 5 subset of data set.

Results and discussion

Reflectance spectra

Figure 2 shows the representative reflectance images of EW and LW flour at 600 nm. Reflectance of EW flour is generally higher than that of LW flours for both dry and wet milled condition.

Fig. 2

Representative reflectance images of wood flours (λ = 600 nm)

Figure 3 shows the mean reflectance spectra of EW and LW flour with SNV treatment. Generally, reflectance of wood increases as longer wavelength at 450–950 nm [12]. As shown in Fig. 3, spectra of EW and LW flour tended to have quite similar shapes for both dry and wet milled conditions. However, reflectance of LW was slightly higher than that of EW around 500 nm. On the solid wood, EW and LW are easily distinguished because LW has a darker color. This color difference is due to the density (cell wall thickness). However, in the case of Norway spruce, lignin content of EW is higher than LW on heartwood [13]. Difference of reflectance around 500 nm might be due to the different amount of lignin. Generally, smaller particle size shows lower aspect ratio [14]. Spectral difference among EW and LW also might be influenced by the morphological difference between EW and LW or the particle orientation inside the frame.

Fig. 3

Mean SNV-treated reflectance spectra of EW and LW flour: a dry milled b wet milled

Mean spectra of different particle size wood flours are shown in Fig. 4. In the case of dry milled wood flour (see Fig. 4a), reflectance around 500 nm and 900 nm increased as increasing mesh size. As shown in Fig. 4b, a similar relationship between particle size and reflectance was observed, however the mean spectra of medium and small wood flours (300 and 212 µm mesh on) almost overlapped in the whole wavelength region. These results suggest that the spectra of large wood flour are similar to that of LW flour, and that of medium and small wood flours are similar to that of EW flour.

Fig. 4

Mean SNV-treated reflectance spectra of wood flours with different size: a dry milled b wet milled

Discriminant accuracy of EW and LW discriminant model

Discriminant accuracy of EW and LW flour discriminant model for dry milled wood flour is summarized in Table 1. Mean p_e and p_l were 0.819 and 0.777, respectively. Sensitivity of LW was about 0.06 higher than that of EW. Mean of overall accuracy was 0.797. Standard deviation of p_e and p_l for 5 subset of dry milled wood four were 0.002 and 0.004, respectively, which means less than 0.6% of coefficient of variation. As shown in Table 2, mean p_e and p_l for wet milled wood flour were 0.684 and 0.685, respectively. Mean sensitivity of LW and EW were almost same. Mean of overall accuracy was 0.685. Standard deviation of p_e and p_l for 5 subset of wet milled wood four were 0.009 and 0.011, respectively, which means over 1% of coefficient of variation. In general, discriminant accuracy for wet milled wood flour was lower than dry milled wood flour and was greatly influenced by sampling condition.

Table 1 Confusion matrix of discriminant model for dry milled EW and LW flour

Table 2 Confusion matrix of discriminant model for wet milled EW and LW flour

PLS-DA is a discriminant model based on linear regression, which predicts a dummy variable. In this study, dummy variables of EW and LW were assigned to 1 and − 1, respectively. Figure 5 shows the mean regression vector of 5 subset for two models. Predicted value was calculated as the inner product of the regression vector and pre-treated spectra. Thus the regression vector represents a wavelength region which contributes a discrimination. For both models, there is a positive peak around 500 nm and negative peak around 550–600 nm. As shown in Fig. 3, reflectance of LW at 500 nm is higher than that of EW. On the other hand, reflectance of LW longer than 550 nm shows lower than that of EW. Those two peaks on the regression vector indicates the spectral difference around 500–600 nm contributes to the discrimination of EW and LW.

Fig. 5

Regression vectors of discriminant model for EW and LW flour generated by PLS-DA: a dry milled b wet milled

As described in Tables 1 and 2, the overall accuracy is not extremely fine, but these models clarified that the simple linear discriminant model is functioning.

Relationship between estimated EW ratio and particle size

Table 3 shows the estimated EW ratio with different particle size. r_e is the ratio of the pixel directly assigned as EW by the PLS-DA model, and R_e is finally estimated EW ratio adjusted by p_e and p_l. In case of dry milled wood flour, more than half of the pixels were assigned to EW for all particle sizes. Instead of dry milled, less than half of the pixels were assigned as EW for wet milled wood flour. The difference of total R_e between dry and wet milled is likely to happen because those three particle sizes did not cover all produced wood flour in this study.

Table 3 Estimated r_e and R_e of dry milled and wet milled wood flour with different size

As shown in Table 3, R_e decreased as a larger particle size for dry milled conditions. On the other hands, there is no noticeable relationships between particle size and R_e for wet milled conditions. This may be due to the lower discriminant accuracy of the model (lower precision increases the difference between R_e and r_e, and R_e becomes closer to 0.5). The cell wall of EW is thinner than LW, therefore EW is easily pulverized when the same cutting force is applied. Kurata et al. reported that there were more hemicelluloses forming strong hydrogen bonds in LW than in EW which produced a strong cell wall framework [15]. This may also contribute the less pulverization of LW by both dry and wet milling.

Conclusions

In this study, EW and LW ratio for 3 different particle sizes of wood flour were estimated from Cryptomeria japonica heartwood. Application of discriminant model developed by hyperspectral imaging revealed that large size of wood flour had higher LW ratio for dry milled condition. The results suggest that when the produced wood flour was separated by a particle size, not only the particle size and morphology, but also the density or mechanical property of wood flour may differ.