The aim of this study was to evaluate the performance of multivariate models using Near infrared (NIR) spectra for predicting total extractives content of solid and powdered wood of planted and native species from tropical savanna. NIR spectra were recorded on the milled wood and radial surface of solid wood specimens of
The chemical composition of wood is complex. Wood tissues are composed of many chemical components that are unevenly distributed as a result of the anatomical structure (
Wood structure has been described as an interpenetrating system of polymers comprising holocellulose and lignin (
Although the extractives contribute to only a small percentage of the wood composition, the solubility of extractives in wood has importance regarding its properties and, therefore, in industrial applications (
The traditional method to determine the extractive content of wood is time-consuming, costly, and operationally inappropriate for large sample quantities. There is no universal solvent that removes all substances; each solvent is selective for one or more classes of extractives, thus the use of a sequence of solvents is needed. In general, an extraction sequence initiated by the ethanol/toluene mixture, followed by ethanol and finished by hot water is used to quantify the extractives (
In regard to the NIR based models for wood extractives of softwoods,
NIR models for extractive content of hardwoods also are reported in the literature.
The above studies demonstrated the potential of spectroscopy in estimating the extractive content in several species of wood using several solvents. However, the influence of the sample preparation (solid or powdered wood) on the fit of the predictive models for extractive content is still not clear. Thus, the first objective of this study was to develop predictive models to monitor the total extractive content of planted and natural forest wood from NIR spectra measured in solid wood and powder, and validate them using independent sets. Moreover, the study also aimed to verify whether NIR models developed for wood specimens from planted forests can predict the extractive content of wood specimens from natural forest species.
The wood samples used in this study were collected from different forest species from the Cerrado (tropical savanna) and commercial plantation biomes in Brazil. The Cerrado species (21° 10′ S, 44° 54′ W) are:
For each plant material, six specimens with dimensions of approximately 35 × 35 × 100 mm (R × T × L) representing native and commercial plantation woods, were prepared as described by
For the determination of the total extractives of the wood, the technical standard NBR 14853 (
After the extraction process, the porous crucibles with extractive-free wood powder were placed in an oven at 103 ± 2 °C until reaching constant mass, and were later weighed to determine their dry mass. The total extractive content of the samples was then quantified by the difference between the dry mass of the material before and after extraction. The volatile organic compounds possibly emitted as gases during the drying were not taken into account in the calculations.
NIR signatures were recorded initially in solid wood and, after processing the material, in the powdered wood before the extraction process. An MPA FT-NIR® spectrometer (Bruker Optik GmbH, Ettlingen, Germany) was used to acquire the NIR spectra, which is a Fourier transform (FT) based spectrometer equipped with an integrating sphere and an optical fiber. The software OPUS ver. 7.5 (Bruker) was used for data storage.
NIR signatures of solid wood were recorded directly on the radial face (transition between core and sapwood) of the specimens, in an air-conditioned room, using the integrating sphere (
The same procedure and equipment for acquisition of the solid wood spectra was used for recording the spectra in the powdered wood, but the two acquisition paths were used: integrating sphere and optical fiber (
Principal Component Analysis (PCA) and Partial Least Square Regression (PLS) analyses were conducted using the free software Chemoface ver. 1.61 (
PCA was used to verify the spectral similarity of the different samples of the species used, both from solid wood and powdered wood spectra. The PLS regression was used to correlate the NIR spectral data (independent variables) and the extractive content obtained by wet-chemistry analyses (dependent variables). NIR spectra obtained by the two acquisition pathways (integrating sphere and optical fiber) on solid wood and powdered wood at equilibrium moisture were used for these analyses.
Calibration models were cross-validated by leave-one-out and external set validation methods. The external validation was carried out using two different approaches: first, PLS models were developed with 75% of the randomly chosen specimens and validated with the remaining 25%. In the second step, the models were calibrated using data from five species and the validation was done against the sixth species (not included in the calibrated model). This procedure was repeated six times, for the six plant materials.
The models were developed with eight latent variables (LV). Preliminary tests indicated that eight LVs minimized the root mean standard error (RMSE) and maximized the coefficient of determination (R²). No samples were considered outliers. The criteria for selection of the prediction models were the values of coefficient of determination of the cross-validation leave-one-out (R²cv) and for prediction (R²p), root mean standard error of the cross-validation (RMSECV) and ratio performance deviation (RPD).
The natural species and the
The descriptive statistics of the values of total extractives determined by traditional laboratory method for each plant material which was used as reference for model calibration are shown in
The mean NIR spectra obtained by diffuse reflection of solid wood through the integrating sphere and of the powdered wood obtained by the two acquisition pathways of the equipment are shown in
The most prominent absorption bands in all plant materials are approximately at 7000 and 5100 cm-1 (
PCA was carried out using the raw spectra obtained by integrating sphere from solid wood and powdered wood (
The contribution of each principal component in the spectral variance obtained in the two approaches (solid and powder) is presented in the graph axis of
The PCA scores resulted in better classification, as regards the grouping (dashed groups) according to the species, when using the spectra obtained through the solid wood (
Overall, a better discrimination between species was obtained when PCA was performed on solid samples than on wood powder data. A possible explanation is that the spectrum measured from solid wood contains physical information from the wood specimens, in addition to variations in their chemical composition. Spectra capture variations due to wood texture, color, brightness, surface quality, etc. On the other hand, the physical component effect is canceled out in the wood powder spectra. A possible physical effect on the wood powder spectra could be the response to the particle size, though in this study the particle dimension was kept constant (40 to 60 mesh) for all wood samples.
The PCA loadings from untreated (
Partial Least Squares (PLS) regression analysis was performed with the aim of obtaining global models for the estimation of total extractive content of the wood using both original spectra and those treated by the first derivative method. The results of these analyses are summarized in
In general, the application of the first derivative pretreatment was not efficient in estimating the extractive content of the wood, since all the models presented lower calibration and validation results when submitted to such method (
All the models based on spectra obtained from integrating sphere had better statistics when compared to the models generated by optical fiber, as seen for Model 3 (sphere) and Model 7 (fiber), which had respectively R²cv values of 0.85 and 0.71, RMSECV values of 1.19% and 1.65% and an RPD of 2.57 and 1.85. Therefore, the integrating sphere is more suitable for the acquisition of spectra when the aim is to estimate the extractive content of the wood. It is important to note that this acquisition path presents a surface area approximately 100 times larger than the surface area of a fiber-optic probe.
Most of the models generated from wood spectra at 0% moisture presented statistics well below the models generated from wood spectra at equilibrium moisture (moisture of approximately 12% -
The most robust statistics in estimating extractive content from wood were obtained using Model 1 (R²cv = 0.87, RMSECV = 1.08%) which is based on spectra acquired through integrating sphere and solid wood at equilibrium moisture. Model 3 (R²cv = 0.85. RMSECV = 1.19%) was based on NIR of wood powder at equilibrium moisture. Despite the difference in the associated statistics, the two models presented satisfactory results and could be used successfully.
In the first independent validation, 75% of the samples were used for the calibration of the models while the remaining 25% were used for their validation. For this, we used original spectral data without mathematical treatments. The results of these analyses are summarized in
Again, the integrating sphere showed better performance, as indicated by the statistics related to the generated models (
Better predictions were obtained by spectra measured in solid wood (model 11: R²p = 0.93, RMSEP = 0.95%) when compared to spectra from ground wood (model 12: R²p = 0.87, RMSEP = 1.40%). However, in both cases these models can be considered promising tools for the prediction of total extractive contents of unknown wood.
A strong correlation between the values predicted by NIR analysis and those measured by wet-chemistry in laboratory was detected both using calibration and external validation. Correlation was higher for spectra recorded by integrating sphere on wood at equilibrium moisture. Regarding the use of wood in the solid or powder condition, a slight improvement in the correlation was observed for solid wood.
The lower dispersion of the predicted values observed for planted species as compared to native species confirmed their lower variability in extractive content (
The models used in this study had close and robust statistics and yielded satisfactory predictions of the extractive content of unknown wood samples. Based on our results, the models developed from NIR signatures taken on solid wood presented more robust statistics, probably because these are based on data from several different species, while the above-mentioned studies were done using a single species.
To assess the model accuracy in predicting the extractive content from unknown wood, calibrated models based on five species were validated against a different species not included in the calibration step. The analyses were carried out in three different steps: (i) from NIR spectra recorded with integrating sphere from solid wood; (ii) from NIR spectra recorded with integrating sphere from powdered wood; and (iii) from NIR spectra recorded with fiber-optic probe in milled wood.
Considering the powdered wood spectra, the highest prediction coefficient and lowest prediction error was obtained using the model validated by the EC species (R²p = 0.66, RMSEP = 1.04%). The models validated by species J (R²p = 0.67, RMSEP = 6.21%) and P (R²p = 0.57, RMSEP = 5.46%), though presenting good prediction coefficients, had errors considered too high.
The models based on spectra obtained through fiber-optic probe did not yield satisfactory results, with low prediction coefficients and/or high prediction errors.
The graphs in
NIR spectroscopy proved to be an efficient and fast technique for estimating the total extractive content of wood. For this purpose, models from spectra acquired by the two acquisition pathways of the equipment can be used. However, the acquisition by integrating sphere showed more robust results in different cases.
The preparation of samples (solid wood or powder wood) affects the acquired spectra and therefore the final results. In this study, models from spectra drawn from solid wood had relatively better results when compared to models from spectra from wood powder. However, in both cases the results were considered satisfactory and the models could be successfully applied to estimate the extractive content of the wood.
Based on PCA results, it is concluded that the NIR is sensitive to the chemical variations of the woods investigated in this study. Although a classification of the species based on extract contents was not fully evident in several cases, a tendency towards the discrimination among species based on such data could be noticed.
The external validation of the NIR models using 25% of the samples to validate the calibration batch presented satisfactory results, as well as their validation using an extra species. In spite of the good predictions obtained in the majority of cases, errors considered high for the content of extractives were observed.
The authors express special thanks to the Wood Science and Technology Laboratory of the Federal University of Lavras (UFLA, Brazil), to CNPQ, FAPEMIG and CAPES for supporting the experimental work. This work was supported by the National Council for Scientific and Technological Development (CNPq, Brazil. grant no. 405085/2016-8), the Higher Education Personnel Improvement Coordination (CAPES, Brazil) and Foundation for Research Support of the State of Minas Gerais (FAPEMIG, Brazil). PRG Hein was supported by CNPq (grant no. 303675/2017-9) grants.
Acquisition of spectra with NIR spectrometer (b) in solid wood using integrating sphere (a and c) and powder wood using optical fiber (c).
Mean spectra untreated (a) and treated by second derivative (b), collected in solid wood through the integrating sphere and in powder wood through the two acquisition pathways of the equipment. (C):
Results of the principal components analysis of NIR spectra obtained in solid wood (a) and wood powder (b) through the integrating sphere. (C):
PCA loadings from untreated (a) and treated (b) NIR spectra.
Graph of calibrations and cross and external validations with spectral data obtained from the integrating sphere in powdered wood (a and c) and solid wood (b and d).
Graphs of calibrations and external validations from a species with spectral data obtained from the solid wood by the integrating sphere acquisition path. (C):
Descriptive statistics of the total content of extractives from different wood species determined by solvent sequence. (CV): coefficient of variation; (N): sample intensity.
Plant material | Common name | Abbr. | Mean(%) | Max(%) | Min(%) | CV(%) | N |
---|---|---|---|---|---|---|---|
Eucalypt | EC | 4.25 | 6.85 | 1.75 | 39.32 | 12 | |
Eucalypt | EV | 5.04 | 6.41 | 3.80 | 19.82 | 12 | |
Cedar | C | 7.03 | 9.33 | 5.94 | 15.35 | 12 | |
Garapa | G | 3.66 | 4.18 | 3.20 | 8.39 | 12 | |
Jacaranda | J | 9.46 | 11.77 | 6.69 | 19.66 | 12 | |
Peroba | P | 10.99 | 14.25 | 8.04 | 18.49 | 12 |
Global calibrations and cross-validations to estimate the total extractive content of wood from NIR spectra. (Treat): mathematical treatment; (“-“): original data; (1d): first derivative; (R²c): coefficient of determination for calibration; (RMSEC): mean square error for calibration (%); (R²cv): coefficient of determination for cross validation; (RMSECV): mean square error for cross validation (%); (RPD): deviation to performance ratio; (EMC): equilibrium moisture content; (0%): 0% moisture.
Model | Via | Processing | Moisture | Treat | R²c | RMSEC | R²cv | RMSECV | RPD |
---|---|---|---|---|---|---|---|---|---|
1 | Sphere | Solid | EMC | - | 0.91 | 0.90 | 0.87 | 1.08 | 2.83 |
2 | Sphere | Solid | EMC | 1d | 0.91 | 0.90 | 0.84 | 1.20 | 2.55 |
3 | Sphere | Powder | EMC | - | 0.90 | 0.95 | 0.85 | 1.19 | 2.57 |
4 | Sphere | Powder | EMC | 1d | 0.87 | 1.11 | 0.81 | 1.34 | 2.30 |
5 | Sphere | Powder | 0% | - | 0.81 | 1.32 | 0.72 | 1.63 | 1.89 |
6 | Sphere | Powder | 0% | 1d | 0.84 | 1.21 | 0.72 | 1.63 | 1.89 |
7 | Fiber | Powder | EMC | - | 0.91 | 0.90 | 0.71 | 1.65 | 1.85 |
8 | Fiber | Powder | EMC | 1d | 0.72 | 1.61 | 0.39 | 2.48 | 1.24 |
9 | Fiber | Powder | 0% | - | 0.86 | 1.14 | 0.61 | 1.96 | 1.57 |
10 | Fiber | Powder | 0% | 1d | 0.77 | 1.45 | 0.53 | 2.14 | 1.43 |
External calibrations and validations performed from 25% of the samples, using spectral data obtained from the sphere and fiber pathways. (R²cv): coefficient of determination for cross validation; (RMSECV): mean square error for cross validation (%); (R²p): coefficient of determination for prediction; (RMSEP): mean square error for prediction (%); (RPD): deviation to performance ratio; (EMC): equilibrium moisture content; (0%): 0% moisture.
Model | Via | Processing | Moisture | R²cv | RMSECV | R²p | RMSEP | RPD |
---|---|---|---|---|---|---|---|---|
11 | Sphere | Solid | EMC | 0.84 | 1.17 | 0.93 | 0.95 | 3.10 |
12 | Sphere | Powder | EMC | 0.84 | 1.16 | 0.87 | 1.40 | 2.12 |
13 | Sphere | Powder | 0% | 0.68 | 1.68 | 0.67 | 1.93 | 1.53 |
14 | Fiber | Powder | EMC | 0.74 | 1.51 | 0.64 | 2.08 | 1.42 |
15 | Fiber | Powder | 0% | 0.67 | 1.71 | 0.58 | 2.19 | 1.35 |
Calibrations and external validations made from only one species using spectral data obtained from spectra recorded with integrating sphere and fiber-optic probe. (R²cv): coefficient of determination for cross validation; (RMSECV): mean square error for cross validation (%); (R²p): coefficient of determination for prediction; (RMSEP): mean square error for prediction (%); (RPD): deviation to performance ratio. (C):
Model | Via /Processing | Validationset | R²cv | RMSECV | R²p | RMSEP | RPD |
---|---|---|---|---|---|---|---|
16 | Sphere / Solid | EC | 0.88 | 1.04 | 0.34 | 1.52 | 2.00 |
17 | EV | 0.87 | 1.16 | 0.63 | 0.80 | 4.02 | |
18 | G | 0.81 | 1.29 | 0.31 | 2.23 | 1.35 | |
19 | C | 0.86 | 1.22 | 0.66 | 3.12 | 1.07 | |
20 | J | 0.89 | 0.99 | 0.36 | 5.87 | 0.51 | |
21 | P | 0.84 | 0.98 | 0.74 | 4.81 | 0.52 | |
22 | Sphere / Powder | EC | 0.83 | 1.25 | 0.66 | 1.04 | 2.93 |
23 | EV | 0.84 | 1.29 | 0.44 | 1.20 | 2.69 | |
24 | G | 0.80 | 1.34 | 0.17 | 7.24 | 0.41 | |
25 | C | 0.86 | 1.24 | 0.001 | 1.78 | 1.87 | |
26 | J | 0.85 | 1.15 | 0.67 | 6.21 | 0.48 | |
27 | P | 0.81 | 1.08 | 0.57 | 5.46 | 0.45 | |
28 | Fiber / Powder | EC | 0.68 | 1.73 | 0.01 | 1.82 | 1.67 |
29 | EV | 0.73 | 1.67 | 0.002 | 1.57 | 2.06 | |
30 | G | 0.72 | 1.58 | 0.002 | 3.00 | 1.00 | |
31 | C | 0.71 | 1.79 | 0.05 | 1.88 | 1.77 | |
32 | J | 0.70 | 1.63 | 0.44 | 5.65 | 0.53 | |
33 | P | 0.62 | 1.56 | 0.39 | 6.63 | 0.37 |