DiscussionTop

Mean spectra (Fig. 1) for both evaluated materials were very similar, probably due to the anatomical and chemical characteristics of the species. There were comparable wood anatomical characteristics in function of kind and distribution of axial parenchyma cells, as was noted by the presence of lozenge-aliform and confluent cells (Inga vera) and aliform confluent cells (Machaerium paraguariense) (Vieira et al., 2019a). Also, all species have similar tangential diameter of vessels.

Similarity of spectral patterns of native species from the same botanical family is frequent. Soares et al. (2017) also reported close similarity of spectra from six Amazon Forest species, while Pastore et al. (2011) reported difficulties of NIR spectroscopy to discriminate native species from Brazil.

Visually, for charcoal, some proximity was observed for mean spectra of Machaerium paraguariense and Muellera campestris. On the other hand, for wood, mean spectra of Machaerium paraguariense had different behavior, principally at wavenumbers less than 8500 cm^-1, which according to Schwanninger et al. (2011), indicate lignin and hemicellulose.

When comparing second derivative spectra (Figure 2), some differences for wood and charcoal were observed in the region from 4400 to 4000 cm^-1, with the presence of bands related to cellulose, hemicellulose and lignin (Schwanninger et al., 2011). These results are in accordance with wood thermal degradation, which can vary in function of carbonization time and temperature, but is principally influenced by species characteristics (Nisgoski et al., 2019).

Pretreatment of data is frequent in studies to verify the potential of NIR spectroscopy for forest species discrimination. Some authors have applied first derivative transformation (Bergo et al., 2016; Soares et al., 2017) and others second derivative (Toscano et al., 2017; Nisgoski et al., 2018; Vieira et al., 2019b), while Ramalho et al. (2018) applied the standard normal variate (SNV) model. In that work, three transformations were tested, but the authors only discussed the best results, obtained from second derivative transformation. Also, in charcoal Sandak et al. (2016) reported that the chemical information after second derivative application was still preserved and the pretreatment effectively removed the scatter and some noise.

Figure 3 shows the modifications in chemical composition after carbonization, changing the NIR spectra, as expected, and consequently the PCA distribution. Carbonization causes degradation of biopolymers, resulting in changes in NIR spectra because these reflect the composition of charcoal, and even when the carbonization process is the same among species, alterations in chemical composition depend on the original material, i.e., the species’ characteristics (Davrieux et al., 2010; Muñiz et al., 2013; Nisgoski et al., 2015).

The second derivative data (Figure 3 C, D) indicated diverse mean behavior of the species, confirmed by the ellipses. Better results for distinction of wood species based on PCA with second derivative spectra also were described by Hwang et al. (2016) in identifying Pinus species from Korean historic architecture.

In the classification models (Table 2), second derivative NIR spectra showed better results in all tested procedures. Horikawa et al. (2015) also described better results for identification of anatomically similar Diploxylon species in Japan based on second derivative NIR spectra and PLS-DA models.

The high values of accuracy and precision of the classification models confirmed the potential of using NIR spectroscopy to differentiate wood and charcoal samples of the species Inga vera, Machaerium paraguariense and Muellera campestris. However, this requires standardization of factors, such as sampling sites and moisture content of the material. Hein et al. (2017) commented that more studies are necessary to evaluate models based on NIR spectroscopy in real situations, such as evaluation of samples with diverse granulometry and surface quality, among other factors.

Other studies have also described adequate results for species distinction with these classification models. With the artificial neural network method (ANN), Esteban et al. (2009), using biometry of anatomical traits, reported 92% probability in differentiation of Juniperus cedrus and Juniperus phoenicea var. canariensis. Also, applying quantitative anatomical characteristics Turhan & Serdar (2013) evaluated SVM to differentiate three Salix species from Turkey and found successful of classification of 95.2% in test group validation. Applying NIR spectroscopy, Zhou et al. (2020) described that SVM produced better results than LDA in discrimination of a green mix of timber from Tsuga heterophylla and Abies amabilis, and Xu et al. (2019) described a correct rate of 92.8% in prediction and identification of origin of Angelica dahurica.

NIR spectroscopy in association with classification models can be applied to a wide range of material. For example, Balabin et al. (2011), using different discriminant techniques, such as regularized discriminant analysis (RDA), soft independent modeling of class analogy (SIMCA), partial least squares regression and classification (PLS), K-nearest neighbor (KNN), multilayer perceptron (MLP), support vector machine (SVM) and probabilistic neural network (PNN), applied NIR spectra to classify motor oils. They found that among the tests evaluated, simple techniques such as k-NN were considered inadequate for classification.

In general, the confusion matrix for wood (Fig. 4) confirmed the behavior of previous analysis. In raw data, discriminant models based on k-NN and SVM indicated the similarity between Inga vera and Muellera campestris, which was also verified in visual analysis of spectra (Fig. 1) and PCA (Fig. 3). For raw data, the ANN model had higher efficacy in distinction of I. vera. Considering second derivative data, models based on k-NN and SVM distinguished all species, and ANN presented confusion of one sample between Muellera campestris and Machaerium paraguariense.

In the confusion matrix based on charcoal samples (Fig. 5), the same tendency of the other analysis was verif ied. Just as for wood, raw data of charcoal based on k-NN and SVM confirmed the similarity between M. campestris and M. paraguariense, illustrated in Figs. 1 and 4. With second derivative data, SVM tended to classify samples of all species as I. vera.

Based on our results, and mainly considering the speed and the possibility of identifying wood based on NIR spectra in association of chemometrics, NIR spectroscopy can be applied for distinction of wood and charcoal of the evaluated Fabaceae species.

Figure 4. Confusion matrix results from test block data based on classification models (k-NN, SVM and ANN) with wood data for studied Fabaceae species.

Figure 5. Confusion matrix results from test block data based on classification models (k-NN, SVM and ANN) with charcoal data from studied Fabaceae species.