Discussiones Mathematicae

Probability and Statistics 29 (2009 ) 155–168

A DOMINANT HEIGHT GROWTH MODEL

FOR EUCALYPTUS PLANTATIONS IN PORTUGAL
Ayana Mateus
CMA, Departamento de Matemática,
Faculdade de Ciências e Tecnologia FCT
Universidade Nova de Lisboa, UNL
Quinta da Torre, 2829–516 Caparica, Portugal
e-mail: [email protected] ; +351 21 2948388

and
Margarida Tomé
CEF-Centro de Estudos florestais
ISA- Instituto Superior de Agronomia
UTLisboa-Universidade Técnica de Lisboa
Tapada da Ajuda, 1349–017 Lisboa, Portugal
e-mail: [email protected] ; +351 21 3653100

Abstract
Eucalyptus globulus Labill is one of the most important economic
forest species in Portugal, occupying an area of 875.103ha in a to-
tal forest area of 3346.103ha (Tomé et al., 2007). The main goal of
this study is to develop a dominant height growth model for Eucalyp-
tus, applicable throughout the country, representing an improve of the
curves that are part of the whole stand model existing in Portugal, the
GLOBULUS model (Tomé et al., 2001). The dominant height growth
model will be built on a biological function formulated as a difference
equation (Algebraic Difference Approach) to an all possible growth
intervals data structure. The dynamic model proposed, obtains non
biased and efficient estimates for the parameters. Comparing to the
already available model, the last has the advantage of expressing the
asymptote in weather variables functions, meaning that it is possible
to reproduce the eucalyptus growth according to weather changes.
Keywords: algebraic difference equation, dynamic model, dominant
height, forest planning, non linear model.
156 A. Mateus and M. Tomé

2000 Mathematics Subject Classification: Primary 62J02;

Secondary 62P12.

1. Introduction

Eucalyptus globulus Labill is one of the most important economic forest

species in Portugal, occupying an area of 875.10 3 ha in a total forest area of
3346.103 ha (Tomé et al., 2007). It is a fast growing specie mainly used by the
pulp industry; the trees are planted at final density-thinning and pruning
practices are unusual at first rotation stands. The stands are intensively
managed in a short rotation coppice system in which the first cycle of plantes
seedlings (single stem) is followed by two or three coppiced stands, with an
average cutting cycle of 10-12 years. In order to contribute to a balanced
and resourceful management of Eucalyptus stand in Portugal, it is necessary
to acquire models that simulate their growth under different environment
conditions and treatments.
Estimating forest productivity is both necessary for effective forest man-
agement and useful for evaluating basic site conditions for ecological field
studies. Site quality is therefore influenced by factors such as available light,
heat, moisture and nutrients along with other soil characteristics such as soil
depth and aeration (Wang and Klinka, 1996). Although it would be best to
directly measure and predict these factors, some of them fluctuate widely
over the course of a day, month or year, whereas others require precise
measurements that may be difficult extrapolate across scales. Therefore, in-
direct methods for evaluating site quality are more frequently used in forest
management (Monserud, 1984).
Site index, defined as dominant height at some fixed base age, is one of
the most commonly used indicators of site productivity because there exists
a close correlation between volume and dominant height growth, and it is
generally accepted that height of dominant trees is only slightly affected by
competition (Clutter et al., 1983).
The main goal of this study is to develop a dominant height growth
model for Eucalyptus, applicable throughout the country, representing an
improve of the curves that are part of the whole stand model existing in
Portugal, the GLOBULUS model (Tomé et al., 2001). This model based on
environmental factors would be seen as a boom for a site quality one. In
particular, these curves should be totally independent on any global climatic
classification and parameterized directly from local climate variables. A
global climatic classification has the disadvantage of not considering each
A dominant height growth model for eucalyptus ... 157

plot unique characteristics, which may lead to wrong conclusions. Nowadays,

the development of geographical information systems facilities the access to
local climatic variables used to parameterize the curves.
The dominant height variable is defined as the average height of the
thickest trunk trees in the plot (keep in mind that the dominant height
quantifies/qualifies the productivity of the field).
Many mathematical functions are available to model dominant height
growth. These functions should satisfy some conditions. Based in biologi-
cally behavior, meaning have a universal application. In fact, these functions
are obtained in the differential form, establishing an hypothesis over absolute
or relative growth rates, obtaining an integration expression for production.
In this way one can assign absolute parameter meanings for this functions.
Consistent behavior with the biological growth, i.e. allowing null height
at the beginning of production and maximum finite height at an advance
stage (existence of a horizontal asymptote). Allow a sigmoidal growth (S
curve): curve slope increases with increasing production at an early stage
and should decrease in the final stage (existence of a inflection point). These
requirements are achieved depending on both the construction method and
the mathematical function used. Among the three general methods for site
index curve construction, the algebraic difference approach (ADA), has the
following advantages: short observations periods can be effectively used and
the structure of equations is base-age invariant (Clutter et al., 1983).
The algebraic difference approach formulates a function as a difference
equation in order to express the dominant height at age t i (hdomi ) according
to dominant height at age ti−1 (hdomi−1 ) and measurement ages ti and ti−1 , i.e.

hdomi = f (hdomi−1 , ti−1 , ti , β) + ei

where, β is the parameters vector to estimate and e i represents the stochastic

part of model.
In this study dominant height growth model will be built on a bio-
logical function formulated as a difference equation (Algebraic Difference
Approach). For the adjustment all observed pairs emerged from the avail-
able measurements will be used. The dynamic model proposed, obtains non
biased and efficient estimates for the parameters. Comparing to the already
available model, the last has the advantage of expressing the asymptote in
weather variables functions, meaning that it is possible to reproduce the
eucalyptus growth according to weather changes.
158 A. Mateus and M. Tomé

1. Data
The data used to model dominant height growth of eucalyptus plantations in
Portugal is slightly different from those used in the existing model, GLOB-
ULUS (Tomé et al., 2001). In fact, several validations were performed in
order to eliminate or correct data unreliable and using new observations
available from recent experiments.
In a previous work, Portugal was divided in eight climatic regions con-
cerning the productivity of eucalyptus (Ribeiro and Tomé, 2000). The data
used has cover this regions, although have a weak representation about the
relationship between site index in successive rotations (coppiced stands).
Rotations represents the number of cuts that a plot has been subject. Also,
measurements for ages above 10 years are little representative in the data.
This problem is more pronounced in coppiced stands. This fact may lead to
low precision of the asymptotes estimates, and consequently an underesti-
mation of growth. We must take this in account in the fase of adjustment.
Table 1 and 2 shows some characteristics of data used in first rotation and
coppiced stands.

Table 1. Characteristics of plots used at first rotation.

Region
1 2 3 4 5 6 7 8
No of measurements 98 1538 8194 1111 11364 836 3600 628
Age (years)
Minimum 1,0 0,6 0,8 2,8 1,1 2,6 2,8 3,0
Mean 5,4 5,6 7,8 6,5 6,7 7,2 6,5 5,4
Maximum 12,1 19,4 34,3 14,2 20,8 11,9 13,5 8,0
hdom (m)
Minimum 2,6 2,8 1,5 4,8 2,1 4,0 3,5 3,0
Mean 16,8 13,7 16,6 14,4 12,3 12,0 11,7 8,2
Maximum 29,2 28,1 36,6 27,9 34,6 25,4 27,4 15,0
Density (ha)
Minimum 901 350 200 421 200 250 200 200
Mean 2094 1297 1075 997 848 949 765 583
Maximum 4745 5000 3750 1850 2600 1875 1829 1575
Site Index
Minimum 16,0 5,2 4,2 8,3 7,0 5,7 9,5 7,8
Mean 24,5 21,5 20,1 19,2 17,0 16,0 15,9 13,6
Maximum 29,6 28,5 35,9 28,1 32,0 25,9 27,3 21,2
A dominant height growth model for eucalyptus ... 159

Table 2. Characteristics of plots used at coppiced stands.

Region
1 2 3 4 5 6 7 8
No of measurements 0 242 1029 131 2477 38 130 0
Age (years)
Minimum 1,7 1,1 4,4 1,9 6,1 2,3
Mean 8,0 7,5 7,8 6,8 7,8 5,8
Maximum 14,9 19,1 11,7 17,1 10,4 10,4
hdom (m)
Minimum 4,8 4,2 9,0 5,1 5,5 5,8
Mean 16,9 16,3 17,1 14,4 13,0 11,1
Maximum 28,2 30,8 25,4 29,5 19,9 21,0
Density (ha)
Minimum 825 328 800 221 550 275
Mean 2048 1594 1320 1374 1405 755
Maximum 3563 4852 2225 4179 2550 1650
Site Index
Minimum 12,2 8,5 11,2 8,1 8,0 10,2
Mean 19,8 19,6 19,0 17,9 15,2 14,5
Maximum 29,3 29,1 25,1 29,0 20,5 22,5

1.2. Selecting a growth function

In Table 3 we present the biological functions, formulated as an algebraic
difference equation, that we will be used in our work. These functions were
selected based on their accurate achievements in previous studies (Tomé
et al., 2001).
Table 3. Candidate functions formulated as an algebraic difference
equation (yi represents dominant height at age ti ).

Growth
Analytical expression
function

(tj /ti )n
Lundqvist-Korf yi = A (yj /A) A > 0, n > 0

)/ ln (1−e−ktj ) y ln (1−e )/ ln (1−e−ktj )

−kti
yi = A1−ln (1−e
−kti
Richards-m j A > 0, k > 0

    b −1
A tj
McDill-Amateis yi = A 1 − 1 − yj ti A > 0, b > 0
160 A. Mateus and M. Tomé

1.3. Data Structure and model fitting

Modeling dominant height growth involves two processes:
• estimating height at base age (hdom p ), given height at some other age.
The main objective is obtaining a site index prediction equation:

hdomp = f (tp , ti , hdomi )

• estimating height at some desired age (hdom i ) given height at base

age. The main objective is obtaining is obtaining a height prediction
equation, having a prior knowledge about site index

hdomi = f (ti , tp , hdomp ).

These two processes may be modeled by individual functions for each pro-
cess, or by one equation that predicts height at any desired age, given height
at any other known age. When using individual functions, height is assumed
to be measured without error when on the right-hand of equation but with
error used on the left- hand side. This assumption causes a bias in the pa-
rameters of the curves; neither the height-prediction equation nor the site
index prediction equation will have a shape that represents the true func-
tional relationship between height and age across levels of site index.
When a unique function is used, the purpose is to simultaneously op-
timize the regression of Y on X and X on Y and avoid parameter bias.
In this study, the model to dominant height growth will be built using a
unique function, in the form of an algebraic difference equation (ADE) to
an all possible growth intervals data structure. For example if plot height
is measured at ages t1 , t2 and t3 , we would have the following set of ob-
servations for the adjustment (hdom 1 , t1 , hdom2 , t2 ), ( hdom1 , t1 , hdom3 ,
t3 ), (hdom2 , t2 , hdom3 , t3 ), (hdom2 , t2 , hdom1 , t1 ), (hdom3 , t3 , hdom1 , t1 ),
(hdom3 , t3 , hdom2 , t2 ). With the all possible growth intervals, there is no
independence between observations. In order to avoid this problem we used
an auto-regressive structure to errors, such as:

hdomi,j = f (hdomj , ti , tj , β) + ei,j com

ei,j = ρei−1,j + γei,j−1 +i,j |ρ| < 1, |γ| < 1

where hdomi,j represents the obtained value for dominant height at
A dominant height growth model for eucalyptus ... 161

age ti using as explanatory variables dominant height at age t j and also

measurements age ti and tj . β is the parameter vector to be estimate and
ei,j is the correspondent error term. The parameter ρ takes into account
the correlation between the current residual and the residual associated to
estimating dominant height at age t i−1 using dominant height at age tj as
a predictor variable. Finally, the parameter γ refers to the autocorrelation
between the current residual and what you get from estimating dominant
height at age ti using the dominant height at age tj−1 as predictor variable.
i,j are independent and identically distributed variables.
Models like this are usually referred as dynamic models and its main
objective is getting unbiased and efficient estimates to the parameters model
(Gallant, 1987) (p. 405–426).
By using all possible differences, the number of observations is artificially
inflated, although no additional information is obtained. As a consequence a
short standard deviation is obtained and needs to be corrected.
p This can be
achieved multiplying the obtained deviation by the factor N (apd/N (f d)),
where N (apd) is the number of observations using all possible differences
and N (f d) is the number of observations using only the first differences.
If k represents the number of trees and n i represents the number of
measurements on tree i, then

k
X k
X
N (apd) = (ni (ni − 1)) and N (f d) = ni .
i=1 i=1

We deal in this way because if a data set is replicated, for instance each
observation reappears
p four times, the mean standard deviation will have a
factor decrease of 1N/4N , where N is the original number of observations.
The fitting were carried using the generalized least square method in
non linear equations (Generalised Least Squares, GLS) (Seber and Wild,
1989). The PROC MODEL procedure on the SAS/STAT software (SAS,
2005) was used.

1.4. Model selection criteria

A three-step procedure is used to evaluate and select the most appropriate

model, which include qualitative and quantitative examinations.
162 A. Mateus and M. Tomé

The first step evaluates the model fitting statistics based on mean sum square
residual and coefficient of determination, represented respectively by:
n
X n
X
2
(yi − ybi ) (yi − ybi )2
i=1 i=1
M SSR = and R2 = 1 − n
n X
(yi − Y )2
i=1

where ybi represents the predict value and Y mean of observations.

The model validation proceeds in the second step. An independent data
set testing was not available, so the the validation criteria were: in plots
with measurements above six, the second and penultimate were removed for
validation. The plots that met this criterion were in number of 178. From
the second observation the penultimate observation was predicted, and both
mean predicted residual (r) and mean absolute predicted residual (|r|) were
computed. The same procedure was used to predict the second observation
based on the penultimate. To evaluate the dispersion of residuals percentiles
5% and 95% were calculated.
The selection was also based on the analysis of parameters obtained
with the expected biological behavior, including: Signs and values of the
coefficients in the model components, especially the asymptotes and height
curve development.

2. Results

An iterative process always requires initial values for the parameters in order
to avoid local solutions and to quickly achieve convergence.
The initial parameters related with growth used in the three selected
functions, were those obtained in the GLOBULUS model (Tomé et al.,
2001). A regression like,

ei,j = ρei−1,j + γei,j−1 |ρ| < 1, |γ| < 1

was used in order to obtain initial information about the parameters, ρ and
γ, related with the stochastic part of the model.
The results achieved by each of the stared functions were the following
ones:
A dominant height growth model for eucalyptus ... 163

(1) Function Lundqvist-Korf MSSR= 1.369, R 2 = 0.9606;

(2) Function McDill-Amateis MSSR= 1.412, R 2 = 0.9595;

(3) Function Richards-m MSSR= 1.497, R 2 = 0.9569.

Table 4 presents some validation statistics on selected models, namely char-

acterization of the error (bias and precision). By the same table, the use of
Richards-m leads to a less precise model. Taking in to account this together
with the fact that this function provides a not suitable curve development,
we decided to not consider this function.

Table 4. Evaluation of the predictive ability of models.

Growth Projection Mean Abs. mean Percentile Percentile

function order r |r| 5% 95%
Lundqvist-Korf t i < tj 0,1005 0,3590 -0,5725 1,0428
tj < t i -0,0767 0,2933 -0,8524 0,4926
McDill-Amateis t i < tj 0,0909 0,3520 -0,6121 0,9861
tj < t i -0,0741 0,3005 -0,8208 0,5069
Richards-m t i < tj 0,0982 0,5053 -0,7801 1,3354
tj < t i -0,0824 0,4791 -1,2347 0,7278

As dominant growth of a certain specie is related with it’s location, we tried

to relate the function parameters with the specifies of the region, in order
to obtain a model reflecting how geography influences growth.
There is a biological relation between temperature (T M ) and vege-
tation growth though it is not linear as there is an exact temperature
associated to an optimal growth, and above this level the growth decreases
substantially. Below we present the mathematical expression utilized:
a + bT M + cT M 2 .
164 A. Mateus and M. Tomé

In order to get a proper biological adjustment, coefficient b has to be

positive and c negative, with the maximum curve slope around the
maximum temperature supported by the tree. The same criteria was applied
to the insolation (IN S) and radiation (RAD).
The remain variables, such as, frost, evapotranspiration, precipitation
and humidity were incorporated linearly. All the coefficients should be
positive regarding evapotranspiration, precipitation and humidity variables.
Regarding frost it is expected a negative coefficient.
In order to avoid over-parametrization problems, we did not introduce
both radiation and insolation as well as temperature and humidity, because
this variables have high correlation. So the variables were picked up ac-
cording to their sign compatibility with the biological growth and the best
adjustment.
Once we have data from stands in first rotation and coppice, we decided
to study the influence of rotation in the dominant height growth, expressing
the rotation as a dummy variable (1 if coppice).
Basically parameter A, related with asymptote value, was estimated in
function of climatic variables. The effect of rotation (Rot) has only been
tested, in a linear way, at growth rate parameter (n, b and k) because the
asymptote expressing the potential of a species is the same in successive
rotations. Table 5 and 6 shows the main results obtained.
In Portugal the conditions that most affect Eucalyptus plantations
productivity are related to water and frost. Conclusions taken after in-
tense analysis to the country climatic conditions state there’s a negative
correlation between precipitation, temperature and radiation. So, the
effects of precipitation are quantified indirectly in the estimates of the
coefficients a2 and a4 used in the models (see Table 5). Inferior asymptote
values will represent low levels showers associated to hight temperatures
and radiations.
Growth rate parameter, estimated in function of rotation, always
came significant with a negative signal compatible with the biological
growth in coppiced stands. That is, in coppice stands the shape of growth
curve is different from stands in first rotation. Initially, there is a faster
growth, consequence of the root system already be installed, becoming
slower then.
A dominant height growth model for eucalyptus ... 165

Table 5. Results from estimating parameters in function of climatic variables.

Growth Parameters Standard error Statistics

function estimates (correct) t*
Lundqvist-Korf

n1=-0,081 0,003 -27,99

A = a0 + a1 T M + a2 T M 2 a0=-1978,11 492,721 -4,01
+a3 RAD + a4 RAD 2 a1=32,389 7,231 4,48
a2=-1,134 0,247 -4,59
a3=24,067 6,603 3,65
a4=-0,0796 0,022 -3,56
ρ=0,287 0,007 44,23
γ=0,621 0,007 88,79
MSE=1,3301

McDill-Amateis

b1=-0,233 0,008 -30,16

A = a0 + a1 T M + a2 T M 2 a0= -971,874 212,173 -4,58
+a3 RAD + a4 RAD 2 a1=17,859 3,020 5,91
a2=-0,604 0,104 -5,82
a3=11,809 2,856 4,13
a4=-0,039 0,010 -4,09
ρ=0,325 0,006 51,13
γ=0,606 0,007 87,90
MSE=1,3696

* p-value < 0, 0001

Table 6. Evaluation of the predictive ability of models expressed in function

of climatic variables.

Growth Projection Mean Mean Percentile Percentile

function order r |r| 5% 95%
Lundqvist-Korf ti < tj 0,0997 0,3530 -0,5426 0,9888
tj < t i -0,0765 0,2941 -0,8006 0,4997
McDill-Amateis ti < tj 0,0849 0,3437 -0,5829 0,9183
tj < t i -0,0676 0,2943 -0,7281 0,4961
166 A. Mateus and M. Tomé

Figure 1. Graphs obtained with different models parameterize in function of

climatic variables.
A dominant height growth model for eucalyptus ... 167

Figure 1 above represent the growth obtained with different models parame-
terize in function of climatic variables. The three main bold curves represent
the growth quality according to the season. Observing the results above we
immediately notice the Lundqvist-Korf function has a far better asymptotic
behavior regarding old ages plots than the McDill-Amateis one. We can
also notice the tendency of each curve to converge at latter ages regardless
the climatic conditions (McDill-Amateis function). To finalize we choose the
Lundqvist-Korf function based of the facts stated above plus it’s adjustment
and predicting capacities.

Acknowledgements

This work was supported by the FP6 EFORWOOD IP project (contract

518128) and the FCT project CarbWoodCork (POCI/AGR/57279/2004).
The authors also thanks Celbiana Silvicaima pulp company that provide
the data used in this study.

References

[1] M. Tomé, S. Barreiro, A. Cortiçada, A. Meyer, T. Ramos and P. Malico,

Inventário florestal 2005-2006. Áreas, volumes e biomassas dos povoamentos
florestais. Resultados Nacionais e por Nut´s II e III. Publicações GIMREF
RT 5/2007 Universidade Técnica de Lisboa. Instituto Superior de Agronomia,
Centro de Estudos Florestais, Lisboa, Portugal 2007.
[2] M. Tomé, F. Ribeiro and P. Soares, O modelo GLOBULUS 2.1-Relatórios
técnico cientı́ficos do GIMREF no 1/2001. Universidade Técnica de Lisboa,
Instituto Superior de Agronomia, Centro de Estudos Florestais, Lisboa, Por-
tugal 2001.
[3] G.G. Wang, K. Klinka and W. John, Use of synoptic variables in predicting
white spruce site index, For. Ecol. Manage 80 (1996), 95–105.
[4] R.A. Monserud, Height growth and site index curves for inland Douglas-
firbased on stem analysis data and forest habitat type, For. Sci. 30 (1984),
943–965.
[5] L. Clutter, J. Fortson, L. Pienaar, G. Brister and L. Bailey, Timber Manage-
ment: A Quantitative approach, John Wiley and Sons 1983.
[6] F. Ribeiro and M. Tomé, Classificação climática de Portugal continental,
baseada em informação do atlas do ambiente, Revista de Ciȩncias Agrárias
2000.
168 A. Mateus and M. Tomé

[7] A.R. Gallant, Nonlinear Statistical Models, John Wiley & Sons (1987), 624p.
[8] The SAS System for windows, 9.1 edition.
[9] G.A.F. Seber and C.G. Wild, Nonlinear Regression, New York, John Wiley
1989.

Received 17 September 2009