discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/317825551
CITATIONS READS
0 124
3 authors:
Harold E. Burkhart
Virginia Polytechnic Institute and State University
229 PUBLICATIONS 4,701 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Decision Support System for Sustainable Forest Management in Galicia (Spain) View project
All content following this page was uploaded by Harold E. Burkhart on 28 June 2017.
The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
For. Sci. 63(3):250 260
FUNDAMENTAL RESEARCH https://doi.org/10.5849/FS-2016-087
Copyright 2017 Society of American Foresters
biometrics
In addition to dbh, total tree height and upper-stem diameter have been commonly used to improve tree stem volume estimation. However, measurement error in these
variables has not been accounted for when realized gains on estimating stem volume are assessed. The aim of this study was to evaluate whether height and upper-stem
diameter actually improve accuracy of volume estimation when measurement error is considered and its effect on prediction. dbh was considered to be observed without
measurement error. For this analysis, a volume equation and three stem taper functions (with two algebraic constraining alternatives) were fitted and applied over four
data sets of stem measurements from loblolly pine stands. Several levels of variability of measurement errors were assessed using a Monte Carlo simulation approach.
These conditional predictor variables improved volume prediction when no measurement error was considered, as expected. However, as measurement error increased,
the accuracy of volume estimation decreased. When the three variables were considered for volume estimation, measurement error in height caused a positive bias,
whereas that corresponding to upper-stem diameter yielded a negative bias. Finally, in terms of volume prediction, inclusion of height clearly justifies the effort needed
to measure it, whereas inclusion of upper-stem diameter is only recommended if its measurement error is low.
Keywords: volume equation, stem taper function, algebraic constraint, coefficient of variation, loblolly pine
E
stimation of tree stem volume, which is closely related to ated variables have been upper-stem diameters (du) (e.g.,
economic value, tree biomass, or carbon content, is a mat- Czaplewski and McClure 1988, Cao and Wang 2015, Sabatia
ter of interest for both researchers and forest practitioners. and Burkhart 2015), gaining even more importance with the
The most common approaches to predict stem volume are using technologic advance in instruments for tree measurement like
volume equations (e.g., Burkhart 1977, Van Deusen et al. 1981) laser dendrometers (Clark et al. 2000) or terrestrial laser scan-
or integrating taper equations (e.g., Max and Burkhart 1976, ning (Henning and Radtke 2006). Upper-stem diameters have
Byrne and Reed 1986), although other methods such as Monte been considered for volume estimation as a predictor (e.g.,
Carlo estimators (Valentine et al. 1992) or nonparametric stem Berger et al. 2014), as a constraint to force stem taper functions
curves (Lappi 2006) have also been used. Both volume and taper to pass through a specific point (algebraic constraint) (e.g., Cao
equations use dbh (d ) and total tree height (h) as independent 2009), or as additional information to localize stem taper func-
variables, although h may not appear in the former (Burkhart and tions fitted by the mixed-effects modeling approach (e.g., Trin-
Tome 2012, p. 44). In addition, the inclusion of additional cado and Burkhart 2006, Arias-Rodil et al. 2015). When alge-
predictors, such as crown variables, upper-stem diameters, or braically constraining stem taper functions through mid-stem
even stand variables has been considered in many studies (e.g., points, authors typically have considered only d or the joint use
Petersson 1999, Leites and Robinson 2004, Sharma and Parton of d and du as the points through which the model is forced to
2009). Among additional predictors, the most commonly evalu- pass (a constraint that the function equals zero at the tree tip is
Manuscript received September 7, 2016; accepted November 12, 2016; published online January 5, 2017.
Affiliations: Manuel Arias-Rodil (manuel.arias.rodil@gmail.com), Unidad de Gestin Forestal Sostenible, Departamento de Ingeniera Agroforestal, Universidade de
Santiago de Compostela. Escuela Politcnica Superior, C/ Benigno Ledo, Campus Terra, 27002 Lugo, Spain. Ulises Dieguez-Aranda (ulises.dieguez@usc.es), Unidad de
Gestin Forestal Sostenible, Departamento de Ingeniera Agroforestal, Universidade de Santiago de Compostela. Escuela Politcnica Superior, C/ Benigno Ledo, Campus
Terra, 27002 Lugo, Spain. Harold E. Burkhart (burkhart@vt.edu), Department of Forest Resources and Environmental Conservation, Virginia Polytechnic Institute and
State University, 319 Cheatham Hall, 310 West Campus Drive, Blacksburg, VA 24061-0324, United States.
Acknowledgements: This research was conducted when the corresponding author was a Visiting Scholar at the Department of Forest Resources and Environmental
Conservation, Virginia Polytechnic Institute and State University in Blacksburg, VA. The corresponding author was in receipt of an FPU grant (AP20125337)
from the Ministry of Education of the Spanish Government. The authors would also like to acknowledge the Forest Modeling Research Cooperative at Virginia Tech
for providing the data used in this study.
volume estimation.
where di is the diameter along the stem (cm), d is the dbh (cm), h is
the total tree height (m), q (hi/h), hi is the height along the stem
Volume Equation (m), x [w/(1 (1.37/h)13)], w 1 q1/3, ai (i0, 1, 2) and
In this study, we initially considered a single-entry volume equa- bi(i 1, , 6) are model parameters, I1 1 if a1 q and zero
tion (Equation 1) (Avery and Burkhart 2002, p. 169), which only otherwise, and I2 1 if a2 q and zero otherwise.
depends on dbh (d ): Stem taper functions predict stem diameter (di) at a specific stem
height (hi) from d and h; i.e., di f(hi, d, h). Total tree volume (v,
v b 1d b2 (1) m3) can then be obtained by integration (Equation 7). The sixth-
where v is the tree volume (m3), d is the dbh (cm), and bi (i 1, 2) degree polynomial and Max and Burkhart (1976) equations are
are the model parameters. Then, we considered the well-known analytically integrable, whereas the Kozak (2004) function must be
expression of Spurr (1952) to include the total tree height: numerically integrated.
v b 0 b 1d 2h (2)
h
where h is the total tree height (m) and the rest of variables as v fh i , d, h 2 dh i (7)
40,000
mentioned above. Finally, and after previous analyses, we added a 0
Equation
Parameter 1 2 3 4 5 6
a0 0.9584
a1 0.7733 1.016
a2 0.1109
b0 4.154 103 2.812 103
b1 6.711 105 3.498 105 5.517 105 11.28 3.480 0.4542
b2 2.706 52.15 1.663 0.3648
b3 133.4 1.088 0.9369
b4 179.4 51.60 0.02835
b5 121.5
b6 32.87
Volume equations: Equation 1 v f(d), Equation 2 v f(d, h), Equation 3 v f(dm h,du). Stem taper functions: Equation 4 (polynomial), Equation 5 (Max and
Burkhart 1976), Equation 6 (Kozak 2004). All parameter estimates were significant at a 95% confidence level.
Table 3. RMSE and bias in volume prediction obtained with volume equations (VE), and stem taper functions applied over different data
sets.
Data set
lowest RMSE and absolute value of bias were obtained when The addition of predictors improved volume estimation (i.e.,
these equations were constrained to pass through d. Accordingly, RMSE): for volume equations, from 42 to 52% when h was in-
the d constraint was considered for further analysis, as it was cluded and from 63 to 76% when both h and du were included; and
generally more accurate. Cao (2009) also reported that the d for stem taper functions, from 16 to 86% when du was added to the
constraint improved the results over those of nonconstrained respective taper function. This contrasts with the results of Sabatia
equations, although only for outside-bark diameters. Between and Burkhart (2015), who found that an algebraically constrained
models, the volume equation provided the best results (in RMSE stem taper function provided poorer results than a nonconstrained
and bias), although the differences compared with stem taper model for one of the two data sets used in their study. In terms of
equations were slight. bias, no clear pattern of over- or underestimation was observed when
60
40
0.003
0.002
0.001
Bias difference (m3)
0.000
0.001
0.002
0.003
1 5 10
CVh (%)
Figure 1. Relative RMSE and absolute bias difference of the models that use only dbh (d ) and total tree height (h) as independent
variables, including ME in h, computed relative to those obtained with corresponding models without ME, and applied over the NIMP data
set. The [i]x-axis represents the coefficient of variation of h-ME (CVh).
independent variables were added. Nevertheless, we observed the those of plantations, which result in different stem shape, this im-
same effect on bias from addition of du across models and data sets: plied that the worst and most biased predictions corresponded to
bias differences were highest for the volume equation and lowest for this data set, but that one also benefited most from including addi-
the polynomial equation, whereas the Max and Burkhart (1976) tional variables, as this extra information became more helpful in
and Kozak (2004) equations yielded similar values. this case. On the other hand, the latter showed the best results (in
The relative reduction in RMSE associated with the inclusion RMSE and bias), which is explained because it has the lowest data
of an upper-stem diameter was always highest for the polynomial variability (Table 1). Finally, it should also be noted that the perfor-
stem taper equation; in fact, it was the most accurate and precise mance of each stem taper function was different for each data set,
stem taper function when du was considered. It appears that without exhibiting any clear hierarchy.
the sixth-degree polynomial function was more responsive to
inclusion of an upper-diameter constraint than the other taper Evaluating ME in Total Tree Height
equation forms evaluated, because they allow for varying tree Figure 1 shows the error statistics obtained from 1,000 simula-
form along the tree stem. tions with all models that considered d and h for volume prediction
There were some remarkable differences among data sets, mainly (not du), with ME in h, for the NIMP data set. The base models used
concerning NS and IMP data sets, that should be mentioned. As the here to compute relative error statistics were those without ME
former comprised trees grown in stand conditions different from considered. As expected, the mean value and variability of RMSE
100
50
0.006
0.004
Bias difference (m3)
0.002
0.000
0.002
0.004
1 5 10 1 5 10 1 5 10
CVd u (%)
Figure 2. Relative RMSE and absolute bias difference, of all the models that use dbh (d ), total tree height (h), and upper-stem diameter
(du) as independent variables, including ME in the h and du, computed relative to those obtained with the corresponding models without
ME, and applied over the NIMP data set. The x-axis represents the coefficient of variation of du-ME (CVdu).
polynomial equation yielded the highest RMSE values for the base 20
model.
The variability of bias increased with CVh, although the mean
value was not affected, except for a slight positive increase observed 30
for the polynomial and the Max and Burkhart (1976) equations for
the extreme case of CVh 10%. This bias increase results from
height entering nonlinearly in the functional forms of the equations 40
mentioned (see Equations 4 and 5), whereas it enters linearly in the
volume equation (see Equation 2); it also entered almost linearly in
1 5 10
the Kozak (2004) function (graph not shown).
CVh (%)
Evaluating ME in Total Tree Height and Upper-Stem Diameter Figure 3. Relative RMSE of all the models that use dbh (d ) and
Figure 2 shows error statistics obtained with the NIMP data total tree height (h) as independent variables, including ME in h,
computed relative to that obtained with the volume equation, which
set from 1,000 simulations for the models that use d, h, and du in depends only on d, and applied over the NIMP data set. The x-axis
volume prediction and with ME in h and du. These error statistics represents the coefficient of variation of h-ME (CVh).
were computed relative to the corresponding models without
ME. In terms of the relative RMSE increase, volume equation
results were affected both by CVh and CVdu variation, whereas shape prediction of stem taper functions and subsequently in
those obtained with stem taper functions were more affected by volume estimation.
CVdu. This differential behavior occurs because du is included in An increase in the variability of ME in height (CVh) caused an
stem taper functions as an algebraic constraint and h as a predic- increase in the positive bias. When h is used together with du, the
tor, whereas both variables are considered as predictors in the h-ME also affects the upper-stem diameter (by 1a error, as explained
volume equation. This made du more influential than h in stem in the Methods section). Accordingly, volume is underestimated
50
50
1 5 10 1 5 10 1 5 10
CVd u (%)
Figure 4. Relative RMSE of all the models that use dbh (d ), total tree height (h), and upper-stem diameter (du) as independent variables,
including measurement error in h and du, computed relative to those obtained with the corresponding models that depend on d and h,
with ME in h, and applied over the NIMP data set. The x-axis is the coefficient of variation of du-ME (CVdu).
when the h-ME 0 (i.e., h* lower than h) and overestimated in the timation or not; and the difference in bias between the tree volume
inverse situation. However, the underestimation is greater than the equation which depends only on d and that depending on d and h
overestimation, which explains the above-mentioned positive bias (ME in h) presented trends similar to those observed in Figure 1
increase. For example, when CVh 10% and CVdu 0%, the with respect to CVh but shifted according to the bias variation
polynomial stem taper function in the NIMP data set showed a showed in Table 3 [f(d, h) versus f(d )] (graph not shown).
volume underestimation of 2.1% when h-ME 0 and overestima-
tion of 0.4% when h-ME 0, relative to the same model without Adding Upper-Stem Diameter in Volume Estimation
ME. In contrast, an increase in CVdu caused a negative bias increase We compared models that consider d, h, and du for volume
in stem taper functions. This is because the volume overestimation estimation (with ME in h and du) with those that consider only d
when du-ME 0 (i.e., d u* greater than du) is greater than the volume and h (with ME in h). The relative RMSE is presented in Figure 4
underestimation in the inverse situation. This situation can be illus- for the NIMP data set. Generally, the models with du became worse
trated for the NIMP data set and the polynomial stem taper equa- than those without it for high variability levels of ME (e.g., CVdu
tion, considering CVh 0% and CVdu 10%, when the overesti- 10%); this pattern is clearer for low variability levels of CVh. Con-
mation over the model without ME was 8.8% when du-ME 0, cerning bias, we omitted the results of bias difference between
whereas the corresponding underestimation (when du-ME 0) was f(d, h*) and f(d, h*, d u*) in Figure 4 as the trends (graphs not shown)
7.8%. were a result of the combination of the bias variation observed in
When stem taper functions were used, du-ME affected the mag- Table 3 and variation with CVh and CVdu (Figure 2).
nitude of the bias, whereas it was not influential when the volume Figure 5 shows the lower limit of ME from which the models that
equation was used. This result may come from the different way in depend on d and h (with ME in h) provided better results than those
which du was considered in these equations: it was included as an which depend on d, h, and du (with ME in h and du); i.e., the error
algebraic constraint in the former, varying the predicted stem taper, levels in h and du from which the upper-stem diameter is no longer
whereas it was included as a predictor variable in the volume useful to increase the prediction accuracy. The ME level of du at
equation. which using only d and h for volume estimation is better than using
d, h, and du increases with CVh, which is explained because ME in h
Assessing Inclusion of Additional Variables influences both estimation alternatives [f(d, h*) and f(d, h*, d u*)] and
Adding Height in Volume Estimation the error increase due to ME in h is higher in f(d, h*) than in
Figure 3 shows the volume RMSE values obtained with the NIMP f(d, h*, d u*) (Figures 1 and 2, respectively).
data set when the models that consider d and h (with ME in h) were The polynomial equation consistently provided better results
used, computed relative to the volume equation in which only d was than f(d, h*,du*) when du was included except for the IMP data set,
considered. Inclusion of total tree height improved volume estimation, and therefore the corresponding line in Figure 5 is only shown for
even with a ME of 10%, which agrees with the findings of Williams and the IMP data set. Concerning the remaining models, the ME level of
Schreuder (2000), who concluded that height measurement is recom- du at which f(d, h*, d u*) got worse than f(d, h*) is generally lower for
mended in volume estimation, even with a ME of 40%. Height the Max and Burkhart (1976) equation, which agrees with the re-
measurement effort is clearly justified in volume estimation. sults observed in Figure 4 and demonstrates its greater sensitivity to
The bias difference computed relative to f(d ) was not deemed MEs in du compared with those in volume equations or the Kozak
informative and therefore not included in Figure 3 for two reasons: (2004) stem taper equation. Among data sets, the lowest limits of
only the volume equation considers the alternative of using d alone error in h and du that define the region where f(d, h*) was better than
as a predictor and therefore bias difference to stem taper functions f(d, h*, d u*) IMP, which implies that du adds less valuable informa-
that considered d and h, with ME in h, has no bearing on the tion to volume prediction in this case, supported also by the results
determination of whether height should be included in volume es- obtained without ME (see RMSE reduction for IMP, Table 3).
6
CVd u (%)
5
Nonintensivelymanaged plantations (NIMP) Intensivelymanaged plantations (IMP)
10
0 2 4 6 8 10
CVh (%)
Figure 5. Lower limits of coefficients of variation of ME in height (CVh) and in upper-stem diameter (CVdu) from which the model that
depends on dbh (d ) and total tree height (h, with ME in this variable) provided better results than the model which includes d, h, and an
upper-stem diameter (du, with ME in h and du), for different data sets.
From the results observed in Figure 5 (supported also by those in volume estimation, even for MEs up to 10%, the recommendation
Figure 4), the use of an upper-stem diameter increased volume pre- about using upper-stem diameter for volume estimation was more
diction accuracy even for relatively high levels of ME (mostly up to dependent on the level of ME considered. The effort needed to
710%); therefore, its use is recommended. Nevertheless, we should measure h is therefore justified, although it may not be warranted for
consider that even for errors above 4 5%, the volume precision gain du if ME is greater than about 5%.
with du may not be worth the measurement effort needed to record
this variable in the field, which is, furthermore, higher in variability
Appendix: Algebraic Constraining of Stem Taper
than that of h, as it requires two steps: locating the stem point and
measuring the diameter. Functions
As procedures for constraining taper functions have been de-
tailed in previous studies (see, e.g., Cao 2009, Cao and Wang 2015,
Conclusions Sabatia and Burkhart 2015), we present here the expressions of the
The present study evaluated whether the measurement of vari- parameters to constrain the polynomial equation and the Max and
ables height and upper-stem diameter improves volume estimation Burkhart (1976) equation. For the Kozak (2004) equation, we used
when they were affected by ME. When no ME was included, the use numerical procedures (OPTIMIZE function and MULTIROOT func-
of h and du in addition to d increased volume prediction accuracy. tion of the ROOTSOLVE package [Soetaert and Herman 2009] of R
The lowest precision was obtained for natural stands, as they exhibit [R Core Team 2015]) as no analytical solution was available.
different stem shapes relative to plantations, although they also ex- Let f(hi, d, h) be the nonconstrained stem taper function, which
hibited the greatest improvements when additional variables were yields squared stem diameter (d2i ) for a stem height (hi), a dbh (d ),
considered. total tree height (h), and bi(i 1, , 6) are model parameters.
The inclusion of ME in h when only d and h were used for
volume estimation increased the volume prediction error, as ex-
dbh Constraint
pected, although bias was only slightly affected. Although the vol-
To force the polynomial equation to pass through d, the b1
ume equation was most accurate when no ME was considered, it was
parameter proved the best in this study, and the expression that
the most affected by ME in h.
should be incorporated in the equation to constrain the model is
When both h and du were used for volume estimation and af-
fected by ME, the prediction accuracy yielded by the volume equa- d 2 f1.37, d, h
tion was affected by ME in h and du, whereas stem taper functions b *1 b 1 (A1)
d 2 q d 1
were more affected by that of du, as it was considered in the latter as
an algebraic constraint. In terms of bias, the h-ME implied a positive where qd (1.37/h).
bias increase, whereas the du-ME caused a negative bias difference. For the Max and Burkhart (1976) equation and constraining the
Finally, although height proved to be valuable information for b3 parameter, we obtained:
CANAVAN, S.J., AND D.W. HANN. 2004. The two-stage method for mea-
d 2u fh u , d, h d 2 f1.37, d, hq d u 1 surement error characterization. For. Sci. 50(6):743756.
d2 d 2 q d 1 CAO, Q.V. 2009. Calibrating a segmented taper equation with two diam-
b2 b2
*
q d 1q d u 1
2 eter measurements. South. J. Appl. For. 33(2):58 61.
q 2d u 1 CAO, Q.V., AND J. WANG. 2015. Evaluation of methods for calibrating a
qd 1 tree taper equation. For. Sci. 61(2):213219.
(A6) CLARK, N.A., R.H. WYNNE, AND D.L. SCHMOLDT. 2000. A review of past
research on dendrometers. For. Sci. 46(4):570 576.
For the Max and Burkhart (1976) equation, we followed the CLARK, N. A., S.J. ZARNOCH, A. CLARK, III, AND G.A. REAMS. 2001.
same procedure for b1 and b3 parameters and obtained Comparison of standing volume estimates using optical dendrometers.
P. 123128 in Proc. of the Second annual Forest Inventory and Analysis
d 2 f1.37, d, h b *3 b 3 a 1 q d 2 I 1, d d 2 symposium, 2000 October 1718, Salt Lake City, UT. USDA For. Serv.,
b *1 b 1
d 2 q d 1 Gen. Tech. Rep. SRS-47, Southern Research Station, Asheville, NC.
CZAPLEWSKI, R.L., AND J.P. MCCLURE. 1988. Notes: Conditioning a seg-
(A7) mented stem profile model for two diameter measurements. For. Sci.
34(2):512522.
d 2u fh u , d, h d 2 f1.37, d, hq d u 1
DIEGUEZ-ARANDA, U., F. CASTEDO-DORADO, J.G. ALVAREZ-GONZALEZ,
d2 d 2 q d 1 AND A. ROJO. 2006. Compatible taper function for Scots pine planta-
b *3 b 3
a 1 q d 2 I 1, d q d u 1 tions in northwestern Spain. Can. J. For. Res. 365:1190 1205.
a 1 q d u I 1, d u
2
HENNING, J.G., AND P.J. RADTKE. 2006. Detailed stem measurements of
qd 1
standing trees from ground-based scanning lidar. For. Sci. 52(1):67 80.
(A8) KALLIOVIRTA, J., J. LAASASENAHO, AND A. KANGAS. 2005. Evaluation of
the Laser-relascope. For. Ecol. Manage. 204(23):181194.
where I1,d 1 if a1 qd. Given that a1 0.7651 in this study, I1,d KANGAS, A. 1996. On the bias and variance in tree volume predictions due
and I1,du will be 1 because relative breast height and relative upper- to model and measurement errors. Scand. J. For. Res. 11(1 4):281290.
stem height (qd and qdu, respectively) are always lower than a1. KOZAK, A. 1998. Effects of upper stem measurements on the predictive