Influence of moisture on the identification of tropical wood species by NIR spectroscopy

2.1 Sample collection and preparation

Native woods were obtained by reusing material from the Amazon region near the municipality of Parauapebas, state of Pará, Brazil (latitude 6°04′18″S, longitude 49°54′08″W and 18 m altitude). Eleven species were selected to prepare the 110 wood specimens (Table 1), totalling 10 repetitions per species (Figure 1A).

Figure 1:

Amazonian species investigated in this study (A), sample preparation diagram (B), and methodological flowchart used for measuring NIR spectra at different moisture content (C).

From the prismatic samples of the Amazonian species mentioned above, radial strips were cut using a circular saw perpendicular to the growth rings, as shown in the diagram in Figure 1B. The dimensions of the samples were 100 mm long, 25 mm wide and 4 mm thick. The samples were duly identified with a pencil and kept in an air-conditioned room at 25 °C and relative humidity of ∼60 %, until they reached equilibrium moisture with the environment. To do this, the samples were monitored by successive weighings until there was no variation in the measurement of their masses.

For this study, the transverse plane would be the most relevant, as it is commonly exposed in sawmills, lumberyards, and wood stacks. However, we used thin wood specimens, which would become brittle if cut into transverse sections. Considering that the wood samples would undergo repeated handling, it was opted to use the radial face in this investigation.

2.2 Determination of moisture content and wood density

The basic density of the wood specimens was obtained from the ratio between the dry mass and saturated volume of the specimens, in accordance with ASTM D2395 (ASTM 2017). The specimens were immersed in water until they reached the saturation point of the fibers and, after reaching equilibrium humidity, they were sent to an oven at 103 ± 2 °C until they reached a constant dry mass. The mass was obtained using an analytical balance and the volume using the hydrostatic method based on Archimedes’ principle.

Subsequently, the density of the species was classified according to the Institute of Agricultural Defense of the State of Mato Grosso – INDEA (2011), which characterizes five classes: very heavy with values greater than or equal to 960 kg/m³; heavy with values between 760 and 960 kg/m³; moderately heavy with values between 560 and 760 kg/m³; light with values between 410 and 560 kg/m³ and very light with values lower than 410 kg/m³.

In sequence, the moisture of the 110 wood samples was determined based on the respective masses measured during the drying phases, from the saturated condition (>FSP) to equilibrium (EMC), totalling four drying stages over 21 days. The ratio of wet mass to dry mass divided by the dry mass of the wood was used, with the result expressed as a percentage.

2.3 NIR spectra records

Two sets of NIR spectral collections were acquired on the radial face of the wood specimens. The first collection step was carried out after preprocessing the wood samples with a circular saw at equilibrium moisture. The second collection was carried out after material saturation in the maximum moisture (MMC). Subsequently, the wood specimens were subjected to air drying in a controlled environment and monitored using a precision analytical balance. From the saturated condition, the mass of each specimen was monitored. The spectral acquisitions and mass measurements were carried out every 10 % loss of water mass from the specimen until reaching the equilibrium condition.

Two spectrometers were used to acquire the absorbance spectra. Table 2 provides some of the technical characteristics of the equipment used for sample analysis. Figure 1C shows the methodological flowchart used for measuring NIR spectra at different moisture content using the two NIR spectrometers.

2.4 Preparation of the spectral matrices and the calibration and validation batches

The first database (Xb, y)b contained Nb = 550 observations obtained at different humidity levels using the benchtop spectrometer (Bruker, MPA model) with the integrating sphere. Only the range between 11,013 and 4,000 cm⁻¹ was considered due to the noise and lack of relevant information in the removed range. The data were ranked in ascending order in relation to basic density values. Then, one out of five samples were systematically placed in the test set (20 %), and the other four were placed in the calibration set (80 %). This procedure led to a calibration set (Xb, y)b, cal and a test set (Xb, y)b, tests corresponding to 4/5 and 1/5 of the data in matrix Nb. This method ensured that both the calibration and test sets had a similar density distribution. Furthermore, the wood drying stages were also provided in the calibration and validation sets for the global models (Figure 2).

Figure 2:

Representative diagram of the preparation of calibration and validation sets.

A second database (Xm, y)m was built from Nm = 550 observations obtained at different moisture levels by the Viavi MicroNIR portable spectrometer on the same samples, considering the total coverage range of the equipment (11,008–5,967 cm⁻¹). The procedure used to select the calibration and validation sets was the same as that adopted for the first database. Thus, the y response values were identical in both validation sets.

2.5 Construction of PCA and PLS-DA models

Principal component analysis (PCA) and partial least squares-discriminant analysis (PLS-DA) was performed using Chemoface software (Nunes et al. 2012) for the determination of the possible correlations between the spectra and the response variables to discriminate and group the timber species according to their spectral signatures.

In PLS-DA, the species were categorized by values, and the samples were grouped into different classes. Thus, all the samples in each category were assigned a value of 0 or 1. A value of 1 was assigned when the sample belonged to the category and a value of 0 when it did not (Pace et al. 2019).

Spectral mathematical transformations proposed by Savitzky and Golay (1964) were also tested using the following schemes: 25.2.1 (25 smoothing points, second order polynomial, first derivative); 15.2.2 (15 smoothing points, second order polynomial, second derivative) as described in Novaes et al. (2023). Standard Normal Variate (SNV) standardization treatments (Barnes et al. 1989) and the Multiplicative Scatter Method (MSC) were also tested (Geladi et al. 1985). For MSC transformation, the average spectrum of the calibration set was used as the reference spectrum. From a practical standpoint, using only the first or second derivative would be preferable. As noted by Sandak et al. (2016), applying derivatives (second derivative, second order polynomial with a 17 smoothing points window) to NIR spectra typically results in reliable predictions for wood-related applications.

Samples presenting large residual variation or spectral noise were previously detected in visual analyses and had their spectra immediately remeasured. Thus, the database did not contain any anomalous samples and no reading was considered an outlier. The experimental design used to construct and validate the discriminant models by PLS is represented in Table 3. In this design, the starting point was considered to be the wavelength of 11,013 cm⁻¹ for the benchtop equipment, which corresponds to the same starting point of the spectra of the portable sensor.

To validate the models, the complete cross-validation method was adopted, and an independent validation in a systematic way, with a maximum of 10 latent variables, was performed. The calibration and independent validation batches were constructed by individual wood specimens (in various moisture content conditions). Each specimen was carefully selected to belong to either the calibration or validation batch, ensuring that the spectra of samples with varying moisture contents remained within the same batch.

To determine the ideal number of latent variables (LV) for each model, the leave-one-out cross-validation was used; each model used the number of LV that maximize the average accuracy. However, this method for selecting the number of latent variables in PLS regression presents a risk of overfitting, especially when working with NIR data, which can contain a large amount of noise and subtle variations. Overfitting occurs when the model captures irrelevant noise or random fluctuations in the spectral data, rather than the true underlying relationships, leading to poor performance on new data. In this study, several cross-validation strategies were simulated and any indications of overfitting in the model was observed. Another reliable approach would involve applying k-fold cross-validation to evaluate the model’s performance across different subsets of the data.

2.6 Classification performances

To select the models, the following parameters were adopted: the number of the latent variables used in calibration (LV) and the percentage of the correct predictions (Cp %). The performance of PLS-DA models was evaluated using the following indicators: sensitivity (SEN), specificity (SPEC) and accuracy (ACCU) as presented by Zhou et al. (2020) and the efficiency rate (EFF) as presented by Oliveri and Downey (2012). It is defined as follows:

where: TP is the true positive, TN is the true negative, FN is the false negative and FP is the false positive.

According to Oliveri and Downey (2012), Sensitivity (also known as recall) is the proportion of samples from the modeled class that are correctly identified, serving as an experimental measure of the confidence level within the class space. Specificity (or true negative rate), on the other hand, represents the proportion of samples that do not belong to the modeled class and are correctly rejected by the model. Accuracy is a common measure of classification performance, typically expressed as the percentage of correctly classified samples. When calculated on an evaluation sample set, it is also referred to as the prediction rate. However, accuracy can be misleading when class sizes are unbalanced, as larger classes may disproportionately influence the overall result. Efficiency is a comprehensive metric calculated as the geometric mean of sensitivity and specificity. It ranges from 0, when either sensitivity or specificity is zero, to 1, when both parameters reach their maximum value of 1.

3.1 Wood density and moisture

The woods presented basic density values ranging from 315 kg/m³ to 882 kg/m³, with an overall average of 694 kg/m³ (Table 4). These woods were classified from very light to heavy, with the Marupá species classified as very light, Aroeira, Angelim, Garapeira, Castanheira, Guajará, Itaúba and Sucupira as moderately heavy and Cumaru, Jatobá and Muiracatiara as heavy, totaling three grouping classes.

This variation in density is typical of tropical woods, due to the high diversity of species and their anatomical characteristics. This, in turn, can help in more accurate identification of species, including realistic variables in the data for later development of classification models. According to Novaes et al. (2023), woods with light density are used indoors, while moderately heavy species have internal structural applications, such as decoration. Heavy woods, on the other hand, are used directly in civil construction to produce beams, rafters, boards, sleepers, frames and others, which have a long useful life.

The moisture content of the wood specimens varied by species and wood density. On the first day, the specimens exhibited the highest saturation levels, with moisture content ranging from approximately 30 % to 70 %. By day 21, all wood specimens had reached equilibrium with the environment (∼12 %).

The lightest wood species, Marupá (Simarouba amara Aubl), had the highest saturated moisture value, approximately 70 %. This can be attributed to its low density, which results in greater porosity and more voids for water molecules to occupy. In contrast, the denser species absorbed significantly less water during saturation, reaching moisture levels close to 40 %, just above the fiber saturation point (FSP).

Just as variations in density add complexity to forest species identification models, moisture content is also a crucial factor. It affects the spectral signal, thereby influencing the reliability of the statistical parameters in multivariate data analysis. In the study by Medeiros et al. (2023b), four types of cellulose pulps with different grammage and moisture conditions were successfully identified through NIR spectroscopy associated with multivariate statistics of spectral data. Thus, it is expected that the same result will be achieved for the tropical woods in this study.

3.2 NIR spectral signature

Figure 3 shows the average spectral signatures of the wood species obtained on the benchtop and portable equipment, with absorbance values initially ranging from approximately 0.2 to 0.6 and peaking between 1.0 and 1.4 on the benchtop sensor and absorbance ranging from 0.1 to 0.4 on the portable NIR. According to Medeiros et al. (2024), the lower absorbance value observed in the portable equipment is attributable to the type, potential and intensity of the lamp’s radiation, as well as the sensor’s acquisition system. Analyzing the spectral behavior of the woods, the Sucupira species stood out with the highest absorbance values at all wavelengths, a behavior opposite to that of the Marupá and Aroeira species, both on the benchtop NIR equipment and on the portable equipment.

Figure 3:

Average spectral signature obtained with the benchtop (MPA FT-NIR) and portable (MicroNIR) spectrometer for the wood species.

The spectral bands corresponding to the interval of 7,000–6,000 cm⁻¹ have O–H and N–H bonds, which generate stretching-type vibrations associated with hydrogen bonds that interact with cellulose, hemicellulose and water (Hein et al. 2017; Soares et al. 2017). The peaks present in the range of 7,000–5,300 cm⁻¹ have a high correlation with the main chemical constituents of wood and can be used to distinguish the wood species (Tsuchikawa and Kobori 2015). Using the range between 4,000 and 6,200 cm⁻¹, Nisgoski et al. (2017) obtained suitable results for the discrimination of the Sugi species (Cryptomeria japonica) coming from various locations in the southern region of Brazil. Pastore et al. (2011) distinguished similar species of mahogany using the same wavenumber region.

According to the Schwanninger et al. (2011) the absorbance peak near the wavenumber of 8,333 cm⁻¹ is attributed to the second overtone of the C–H stretching vibration in cellulose or lignin. The absorption peak present at approximately 6,848 cm⁻¹ corresponds to the first overtone of the OH stretch in cellulose, hemicellulose and water. Wavenumbers of 8,749 and 8,547 cm⁻¹ are related to the aromatic groups of lignin. The wavenumbers of 7,000 cm⁻¹ and 6,287 cm⁻¹ correlate with the amorphous and crystalline regions of cellulose, respectively. The peaks between 6,800 and 4,401 cm⁻¹ are attributed to hemicelluloses; regions close to 6,110–5,697 cm⁻¹ and 4,335–4,146 cm⁻¹ are related to all components of the cell wall, and the wavenumber of 5,995 cm⁻¹ is related to extractives (Pasquini 2018; Schwanninger et al. 2011).

Figure 4 presents the averaged spectral signature of all species obtained from the benchtop and portable equipment in different moisture ranges, corresponding to the desorption of water in the wood. The main absorbance peaks were observed in the most humid conditions of the samples (30–40 % moisture), with the lowest values ​​recorded when the wood reached its equilibrium moisture content (∼12 %). According to Medeiros et al. (2023b), this behavior is attributed to the vibration of the hydroxyl groups, which becomes more pronounced when water is present. E1 and E5 spectra overlap because they were collected under the same equilibrium moisture condition. The key difference is that E1 was recorded at the initial equilibrium moisture content (before saturation), while E5 was obtained after the samples underwent saturation and the subsequent drying process, returning to equilibrium moisture.

Figure 4:

Averaged spectral signature (all species) per drying stage obtained with the benchtop spectrometer, considering the initial moisture content (Stage E1) to the equilibrium moisture content after saturation (Stage E5).

Soares et al. (2017) stated that absorbance spectra are influenced by the free and absorbed water present in the wood. According to Tsuchikawa and Kobori (2015) the NIR range is useful for materials with high moisture values. Thus, NIR has the potential to be used in wood with moisture content above equilibrium (∼12 %). Nisgoski et al. (2016) indicate that bands between 5,200 and 5,050 cm⁻¹ correspond to water present in the wood. Bands from 7,000 to 6,800 cm⁻¹ may refer to variations in water content (Schwanninger et al. 2011).

3.3 Principal component analysis (PCA)

The PCA scores of the Bruker and MicroNIR equipment are presented in Figure 5, considering the original and SNV-treated spectra. The first principal component of the raw spectral data obtained from the Bruker spectrometer captured 99.32 % of the spectral variation, while the PC1 of the spectral data obtained from the MicroNIR spectrometer captured 98.87 % of the variability between samples. These values were similar to those reported by Novaes et al. (2023) in the identification of Manilkara elata, Dinizia excelsa, Goupia glabra, Hymenaea sp., Micropholis melinoniana, and Copaifera sp. using benchtop NIR equipment. Their models achieved a PC1 of 98.98 % and a PC2 of 0.87 %, based on spectra processed with the first derivative.

Figure 5:

PCA of the data obtained from the Bruker (A) and MicroNIR (B) spectrometers of the species in the drying stages E1 to E5.

Figure 5 displays the dispersion of spectra from wood samples of different species and varying moisture content. These plots are provided primarily to visualize the effect of sample moisture on dispersion. A direct comparison between the two instruments would require data normalization. When analyzed on the same scale, the Bruker appears to be more sensitive to moisture content variations, as evidenced by the greater dispersion of the points. The scores measured by the Bruker range from −20 to −5 on the X-axis, while the MicroNIR spectra range from −6 to −1 on the same axis. Notably, in Figure 5, the scores of the same species seem to vary more in the MicroNIR spectra. For instance, species S (in blue) is more widely distributed in the plot obtained with the portable MicroNIR compared to the Bruker benchtop equipment. Due to the significant variation in moisture content, the variation in wood density does not appear to systematically or logically affect the species’ positioning in the plot.

In the scores, it was observed that few species present defined grouping, with overlapping occurring in most of the analyzed woods. One hypothesis for this is the basic density of the wood, which presented similar values, being categorized mainly as moderately heavy. In addition, the moisture of the samples may have caused modification in the spectral signal, as well as the anatomical plane of the wood. This trend was maintained after the mathematical treatment of standard normal variation and in the two NIR sensors evaluated.

Pace et al. (2019) used NIR spectroscopy with benchtop equipment to identify 12 native Atlantic Forest wood species. Their models achieved 99.98 % discrimination based on the sum of PC1 and PC2, using the original data and spectral acquisition from the transverse surface of the wood samples. Medeiros et al. (2023a) also distinguished native species using benchtop NIR, specifically for Amazonian woods such as Manilkara huberi, Caryocar villosum, Bowdichia nitida, and Piptadenia suaveolens. They collected spectra from the radial and transverse surfaces, achieving a total of 93.40 % data explanation with the sum of PC1 and PC2, using NIR spectra processed by the second derivative.

3.4 PLS-DA for classifying wood at specific moisture levels

Table 5 presents the accuracy rates for predicting 11 species across five different moisture stages, demonstrating the effectiveness of both spectrometers in species discrimination. The analysis considers various moisture conditions, including initial moisture, maximum moisture content after saturation, fiber saturation point, and equilibrium moisture content.

The PLS-DA models developed for each drying stage achieved calibration accuracy values exceeding 90 % (Table 5). During validation, accuracy remained above 85 %, reaching 100 % in models based on spectra obtained with the MPA FT-NIR spectrometer for stages 1 and 5, where the species were close to equilibrium moisture. These results indicate that the models performed better for wood with lower moisture content, allowing for accurate species identification through PLS-DA models. The MicroNIR spectrometer also showed satisfactory prediction values, with percentages above 90 % accuracy across the five moisture stages, even though it has a smaller spectral range compared to the MPA.

Lima et al. (2022) developed models for Amazonian species using spectra at equilibrium moisture and reported accuracy rates exceeding 90 %. Soares et al. (2017) employed a portable spectral device to develop PLS-DA models at equilibrium moisture for six native Amazonian species, achieving high efficiency rates, with accuracy ranging from 93 % to 99 %.

3.5 PLS-DA for classifying wood species at variable moisture contents

In this study, the obtained percentages demonstrate that both devices were effective for developing models at different wood moisture stages for the discrimination of Amazonian species. Table 6 presents the percentage values of predictions for the global models of each spectrometer for the discrimination of Amazonian species.

The models were developed using the five previously mentioned moisture stages. For the MPA FT-NIR spectrometer (B), the application of the Standard Normal Variate (SNV) treatment resulted in a model with a higher prediction accuracy percentage (85 %). In the global models developed from spectra obtained with the portable spectrometer, the mathematical treatment of the 1st derivative of Savitzky and Golay (1964) improved species prediction, achieving a prediction accuracy of 75 %.

The models developed from the bench-top spectrometer showed prediction accuracy values ranging from 69 % to 85 %. The models developed from the portable spectrometer had accuracy values between 63 % and 75 % in species discrimination. In qualitative analyses, restricted regions of the spectrum are used to classify or identify the unknown sample, combining the location and strength of absorbance peaks with those of known substances (Tsuchikawa et al. 2003). Ramalho et al. (2018) developed PLS-DA models to classify wood samples of native species Jacaranda sp., Aspidosperma sp., and Apuleia sp. from untreated NIR spectra and reported efficient models with high classification rates, ranging from 86 % to 100 % accuracy, results comparable to those obtained in this study. However, it is evident that incorporating varying moisture conditions into the model leads to lower accuracy percentages in both calibration and prediction, compared to models developed for specific moisture levels. This underscores the significant impact of moisture on the differentiation of native species.

The peaks of higher absorbances found at the stage of maximum water content in the wood, where hydroxyl (–OH) bonds cause interference in the intensity of electromagnetic energy absorption (Igne et al. 2014; Zhou et al. 2020) and variation in the spectrum. These spectral variations disrupt the signal and reduce the precision of the models. Cooper et al. (2011) highlight that while the broad applicability and sensitivity of the technique to various factors are advantageous, they can also pose challenges. Variations in surface texture, density, and moisture content of the wood can significantly impact the response variables.