Classifying Wood Properties of Loblolly Pine Grown in Southern Brazil Using NIR-Hyperspectral Imaging

: Loblolly pine ( Pinus taeda L.) is one of the most important commercial timber species in the world. While the species is native to the southeastern United States of America (USA), it has been widely planted in southern Brazil, where it is the most commonly planted exotic species. Interest exists in utilizing nondestructive testing methods for wood property assessment to aid in improving the wood quality of Brazilian grown loblolly pine. We used near-infrared hyperspectral imaging (NIR-HSI) on increment cores to provide data representative of the radial variation of families sampled from a 10-year-old progeny test located in Rio Negrinho municipality, Santa Catarina, Brazil. Hyperspectral images were averaged to provide an individual NIR spectrum per tree for cluster analysis (hierarchical complete linkage with square Euclidean distance) to identify trees with similar wood properties. Four clusters (0, 1, 2, 3) were identiﬁed, and based on SilviScan data for air-dry density, microﬁbril angle (MFA), and sti ﬀ ness, clusters di ﬀ ered in average wood properties. Average ring data demonstrated that trees in Cluster 0 had the highest average ring densities, and those in Cluster 3 the lowest. Cluster 3 trees also had the lowest ring MFAs. NIR-HSI provides a rapid approach for collecting wood property data and, when coupled with cluster analysis, potentially, allows screening for desirable wood properties amongst families in tree improvement programs.


Introduction
Increasingly, plantation forests are being relied upon to supply society with a multitude of fiber products. The shift from harvesting native forests to plantations has occurred in many countries, and subsequently, there has been a substantial increase in the total area of plantations globally (estimated to be 278 million ha in 2015, up from 167.5 million ha in 1990) [1]. At 7.83 million ha, Brazil has one of the largest plantation estates in the world and also one of the highest utilization rates (91%) of plantation-grown wood for industrial needs. A number of exotic species are planted in Brazil, with pines representing over 1.6 million ha of the resource [2]. The largest conifer plantation areas are in southern Brazil, with a majority in the states of Paraná (42%), Santa Catarina (34%), Rio Grande do Sul (12%), and São Paulo (8%) [2]. Loblolly pine (Pinus taeda L.), a native of the southeastern United States of America (USA), is the most frequently planted species with a plantation area exceeding one all spectra within an image can be averaged, providing a single spectrum that represents the whole area scanned.
As we are interested in both rapidly screening trees and understanding how wood properties vary radially amongst trees within families, we utilized NIR hyperspectral imaging (NIR-HSI) of increment cores coupled with cluster analysis to identify families possessing similar wood properties. SilviScan provided high-resolution wood property data, which were used to verify the properties of trees identified by the cluster analysis. We are not aware of NIR-HSI being used to characterize the wood properties of trees of recognized families within a species. Hence, the objectives of this study were: • To utilize NIR-HSI and cluster analysis to identify Brazilian loblolly pine trees sharing similar characteristics; • To analyze wood properties amongst clusters based on density, microfibril angle (MFA), and stiffness data provided by SilviScan; • To provide an approach for identifying specific families for future commercial plantations based on the potential to produce a range of desirable forest products.

Sample Origin
Wood samples were collected from a 10-year-old progeny trial ( Figure 1) established by BattiStella in the Rio Negrinho municipality, Santa Catarina, Brazil (coordinates 26 • 40 07.24" S 49 • 37 37.38" W). The trial was planted in randomized blocks with 120 families, five replicates, and five plants per linear plot. Stem form, growth, and internode distribution and distance were assessed, and the best 600 trees identified for nondestructive (breast height increment cores from the best performing individuals) and destructive (breast height disc) sampling.
As we are interested in both rapidly screening trees and understanding how wood properties vary radially amongst trees within families, we utilized NIR hyperspectral imaging (NIR-HSI) of increment cores coupled with cluster analysis to identify families possessing similar wood properties. SilviScan provided high-resolution wood property data, which were used to verify the properties of trees identified by the cluster analysis. We are not aware of NIR-HSI being used to characterize the wood properties of trees of recognized families within a species. Hence, the objectives of this study were:


To utilize NIR-HSI and cluster analysis to identify Brazilian loblolly pine trees sharing similar characteristics;  To analyze wood properties amongst clusters based on density, microfibril angle (MFA), and stiffness data provided by SilviScan;  To provide an approach for identifying specific families for future commercial plantations based on the potential to produce a range of desirable forest products.

Sample Origin
Wood samples were collected from a 10-year-old progeny trial ( Figure 1) established by BattiStella in the Rio Negrinho municipality, Santa Catarina, Brazil (coordinates 26°40′07.24″ S 49°37′37.38″ W). The trial was planted in randomized blocks with 120 families, five replicates, and five plants per linear plot. Stem form, growth, and internode distribution and distance were assessed, and the best 600 trees identified for nondestructive (breast height increment cores from the best performing individuals) and destructive (breast height disc) sampling.

Sample Preparation: Radial Strips
For discs obtained from destructive sampling, pith-to-bark radial sections were cut using a bandsaw. Section dimensions were 12.5 mm longitudinally (L), and 12.5 mm tangentially (T); tree diameter determined radial (R) length. Sections were dried, glued into core holders, and cut using a

Sample Preparation: Radial Strips
For discs obtained from destructive sampling, pith-to-bark radial sections were cut using a bandsaw. Section dimensions were 12.5 mm longitudinally (L), and 12.5 mm tangentially (T); tree diameter determined radial (R) length. Sections were dried, glued into core holders, and cut using a twin-blade saw to give radial strips with dimensions: 2 mm (T) × 12.5 mm (L), R varied as described earlier in Section 2.2. Similar strips were prepared using the cores sampled non-destructively.

Wood Property Analysis
From the 600 samples, a subsample of 52 was available for SilviScan and NIR-HSI analysis.

SilviScan
SilviScan analysis was conducted at FPInnovations, Canada, on unextracted radial strips having dimensions, 2 mm tangentially, 7 mm longitudinally, with the radial dimension varying owing to variation in the diameter of trees sampled. The following wood properties were determined:

1.
Air-dry density (referred to as density in the following text) was measured in 25-micron steps using X-ray densitometry [25]; 2.
Wood stiffness estimated at the same resolution as MFA [28].
A controlled environment of 40% relative humidity and 20 • C was used for all measurements.

NIR-HSI Analysis
Hyperspectral images were collected using a Specim FX17 camera (wavelength range 900-1700 nm, 224 wavelengths) fitted with a Specim (OLET 17.5 F/2.1) focusing lens (Specim, Spectral Imaging Ltd., Oulu, Finland). Tungsten halogen lamps illuminated samples, while sample motion and image acquisition (which included a dark current and white reference before collecting an image of each set of samples) were controlled using Lumo software supplied with the system. Samples were analyzed on a black background, and care was taken to minimize exposure of the samples to the intense light of the lamps.
Relative reflectance values for each image was calibrated with the corresponding dark current and white reference data and then transformed to absorbance (A), where A = log 10 1/R. Regions of interest (ROI), corresponding to each of the 4 samples analyzed per set, were identified, cropped, and these images averaged to give an individual spectrum per sample using MATLAB. The resultant spectra were used for the cluster analysis.

Cluster Analysis of Wood Properties Based on NIR-HSI Spectra
To group the 52 trees into groups with similar characteristics, a cluster analysis (hierarchical complete linkage with square Euclidean distance), was performed on the hyperspectral data with a matrix that consisted of 224 columns (variables-wavelengths) and 52 rows (trees) using The Unscrambler X, version 10.2 (64 bit). (Camo Software AS, Oslo, Norway). Clustering is an agglomerative method that begins by treating each sample as a single cluster and, from there, begins to group the samples based on their similarity, as measured by their square Euclidian distance, until they form a large cluster. The smaller groups, formed from the larger dataset, are known as clusters. Grouping allows identification of similar observations that can potentially categorize them. Prior to cluster analysis, the reflectance values were normalized for each wavelength by dividing by the standard deviation calculated for each wavelength. Three parameters were used to determine the final number of clusters identified. These included the cubic clustering criterion, pseudo F statistic, and pseudo t 2 statistic [29]. Two and four clusters were most frequently selected, but we decided to use four as we wanted to have an indication of the most and least suitable trees, something we could not have achieved using two clusters. Prior to cluster analysis, the reflectance values were normalized for each wavelength by dividing by the standard deviation calculated for each wavelength.

Cluster Analysis Based on NIR Spectra
The cluster analysis using the spectral data obtained for each tree identified four groups ( Figure 2). Based on each cluster, mean wood and tree growth values were determined for each measured wood property and summarized in Table 1. Cluster 0 trees (six in total) had superior wood quality having the highest mean density and stiffness. Average height (10.9 m) for Cluster 0 exceeded that of the other clusters, while the average diameter at breast height (DBH) (213 mm) was similar amongst clusters.

Cluster Analysis Based on NIR Spectra
The cluster analysis using the spectral data obtained for each tree identified four groups ( Figure  2). Based on each cluster, mean wood and tree growth values were determined for each measured wood property and summarized in Table 1. Cluster 0 trees (six in total) had superior wood quality having the highest mean density and stiffness. Average height (10.9 m) for Cluster 0 exceeded that of the other clusters, while the average diameter at breast height (DBH) (213 mm) was similar amongst clusters.    In terms of end-use, trees identified in Cluster 0 have the greatest potential (at age 10) for producing structural lumber. In contrast, Cluster 3, which included four trees, had the least desirable solid-wood properties, having the lowest average density and second-lowest stiffness (Cluster 2 had a lower stiffness owing to its higher MFA). Owing to low density and stiffness, trees in Cluster 3 have a lower potential for producing structural lumber and would be better suited for the manufacture of pulp or composites, such as particleboard and medium-density fiberboard. In such products, lower wood density may be beneficial in avoiding the manufacture of panels that are overly heavy, which can be problematic if density becomes too high. Clusters 1 and 2, which included the majority of trees (42), had a very similar average density, while Cluster 1 had a higher average stiffness owing to its lower MFA. These characteristics indicate that wood from these clusters may be suited to low-value solid wood applications, such as pallet stock and furniture parts. However, the characteristics of the two groups are so similar that the lower-value uses identified for both clusters could be interchanged.
The cluster analysis relies on NIR spectra, reflecting differences in wood chemical and physical properties (no wood property information was utilized in the analysis). The average NIR spectra of each cluster show clear differences ( Figure 3) with Cluster 0 trees demonstrating the largest upward baseline shift relative to the other clusters, with Cluster 3 trees having the smallest. In their study of alpine ash (Eucalyptus delegatensis R. T. Baker) solid wood samples, Schimleck et al. [34] associated such shifts with changes in wood density. In terms of end-use, trees identified in Cluster 0 have the greatest potential (at age 10) for producing structural lumber. In contrast, Cluster 3, which included four trees, had the least desirable solid-wood properties, having the lowest average density and second-lowest stiffness (Cluster 2 had a lower stiffness owing to its higher MFA). Owing to low density and stiffness, trees in Cluster 3 have a lower potential for producing structural lumber and would be better suited for the manufacture of pulp or composites, such as particleboard and medium-density fiberboard. In such products, lower wood density may be beneficial in avoiding the manufacture of panels that are overly heavy, which can be problematic if density becomes too high. Clusters 1 and 2, which included the majority of trees (42), had a very similar average density, while Cluster 1 had a higher average stiffness owing to its lower MFA. These characteristics indicate that wood from these clusters may be suited to low-value solid wood applications, such as pallet stock and furniture parts. However, the characteristics of the two groups are so similar that the lower-value uses identified for both clusters could be interchanged.
The cluster analysis relies on NIR spectra, reflecting differences in wood chemical and physical properties (no wood property information was utilized in the analysis). The average NIR spectra of each cluster show clear differences ( Figure 3) with Cluster 0 trees demonstrating the largest upward baseline shift relative to the other clusters, with Cluster 3 trees having the smallest. In their study of alpine ash (Eucalyptus delegatensis R. T. Baker) solid wood samples, Schimleck et al. [34] associated such shifts with changes in wood density.

Radial Variation in Wood Properties
Ring averages based on SilviScan data were determined for each tree to examine wood property differences amongst clusters. Figure 4 (density), shows radial variation for each cluster, with fitted Loess regression lines. Based on Figure 4, trees from the two extreme clusters were clearly different in terms of average ring densities, while Clusters 1 and 2 were very similar. After ring 2, average density values for Cluster 0 were higher than the other clusters, and the difference between Cluster 0 and Clusters 1 and 2 became more pronounced as age increased. In contrast, the average ring densities of the Cluster 3 samples, with the exception of ring 1, were clearly lower than the other clusters. Despite similar initial densities (470 to 480 kg/m 3 ), the ring averages for clusters showed different trends. Cluster 0 ring densities increased rapidly compared to the other clusters, reaching an average of 630 kg/m 3 by age 9. Clusters 1 and 2 also increased, however the rate of increase was inferior to Cluster 0, and ring 9 had an average density of approximately 570 kg/m 3 for both clusters.

Radial Variation in Wood Properties
Ring averages based on SilviScan data were determined for each tree to examine wood property differences amongst clusters. Figure 4 (density), shows radial variation for each cluster, with fitted Loess regression lines. Based on Figure 4, trees from the two extreme clusters were clearly different in terms of average ring densities, while Clusters 1 and 2 were very similar. After ring 2, average density values for Cluster 0 were higher than the other clusters, and the difference between Cluster 0 and Clusters 1 and 2 became more pronounced as age increased. In contrast, the average ring densities of the Cluster 3 samples, with the exception of ring 1, were clearly lower than the other clusters. Despite similar initial densities (470 to 480 kg/m 3 ), the ring averages for clusters showed different trends. Cluster 0 ring densities increased rapidly compared to the other clusters, reaching an average of 630 kg/m 3 by age 9. Clusters 1 and 2 also increased, however the rate of increase was inferior to Cluster 0, and ring 9 had an average density of approximately 570 kg/m 3 for both clusters. Cluster 3 trees had no change in average density (470 kg/m 3 ) for rings 1 to 3. By ring 6, the average density was 520 kg/m 3 , after which little change was observed.
range. However, they reported basic specific gravity (SG) values (determined from oven dry weight, green volume, and density of water), and the trees sampled were older (average age for the different physiographic regions sampled ranged from 22.4 to 24.2 years). To facilitate direct comparison, studies that utilized SilviScan to analyze loblolly pine breast height cores of similar age were identified. For example, Isik et al. [36] reported mean SilviScan density values for two sites each in Georgia aged 15 and 16 years (486 kg/m 3 ), North Carolina aged 18 and 19 years (542 kg/m 3 ), and South Carolina aged 14 and 15 years (499 kg/m 3 ). These density values are similar to the average densities reported for the respective clusters in this study (Table 1), and the trees were also older. Despite the differences in age, DBH values (185 mm Georgia, 228 mm North Carolina, and 235 mm South Carolina) were similar owing to the improved growth rates of loblolly pine in southern Brazil. While it is not possible to compare ring densities, percent latewood (% LW) may provide useful comparative data, and, as noted by Jordan et al. [35], % LW is of critical importance in determining ring density and subsequently, whole-tree density. Jordan et al. [35] used an SG of 0.48 to separate earlywood (EW) from LW, while we used a density of 480 kg/m 3 [37]. Hence, the point within a ring at which the commencement of LW production may be slightly different, but the practical significance of the difference will be small as EW changes rapidly to LW in loblolly pine [37,38]. For Cluster 0 trees, average % LW was initially low (ring 1 = 17.5%, ring 2 = 24.8%) but increased quickly (% LW for rings 6 to 8 was approximately 50%) and reached 65.3% for ring 9. In comparison, Cluster 3 had similar % LW near the pith (ring 2 = 18.1%, ring 3 = 21.3%), but it increased slowly with rings 4 to 9 having % LW in the range 26 to 34%. Clusters 1 and 2 had an initial % LW consistent with the other clusters and reached 40% LW by ring 6 and approximately 54% LW by ring 9. Pictures of the radial samples for trees in Cluster 0 and 3 are shown in Figure 5. Direct comparisons of our density values with loblolly pine grown in the SE USA is imperfect owing to a different measure of wood density often being employed (specific gravity versus air-dry density), different sampling ages, and differences in what an estimate of average density represents. Jordan et al. [35] provide the most comprehensive study of density in loblolly pine across its natural range. However, they reported basic specific gravity (SG) values (determined from oven dry weight, green volume, and density of water), and the trees sampled were older (average age for the different physiographic regions sampled ranged from 22.4 to 24.2 years). To facilitate direct comparison, studies that utilized SilviScan to analyze loblolly pine breast height cores of similar age were identified. For example, Isik et al. [36] (Table 1), and the trees were also older. Despite the differences in age, DBH values (185 mm Georgia, 228 mm North Carolina, and 235 mm South Carolina) were similar owing to the improved growth rates of loblolly pine in southern Brazil.
While it is not possible to compare ring densities, percent latewood (% LW) may provide useful comparative data, and, as noted by Jordan et al. [35], % LW is of critical importance in determining ring density and subsequently, whole-tree density. Jordan et al. [35] used an SG of 0.48 to separate earlywood (EW) from LW, while we used a density of 480 kg/m 3 [37]. Hence, the point within a ring at which the commencement of LW production may be slightly different, but the practical significance of the difference will be small as EW changes rapidly to LW in loblolly pine [37,38]. For Cluster 0 trees, average % LW was initially low (ring 1 = 17.5%, ring 2 = 24.8%) but increased quickly (% LW for rings 6 to 8 was approximately 50%) and reached 65.3% for ring 9. In comparison, Cluster 3 had similar % LW near the pith (ring 2 = 18.1%, ring 3 = 21.3%), but it increased slowly with rings 4 to 9 having % LW in the range 26 to 34%. Clusters 1 and 2 had an initial % LW consistent with the other clusters and reached 40% LW by ring 6 and approximately 54% LW by ring 9. Pictures of the radial samples for trees in Cluster 0 and 3 are shown in Figure 5.    [35] indicates that % LW increases rapidly with age in the SE USA, with regional averages approaching 40% by age 5 and at least 45% by age 9 (at this age trees from stands from the south Atlantic coastal plain had 52 to 53% LW). This data indicates that % LW (and note this is on a regional basis and represents the average of many trees and sites) for trees grown in the USA south would be well above the values reported here for Cluster 3, similar to those reported for Clusters 1 and 2 and lower than those for the Cluster 0 trees. Hart [7], in his comparison of loblolly pine grown in Santa Catarina and the USA South, observed that the ratio of EW to LW was a major difference between the two regions. On a volume basis, trees from Santa Catarina at age 17 were approximately 21 to 22% latewood, while the trees from the USA south (aged 18-21 years) had a much higher proportion (54 to 55%). When making comparisons with Hart [7], it is important to recognize that the families from the breeding trial sampled for this study are not typical of what is planted in industrial plantations. The families also represent a wide range of variability in terms of wood properties and physiology. This is apparent upon a detailed examination of the wood properties of the different clusters. The variation in % LW provides an explanation for the separation of trees into different clusters, but for trees of different families to produce such large differences in % LW while growing at the same location indicates that the trees responded very differently to the growing conditions they experienced.
The average ring MFA of Cluster 3 was consistently lower than the other clusters ( Figure 6). The MFA of ring 1 was 30°, and similar to the other clusters, but by ring 6, it had decreased to approximately 16° and was 5 to 8° lower than Clusters 0 to 2. By ring 9, Clusters 0 to 2, which demonstrated similar radial trends of decreasing MFA, had similar angles (18°) and close to that of Cluster 3 (16°). Three samples (one each in rings 3 to 5) with MFAs close to or above 40° were much higher than the other values observed for these rings, and likely indicate the presence of compression wood in one tree.   [35] indicates that % LW increases rapidly with age in the SE USA, with regional averages approaching 40% by age 5 and at least 45% by age 9 (at this age trees from stands from the south Atlantic coastal plain had 52 to 53% LW). This data indicates that % LW (and note this is on a regional basis and represents the average of many trees and sites) for trees grown in the USA south would be well above the values reported here for Cluster 3, similar to those reported for Clusters 1 and 2 and lower than those for the Cluster 0 trees. Hart [7], in his comparison of loblolly pine grown in Santa Catarina and the USA South, observed that the ratio of EW to LW was a major difference between the two regions. On a volume basis, trees from Santa Catarina at age 17 were approximately 21 to 22% latewood, while the trees from the USA south (aged 18-21 years) had a much higher proportion (54 to 55%). When making comparisons with Hart [7], it is important to recognize that the families from the breeding trial sampled for this study are not typical of what is planted in industrial plantations. The families also represent a wide range of variability in terms of wood properties and physiology. This is apparent upon a detailed examination of the wood properties of the different clusters. The variation in % LW provides an explanation for the separation of trees into different clusters, but for trees of different families to produce such large differences in % LW while growing at the same location indicates that the trees responded very differently to the growing conditions they experienced.
The average ring MFA of Cluster 3 was consistently lower than the other clusters ( Figure 6). The MFA of ring 1 was 30 • , and similar to the other clusters, but by ring 6, it had decreased to approximately 16 • and was 5 to 8 • lower than Clusters 0 to 2. By ring 9, Clusters 0 to 2, which demonstrated similar radial trends of decreasing MFA, had similar angles (18 • ) and close to that of Cluster 3 (16 • ). Three samples (one each in rings 3 to 5) with MFAs close to or above 40 • were much higher than the other values observed for these rings, and likely indicate the presence of compression wood in one tree. MFA variation for loblolly pine grown in the SE USA is not as extensively studied as density (or SG) owing to the cost of analysis. Jordan et al. [39] and Jordan et al. [40] reported MFA within-tree variation for a small number of trees from five physiographic regions (the upper coastal plain was not represented). Jordan et al. [40] reported average MFAs near the pith of approximately 35° for trees from the north Atlantic coastal plain and Piedmont, whereas MFA for the other regions was closer to 30°. For all regions in the SE USA, MFA decreased rapidly in the first few years of growth, and at age 10, trees from the north Atlantic coastal plain and Piedmont had the highest average breast height MFAs (approximately 23°), whereas MFAs for the remaining three regions (hilly, gulf, and south Atlantic coastal plains) were all similar (16 to 17°).
Radial variation in stiffness (Figure 7) was similar across clusters. Stiffness of rings 1 to 3 for all clusters was very low (5 GPa or less) but increased rapidly and by ring 7, averaged 10 GPa. Trees in Cluster 0 had the highest ring 7 stiffness with differences between it and the other clusters becoming more pronounced as the trees grew older. By Ring 9, average stiffness for the Cluster 0 trees was 15 GPa compared to approximately 12.5 GPa for the other 3 clusters. The marked increase in stiffness for Cluster 0 can be related to the higher density of this cluster ( Figure 4) and a decreasing trend in MFA particularly after ring 5 when its average ring MFA became, and stayed, lower than the MFA of Clusters 1 and 2 ( Figure 6). The comparatively high stiffness of trees in Cluster 3 up until ring 6 relates directly to the lower initial MFAs of this cluster. After ring 6, differences in MFA became progressively smaller with age, and while Cluster 3 trees still had the lowest ring MFAs (Figure 6), their lower ring densities (Figure 4) resulted in stiffness increasing at a slower rate than for Clusters 0 to 2. It is also apparent in Figure 7 that a small number of trees have quite high ring stiffness values for rings 8 and 9, with stiffness values being close to 20 GPa. MFA variation for loblolly pine grown in the SE USA is not as extensively studied as density (or SG) owing to the cost of analysis. Jordan et al. [39] and Jordan et al. [40] reported MFA within-tree variation for a small number of trees from five physiographic regions (the upper coastal plain was not represented). Jordan et al. [40] reported average MFAs near the pith of approximately 35 • for trees from the north Atlantic coastal plain and Piedmont, whereas MFA for the other regions was closer to 30 • . For all regions in the SE USA, MFA decreased rapidly in the first few years of growth, and at age 10, trees from the north Atlantic coastal plain and Piedmont had the highest average breast height MFAs (approximately 23 • ), whereas MFAs for the remaining three regions (hilly, gulf, and south Atlantic coastal plains) were all similar (16 to 17 • ).
Radial variation in stiffness (Figure 7) was similar across clusters. Stiffness of rings 1 to 3 for all clusters was very low (5 GPa or less) but increased rapidly and by ring 7, averaged 10 GPa. Trees in Cluster 0 had the highest ring 7 stiffness with differences between it and the other clusters becoming more pronounced as the trees grew older. By Ring 9, average stiffness for the Cluster 0 trees was 15 GPa compared to approximately 12.5 GPa for the other 3 clusters. The marked increase in stiffness for Cluster 0 can be related to the higher density of this cluster ( Figure 4) and a decreasing trend in MFA particularly after ring 5 when its average ring MFA became, and stayed, lower than the MFA of Clusters 1 and 2 ( Figure 6). The comparatively high stiffness of trees in Cluster 3 up until ring 6 relates directly to the lower initial MFAs of this cluster. After ring 6, differences in MFA became progressively smaller with age, and while Cluster 3 trees still had the lowest ring MFAs (Figure 6), their lower ring densities ( Figure 4) resulted in stiffness increasing at a slower rate than for Clusters 0 to 2. It is also apparent in Figure 7 that a small number of trees have quite high ring stiffness values for rings 8 and 9, with stiffness values being close to 20 GPa. Antony et al. [41] reported stiffness values for the same regions examined by Jordan et al. [35]. Corewood stiffness values for static bending samples from a height of 2.4 m ranged from 3.6 GPa (North Atlantic coastal plain) to 5.6 GPa (gulf coastal plain), while outerwood stiffness ranged from 8 GPa (North Atlantic coastal plain) to 10.9 GPa (gulf coastal plain). The stiffness averages for the different clusters (Table 1) are comparable to the corewood averages reported by Antony et al. [41]. However, it is important to note that the corewood samples examined by Antony et al. [41] generally represented rings two to four. Antony et al. [40] also measured stiffness on short, clear bending specimens (static stiffness) versus the dynamic stiffness used here. Jordan et al. [35] observed that outerwood production commences in loblolly pine by year 13 in the SE USA. Hence, the trees examined in this study (aged 10 years) largely comprise corewood.
We have shown that within a Brazilian-grown loblolly pine progeny test, it is possible to use NIR-HSI data to identify groups (or clusters) of trees. Our hypothesis was that clusters would differ in terms of their wood properties and SilviScan data for density, MFA, and stiffness when averaged for trees within clusters, confirmed wood property differences among clusters. Four clusters were identified, with trees identified in Clusters 0 and 3 having the highest and lowest average densities, respectively. Clusters 1 and 2 had very similar average wood properties, with densities closer to that of Cluster 0 than Cluster 3. Radial trends of increasing density and stiffness and decreasing MFA observed for the clusters are consistent with trends reported for loblolly pine [7,35,41,42]. The greatest differences in trends amongst clusters were observed for density (clearly highest for Cluster 0 and lowest for Cluster 3) and MFA (clearly lowest for Cluster 3). The higher densities for Cluster 0 were related to a higher % LW compared to the other clusters, with Cluster 3 having the lowest.
NIR-HSI provides a rapid approach for collecting wood property data and, when coupled with cluster analysis, potentially allows screening for desirable wood properties amongst families in tree improvement programs. Selection will be company-specific, with emphasis likely to be on improving solid wood properties. Hence, as an initial screening (based on NIR-HSI cluster analysis and supported by corresponding SilviScan data), trees in Cluster 0 would be likely candidates for further evaluation and testing. Regardless of the target of the breeding program, NIR-HSI can focus effort on trees with desirable properties. The greater provision of wood property data would allow tree breeders to have the option of selecting fewer families for field experiments, or they could decide which families in existing trials would be the most suitable for meeting the objectives of a company Antony et al. [41] reported stiffness values for the same regions examined by Jordan et al. [35]. Corewood stiffness values for static bending samples from a height of 2.4 m ranged from 3.6 GPa (North Atlantic coastal plain) to 5.6 GPa (gulf coastal plain), while outerwood stiffness ranged from 8 GPa (North Atlantic coastal plain) to 10.9 GPa (gulf coastal plain). The stiffness averages for the different clusters (Table 1) are comparable to the corewood averages reported by Antony et al. [41]. However, it is important to note that the corewood samples examined by Antony et al. [41] generally represented rings two to four. Antony et al. [40] also measured stiffness on short, clear bending specimens (static stiffness) versus the dynamic stiffness used here. Jordan et al. [35] observed that outerwood production commences in loblolly pine by year 13 in the SE USA. Hence, the trees examined in this study (aged 10 years) largely comprise corewood.
We have shown that within a Brazilian-grown loblolly pine progeny test, it is possible to use NIR-HSI data to identify groups (or clusters) of trees. Our hypothesis was that clusters would differ in terms of their wood properties and SilviScan data for density, MFA, and stiffness when averaged for trees within clusters, confirmed wood property differences among clusters. Four clusters were identified, with trees identified in Clusters 0 and 3 having the highest and lowest average densities, respectively. Clusters 1 and 2 had very similar average wood properties, with densities closer to that of Cluster 0 than Cluster 3. Radial trends of increasing density and stiffness and decreasing MFA observed for the clusters are consistent with trends reported for loblolly pine [7,35,41,42]. The greatest differences in trends amongst clusters were observed for density (clearly highest for Cluster 0 and lowest for Cluster 3) and MFA (clearly lowest for Cluster 3). The higher densities for Cluster 0 were related to a higher % LW compared to the other clusters, with Cluster 3 having the lowest.
NIR-HSI provides a rapid approach for collecting wood property data and, when coupled with cluster analysis, potentially allows screening for desirable wood properties amongst families in tree improvement programs. Selection will be company-specific, with emphasis likely to be on improving solid wood properties. Hence, as an initial screening (based on NIR-HSI cluster analysis and supported by corresponding SilviScan data), trees in Cluster 0 would be likely candidates for further evaluation and testing. Regardless of the target of the breeding program, NIR-HSI can focus effort on trees with desirable properties. The greater provision of wood property data would allow tree breeders to have the option of selecting fewer families for field experiments, or they could decide which families in existing trials would be the most suitable for meeting the objectives of a company investing in commercial plantations on a large-scale. It also provides a rapid approach for collecting wood property-related data from increment cores or discs. The same outcome could likely be achieved using a standard NIR spectrometer [11][12][13] where spectra at a given spatial resolution are collected and then averaged, but the time required to collect data is greatly reduced using a NIR-HSI system.

Conclusions
Near-infrared hyperspectral images (NIR-HSI) were collected from 52 radial strips obtained from a 10-year-old loblolly pine progeny trial located in Brazil. The NIR data were subject to a hierarchical complete linkage with square Euclidean distance cluster analysis. Four clusters (0, 1, 2, 3) were identified, and we demonstrated that clusters differed in terms of average air-dry density, microfibril angle (MFA), and stiffness. Average ring data were used to compare clusters, and Cluster 0 trees had the highest average ring densities, whereas those in Cluster 3, the lowest. Cluster 3 trees also had the lowest ring MFAs. Funding: National Council for Scientific and Technological Development-CNPq for granting support for the execution of the research project. FINEP for the grant to develop the project "Use of biotechnological tools to improve the genetic quality of Pinus taeda and alternatives species".