Abstract

Sample origin and preparation

In the 2000s, the Québec provincial government implemented a large-scale forest inventory program to develop site index and taper equations for the main commercial species. In total, 1,648 plots were sampled in the main ecological regions. About 14,000 trees were harvested and cut into disks for ring analysis. At the end of the program, thousands of disks were selected and kept frozen by the federal government for future research. For this study, 127 wood disks cut at breast height were selected. They were from 50 sites located in the black spruce–moss domain. The sites were chosen to take into account the ecological variability of the study area. The trees were between 45 and 148 years old. The 127 samples were cut in radial sticks of 17 by 10 mm (tangentially by longitudinally). NIRS measurement is sensitive to wood moisture content (Xu et al. 2011), surface preparation (Schimleck et al. 2005a), and cell orientation (Schimleck et al. 2005b). Therefore, more attention was paid to these parameters. Wood surface was sanded using three sandpaper grits (120, 220, and 320) and then cleaned up to remove dust. The radial sticks were conditioned for several days at 60 percent relative humidity and 20°C to reach a wood moisture content of 12 percent.

Near-infrared spectroscopy

Fourier Transform (FT) NIR spectra were acquired with a Perkin Elmer Spectrum 400 FT-MIR/FT-NIR spectrometer, using near-infrared reflectance accessory (NIRA). The diameter of the NIRA opening was 15 mm. Black electrical masking tape, which fully absorbs NIR light, was used to reduce the opening to 5 by 5 mm in order to be closer to the size of increment cores as usually collected in eastern Canada. Only the transverse face of cores is usually prepared and sanded for ring analysis. Hence the NIRS calibrations were developed with the transverse face even though SilviScan works with the radial face. Furthermore Schimleck et al. (2005b) tested NIRS calibrations on both faces of SilviScan samples for air-dry density, MFA, MOE, and several tracheid morphological characteristics. They concluded that the differences between the two sets were small, indicating that either face could be used for NIR analysis. The spectra were collected from the bark side to the pith in 7.5-mm increments. These spectra were initially collected for other reasons. For the purposes of this study, a 5-mm increment size would have been more justified, but the samples were already cut in sticks according to SilviScan's specifications. The spectra were collected over the wavelength range of 680 to 2520 nm, at 1-nm intervals. Thirty-two scans were accumulated and averaged to give a single spectrum by spot. A ceramic standard was used as instrument reference. In total, 1,332 spectra were collected from the 127 samples.

SilviScan analysis

The same 127 radial sticks were then analyzed by SilviScan-3 at FPInnovations in Vancouver, British Columbia. SilviScan processing was precisely described by Cieszewski et al. (2013). The samples were cut into radial strips of 2 by 7 mm (tangentially by longitudinally) and conditioned at 40 percent relative humidity and 20°C to reach a wood moisture content of 8 percent (dry basis). No acetone extraction was performed because the NIR models are intended to be used on increment cores with no extraction. Wood density was determined at a 25-μm radial resolution using the X-ray densitometer unit of SilviScan. MFA was determined at a 5-mm radial resolution using the X-ray diffractometer unit. MOE was estimated using the density and the coefficient of variation of the X-ray diffraction profile intensity (Evans 1999). The density at 8 percent moisture content was converted into basic density as follows (Siau 1995):

where BD is basic density and D₈ is density at 8 percent moisture content. It is assumed that the fiber saturation point is 30 percent and that the water density is 1,000 kg/m³.

Development of NIR calibrations

SilviScan values were matched with the corresponding NIR spots according to measurement position. Data processing was used to prepare and match the databases (PROC SQL, SAS 9.3). For statistical purposes, 34 sites were randomly chosen for the calibration set (88 samples, 918 spectra) and 16 others for the validation set (39 samples, 414 spectra). Multivariate data analysis was performed using The Unscrambler X software (version 10.2, Camo, Oslo, Norway). The presence of outliers was evaluated using the Hotelling T² ellipse on principal components analysis (PCA) scores plots with a 99 percent confidence interval. Model calibration and validation were performed using partial least-squares (PLS) regression. Calibration models were developed with four random cross-validation segments. Several combinations of wavelength ranges and mathematical pretreatments (smoothing, baseline offset, Savitzky-Golay derivatives, and multiplicative scatter correction) were tested. The optimal number of principal components was suggested by the software, based on the significance of the change in explained variance. In addition, to test the model on a validation set, a full cross-validation procedure was also applied on the calibration set. The calibration performance was assessed using the coefficient of determination (R²c), the standard error of calibration (SEC), the standard error of cross-validation (SECV), and the ratio of performance to deviation (RPDc) calculated as the ratio of standard deviation of the reference data to SECV (Williams and Sobering 1993, Jones et al. 2005). The calibration model was then tested on the validation set. The standard error of prediction (SEP) gave a measure of how well the calibration predicted the wood property of the remaining 414 spots (validation set). The predictive ability of the calibration was also assessed using the ratio of performance to deviation (RPDv).

Segmented regression analysis

TSLRs were applied to MFA profiles to determine the transition age from juvenile to mature wood for the 127 samples. Models were built with MFA values measured by SilviScan and MFA values predicted by NIRS. The transition age is frequently expressed in terms of number of rings from the pith (Wang and Stewart 2012). Each ring MFA value was determined from the 5-mm increment MFA profile using data processing (PROC SQL, SAS 9.3). A mathematical pretreatment was applied on MFA profiles to remove the effect of compression wood. Any sudden MFA increase above 0.025° per mm was removed. This threshold was chosen by trial and error.

A TSLR model can be written as follows (Ryan and Porth 2007):

and

where y_i is the MFA value of the ith ring from the pith, x_i is the ith ring, a₁ is the intercept of linear fit to data below the estimated breakpoint, b₁ is the slope of linear fit to data below the estimated breakpoint, a₂ is the intercept of linear fit to data above the estimated breakpoint, b₂ is the slope of linear fit to data above the estimated breakpoint, c is the estimated breakpoint, and ɛ is the error term. The transition point (c) is the intersection point of two lines, which is obtained as follows:

The first step in applying a segmented regression is to graph the data to estimate visually where the break occurs (Ryan and Porth 2007). The transition age seemed to vary between 10 and 40 years for the 127 MFA profiles obtained from SilviScan. An automatic procedure was developed (1) to estimate all the possible starting parameters (a₁, b₁, b₂, c) for transitions between 10 and 40 years using regression analysis (PROC REG, PROC NLIN, SAS 9.3) and (2) to choose the model with the smallest mean squared error using data processing (PROC SQL, SAS 9.3). Once the transition point was determined, the proportions of juvenile and mature wood were calculated, as well as the average values of BD, MOE, and MFA. These juvenile and mature wood properties were obtained by calculating the area-weighted mean, assuming a circular shape of the rings.

Methods comparison for the determination of wood properties

The ability of a model to precisely predict the value from a single sample can be assessed using statistics such as RPD for NIRS models. However, the analysis of population statistics is also a good way to assess the calibration performance. Average wood properties and transition ages determined using SilviScan data were thus compared with values obtained using NIRS predicted data. The comparison was assessed using the Pearson correlation coefficient (r), the mean prediction error (bias), and the 95 percent limits of agreement. High correlation means that data from the two methods are linearly related, even if they cannot agree. The mean prediction error is the mean difference between the measurements of the two methods (Bland and Altman 1995). The closer the mean prediction error is to zero, the less bias there is. The lower and upper limits represent the 95 percent limits of agreement of the mean difference (±1.96 × SD). The smaller the range between these two limits, the better the agreement.

SilviScan wood properties

The statistics of wood properties were calculated using SilviScan data of the 127 samples. Average values were 431.8 ± 31.2 kg/m³ in BD, 14.1 ± 2.3 GPa in MOE, and 14.0° ± 3.1° in MFA. The average BD value was similar to that measured by Jessome (1995). Lessard (2013) obtained similar MOE and MFA values for black spruce samples from 77 sites across Newfoundland, with average values of 13.9 ± 2.3 GPa in MOE and 15.6° ± 3.3° in MFA. The statistics of wood properties were also calculated for the calibration set (88 samples, 918 spectra) and for the independent validation set (39 samples, 414 spectra) for the purpose of comparison (Table 1). Average values for the calibration set were slightly lower for MFA (16.9° vs. 18.1°) but very similar to those of the validation set for BD (428.3 vs. 425.8 kg/m³) and MOE (12.4 vs. 12.0 GPa).

Table 1. Statistics of wood properties for the near-infrared calibration and validation sets.^a

View Fullscreen

Development of NIR calibrations

Table 2 presents the calibration statistics obtained for BD, MOE, and MFA by PLS regression. No outlier was observed or removed. None of mathematical pretreatments significantly improved the precision of the models. Calibration statistics (R²c, SECV) were higher for MOE (0.89, 1.3) than for BD (0.82, 25.0) or MFA (0.82, 3.1). The wavelength range and the number of principal components (PC) were also smaller for MOE (900 to 1,150 nm, 6 PC) than for BD (900 to 2,500 nm, 8 PC) or MFA (680 to 2,500 nm, 9 PC). The models were then tested on the validation set (39 samples, 16 sites, 414 spectra; Table 2). The statistics (R²v, SEP, RPDv) did not drop off when predicting BD (0.85, 26.0, 1.8), MFA (0.79, 3.6, 2.2), or MOE (0.88, 1.3, 2.9). SEC and SEP values were very similar for all wood properties.

Table 2. Summary of near-infrared calibration and validation statistics.^a

View Fullscreen

Good to excellent calibrations were thus obtained for BD, MOE, and MFA. In other words, in this study, strong linear relationships were obtained between the studied wood properties and the absorption values. The validation statistics (R²v, SEP, RPDv) obtained by Xu et al. (2011) for black spruce and balsam fir were slightly higher for air-dried density (0.87, 25, 2.7), but lower for MOE (0.72, 1.0, 1.9) and MFA (0.56, 3.2, 0.9). A value of RPDv above 2.5 is usually defined as satisfactory for screening in a breeding program, whereas values of 5 to 10 are defined as adequate for quality control (Williams and Sobering 1993). However, based on the references cited in this article, the RPDv value is rarely above 3 for calibrations applied to solid wood. The calibrations obtained for BD, MOE, and MFA can thus be considered robust enough for operational use.

Figure 1 shows the correlations between SilviScan data (BD, MFA, MOE) and NIRS values obtained from the calibration and validation sets. Small deviations from linearity were observed for the highest values of BD (500 kg/m³ and more). In this range, BD was underestimated by NIRS. For MFA, small deviations also occurred in extreme values (7° and less, 30° and more). In these ranges, MFA was also underestimated by NIRS. The same trends were observed with the prediction set. For MOE, no deviation was obvious. Several authors have also observed slight deviations from linearity at extreme values in density (Schimleck et al. 2001, Via 2010) and in MFA (Jones et al. 2005, So et al. 2013). The source of nonlinearity for these extreme values seems to be intrinsic to wood material given that the mathematical pretreatments were ineffective (Pérez-Marín et al. 2007). However, these deviations from linearity concern only a few samples in this study.

View Figure

• Download as Powerpoint
• Download Figure

A large part of the variance in BD (85%), MFA (79%), and MOE (88%) was thus explained by FT-NIR spectra. The unexplained variance may be attributed to many sources. NIR light penetration is limited to a few millimeters from the surface as opposed to X-ray beam, which passes through the sample. NIRS may thus have not exactly captured the same properties as SilviScan. Furthermore, the transverse face was scanned by NIRS, whereas SilviScan scans samples on the radial face as previously discussed. Finally, in this study, MFA and MOE were estimated in 5-mm increments using Silviscan and in 7.5-mm increments using NIRS, which may have caused some approximations.

The autocorrelation between the trees from the same sites (site random effect) and between the spots from the same samples (sample random effect) was investigated. The classic PLS regression method proposed by The Unscrambler X software does not take into account the dependence between observations. The random effects were thus assessed on the residuals using mixed regressions (Proc MIXED, SAS 9.3). The random effect was highly significant for the sample and not significant for the site. However, the dependence between the measurements from a same sample is not surprising. This random effect was not considered in the development of calibrations that may have influenced the NIRS predictions at the spot scale and was certainly not considered for the MFA profile, which is the most important to determine the transition point.

A fixed effect related to the position of the measurement (spot) in the increment cores was also highly significant. The higher biases in BD and MOE were observed for spots collected close to the pith and bark. The extractive content is usually higher in these wood parts. Via et al. (2003) observed that MOE and MOR were poorly related to NIR spectra for pith wood in Pinus palustris, while density was strongly correlated. The authors explained that this lack of significance was perhaps attributable to the high concentration of resinous extractives near the pith. Furthermore, compression wood is also very frequent close to the pith. It could thus be interesting to consider the spot position as covariable in a PLS regression.

Two segment linear regressions

Table 3 shows the results of TSLR applied to the 127 MFA profiles obtained with SilviScan. The average transition age was 23 ± 7 years, varying between 8 and 43 years. The average MFA value at the transition point was 13.9° ± 3.2°. All the MFA profiles presented the same inverted J shape (Fig. 2). On average, the MFA value close to the pith was 32.3° ± 5.7°, whereas the MFA value close to the bark was 10.5° ± 2.1°. The average MFA value at the transition point was 2.3 times smaller than the MFA value close to the pith. Alteyrac et al. (2006) applied TSLR to ring area, maximum ring density, and ring MFA to determine the transition point of 36 black spruce trees harvested near Chibougamau, Québec. The transition age at 2.4 m was estimated at 14.0, 17.6, and 15.9 years, respectively. Yang and Hazenberg (1994) applied TSLR to tracheid length to determine the transition point of 10 black spruce trees harvested near Thunder Bay, Ontario. The transition varied between 11 and 21 years among the samples taken at breast height. The transition age estimated at 23 years in the present study is thus relatively similar to results presented in the literature.

Table 3. Statistics of juvenile and mature wood properties calculated from SilviScan data (n = 127).^a

View Fullscreen

View Figure

• Download as Powerpoint
• Download Figure

Once the transition age became known, juvenile–mature wood properties were calculated. The proportion of mature wood was 82.5 ± 13.3 percent. There were highly significant differences in BD, MFA, and MOE between juvenile and mature wood using t tests for paired observations (Proc TTEST paired, SAS 9.3). Average values for mature wood were significantly higher than those for juvenile wood for BD (436.6 vs. 417.9 kg/m³) and MOE (15.0 vs. 10.4 GPa), whereas for MFA, average value for mature wood was significantly lower than that for juvenile wood (12.6° vs. 20.2°). These results are consistent with the definition of juvenile and mature wood (Panshin and De Zeeuw 1980).

Table 4 shows the results of TSLR applied to the 127 MFA profiles obtained by NIRS. Average transition age was 21 ± 7 years, varying between 5 and 40 years. Average MFA value at the transition point was 15.4° ± 3.5°. All the MFA profiles presented the same inverted J shape (Fig. 2). The proportion of mature wood was 85.5 ± 11.3 percent. There were highly significant differences in BD, MFA, and MOE between juvenile and mature wood using t tests for paired observations (Proc TTEST paired, SAS 9.3). Average values for mature wood were significantly higher than those for juvenile wood for BD (435.4 vs. 426.4 kg/m³) and MOE (14.6 vs. 10.5 GPa), whereas for MFA, average value for mature wood was significantly lower than that for juvenile wood (13.0° vs. 20.5°).

Table 4. Statistics of juvenile and mature wood properties calculated from NIRS data (n = 127).^a

View Fullscreen

Of the 127 TSLR models, four and five models failed to converge using SilviScan data and NIRS data, respectively. Each of these cases was investigated. The presence of compression wood seems to be the main reason for nonconvergence when SilviScan data were used. Therefore, the data pretreatment applied to MFA profiles to remove the effect of compression wood was not sufficient for these samples. However the threshold filtering method has corrected 28 percent of the MFA values and helped the convergence of other models while increasing the measurement accuracy of the transition point (Fig. 3). Such pretreatment is thus strongly recommended for operational use. Moreover, it would be relevant to optimize this threshold filtering in function of transition age results rather than by trial and error. This pretreatment has also made it possible to remove some unlikely NIRS predictions in the MFA profile. Nevertheless, a succession of prediction errors was the main reason for nonconvergence when NIRS data were used. A NIRS calibration is a predictive model. Even with an excellent calibration, errors due to variance (called prediction errors) and slight biases occur. However, after inspection, the transition ages of these nonconvergent models did not seem unlikely and the values were kept.

View Figure

• Download as Powerpoint
• Download Figure

It should also be noted that TSLR was only applied in the present study to mature trees aged 45 years and older. Younger trees, which have less or no mature wood, could be sampled in an inventory. In this case, a standard linear regression should give a better fit than a TSLR, which could signal the absence of mature wood. Furthermore, MFA was estimated in 5-mm increments with SilviScan and in 7.5-mm increments with NIRS. The average ring width of the 127 samples was 0.95 mm. Consequently, such resolutions could not give a precision to the nearest ring. The determination of the transition age could also be more accurate with MFA determined or predicted at a resolution of 2 mm or less, particularly in the presence of very narrow growth rings.

Methods comparison for the determination of wood properties

The transition age and juvenile–mature wood properties obtained from SilviScan data (Table 3) were compared with values obtained from NIRS predicted data (Table 4). The two methods were compared using the correlation coefficient and the agreement. Table 5 shows that the correlations between the two data sets were moderate to high. The lowest correlations were observed for transition age (r = 0.59, P < 0.001) and juvenile–mature wood proportion (r = 0.65, P < 0.001). The lowest agreements were also observed for these properties. The transition age was slightly underestimated by NIRS with a mean prediction error (and 95% limits of agreement) of −2.2 ± 6.3 years (−14.6/10.1). The juvenile wood proportion was also slightly underestimated by NIRS with a mean prediction error (and 95% limits of agreement) of −3.0 ± 10.4 percent (−23.4/17.5). Excellent correlation (r = 0.95, P < 0.001) and agreement were obtained for transition MFA with −1.5° ± 1.1° (−3.7/0.7). Good to excellent correlations and agreements were also observed for juvenile–mature wood properties (Table 5).

Table 5. Comparison between the SilviScan and NIRS estimates (n = 127).^a

View Fullscreen

The correlation and the agreement were thus notably higher for transition MFA than for transition age. Close to the transition point, MFA presents very few variations, as opposed to the transition age, which was strongly influenced by the measurement resolution (Fig. 2). For 86 of 127 (68%) samples, the transition age was similar (±5 y) to that obtained using SilviScan data. A few cases with greater discrepancies occurred despite the good calibration statistics for MFA. Discrepancies above 10 years for the transition age were observed for 12 of 127 (9%) samples, including extreme discrepancies above 15 years for three samples (2%). These cases were investigated, and the main observed cause also was the succession of prediction errors, which disturbed the MFA profile. For operational use, the performance of the NIRS calibration should not be inferior to that obtained for MFA in this study. In addition, some rules should be defined to detect potential outliers. Using the 0.05 and 0.95 percentiles of SilviScan measurements as thresholds, a transition MFA below 9° or above 21° could be considered as suspicious, as well as a transition age below 10 years or above 32 years. A better knowledge of the natural range of black spruce wood properties could help to define stronger rules. A larger sampling of all the natural ranges of growth conditions could thus be required.