UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Assessment of some wood properties by near infrared spectroscopy Ayanleye, Samuel 2020

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
24-ubc_2020_november_ayanleye_samuel.pdf [ 1.6MB ]
Metadata
JSON: 24-1.0392877.json
JSON-LD: 24-1.0392877-ld.json
RDF/XML (Pretty): 24-1.0392877-rdf.xml
RDF/JSON: 24-1.0392877-rdf.json
Turtle: 24-1.0392877-turtle.txt
N-Triples: 24-1.0392877-rdf-ntriples.txt
Original Record: 24-1.0392877-source.json
Full Text
24-1.0392877-fulltext.txt
Citation
24-1.0392877.ris

Full Text

1  Assessment of Some Wood Properties by Near Infrared Spectroscopy    by  Samuel Ayanleye  B.Agric.Tech., Federal University of Technology, Akure, 2017      A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF  MASTER OF SCIENCE    in  The Faculty of Graduate and Postdoctoral Studies  (Forestry)    THE UNIVERSITY OF BRITISH COLUMBIA  (Vancouver)    August 2020  ©Samuel Ayanleye, 2020 ii  The following individuals certify that they have read, and recommend to the Faculty of Graduate and Postdoctoral Studies for acceptance, a thesis entitled:     Assessment of Some Wood Properties by Near Infrared Spectroscopy     submitted by  Samuel Ayanleye  in partial fulfillment of the requirements for  the degree of   Master of Science    in    Forestry   Examining Committee:  Dr. Stavros Avramidis, Professor, Forestry, UBC Supervisor Dr. Shawn Mansfield, Professor, Forestry, UBC Supervisory Committee Member Dr. Simon Ellis, Assistant Professor, Forestry, UBC Supervisory Committee Member   Dr. Savvas Hatzikiriakos, Professor, Chemical and Biological Engineering, UBC  Additional Examiner    iii  Abstract  Near-infrared spectroscopy (NIRS) is a suitable technique for characterizing many materials, including wood, and has been used to predict several wood properties. However, existing reports on this use of NIRS have paid little attention to the effect of wood surface condition and grain orientation. This study therefore used NIRS to assess wood density, modulus of elasticity, modulus of rupture, grain angle, and annual ring width, studying whether and how surface condition and grain orientation affected the measurement of these properties. The research focused on using NIRS coupled with partial least squares regression (PLS-R) to predict the properties of two softwoods (Western hemlock and Douglas-fir). PLS-R models were calibrated and validated using the test-set validation method. The predictive accuracies based on grain orientation (quarter-sawn and flat-sawn) and wood surface condition (rough and smooth) were compared. Models developed using reduced wavelengths also showed the possibility of predicting these properties using a narrow spectral range. The results of this study showed that calibrations based on mixed sets, which included both cross-sections, were inferior to those based on these cross-sections separately. Promising predictive models were obtained for density (Rp2 = 0.66), modulus of elasticity (Rp2 = 0.78), and modulus of rupture (Rp2 = 0.82), with poor correlations for grain angle and annual ring width (Rp2 ≤ 0.50). Further, the rough surface predictions outperformed those from the smooth surface for all properties. The quarter-sawn sections also showed better predictive ability than the flat-sawn sections for both surface conditions. The only exception was for modulus of rupture, where the trend was reversed. The results therefore show the potential for using NIRS as a non-destructive technique to predict the properties of wood.   iv  Lay Summary  The main goal of this thesis was to assess the feasibility of using NIRS to predict the properties of two softwoods based on different surface conditions and grain orientations. NIRS is a rapid analysis technique useful for predicting wood properties. This technique has several advantages, including scant need for sample preparation, and the non-destructive nature of sampling. However, the reliability of this method depends on the accuracy of spectral acquisition and reference measurements. While plenty of research has used NIRS to assess wood properties, little attention has been paid to applying the method to two species in Canada: Western hemlock and Douglas-fir. This study therefore focused on using NIRS to predict the properties of these softwoods. Using the partial least squares procedure, I built NIRS calibrations to predict certain wood properties and observed the influence of surface condition and grain orientation on NIRS performance.    v  Preface  I, Samuel Ayanleye affiliated to the Wood Physics and Drying Laboratory at the University of British Columbia, Point Grey Campus, conducted the work presented in this thesis. I was responsible for all the experiments, data collection, analyses, result interpretation and wrote the thesis.   vi  Table of Contents Abstract ....................................................................................................................................................... iii Lay Summary ............................................................................................................................................. iv Preface .......................................................................................................................................................... v Table of Contents ....................................................................................................................................... vi List of Tables ............................................................................................................................................ viii List of Figures ............................................................................................................................................. ix Acknowledgments ...................................................................................................................................... xi Dedication .................................................................................................................................................. xii 1.0 INTRODUCTION ........................................................................................................................... 1 1.1 Outline of thesis ................................................................................................................................. 2 2.0.  LITERATURE REVIEW............................................................................................................... 4 2.1. Non-destructive Evaluation of Wood ............................................................................................. 4 2.2. The Physics of Light and Spectroscopy .......................................................................................... 5 2.2.1 Physics of light ............................................................................................................................ 5 2.2.2. Spectroscopy .............................................................................................................................. 5 2.2.2.1. Infrared spectroscopy (IRS) .............................................................................................. 6 2.2.2.2. Near infrared spectroscopy (NIRS) .................................................................................. 7 2.2.2.3. Effect of surface roughness on NIR spectra .................................................................... 8 2.3. Western Hemlock and Douglas-fir ................................................................................................. 9 2.3.1. Distribution, properties, and uses of Western hemlock ......................................................... 9 2.3.2. Distribution, properties, and uses of Douglas-fir ................................................................. 10 2.4. Selected Wood Properties .............................................................................................................. 11 2.4.1 Density ....................................................................................................................................... 11 2.4.2. Mechanical properties ............................................................................................................ 12 2.4.3. Slope of grain (SOG) ............................................................................................................... 12 2.4.4. Annual ring width (ARW) ...................................................................................................... 13 2.5. NIRS Wood Assessment  ............................................................................................................... 13 2.5.1. Density ...................................................................................................................................... 13 2.5.2. Mechanical properties ............................................................................................................ 14 2.6. Chemometrics and Multivariate Calibration .............................................................................. 15 2.6.1. Effect of spectral pre-treatment on chemometrics ............................................................... 16 2.6.2. Effect of grain orientation and surface roughness on NIRS calibrations .......................... 17 vii  2.6.3. Principal component analysis (PCA) ..................................................................................... 18 2.6.4. Partial least squares regression (PLS-R) .............................................................................. 19 2.6.5. Validation of chemometrics model ........................................................................................ 19 2.7. Objective ......................................................................................................................................... 20 3.0  MATERIALS AND METHODS ................................................................................................ 21 3.1. Sample Preparation ....................................................................................................................... 21 3.2. NIRS Measurements ...................................................................................................................... 23 3.3. Determination of Wood Density (ρ) .............................................................................................. 25 3.4. Determination of MOE and MOR ................................................................................................. 26 3.5. Annual Ring Width Measurements .............................................................................................. 27 3.6. Slope of Grain Measurements ....................................................................................................... 28 3.7. Multivariate Data Analysis ........................................................................................................... 29 4.0 RESULTS AND DISCUSSION ................................................................................................... 31 4.1 NIRS Calibration ............................................................................................................................ 31 4.2 Density and Mechanical Properties ............................................................................................... 32 4.2.1 Density based on calibrations for all samples ........................................................................ 32 4.2.2. Density based on calibrations for the Qs and Fs samples .................................................... 38 4.3 Modulus of Elasticity ...................................................................................................................... 41 4.3.1 MOE based on calibrations for all samples ........................................................................... 41 4.3.2 MOE based on calibrations for Qs and Fs samples .............................................................. 45 4.4 Modulus of Rupture ........................................................................................................................ 48 4.4.1 MOR based on calibrations for all samples ........................................................................... 48 4.4.2 MOR based on calibrations for Qs and Fs samples .............................................................. 52 4.5 Slope of Grain (SOG) ...................................................................................................................... 56 4.5.1 SOG based on calibrations for all samples ............................................................................ 56 4.5.2 SOG based on calibrations for Qs and Fs samples ............................................................... 58 4.6 Annual Ring Width (ARW) ............................................................................................................ 59 4.6.1 ARW based on calibrations for all samples ........................................................................... 59 4.6.2 ARW based on calibrations for Qs and Fs samples .............................................................. 62 5.0 CONCLUSION ............................................................................................................................. 64 6.0 FUTURE RESEARCH ................................................................................................................. 66 REFERENCES .......................................................................................................................................... 68  viii  List of Tables  Table 3.1: Sample categorization for NIR analysis, density and mechanical properties…….…..23 Table 4.1: Descriptive Statistics of the wood properties from the calibration and prediction set………………………………………………………………………………………………….…..….31  Table 4.2 Calibration statistics for ρ based on all samples (n = 240) and Qs and Fs samples (n = 120) .…………………………………………………………………………………………………….…. 35 Table 4.3: Calibration statistics for MOE based on all samples (n = 240) and Qs and Fs samples (n = 120) …………………………………………………………………………………………….….. 43 Table 4.4: Calibration statistics for MOR based on all samples (n = 240) and Qs and Fs samples (n =120) .………………………………………………………………………………………... 48 Table 4.5: Calibration statistics for SOG based on all samples (n =240) and Qs and Fs samples (n =120) ………………………………………………………………………………………...  55 Table 4.6: Calibration statistics for ARW based on all samples (n = 240) and Qs and Fs samples (n = 120) ………………………………………………………………………………………... 59   ix  List of Figures  Figure 2.1: The distribution map of western hemlock and Douglas-fir in the Pacific Northwest (Natural Resources Canada, 2015) ………………………………………………………………….. 10 Figure 3.1: Schematic of cutting pattern of samples ………………………………………………. 21 Figure 3.2: Samples from the rough and smooth surfaces ………………………………………………………………………………………………………….... 22 Figure 3.3: Cutting pattern of samples for MOE and MOR measurements .…………………….. 22 Figure 3.4: Cutting pattern of samples for ρ measurements …………………………………........ 23 Figure 3.5: Schematic diagram of the NIRS system (adapted from Zhou et al. 2019) .…………………………………………………………………………………………………… 24 Figure 3.6: The NIR system setup used for spectral acquisition …………………………………………………………………………………………………………… 25 Figure 3.7: Material Test System (model 810) for measuring MOE and MOR ………………………………………………………………………………………………………….. 27 Figure 3.8: Measuring the SOG of wood …………………………………………………………… 28 Figure 4.1: A typical graphical representation of raw spectrum …………………………………. 32 Figure 4.2: PCA plot for raw spectra obtained from a rough surface (top) and smooth surface (bottom), where DF: Douglas-fir flat-sawn, DQ: Douglas-fir quarter-sawn, HF: Hemlock flat-sawn, HQ: Hemlock quarter-sawn …………………………………………………………………… 34 Figure 4.3: The relationship between measured and NIR-predicted values of ρ (kg/m3), based on rough (top) and smooth (bottom) surfaces for two softwoods (hemlock and D-fir) .…………………………………………………………………………………………………………... 37 Figure 4.4: PCA plot for the best model based on the rough surface and quarter-sawn section (top) and the smooth surface and quarter-sawn section (bottom) .…………………………………………………………………………………………………………... 39 Figure 4.5: The relationship between measured and NIR-predicted values of ρ (kg/m3), based on the rough (top) and smooth (bottom) surfaces (quarter-sawn section) for the softwoods …………………………………………………………………………………………………………... 41 Figure 4.6: The relationship between measured and NIR-predicted values for MOE (MPa), based on the rough surface (top) and smooth surface, 1300–2300 nm (bottom) for the softwoods …………………………………………………………………………………………………………… 44 Figure 4.7: The relationship between measured and NIR-predicted values for MOE (MPa), based on the quarter-sawn (top) and flat-sawn (bottom) sections for the softwoods …………………………………………………………………………………………………………… 48 Figure 4.8: Regression coefficients based on raw data for the rough surface (top) and the smooth surface (bottom) .………………………………………………………………………………………. 51 Figure 4.9: The relationship between measured and NIR-predicted values for MOR (MPa), based on the rough surface (top) and smooth surface (bottom) for the softwoods …………………………………………………………………………………………………………… 52 Figure 4.10: The relationship between measured and NIR-predicted values for MOR (MPa), based on the rough surface (flat-sawn) (top) and smooth surface (quarter-sawn) (bottom) for the x  softwoods ………………………………………………………………………………………………. 55 Figure 4.11: PCA plot based on raw spectra for SOG .…………………………………………………………………………………………………………... 57 Figure 4.12: The relationship between measured and NIR-predicted values for SOG (), based on the raw spectra for the softwoods ………………………………………………………………… 57 Figure 4.13: The relationship between measured and NIR-predicted values for SOG (), based on the flat-sawn section for the softwoods ………………………………………………………….. 59 Figure 4.14: The relationship between measured and NIR-predicted values for ARW (mm), based on the rough surface (top) and smooth surface (bottom) for the softwoods …………………………………………………………………………………………………………… 61 Figure 4.15: The relationship between measured and NIR-predicted values for ARW (mm), based on the rough surface (top) and smooth surface (bottom) in the quarter-sawn section for the softwoods ………………………………………………………………………………………………. 63    xi  Acknowledgments   I would like to thank the Almighty God who gave me life, wisdom, and strength to complete my Master of Science in Forestry. My deepest appreciation goes to my supervisor Dr. Stavros Avramidis for accepting me into his research group which I considered a privilege. I also thank him for his time, guidance and support throughout my studies. I would also like to sincerely appreciate my supervisory committee members, Dr. Mansfield and Dr. Ellis for their support and contributions toward the completion of this thesis.   I would like to also thank all my Lab mates, Poorya Koochebaghy, Sohrab Rahimi, and Dessie Tibebu for their friendship, and support during my program. I say big thank you to Sohrab Rahimi for his help during the sample preparation stage.   Special thanks to the Mastercard community for the high level of support and assistance during my degree. I am also grateful to all my friends, family and relatives in Nigeria, whose unending love, care and support gave me the push to finish this program.  My heartfelt gratitude goes to the donors of the Mastercard Foundation Scholarship, International Student Award, Jack and Mavis Toovey Scholarship, David and Macaree Scholarship in Wood Science for their financial support.   xii  Dedication  To the Ayanleye family, who has always been a continuous source of support and motivation to me during the tough moments of life.  I will be forever thankful for your unconditional love towards me. 1  1.0 INTRODUCTION  Wood is an important construction material that requires high degree of structural performance and reliability (Liang et al. 2016). Consequently, it has a broad range of applications (Nguyen et al. 2017). Wood is an anisotropic, hygroscopic, and heterogenous bio-based material with a porous structure that presents variations between species at both microscopic and macroscopic levels (Hans 2014). This variation is one of the most important challenges that wood product producers face when trying to accurately characterize quality for appropriate end product design and manufacturing (Schajer and Orhan 2006, Fujimoto et al. 2008).   Western hemlock is the single most plentiful tree species found on the coast of British Columbia (BC), making up 18% of the volume of BC’s growing stock (Middleton and Munro 2001). Hemlock wood is of medium density, has moderate strength and stiffness, and is mostly used for construction purposes (Kenneth 2001). Douglas-fir (D-fir) makes up to 8.8% of the provincial growing stock (D-fir species profile n.d.) and grows along the BC coast and in the interior of the province. Coastal D-fir wood is known for its superior strength, stiffness, and higher density (Cherry et al. 2008) and is mostly used as dimensional timber, plywood, flooring, and in construction (Johnson and Gartner 2006).  Wood density (ρ) influences its physical and mechanical properties and is regarded as an important measure of wood quality (Jozsa et al. 1989). The ρ within a tree increases radially from pith to bark and decreases vertically with height in the stem (Kiaei et al. 2015, Tsoumis 1991, Todoroki et al. 2012). The moduli of elasticity and rupture (MOE and MOR) are two important properties used in choosing wood as a structural material (Missanjo and Matsumura 2016). MOE is an indication of stiffness, while MOR is an indication of the strength of a board or structural member (Johnson and Gartner 2006). The annual ring width (ARW), which is composed of earlywood with thinner walls and latewood with thicker walls—and thus has a smaller lumen size and higher ρ (Fromm et al. 2001)—has been reported as an important variable in the prediction of MOE and MOR of wood (Haartveit and Flæte 2002). The slope of grain (SOG) relative to the length of a timber is another factor that contributes to understanding wood’s strength properties. The strength varies with the orientation of the grain, making the measurement of SOG an important requirement for machine grading and structural analysis (Ingemi and Yu 2019).  2  Uncertainty about some of the properties of wood has led to the development of prediction methods by indirect observation—that is, obtaining measurements without damaging the material or altering its properties (Romer 2013). Measurements of ρ and SOG, which directly affect a product’s mechanical properties, are important to the wood industry. Although traditional methods have been employed to estimate these properties, a suitable method that is non-destructive, swift, easy to implement within the production line, and requires minimal sample preparation is highly desirable.   Near infrared spectroscopy (NIRS) is the measurement of the wavelength and intensity of a material’s absorption of near-infrared light. NIR light spans the 800–2500 nm range and is energetic enough to excite overtones and combinations of molecular vibrations to higher energy levels. NIRS is normally used for the quantitative measurement of functional groups, especially O-H, N-H, and C-O, containing physical and chemical information about a sample (Schimleck et al. 2001). The advantages of NIRS include contactless application, fast acquisition times, non-destructive sampling, and the potential for on-line or portable applications (Druckenmüller et al. 2018). NIRS has been successfully applied to estimate several properties of wood, such as moisture content (M), density (ρ), and mechanical properties (Leblon et al. 2013, Acquah et al. 2018).   This study examines the use of NIRS to predict the density, bending strength, stiffness, grain angle, and annual ring width of hemlock and D-fir. The hypothesis is that NIRS will be a swift and effective tool to predict the aforementioned properties using multivariate analysis of the acquired spectra in real time. Such capability may allow for sorting before processing and/or property evaluation, thereby increasing product quality and decreasing processing costs.   1.1 Outline of thesis  Chapter 1 introduces what the research was about and why it was carried out. Chapter 2 provides background information on the non-destructive evaluation of wood, as well as spectroscopy, specifically explaining the NIRS utilized in this study. It also explains Western hemlock and D-fir properties and applications, and describes in detail the application of NIRS technology in wood property assessment. Chapter 2 ends with a statement of the thesis’ objective. Chapter 3 provides 3  information about the study material and methods: sample preparation, NIR measurements, procedures for testing wood properties (density, MOE, MOR, grain angle, and annual ring width), and the method of analysis. Chapter 4 presents the results of the study and discusses them in relation to the objective of the thesis. Finally, Chapter 5 draws general conclusions from this research and makes recommendations for further studies.    4  2.0.  LITERATURE REVIEW  2.1. Non-destructive Evaluation of Wood  Non-destructive evaluation (NDE) is a general term that encompasses various methods or techniques used to assess material properties, components, or entire process units without damaging the material in question. These techniques can also detect, characterize, or measure the presence and magnitude of natural and processing defects in wood (ASNT 2016). The importance of these methods increases as the timber industry continues to seek better ways to test wood products in a non-destructive manner. These approaches have also gained attention for the evaluation of complete wooden structures (Dunn 1992).  The instruments used for NDE may be quite diverse, and some of their benefits and limitations have been reported in detail (Bucur 2003, Niemz and Mannes 2012). NDE as a method of identifying the physical and mechanical properties of wood, from standing trees to in-place structures, does so without altering the material’s end-use capabilities. The resulting information can be used to make decisions about appropriate applications and design changes (Ross et al. 1998). Research on NDE arose mainly from the need to provide methodologies or techniques for assessing wood and wood-based materials and thereby to promote better decision-making processes regarding the use of the products (Brashaw et al. 2009).   Historically, the wood community has developed and used various NDE techniques in applications such as sorting or grading structural products. Two common examples are machine stress rating (MSR) of timber and ultrasonic grading of veneer. In North America, MSR grading couples visual sorting criteria with NDE practices to assign a timber to an established grade (Galligan et al. 1977). Laminated veneer manufacturing facilities use stress wave NDE techniques for sorting incoming veneer into strength classes before final product assembly. These are established through empirical relationships between the travel velocity of induced stress waves through the veneer, and its strength (Sharp 1985). A detailed review of NDE techniques, especially those utilized for the assessment of wood properties, was conducted by Ross (2015).    5  2.2. The Physics of Light and Spectroscopy   2.2.1 Physics of light  Light is defined as any natural agent that stimulates the sight and improves the visibility of things (Sliney 2016). It has also been regarded as a fraction of the radiation that can excite the human visual system (Zwinkels 2015). Visible light is the electromagnetic radiation (EMR) within the electromagnetic spectrum that is easily observed by the human visual system (CIE 1987). In physics, “light” sometimes refers to EMR of any wavelength, whether it is visible or not (Smith 2006). The known properties of light include intensity, direction of propagation, frequency or wavelength, polarization, and speed (Uzan and Leclercq 2008). Like other EMR, light propagates as waves, and the energy imparted by these waves is absorbed, reflected and transmitted. The absorbed energy of these waves is called a photon and represents quanta of light.  For several centuries, scientists disagreed about whether light should be regarded as particles or waves. Sir Isaac Newton, after conducting various experiments in the 17th century, argued that the nature of refraction and reflection could only be explained if light were composed of particles, as waves would not travel in straight lines. Although many physicists subsequently attempted to articulate a wave theory of light, this concept was not accepted until Sir Thomas Young experimentally demonstrated wave interference (Young 1807). Physicists later reached a consensus that light could be characterized concurrently as both a stream of particles and a wave, depending on the type of experiment (Sliney 2016).   There are three main methods of dispersing radiation: refraction, diffraction, and interference. The interactions between particles such as protons and ions, and their interactions with other particles due to collision energy, is a field of study known as “spectroscopy.”   2.2.2. Spectroscopy  The field of spectroscopy is concerned with studying the interactions between electromagnetic waves and matter. Light interacts with the atoms and molecules of matter in several ways, through absorption and emission. Measurement of these interactions permits us to identify specific wavelengths of light that interact with the atoms and molecules, and to measure the amount of light that is absorbed or emitted at certain wavelengths (Spectroscopy 2017).   6   Spectroscopy falls into two types: atomic spectroscopy (AS) and molecular spectroscopy (MS). AS comprises atomic absorption and atomic emission spectroscopy, while MS comprises mass spectroscopy, infrared spectroscopy, nuclear magnetic resonance, and ultraviolet and visible spectroscopy. In AS, every element has a distinct atomic structure characterized by a positively charged nucleus surrounded by the number of electrons necessary to maintain neutrality (Sahin 2019). In MS, EMR interacts with a material and produces a spectrum from which relevant compositional and structural information about the material can be deduced (Goodman 1994).   Spectroscopic experiments primarily use light sources that emit radiation of discrete and well-defined frequencies (Chu et al. 2018) and its data are represented by a range —for instance, a plot of the material’s reflectance/absorbance as a function of wavelength/frequency (Sandak et al. 2016). This spectrum is indicative of the material’s molecular structure and has unique properties, including band position, width, and intensity, which are all used to obtain information about functional groups or to monitor molecules under different conditions (Mollaoglu et al. 2018). The spectral response is used to either qualitatively determine the chemical constituents or quantitatively determine the amounts of specific compounds present in the material.   2.2.2.1. Infrared spectroscopy (IRS)                                                                                                                        S IRS is an analytical technique for the characterization of a wide range of materials and is commonly used in research and production for quality control purposes (Theophanides 2012). This technique involves the interaction between infrared radiation (IR) and matter and is mostly used to identify and quantify various organic compounds present in a material, via their selective IR absorption. The history of IRS dates back to the 19th century, and the first spectrometer was built in 1835. By the beginning of the 20th century, IRS had become a valuable tool for identifying and characterizing chemical compounds and materials (Theophanides 2012). Thereafter, it was applied to more complicated molecules such as proteins (Elliot and Ambrose 1950) and later, DNA (Colthup et al. 1990).   IRS is a form of vibrational spectroscopy. The absorption of IR radiation by matter results in molecular bond vibrations. Each molecule possesses unique vibrational characteristics, and the resulting IR spectra make vibrational spectroscopy a good method for characterizing molecular 7  structures. It is a reliable way to investigate functional, structural, and compositional changes in cells, tissues, and biological molecules. It is also one of the most important analytical techniques used to study almost all types of liquids, solutions, and pastes (Mollaoglu et al. 2018).   The IR range refers to the part of the electromagnetic spectrum between the visible and microwave regions, and it is divided into three regions: near-infrared, mid-infrared, and far-infrared (Anastasopoulou and Theophanides 1972). The IR energy excites the molecules at the surface of the material, and part of the energy is absorbed selectively and proportionally to the quantity of the excited molecules (Sandak 2013). Different molecular interactions, such as, C–H, O–H, and N–H, are stimulated and often result in phenomena such as stretching (symmetrical or non-symmetrical), scissoring, rocking, wagging, and twisting—phenomena that depend on factors such as the molecular structure, chemical composition, or physical properties of the surface measured (Coates 2000).   An instrument called a “spectrometer” measures the light that is reflected or transmitted from a sample and uses a detector to plot transmittance through the molecules against the wavelength of the radiation. The resulting spectrum is better visualized in graphs of transmittance/reflectance plotted against wavelength (Infrared Spectroscopy n.d.). This type of information can be obtained from a single experiment, and from small samples or powders.    2.2.2.2. Near infrared spectroscopy (NIRS) NIRS is a non-destructive technique for analyzing organic materials that is rapidly growing in popularity. Although the use of spectrophotometers to measure spectra is an old technique, the NIR region of the electromagnetic spectrum was not properly utilized, as it was considered to contain non-structural information (Barton 2004). The earliest analytical application of NIRS was to determine the moisture content (M), crude protein, and oil concentrations of cereal grains (Ben-Gera and Norris 1968a); it was also used in the agriculture, food, paper, polymer, and textile industries (Ciurczak 1992, Barton 2004).  NIRS is a potentially rapid tool that can be used to get information about wood products in the industry. NIRS has been successfully applied in wood science and technology, as reported elsewhere (So et al. 2004, Tsuchikawa 2007, Leblon et al. 2013). It has been used to assess M 8  (Hoffmeyer and Pedersen 1995, Adedipe and Dawson-Andoh 2008), ρ (Thygesen 1994, Schimleck et al. 1999), wood chemical properties (Schimleck and Evans 2003, Jones et al. 2005a), and mechanical properties (Gindl et al. 2001; Kelley et al. 2004a; Kothiyal and Raturi 2011). It has also been used to estimate the properties of normal and modified wood as well as wood composites (Schimleck et al. 2002, Blanco and Villarroya 2002, Li et al. 2010).   As solid wood is an opaque material, EMR has difficulty passing through thick samples, so NIR spectra are collected via reflection (Hans 2014). In reflectance spectroscopy, there is an interaction between the radiation energy sent by the source and the material, producing scattering, transmission, and absorption (Leblon et al. 2013). The radiation energy reflected from the sample is then measured. Details about reflectance and absorbance spectroscopy with their equations have recently been reported by Hans (2014). Detailed reviews of the application of NIRS to wood and forest products have been presented by So et al. (2004) and Tsuchikawa (2007).  2.2.2.3. Effect of surface roughness on NIR spectra Wood is a porous material whose surface features are the result of machining conditions and the wood’s anatomy (de Moura Palermo et al. 2014, Zhang et al. 2015). Surface roughness is difficult to control, as it is an important index for wood product quality that is directly affected by wood machining and bonding performance (Ozdemir and Hiziroglub 2007). The surface quality of wood is important in developing NIR-based models for predicting wood properties, as previously reported (Hoffmeyer and Pederson 1995, Costa et al. 2018). However, NIR-region spectral absorbance is reported to be inversely proportional to wood surface roughness (Schimleck et al. 2005a).  Previous studies have shown the influence of surface roughness on the NIR spectra of wood. Some authors found the predicted results of NIR-based models for smooth surfaces to be better than those for rough surfaces (Schimleck et al. 2005b, Liu et al. 2006). Contrasting results have also been reported (Costa et al. 2018, dos Santos et al. 2020), where rough surface calibrations outperformed those from smooth surfaces. Better predictions from samples of greater roughness could be attributed to the capture of more information by the EMR during NIR spectral acquisition (Costa et al. 2018).  9  2.3. Western Hemlock and Douglas-fir  2.3.1. Distribution, properties, and uses of Western hemlock                                                                      n Western hemlock (Tsuga heterophylla) is the most abundant species along the coast of BC, and it also grows in the interior wet belt of the Rocky Mountains (Parish 1995). The standing volume of matured hemlock is about 1.3 billion m3 and represents about 18% of the total growing stock in BC (Shahverdi 2015). The best climate for its growth is mild or humid weather that is characterized by frequent fog and precipitation during the growth season in the coastal region (Pojar et al. 1991). It grows in pure stands or in mixture with other species, in even-aged stands where the climate is cool and moist.  Hemlock is a strong wood which is suitable for structural applications (Rohrbach 2008). It also has a straight and even grain with a medium to fine, coarse, and lustrous structure. The sapwood and heartwood are often difficult to distinguish (Nourian 2018). In green wood, the sapwood appears to be a little darker due to the large amount of free water in it (M is sometimes over 200%). Once the free water is evaporated when the wood is dry, there is no colour difference between the two areas, making it impossible to visually separate them. However, the growth rings are distinct and uniform in width, albeit narrow, delineated by a band of darker latewood with a purplish to reddish-brown tinge that is visible to the eyes. Approximately two-thirds or more of the growth ring is occupied with earlywood, and the transition from earlywood to latewood is gradual (Hoadley 1990).     Western hemlock  Douglas-fir  Figure 2.1: The distribution map of western hemlock and Douglas-fir in the Pacific Northwest (Natural Resources Canada, 2015).  10  Hemlock is also known for the presence of wetwood, which is a portion of the heartwood zone with higher M and ρ than adjacent parts. This zone is related to bacterial activity and is difficult to dry (Shahverdi 2015, Nourian 2018). The average ρ of hemlock is 470–490 kg/m3 (Rohrbach 2008). The literature also stated an average MOE value ranging from 10,000 to 12,300 MPa and an average MOR value ranging from 75 to 81.1 MPa (Coast Forest Products Association 2003).  The strength and the nailing characteristics of hemlock have made it a popular construction material in North America and overseas (Shahverdi 2015). It is commonly used for general construction, decking, plywood manufacturing, stock lamination, and the production of glue-laminated and solid beams. Other uses of hemlock include the manufacture of doors, windows, and floors, and in applications where a high-grade softwood is needed (Parish 1995). It is also exported to Japan for posts and beams in roof rafters and traditional housing (Lazarescu and Avramidis 2012) and is the most important pulpwood species in BC (Softwood Export Council 2004, Rohrbach 2008).   2.3.2. Distribution, properties, and uses of Douglas-fir                                                          Douglas-fir (Pseudotsuga menziesii) is one of the most popular softwood species in North America, originally named after a Scottish botanist (David Douglas) whom the Royal Horticultural Society in the late 1700s directed to study this tree (Hebda 1995). Two common native species of Pseudotsuga menziesii are P. menziesii var. menziesii, called coastal D-fir, and P. menziesii var. glauca, referred to as interior D-fir (Hermann 1982). The range of coastal D-fir extends from central BC (55N) south along the Pacific Coast ranges into central California to a latitude of 34.44N (Hermann and Lavender 1999), while the interior D-fir ranges from northern Mexico (19N) to the north of BC (55N) (Zustovic 2015).  D-fir is characterized by a straight or slightly wavy grain with a medium to coarse texture. The sapwood and heartwood of D-fir are not difficult to distinguish, unlike hemlock, as the sapwood is light in colour and the heartwood ranges from yellowish to reddish-brown with sharply defined bands (Leavengood 1998). Also, in contrast to hemlock, the transition from earlywood to latewood is abrupt (Osborne et al. 2016). D-fir wood is very hard and resistant to abrasion, making it very suitable for applications where wear is a factor, as in bridge parts and commercial buildings. An 11  average ρ of 450 kg/m3, and MOE and MOR values of 13,500 MPa and 88.6 MPa, have been reported for D-fir (Jozsa et al. 1998).  The higher strength of D-fir and the availability of large dimensions from old-growth trees make it a good choice of material for the construction industry. It has been identified as one of the finest timbers for heavy structural purposes, including glulam beams, laminated arches, and roof trusses (Douglas-fir Species Profile n.d). D-fir is used primarily in general construction and as dimension timber, plywood, and laminated veneer timber (Johnson and Gartner 2006, Ukrainetz et al. 2008).  2.4. Selected Wood Properties  2.4.1 Density The density of wood (ρ) is defined as the mass or weight per unit of volume (both at the same M), usually expressed in kilograms per cubic meter (kg/m3) or grams per cubic centimeter (g/cm3) (Haygreen and Bowyer 1996). Density is known to be influenced by the amount of water in the hygroscopic range, namely, 0 to 30% M (Antwi-Boasiako and Atta-Obeng 2009), which therefore allows for comparison of values only under the same M conditions (Kollman and Cote 1968).   Density is one of the most important properties of wood and is known to vary significantly within and between trees (Louzada 2003). This variation can be attributed to several factors, such as differences in cellular structure resulting from physiological and hereditary characteristics, and the presence of extraneous components (Rathgeber et al. 2006). For instance, the ρ of heartwood is usually higher than that of sapwood due to a higher concentration of extractives such as terpenes, resins, and polyphenols in heartwood (Hans 2014). In addition, past studies show that ρ variation in wood is usually affected by annual ring width (Fabris 2000) and varies from the earlywood to the latewood within rings (Gartner et al. 2002).  Density is considered important because it influences wood attributes such as shrinkage, swelling, and mechanical properties (Haygreen and Bowyer 1989), and because of its effect on the general quality of final products (Anjos et al. 2010). In many cases, ρ is particularly important to tree breeders and is usually evaluated when there is large within-species variation, a high degree of genetic control, and easy assessment (Santos et al. 2012).  12  2.4.2. Mechanical properties The mechanical properties of wood reflect its ability to resist applied forces that tend to deform it in any manner. Wood with greater resistance to deformation has a higher strength, and the ability of wood to resist applied load depends on the magnitude and direction of the force in relation to its fiber orientation. Wood’s mechanical properties therefore affect the quality of the products used in load-carrying applications (Haartveit and Flæte 2006).  Many methods have been identified for measuring the mechanical properties of wood, such as the modulus of rupture (MOR) and the modulus of elasticity (MOE). MOR is a measurement of the maximum stress that the material can endure before failure occurs and is measured in Newtons per square meter (N/mm²). MOE measures the deformation that results from a given level of stress before the point of total failure and is also measured in N/mm². Apart from their importance in structural lumber, MOR and MOE are also valuable traits in standing trees because they reduce mortalities that results from stem failures and the uprooting of trees during harsh weather conditions (Lachenbruch et al. 2011).  Destructive and NDE techniques are commonly used to measure the mechanical properties of wood. The destructive technique is based on the direct estimate of the MOE and MOR of wood using static bending tests, where a load is applied at mid-span to a piece of wood supported at its ends (ASTM 2005). The NDE tests are based on propagation speed using either acoustic or ultrasonic waves, or on acoustic spectral analysis in which mechanical properties can easily be estimated. In studying the mechanical properties of wood, it is important to know that those relating to bending strength include, among others, compressive strength, shear strength, and MOR, while the properties relating to wood elasticity include MOE, plasticity, resilience, and Poisson coefficient (Andrade et al. 2010).   2.4.3. Slope of grain (SOG) The SOG of wood refers to deviations other than those due to spiral grain or its related phenomena, such as interlocked and wavy grain, and is commonly referred to as cross-grain in sawn timber (Anderson et al. 1945). As wood is highly anisotropic, SOG exerts a strong influence on its material properties (Bodig and Jayne 1982). For instance, at a SOG of 45, Young’s modulus and axial strength are reduced to 10–25% of the value at 0 (Kollmann and Cote 1968), whereas 13  longitudinal shrinkage increases considerably. Therefore, SOG remains a considerable factor when lumber is graded, especially for load-bearing applications (Piazza and Riggio 2008).    SOG has been previously evaluated using thermal anisotropy (Naito et al. 2000), X-ray computed tomography (Sepulveda 2001), transmission measurements (Malik et al. 2005), and microwave technology (Schajer and Orhan 2006). In addition, past researchers have proposed a SOG indicator using a rotating capacitance-type transducer sensor (McDonald and Bendtsen 1986, Samson et al. 1993). While measuring the SOG of wood is often difficult (Sepúlveda 2001), its potential for predicting the mechanical properties of timber has been established (Olsson et al. 2013, Viguier et al. 2015).   2.4.4. Annual ring width (ARW)  Wood structure is mainly composed of tubular fibre cells that are cemented together and normally created by the way the trees grow. Every growing season, two layers are formed on the outside of the trunk, resulting in earlywood and latewood. Annual ring width (ARW), defined as the sum of earlywood and latewood for that season, is important because of its correlation to wood mechanical properties (Alteyrac et al. 2006). ARW is also one of the most important factors that directly affect wood’s physical and mechanical properties (Nathalie and Malo 2010, Ibrahim and Ayşenur 2016) and its durability (Takata et al. 2002).   Past researchers have reported the influence of ARW on the properties of wood. For instance, Karlman et al. (2005) showed in their work on different species of larch that an ARW up to 2.5 mm had the highest ρ, while there was a marked reduction in ρ where the ARW was wider than 3 mm. Likewise, Ethington et al. (1996) incorporated variations in ARW and observed that the ARW has a significant effect on the compressive strength perpendicular to the grain of the wood. Lastly, statistical analyses performed on five different species showed the effect of SOG and ARW on the shear strength and MOE of the species (Lang et al. 2000, 2002).   2.5. NIRS Wood Assessment   2.5.1. Density                                                                                                                                 T The first known study that reported the correlation between NIR spectra and ρ was presented by Thygesen (1994), where solid wood of Norway spruce was used to predict the ρ of wood. Several 14  studies have since reported that NIRS can be used to estimate the ρ of wood (Hoffmeyer and Pedersen 1995, Viana et al. 2009, Hein et al. 2009a).  A number of studies have reported good calibration of ρ using NIRS. Alves et al. (2012) reported PLS-R models for ρ based on X-ray micro-density data for each species of Pinus pinaster and Larix × eurolepis. The common model provided a residual prediction deviation (RPD) of 3.1, and the single models for Pinus pinaster and Larix × eurolepis provided RPDs of 3.5 and 3.2, respectively. Also, Fujimoto et al. (2012) used NIRS to estimate the ρ of wood independently of its M regardless the fact that the increase in M is reported to greatly affect the variation of NIR spectra (Fujimoto et al. 2012, Via et al. 2003). In the same way, Inagaki et al. (2010) developed a PLS-R model for the prediction of the air-dried ρ of Eucalyptus camaldulensis with an RPD of 3.8. In their experiment, they proved that the model is quite robust and stable by removing 40% of the samples during the cross-validation step and ended up with an RPD of 3.2. The use of NIRS and multivariate analysis to predict ρ has been fully established (Hoffmeyer and Pederson 1995, Via et al. 2003).   2.5.2. Mechanical properties Several reports have previously described the use of NIRS to estimate the mechanical properties of wood (Schimleck et al. 1999, Fujimoto et al. 2008, Hein 2010). For instance, NIRS was used to predict the MOR of Eucalyptus grandis and Eucalyptus urophylla (Thumm and Meder 2001). In the same way, NIRS was also utilized to predict the strength and elasticity in compression parallel to the grain of wood with good correlation coefficients of 0.78 and 0.75 (Hein et al. 2009b). Past researchers have also modeled the MOE and MOR of Pinus palustris of different wood types—juvenile, mature, and pith (Via et al. 2003). The poor performance of the models was linked to the narrow range of values for the species utilized in the study. Similarly, others (Rials et al. 2002, Adedipe and Dawson-Andoh 2008) have applied NIRS to estimate the MOE and MOR of solid wood and wood composites. Hence, this technique has been established as a very effective, reliable, and practical method of estimating the mechanical properties of softwoods (Kelley et al. 2004a).    15  2.6. Chemometrics and Multivariate Calibration  Chemometrics has been fully explored for the extraction of physical, chemical, and mechanical information from spectral data by employing multivariate data analysis techniques in the estimation of wood properties. Multivariate analysis methods such as PCA, PLS-R, principal component regression (PCR), and soft independent modelling of class analogy (SIMCA) are useful in the extraction of relevant information from spectroscopic measurements. This is possible over a broad wavelength range, allowing all data to be used in the analysis (Leblon et al. 2013). The advantage of using multivariate analysis techniques is that the wavelength variables tend to be collinear, and the information extracted or mined is in the correlation patterns instead of individual data points (Eriksson et al. 2001). Most of the published studies and reviews on the application of NIRS in the forest products industry have applied chemometrics to predict several properties of wood and wood products from NIR spectra (So et al. 2004, Tsuchikawa 2007).   Several regression methods have been used to quantitatively derive information from reflectance or absorbance spectra. Some of the methods, such as MLR, support vector machine (SVM), PCR, and PLS-R, have been reported on (Steven et al. 2010, Atzberger et al. 2010, Vohland and Emmerling 2011). A detailed comparison of many of these methods has been attempted in the past (Viscarra Rossel and Behrens 2010). The PLS-R method has become one of the most popular for chemometrics in recent years (Wold et al. 2001) and was the method utilized in this study.  There are many kinds of multivariate data modelling, such as PCA modelling, which models only the X matrix. In PCA, a principal component model of the essential structure of the data is built to gain an overview of the data structure. Multivariate calibration is concerned with the X and Y matrices, with X being the independent variable (spectra) and Y being the dependent variable (property of interest). The multivariate model for (X, Y) is a logistic relationship between the empirical X and Y relations, and establishing and calibrating such a model is the first stage in multivariate modelling (Esbensen et al. 2002). The second stage is prediction.   In multivariate modeling, a set of known X and Y values are first used to develop the model, which is then used to predict the Y values from new X measurements (spectra which is also the predictor variable). The interesting point here is that the model makes it possible to use only X measurements (spectra) in future predictions instead of taking more Y measurements (ρ, MOE, MOR, among 16  others) that are very expensive, time-consuming, laborious, dangerous, and in many cases destructive. These characteristics of the Y values make it desirable to replace them with X measurements, such as spectra, that are simple to use, faster than other methods, and cheaper to collect. NIRS could therefore be used to replace slow and cumbersome experiments such as physical, mechanical, and wet chemistry measurements.  The calibration data must meet several requirements, such as being representative of the future population from which the new X-measurements are to be sampled, and the measuring conditions should also be broadly similar. However, they should not be too similar, because the only difference will then be the sampling variance, that is, the variance due to two independent samplings from the target population. The determination of the predictive ability of a PLS-R model must be accompanied by the separation of the data sets of the property (y variable) being studied in the training or calibration set used in developing the model. An external or independent test set based on the calibration set is then used to validate the model and determine its predictive ability (Hauksson et al. 2001).  2.6.1. Effect of spectral pre-treatment on chemometrics                                                                  The The use of spectral pre-treatments in chemometrics has been widely studied. For instance, derivatives were applied to absorption spectra to gain enhanced spectral information for the spectral ranges during a decomposition process (Ben-Dor et al. 1997). In the same way, min-max normalization, first derivative, and vector normalization after mean centering have been used to improve spectral quality (Udelhoven et al. 2003). Vasques et al. (2008) also applied various pre-treatments such as derivatives, normalization, and non-linear transformations on 554 soil sample spectra from Florida to estimate their organic carbon content. The effect of log 1/R transformation, first derivative, and SNV-DT on soil diffuse reflectance spectra (Stenberg and Viscarra 2010), untransformed spectra, and first and second derivative with gaps of one to 64 bands have also been analyzed to estimate several variables for airborne hyperspectral data (Hively et al. 2011).   A detailed review of the most common pre-treatments for near-infrared spectra in chemometrics was published in the last decade (Rinnan et al. 2009). Most of the work done on NIR analysis of wood and wood products has used simple spectral pre-treatments such as 1st or 2nd derivatives, or 17  multiplicative scattering correction. Other analytical techniques have been used to improve the quality of the models constructed using NIR spectral data (Axrup et al. 2000).  2.6.2. Effect of grain orientation and surface roughness on NIRS calibrations One of the factors that influences NIRS calibrations is wood anisotropy (Hans 2014). There are three sections in a piece of wood: transverse, radial (quarter-sawn, Qs), and tangential (flat-sawn, Fs). The effect of these sections on the NIR spectra of wood has been reported (Leblon et al. 2013). The sections of wood to be used for calibration must be carefully selected to suit the application so that the section from which spectra are collected contributes to the performance of NIR-based models. For instance, calibrations developed based on transverse sections to estimate the M of red oak lumber were better than those developed from Qs and Fs (Defo et al. 2007).   The transverse section has been reported to result in better predictions, which is attributable to the distinct anatomical differences within the growth rings. The differences between the measurements for the various sections could also be explained by the way scanning is carried out. With a transverse section, incident radiation directly interferes with free water in the cell lumen, which is not the case when scanning other sections, since the radiation first has contact with the cell wall (Tsuchikawa et al. 1996). The models based on Fs sections are known to have the poorest performance. This is because of the presence of only earlywood or latewood portions when scanning is done which thus affects the calibrations (Hans 2014).   An additional factor influencing NIRS calibration is surface roughness. Roughness of wood on transverse sections is usually higher than on Qs and Fs sections (Tsuchikawa et al. 1996). Also, transversal sections have tracheid cells with a longitudinal axis that is parallel to the direction of the NIR incident radiation, which results in greater absorbance compared to Qs and Fs surfaces (Fujimoto et al. 2008). Absorbance is higher for Qs than Fs samples because the former coincides with the longitudinal orientation of the medullary rays that form the tissue radiating between pith and bark (Leblon et al. 2013). Generally, an increase in surface roughness decreases the absorbance of NIR radiation (dos Santos et al. 2020).   18  2.6.3. Principal component analysis (PCA) PCA is a multivariate analysis technique that is often applied to spectroscopy for data reduction and clustering visualization. It decomposes a linear combination of original variables into a few principal components (PCs) (Yu et al. 2016). The PCs show the structure of the dataset and provide information on the main spectral components causing differences between samples (Sandak et al. 2016). PCA and PCR are common algorithms used for the calibration of spectrometers and the evaluation of unknown measurement spectra (Vogt and Tacke 2002). The calculation of PCs is often accompanied by singular value decomposition (SVD) of the data matrix (Press et al. 1992). The k calibration spectra contain n measurements at a time. In addition, the rows of the matrix VT build up the preferred orthonormal basis of the vector subspace spanned by the calibration spectra (Vogt and Tacke 2002). This is represented mathematically as  M (K X n) = U (K X K). S (K X K). VT (K X n)                                                                                           (1)  SVD and PCA are common techniques in the analysis of multivariate data (Wall et al. 2003) and were utilized in this study. A large dataset is widely used in many disciplines and applications, and methods are often required to drastically reduce the dimensionality of the data in an interpretable way to get information from such datasets. While many techniques have been identified for this purpose, PCA is the oldest and the most widely used (Jolliffe and Cadima 2016).   PCA is based on the principle of reducing the dimensionality of the dataset while preserving as much variability as possible (Jolliffe and Cadima 2016). The variability is preserved by finding new variables that are linear functions of those in the original dataset. The earliest literature on PCA dates from Pearson (1901), but not until computers became available decades later was it computationally feasible to use PCA on large datasets. It can be based on either the covariance matrix or the correlation matrix, and choosing between these analyses has been previously described (Jolliffe and Cadima 2016). In either case, the new variables (the PCs) depend on the dataset rather than being pre-defined basis functions and are therefore adaptive in the broad sense. The main uses of PCA are descriptive rather than inferential.  19  2.6.4. Partial least squares regression (PLS-R) PLS-R models have been widely used to extract useful information from spectroscopic data. A PLS-R model relates the spectral information to quantitative information about the measured samples (Swierenga et al. 1998). PLS-R has proven to be a popular and effective approach to solving several problems in chemometrics. The algorithms are resistant to overfitting and are easy to implement and tune, and their ability to produce inferences in high-dimensionality conditions makes them ideal for a kernel approach (Bennett and Embrechts 2003).   A PLS-R model is better at finding the relationship between input variables and output variables; a correlation model is then built using this procedure to predict the output variables by utilizing the input variables (Wang et al. 2015). Although the accuracy of PLS-R may not be significantly higher than that of PCR, it is considered better because there are fewer PCs in the final calibration (Naes et al. 1986). PLS-R is also preferred for its faster algorithms, higher precision, and more harmonious calibration models (Kalivas and Gemperline 2006).    While PLS-R is reliable in linear conditions, most industrial processes have explicit non-linear qualities that can be hard to ignore (Ding et al. 2013). Hence, a new method that can be utilized in highly non-linear processes, called the Kernel PLS (KPLS), has been proposed (Rosipal and Trejo 2001). The KPLS algorithm was introduced to speed up calibrations, handle very wide and tall data, and handle multivariate Y according to the nonlinear iterative PLS (NIPALS) standard, as reported by Andersson (2009). KPLS has also been shown to be numerically stable and one of the fastest algorithms for PLS (Alin 2009, Anderson 2009).  2.6.5. Validation of chemometrics model NIR calibration models are often used for the correlation of either raw or pre-processed spectra with one or more physical-chemical properties of a set of samples. As complicated as it sounds, numerous well-developed calibration techniques have proven to work with most NIR applications, and these techniques are usually included in chemometric software packages (Agelet and Hurburgh 2010).   An adequate validation of calibration models is a crucial step to determine the suitability of models to predict new samples; this is the main reason for developing NIR calibrations. There are different 20  approaches to estimate the suitable number of PCs to be included in the calibration. One of the most popular is cross-validation (CV), a technique used in the assessment of how the results of a statistical analysis will generalize to an independent data set. The goal of CV is to test the model’s ability to predict new data that were not included in the calibration to avoid overfitting or selection bias (Cawley and Talbot 2010). It is often used for the basic assessment of calibration performance. A description of this method can be found elsewhere (Seni and Elder 2010).  Another approach used in the validation of chemometric models is called test-set validation, which is considered the ultimate way of testing calibration performance. Here, two-thirds of the samples are regarded as the training set and one-third as the validation set; only the training set is used in the calibration, while the validation set is used for the prediction of new samples based on the PLS-R models tuned with the calibration dataset (Sandak et al. 2016). This method (test-set) is regarded as the most conservative validation for testing models on a representative independent test-set, especially when the purpose is to establish calibration models that can predict quantities (Westad and Marini 2015). The quality of the predictive model can be measured in several ways, some of which have been described (Martens and Naes 1991, Schimleck et al. 2001).  2.7. Objective                                                                                                                                                  T This study aims to examine the ability of NIRS to predict specific properties (density, grain angle, annual ring width, and mechanical properties) of Western hemlock and D-fir. The study attempts to develop robust models for within- and between-species variation based on PLS-R to evaluate the predictability based on orientation—that is, quarter-sawn, flat-sawn, and surface roughness (rough and smooth). The results are expected to provide the wood industry with information on using NIRS for swift and accurate analysis of wood properties and on the influence of surface conditions when assessing wood properties.    21  3.0  MATERIALS AND METHODS  3.1. Sample Preparation  Kiln-dried lumbers of Western hemlock and coastal D-fir were obtained from Interfor-Acorn Sawmill, Surrey, BC and Western Forest Products Sawmill, Ladysmith, BC, respectively. Sixty quarter-sawn (QS) and sixty flat-sawn (Fs) samples 1220 mm in length and 50 x 100 mm in cross-section were selected for each wood species. These samples were further sawn into dimensions of 25 mm × 90 mm × 410 mm (longitudinal [L]) by a sliding table saw (T75, MARTIN Woodworking Machines Corp., USA) and conditioned at a temperature and relative humidity of 20C and 65% H. A schematic of the cutting pattern of samples is shown in Figure 3.1.   Figure 3.1: Schematic of the cutting pattern of samples.  One of the longitudinal-tangential surfaces of the QS samples and one of the longitudinal-radial surfaces of the FS samples were then planed in order to have two types of surface conditions: one rough and the other smooth (Figure 3.2). This was done to examine the effect of surface condition on NIR spectra. NIRS measurements were done on the rough and smooth surfaces of all Qs and Fs samples. Thus, for each type of surface condition, a total of 240 samples were analysed. Two different species and two orientations were considered in this study, resulting in four combinations (two wood species x two orientations), for a total of 240 samples of mixed wood type (60 for each combination).   22   Figure 3.2: (a) Hemlock rough, (b) D-fir rough, (c) hemlock smooth, and (d) D-fir smooth.  Each NIRS sample, with dimensions of 25 x 90 x 410 mm (thickness, width, and length, respectively) was converted into two strips (samples) of 25 mm x 25 mm x 410 mm (length) for destructive MOE and MOR measurements (Figure 3.3) using an MTS dynamic material testing machine (model 810, MTS Systems Corp, Minnesota, USA), while 25 mm x 25 mm x 25 mm (length) samples were prepared for density determination (Figure 3.4). A total of 480 samples were prepared.  Figure 3.3: Cutting pattern of samples for MOE and MOR measurements. 23    Figure 3.4: Cutting pattern of samples for ρ measurements.   Table 3.1: Sample categorization for NIR analysis, density, and mechanical properties. N.B.: Only 120 samples were prepared for each wood species for the NIR analysis because one surface was rough and the other was smooth.   3.2. NIRS Measurements  Wood sample NIR spectra were acquired using an NIR spectrometer system, which consisted of a light source, an optic fiber, a fiber spectrometer (ASD Lab Spec® Pro, Analytical Spectral Devices Inc. Boulder, CO, USA), a sample holder, and a computer.   Species Orientation NIR Analysis ρ, MOE and MOR SOG and ARW Hemlock Quarter-sawn (Rough & Smooth) 60 120 60  Flat-sawn (Rough & Smooth) 60 120 60 Douglas-fir Quarter-sawn (Rough & Smooth) 60 120 60  Flat-sawn (Rough & Smooth) 60 120 60 Total  240 480 240 24   The optic fiber connected to the spectrometer was oriented at 30º above the sample surface at a distance of 130 mm and aligned in parallel with the longitudinal axis of the sample. The samples were illuminated with a tungsten halogen bulb (ASD Pro Lamp) oriented perpendicular to the sample surface. The distance between the sample surface and the bulb was 168 mm, resulting in an NIR spot size of approximately 80 mm (Figure 3.5). For each sample, four spectra were collected with four subsets that consisted of 50 scans (to reduce the noise), which were averaged into a single spectrum. The averaged spectra for the samples were used to predict the wood properties.  145152°25°1.Light source  2.Optic fiber  3.Spectrometer  4.Computer  5.Specimen holder.  6.Specimen155mm236 Figure 3.5: Schematic diagram of the NIRS system (adapted from Zhou et al. 2019).  All samples used in this study were scanned in the full wavelength range of 350–2500 nm at intervals of 1 nm to cover the visible spectroscopy (Vis) and NIR spectroscopy ranges. The effects of reduced spectral ranges (900–1900 nm, 1100–2500 nm, 1300–2300 nm) on spectral information were also assessed. A piece of commercial microporous Teflon was chosen as the reference material, and a background reference scan was taken every 30 minutes to correct for potential drift over time.   30 80mm 168mm 25  The reference spectra were measured and stored before the test sample spectra collection. Two readings were taken (top and bottom face) at quarter points per sample to give four spectra (two rough and two smooth spectra per sample). The two spectra per surface were averaged to give a single spectrum, and a total of 960 spectra (4 x 240 samples) were acquired for the experiment. All NIRS measurements were taken in a lab location where the temperature ranged between 20 - 250C and relative humidity between 40 - 55%. Effort was also made for all NIRS measurements to be completed in a short time span to avoid M fluctuations of the samples. The set-up of the NIR system used for spectral acquisition is shown in Figure 3.6.  Figure 3.6: The NIR system setup used for spectral acquisition  (1. light source, 2. optic fiber, 3. spectrometer, 4. specimen).  3.3. Determination of Wood Density (ρ)  The wood density (ρM) is defined as the ratio between the mass and the volume of the sample at a given M (Williamson and Wiemann 2010). The samples used for ρ were conditioned to a target  equilibrium moisture content (Memc) of 13.7% and 13.3% (for hemlock and D-fir) in a walk-in chamber (Constant Temperature Control Limited, Aurora, ON, Canada) at a temperature and relative humidity of 20C and 65% H, respectively. The difference in the Memc of the species 26  could be attributed to the wood chemistry and specimen moisture history which thus affects the sorption properties of wood (Glass et al. 2014). The initial volumes and weight of samples were measured using a digital caliper and digital balance, respectively.   The ρ of the samples was calculated by  ρM  = WM/VM                                                                                                                                                                                                                   (2)                                                                                                                                                                                                                                                                                                                                             where, ρM  (kg/m3) is the density, W (kg) is the sample weight, and V (m3) is the volume of the sample at the respective M for the species.  3.4. Determination of MOE and MOR  The mechanical tests were done via a 3-point static bending machine loaded with a 10 kN load cell, a span of 360 mm, and loading rate of 1.3 mm/min according to ASTM D143-94. The samples were conditioned at a temperature of 20C and 65% H in a walk-in chamber (Constant Temperature Control Limited, Aurora, Ontario, Canada), and MOE and MOR were determined using an MTS dynamic material testing machine (model 810, MTS Systems Corp, Minnesota, USA) (Figure 3.7). A total of 480 strips were used for the mechanical tests, and spectral information for the mechanical properties of wood was also collected on the samples for NIR analysis.  27   Figure 3.7: Material Test System (model 810) for measuring MOE and MOR.  The MOE and MOR of the samples were calculated by  𝑀𝑂𝑅 (N/mm2) =3𝑃𝐿2𝑏ℎ²                                                                                                                               (3) 𝑀𝑂𝐸 (N/mm2) =𝑃𝐿³ 4∆𝒃𝒉³                                                                                                                             (4) where P (N) is the maximum load (MOR), L (mm) is the span length, P (N) is the applied load at elastic limit (MOE), ∆ (mm) is the deflection at elastic limit, b (mm) is the sample width, and h (mm) is the sample depth or thickness.  3.5. Annual Ring Width Measurements  ARW is important in the manufacturing of wood products because of its correlation to the strength properties of wood (Alteyrac et al. 2006). The samples used for the NIR test were conditioned at 20C and 65% H for the ARW measurements, and the visual inspection of the ARW were carried out on them.   28  The rings were counted over a length of 90 mm and reported as number of rings per distance (mm).  A total of 240 samples were assessed, and all measurements were done visually with a ruler. The ARW in mm was calculated by   𝐴𝑅𝑊 (𝑚𝑚) = ∑𝑊𝑖𝑛𝑛𝑖=1                                                                                                                           (5)                                                                                                                                                where Wi (mm) is the ith annual ring width and n is the number of rings on the Qs and Fs samples used for the NIRS test.  3.6. Slope of Grain Measurements   The SOG is defined as the angle between the wood grain (i.e., the direction of the wood fibres) and the main axis of lumber (Gindl and Teischinger 2002). The SOG was measured with a shop-built scratch gauge with a pivoting angle for inscribing a light groove parallel to the grain (Figure 3.8 left), and a MITE-R-GAGE adjustable Lexan protractor (Figure 3.8 right) was used to measure the SOG with respect to specimen edge, measurable to 0.5. SOG measurement via this method is called apparent or surface rather than machine-measured internal SOG.  The samples used for the NIR test also had their SOG measured only on the smooth surface of the samples rather than the rough surface to increase the accuracy of the SOG measurements. For each specimen, measurements were carried out at three points: the front quarter end, the middle point and the rear quarter end. The groove made had a minimum length of 300 mm so that the average SOG could be measured. A total of 240 samples were tested.  29   Figure 3.8: Measuring the SOG of wood (left: shop-built scratch gauge; and right:  MITE-R-GAGE adjustable Lexan protractor).  3.7. Multivariate Data Analysis  Once all spectra were collected, multivariate calibrations were developed for each wood property by PLS-R, and the analysis was performed using Unscrambler® version 10.5.1 (CAMO, Norway) software to find possible correlations between the X components (spectra) and the Y components (response variables). Also, the NIR data were mean-centered and normalised before the PLS-R analysis. As a rule, in multivariate modelling, the first stage is to subtract the average from each variable, which is referred to as mean centering. This ensures that all results will be interpretable in terms of variation around the mean.  The spectral data were randomly split into calibration and prediction sets consisting of 36 and 24 samples in each combination of species and surface conditions. Samples of each combination were then divided into two data sets according to the sample set partitioning, based on the joint x-y distance (SPXY) algorithm (Galvão et al. 2005). Three-fifths (36 samples in each combination) was used as the calibration set and the remaining two-fifths (24 samples) as the prediction set. Thereafter, the samples in the calibration and prediction sets of each combination were merged into the final calibration (144 samples) and prediction (96 samples) sets, respectively. Sample calibrations were carried out using a mixed set with 960 datapoints (4 datapoints x 240 samples), which included both types of cross-sections, and another set with 480 datapoints (4 datapoints x 30  120 samples), which consisted of samples from Qs and Fs sections, separately. The methods used for the pre-treatments of spectra before model development included Savitzky-Golay (SG) smoothing, standard normal variate (SNV), and first- and second-order derivatives. NIR reflectance data are usually transformed into log of inverse reflectance (log 1/R) (absorbance); this is often done to ensure good linear modelling.  The development of the models was based on the spectrum of wavelengths 800–2500 nm (full NIR range). Reduced wavelength regions were also used to build models to check the influence of defined spectral ranges on model quality. Reducing the range allowed for the potential use of lightweight, portable, and inexpensive spectrometers for predicting wood properties. The best models were selected based on the coefficient of determination of calibration (Rc2) and the coefficient of determination of prediction (Rp2), which reflect the ability of the model to predict new samples, and RPDp, which uses the standard deviation of the reference data used in the prediction and the standard error of prediction (SEP).   Statistical analysis provided correlations between the wood properties and the spectra for the full and reduced spectral ranges. For each of the wood properties, several data sets were created, including the raw spectral data and the first- and second-derivative data. The regression coefficients plots were also used to check the wavelength contributing most to the models, models were developed based on those variables, and comparisons were made with the full spectrum. Outliers were detected using the influence plot generated by PCA modelling, and they were removed before re-modelling to improve the predictive ability of the final model.  31  4.0 RESULTS AND DISCUSSION  4.1 NIRS Calibration  Wood property calibrations were performed using the test-set validation method, regarded as the best way to validate a model. It is based on testing the model on a subset of the available samples that were not included in the computation of the model components (Sandak et al. 2016). During the NIR calibration stage, the models were developed based on the training set, and some samples (40%) of the dataset were randomly selected to represent the entire population (Galvao et al. 2005). These samples, which are sometimes called the validation or test set, were then used to determine the ability of the model to predict new samples (Hauksson et al. 2001).   Validation of the chemometric model is important for determining modelling and prediction errors, outliers, an optimal number of PLS components, and trends (Tounis 2009). The number of factors or PCs used in the model development was selected by observing the response of the residual Y-variance with additional factors. The number of factors used in the final calibration is crucial to avoid overfitting, a phenomenon where too many factors are included and a fraction of noise is modelled (Agelet and Hurburgh 2010), and underfitting, when too few factors are included with higher bias and lower variance (Gowen et al. 2011). The optimal number of PCs used in this study was suggested by the Unscrambler® software.   The best calibration model was selected using high R2, reduced error, and rank (number of PCs). All PLS-R models used in this study were based on the KPLS algorithm. The KPLS is a sample-centric approach where the relationship between every sample is characterized by a kernel function that maps the data into higher dimensional space, where the linear regression is performed (Mora and Schimleck 2010). A full description of its derivation can be found elsewhere (Rosipal and Trejo 2001).  The pre-treatment of spectra is also an important step in NIRS calibration, done to improve the model quality. SG smoothing using a 25-point filter width and a second-order polynomial was used during model development (Savitzky and Golay 1964, Gaspar et al. 2009). A common pre-treatment known as the derivatives—such as first derivative, which removes any offset difference, and the second derivative, which removes any slope effect in the data—were also applied to 32  improve spectra quality (Rinnan 2014). The derivatives used in this study were calculated by applying the SG algorithm using 17 smoothing points and a second-order polynomial (Schwanninger et al. 2004).   4.2 Density and Mechanical Properties  4.2.1 Density based on calibrations for all samples                                                                    T The ρ ranged from 377–725 kg/m3 and 418–794 kg/m3, with averages of 491.97 kg/m3 and 552.89 kg/m3 for hemlock and D-fir, respectively. Others have found similar results for hemlock (Gonzalez 1990, Nourian 2018) and D-fir (Jozsa and Kellogg 1986, Gonzalez 1990). As expected, higher ρ values were observed for D-fir than for hemlock (Softwood Species Profile n.d). The variations in the ρ of the wood could be linked to the differences in the amounts of cell wall and extraneous materials present per unit of volume (Henze 2006). For ρ modelling, 144 and 96 samples of both species were used for the calibration and prediction sets. The descriptive statistics for this property from the calibration and prediction sets are shown in Table 4.1.   Table 4.1: Descriptive statistics of the wood properties from the calibration and prediction sets.  Min is the minimum value, max is the maximum value, SD is the standard deviation, and CV is the coefficient of variation.   Calibration Set (n = 144) Prediction Set (n = 96) Property Mean Min Max SD CV Mean Min Max SD CV ρ (kg/m3) 507 377 725 60 11.83 546 389 794 72.21 13.23 MOE (MPa) 5608 1587 13247 2095 37.36 6199 1601 11792 2270 36.63 MOR (MPa) 63 18 110 18 28.57 70 33 114 17.33 24.76 SOG () 8.9 1.5 16.7 3.2 35.96 9.2 1.3 16 3.4 36.96 ARW (mm) 1.8 0.1 12.6 1.8 100 1.7 0.2 13.1 1.6 94.12 33  All ρ models gave moderate predictions, with the Rc2 ranging from 0.46–0.60 and the Rp2 ranging from 0.51–0.63 (Table 4.2). The effect of pre-treatment on model quality was observed, especially when SG smoothing with a 25-point filter and a second-order polynomial were used (Savitzky and Golay 1964). The Rp2 improved when SG smoothing was applied to the spectral data (Table 4.2), which is in line with others’ work (Gaspar et al. 2009) where pre-treatments improved the model performance. A typical graphical representation of a raw spectrum is shown in Figure 4.1.   Figure 4.1: A typical graphical representation of a raw spectrum.  The ρ models in this study were considered robust with few factors (no more than five). This is interesting because past authors (Sandak et al. 2016, Schimleck et al. 2018) reported that fewer factors in a PLS-R model is a better strategy for avoiding data overfitting. Although standard normal variate (SNV) as well as first and second derivatives were applied to the raw data, no improvement in the predictive ability of the model was shown. This was not surprising, as the best correlations were previously obtained from raw spectra (Andrade et al. 2010). Overall, results based on rough surfaces were better than those from smooth surfaces. This could be explained by the fact that EMR on a rough surface captures more information than on a smooth surface during scanning, which contributes to the predictions (Costa et al. 2018).  34  The PCA was used to distinguish between the grain orientations of the two softwoods (hemlock and D-fir) and to reveal any unusual variation in the PCA model. The PCA scores plots (Figure 4.2) show the four classes of samples earlier described with the first two components, PC1 and PC2 (the highest model contributors), which explained 96% of the variation in the raw data. It can be seen that some samples are clustered together, meaning they are similar and highly correlated with respect to the PCs, while those far apart from one another are negatively correlated (Figure 4.2).    Figure 4.2: PCA plot for raw spectra obtained from a rough surface (top) and smooth surface (bottom), where DF: Douglas-fir flat-sawn, DQ: Douglas-fir quarter-sawn, HF: Hemlock flat-sawn, HQ: Hemlock quarter-sawn.  35  For instance, PC1 describes more of the variation in the Fs samples, while PC2 shows the variation in the Qs samples for the rough surface (Figure 4.2, top). In the same way, for the smooth surface, both PC1 and PC2 fully describe the variation in grain orientations for the two species (Figure 4.2, bottom). A few samples are left out of the eclipse and are considered to have contributed little to the PCA model (Figure 4.2).   This showed that NIRS coupled with PCA analysis is a useful tool for identifying and differentiating between the Qs and Fs samples of the two softwoods, which agrees with previous findings (Everard et al. 2012). The plots showing the relationship between NIR predicted and measured values for the best models are shown in Figure 4.3 for both surfaces. The closeness of the data points to the target line (black line) is often used to determine the model’s goodness of fit. These plots show the strength of correlation between the calibration data (blue dots) and validation data (red dots) for the property measured.  The calibrations in this study also involved selecting the optimal spectral range for the prediction of ρ. This is considered important to remove the spectral regions that make little contribution to the predictive ability of the models. It has been reported that not all spectral regions contribute to the predictive ability of models, hence the need for a narrow spectral range to optimize the model (Leardi et al. 2002, Mehmood et al. 2012). Although spectral ranges of 1100–2500 nm and 1300–2300 nm were used to develop calibrations, no model performance improvement was observed.  The Rp2 of ρ for the rough (0.57–0.63) and the smooth (0.51) surfaces were obtained. Although these values are lower than what has been reported for the ρ of Pinus taeda (R2p of 0.69; Schimleck et al. 2002), they are higher than the ρ predictions found in another study (Jones et al. 2007) for the same species. In this study, the rough surface predictions were considered best for ρ due to the high Rp2 values and reduced prediction errors. This meant that better predictions could be obtained for the ρ of hemlock and D-fir by taking the NIRS readings on the rough surface of lumber. The summary of the calibration statistics for ρ of the softwoods is shown in Table 4.2.  The difference in the calibration statistics for these studies could be attributed to the range of ρ observed and the species used. It has also been shown that mixed species calibrations may not be as accurate as those based on single species, such as in the aforementioned study (Schimleck et al. 36  Table 4.2: Calibration statistics for ρ based on all samples (n = 240)  and Qs and Fs samples (n = 120).  2010). Likewise, the poor performance of the ρ models could be attributed to the variation in ρ for different sections of wood along the stem, as compared to mechanical properties (Acquah et al. 2018). Again, the generally poor predictive statistics of ρ could be ascribed to the diverse genetic makeup of the samples in the prediction set. Previous researchers have also revealed that more precision and accuracy are needed in testing procedures (spectral acquisition and reference methods) to improve the calibration statistics from NIRS (Hein 2010).  A direct comparison of the prediction regression coefficients generated from the PLS-R model was also performed. Comparisons were made between the regression coefficients of ρ predicted for both surface conditions (rough and smooth). Some of the absorption bands related to the major wood components (cellulose, hemicelluloses, and lignin) were found to have contributed to ρ predictions, which is in line with the work of other researchers (Fujimoto et al. 2008). Similarly, Property Roughness Spectral Range Rc2 Rp2 SEC SEP RPD ρ  (All) Rough Raw Data (Full Range)              Smoothed data 0.46  0.44 0.57  0.63 44.42  44.95 48.97  45.74 1.5  1.6  Smooth Full range 0.60 0.51 38.55 54.79 1.3 ρ  (Qs) Rough Raw Data (Full Range)  Smoothed (1100–2500) 0.52  0.52 0.53  0.57 45.58  45.49 40.59  41.61 1.5  1.4  Smooth Raw Data (Full Range)  Raw Data (1100–2500) 0.64  0.74 0.56  0.63 39.36  33.56 43.70  40.23 1.4  1.5 ρ  (Fs) Rough Raw Data (Full Range)  Raw Data (1100–2500) 0.32  0.60 0.63  0.66 42.53  32.62 52.67  52.28 1.1  1.1  Smooth Raw Data (Full Range)  Raw Data (1300–2300) 0.30  0.47 0.59  0.59 43.23  37.76 54.12  54.14 1.5  1.5 37  the cellulose- and lignin-associated wavelengths were correlated with ρ, suggesting that wood chemistry influences ρ independent of baseline shift (Hein 2010).     Figure 4.3: The relationship between measured and NIR-predicted values of ρ (kg/m3), based on rough (top) and smooth (bottom) surfaces for two softwoods (hemlock and D-fir). The blue and red dots indicate the calibration and prediction sets, respectively.  The results of the regression coefficients showed that similar wavelengths contributed to the ρ prediction of both surfaces. The important wavelengths found to have contributed to ρ included: 38  1480, 1484–1493, 1534, 1550, 1540, 1579, 1580–1597, 1793, 2110 (due to crystalline and semi-crystalline cellulose; Schwanninger et al. 2011), 1471, and 1724 (due to the presence of hemicelluloses; Fujimoto et al. 2007, Fackler and Schwanninger 2010), and 1447, 1448, and 1672–1674 (due to the presence of lignin; Mitsui et al. 2008). The wavelengths contributing to the predictions were not surprising, as past authors have also found the spectral range of 1100–1700 nm to contain the most important information for predicting wood properties (Schimleck et al. 2003, Todorovic et al. 2015). The specific assignments of NIR spectral features to different wood components have been reported by several authors (Bokobza 2002, Tsuchikawa and Siesler 2003).  4.2.2. Density based on calibrations for the Qs and Fs samples For the modelling of ρ, 72 and 48 samples of both species were used for the calibration and prediction sets, respectively. Most of the models for ρ gave moderate predictions, with Rc2 ranging from 0.52–0.74 and 0.32–0.60 for the Qs and Fs sections, respectively. The Rp2 ranged from 0.53–0.63 and 0.59–0.66 for the former and latter, respectively. A comprehensive summary of the model statistics is provided in Table 4.2.   For the rough surface, the Qs section yielded better predictions (Rc2 of 0.52 and Rp2 of 0.53) compared to the Fs section (Table 4.2). The rough surface prediction was considered best due to the SEP (40.6 kg/m3 for Qs section and 52.3 kg/m3 for Fs section) and the RPD (1.5 for Qs section and 1.1 for Fs section). Several pre-treatments were applied, but no significant improvement was found in the model performance. The optimal spectral range for ρ prediction was also investigated in this study. The spectral range of 1100–2500 nm was used to develop the ρ model, and slight improvement was observed in the predictions for the grain orientations (Table 4.2). While the ρ prediction values are considered reasonable, better results have been reported by other researchers (Schimleck et al. 2002). A likely reason for this discrepancy may be related to the single species used for calibrations in their study. This is because single-species calibrations give better results than mixed-species calibrations such as those used in this study (Alves et al. 2012).  The smooth surface predictions were also similar to those from the rough surface (Qs section reported better statistics). The Rc2 ranged from 0.64–0.74 and 0.30–0.47 for the Qs and Fs sections, respectively. In the same way, the Rp2 ranged from 0.56–0.63 for the Qs section and was 0.59 for the Fs section. A similar spectral range (1100–2500 nm) also improved the calibration for the grain 39  orientations based on the smooth surface (Table 4.2). Overall, calibrations based on the Qs section were better than those from the Fs section, which is also in line with past work (Gindl et al. 2001).      Figure 4.4: PCA plot for the best model based on the rough surface and quarter-sawn section (top) and the smooth surface and quarter-sawn section (bottom).  The difference in the spectral information of the surfaces, especially for ρ, has been previously explained (Fujimoto et al. 2008) and was attributed to the interaction between the chemical and 40  anatomical properties of wood. Another reason could be the higher exposure of parenchyma cells in the Qs section compared to the Fs section, which may have affected calibrations (Hein et al. 2009a). The best prediction for ρ was observed when the rough surface was utilized. The highest prediction value from the ρ calibrations was found by utilizing the Fs section, a finding similar to a past report (Ribeiro 2009). The potential for predicting the ρ of wood has been revealed in past work (Mora et al. 2008), where strong correlations (Rp2 = 0.80–0.84) were reported between the measured values and NIR-predicted values for softwood.  A PCA was also used to distinguish between the Qs and Fs sections of the two softwoods. The PCA scores plots of the best model explained 98% of the variation in the data for both rough and smooth surfaces (Figure 4.4). For the rough surface, PC1 showed more variability in the Qs samples, which were clustered together (93% variation explained) than did PC2, where the samples were far apart and explained only 5% of the variation in the data (Figure 4.4, top). In the same way, the PCA obtained from the smooth surface explained more variation (97% for PC1, 1% for PC2) in the model with samples found to be highly correlated to one another (Figure 4.4, bottom). Even though some of the samples overlap (falling between PC1 and PC2), the results of the PCA model shows that 93% of the data collected from the Qs section of both species could be separated with certainty, especially when they fell outside the overlapping zone (Figure 4.4).  The plots showing the relationship between NIR predicted and measured values for the best models selected are shown (Figure 4.5) for both surface conditions. From the figures, it can be seen that the data points are not close to the target line, and the calibration and validation data are not tightly packed together. This shows that the predictions obtained for the  ρ of wood were only moderate. The regression coefficients of the grain orientations for the rough surface were similar, suggesting that the same wavelengths contributed to ρ prediction. Similar wavelengths described earlier in this thesis contributed to the ρ prediction of the smooth surface. The assignment of bands that contributed to ρ prediction has been described earlier in this thesis (Section 4.2.1).  41    Figure 4.5: The relationship between measured and NIR-predicted values of ρ (kg/m3), based on the rough (top) and smooth (bottom) surfaces (quarter-sawn section) for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.   4.3 Modulus of Elasticity  4.3.1 MOE based on calibrations for all samples                                                                                             T The MOE values ranged from 1587–6182 MPa and 5532–13247 MPa for hemlock and D-fir, respectively. Similar values have been reported for the MOE of D-fir (FPL 2010, Mousavi 2016), 42  although higher values were found for hemlock (Rohrbach 2008, FPL 2010). The reason for this could be the lower density values (Johnson and Gartner 2006) and the possible effect of knots observed in a few samples, which might have caused decreases in the MOE values (Koman et al. 2013). Higher MOE values were observed for D-fir than for hemlock, as expected. This could be explained by the reduction in its M (13.3) as compared to hemlock (13.7), as previously reported in this thesis (Section 3.3). Past researchers have also revealed that an increase in the MOE of wood could be linked to the decrease in its M below the fibre saturation point (FSP) (Skaar 1988, Güntekin and Aydin 2013). For the modelling of MOE, 144 and 96 samples were used for the calibration and prediction sets, respectively. The descriptive statistics for this property from the calibration and prediction sets are shown in Table 4.1.  The raw spectra were used to develop calibrations for MOE, and moderate predictions were obtained. An Rc2 of 0.69 and an Rp2 of 0.57 were reported for the rough surface, whereas lower values with high prediction errors were observed for the smooth surface (Table 4.3), while a reduced spectral range (1100–2300 nm) gave the best prediction for the rough surface with six factors, which makes the model less robust. The best model (Rp2 of 0.67) with reduced prediction error was observed when the spectral range was 1300–2300 nm for the smooth surface. These values are considered reasonable due to the robustness of the model and the fact that the model was validated using separate samples.  Pre-treatments improved the model quality for the rough surface, whereas no improvement was observed for the smooth surface. The first derivative slightly improved the prediction value (from 0.57 to 0.60) and reduced the prediction error (Table 4.3). Improvement in the model quality with reduced spectral range was not unexpected, especially in the spectral range of 1000–2500 nm. It has been established that this range contains the most distinct spectral information on the first overtone and combination bands that make up the NIR spectra (Kelley et al. 2004b). Reducing the spectral range for calibrations allows for the use of lightweight, portable, and inexpensive spectrometers for field applications (Hedrick et al. 2004). The calibration statistics are summarized in Table 4.3.   Most studies using independent samples rather than cross-validations to validate prediction models generally indicate that cross-validations are overly optimistic (Brown et al. 2005, Brunet et al. 43  2007). However, the PLS-R models developed for the MOE of agro-based particleboards by NIRS have yielded Rp2 values ranging from 0.42 to 0.62 (Hein et al. 2011), which are lower than the results in this study. A ratio of performance to deviation (RPD) of between 1.3 and 1.6 was reported, making it suitable for initial screening. NIR spectra, especially from complex materials like wood and panel products, are driven not only by the chemistry of the material but also by the solid structure (Gierlinger et al. 2004).   Table 4.3: Calibration statistics for MOE based on all samples (n = 240)  and Qs and Fs samples (n = 120).     Property Roughness Spectral Range Rc2 Rp2 SEC SEP RPD MOE  (All) Rough Raw Data (Full Range)  1st Derivative  Smoothed (1100–2300) 0.69  0.74  0.64 0.57  0.60  0.71 1166  1070  1250 1538  1486  1263 1.5  1.5  1.8  Smooth Raw Data (Full Range)  Reduced (1300–2300)  Reduced (1100–2500) 0.58  0.66  0.75 0.41  0.67  0.64 1364  1229  1049 1779  1358  1416 1.3  1.7  1.6 MOE  (Qs) Rough Raw Data (Full Range) 0.77 0.78 1149 955 2.1  Smooth Raw Data (Full Range)  Raw Data (900–1900) 0.85  0.84 0.68  0.72 907  963 1160  1087 1.7  1.8 MOE  (Fs) Rough Raw Data (Full Range)  Raw Data (1000–2500) 0.67  0.72 0.57  0.63 991  917 1719  1590 1.5  1.6  Smooth Raw Data (Full Range) 0.75 0.69 856 1453 1.7 44  The RPD values were used to evaluate the model performance in this study. All RPD values for MOE met the 1.5 criteria, except for those based on raw spectra for the smooth surface, where the RPD was 1.3 (Table 4.3). It has been shown that NIRS can be used as an initial screening tool with an RPD of approximately 1.5 (Schimleck et al. 2003, Acquah et al. 2018). It has also been suggested that an RPD between 1.5 and 2.5 is enough for estimating wood properties (Hein 2010).     Figure 4.6: The relationship between measured and NIR-predicted values for MOE (MPa), based on the rough surface (top) and smooth surface, 1300–2300 nm (bottom) for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.  45  The regression coefficients showing the important variables for these models were observed. The important wavelengths identified for MOE prediction are 1212–1225, 1480, 1477–1484, 1484–1493, and 1586–1596, which have been ascribed to cellulose, while 1471 is related to hemicelluloses and 2200 to lignin. For small clear specimens, high correlations for the MOE and MOR were found at the absorption bands due to -OH and -CH in the semi-crystalline or crystalline regions in cellulose. These results explain that an increase in the semi-crystalline or crystalline regions of cellulose is related to an increase in the MOE and MOR (Fujimoto et al. 2007). The plots showing the relationship between NIR predicted and measured values for the best models are shown in Figure 4.6 for both surface conditions.  The results obtained in this study using NIRS to predict the MOE of wood were better than those reported by Acquah et al. (2018), with cross-validated models for predicting this property for new samples. The Rp2 reported was 0.45 for MOE, even though the samples were taken from many trees (450). The errors reported in their study were also higher than what were obtained in this study. Higher errors are usually associated with predicting an independent test set from calibrated models, because how worse a model will perform is considered when applied to the test set not originally used in model training. Overall, the results based on the rough surface were better than those based on the smooth surface (Table 4.3).   4.3.2 MOE based on calibrations for Qs and Fs samples                                                     T               For the modelling of MOE, 72 and 48 samples of both species were used for the calibration and prediction sets, respectively. Most of the models for MOE gave strong correlations of Rc2, ranging from 0.77–0.85 and 0.67–0.75 for the Qs and Fs sections, respectively. The Rp2 ranged from 0.68–0.78 and 0.57–0.69 for the two sections, respectively. These results are in line with the MOE calibration statistics previously reported for spectra collected from similar grain orientations (Thumm and Meder 2001, Andrade et al. 2010, Kothiyal and Raturi 2011).  Several pre-treatments were applied to improve the model quality but showed no effect. Thus, all models for MOE were based on raw spectra. For the rough surface, the results based on the Qs section yielded better calibration statistics than the Fs section (Table 4.3). A good Rc2 of 0.77 and an Rp2 of 0.78 were reported for the Qs section, while an Rc2 of 0.67 and an Rp2 of 0.57 were reported for the Fs section. These results show that the former provides better predictions for MOE than the 46  latter, which is similar to past reports (Thumm and Meder 2001, Zhao et al. 2009). The Qs section showed the potential for predicting new samples better (RPD > 2.0). As reported previously for agricultural materials, an RPD between 2.0 and 2.5 indicates very good prediction (Mouazen et al. 2005). Reducing the spectral range (from 1000 to 2500 nm) slightly improved the predictions, but only for the Fs section (Rp2 from 0.57 to 0.63; see Table 4.3).  Similar predictions to the rough surface were also observed for the smooth surface, where the Qs section reported better statistics than the Fs section. Reasonable predictions were observed for the Qs (Rc2 = 0.85, Rp2 = 0.68) and Fs (Rc2 = 0.75, Rp2 = 0.69) sections. The same trend was observed for the prediction error of the Qs (1160 MPa) and Fs (1453 MPa) sections, respectively. It was evident that better predictions were obtained from the Qs section. Calibrations based on the Qs and Fs sections previously reported (Kothiyal et al. 2014) showed that the former had the stronger calibrations. The difference in the accuracy of prediction could be linked to the surface condition of the sample, since earlywood and latewood are better represented in the Qs section than the Fs section. The resulting spectrum from the former usually has a better representation of the total wood characteristics (Thumm and Meder 2001) than the latter, in which calibrations depend on whether earlywood or latewood was present when spectra were collected (Dahlen et al. 2017).  The narrow spectral range (900–1900 nm) created for the Qs section improved the model (Rp2 increased from 0.68 to 0.72), and the SEP dropped from 1160 to 1087 MPa. Spectral ranges of 900–1900 nm and 1000–2500 nm also improved the model quality. Reducing the spectral range is deemed important because it is less expensive than full spectral analysis (Yu et al. 2009). The aforementioned spectral ranges are considered effective for predicting the MOE of wood. Overall, the Qs section reported better statistics than the Fs section for rough and smooth surface conditions, as shown in Table 4.3. The highest predictive value for MOE was obtained from the Qs section (Rp2 = 0.78) based on the rough surface, which agrees with past work when calibrations were developed using the same section (Fujimoto et al. 2008).    The models for MOE in this study were constructed with no more than five factors. This is in contrast with previous claims (Fujimoto et al. 2008) that six to eight factors are needed to build good models that predict mechanical properties for small clear samples. The calibrations presented in this study demonstrate that NIRS has the potential to predict the mechanical properties of wood 47  samples with considerable accuracy. NIRS has also been used to predict the mechanical properties of wood-based materials such as medium-density fibreboard, with an Rp2 of 0.80–0.82 (Rials et al. 2002).     Figure 4.7: The relationship between measured and NIR-predicted values for MOE (MPa), based on the quarter-sawn (top) and flat-sawn (bottom) sections for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.  48  The PCA was used to distinguish between the grain orientations of the two softwoods and show variations in the data. The PCA scores plots of the best model selected explain 98% and 96% variation in the data for the rough surface, and 99% and 96% explained for the smooth surface. The plots showing the relationship between NIR predicted and measured values for the best models selected for the Qs and Fs sections are shown in Figure 4.7. The wavelengths contributing to the MOE prediction of the former and latter are similar to those reported earlier in this thesis (Section 4.3.1). The assignment of NIR peaks has been reported by several authors (Via et al. 2003, Tsuchikawa and Siesler 2003).  4.4 Modulus of Rupture   4.4.1 MOR based on calibrations for all samples The MOR assessed in this study ranged from 18–83 MPa and 54–114 MPa for hemlock and D-fir, respectively. Similar values have been reported for the MOR of hemlock (Rohrbach 2008) and D-fir (FPL 2010, Mousavi 2016). Higher MOR values were observed for D-fir than for hemlock, which is similar to past reports (FPL 2010). This could be attributed to the difference between the M and ρ of the species. Past reports have revealed that a lower M and a higher ρ, as observed in D-fir, is accompanied by an increase in the mechanical properties (Tsoumis 2020). This could have contributed to the higher MOR values observed in D-fir. For the modelling of MOR, the same number earlier described (Table 4.1) were used for the calibration and prediction sets.  Several pre-treatments were applied, with no significant improvement observed, so raw spectra were used to build the model for both surface conditions. For the rough surface, moderate predictions (Rp2 = 0.56, SEP = 12.24 MPa) were reported, while lower values (Rp2 = 0.42, SEP = 13.64 MPa) were reported for the smooth surface. The 1300–2300 nm range based on first-derivative spectra gave a good prediction for the smooth surface, with Rc2 increasing from 0.55 to 0.74 and Rp2 from 0.42 to 0.66. The SEC dropped from 12.05 to 9.15, while the SEP went from 13.64 to 10.76 MPa (Table 4.4). Previous works have revealed similar values for the MOR of lumber (Schimleck et al. 2018). These findings also agree with the work of others, where similarities were found in the prediction of lumber and small clear samples (Fujimoto et al. 2008).    49  Table 4.4: Calibration statistics for MOR based on all samples (n = 240)  and Qs and Fs samples (n = 120).   However, stronger correlations (Rp2 = 0.81–0.86) were previously reported for small clear samples of Pinus taeda, regardless of whether the cross-section or radial surface was rough or sanded (Schimleck et al. 2005b). Fujimoto et al. (2007) also reported an Rp2 of 0.84 when the MOR of hybrid larch was predicted using NIRS. The lower values that were observed in this study could be attributed to not separating the samples into juvenile and mature wood. This could have contributed to poorer calibrations, as samples containing a wide variety of juvenile and mature wood have been reported to have lower calibration performance (Dahlen et al. 2017). Another reason for the lower values could be the range of values of MOR investigated in this study, since a wide range of values has been linked to improved model quality (Gindl et al. 2001).   Property Roughness Spectral Range Rc2 Rp2 SEC SEP RPD MOR  (All) Rough Raw Data (Full Range)  Smoothed 0.68  0.68 0.56  0.57 10.18  10.25 12.24  12.10 1.4  1.4  Smooth Raw Data (Full Range)  1st Derivative (1300–2300) 0.55  0.74 0.42  0.66 12.05  9.15 13.64  10.76 1.3  1.6 MOR  (Qs) Rough Raw Data (Full Range) 0.77 0.52 9.52 11.48 1.4  Smooth Raw Data (Full Range)  Raw Data (1100–2500) 0.87  0.85 0.57  0.65 7.20  7.68 10.72  9.85 1.5  1.6 MOR  (Fs) Rough Raw Data (Full Range)  Smoothed (1300–2300) 0.77  0.66 0.76  0.82 7.42  9.05 9.87  8.61 1.9  2.2  Smooth Raw Data (Full Range)  Raw Data (900–1900) 0.72  0.73 0.67  0.71 8.24  7.97 11.14  11.10 1.7  1.7 50  The calibrations for MOR were lower than those for MOE, which could be because the ρ of wood samples strongly affects the MOR rather than the MOE (Fujimoto et al. 2007). The relationship between MOE and MOR and several properties of wood have been described in detail (Kliger et al. 1995). Overall, the smooth surface gave better predictions than the rough surface, a reverse trend to the models built for ρ and MOE.  The important wavelengths identified for MOR prediction were 1212–1225, 1480, 1477–1484, 1484–1493, 1586, 2110 (due to cellulose), 1907, 1910, 2134 (due to hemicelluloses), and 2200 (due to lignin; see Figure 4.8). Significant correlations for the MOR were found at the absorption bands due to the -CH in hemicelluloses, meaning that increase in hemicelluloses is related to an increase in the MOR of wood (Fujimoto et al. 2007). Some wavelengths show negative correlations at bands in the 1868–1981 nm range, which are associated with water and known to play an important role in the NIR spectrum of wood (Thygesen and Lundqvist 2000, Bokobza 2002). All band assignments for the spectra of wood can be found elsewhere (Schwanninger et al. 2011).  Plots showing the relationship between NIR predicted and measured values for the best models for both surfaces are shown in Figure 4.9. NIRS coupled with PLS-R was found to be a suitable technique for predicting the ρ, MOE, and MOR of wood. Overall, the best prediction was obtained for MOE, followed by MOR and then ρ, which is in accordance with what was reported in previous studies (Haartveit and Flæte 2006). 51    Figure 4.8: Regression coefficients based on raw data for the rough surface (top) and the smooth surface (bottom).    52    Figure 4.9: The relationship between measured and NIR-predicted values for MOR (MPa), based on the rough surface (top) and smooth surface (bottom) for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.   4.4.2 MOR based on calibrations for Qs and Fs samples For the modelling of MOR, 72 and 48 samples of both species were used for the calibration and prediction sets, respectively. Most of the models for MOR gave strong correlations, with Rc2 53  ranging from 0.77–0.87 and 0.66–0.77 for the Qs and Fs sections, respectively. The Rp2 ranges were from 0.52–0.65 and 0.67–0.82 for the former and latter, respectively. Similar reports for MOR calibrations have been found in the literature (Fujimoto et al. 2007). A summary of the calibration statistics is presented in Table 4.4.  Pre-treatments were applied to improve the model quality, but raw spectra gave the best predictions for the rough surface. The models based on the Fs section reported better performance than the Qs section (Table 4.4). A good Rc2 of 0.77 and a moderate prediction value (Rp2 = 0.52) were obtained for the Qs section, while the Fs section yielded an Rc2 of 0.77 and a higher value of Rp2 (0.76). The results obtained for the Qs section are considered moderate, as very low prediction value (Rp2 = 0.44 and RPD = 1.1) have been reported (Andrade et al. 2010). The values obtained for the Fs section are similar to those in past reports (Fujimoto et al. 2008, Andrade et al. 2010). Reducing the spectral range to 1100–2500 nm, 900–1900 nm, and 1300–2300 nm did not affect the model performance for either grain orientation. An exception was found when smoothed data in the range of 1300–2300 nm was used to predict the MOR for the Fs section, which slightly improved the Rp2 from 0.76 to 0.82 and reduced the SEP from 9.87 to 8.61.  A similar trend in the results for the rough surface was observed for the smooth surface where the Fs section yielded better statistics than the Qs section. A good Rc2 of 0.87 and an Rp2 of 0.57 for the Qs section and an Rc2 of 0.72 and Rp2 of 0.67 for the Fs section were obtained. Models based on the latter were chosen as the best due to the high Rp2 and RPD values (Fujimoto et al. 2008). Pre-treatments such as smoothing, SNV, first derivative, and second derivative reduced the model quality. Reducing the spectral range to 1000–2500 nm for the Qs section slightly improved the predictive ability (Rp2 from 0.57 to 0.65) and reduced the error (SEP from 10.72 to 9.85; see Table 4.4). For the Fs section, the prediction slightly improved in the 900–1900 nm range (Rp2 increased from 0.67 to 0.71, while the SEP dropped from 11.14 to 11.10). The improvement in the model performance due to the reduced spectral ranges was not unexpected. The work of other researchers also revealed that the 900–1900 nm range is useful for predicting wood’s mechanical properties (Liang et al. 2016).  All RPD values for MOR meet the 1.5 criteria except for a model based on the rough surface and Qs section, where the RPD was 1.4 (Table 4.4). A considerably higher RPD of 2.2 was observed 54  for the best model selected for the Fs section, which is in line with the work of Fujimoto et al. (2008) for the same grain orientation. It has been established that a model must have an RPD greater than 1.5 to be considered as a preliminary screening tool (Hein et al. 2009a). Overall, the Fs section yielded better results than the Qs section for both surface conditions, and the best prediction for MOR was observed when the rough surface was utilized.  The PCA was used to distinguish between the grain orientations of the two softwoods and show variations in the data. The PCA scores plots of the best model selected explain 95% and 99% of the variation in the data for the grain orientations based on a rough surface. The PCA for the smooth surface explains 99% and 98% of the variation in the data for both grain orientations. Plots showing the relationship between NIR predicted and measured values for the best models selected for the Qs and Fs sections are shown in Figure 4.10. The interpretation of PCA and PLS-R plots has been reported earlier in this thesis (Sections 4.2.1 and 4.2.2).  The significant wavelengths contributing to MOR prediction for both grain orientations were similar to those reported earlier in this thesis. The bands at 1674 and 1684 nm were found to be negatively correlated to MOR. These bands have previously been assigned to extractives and lignin (Michell and Schimleck 1996); the lignin content of wood has been reported to influence NIR absorbance (Schimleck et al. 1997). The relationship between mechanical properties such as MOR and NIR spectra could be attributed to how lignin content and composition affected NIR absorbance (Gindl et al. 2001). The absorption bands related to cellulosic features were found to be the most important contributors in model building for MOR, which is in accordance with what others have reported (Fujimoto et al. 2008). 55     Figure 4.10: The relationship between measured and NIR-predicted values for MOR (MPa), based on the rough surface (flat-sawn) (top) and smooth surface (quarter-sawn) (bottom) for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.    56  4.5 Slope of Grain (SOG)  4.5.1 SOG based on calibrations for all samples                                                                                        T The SOG assessed in this study ranged from 1.33–16.67 for hemlock and 1.67–15 for D-fir. For the modelling of SOG, the same number of samples as described above (see Table 4.1) were used for the calibration and prediction sets.     SOG is highly important for load-bearing applications because of its influence on wood’s MOE and MOR. PLS-R analysis coupled with NIR absorbance spectra were used to determine the wood’s SOG. The NIR spectra were taken from the smooth surface to increase the accuracy of prediction. Raw spectra were used to develop calibrations but gave poor predictions, with an Rc2 of 0.13 and an Rp2 of 0.21. An SEC of 2.99 and an SEP of 3.03 were obtained (Table 4.5).   Table 4.5: Calibration statistics for SOG based on all samples (n =240)  and Qs and Fs samples (n =120).  The regression coefficients show the relationship between SOG and absorbance spectra. The correlation was variable over the observed wavelength range in this study, with positive and negative values, which is in accordance with past research (Gindl and Teischinger 2002). The PCA plot explains the variation in the data (96%) (Figure 4.11), with PC1 explaining more variability in the samples (93%) than PC2 (3%). A plot showing the relationship between NIR predicted and measured values is provided in Figure 4.12, where it is evident that the correlation between the Property Roughness Spectral Range Rc2 Rp2 SEC SEP RPD SOG  (All) Smooth Raw Data 0.13 0.21 2.99 3.03 1.1 SOG  (Qs) Smooth Raw Data (Full Range) 0.15 0.21 2.64 2.84 1.1 SOG  (Fs) Smooth Raw Data (Full Range)  Raw Data (1000–1960) 0.38  0.36 0.17  0.24 2.03  2.06 2.16  2.16 1.1  1.1 57  calibration and validation data was weak and the model is a poor fit (data points far from the target line).  The pre-treatments earlier described in this thesis were applied to the spectra but yielded no improvement in model quality. Reducing the spectral range based on observation of the regression coefficient also resulted in no improvement. The calibration statistics reported in this study were very low. This may be due to the method used to measure the SOG of wood. More precise methods and equipment should be used in future studies to check the predictive value of SOG using NIR. Another reason for the poor prediction may have been the range of SOG assessed in this study. A very strong correlation between NIR and the SOG of wood has been reported, with an Rp2 of 0.80 (Gindl and Teischinger 2002). The range of observed values may have contributed to the calibration performance, as the range of SOG in their study was 0–40 compared to 1.33–16.67 in this study. A wide range of values has been reported to improve NIRS calibrations.    Figure 4.11: PCA plot based on raw spectra for SOG. 58   Figure 4.12: The relationship between measured and NIR-predicted values for SOG (), based on the raw spectra for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.  4.5.2 SOG based on calibrations for Qs and Fs samples For the modelling of SOG, 72 and 48 samples of both species were used for the calibration and prediction sets, respectively. All models for SOG gave poor correlations, with Rc2 ranging from 0.15–0.38 and Rp2 from 0.17–0.24 for the Qs and Fs sections, respectively.  Raw spectra were used to develop calibrations, as pre-treatments failed to improve the model performance. The results based on Fs section were considered better (Rp2 = 0.24) than those based on Qs section (Rp2 = 0.21). An optimal spectral range of 1000–1960 nm has been reported to yield good predictions for SOG (Gindl and Teischinger 2002). This range slightly improved the model for the Fs section (Rp2 increased from 0.17 to 0.24). The poor predictions obtained for SOG may be due to the narrow range of values and the precision of the equipment used. A wider range of values and more accurate measurements should be considered in future research.  The PCA was used to distinguish between the grain orientations of the two softwoods. The PCA scores plots of the best model selected explain 97% and 99% of the variation in data for the Qs and Fs sections, respectively. The plot of the relationship between NIR predicted and measured values for the best model, selected based on the latter, is shown in Figure 4.13. For the regression 59  coefficients, negative correlations were observed at bands between 1432 and 1613 nm, which have been associated with cellulose, hemicelluloses, and lignin (Schwanninger et al. 2011). This is also similar to previous findings (Gindl and Teischinger 2002).   Figure 4.13: The relationship between measured and NIR-predicted values for SOG (), based on the flat-sawn section for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.  4.6 Annual Ring Width (ARW)  4.6.1 ARW based on calibrations for all samples The range of ARW observed in this study was 0.66–7.40 mm for hemlock and 0.10–13.10 mm for D-fir. For the modelling of ARW, the same numbers of samples as described in Table 4.1 were used for the calibration and prediction sets.     Raw spectra were used to develop calibrations for the ARW. Raw spectra gave poor predictions (Rp2 = 0.18) for the rough surface and worse results for the smooth surface (Rp2 = 0.04). Pre-treatments were applied but did not improve the model’s accuracy. However, reducing the spectral range to 800–1400 nm improved the Rc2 from 0.36 to 0.44 and reduced the calibration error from 1.40 to 1.31 mm for the rough surface (Table 4.6). The reason for these suboptimal results might be the narrow range of values observed in this study.   60  Table 4.6: Calibration statistics for ARW based on all samples (n = 240)  and Qs and Fs samples (n = 120).  The poor calibration statistics observed in this study could also be attributed to the method used in determining the ARW. No past studies were found in the literature to serve as a basis of comparison. Further studies should be done using more accurate measurements and a wide range of reference values to improve calibration performance. The PCA plot for the best models selected explains the variation in the data (96% and 96%) for both surface conditions. The plots showing the relationship between NIR predicted and measured values for both surface conditions are provided in Figure 4.14. Property Roughness Spectral Range Rc2 Rp2 SEC SEP RPD  ARW  (All)   Rough  Raw Data (Full Range)  Reduced (800–1400) * 0.36  0.44 0.18  0.19 1.40  1.31 1.43  1.45 1.2  1.2  Smooth Raw Data (Full Range) 0.10 0.04 1.67 1.58 1.0  ARW  (Qs)    Rough  Raw Data (Full Range) 0.56 0.34 1.06 0.81 1.2  Smooth Raw Data (Full Range)  Smoothed 0.54  0.67 0.25  0.30 1.08  0.95 0.81  0.84 1.2  1.2   ARW  (Fs)    Rough  Raw Data (Full Range)  Raw Data (800–1400) 0.27  0.44 0.09  0.24 1.62  1.42 1.81  1.78 1.1  1.2  Smooth Raw Data (Full Range) 0.02 0.04 1.88 2.01 1.0 61    Figure 4.14: The relationship between measured and NIR-predicted values for ARW (mm), based on the rough surface (top) and smooth surface (bottom) for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.   62  4.6.2 ARW based on calibrations for Qs and Fs samples For the modelling of ARW, 72 and 48 samples of both species were used for the calibration and prediction sets, respectively. Most of the models for ARW gave fair calibration values, with Rc2 ranging from 0.02–0.67 and Rp2 from 0.04–0.34 for the Qs and Fs sections, respectively.  Pre-treatments were applied to improve the model performance, but no effect was observed. For the rough surface, results based on the Qs section gave fair predictions (Rp2 = 0.34) compared to the Fs section (Rp2 = 0.09). Reducing the spectral range to 800–1400 nm and 1300–2300 nm failed to improve the model’s performance except when the range of 800–1400 nm was utilized for the Fs section (Rp2 increased from 0.09 to 0.24, and SEP dropped from 1.81 to 1.78).   The smooth surface predictions were similar to those based on the rough surface. The Qs section gave a better prediction value (Rp2 = 0.25) than the Fs section (Rp2 = 0.04). SG smoothing also improved the prediction for the former (Rp2 from 0.25 to 0.30). A comprehensive report on the influence of pre-treatments on model performance has been published (Rinnan et al. 2009). Overall, the Qs section outperformed the Fs section for both surfaces. The best prediction for ARW was observed when the rough surface was utilized.  Calibrations were also developed using a CV method with 10 segments to test the predictive ability of the ARW model. The combination of SNV and second derivative was applied to the spectra to improve the model performance for the Qs section, and a reasonable prediction was obtained, with Rc2 = 0.77 and Rcv2 = 0.50. This further shows it is possible to predict the ARW of wood using NIR with more accurate measurements and a wide range of values. Although new samples were not predicted, a CV method has been used by several researchers to determine the predictive ability of models (Gindl et al. 2001, Kobori et al. 2015).  The PCA was used to distinguish between the grain orientations of the two softwoods. The plots showing the relationship between NIR predicted and measured values for the best models selected based on the Qs section are provided in Figure 4.15. As observed from the regression coefficients generated by the PLS-R models, the relationship between the ARW and absorbance was highly variable, changing from positive to negative over the investigated wavelength range. The bands at 63  2200–2300, 1877–1930, and 1687–1700 nm were found to be fairly correlated with ARW. All band assignments for the NIR spectra of wood can be found elsewhere (Schwanninger et al. 2011).    Figure 4.15: The relationship between measured and NIR-predicted values for ARW (mm), based on the rough surface (top) and smooth surface (bottom) in the quarter-sawn section for the softwoods. The blue and red dots indicate the calibration and prediction sets, respectively.    64  5.0 CONCLUSION   NIRS coupled with PLS-R was found to be a suitable technique for the prediction of wood properties. While a good number of studies have utilized this technique to estimate several properties of wood, the current study focused on two coastal softwoods (Western hemlock and Douglas-fir), which have not received much attention from past researchers and thus make this essentially new research. The results of this study demonstrate that NIRS has the potential to predict some wood properties, although poor calibrations were developed for the SOG and ARW of wood, which could have been due to the narrow range of values investigated and the testing procedures. Differences between the calibration models for the separate sets (the Qs and Fs sections) of the two species utilized had not previously been explored and hence were the focus of this study.  Two sets of predictive models were developed in this thesis: one included all the samples (n = 240) and another consisted of samples from the Qs and Fs sections separately (n = 120) for both surface conditions (rough and smooth). Different surface conditions were used to simulate the actual situation in a sawmill, where the surface quality of lumber varies due to factors such as roughness, grain, and knots. The models that included all samples were inferior to the ones with separation of the samples into Qs and Fs sections for the assessed properties. For the first set, the best prediction was obtained for MOE, followed by MOR and then ρ, with poor results for SOG and ARW. The rough surface predictions were better than the smooth surface for all properties except MOR, which yielded better predictions on the smooth surface. The second set of predictions followed the same trend, with the best predictions obtained in the order of MOE, MOR, and ρ, and poor results for SOG and ARW. For this set, the Qs section gave better results than the Fs section for both surface conditions (rough and smooth), except for MOR based on a rough surface, which showed the opposite trend. The Fs section reported better results for SOG based on the smooth surface. Most of the models reported in this thesis met the 1.5 RPD criteria for a model to be considered as a preliminary screening tool.  The models developed using the reduced spectral ranges also showed slight improvement in model performance compared to those based on the full NIR range. These findings therefore show the feasibility of using portable, lightweight, and inexpensive spectrometers for predicting wood 65  properties by NIRS. The results of this study also show that the application of pre-treatments to the NIR spectra helped to optimize the model. It was further concluded that the same wood chemistry associated with wavelengths related to ρ was also found to be related to MOE and MOR. This could have been the result of the strength of the relationship established between these properties; hence, this study found that the same underlying chemical constituent that contributes to ρ prediction is also responsible for both MOE and MOR. The contributions of these polymers to mechanical properties have been studied extensively but are rather controversial. It is not entirely possible to comprehensively interpret the relationship between chemical components and mechanical properties. However, the absorption bands related to cellulosic features are certainly the main contributors to building robust models for predicting the mechanical properties of wood.                      66  6.0 FUTURE RESEARCH  In light of this investigation, and acknowledging the constraints and limitations of the experimental design and available equipment, the following are suggested as areas for future research related to the use of NIRS as a non-destructive evaluation method to assess wood properties.  1. The lower statistics reported in this thesis could have been a result of the samples used when developing calibrations; specifically, mixed-species calibrations could have caused the lower calibration statistics. Therefore, single-species calibrations with diverse samples should be considered in future studies to improve the quality of the model. 2. Although clear wood samples were used in this study, a negligible number of samples were found to have small knots, which could have affected the calibrations. Studies should be conducted to determine the differences in calibrations for wood of higher and lesser grades. 3. One of the most limiting factors in NIRS calibration is the accuracy of the reference data. Several attempts have been made to determine the best analytical conditions (sample preparation and spectrum acquisition technique) that minimize scatter in NIRS methods. This is important because not even the most powerful NIRS calibration software can give reliable predictions from low-quality spectra. Therefore, work and time need to be spent on spectral collection and reference measurements to improve the overall statistical models from NIRS calibrations.  4. The NIR spectra in this study were collected in a static condition. However, it would be helpful to understand the influence of measurement conditions on calibrations. Combining the NIR spectra from moving measurement conditions and those from static conditions could improve the performance of the model.  5. NIRS calibrations based on samples with a wide range of values could be done to build more robust and reliable models. This is because smaller datasets and a narrower range of properties could compromise the model’s quality more than the limitations inherent to particular species.  6. Several studies have linked the use of reduced wavelengths in NIRS studies to much lower cost, which necessitates the prediction of wood properties with narrow spectral ranges. So, 67  more studies should be carried out to assess calibration performance within narrow spectral ranges. 7. Further studies with well-defined grain orientation are also needed to get better results. Only two sections (Qs and Fs) of the three found in wood were investigated in this study. A study that utilized all three sections, that is the transverse, Qs, and Fs sections, should be considered. The direction of scan—for instance, with the transverse section, where variations in wood properties can be accurately captured—contributes to better predictions. 8. Finally, other technologies such as hyperspectral imaging, which combines NIRS and an imaging system and allows the accurate measurement of wood properties, should be explored.                     68  REFERENCES  Acquah, G. E., Essien, C., Via, B. K., Billor, N., & Eckhardt, L. G. (2018). Estimating the basic  density and mechanical properties of elite loblolly pine families with near infrared  spectroscopy. Forest Science, 64(2): 149-158. Adedipe, E. O. & Dawson-Andoh, B. (2008). Predicting moisture content of yellow-poplar  (Liriodendron tulipifera L.) veneer using near infrared spectroscopy. Forest Products  Journal, 56(4): 28-33. Agelet, L. E. & Hurburgh Jr, C. R. (2010). A tutorial on near infrared spectroscopy and its  calibration. Critical Reviews in Analytical Chemistry, 40(4): 246-260. Alin, A. (2009). Comparison of PLS algorithms when number of objects is much larger than  number of variables. Statistical papers, 50(4): 711-720.                                                                                                                                                                                                      Alteyrac, J., Cloutier, A., Ung, C. H., & Zhang, S. Y. (2006). Mechanical properties in relation to  selected wood characteristics of black spruce. Wood and Fiber Science, 38(2): 229-237. Alves, A., Santos, A., Rozenberg, P., Pâques, L. E., Charpentier, J. P., Schwanninger, M., &  Rodrigues, J. (2012). A common near infrared—based partial least squares regression  model for the prediction of wood density of Pinus pinaster and Larix× eurolepis. Wood  Science and Technology, 46(1-3): 157-175.                                                             Anastasopoulou, J. & Theophanides, T. (1972). Chemistry and Symmetry, In Greek National  Technical University of Athens, NTUA, (1997), 94 p.  Anderson, E. A., Koehler, A., & Krone, R. H. (1945). Instruments for rapidly measuring slope of  grain in lumber. US Forest Service Report, Forest Products Laboratory, Madison,  Wisconsin, 1592, 12 pp.  Andersson, M. (2009). A comparison of nine PLS1 algorithms. Journal of Chemometrics: A  Journal of the Chemometrics Society, 23(10): 518-529. Andrade, C., Trugilho, P., Napoli, A., Vieira, R., Lima, J., & Sousa, L. (2010). Estimation of the  mechanical properties of wood from Eucalyptus urophylla using near infrared  spectroscopy. Cerne, 16(3): 291-298.  Anjos, O., Pereira, H., & Rosa, M. (2010). Tensile properties of cork in the tangential direction:  Variation with quality, porosity, density and radial position in the cork plank. Materials  and Design, 31(4): 2085-2090.                                                                                                                                                                                        ASNT. (2016) Recommended Practice No. SNT-TC-1A: Personnel Qualification and  Certification in Non-destructive Testing (2016).                                                                     ASTM D-198 (2005). Standard test methods of static tests of lumber in structural sizes. ASTM  International, Philadelphia.                                                                                           ASTM D143-94. (2007). Standard test methods for small clear specimens of timber. ASTM  International, West Conshohocken, PA, USA.                                                                         Antwi-Boasiako, C. & Atta-Obeng, E. (2009). Vessel-Fibre Ratio, Specific Gravity and  Durability of Four Ghanaian Hardwoods. Journal of Science and Technology  (Ghana), 29(3).                                                                                                               Atzberger, C., Guerif, M., Baret, F., & Werner, W. (2010). Comparative analysis of three  chemometric techniques for the spectroradiometric assessment of canopy chlorophyll  content in winter wheat. Computers and Electronics in Agriculture, 73(2): 165-173. Axrup, L., Markides, K., & Nilsson, T. (2000). Using miniature diode array NIR spectrometers  for analysing wood chips and bark samples in motion. Journal of Chemometrics, 14(5‐6),  561-572.                                                                                                                            69  Barton, F. E. II. (2004). Progress in near infrared spectroscopy the people, the instrumentation,  the applications. Pp. 13–18 in Davies, A.M.C.; Garrido-Varo, A. (Ed.) “Near Infrared  Spectroscopy: Proceedings of the 11th International Conference, Córdoba, Spain. NIR  Publications, Chichester.                                                                                                   Ben-Dor, E., Inbar, Y., & Chen, Y. (1997). The reflectance spectra of organic matter in the  visible near-infrared and short-wave infrared region (400–2500 nm) during a controlled decomposition process. Remote Sensing of Environment, 61(1): 1-15.                           Ben-Gera, I. & Norris, K. H. (1968a). Determination of moisture content in soybeans by direct  spectrophotometry. Israel Journal of Agricultural Research, 18: 125-132.                 Bennett, K. P. & Embrechts, M. J. (2003). An optimization perspective on kernel partial least  squares regression. NATO Science Series sub series III computer and systems  sciences, 190: 227-250.                                                                                                                              Blanco, M. & Villarroya, I. (2002). NIR spectroscopy: A rapid-response analytical tool. Trends  in Analytical Chemistry, 21(4): 240-250.                                                                               Bodig, J. & Jayne, B. A. (1982). Mechanics of wood and wood composites (Volume. 712). New  York: Van Nostrand Reinhold.                                                                                                Bokobza, L. (2002). Origin of Near-Infrared Absorption Bands, in Near Infrared Spectroscopy,  H. W. Siesler, Y. Ozaki, S. Kawata, and H. M. Heise, Eds. (Wiley-VCH, Weinheim,  Germany), p. 11.                                                                                                                                                                                                                                                                                                                                                                                                                                          Brashaw, B., Bucur, V., Gonçalves, R., Lu, J., Meder, R., Pellerin, R., et al. (2009). Non- destructive Testing and Evaluation of Wood: A Worldwide Research Update. Forest  Products Journal, 59: 7-14.                                                                                                      Brown, D. J., Bricklemyer, R. S., & Miller, P. R. (2005). Validation requirements for diffuse  reflectance soil characterization models with a case study of VNIR soil C prediction in  Montana. Geoderma, 129(3-4): 251-267.                                                                                  Brunet, D., Barthès, B., Chotte, Jean-Luc., & Feller, C. (2007). Determination of carbon and  nitrogen contents in Alfisols, Oxisols and Ultisols from Africa and Brazil using NIRS  analysis: Effects of sample grinding and set heterogeneity. Geoderma, 139: 106-117.                                                                                       Bucur, V. (2003). Non-destructive Characterization and Imaging of Wood. Springer Series  in Wood Science.                                                                                                                   Cawley, G. & Talbot, N. (2010). On Over-fitting in Model Selection and Subsequent Selection  Bias in Performance Evaluation. Journal of Machine Learning Research, 11: 2079-2107.                                                                                                                          Cherry, M. L., Vikram, V., Briggs, D., Cress, D. W., & Howe, G. T. (2008). Genetic variation in  direct and indirect measures of wood stiffness in coastal Douglas-fir. Canadian Journal  of Forest Research, 38(9): 2476-2486.                                                                                      Chu, S., Graybeal, J. D., Stoner, J. O., & Hurst, G.S. (2018). Spectroscopy. Encyclopedia  Britannica Inc. Retrieved from https://www.britannica.com/science/spectroscopy (Date  Accessed: February 05, 2020).                                                                                                        CIE. (1987). International lighting vocabulary. CIE Publication, 17.4, Central Bureau of the CIE,  Vienna. ISBN 978-3-900734-07-7.                                                                                       Ciurczak, E. W. (1992). Analytical applications of near infrared spectroscopy. Chemtech, 22:  374-380.                                                                                                                                       Coast Forest Products Association. (2003). Hem-Fir, British Columbia’s most abundant coastal  species.                                                                                                                           Coates, J. (2000). Interpretation of infrared spectra, a practical approach. In Encyclopedia of  Analytical Chemistry, Meyers RA (ed.). John Wiley & Sons Ltd: Chichester; 10815-70   10837.                                                                                                                                                                                                                                                                                                        Colthup, N. B., Daly, L. H., & Wiberley, S. E. (1990). Introduction to Infrared and Raman  Spectroscopy. Third Edition. London: Academic press Ltd, 547 p.                                     Costa, E. V., Rocha, M. F., Hein, P. R., Amaral, E. A., Santos, L. M., Brandão, L. E. et al.  (2018). Influence of spectral acquisition technique and wood anisotropy on the statistics  of predictive near infrared–based models for wood density. Journal of Near Infrared  Spectroscopy, 26(2): 106-116.                                                                                     Dahlen, J., Diaz, I., Schimleck, L., & Jones, P. D. (2017). Near-infrared spectroscopy prediction  of southern pine No. 2 lumber physical and mechanical properties. Wood Science and  Technology, 51(2): 309-322.                                                                                                      Defo, M., Taylor, A.M., & Bond, B. (2007). Determination of moisture content and density of  fresh-sawn red oak lumber by near infrared spectroscopy. Forest Products Journal, 57: 68-72.                                                                                                                               Ding, S. X., Zhang, P., Yin, S., & Ding, E. L. (2013). An integrated design framework of fault- tolerant wireless networked control systems for industrial automatic control  applications. IEEE Transactions on Industrial Informatics, 9(1): 462-471.                                                                                       de Moura Palermo, G. P., de Figueiredo Latorraca, J. V., de Moura, L. F., Nolasco, A. M.,  Monteiro de Carvalho, A., & Garcia, R. A. (2014). Surface roughness of heat-treated  Eucalyptus grandis wood. Maderas. Ciencia y tecnología, 16(1), 03-12.                                                                                                                   dos Santos, L. M., Amaral, E. A., Nieri, E. M., Costa, E. V., Trugilho, P. F., Calegário, N. et al.  (2020). Estimating wood moisture by near infrared spectroscopy: Testing  acquisition  methods and wood surfaces qualities. Wood Material Science and Engineering, 1-8. Douglas-fir Species Profile (n.d). Retrieved July 02, 2019 from:  https://www.naturallywood.com/resources/douglas-fir-species-profile                                   Druckenmüller, K., Günther, K., & Elbers, G. (2018). Near-infrared spectroscopy (NIRS) as a  tool to monitor exhaust air from poultry operations. Science of The Total  Environment, 630: 536-543.                                                                                                    Dunn, D. (1992). A preliminary assessment of the Metriguard 239A stress wave timer.  Dissertation, University of Canterbury, New Zealand.                                                            Elliot, A. & Ambrose, E. (1950). Nature, Structure of Synthetic Polypeptides, 165, 921; D.L.  Woernley, Infrared Absorption Curves for Normal and Neoplastic Tissues and Related  Biological Substances, Current Research, 12: 516 p.                                                        Eriksson L., Johansson, E., Kettaneh-Wold, N., & Wold, S. (2001). Multi- and megavariate  data analysis. Principles and Applications. Umetrics Academy, Umeä, Sweden. Esbensen, K. H., Guyot, D., Westad, F., & Houmoller, L. P. (2002). Multivariate data analysis:  in practice: an introduction to multivariate data analysis and experimental design.  Multivariate Data Analysis.                                                                                             Ethington, R. L., Eskelsen, V., & Rakesh, G. (1996). Relationship between compression strength  perpendicular to grain and ring orientation. Forest Products Journal, 46(1): 84-86. Everard, C. D., Fagan, C. C., & McDonnell, K. P. (2012). Visible-near infrared spectral sensing  coupled with chemometric analysis as a method for on-line prediction of milled biomass  composition pre-pelletising. Journal of Near Infrared Spectroscopy, 20(3): 361-369. Fabris, S. (2000). Influence of cambial ageing, initial spacing, stem taper and growth rate on  wood quality of three coastal conifers. Faculty of Graduate Studies, Department of  Forestry, The University of British Columbia, Vancouver, 250 pp. 71  Fackler, K. & Schwanninger, M. (2010). Polysaccharide degradation and lignin modification  during brown rot of spruce wood: a polarised Fourier transform near infrared study. Journal of Near Infrared Spectroscopy, 18(6): 403-416.                                    Forest Products Laboratory. (2010). Wood handbook – wood as an engineering material. General  Technical Report FPL-GTR-190. U.S. Department of Agriculture, Forest Service, Forest Products Laboratory, Madison, Wisconsin, USA.                                                                                                     Fromm, J. H., Sautter, I., Matthies, D., Kremer, J., Schumacher, P., & Ganter, C. (2001). Xylem  water content and wood density in spruce and oak trees detected by high-resolution computed tomography. Plant Physiology, 127(2): 416-425.                                                  Fujimoto, T., Yamamoto, H., & Tsuchikawa, S. (2007). Estimation of Wood Stiffness and  Strength Properties of Hybrid Larch by Near-Infrared Spectroscopy. Applied  Spectroscopy, 61(8): 882-888.                                                                                                                                      Fujimoto, T., Kurata, Y., Matsumoto, K., & Tsuchikawa, S. (2008). Application of near infrared  spectroscopy for estimating wood mechanical properties of small clear and full-length  lumber specimens. Journal of Near Infrared Spectroscopy, 16(6): 529-537.                   Fujimoto, T., Kobori, H., & Tsuchikawa, S. (2012). Prediction of wood density independently of  moisture conditions using near infrared spectroscopy. Journal of Near Infrared  Spectroscopy, 20(3): 353–359.                                                                                                      Galligan, W. L., Snodgrass, D. V., & Crow, G. W. (1977). Machine stress rating: practical  concerns for lumber producers. General Technical Report. FPL–GTR–7. Madison, WI:  U.S. Department of Agriculture, Forest Service, Forest Products Laboratory.                Galvão, R. K., Araujo, M. C., José, G. E., Pontes, M. J., Silva, E. C., & Saldanha, T. C. (2005).  A method for calibration and validation subset partitioning. Talanta, 67(4): 736-740.                                                                  Gartner, B. L., North, E. M., Johnson, G. R., & Singleton, R. (2002). Effects of live crown on  vertical patterns of wood density and growth in Douglas-fir. Canadian Journal of Forest  Research, 32(3): 439-447.                                                                                              Gaspar, F., Lopes, J., Cruz, H., Schwanninger, M., & Rodrigues, J. (2009). Application of near  infrared spectroscopy and multivariate data analysis for the evaluation of glue lines of  untreated and copper azole treated laminated timber before and after ageing. Polymer  degradation and stability, 94(7): 1061-1071.                                                                         Gierlinger, N., Schwanninger, M., & Wimmer, R. (2004). Characteristics and classification of  Fourier-transform near infrared spectra of the heartwood of different larch species (Larix  sp.). Journal of Near Infrared Spectroscopy, 12(2): 113-119.                                                       Gindl, W., Teischinger, A., Schwanninger, M., & Hinterstoisser, B. (2001). The relationship  between near infrared spectra of radial wood surfaces and wood mechanical properties.  Journal of Near Infrared Spectroscopy, 9(4): 255-261.                                                      Gindl, W. & Teischinger, A. (2002). The potential of Vis-and NIR-spectroscopy for the non- destructive evaluation of grain-angle in wood. Wood and Fiber Science, 34(4): 651- 656. Glass, S. V., Zelinka, S. L., & Johnson, J. A. (2014). Investigation of historic equilibrium  moisture content data from the Forest Products Laboratory. USDA Forest Service, Forest  Products Laboratory, General Technical Report, FPL-GTR-229, 2014; 37 p., 229, 1-37.                                                                                                                                                                                                                                                                                              Gonzalez, J. S. (1990). Wood density of Canadian tree species. Information Report NOR-X-315.  Forestry Canada, Northwest Region, Northern Forestry Centre, Edmonton, Alberta.                                                                                                     Goodman, B.A. (1994). Molecular spectroscopy: introduction and general principles. In: 72   Wilson M.J. (eds) Clay Mineralogy: Spectroscopic and Chemical Determinative  Methods. Springer, Dordrecht.  Gowen, A. A., Downey, G., Esquerre, C., & O'Donnell, C. P. (2011). Preventing over‐fitting in  PLS calibration models of near‐infrared (NIR) spectroscopy data using regression  coefficients. Journal of Chemometrics, 25(7): 375-381.                                                                                                        Güntekin, E. & Aydin, T. Y. (2013). Effects of moisture content on some mechanical properties  of Turkish Red pine (Pinus brutia Ten.). In Proceedings of the International Caucasian  Forestry Symposium (Artvin, Turkey) (pp. 878-883).                                                                Hans, G. (2014). Real-time and non-destructive characterization of wood log properties using  near-infrared and ground penetrating radar sensors. PhD Thesis, University of News  Brunswick, Canada. Retrieved from:  https://unbscholar.lib.unb.ca/islandora/object/unbscholar%3A9453                       Haartveit, E. Y. & Flæte, P. O. (2002). Mechanical properties of Norway spruce lumber from  monocultures and mixed stands—modelling bending stiffness and strength using stand  and tree characteristics. Proceedings of the Connection between Silviculture and Wood  Quality: Modelling Approach and Simulation Software, IUFRO WP S, 5, 01-04. Haartveit, E. Y. & Flæte, P. O. (2006). Rapid prediction of basic wood properties by near  infrared spectroscopy. New Zealand Journal of Forestry Science, 36(2/3): 393.  Hauksson, J., Bergqvist, G., Bergsten, U., Sjostrom, M., & Edlund, U. (2001). Prediction of  basic wood properties for Norway spruce. Interpretation of Near Infrared Spectroscopy  data using partial least squares regression. Wood Science and Technology, 35(6): 475- 485.                                                                                                                                   Haygreen, J.G. & Bowyer, J.L (1989). Forest Products and Wood Science, 2nd edition. Iowa  State University Press, Iowa, USA. 500 p.                                                                                                       Haygreen J.G. & Bowyer J.L. (1996). Forest Products and Wood Science: An Introduction, 3rd  Edition. Iowa State University Press, Iowa, USA (1996).                                             Hebda, R. (1995). Natural History of Douglas-fir (Pseudotsuga menziesii).  Retrieved June 22,  2019 from: http://rbcm1.rbcm.gov.bc.ca/nh_papers/nativeplants/pseumenz.html   Hedrick, S. E., Bennett, R. M., Kelley, S. S., & Rials, T. G. (2004). Determination of Wood   Properties Using Near Infrared Spectroscopy. In Structures 2004: Building on the Past,  Securing the Future (pp. 1-8).                                                                                           Hein, P. R., Lima, J. T., & Chaix, G. (2009a). Robustness of models based on near infrared  spectra to predict the basic density in Eucalyptus urophylla wood. Journal of near  infrared spectroscopy, 17(3): 141-150.                                                                                    Hein, P. R., Campos, A. C., Lima, J. T., Trugilho, P. F., & Chaix, G. (2009b). Estimativa da  resistência e da elasticidade à compressão paralela às fibras da madeira de Eucalyptus  grandis e. urophylla usando a espectroscopia no infravermelho próximo. Scientia  Forestalis, 37(82): 119-129.                                                                                                        Hein, P. R. (2010). Multivariate regression methods for estimating basic density in Eucalyptus  wood from near infrared spectroscopic data. Cerne, 16: 90-96.                                           Hein, P. R., Campos, A. C., Mendes, R. F., Mendes, L. M., & Chaix, G. (2011). Estimation of  physical and mechanical properties of agro-based particleboards by near infrared  spectroscopy. European Journal of Wood and Wood Products, 69(3): 431-442.            73  Henze, K. J. (2006). Visualization of the spatial variation of wood density in Western  Hemlock (MSc dissertation), The University of British Columbia, Vancouver, BC,  Canada.                                                                                                                                        Hermann, R. K. (1982). The Genus Pseudotsuga: Historical Records and Nomenclature.  Corvallis, Oregon. Forest Research Laboratory.                                                                   Hermann, R. K. & Lavender, D. P. (1999). Pseudotsuga menziesii (Mirb.) Franco. Retrieved  June 22, 2019:  http://willow.ncfes.umn.edu/silvics_manual/Volume_1/pseudotsuga/menziesii.htm Hively, W., Reeves, J., Lang, M., Oesterling, R., & Delwiche, S. (2011). Use of Airborne  Hyperspectral Imagery to Map Soil Properties in Tilled Agricultural Fields. Applied and  Environmental Soil Science, 10: 1155/2011/358193.                                              Hoadley, R. B. (1990). Identifying wood, accurate results with simple tools. Newtown, Taunton  Press.                                                                                                                                 Hoffmeyer, P. & Pedersen, J. G. (1995). Evaluation of density and strength of Norway spruce  wood by near infrared reflectance spectroscopy. Holz als Roh-und werkstoff, 53(3): 165- 170.                                                                                                                                    Ibrahim, B. & Ayşenur, K. A. (2016). Determination of the Relationships between Site Index and  Annual Ring of Turkish Calabrian Pine Wood. 1st International Mediterranean Science  and Engineering Congress (IMSEC 2016), Çukurova University, Congress Center October 26-28, 2016, Adana / TURKEY. Pp. 1742-1746.                                                       Inagaki, T., Schwanninger, M., Kato, R., Kurata, Y., Thanapase, W., Puthson, P. et al. (2010).  Eucalyptus camaldulensis density and fiber length estimated by near-infrared  spectroscopy. Wood Science and Technology, 46(1): 143–155.                                                                                                                                     Ingemi, C. M. & Yu, T. (2019). Detection of grain angle in wood specimens using synthetic  aperture radar imaging.  In Non-destructive Characterization and Monitoring of  Advanced Materials, Aerospace, Civil Infrastructure and Transportation XIII (Vol.  10971, p. 109710U). International Society for Optics and Photonics.                                 Johnson, G. R. & Gartner, B. L. (2006). Genetic variation in basic density and modulus of  elasticity of coastal Douglas-fir. Tree Genetics & Genomes, 3(1): 25-33.                           Jolliffe, I. & Cadima, J. (2016). Principal component analysis: A review and recent  developments. Philosophical Transactions of the Royal Society A: Mathematical,  Physical and Engineering Sciences, 374, 20150202.                                                                Jones, P. D., Schimleck, L. R., Peter, F. G., Daniels, F. R., & Clark, A. (2005a). Non-destructive  estimation of Pinus taeda L. tracheid morphological characteristics for samples from a  wide range of sites in Georgia. Wood Science and Technology, 39(7): 529–545.                    Jones, P. D., Schimleck, L. R., So, C. L., Clark III, A., & Daniels, R. F. (2007). High resolution  scanning of radial strips cut from increment cores by near infrared spectroscopy. IAWA  Journal, 28(4): 473-484.                                                                                                             Jozsa, L. A. & Kellogg, R. M. (1986). An Exploratory Study of the Density and Annual Ring  Weight Trends in Fast-growth Coniferous Woods in British Columbia. Forintek Canada  Corporation.                                                                                                                                Jozsa, L. A., Richards, J., & Johnson, S. G. (1989). Relative density. In: R.M. Kellogg (Editor),  Second Growth Douglas- Fir: Its Management and Conversion for Value. Forintek 74   Canada Corporation. Vancouver, B.C., Special Publication. SP-32, 5, p. 22.                   Jozsa, L.A., Munro, B.D., & Gordon, J.R. (1998). Basic wood properties of second-growth  western hemlock. B.C Ministry of Forests Special Publication. No SP-38.                         Kalivas, J. H. & Gemperline, P. J. (2006). Calibration, in Practical guide to chemometrics,  ed. P. J. Gemperline (Boca Raton: CRC, Taylor and Francis group, 2006), chapter. 6,  pp.105-166.                                                                                                                                                                  Karlman, L., Tommy, M., & Owe, M. (2005). Wood Density, Annual Ring Width and Latewood  Content in Larch and Scots Pine. Eurasian Journal of Forest Research, 8(2): 91-96.                Kelley, S., Rials, G. T., Groom, L. R., & So, C. L. (2004a). Use of near infrared spectroscopy to  predict the mechanical properties of six softwoods. Holzforschung, 58(3): 252-260. Kelley, S., Rials, T. G., Snell, R., Groom, L. H., & Sluiter, A. (2004b). Use of near infrared  spectroscopy to measure the chemical and mechanical properties of solid wood. Wood  Science and Technology, 38(4): 257-276.                                                                                  Kenneth, B. J. (2001). Long term Patterns of Douglas-fir and Western hemlock Mortality in the  Western Cascade Mountains of Washington and Oregon. PhD Dissertation, University of  Oregon, USA. Retrieved from https://andrewsforest.oregonstate.edu/publications/3034 Kiaei, M., Bakhshi, R., Saffari, M., & Golkari, S. (2015). Juniperus polycarpos. Journal of forest  and environmental science, 31(4): 267-271.                                                                      Kliger, I. R., Perstorper, M., Johansson, G., & Pellicane, P.J. (1995). Quality of timber products  from Norway spruce. Wood Science and Technology, 29(3): 397-410.                                     Kobori, H., Inagaki, T., Fujimoto, T., Okura, T., & Tsuchikawa, S. (2015). Fast online NIR  technique to predict MOE and moisture content of sawn lumber. Holzforschung, 69(3):  329-335.                                                                                                                                                                              Kollmann, F. P. & Cote Jr, W. A. (1968). Principles of wood science and technology: solid  wood. Berlin: Springer- Verlag, 1: 592 p.                                                                                   Koman, S., Feher, S., Abraham, J., & Taschner, R. (2013). Effect of knots on the bending  strength and the modulus of elasticity of wood. Wood Research, 58(4): 617-626.                                                                           Kothiyal, V. & Raturi, A. (2011). Estimating mechanical properties and specific gravity for five- year-old Eucalyptus tereticornis having broad moisture content range by NIR  spectroscopy. Holzforschung, 65(5): 757-762.                                                                 Kothiyal, V., Raturi, A., & Dubey, Y. (2014). Enhancing the applicability of near infrared  spectroscopy for estimating specific gravity of green timber from Eucalyptus tereticornis  by developing composite calibration using both radial and tangential face of wood. Holz  als Roh- und Werkstoff, 72(1): 11-20.                                                                                                                                                                                                                                                                                          Lachenbruch, B., Moore, J. R., & Evans, R. (2011). Radial variation in wood structure and  function in woody plants, and hypotheses for its occurrence. In Size-and age-related  changes in tree structure and function (pp. 121-164). Springer, Dordrecht.                     Lang, E., Bejo, L., Szalai, J., & Kovacs, Z. (2000). Orthotropic strength and elasticity of  hardwoods in relation to composite manufacture. Part I. Orthotropy of shear strength.  Wood and Fiber Science, 32(4): 502-519.                                                                                  Lang, E., Bejo, L., Szalai, J., Kovacs, Z., & Anderson, R. B. (2002). Orthotropic strength and  elasticity of hardwoods in relation to composite manufacture. Part II. Orthotropy of  compression strength and elasticity. Wood and Fiber Science, 34(2): 350-365.                       75  Lazarescu, C. & Avramidis, S. (2012). Heating characteristics of western hemlock (Tsuga  heterophylla) in a high frequency field. European Journal of Wood and Wood Products,  70(4): 489-496.                                                                                                                 Leavengood, S. A. (1998). Identifying common northwest wood species: a woodworker's guide.                                                                                                                 Leardi, R., Seasholtz, M. B., & Pell, R. J. (2002). Variable selection for multivariate calibration  using a genetic algorithm: prediction of additive concentrations in polymer films from  Fourier transform-infrared spectral data. Analytica Chimica Acta, 461(2): 189-200. Leblon, B., Adedipe, O., Guillaume, H., Haddadi, A., Tsuchikawa, S., Burger, J. et al. (2013). A  review of near-infrared spectroscopy for monitoring moisture content and density of solid  wood. Forestry Chronicle, 89(5): 595-606.                                                                                           Liang, H., Cao, J., Tu, W., Lin, X., & Zhang, Y. (2016). Non-destructive determination of the  compressive strength of wood using near-infrared spectroscopy. Bioresources, 11(3),  7205-7213.                                                                                                                                   Liu, Z., Liu, Y., Yu, H., & Juan, J. (2006). Measurement of the dynamic modulus of elasticity of  wood panels. Frontiers of Forestry in China, 1(4): 425-430.                                                  Li, P., Zhang, F. H., Li, X. Y., Zhang, Z., & Zhang J. (2010). Use of BP neural network in near  infrared spectroscopy calibrations for predicting of wood density, Advanced Materials  Research, 129-131: 306-311.                                                                                                    Louzada, J. L. (2003). Genetic correlations between wood density components in Pinus pinaster  Ait. Annals of Forest Science, 60(3): 285-294.                                                                              Malik, S., Ghodgaonkar, D., Hambaly, A., Majid, W., & Nuruddin, M. (2005). Measurement of  wood grain angle using free-space microwave measurement system in 8–12 GHz  frequency range. Proceedings of Asian Conference on Sensors and the International  Conference on New Techniques in Pharmaceutical and Biomedical Research, pp. 213- 218.                                                                                                                                       Martens, H. & Naes, T. (1991). Multivariate Calibration. Wiley, Chichester.  419 p.                    Mcdonald, K. A. & Bendtsen, B. A (1986). Measuring localized slope of grain by electrical  capacitance. Forest Products Journal, 36(10): 75-78.                                                           Mehmood, T., Liland, K. H., Snipen, L., & Sæbø, S. (2012). A review of variable selection  methods in partial least squares regression. Chemometrics and Intelligent Laboratory  Systems, 118: 62-69.                                                                                                                  Michell, A. J. & Schimleck, L. R. (1996). NIR spectroscopy of woods from Eucalyptus  globulus. Appita Journal, 49(1): 23-26.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  Middleton, G. R. & Munro, B. D. (Eds.). (2001). Second-growth western hemlock: product  yields & attributes related to stand density. Forintek Canada Corporation.                               Missanjo, E. & Matsumura, J. (2016). Wood density and mechanical properties of Pinus kesiya  Royle ex Gordon in Malawi. Forests, 7(7): 135.                                                                       Mitsui, K., Inagaki, T., & Tsuchikawa, S. (2008). Monitoring of hydroxyl groups in wood during  heat treatment using NIR spectroscopy. Biomacromolecules, 9(1): 286-288.                 Mollaoglu, A. D., Ozyurt, I., & Severcan, F. (2018). Applications of Infrared Spectroscopy and  Microscopy in Diagnosis of Obesity. In Infrared Spectroscopy-Principles, Advances, and  Applications. IntechOpen.                                                                                                            Mora, C. R., Schimleck, L. R., & Isik, F. (2008). Near infrared calibration models for the 76   estimation of wood density in Pinus taeda using repeated sample measurements. Journal  of Near Infrared Spectroscopy, 16(6): 517-528.                                                             Mora, C. & Schimleck, L. (2010). Kernel regression methods for the prediction of wood  properties of Pinus taeda using near infrared spectroscopy. Wood Science and  Technology, 44: 561-578.                                                                                                       Mouazen, A. M., Saeys, W., Xing, J., De Baerdemaeker, J., & Ramon, H. (2005). Near infrared  spectroscopy for agricultural materials: an instrument comparison. Journal of Near  Infrared Spectroscopy, 13(2): 87-97.                                                                                       Mousavi, S.Y. (2016). Performing different mechanical tests on Douglas-fir wood. Retrieved  from: http://works.bepress.com/seyyedyahya-mousavi/6/                                                        Naes, T., Irgens, C., & Martens, H. (1986). Comparison of linear statistical methods for  calibration of NIR instruments. Journal of the Royal Statistical Society Series C, Applied  Statistics, 35(2): 195-206.                                                                                                 Naito, S., Fuiii, Y., Sawada, Y., & Okumura, S. (2000). Thermographic measurement of slope of  grain using the thermal anisotropy of wood. Mokuzai Gakkaishi, 46(4): 320-325. Nathalie, L. & Malo, K. A. (2010). Effect of annual ring patterns on Norway spruce resulting  material properties. 11th World Conference on Timber Engineering 2010, WCTE 2010.                                                                                                                             Natural Resources Canada (2015). Western hemlock distribution map. Retrieved April 20, 2019,  from: https://tidcf.nrcan.gc.ca/en/trees/factsheet/119                                                    Nguyen, T. T., Ji, X., Van Nguyen, T. H., & Guo, M. (2017). Wettability modification of heat- treated wood (HTW) via cold atmospheric-pressure nitrogen plasma jet  (APPJ). Holzforschung, 72(1): 37-43.                                                                                           Niemz, P. & Mannes, D. (2012). Non-destructive testing of wood and wood-based materials.  Journal of Cultural Heritage, 13(3): S26-S34.                                                                            Nourian, S. (2018). Thermal modification of Western hemlock (Tsuga heterophylla).                      MSc thesis. The University of British Columbia. Retrieved from  https://open.library.ubc.ca/cIRcle/collections/ubctheses/24/items/1.0368548                         Olsson, A., Oscarsson, J., Serrano, E., Kallsner, B., Johansson, M., & Enquist, B. (2013).  Prediction of timber bending strength and in-member cross-sectional stiffness variation  on the basis of local wood fibre orientation. European Journal of Wood and Wood  Products, 71(3): 319–333.                                                                                                                                                                                                                 Osborne, N. L., Høibø, Ø. A., & Maguire, D. A. (2016). Estimating the density of coast Douglas- fir wood samples at different moisture contents using medical X-ray computed  tomography. Computers and Electronics in Agriculture, 127: 50-55.                                     Ozdemir, T. & Hiziroglu, S. (2007). Evaluation of surface quality and adhesion strength of  treated solid wood. Journal of Materials Processing Technology, 186(1-3): 311-314.                                                                                                                                Parish, R. (1995). Tree book- Learning to recognize trees in British Columbia, B.C. Ministry of Forests, 183 p.                                                                                                      Pearson, K. (1901). On lines and planes of closest fit to systems of points in space.  Philosophical Magazine, 2: 559–572. (doi:10.1080/14786440109462720).                                                                                                                                                                                                                                            Piazza, M. & Riggio, M. (2008). Visual strength-grading and NDT of timber in traditional  structures. Journal of Building Appraisal, 3(4): 267-296.                                                          77  Pojar, J., Klinka, K., & Demarchi, D. A. (1991). Coastal Western hemlock zone. In Ecosystems  of British Columbia. Special Reports Series 6, BC Ministry of Forests, Victoria, BC,  Canada. pp. 95-111.                                                                                                                     Press, W., Teukolsky, S., Vetterling, W., & Flannery, B. (1992). Numerical Recipes in C, 2nd  edition, Cambridge University Press, New York.                                                                                                                                                                                           Rathgeber, C. B., Decoux, V., & Leban, J. M. (2006). Linking intra-tree-ring wood density  variations and tracheid anatomical characteristics in Douglas-fir (Pseudotsuga menziesii  (Mirb.) Franco). Annals of Forest Science, 63(7): 699-706.                                                    Rials, T., Kelley, S., & So, C.-L. (2002). Use of advanced spectroscopic techniques for  predicting the  mechanical properties of wood composites. Wood and Fiber Science, 34  (3): 398-407.                                                                                                                           Ribeiro, P. G. (2009). Utilização de técnicas alternativas para caracterização tecnológica de  madeiras. 2009. 108 p. Dissertação (Mestrado em Engenharia Florestal) – Universidade  de Brasília, Brasília.                                                                                                                  Rinnan, Å., Van Den Berg, F., & Engelsen, S. B. (2009). Review of the most common pre- processing techniques for near-infrared spectra. Trends in Analytical  Chemistry, 28(10): 1201-1222.                                                                                                Rinnan, Å. (2014). Pre-processing in vibrational spectroscopy – when, why and how. Analytical  Methods, 6(18): 7124-7129.                                                                                                                                                                                                                                                                                                                         Rohrbach, K. (2008). Schedule and post-drying storage effects on Western Hemlock squares  quality (MSc dissertation), The University of British Columbia, Vancouver, BC,  Canada.                                                                                                                         Romer, J. (2013). Predicting mechanical properties of southern pine lumber with non-destructive  measurements, MSc Thesis, Louisiana State University and Agricultural and Mechanical  College, Louisiana.                                                                                                            Rosipal, R. & Trejo, L. J. (2001). Kernel partial least squares regression in reproducing kernel  Hilbert space. Journal of Machine Learning Research, 2: 97-123.                                Ross, R., Brashaw, B., & Pellerin, R. (1998). Non-destructive evaluation of wood. Forest  Products Journal, 48: 101-105.                                                                                             Ross, R.J. (2015). Non-destructive evaluation of wood. 2nd edition. Gen. Tech. Rep. FPL–GTR– 238. Madison, WI: U.S. Department of Agriculture, Forest Service, Forest Products  Laboratory: 167 p.                                                                                                                         Sahin, D. (2019). Atomic Spectroscopy. Intech Open, pp. 1-11.                                                          Samson, M., Tremblayan, C., & Langlais, D. A. (1993). Measuring slope of grain by electrical  capacitance at moisture contents above fiber saturation. Forest Products Journal, 43(2):  58-60.                                                                                                                                            Sandak, A. (2013). Establishing NIR measurement protocol. SLOPE Project, Deliverables 4.03, FP7-NMP-2013-SME-7—604129.                                                                             Sandak, J., Sandak, A., & Meder, R. (2016). Assessing trees, wood and derived products with  NIR spectroscopy: Hints and tips. Journal of Near Infrared Spectroscopy, 24: 485-505. Santos, A. J., Alves, A. M., Simões, R. M., Pereira, H., Rodrigues, J., & Schwanninger, M.  (2012). Estimation of wood basic density of Acacia melanoxylon (R. Br.) by near  infrared spectroscopy. Journal of Near Infrared Spectroscopy, 20(2): 267-274.                   78  Savitzky, A. & Golay, M. J. (1964). Smoothing and differentiation of data by simplified least  squares procedures. Analytical chemistry: 36(8): 1627-1639.                                                     Schajer, G. S. & Orhan, F. B. (2006). Measurement of wood grain angle, moisture content and  density using microwaves. Holz als Roh-und Werkstoff, 64(6): 483-490.                                    Schimleck, L. R., Wright, P. J., Michell, A. J., & Wallis, A. A. (1997). Near-infrared spectra and  chemical compositions of E. globulus and E. nitens plantation woods. Appita  Journal, 50(1): 40-46.                                                                                                         Schimleck, L. R., Michell, A. J., Raymond, C. A., & Muneri, A. (1999). Estimation of basic  density of Eucalyptus globulus using Near-Infrared Spectroscopy. Canadian Journal of Forest Research, 29(2): 194-201.                                                                          Schimleck, L. R., Evans, R., & Ilic, J. (2001). Estimation of Eucalyptus delegatensis wood  properties by near infrared spectroscopy. Canadian Journal of Forest Research, 31(1): 671-1675.                                                                                                                     Schimleck, L. R., Evans, R., & Matheson, A. C. (2002). Estimation of Pinus radiata D. Don  clear wood properties by near-infrared spectroscopy. Journal of Wood Science, 48(2):  132-137.                                                                                                                          Schimleck, L. R. & Evans, R. (2003). Estimation of air-dry density of increment cores by near  infra-red spectroscopy. Appita Journal, 56(4): 312–317.                                                    Schimleck, L., Doran, J., & Rimbawanto, A. (2003). Near Infrared Spectroscopy for Cost  Effective Screening of Foliar Oil Characteristics in a Melaleuca cajuputi Breeding  Population. Journal of Agricultural and Food Chemistry, 51(9): 2433-2437.                    Schimleck, L. R., Stürzenbecher, R., Mora, C., Jones, P., & Daniels, R. (2005a). Comparison of  Pinus taeda L. wood property calibrations based on NIR spectra from the radial- longitudinal and radial-transverse faces of wooden strips. Holzforschung, 59: 214-218. Schimleck, L. R., Jones, P.D., Clark, A., Daniels, R. F., & Peter, G. F. (2005b). Near infrared  spectroscopy for the non-destructive estimation of clear wood properties of Pinus taeda  L. from the Southern USA. Forest Products Journal, 55(12): 21-28.                                Schimleck, L., Hodge, G., & Woodbridge, W. (2010). Toward global calibrations for estimating  the wood properties of tropical, sub- tropical and temperate pine species. Journal of Near  Infrared Spectroscopy, 18(6): 355-365.                                                                             Schimleck, L., Matos, J. L., Trianoski, R., & Prata, J. (2018). Comparison of Methods for  Estimating Mechanical Properties of Wood by NIR  Spectroscopy. Journal of  Spectroscopy, 218.                                                                                                                                                                                                             Schwanninger, M., Hinterstoisser, B., Gierlinger, N., Wimmer, R., & Hanger, J. (2004).  Application of Fourier Transform Near Infrared Spectroscopy (FT-NIR) to thermally  modified wood. Holz Als Roh-Und Werkstoff, 62(6): 483-485.                                    Schwanninger, M., Rodrigues, J. C., & Fackler, K. (2011). A review of band assignments in  near infrared spectra of wood and wood components. Journal of Near Infrared  Spectroscopy, 19(5): 287- 308.                                                                                                                                             Sepulveda, P. (2001). Measurement of spiral grain with computed tomography. Journal of Wood  Science, 47(4): 289-293.                                                                                                            Seni, G. & Elder, J.F. (2010). Ensemble methods in data mining: improving accuracy through  combining predictions. Synthesis lectures on data mining and knowledge discovery, 2(1):79   1-126.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 Shahverdi, M. (2015). Kiln drying optimization for quality hem-fir lumber (Msc dissertation,  University of British Columbia).                                                                                           Sharp, D.J. (1985). Non-destructive testing techniques for manufacturing LVL and predicting  performance. In: Proceedings, 5th non-destructive testing of wood symposium; 1985  September 9–11; Pullman, WA. Pullman, WA: Washington State University: 99–108. Skaar, C. (1988). Wood-water relations Springer, Berlin, Heidelberg, New York. Pages 22-31. Sliney D. H. (2016). What is light? The visible spectrum and beyond. Eye (London,  England), 30(2), 222–229. doi:10.1038/eye.2015.252.                                                                 Smith, G. H. (2006). Camera Lenses: From Box Camera to Digital (SPIE Press Monograph  Vol. PM158) (Press Monograph). SPIE-International Society for Optical Engineering. So, C. L., Via, B. K., Groom, L. H., Schimleck, L. R., Shupe, T. F., Kelley, S. S. et al. (2004).  Near infrared (NIR) spectroscopy in the forest products industry. Forest Products  Journal, 54(3): 8–18.                                                                                                                        Softwood Export Council (SEC). (2004). Hem-fir. Retrieved April 10, 2018, from:  http://www.softwood.org/publications-library/category/hemfir.                                          Softwood Species Profile (n.d). Retrieved June 10, 2020 from:  https://www.naturallywood.com/forest-products-species/softwood-species                                                                                                                                                                    Spectroscopy, 101. (2017). Introduction to Spectroscopy and its applications. Ocean Optics, pp.  1 -16.                                                                                                                                       Stenberg, B. & Viscarra, R. (2010). Diffuse reflectance spectroscopy for high-resolution  soil  sensing, in Proximal Soil Sensing, R. A. Viscarra Rossel, B. A. McBratney, and B.  Minasny, Eds., pp. 29–47, Springer Science+Business, Dordrecht, The Netherlands. Stevens, A., Udelhoven, T., Denis, A., Tychon, B., Lioy, R., Hoffmann, L. et al. (2010).  Measuring soil organic carbon in croplands at regional scale using airborne imaging  spectroscopy. Geoderma, 158(1-2): 32-45.                                                                         Swierenga, H., Groot, P. J., Weijer, A. P., Derksen, M. W., & Buydens, L. (1998).  Improvement  of PLS model transferability by robust wavelength selection.  Chemometrics and Intelligent Laboratory Systems, 41(2): 237-248.                                                                                                                                                                                                     Takata, K., Koizumi, A., Yamashita, K., & Nakada, R. (2002). Mechanical and anatomical  properties of Siberian larch grown in southern-central Siberia. In Improvement of larch  for better growth, stemform and wood quality, INRA, 396-404.                                                  Theophanides, T. (2012). Introduction to Infrared Spectroscopy. Materials Science, Engineering  and Technology, Rijeka, Croatia, Intech Publications, 10.5772/49106.                                     Thumm, A. & Meder, R. (2001). Stiffness prediction of radiata pine clearwood test pieces using  near infrared spectroscopy. Journal of Near Infrared Spectroscopy, 9(2): 117-122. Thygesen, L. G. (1994). Determination of dry matter content and basic density of Norway spruce  by near infrared reflectance and transmittance spectroscopy. Journal of Near Infrared  Spectroscopy, 2(3): 127-135.                                                                                              Thygesen, L. G. & Lundqvist, S. O. (2000). NIR measurement of moisture content in wood  under unstable temperature conditions. Part 1. Thermal effects in near infrared spectra  of wood. Journal of Near Infrared Spectroscopy, 8(3): 183-189.                                       Todoroki, C. L., Lowell, E. C., Dykstra, D. P., & Briggs, D. G. (2012). Maps and models of 80   density and stiffness within individual Douglas-fir trees. SCION. 42: 1-13.            Todorovic, N., Popovic, Z., & Milic, G. (2015). Estimation of quality of thermally modified  beech  wood with red heartwood by FT-NIR spectroscopy. Wood science and  Technology, 49(3): 527-549.                                                                                                                                                                                                                      Tounis, E. (2009). Investigation of NIR spectroscopy for identifying and sorting wood with  respect to species, moisture content, and weathering. Theses and dissertations. Ryerson  University, Toronto, Canada.                                                                                                 Tsoumis, G. (1991). Science and technology of wood: structure, properties, utilization (Vol.  115). New York: Van Nostrand Reinhold.                                                                                                                                                                                                                                                                                                                                       Tsoumis, G.T. (2020). Wood plant tissue. Retrieved from:  https://www.britannica.com/science/wood-plant-tissue/Properties-of-wood                 Tsuchikawa, S., Hayashi, K., & Tsutsumi, S. (1996). Non-destructive measurement of the  subsurface structure of biological material having cellular structure by using near-infrared  spectroscopy. Applied Spectroscopy, 50(9): 1117-1124.                                                   Tsuchikawa, S. & Siesler, H. (2003). Near-Infrared Spectroscopic Monitoring of the Diffusion  Process of Deuterium-Labeled Molecules in Wood. Part II: Hardwood. Applied  Spectroscopy, 57(6): 675-681.                                                                                                                                                   Tsuchikawa, S. (2007). A review of recent near infrared research for wood and paper. Applied  Spectroscopy Reviews, 42(1): 43–71.                                                                    Udelhoven, T., Emmerling, C., & Jarmer, T. (2003). Quantitative analysis of soil chemical  properties with diffuse reflectance spectrometry and partial least-square regression:  feasibility study, Plant and Soil, 251(2): 319–329.                                                                                                                                                                                                                          Ukrainetz, N. K., Ritland, K., & Mansfield, S. D. (2008). Identification of quantitative trait loci  for wood quality and growth across eight full-sib coastal Douglas-fir families. Tree  Genetics & Genomes, 4(2): 159-170.                                                                                       Uzan, J.P. & Leclercq, B. (2008). The Natural Laws of the Universe: Understanding  Fundamental Constants. 10.1007/978-0-387-74081-2.                                                         Vasques, G. M., Grunwald, S., & Sickman, J. O. (2008).  Comparison of multivariate methods  for inferential modeling of soil carbon using visible/near-infrared spectra, Geoderma,  146(1- 2): 14-25.                                                                                                                       Via, B. K., Shupe, T. F., Groom, L. H., Stine, M., & So, C. L. (2003). Multivariate modelling of  density, strength and stiffness from near infrared spectra for mature, juvenile and pith  wood of longleaf pine (Pinus palustris). Journal of Near Infrared Spectroscopy, 11(5):  365-378.                                                                                                                                  Viana, L. C., Trugilho, P. F., Hein, P. R., Lima, J. T., & da Silva, J. R. (2009). Predicting the  morphological characteristics and basic density of Eucalyptus wood using the NIRS  technique. Cerne, 15(4): 421-429.                                                                              Viguier, J., Jehl, A., Collet, R., Bleron, L., & Meriaudeau, F. (2015). Improving strength grading  of timber by grain angle measurement and mechanical modeling. Wood Material  Science and Engineering, 10(1): 145-156.                                                                             Viscarra, R. & Behrens, T. (2010). Using data mining to model and interpret soil diffuse  reflectance spectra, Geoderma, 158(1-2): 46-54.                                                          Vogt, F. & Tacke, M. (2002). Fast Principal Component Analysis of Large Data Sets Based on 81   Information Extraction. Journal of Chemometrics, 16(11): 562 - 575.                                   Vohland, M. & Emmerling, C. (2011). Determination of total soil organic C and hot water- extractable C from VIS-NIR soil reflectance with partial least squares regression and spectral feature selection techniques, European Journal of Soil Science, 62(4): 598–606. Wall, M.E., Rechtsteiner, A., & Rocha, L.M. (2003). Singular value decomposition and principal  component analysis. In A Practical Approach to Microarray Data Analysis, eds Berrar  DP, Dubitzky W, Granzow M (Kluwer, Boston), pp 91–109.                                                    Wang, M., Yan, G., & Fei, Z. (2015). Kernel PLS based prediction model construction and  simulation on theoretical cases. Neurocomputing, 165: 389-394.                                              Westad, F. & Marini, F. (2015). Validation of chemometric models - a Tutorial. Analytica  Chimica Acta,  893: 14-24.                                                                                             Williamson, G. B. & Wiemann, M. C. (2010). Measuring wood specific gravity…  correctly. American Journal of Botany, 97(3): 519-524.                                                                                                                                                                                               Wold, S., Sjostrom, M., & Eriksson, L. (2001). PLS-regression: a ¨ basic tool of chemometrics,”  Chemometrics and Intelligent Laboratory Systems, 58(2): 109–130.                                                                                                                      Young, T. A. (1807). Course of Lectures on Natural Philosophy and the Mechanical Arts, vol. 1.  William Savage: Bedford, UK, 1807, pp 463–464 (lecture 39).                                                     Yu, H., Zhao, R., Fu, F., Fei, B., & Jiang, Z. (2009). Prediction of mechanical properties of  Chinese fir wood by near infrared spectroscopy. Frontiers of Forestry in China, 4(3):  368-373.                                                                                                                                Yu, F., Qiu, F., & Meza, J. (2016). Design and Statistical Analysis of Mass-Spectrometry-Based  Quantitative Proteomics Data. In Proteomic Profiling and Analytical Chemistry (pp. 211- 237). Elsevier.                                                                                                                       Zhang, M., Liu, Y., & Yang, Z. (2015). Correlation of near infrared spectroscopy measurements  with the surface roughness of wood. BioResources, 10(4): 6953-6960.                                    Zhao, R.J., Huo, X.M., & Zhang, Li. (2009). Estimation of Modulus of Elasticity of Eucalyptus  Pellita Wood by Near Infrared Spectroscopy. Spectroscopy and spectral analysis, 29(9):  2392-2395.                                                                                                                                    Zhou, Z., Rahimi, S., & Avramidis, S. (2019). On-line species identification of green hem-fir  timber mix based on near infrared spectroscopy and chemometrics. European Journal of  Wood and Wood Products, 78(1): 151-160.                                                                                                                                                                                                                                                                                        Zustovic, M. (2015). Forest canopy gap size affects regeneration potential of interior Douglas- fir (Msc thesis, University of British Columbia).                                                              Zwinkels, J. (2015). Light, Electromagnetic Spectrum. Encyclopedia of Color Science and  Technology, New York.    

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0392877/manifest

Comment

Related Items