3.1 Scanning electron microscopy of wood
Figure 2 presents micrographs that highlight surface irregularities observed on a microscopic scale, including ripples, radii, and defects in the cross-sections of Diplotropis racemosa (Hoehne) Amshoff samples.
Figure 2A shows the surface under field conditions, characterized by biodegradation and oxidation processes, with minimal visible porosity, likely due to dust or fungal colonization from environmental exposure. Figure 2B illustrates the surface processed with a chainsaw, featuring deeper cuts, diagonal scratches, and more visible pores. Figure 2C shows the surface processed with a circular saw, with shallower diagonal scratches and visible pores. Figure 2D presents the surface cut with a band saw, notable for its narrower grooves compared to the circular saw, an 'alfombra' pattern, a higher proportion of pores, and less clogging. Finally, Fig. 2E depicts the sanded surface, which is more homogeneous; however, sanding can clog and mask pores, compress or deform the cell wall, and obstruct the cell lumen with dust (Santoni and Pizzo 2011; Sulaiman et al. 2009). All these surface conditions can interfere with the spectral signal during the development of NIR models by either affecting the absorbed radiation or causing information loss in the reflection.
According to Schimleck et al. (2023), the roughness of the wood surface is the result of the properties of the material itself and several variables associated with the tools used to prepare it. And increasing surface roughness has been shown to reduce the absorption of NIR radiation in wood samples, which can influence the performance of NIR calibration models (Schimleck et al. 2005; Zhang et al. 2015). This phenomenon causes light to scatter at different angles depending on the wavelength, affecting the resulting spectra (Sandak et al. 2016).
3.2 NIR spectra and absorption band assignments
The averaged spectral signatures with the original data from the 16 tropical species and the wood machining treatments are presented in Fig. 3A and Fig. 3B, respectively. In Fig. 3A, each spectrum represents the combination of several spectra collected from surface samples under field conditions, classified according to their bulk density. On the color scale, the darker lines in the upper region of the signatures represent the spectra of high-density species, and the lighter colored lines in the lower region of the spectral profile correspond to low-density species (Fig. 3A). In absorbance values, overlap was observed in most species, with minimal variations in intensity, attributable to their densities, which are very close to each other. However, species such as H. oblongifolia, Dalbergia sp., C. odorata and Cedrela sp., present absorption peaks with more evident differences. The first two species may be related to a possible higher extractive content (dark-colored wood), while the last two species may be related to their lower density compared to the other species.
The similarity between the different species is attributed to the predominance of cellulose in the wood. This chemical uniformity makes precise distinction difficult by visual inspection, making it essential to employ advanced techniques, such as chemometrics, for accurate differentiation. (Russ et al. 2009). Second Santos et al. (2021a) and Lima et al. (2022a), the main spectral regions for identifying tropical species are: 6500–7000 cm− 1 (1), 5400–5900 cm− 1 (2) and 4600–5200 cm− 1 (3). Lima et al. (2022a) reported that these bands have the potential to distinguish species from various genera, such as Dinizia, Licania, Manilkara, Pseudopiptadenia, Pouteria, Caryocar, Eschweilera, Brosimum, Simaba, Parkia and Tapirira. Santos et al. (2021a) also found that regions 2 and 3 are effective in differentiating species from genera Ocotea, Nectandra, Roupala, Euplassa, Mezilaurus and Aniba. Vieira et al. (2021) emphasized the importance of the second region in distinguishing four native Brazilian species of the genus Myrtaceae.
The mentioned absorption bands correspond to specific functional groups present in wood (Brereton, 2003). In particular, bands in the range 6500 to 7000 cm− 1 are related to the amorphous and crystalline regions of cellulose. The range from 5400 to 5900 cm− 1 it is associated with all components of the cell wall, mainly cellulose, lignin and hemicellulose. In addition to these structural components of the cell wall, the ranges between 4600 and 5200 cm− 1 are associated with extractive content (Schwanninger et al. 2011). It is important to note that many of the identified features are associated with well-known bond vibration bands. However, we cannot predict that these models are based directly on wood chemistry, as components such as cellulose, lignin and others contribute to these characteristics only to a certain extent. (Schimleck et al. 2023).
To facilitate the understanding of the NIR spectra by surface treatment, two species with different densities were selected (Fig. 3B). In benchtop equipment, for species with low density (Cedrela sp.), the minimum absorbance was 0.2 on the surface processed with a band saw and the maximum absorbance was 0.9 in field conditions and processed with a chain saw. For the high-density species (Dypteryx sp.), a minimum absorbance of 0.3 was recorded under field conditions and a maximum of 1.3 on the surface processed with a circular saw. In portable equipment, for the type of Cedrela sp., the minimum absorbance was 0 under field conditions and on the surface processed with a band saw, while the maximum absorbance was 0.3 on the surface processed with a chain saw. The species Dypteryx sp. it showed a minimum absorption of 0.1 in field conditions and a maximum of 0.5 on the surface processed with a circular saw and finished with sandpaper (Fig. 3B).
Analyzing the intensity of NIR absorption bands in Cedrela sp., the highest absorbance was on the surface produced by chainsaws in both NIR equipment. Furthermore, in the bench equipment, greater absorbance was also observed on the surface produced by the circular saw. For the species Dipteryx sp., higher absorbance was observed on surfaces produced by circular sawdust in both equipment. However, the sandpaper-finished surface also showed significant absorbance peaks in portable NIR. Apparently, finer textures with lower roughness and higher gloss indexes favor greater absorbance, indicating a high influence of the surface on spectral acquisition. Novaes et al. (2023) found higher absorptions on surfaces produced with circular saws, suggesting that these surfaces have better potential to discriminate species. Although visual analysis showed differences, it is not sufficient to identify species.
3.3 Principal component analysis
The contributions of the main components by type of wood surface processing obtained a higher explanatory percentage of the data variation with the spectra pre-treated by normalization on the bench equipment and with the original spectra in the portable NIR (Table 2).
Table 2
Explanation of the variance of the main components, applied in other mathematical treatments, for the five surface qualities.
| Equipment | Pretreatment | PCA- Wood processing Half hits (%) |
| Raw | Chainsaw | Circular saw | Band saw | Sanding |
| PC1 | PC2 | PC1 | PC2 | PC1 | PC2 | PC1 | PC2 | PC1 | PC2 |
| Benchtop | Untreated | 99.24 | 0.68 | 99.32 | 0.61 | 99.02 | 0.92 | 99.3 | 0.65 | 99.1 | 0.84 |
| Normalize | 99.27 | 0.66 | 99.37 | 0.57 | 99.12 | 0.83 | 99.37 | 0.58 | 99.23 | 0.71 |
| 1d | 98.33 | 0.45 | 98.2 | 0.38 | 98.15 | 0.59 | 97.88 | 0.58 | 97.92 | 0.54 |
| 2d | 80.06 | 2.38 | 72.8 | 3.79 | 78.95 | 2.68 | 75.84 | 3.19 | 73.96 | 3.35 |
| MSC | 78.02 | 2.76 | 70.98 | 3.96 | 77.48 | 2.79 | 74.23 | 3.46 | 72.41 | 3.61 |
| SNV | 79.21 | 2.44 | 72.88 | 3.6 | 78.77 | 2.62 | 75.56 | 3.12 | 73.81 | 3.21 |
| Portable | Untreated | 98.43 | 1.37 | 99.3 | 0.61 | 99.08 | 0.82 | 98.97 | 0.91 | 98.79 | 1.08 |
| Normalize | 97.83 | 1.90 | 99.15 | 0.74 | 98.86 | 1.02 | 98.73 | 1.13 | 98.5 | 1.35 |
| 1d | 97.2 | 2.31 | 98.28 | 1.36 | 98.35 | 1.36 | 98.07 | 1.63 | 97.69 | 1.97 |
| 2d | 95.64 | 3.36 | 97.11 | 2.12 | 97.37 | 1.97 | 96.26 | 2.96 | 96.79 | 2.58 |
| MSC | 94.22 | 4.81 | 96.45 | 2.82 | 96.57 | 2.77 | 95.85 | 3.39 | 95.23 | 4.13 |
| SNV | 95.19 | 3.83 | 96.67 | 2.6 | 96.89 | 2.46 | 96.34 | 2.9 | 96.08 | 3.27 |
| Where: 1d = first derivative, 2d = second derivative, MSC = Multiplicative scatter correction and SNV= Standard normal variate. | |
For the original data with surface under field conditions, recorded with the bench equipment, PC1 explained 99.24% and PC2 0.68%, while PC1 from the portable equipment explained 98.43% and PC2 1.37%. On the surface processed with a chainsaw, the sum of the main components in the bench equipment was 99.93%, while in the portable equipment it was 99.91%. When evaluating the total percentage of PCA in wood prepared with a circular saw, band saw and sandpaper, slightly higher values were found in the bench equipment, when compared to the results of the portable NIR: 99.94%, 99.95% and 99.94%, respectively.
Figure 4 illustrates the score plots from both benchtop and portable equipment, showing the grouping tendencies of tropical wood species. In the natural field condition, distinct groups were observed for Cedrela sp., C. odorata, C. ferrea, and H. oblongifolia (Fig. 4A). Using portable equipment under the same surface condition, the primary groups identified were Dalbergia sp., C. ferrea, and H. oblongifolia, with other species showing overlap, indicating spectral similarity (Fig. 4A and 4B). However, visual discrimination of the groups improved as the wood underwent machining processes, particularly with the circular saw and sander, in both benchtop and portable equipment.
The overlapping behavior of the species in field conditions can be explained due to changes in the appearance of the samples, as they were stored in yards for a long period, which implies several aging processes of the material. According to Reis et al. (2023), the most significant aging process is the oxidation of the wood surface, as it influences the color. Santos et al. (2022) classified different species of Tauari without surface refreshment and obtained adequate identification percentages, corroborating the present study.
When machining with a chainsaw, visually the scores revealed a notable increase in overlapping areas for fourteen species, making their identification difficult, except for Cedrela sp and C. odorata (Fig. 4C and 4D). Novaes et al. (2023) performed PCA on spectra of wood processed with a chainsaw and treated by second derivative, which concluded that M. raised and D. excellent presented partial similarities in their spectra, while G. glabra, Himenea sp., M. melinoniana and Copaifera sp. presented overlaps, which prevented their correct discrimination. In this study, chainsaw samples did not achieve satisfactory species identification, possibly due to the coarser texture and greater surface roughness of the wood.
On the surface produced with a circular saw, a reduction in overlapping areas between species was observed, in relation to previous graphs. The bench and portable equipment showed a slight separation of scores into distinct groups in Cedrela sp., Dalbergia sp., C. odorata., V. araroba., Machaerium sp., D. odorata., D. racemosa., H. oblongifolia and Dipteryx sp. (Fig. 4E and 4F). Pace et al. (2019) identified the formation of four different groups in a set of twelve wood species cut with a circular saw, where the score plots explained 99% of the variance in PC1 of the original spectra. By contrast, Novaes et al. (2023) discriminated spectra collected on surfaces produced with circular saws for species such as M. elata, D. excelsa, M. melinoniana, G. glabra and Himenea sp., showing clustering trends with minimal overlap between spectra, which allowed better visual discrimination compared to the chainsaw results. Thus, circular saw-cut surfaces facilitated species identification compared to other surface conditions. The benchtop equipment produced denser clusters without noticeable areas of overlap, while the portable equipment showed less defined clusters.
On the surface produced with a band saw, the benchtop equipment showed that PC1 explained 99.30% of the data variance and PC2 0.65%, while on the portable equipment PC1 explained 98.79% and PC2 0.91%. Visually, the bench equipment scores showed separation only for Cedrela sp. and Dalbergia sp. and, the portable equipment clearly discriminated the species Cedrela sp. and H. oblongifolia (Fig. 4G and 4H). Wood with a sanded surface of 80 grain presented PC1 of 99.10% and PC2 of 0.84% on the bench equipment. In portable equipment, PC1 explained 98.79% and PC2 1.08%. The scores on the bench NIR separated the groups Cedrela sp., Dalbergia sp., C. odorata, C. brasiliense, R. montana, D. iron and H. oblongifolia. On the laptop, separation was observed for Cedrela sp., Dalbergia sp., D. iron, H. oblongifolia and Dipteryx sp. (Fig. 4I and 4J).
Reis et al. (2023) achieved 98% explanation of the variation in PC1 for wood with sanded surfaces and with original spectra collected on benchtop equipment. According to Costa et al. (2018), the greater dispersion of scores in the cross-sectional area of wood is due to the extensive volume of information from the cell wall. NIR radiation penetrates deeper into this surface, where the fibers function as light tubes, allowing a greater volume of wood to be evaluated. Although surface sanding slightly improved the identification of a greater number of species, it was not sufficient to identify all of them.
3.4 Partial Least Squares Discriminant Analysis
The PLS-DA models showed a high percentage of correct classification for tropical species (Table 3). The model implemented to analyze surfaces under field conditions, using original spectra, achieved accuracy levels of 96.5% and 93.4% for bench and portable instruments, respectively. After data processing, normalization maintained a success level of 96.5% for the bench instrument. In portable equipment, a slight improvement was observed (95.3%). To date, there are no reports in the literature of studies using NIR spectroscopy for this surface condition in the field. Although there is no research on the analysis of spectra on non-homogenized or untreated surfaces, it is known that such conditions affect the reliability of NIR spectroscopy and the performance of prediction models (Schwanninger et al. 2004; Sandak et al. 2016). Thus, this study indicates that NIR spectra obtained without prior treatment on the surface of materials, analyzed with multivariate statistics, can discriminate wood species with high efficiency.
Table 3
PLS-DA summary with percentage of cross-validation hits, considering acquisition equipment and mathematical pre-treatment used.
| Equipment | Pretreatment (10LV) | NIR - Species classification |
| Average hits (%) |
| Raw | Chainsaw | Circular saw | Band saw | Sanding |
| Benchtop | Untreated | 96.5 | 85.9 | 96.5 | 91 | 91.8 |
| Normalize | 96.5 | 88.7 | 96.9 | 96.9 | 92.2 |
| 1d | 88.7 | 84 | 87.1 | 93.4 | 84 |
| 2d | 60.6 | 47.3 | 69.5 | 71.1 | 71.5 |
| MSC | 46.9 | 37.1 | 56.3 | 61.3 | 64.8 |
| SNV | 53.1 | 46.1 | 65.2 | 71.9 | 67.2 |
| Portable | Untreated | 93.4 | 84 | 92.6 | 94.9 | 92.6 |
| Normalize | 95.3 | 78.5 | 95.7 | 95.3 | 97.3 |
| 1d | 93.8 | 87.5 | 90.2 | 93 | 98.4 |
| 2d | 94.9 | 88.7 | 93.8 | 96.1 | 98.4 |
| MSC | 92.2 | 89.8 | 99.2 | 97.3 | 99.2 |
| SNV | 94.5 | 93.8 | 96.5 | 95.3 | 99.2 |
Where: LV = latent variable, 1d = first derivative, 2d = second derivative, MSC = Multiplicative scatter correction and SNV = Standard normal variate.
In models on surfaces generated by chainsaws, the raw data reflected lower percentages of success, reaching 85.9% and 84%, in bench and portable instruments, respectively. By applying mathematical normalization pretreatment, improvement was observed in the benchtop device, while in the portable device the standard normal variation achieved better optimization. Novaes et al. (2023), obtained a better result when discriminating four Amazonian forest species using PLS-DA models based on spectral signatures acquired on wooden surfaces produced with a chainsaw, achieving 90.09% success in classifying original data and 97.6% in original data. treated by second derivative.
The best classification performances were obtained on surfaces produced by circular saws, reaching 96% with the original data and 98.4% with data pre-treated with a second derivative. Possibly, this difference could be related to the increase in surface roughness due to machining with the cutting tool (Sandak et al. 2009), or variations in the density of the material, which can affect the dispersion of light recorded in the spectra, this being important aspect for wood classification (Ayanleye et al. 2021).
The surfaces produced with a circular saw showed a success rate of 96.5% for the bench instrument and 92.6% for the portable instrument, using original spectra. With the application of mathematical treatments, on the bench and portable instrument, a slight increase was observed with normalization. Novaes et al. (2023), analyzed cube-shaped samples of six Amazonian species, achieving 98.3% correct classification for original data and 99.2% for data pre-treated with second derivatives. Likewise, Pace et al. (2019) categorized samples of 12 species native to the Atlantic Forest in Brazil, obtaining a correct classification of 93.2% with PLS-DA models adjusted by cross-validation for original spectral data. In both cases, the spectra were acquired on bench equipment. In this study, comparable results were obtained, with better successes using the portable instrument, of which there are no previous reports of its use on surfaces processed with a circular saw.
Surface models produced with a band saw showed lower percentages of success with the original data, achieving 91% success with benchtop equipment and 94.9% with portable equipment. After data processing, a significant improvement was observed, reaching 96.9% through normalization for bench equipment and 97.3% with multiplicative correction for portable equipment. Therefore, the results show that after the mathematical treatment of the data, the difference in the percentages of successes between the finishes with circular saw and band saw was minimal, thus favoring the application of the NIR technique in the field for these types of surface finishes.
On sanded surfaces, the original and treated spectra showed higher results with the portable equipment, mainly with the application of multiplicative signal correlation pre-treatment, totaling 99.2% success. The discrimination of tropical Amazonian wood on sanded surfaces has been addressed in several studies. Santos et al. (2021a) achieved 97% accuracy using 80-grit sandpaper to distinguish eight wood species. Lima et al. (2022a, b) used 50, 80, and 100 grit sandpaper and found that spectra treated with the second derivative of the cross-sectional surface achieved a classification rate of 96.5% from residues of twelve wood species. Both studies used benchtop equipment to acquire the spectra. The use of portable equipment was also reported. Soares et al. (2017) classified six wood species and achieved efficiency greater than 90%. Silva et al. (2018) determined the origin of Mahogany wood, reporting success between 90% and 100%, while Snel et al. (2018) identified seven species of the genus Dalbergia with a rating above 90%. Our results on both NIR devices were satisfactory and in line with the percentages reported in the aforementioned studies.
The PLS-DA confusion matrix, obtained from the bench instrument with the data treated by normalization to classify tropical species, shows that out of 256 samples analyzed, only 8 samples were incorrectly classified, corresponding to the species A. lecointei (7) and D. odorata. (1) (Table 4). On the other hand, the model provided classification with 100% accuracy for 14 species: Cedrela sp., Dalbergia sp., C. odorata, Couratari sp., C. brasiliense, V. araroba, R. montana, Machaerium sp., G. glabra, A. leiocarpa, D. racemosa, C. ferrea, H. oblongifolia and Dipteryx sp.
Table 4
PLS-DA confusion matrix obtained by cross-validation, for classification of tropical species with spectra treated by normalization and collected on bench equipment on wood processed by circular saws.
| Species | Classification of wood predicted by Benchtop NIRS | Correct classification (10LVs) |
| CV | CR | CM | JA | MU | GU | TA | CU | AA | JP | SU | GA | CA | PF | JB | CB | | n | (%) |
| CV | 16 | | | | | | | | | | | | | | | | | 16 | 100 |
| CR | | 16 | | | | | | | | | | | | | | | | 16 | 100 |
| CM | | | 16 | | | | | | | | | | | | | | | 16 | 100 |
| JA | | | | 16 | | | | | | | | | | | | | | 16 | 100 |
| MU | | | 6 | | 9 | | | | | | 1 | | | | | | | 9 | 56 |
| GU | | | | | | 16 | | | | | | | | | | | | 16 | 100 |
| TA | | | | | | | 16 | | | | | | | | | | | 16 | 100 |
| CU | | | | | | | | 16 | | | | | | | | | | 16 | 100 |
| AA | | | | | | | | | 16 | | | | | | | | | 16 | 100 |
| JP | | | | | | | | | | 16 | | | | | | | | 16 | 100 |
| SU | | | | | | | | | | | 16 | | | | | | | 16 | 100 |
| GA | | | | | | | | | | | | 16 | | | | | | 16 | 100 |
| CA | | | | | | | | | | | | | 15 | | | 1 | | 15 | 94 |
| PF | | | | | | | | | | | | | | 16 | | | | 16 | 100 |
| JB | | | | | | | | | | | | | | | 16 | | | 16 | 100 |
| CB | | | | | | | | | | | | | | | | 16 | | 16 | 100 |
| Overall classification | 248 | 96.9 |
Where: CV = Cedrela sp., JA = Dalbergia sp., CR = Cedrela odorata, MU = A. lecointei, TA = Couratari sp., GU = C. brasiliense, AA = V. araroba, CM = R. montana, JP = Machaerium sp., CU = G. glabra, GA = A. leiocarpa, CA = D. odorata, SU = D. racemosa, PF = C. ferrea, JB = H. oblongifolia, and CB = Dipteryx sp.
The species A. lecointei (MU) was often classified as R. montana (CM) and, to a lesser extent, as D. racemosa (SU), being the species that presented the lowest correct classification rate of 56%. This confusion can be attributed to the color variation (from “yellowish-brown” to reddish brown), due to the storage of extractives in tylose saturation. This phenomenon occurs close to the core and in the intermediate region of the wood (Melo et al. 2013; Abreu et al. 2024). Chromophobic compounds, such as extractives, are highly absorbent of radiation (Bisht et al. 2021), and could lead to the reflection of a greater amount of incident light, resulting in a lower average absorbance than expected.
D. odorata (CA) was misclassified to a lesser extent as Dipteryx sp. (CB), it is possible that this slight confusion is influenced by cross-sectional anatomical similarities between these species of the same genus. This result is consistent with previous research; Santos et al. (2021a) reported confusion in samples of Roupala sp. and Eupla sp., which share similar anatomical characteristics as they belong to the Proteaceae family. Furthermore, Novaes et al. (2023), presented erroneous classifications in Hymenea sp. and Dinizia excelsa, where these wood species had the worst accuracy performance (85.8% and 90.2%, respectively), due to the anatomical similarity of the two species.
In the PLS-DA matrix, for categorization of species with portable equipment, the models with spectra treated with normalization correctly classified 254 of the 256 wood samples. Only 2 incorrectly classified species were recorded: A. lecointei (1) and Machaerium sp. (1) (Table 5).
Table 5
Confusion matrix of the classification of the model resulting from PLS-DA through cross-validation, of spectra collected on the surface produced with a circular saw with the portable equipment treated with normalization.
| Species | Classification of predicted wood by portable NIRS | Correct classification (10LVs) |
| CV | CR | CM | JA | MU | GU | TA | CU | AA | JP | SU | GA | CA | PF | JB | CB | | n | (%) |
| CV | 16 | | | | | | | | | | | | | | | | | 16 | 100 |
| CR | | 16 | | | | | | | | | | | | | | | | 16 | 100 |
| CM | | | 16 | | | | | | | | | | | | | | | 16 | 100 |
| JA | | | | 16 | | | | | | | | | | | | | | 16 | 100 |
| MU | | | 1 | | 15 | | | | | | | | | | | | | 15 | 94 |
| GU | | | | | | 16 | | | | | | | | | | | | 16 | 100 |
| TA | | | | | | | 16 | | | | | | | | | | | 16 | 100 |
| CU | | | | | | | | 16 | | | | | | | | | | 16 | 100 |
| AA | | | | | | | | | 16 | | | | | | | | | 16 | 100 |
| JP | | | | 1 | | | | | | 15 | | | | | | | | 15 | 94 |
| SU | | | | | | | | | | | 16 | | | | | | | 16 | 100 |
| GA | | | | | | | | | | | | 16 | | | | | | 16 | 100 |
| CA | | | | | | | | | | | | | 16 | | | | | 16 | 100 |
| PF | | | | | | | | | | | | | | 16 | | | | 16 | 100 |
| JB | | | | | | | | | | | | | | | 16 | | | 16 | 100 |
| CB | | | | | | | | | | | | | | | | 16 | | 16 | 100 |
| Overall classification | 254 | 99.2 |
Where: CV = Cedrela sp., JA = Dalbergia sp., CR = Cedrela odorata, MU = A. lecointei, TA = Couratari sp., GU = C. brasiliense, AA = V. araroba, CM = R. montana, JP = Machaerium sp., CU = G. glabra, GA = A. leiocarpa, CA = D. odorata, SU = D. racemosa, PF = C. ferrea, JB = H. oblongifolia, and CB = Dipteryx sp.
In portable MicroNIR, only a sample of the species A. lecointei (MU) was confused with R. montana (CM). This confusion can be attributed to the surface quality generated by the circular saw, which makes the wood samples more similar. An erroneous classification was also recorded in a sample of Machaerium sp (JP) classified as Dalbergia sp (JA). These results are consistent with previous studies that discriminated species of the genus Dalbergia. Snel et al. (2018) achieve efficiencies above 90% in PLS-DA models, while Raobelina et al. (2023) report percentages of 82.3% and 81.8%. This confusion in the classification is justified by the similar transverse anatomical characters, mainly in the axial parenchyma, in both species.
These results corroborate previous studies identifying native species. The results revealed that the portable equipment presented better classification performance, compared to bench-top NIR. According to Santos et al. (2021b), the decrease in the predictive performance of benchtop NIR equipment models could be attributed to the spectral reading area. The integrating sphere used in the benchtop equipment was approximately 10 mm in diameter, while the portable optical system had a punctual area, resulting in a smaller representation of the wood surface. This disparity in spectral reading areas could explain the greater effectiveness of portable NIR in discriminating wood species. Despite this difference in classification performance, both instruments proved to be efficient for identifying and classifying timber species.
3.5 Limitations
It is important to note that this study has some limitations. The number of species could have been greater, as there are thousands of tropical species. However, the authors chose these species due to their commercial importance in the State of Minas Gerais, where the research was conducted.
In treatment 1, there was heterogeneity in surface quality, as the samples were used as they were found. Thus, there was no type of control on surface quality in terms of processing time between samples (from months to years) and processing machine (some chainsaw, others circular saw or band saw. The fact that it is the first processing is completely unknown; it is not known how long ago it was carried out, nor which machine was used. From the second treatment onwards, the processing was fully controlled and standardized between the wood samples.
The initial idea of this study was to get an idea of the model's performance depending on the type of processing. The next steps in this line of research will be to produce a collection of wood samples in a real field or yard situation, to provide support for control and inspection actions.