Computers and Electronics in Agriculture 127 (2016) 5055

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Estimating the density of coast Douglas-fir wood samples at different

moisture contents using medical X-ray computed tomography
Nathaniel L. Osborne a,, lav A. Hib b, Douglas A. Maguire c
a
Weyerhaeuser NR Company, Centralia, WA, USA
b
Norwegian University of Life Sciences, Akershus, Norway
c
Oregon State University, Corvallis, OR, USA

a r t i c l e i n f o a b s t r a c t

Article history: Wood density (q) is an indicator of Douglas-fir (Pseudotsuga menziesii) forest product performance and
Received 29 December 2015 past and present tree ecophysiology. Models describing the spatial variation in this wood property will
Received in revised form 1 June 2016 require a considerable sampling effort. Medical X-ray computed tomography (CT) has been identified
Accepted 2 June 2016
as one technology for rapidly estimating q of Douglas-fir wood. The density of Douglas-fir can be pre-
Available online 9 June 2016
dicted from CT Hounsfield units through a linear relationship (R2 : 96%). The moisture content of wood
samples has an additional linear effect on estimating Douglas-fir wood density (0:0015 g=cm3 ) and also
Keywords:
has a practically minor (2:8E
06 g=cm3 ), but significant, interactive relationship with CT Hounsfield
Wood density
Medical imaging
units. While the effect of moisture content explains only a small percentage of the variance in q, account-
Douglas-fir ing for this effect my be important to avoid prediction biases. Finally, X-ray tube current (mA) may also
impose a small effect (0:00003 g=cm3 ) on estimating wood density. In contrast to other factors, the fil-
tered back-projection algorithm used to produce CT scanning images does not have an strong effect on
estimating q. While it is important to account for scanner settings and moisture content, 74% of the vari-
ance in predicting q can be explained by CT Hounsfield units with 21% explained by accounting for mois-
ture content and X-ray tube current. Independent estimates of wood sample volume for validation can be
achieved in several ways, each with possible systematic biases. This experiment found volume of wood
samples conditioned to different moisture content could be estimated similarly using volumetric dis-
placement or dimension measurement by caliper. The absolute mean deviance of estimated sample vol-
ume from caliper measurement relative to volumetric displacement was 0:45 cm3 or 2.6%. CT scanning
can be used to rapidly estimate Douglas-fir at a resolution of 1-mm using unprepared samples.
2016 Elsevier B.V. All rights reserved.

1. Introduction spring, and is characterized by relatively short longitudinal tra-

Wood density (q) in coast Douglas-fir (Pseudotsuga menziesii)
results from a complex of interacting ecophysiological processes
that influence rate of meristematic cell division and the growth
in cell diameter, cell length and cell wall thickness (Larson,
1969). Wood density integrates the effect of these attributes on
important wood properties that in turn affect the performance of
most wood products. In the wood of Douglas-fir, q largely depends
on the dimensions of longitudinal tracheids, which are long,
skinny, tubular cells with pinched ends, usually arranged parallel
to the stem, branch or root axis (Ct, 1968). Within the annual
growth ring, tracheid development and final dimensions vary with
season of formation. Earlywood is formed at the onset of growth in
Corresponding author.
E-mail address: nathaniel.osborne@weyerhaeuser.com (N.L. Osborne).
cheid cells with large lumens and thin cell walls. Earlywood tran-
sitions to latewood during the latter part of the growing season
and is apparently driven by one or more environmental factors that
also synchronize with the end of terminal growth (Larson, 1969).
For example, latewood production can be related to the onset of
soil moisture deficit (Kantavichai et al., 2010). The timing of tran-
sition to latewood is probably an adaptation to water and other
environmental stresses, so this trait would be expected to exhibit
some level of genetic heritability along environmental gradients
(Rozenberg et al., 2001). Latewood is characterized by relatively
long, thick walled longitudinal tracheids with decreased cell diam-
eter and lumen size. Within an annual growth ring, q is less in ear-
lywood than latewood. Tracheid developmental pattern also varies
spatiotemporally along the longitudinal and radial direction within
the tree bole. As the longitudinal distance from the tree tip

http://dx.doi.org/10.1016/j.compag.2016.06.003
0168-1699/ 2016 Elsevier B.V. All rights reserved.
 N.L. Osborne et al. / Computers and Electronics in Agriculture 127 (2016) 5055
increases, the transition from earlywood to latewood is found to be
more abrupt in Douglas-fir (Emmingham, 1977).
The radial profile of q at breast height (1.37 m) for managed
Douglas-fir trees in the Pacific Northwest United States can be pre-
dicted partly from conventional tree dimensions and growing site
conditions (Filipescu et al., 2013; Kantavichai et al., 2010). Future
goals in modeling q for managed Douglas-fir trees are to predict
spatial variation within the tree bole in the longitudinal and radial
direction, to represent response of q to silvicultural treatment, and
to account for geographic trends induced by environmental influ-
ences. Meeting these goals may require deploying a mixture of
technologies to estimate density of a large number of wood sam-
ples. X-ray densitometry has supported the development of cur-
rent models for predicting density in Douglas-fir (Filipescu et al.,
2013). This high resolution scanning technique (usually at 25l)
will continue to support modeling of within and between growth
ring variation of q, especially to establish causal mechanisms
related to local climate and other site conditions like soil attri-
butes. Many other approaches could be deployed to estimate the
density of Douglas-fir wood, each with its own unique costs and
benefits (Wei et al., 2011). Several authors have demonstrated that
medical X-ray computed tomography (CT) can be used to rapidly
estimate the density of wood, when a 1-mm scanning resolution
is acceptable (Freyburger et al., 2009; Steffenrem et al., 2014).
The goal of this study was to provide a regionally applicable
equation to estimate Douglas-fir wood density g=cm3 based upon
imagery produced by medical CT scanning. Output of CT scanning
includes Hounsfield units (H), a transformation of a radiation
attenuation coefficient into a linear scale between distilled water
(0 m) and air (
1000 m). The goal of the study would clearly be
met most expediently if q could be estimated on unprepared
Douglas-fir wood cores, sample blocks, or whole tree sections.
Most significantly, use of unprepared samples would mean that
moisture content would not confound density estimates to the
extent that the quick and efficient potential of CT scanning would
be negated by complicated, time consuming, and expensive char-
acterization of average moisture content or spatial variation of
moisture content in larger samples. Moisture content was
expected to have a significant effect on predicting q, due to differ-
ent X-ray absorption coefficients for water and wood (Lindgren,
1991). Our work follows that of Freyburger et al. (2009) who devel-
oped an equation to predict q using wood from several tree spe-
cies, but not including Douglas-fir.
The specific objectives of the work reported here were to: (1)
test the hypothesis that CT scanning would provide a significantly
accurate estimate of Douglas-fir wood density, q; and (2) develop
an estimation equation that could be applied to a large regional
sample of Douglas-fir to create a database for modeling radial
and longitudinal variation in q under varying site conditions and
silvicultural regimes. A regional model predicting the magnitude
and spatial pattern of wood density in Douglas-fir would be useful
for marketing logs and for developing manufacturing technology
that would optimize the value recovered from a given log.
2. Methods and materials
Developing an equation to estimate Douglas-fir q from CT
Hounsfield units required four steps: (1) collecting wood core
and cube samples covering the expected range of q for intensively
managed Douglas-fir trees; (2) conditioning samples to three tar-
geted moisture contents; (3) scanning those samples using a com-
bination of CT scanner settings at each targeted moisture content;
(4) measuring sample volume and weight; and (5) converting CT
scanning images into a modeling dataset for regression analysis.
As described below, targeted moisture contents were specified to
51
ensure the desired range, but moisture content of each sample
was measured to determine its specific content at time of CT
scanning.
2.1. Collecting wood core and block samples
Thirty 1.27-cm diameter wood cores were drawn from a 30-
year old coastal Douglas-fir spacing trial established by the Stand
Management Cooperative near Corvallis, Oregon (914-Lewisburg
Saddle) (Maguire et al., 1991). Ten samples were collected from
each of three spacing blocks planted at

1
1
1
485 trees ha ; 190 trees ha , and 40 trees ha . Samples were
drawn across the widest possible range in initial spacing to repre-
sent the potentially largest range in q, owing to differences in tim-
ing of crown closure, average rate of diameter growth, degree of
stem differentiation, and onset of suppression mortality (Fig. 1).
Within each experimental plot, sample trees were randomly
selected across the diameter distribution, so that all social classes
within each spacing were represented. At breast height (1.37 m)
on each sample tree, a single 12.7-mm diameter wood core was
extracted with an increment borer from a random azimuth around
the bole.
Eighty Douglas-fir wood blocks of approximately the same
dimensions (2.5-cm3) were cut from sawn lumber to extend the
range of q for analysis (Fig. 1). The wood cubes were sawn from
several Douglas-fir studs purchased at a local hardware store and
planks from a local wood pile. The wood blocks were purposefully
selected to represent the widest possible range in Douglas-fir q,
implied by relative proportions of earlywood and latewood along
the exposed cross-sectional area. Between the 30 wood cores and
80 wood cubes, the sample q ranged from 0.40 to 0:84 g=cm3
(Table 1).
2.2. Laboratory methods
Each of the wood cores and cubes were scanned using a
TOSHIBA Aquillion medical CT scanner, located at the Oregon State
University, College of Veterinary Medicine (VETMED). For each
scan the X-ray voltage (120 kVp), slice thickness and interval (1-
mm) and pixel resolution (0.544-mm2) was fixed. The targeted
level of sample moisture content and X-ray tube current (mA)
and reconstructive backfilter algorithm used to process imagery
was varied (Table 2). As Freyburger et al. (2009) mention, filters
are provided by the CT scanner manufacturer on an as-is basis
without technical details and are designed for imaging different
parts of the human body. For each scanning session, wood samples
were arranged as accurately as possible to move through the scan-
ner with the growth rings oriented in a vertical position, i.e., with
the sample cross-section facing up. Wood cores were scanned in a
balsa wood cassette, using a method adapted from Steffenrem et al.
(2014), and wood cubes were oriented in a square matrix on a glass
plate placed on the scanner bed (Fig. 2).
The 0-percent moisture content was achieved by drying sam-
ples in an air-circulated oven set to 103  C for 24-hours (Table 1).
To maintain a 0-percent moisture content during transportation
to the medical imaging facility, samples were transferred directly
from the oven into a glass desiccator lined with silica cobbles.
The time between unloading samples from the desiccator and
completing the CT scans was around ten minutes. After the first
scanning session was completed, wood samples were conditioned
to a targeted moisture content of 10-percent. A 10-percent mois-
ture content was achieved by allowing samples to reach equilib-
rium in a climate chamber maintained at 65-percent relative
humidity and temperature of 20  C. To maintain the final moisture
condition, samples were placed in the glass desiccator without sil-
52 N.L. Osborne et al. / Computers and Electronics in Agriculture 127 (2016) 5055

Fig. 1. Douglas-fir wood cubes (A) and 12.7-mm diameter cores (B) representative of the range of densities in the modeling dataset.

Table 1 Table 2
Summary of sample density (g=cm3 ), moisture content (%) and average CT Hounsfield All combinations of targeted moisture content (%), X-ray tube current (mA), and
number after conditioning to different moisture content targets. image filter used for scanning Douglas-fir wood samples with medical X-ray
computed tomography.
Minimum Mean Standard Maximum
error Scan Targeted moisture content X-ray tube current Filter
number (%) (mA)
0% Target moisture
content 1 0 50 Lung
Moisture content 0 0 0 0 2 Bone
Density 0.4040 0.5401 0.0076 0.7369 3 80 Lung
CT Hounsfield number
657.43
500.32 8.97
261.81 4 Bone
5 200 Lung
10% Target moisture
6 Bone
content
Moisture content 5.42 8.98 0.07 10.60
7 10 50 Lung
Density 0.4278 0.5654 0.0083 0.7874
8 Bone
CT Hounsfield number
640.91
460.32 9.81
190.88
9 80 Lung
20% Target moisture 10 Bone
content 11 200 Lung
Moisture content 14.09 19.54 0.22 26.78 12 Bone
Density 0.4455 0.5898 0.0089 0.8444
CT Hounsfield number
617.31
449.27 9.80
163.12 13 20 50 Lung
14 Bone
15 80 Lung
16 Bone
ica cobbles for transportation from the climate chamber to 17 200 Lung
18 Bone
VETMED. After scanning samples at a 10-percent targeted moisture
content, the wood samples were placed in a glass desiccator with a
lining of water about 2-cm deep. The interior of the desiccator was
then allowed to maintain a relative humidity of 95-percent and its weight and volume estimated from both sample dimensions
temperature of 20  C over four weeks with the objective of raising and water displacement.
moisture content to 20-percent. As before, the time between Images (DICOM) from computed tomography were analyzed
unloading samples and completing the scans at VETMED was using the ImageJ software (Abrmoff et al., 2004). Each unique
around ten minutes. combination of CT scanning specifications (Table 2) had a stack
After each scanning session was completed, wood cores and of flat and pixelated imagery files (Fig. 2). These files were ordered
cubes were measured for weight (nearest 0.1-g) and volume (near- in sequential slices, from the sample top-to-base. The center slice,
est 0.1-mm). Sample volume was estimated using two approaches. which was consistently positioned, was used for image analysis.
The first approach was to measure the main dimensions of each For each DICOM image, a rectangle was manually positioned on
sample using electronic calipers and applying standard geometric each wood core or cube, carefully arranged to avoid inclusion of
formulas (rectangular prism or cylinder) to estimate volume airspace or fuzzy borders of a given sample. Within the sampling
(VOLc ). The second approach was to estimate volume through area, the average CT Hounsfield number was dynamically con-
water displacement. Each sample was blotted with a very small verted from a grey value using the DICOM header file and was
amount of water, dried with a cloth, submerged under water in a measured. A database for regression analysis was constructed by
volumetric cylinder and weighed on a scale to estimate displaced merging estimates of sample density, scanner settings, target
volume (VOLd ) within one minute after the samples were blotted moisture content, actual sample moisture content, and CT Houns-
with water. Wood density (q) for each sample was estimated from field numbers measured from image analysis.
 N.L. Osborne et al. / Computers and Electronics in Agriculture 127 (2016) 5055 53

above. Non-linear effects of moisture content were tested by allow-

ing a varying exponent (h) on MC, but it was not significantly differ-
ent from 1.

3. Results

On average, there was less than 0.5-cm3 absolute deviance (AD)

or
2:6% deviance (DP) between volumes estimated by caliper
dimensions and volume from water displacement (Table 3). When
sample volume was estimated from dimensions, it was always less
volume than when it was estimated from water displacement.
Deviance (D) between the two methods of volume estimation
was least when the samples were conditioned to zero-percent tar-
geted moisture content. Over half of the total explained variance in
Eq. (1) was attributed to CT Hounsfield units (H) (Table 4). Other
significant parameters (p-value / 0.05) explaining variation in
sample density were the X-ray tube current, sample moisture con-
tent, and the interaction between CT Hounsfield units and sample
Fig. 2. Flat X-ray computed tomography images for wood cores (left) and cubes
moisture content. Sample type (cube or core: Z 1 ) yielded signifi-
(right) conditioned to a targeted 10% moisture content, using a standard bone filter,
with X-ray scanner settings at 120 kVp and 80 mA. cantly different densities (q^ ) when all other predictors were held
constant, but the parameter estimate indicated that a core would
2.3. Equation development be predicted to have a density that is only 0.00734 g=cm3 less than
a sample cube (Table 4). This magnitude of difference probably has
The differences (D) between sample volumes estimated using little practical significance for sample densities that range from
sample dimensions and water displacement were compared. The about 0.40 to 0.85 g=cm3 . To gain further perspective, this differ-
purpose of making this assessment was to test for significant dif- ence translates into about 7.34 kg=m3 , or about 1.4% of average
ferences between methods, and to cross-validate estimation of Douglas-fir density if it is assumed to be about 544 kg=m3 . The
sample volume (VOLc and VOLd ) prior to equation fitting. The vali- source of this difference is unknown, but may be related to a bias
dation statistics were calculated as follows: in the volume estimation for wood cubes and cores (Table 3) or dif-
ferences in X-ray artifacts generated during scanning (see Fig. 3).
Statistic Description Definition Eq. (1) was reduced to Eq. (2) and a test was performed to check
for non-linearity in the moisture content term using a floating
D = difference = VOLc
VOLd
exponent (h). There was no evidence of non-linearity in the mois-
AD = absolute difference = jDj
ture content term (p-value for an exponent on MC was 0.466), so
DP = difference percent = D=VOLc  100
the linear function represented in Eq. (2) seemed to be the most
parsimonious model. It showed little evidence of bias, and
explained 96.6% percent of variation in sample q (Table 4). This
A series of models, both linear and non-linear, were fit using the final equation explained sample density as a function of CT Houns-
functions lm and nls in the R statistical programming language (R field units, X-ray tube current, sample moisture content and inter-
Core Team, 2015). A full linear model was specified to test for action between CT Hounsfield units and sample moisture content.
potential explanatory variables of q, interactions between vari- Most of the explained variance in Eq. (2) was attributable to CT
ables, and the relative importance (LMG) (Grmping, 2006) of each Hounsfield numbers (74%) and moisture content (12%) (Fig. 4). X-
model parameter: ray tube current explains a significant amount of the variability
q b10 b11 H b12 mA b13 MC b14 Z 1 b15 Z 2 b16 H in q, but only accounts for 0.04% of the variability in q, so has only
a minor influence on q ^ (Fig. 5).
 MC b17 H  mA 1 1

where q is the average wood density (g=cm3 ) for a given sample 4. Discussion
based on weight (g) and volume (cm3 ) estimated through volumet-
ric displacement (VOLd ), b10
b17 are parameters to be estimated Systematic biases are possible using both the water displace-
from the data, H is the average CT Hounsfield unit for a given sam- ment and caliper measurement approach to volume estimation.
ple based on analysis of a tomographic image, mA is the X-ray tube
current, MC is the sample moisture content percentage based on Table 3
weight after drying the samples in an oven at 103  C and weight Mean differences, mean absolute differences, and mean percent differences between
after conditioning the samples (not the targeted moisture content), sample volume (cm3 ) estimated by means of calipers (VOLc ) and water displacement
(VOLd ) at three targeted levels of moisture content.
Z 1 is an indicator variable set equal to one if the sample was a wood
core and zero otherwise, Z 2 is an indicator variable set equal to one Target moisture Mean Mean Mean difference
if the lung filter was used and zero otherwise, and 1 was the error content (%) difference jDifferencej percent (%)
term with 1 : N0; r1 . cm3 cm3

Eq. (1) was reduced to the following more parsimonious model Core 0
0.268 0.456
2.244
on the basis of parameter significance, variance inflation factors, 10
0.378 0.406
2.599
20
0.58 0.58
4.381
and interpretation of trends in data:
Cube 0
0.285 0.384
1.81
q b20 b21 H b22 mA b23 MC b24 H  MC 2 2 10
0.511 0.511
3.119
20
0.451 0.451
2.619
where b20
b24 are parameters to be estimated from the data, 2

0.413
2.648
was the error term with 2 : N0; r2 and all other terms are defined Overall 0.456
54 N.L. Osborne et al. / Computers and Electronics in Agriculture 127 (2016) 5055

Table 4
Parameter estimates and statistics for Eqs. (1) and (2), including relative importance
of each parameter (LMG) (Grmping, 2006) expressed as a percentage of explained
variation in sample density.

Parameter Term Estimate Standard error p-value LMG

Eq. (1)
R2 : 0:9672
b10 Intercept 0.9423378 0.0044181 <0.001
b11 H 0.0008060 0.0000091 <0.001 55.45
b12 mA 0.0000846 0.0000280 0.002 6.91
b13 MC 0.0016989 0.0002238 <0.001 8.72
b14 Z1
0.007343 0.0009575 <0.001 9.79
b15 Z2 0.0012805 0.0007413 0.084 0.02
b16 H  MC 0.0000031 0.0000005 <0.001 5.58
b17 H  mA 0.0000001 0.0000001 0.062 10.26
Eq. (2)
R2 : 0:9660
b20 Intercept 0.9564068 0.0031639 <0.001
b21 H 0.0008380 0.0000061 <0.001 74.71
b22 mA 0.0000343 0.0000057 <0.001 0.04
b23 MC 0.0015412 0.0002260 <0.001 12.95
b24 H  MC 0.0000028 0.0000004 <0.001 8.89 Fig. 4. Response surface of sample density q predicted by Eq. (2) and moisture
content percentage (MC) and CT Hounsfield number (H).

Fig. 3. Observed sample density plotted on measured CT Hounsfield units (H) with
a line representing sample density (g=cm3 ) predicted by Eq. (2) (mA = 80, MC = 12%).
When the water displacement method is used, the rate of swelling,
air bubbles and other sources of error can produce systematic
biases, especially for dry samples (Mantanis et al., 1994). Biases
are also possible from estimating sample volume from basic
dimensions and assuming conformity to a rectangular prism (cube)
or cylinder (core). The angle at which calipers are placed on a sam-
ple, the force exerted on the caliper during measurement, and
actual shape of the sample all can affect volume estimation. While
both methods have shortcomings and possible biases, cross-
validation provides some evidence in support of using sample den-
sities calculated using VOLd for developing Eq. (2). X-ray tube cur-
rent was found to have a weak but significant effect on density
estimation (Table 4). This finding corroborates those of
Freyburger et al. (2009) who also observed a weak effect of mA
on estimating sample density. For each unit increase of X-ray tube
current (mA), the predicted density (q ^ ) increases at a rate of only
0.00003 g=cm3 . To minimize the chance of bias when predicting
q, it is suggested that experimenters control CT scanner settings,
Fig. 5. Response surface for sample density q predicted by Eq. (2) and X-ray tube
current (mA) and CT Hounsfield number (H).
like X-ray tube current. Moisture content was a significant predic-
tor of q in Eq. (2). As moisture content increases, so does q and
there is a significant interaction between MC and H in this relation-
ship. These findings corroborate those of Hoag (1988) and Lindgren
(1991), who also found a small but significant interaction between
moisture content of wood and CT Hounsfield units. In Eq. (2) the
measured sample moisture content was used instead of the tar-
geted moisture content. However, fitting Eq. (2) using the actual
moisture content instead of the targeted moisture content only
increased the R2 value by 0.03% and did not change parameter esti-
mates at the level of precision reported (Table 4). Filter choice in
processing CT scanning images was not found to be significant in
predicting sample density. However, filter choice could be impor-
tant when analyzing the profile of density along a sample. Any
smoothing of pixel values between CT scanner slices may slightly
alter location of important boundaries, like the delineation
between earlywood and latewood.
To minimize the chance of systematic biases in processing CT
imagery, automated image analysis software should be considered
(Steffenrem et al., 2014; Roussel et al., 2014). Some image process-
 N.L. Osborne et al. / Computers and Electronics in Agriculture 127 (2016) 5055
ing software could be insensitive to anomalies in wood samples, so
manual processing of imagery may still be necessary. For example,
one wood core drawn for this analysis had a staple embedded,
which was only revealed by looking carefully through the stacks
of DICOM imagery (Fig. 2).
5. Conclusions
Results from this study confirm the efficacy of medical X-ray
computed tomography to rapidly and robustly predict Douglas-
fir wood density. While the results of this analysis are interesting
to a wide audience, the specific models presented may be applica-
ble to only Douglas-fir and the CT scanner which was used. One
interesting topic to consider for further research would be the val-
idation of Eq. (2) to predict Douglas-fir q using different CT scan-
ning machines.
CT scanning technology has seen minimal applications in the
biometry of managed Douglas-fir in the Pacific Northwestern Uni-
ted States. Medical computed tomography, along with high resolu-
tion X-ray densitometry, could be useful in the development of a
comprehensive model to predict the effects of site conditions and
silvicultural regime on q in coastal Douglas-fir. Researchers should
now identify the optimal way to sample for wood properties
(including q), with consideration for utilizing different wood scan-
ning technologies.
Acknowledgements
Thanks are extended to Doug Mainwaring, Derek Gourley (Ore-
gon State University, College of Forestry) and Jason Wiest (Oregon
State University, College of Veterinary Medicine). Doug and Derek
helped prepare and collect many of the wood samples and aided in
their measurement. Jason scanned all of the wood cubes and cores
using computed tomography. This research was made possible by
the Veterinary Medicine College at Oregon State University, which
allowed use of their diagnostic imagery facility. Establishment of
the Giustina Fertilization trials was funded by a Giustina Innova-
tion Grant. Additional research on these trials was supported by
55
an endowment associated with a chaired faculty position, the N.
B. (Nat) and Jacqueline Giustina Professor of Forest Management.
References
Abrmoff, M., Magalhes, P., Ram, S., 2004. Image processing with ImageJ.
Biophotonics Int. 11, 3642.
Ct, W., 1968. The structure of wood and the wood cell wall. In: Principles of Wood
Science and Technology. Springer.
Emmingham, W., 1977. Comparison of selected Douglas-fir seed sources for cambial
and leader growth patterns in four western Oregon environments. Can. J. For.
Res. 7, 154164.
Filipescu, C., Lowell, E., Koppenaal, R., Mitchell, A., 2013. Modeling regional and
climatic variation of wood density and ring width in intensively managed
Douglas-fir. Can. J. For. Res. 44, 220229.
Freyburger, C., Longuetaud, F., Mothe, F., Constant, T., Leban, J., 2009. Measuring
wood density by means of x-ray computer tomography. Ann. For. Sci. 66, 804.
Grmping, U., 2006. Relative importance for linear regression in R the package
relaimpo. J. Stat. Softw. 17, 127.
Hoag, M., 1988. Measurement of Within Tree Density Variations in Douglas-fir
(Pseudotsuga menziesii (Mirb.) Franco) using Direct Scanning X-ray Techniques:
M.S. Thesis. Oregon State University, Department of Forest Products.
Kantavichai, R., Briggs, D., Turnblom, E., 2010. Modeling effects of soil, climate, and
silviculture on growth ring specific gravity of Douglas-fir on a drought-prone
site in Western Washington. For. Ecol. Manage. 259, 10851092.
Larson, P., 1969. Wood formation and the concept of wood quality. Yale Univ.
School Forestry Bull. 74, 154.
Lindgren, L., 1991. The accuracy of medical CAT-scan images for non-destructive
density measurements in small volume elements within solid wood. Wood Sci.
Technol. 25, 425432.
Maguire, D., Bennett, W., Kershaw, J., Gonyea, R., Chappell, H., 1991. Establishment
Report: Stand Management Cooperative Silviculture Project Field Installations.
College of Forestry, University of Washington, Seattle, WA.
Mantanis, G., Young, R., Rowell, R., 1994. Swelling of wood. Wood Sci. Technol. 28,
119134.
R Core Team, 2015. R: A Language and Environment for Statistical Computing. R
Foundation for Statistical Computing, Vienna, Austria, 2012.
Roussel, J., Mothe, F., Krhenbhl, A., Kerautret, B., Debled-Rennesson, I.,
Longuetaud, F., 2014. Automatic knot segmentation in ct images of wet
softwood logs using a tangential approach. Comput. Electron. Agric. 104, 4656.
Rozenberg, P., Franc, A., Bastien, C., Cahalan, C., 2001. Improving models of wood
density by including genetic effects: a case study in Douglas-fir. Ann. For. Sci.
58, 385394.
Steffenrem, A., Kvaalen, H., Dalen, K., Hib, O., 2014. A high-throughput x-ray-
based method for measurements of relative wood density from unprepared
increment cores from Picea abies. Scand. J. For. Res. 29, 506514.
Wei, Q., Leblon, B., La Rocque, A., 2011. On the use of x-ray computed tomography
for determining wood properties: a review. Can. J. For. Res. 41, 21202140.