Wood specimens from natural forest (cerrado biome) and commercial plantations typically found in Brazil were used in this study. The wood specimens from natural forests belongs to the following genera: Cedrela sp. (Cedar, C), Apuleia sp. (Garapa, T), Aspidosperma sp. (Peroba, P) and Jacaranda sp. (Rosewood, J). The natural forest is located at Lavras municipality, Minas Gerais state, southeastern Brazil (21°10'S, 44°54'W).

On the other hand, Eucalyptus wood represented the planted forests, the main genera cultivated for industrial application in Brazil. In order to represent the large variability in terms of genetic material used in reforestation in Brazil, Eucalyptus clones commercially used by two forestry companies were selected: i) Vallourec (V&M do Brasil) that utilizes hybrids of Eucalyptus grandis × E. urophylla at 6.5 years old located at Paraopeba city (19°16'S, 44°24'W) with focus on charcoal for bioenergy purposes and ii) Cenibra Nipo Brasileira S.A. that utilizes hybrids of Eucalyptus grandis × E. urophylla at 6 years old located at Belo Oriente (19°17'S, 42°23'W) state of Minas Gerais, Brazil, for cellulosic pulp production.

The total wood sampling was composed by six vegetal materials: 4 genera from native forest and 2 different types of Eucalyptus . The wood specimens used in this study came from the same experiment described in Ramalho et al. (2017) who investigated the charcoal produced from these wood specimens.

Ninety wood specimens with nominal dimensions of 35 mm × 35 mm × 100 mm (R × T × L) were removed from the first log of felled trees of each vegetable material, except for Cedrela, Aspidosperma and Apuleia genera that was represented by 45, 88 and 87 specimens, respectively. The total sampling was composed by 490 clear specimens [(90 · 3) + 45 + 88 + 87)] free of defects and well defined tangential-longitudinal and radial-longitudinal surfaces.

Specimens were conditioned in a room set for 20°C and 65% relative humidity to maintain the equilibrium moisture content (EMC) of ∼14% prior to NIR spectra recordings.

The near-infrared spectra were acquired using the Bruker spectrometer (model MPA, Bruker Optik GmbH, Ettlingen, Germany) based on a Fourrier transform and equipped with an integrating sphere. Spectral acquisition was performed within the 12,500 - 3,600 cm ^-1 with a spectral resolution of 8 cm ^-1 in diffuse reflection mode. Each spectrum represented the average of 32 scans. NIR spectra were recorded in the center of the two sides of the longitudinal-radial surfaces of each wood samples and averaged. Previous papers have indicated that the radial surface is more suitable for recording NIR spectra because is more representative ( e.g. Hein et al. , 2009 ; Costa et al. , 2018 ). First derivatives (13-point filter and a second order polynomial) using the Savitzky & Golay (1964) algorithm and Standard Normal Variate (SNV) transformations were applied on the NIR spectra to enhance the quality of the information and analyses.

PCA were calculated for compressing the main information in a set of variables into a lower number of new variables by means of the Chemoface ( Nunes et al. , 2012 , v. 1.61, http://ufla.br/chemoface/ ) software. The PCA were calculated using full model size and maximum of eight principal components. For cross-validations 8 segments of 60 samples per segment were adopted while for test set validation were done using 2/3 of samples in calibration set and the remaining 1/3 in test set. Martens' uncertainty test ( Westad & Martens, 2000 ) was used to select the wavenumbers with significant x-loadings. Outliers were identified from the analyses of residuals x-variances and leverage value plot. The number of latent variables (LVs) adopted for each model corresponded to the first minimal residual x-variance.

Partial least squares discriminant analyses (PLS-DA) were carried out for developing models for classifying wood species. The PLS-DA classification method, which is based on the PLS-R (partial least squares regression) approach, correlated two blocks of variables, X (independent variables) and Y (dependent variables). The independent variables (X) were the NIR spectra matrix while the dependent variables (Y) were the classes of wood species. For evaluating PLS-R regressions, the root mean square error (RMSE), ratio of performance to deviation (RPD) and range error ratio (RER) were used. As described in Costa et al. (2018) , the RMSE measures the efficiency of the calibration model to predict the property of interest in many unknown sample, while the RPD is a way to identify the accuracy of the calibration and is the ratio between the standard deviation of the reference values and RMSE. The RER is the ratio between range of reference data and standard error of prediction. According the AACC (1999) any model that has a RER ≥ 4 is qualified for screening calibration. When RER ≥ 10 the model is acceptable for quality control, and if RER ≥ 15 the model is very good for research quantification.

The samples were grouped in six different cla­s­s­es: ­EC ( Eucalyptus clones commercially used by Cenibra Nipo Brasileira S.A.), EV ( Eucalyptus clones commercially used by Vallourec do Brasil), C ( Cedrela sp. from natural forest), J ( Jacaranda sp. from natural forest), P ( Aspidosperma sp. from natural forest) and T ( Apuleia sp. from natural forest). The classification of the wood specimens into classes of vegetal material was carried out through PLS-DA and the percentages of correct predictions were analyzed. The PLS-DA models were validated using cross- and test set validation methods. The cross-validation was carried out using eight segments while for the independent test set validation 2/3 of the samples were used in calibration set and 1/3 in validation set.

Fig. 1 displays NIR spectra signatures taken from 9,000 to 4,000 cm ^-1 on the radial surface of wood specimens. According to Brereton (2003) , this absorption bands correspond to the specific functional groups present in wood samples. Most band assignments for wood has been presented by Schwanninger et al. (2011) , in which they described band locations in wavenumber, the component likely to absorb at this band location and also the bond vibration.

Here, after treatment of the spectra using standard normal variate and first derivative (SNV + 1d), the existence of the absorption bands became more evident at certain wavenumbers (Fig. 1B), although the peaks are inflection on raw spectra. From a visual analysis, small differences between NIR spectra of native woods can be observed in terms of intensity of the absorption bands at certain wavenumbers (indicated by narrows). For instance, line J (black fine line) decreases less intensely than lines T, P and C (native woods) in the zones indicates by the narrows. However, no difference can be detected between NIR spectra measured from Eucalyptus wood (dotted lines).

An exploratory analysis using NIR spectroscopy is useful and important for wood science and industry as well as quantitative analysis ( Tsuchikawa, 2007 ). PCA was performed with the treated and untreated spectra. The mathematical treatment did not improve the results in terms of increasing the explained variance (Table 1) but from a visual analysis of the score plots, treatment on NIR seems to clarify the separation of different types of wood.

It was possible to identify two wood groups with the PCA of the untreated NIR spectra (Fig. 2A). These clusters were formed because each vegetal material presents different characteristics and properties, such as the chemical composition and anatomical features ( Kollmann & Coté, 1968 ). Thus, the two types of Eucalyptus wood (EC and EV) clearly could be distinguished from most of wood specimens from natural forests (J, P and T). However, the samples of Cedrela (C, gray squares) are mixed with Eucalyptus specimens in this exploratory analysis done from untreated NIR spectra.

PCA of the treated spectra promoted the separation of the species into groups according to the genetic material (Fig. 2B). Cedrela samples were correctly grouped as native species. NIR spectroscopy is very sensitive to variations in chemical properties and it is difficult to discriminate materials with chemical or anatomical similarity. Wood specimens from natural forests clearly differ from commercial Eucalyptus samples for any reasons, such as density, color, grain. In this study, all specimens present same moisture and were processed using the same procedure, so the differences captured by NIR sensor is due to intrinsic properties of each material. In short, NIR signature is able to perceive these small differences and mathematical treatments can make this information clearer. As pointed out by Yang et al. (2015) the non-destructive and fast separation of wood species can improve the speed of production and enable the processes to be more efficient.

In both Figs. 2A and 2B, the two types of Eucalyptus (EC, managed for pulp and paper industry and EV, managed for bioenergy) were grouped together and no difference between them could was highlighted.

In order to generate PLS-DA models capable of classifying the wood specimens by vegetal material based on NIR spectra, preliminary PLS-R models were calibrated and tested by cross-validation. The models from natural forest presented better values of statistical parameters than the models of planted forests (Table 2). The RER of PLS-R models for native woods were above 6 while the models for Eucalyptus were approximately 3.5 (Table 2). The PLS-R model for rosewood (J) wood presented better statistics (higher RPD and RER and lower RMSEP) and consequently the wood specimens were classified better by PLS-DA.

The classification of wood specimens from PLS-DA models both for the cross-validation and the test set yielded findings similar to the results observed by the PCAs (Table 3). The PLS-DA models presented a large percentage of correct classifications for wood specimens, except for the Eucalyptus that were confused between them. However, the misclassifications between Eucalyptus type specimens do not devalue the method since the main goal of the research was to develop models able to distinguish between the two large groups: natural (native species) and planted forests ( Eucalyptus ). Moreover, only one specimen from natural forest (Ceder) was confounded with Eucalyptus , indicating that this native wood has characteristics similar, in some extent, to those of planted forest.

Among the specimens of natural forests there were incorrect classifications between T and P specimens (Table 3). Seven from 87 specimens belonging to Apuleia sp. group, T) was confounded and classified as Aspidosperma sp. wood specimens by the cross-­valida­ted models. In the model validated by test set, 14% of the T samples were misclassified as P wood samples. Besides obtaining high index of correct classification, the native species J ( Jacaranda sp.), P ( Aspidosperma sp.) and T ( Apuleia sp.) were not confused with Eucalyptus . To sum up, the classification of native wood samples from PLS-DA carried out using untreated NIR spectra present high percentage of success (86% to 100%) indicating good efficiency of the classification models.

The PLS-DA models of Table 4 shown to be robust for classifying specimens from natural and planted forests; In this analysis, no specimen of planted forest was classified as a wood specimen from natural forest and vice versa. These results indicate that mathematical treatment can improve the signal quality of NIR spectra, allowing good classification by PLS-DA.

Again, these PLS-DA models (Table 4) present faults for classifying the varieties of the Eucalyptus genus. Compared to classifications of Table 3, the EV became even more mixed with EC resulting in the higher rate of misclassification (74%) while only 7 of 90 specimens of EC was confounded with EV. It is likely that these misclassifications occur because these Eucalyptus samples are very similar from a chemical point of view although no chemical characterization was performed in this study.

In short, the aim of distinguishing vegetal material, especially from planted or planted forests was successfully achieved. The similar findings obtained from PCAs and PLS-DAs show that it is possible to use NIR spectra for distinguishing different wood species, especially for separating unknown wood samples from natural to planted wood, but with some limitations.

The specimens of Cedrela sp. wood had a greater number of samples mistakenly grouped as Eucalyptus wood in the PCA carried out from untreated NIR spectra. This deportment shows that the wood of some native species can be mistaken with Eucalyptus woods. The chemical similarity between wood from natural and planted forests can be an barrier in implementing this approach in real condition, since the chemical composition of some woods can be suchlike. This work is an approach to a preliminary study to develop a system for rapid identification of the type of wood that has been transported and sold illegally in many countries.

According to Yang et al. (2015) the identification of wood species is of great significance for industrial utilization as well as the quality of the final product. Moreover, this technology can be used for monitoring illegal logging. The inspection agent can record a spectrum of the sample of unknown species and compare with the database. Thus, depending on the group in which the spectrum of wood sample frame, it will be possible to classify it as originating from native or planted forest.

More studies applying this approach may include a wide range of planted and native forest species in models. Moreover, models can be developed with wood samples in different moisture content and with different surface qualities in order to generate a comprehensive database that can be efficient when used to identify the source of wood.

This study demonstrates that NIR spectroscopy has sensibility for clearly separating Eucalyptus from native wood and can be used for quickly and reliably identify the origin of unknown wood samples.

PCA based on untreated NIR spectra is able to clearly distinguish Eucalyptus wood samples from Jacaranda (J), Aspidosperma (P) and Apuleia (T) genus; however, most of specimens of Cedrela (C) wood were mixed with Eucalyptus. PCA of treated NIR spectra by normalization (SNV) and first derivative clearly showed two groups formed by wood samples from Eucalyptus and natural forests.

PLS-DA models carried out from untreated NIR spectra presented a large percentage of correct classifications (86 to 100%) for wood specimens of natural forest, except for the Eucalyptus samples that were confused between the two varieties. PLS-DA models based on treated NIR signatures are robust for classifying wood specimens from natural and planted forests and no specimen of planted forest was classified as specimen of natural forest and vice versa.