Sie sind auf Seite 1von 9

Renewable and Sustainable Energy Reviews 66 (2016) 843–851

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews


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

The estimation of basket willow (Salix viminalis) yield – New approach,


Part II: Theoretical model and its practical application
Wojciech Jakubowski a, Wiesław Szulczewski a, Andrzej Żyromski b,
Małgorzata Biniak-Pieróg b,n
a
Department of Mathematics, Wrocław University of Environmental and Life Sciences, ul. Grunwaldzka 53, 50-357 Wrocław, Poland
b
Institute of Environmental Protection and Development, Wrocław University of Environmental and Life Sciences, pl. Grunwaldzki 24, 50-363 Wrocław,
Poland

art ic l e i nf o a b s t r a c t

Article history: The paper presents an innovative approach to probabilistic modelling of the amount of biomass of fast-
Received 29 October 2015 growing bushes on the example of basket willow (Salix viminalis). The model was developed on the basis
Received in revised form of results of biometric measurements of randomly selected shoots of basket willow, cut periodically from
15 July 2016
bushes at least 0.5 m tall, cultivated on an experimental plot. The description of the random variation of
Accepted 24 August 2016
willow shoot and bush was made with the use of a two-dimensional gamma distribution describing the
shoot mass and volume, as well as the Pascal distribution characterising the number of shoots in a plant.
Keywords: The result was an estimation of the variation of plant mass under the conditions of observation of one or
Basket willow (Salix viminalis) biomass more shoots. The very good fit of the model to the experimental data indicates that it is justified to accept
Two-dimensional gamma distribution
the model as a basic tool for the estimation of the amount of biomass on a plantation.
Pascal compound distribution
& 2016 Elsevier Ltd. All rights reserved.
Bush mass prediction
Yield calculator

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 843
2. The probabilistic model and its estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845
2.1. Materials and methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845
2.2. Two-dimensional distribution of mass and volume index of basket willow shoot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 846
2.3. Conditional distribution of basket willow bush mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 848
3. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 849
4. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 850
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 850
Appendix A. Some properties of two-dimensional gamma distribution of shoot mass and volume index (2) and conditional bush mass (7) . . 850
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 851

1. Introduction source is atmospheric precipitation. Due to its random character, it


determines to a significant extent the level of yields of energy
Environmental conditions have a significant effect on the pro- crops, and the kind of soil and the level of fertilisation can con-
ductivity of plants used for power generation purposes. The most tribute additionally to the variation in the yields of those plants.
important factors include the availability of water whose main Those factors are often used as input information for the model-
ling of biomass productivity. However, they require expert
knowledge for their interpretation.
n
Corresponding author. One of the plant species most frequently grown on energy crop
E-mail addresses: wojciech.jakubowski@up.wroc.pl (W. Jakubowski),
plantations in Poland and in the world is Salix viminalis L. [1,2]. Its
wieslaw.szulczewski@up.wroc.pl (W. Szulczewski),
andrzej.zyromski@up.wroc.pl (A. Żyromski), biomass can be acquired in one-, two-, three-, or four-year cycles
malgorzata.biniak-pierog@up.wroc.pl (M. Biniak-Pieróg). of harvest over a period of 20–25 years. Apart from energy

http://dx.doi.org/10.1016/j.rser.2016.08.048
1364-0321/& 2016 Elsevier Ltd. All rights reserved.
844 W. Jakubowski et al. / Renewable and Sustainable Energy Reviews 66 (2016) 843–851

Nomenclature distribution
H (q , v ) distribution function of binomial distribution of the
n total theoretical number of shoots mixture
Qj random variable – mass of j-th shoot (j = 1, … , n) a parameter of the mixture
Vj random variable – volume index of j-th shoot mqi , mvi , ri, βi , i = 1, 2 parameters of i-th gamma distribution
(j = 1, … , n) f1V (v ) , f2V (v ) distribution functions of marginal distributions of
W random variable – number of shoots in plant plant volume index, determined from density func-
W tions h1(q , v ) , h2(q , v )
M = ∑ j = 1 Q j random variable – mass of plant
(Q j, Vj ) binomial random variable – mass and volume index of E1(Q |V = v0) , E2(Q |V = v0) conditional expected value of binomial
j-th shoot gamma distribution
(Q j, |Vj = v0) conditional random variable – mass of j-th shoot D12(Q |V = v0) , D22(Q |V = v0) conditional variance of binomial
under the condition that its volume index is v0; in the gamma distribution
paper it is assumed that the random variables de- E (Q |V = v0) conditional expected value of shoot mass under the
scribed above have identical distributions, therefore condition that its volume index is v0
further on in the text the index j will be omitted D2(Q |V = v0) conditional variance of shoot mass under the
h1(q , v ) , h2(q , v ) density functions of binomial gamma condition that its volume index is v0
distribution m number of measured shoots
h(q , v ) density function of binomial distribution of the (qj , vj ) observed pairs of values of mass and volume index of
mixture j-th shoot (j = 1, … , m)
H1(q , v ) , H2(q , v ) distribution functions of binomial gamma wk observed number of shoots in k-th plant

generation, its wood is also used in soil reclamation, wastewater data concerning the weather conditions. Also necessary is the in-
purification and in plantings in protective green belts. formation on the diurnal concentration of CO2 and on nitrogen
Due to the worldwide interest in various cultivars of basket compounds. Another example is the model CANEGRO [19] which,
willow, and especially in their productivity under given environ- among the input data, apart from the meteorological and soil
mental conditions, that issue has been the subject matter of re- elements, requires also information on cultivation treatments ap-
search for many years. Models are developed that, due to their plied. A similar set of input data is required for the model EPIC [20]
structure, can be used for many plants, as well as such that are which has become the basis for the development of models
dedicated to specific plant species or even cultivars. Problems re- dedicated to specific plants, unfortunately with even greater de-
lated with the effect of local conditions on productivity are ad- gree of complexity, e.g. models ALMANAC or AUSCANE. Relatively
dressed, even on a regional scale, due to the need for developing the least data, limited to only several diurnal weather data, are
algorithms estimating the amount of biomass on willow planta- required for the model EPI, dedicated to opuntia and agave [21].
tions [3]. In general, the function of such models is the estimation The model developed for willow, model 3PG, requires detailed
of profitability of cultivation of selected plants, and identification weather data, and information on the leaves, shoots and root
of factors affecting biomass productivity. Research is also con- system of the plants, as well as soil parameters [22].
ducted in the aspect of the effect of specific plant cultivars on their On the other hand, the researchers of the problem still seek a
productivity, and especially on new willow cultivars with desirable possibility of reducing the number of input variables for the
genetic traits [4–6]. Also important is the problem of variation in models, conducting analyses aimed at the identification of domi-
plantation productivity in successive years of its exploitation and nant factors that determine plant growth and development. Evans
in relation to the variant of cultivation (1-, 2-year and longer et al. [23] indicated a high importance of precipitation totals and
harvest cycles) [7,8]. The above research issues are solved with the air temperatures in the period of willow growth, and a significant
use of empirical and mechanistic methods with varying degree of effect of annual precipitation totals and of soil pH and texture.
detail, both with respect to the model input data and to the time Aylott et al. [9] tried to identify the dominant factors of growth for
step of plant growth process simulation. Empirical models are three willow varieties, and in effect indicated the importance of
often limited to the environmental conditions at various geo- hydrological factors and soil structure. Likewise, Labrecque and
graphic locations. Most frequently they are based on field experi- Teodorescu [24], Bergante [25] indicated in their studies the sig-
ments (e.g. [9,10]), and thus they allow understanding of the effect nificance of climate data, and precipitations in particular, and of
of the local environmental parameters on plant yielding. soil structure, in relation to willow biomass increments.
Whereas, mechanistic models refer to biological and morpho- The models reviewed above, concerning primarily willow, and
logical processes to determine plant growth (e.g. [11–16]). At the studies on the effect of environmental parameters on yields,
present there is more than a dozen models of that type, permitting are focused on the analysis of cause-effect relations that transform
the estimation of biomass of selected energy crop plant species, i.e. environmental features on biomass increment. There is also a
miscanthus, switchgrass, sugar cane, willow, poplar, opuntia or group of models that, on the basis of biometric measurements,
agave. An extensive review of those models has been presented by most frequently the diameter and mass of dried shoots of willow
Nair et al. [17]. The models differ in terms of the number and [26,27,15], seek the correlation between shoot mass and diameter.
quality of input data, and of the degree of complexity. In a great In the literature, models representing such an approach are de-
majority the application of those models is based on meteor- fined as allometric models. A similar idea of model construction
ological data, even 24-h, detailed properties of a specific soil, e.g. was presented by Lupi et al. [28] who determined the mass of a
knowledge of the permanent wilting point of plants, field water bush using the mean lengths of the tallest shoots and their
capacity, porosity. An extreme example can be the model Agro- number in a bush. Another method in use is that of digital pho-
BGC [18] which – as input data – requires the determination of tography, in which – on the basis of digital photographs – the
several dozen constant parameters of vegetation, nine parameters density of a bush is determined and thus its mass is estimated
approximated at the stage of model identification, and six diurnal [29,30]. However, both the processing of digital photos and the
W. Jakubowski et al. / Renewable and Sustainable Energy Reviews 66 (2016) 843–851 845

direct biometric measurements in the field are time-consuming Telenius [37] conducted such measurements at the base of the
and burdened with an error that is hard to estimate. shoots.
In view of the above, the authors made an attempt at devel- The consequence of non-systematised methods of measure-
oping a model of growth of basket willow (Salix viminalis L.) on the ment of biometric features is the lack of a probabilistic model of
basis of its simple biometric features. The basis for this approach correlation of those features. Some authors restrict themselves to
was the conclusion that the current value of the biometric features only the elementary methods of statistics [27,15,31]. There is also
is determined by parameters affecting plant growth, i.e. that the an absence of statistically substantiated studies on the correlation
current amount of biomass had been determined by earlier me- between shoot features such as the length, diameter or mass,
teorological, environmental conditions, etc. Another assumption while such correlations are natural. This suggests the application
building the foundations of the study was that the number of of multi-dimensional distributions for their probabilistic descrip-
shoots in a plant is a function of the age of the plantation. For this tion. Unfortunately, serious difficulties appear during the de-
reason, detailed probabilistic analyses were undertaken concern- termination of the form of such a compound distribution. In such a
ing the correlation of simple non-destructive biometric measure- case not only the marginal distributions are of different types, but
ments and shoot mass and the number of shoots in a bush in also the significant relationships between the individual features
consecutive years of willow vegetation. are not always linear [26,38].
An earlier study by these authors [39] was concerned with a
one-dimensional probabilistic description of two biometric fea-
2. The probabilistic model and its estimation tures, i.e. the volume index determined on the basis of the length
and the diameter at its mid-point and of the mass of an individual
2.1. Materials and methods shoot of basket willow (Salix viminalis). Statistical tests demon-
strated that the distribution of those features is a mixture of two
One of the features of basket willow plantation is its known gamma distributions independent of the age of the plantation and
spatial structure, i.e. the method of bush plantings, which – to- of the time of measurement in the course of the vegetation period.
gether with the known number of bushes – permits the estimation The features are described correctly by a mixture of gamma dis-
of the mass of the entire plantation. In relation with the above, it is tributions for the second, third and the fourth year of cultivation.
natural to assume that the probabilistic properties of all bushes are An analysis of the distribution of the number of shoots in a bush in
similar and do not depend on their situation in the plantation. In relation to the age of the plantation was also performed. The
such a case it becomes necessary to exclude bushes growing on number of shoots in a bush was described by a shifted negative
the plantation perimeter from the modelling. Therefore, those binomial distribution. In each of the years of cultivation that dis-
bushes were treated as the protective belt of the plantation. In tribution described their number in a statistically correct manner.
consequence, the construction of a model of a single bush of Those analyses became the basis for the construction of an
basket willow is sufficient for the determination of a probabilistic innovative two-dimensional probabilistic model, presented fur-
model of the mass of the entire plantation. Therefore, a number of ther on in this paper, permitting the estimation of biomass on a
statistical methods whose main function is the analysis of spatial plantation. In consequence this allows the creation of an easy to
variability of plantings (see e.g. [31]) cannot find an application in use calculator for the estimation of willow biomass during the
this case. vegetation period for successive years of the plantation.
When building a model of a bush the fundamental thing is to In the field experiment plants were cultivated on a plot of
take into account the distribution of the number of its shoots. 4.9 m × 22.8 m , in 7 rows spaced at 70 cm. The spacing of plants in
Then, for the determination of the distribution of the mass of the the rows was 40 cm, which gave a potential stand of 392 plants.
whole bush it becomes natural to apply compound distributions, Their cultivation was conducted with the extensive method. The
taking into account also the features of an individual shoot. Dis- experimental plantation was established in 2010, while the model
tributions of that type [32] are often used in ecology, e.g. for the of biomass increments was developed on the basis of measure-
estimation of fish biomass ([33,34] or [35]), or of the biomass of a ments conducted on the plants in the period of 2011–2013 (2nd–
selected plant species [36]. Two distributions are estimated: a 4th year of the plantation). In the years adopted for the analyses
discrete one – of the number of individuals (fish shoals, plants etc.) varied numbers of plants started their vegetation – 364 plants in
and a continuous one – of the value of their biomass. Their mixture 2011 and 347 in the consecutive years.
estimates the biomass (of fish, plants) within the area under The measurements were performed for plants with the exclu-
analysis. In ecology the problem is the discontinuity of the dis- sion of those growing within the buffer belt on the perimeter of
tribution obtained in that manner. There exists a probability of the plot. Three biometric parameters were measured during the
zero mass, that corresponds to non-observation of any individual vegetation periods to determine the variation of basket willow
(e.g. no fish shoal, no plant) within a given area. That problem shoots. The measurements consisted in periodical cutting of ran-
does not appear in the estimation of the biomass of a bush of dom selected shoots with length over 0.5 m and measuring the
basket willow. length of the shoots cut, their diameter in the middle of the length,
In relation with the above, the key issue is to develop a method and then their weighing to determine the mass of the individual
of estimation of the features of an individual shoot on the basis of shoots. Every time a sample of similar numbers of shoots was
biometric measurements. The literature does not provide devel- taken, at uniform time intervals. Over the entire vegetation period
oped methods for conducting such measurements that would be the numbers of shoots cut and measured as described above
adopted as a standard. An example of this can be the lack of a rule amounted to 192 shoots in 2011, 166 shoots in 2012 and 175 shoots
for the adoption of the shoot height at which its diameter should in 2013.
be measured in the course of biometric measurements during the At the end of the vegetation period the number of all shoots
whole period of vegetation. Amichew et al. [15] measured shoot with length over 0.5 m was counted for all the plants growing on
diameter at the height of 30 cm above ground surface, Stolarski the plantation. The numbers varied from 1312 in 2011 (2nd year of
et al. [4] adopted the height of 50 cm, while Sevel et al. [5] – 90 cm. vegetation) to 2948 in 2013 (4th year of vegetation).
Nordh and Verwijst [26], for the estimation of willow productivity In the study the assumption was adopted that an individual
with a destructive and a non-destructive method, measured shoot shoot is characterised by a correlated two-dimensional random
diameters at 3 heights: 55, 85 and 105 cm, while Verwijst and variable (Q , V ), where Q is the shoot mass (including the mass of
846 W. Jakubowski et al. / Renewable and Sustainable Energy Reviews 66 (2016) 843–851

leaves) and V is the shoot volume index, what resulted from [39]. where Ir(u) is a Bessel function of the first kind. In formulae (1), (2)
The distribution of the variable is a mixture of two two-dimen- variable q denotes mass while v is the volume index of a single
sional gamma distributions. (More information on distributions of shoot. Parameters mqi – mean mass, mvi – mean volume index and
this type can be found in [40].) Chi-squared tests confirmed the ri – correlation coefficient of i-th (i = 1, 2) population of shoots, βi
appropriateness of the fitted distribution. Such an approach per- are scale parameters.
mitted the determination of shoot mass distribution (destructive The function of conditional density of the mass of a single
measurement) under the condition of its volume (non-destructive shoot, at a determined value of its volume Q |V = v0 has the form:
measurement). This is of particular importance in view of the
ah1(q, v0) + (1 − a)h2(q, v0)
application consequences of the method proposed. It was also h(q|v = v0) =
demonstrated [39] that the number of shoots in a basket willow af1V (v0) + (1 − a)f2V (v0) (4)
bush is correctly described by the shifted negative binomial dis-
where af1V (v0) + (1 − a)f2V (v0) is a marginal distribution – of the
tribution (Pascal distribution). The combination of those two dis-
mixture of gamma distributions – Annex (A.4).
tributions: the distribution of the number of shoots per plant and
Using the conditional expected values and the conditional
the conditional probability distribution determines the conditional
variances (A.9) the conditional expected value and the conditional
mass of basket willow bush with the assumption of the volume
variance of shoot mass were determined:
index of an individual shoot. Unfortunately, the variance of such a
distribution is large. This means that the application of such a af1V (v0)E1(Q |V = v0) + (1 − a)f2V (v0)E2(Q |V = v0)
E(Q |V = v0) =
method, consisting in the measurement of biometric features of a af1V (v0) + (1 − a)f2V (v0)
single shoot and concluding on that basis about the mass of an
average willow bush is burdened with an excessive error – prac-
tical application of the method would not make sense. That error af1V (v0)D12(Q |V = v0) + (1 − a)f2V (v0)D22(Q |V = v0)
D2(Q |V = v0) =
was reduced analysing the conditional distribution of mean mass af1V (v0) + (1 − a)f2V (v0) (5)
of n bushes with the assumption of the volume index of k shoots.
As expected, the results obtained permitted the reduction of the
scale of the error to a level that can be accepted in the practice of 2
a(1 − a)f1V f2V ( E1(Q |V = v0) − E2(Q |V = v0))
agricultural production. + 2
Attempts were also made at applying other types of compound ( af1V (v0) + (1 − a)f2V (v0))
distributions, e.g. based on the Poisson, Pascal or Tweedie dis-
Both values are the functions of moments (A.9). Let us note that
tributions [32]. Good fit was obtained only in the case of the Pascal
the expected values E1(Q |V = v0), E2(Q |V = v0) and the corresponding
distribution. In the case of the other models the results obtained
were not satisfactory, as e.g. in the model based on Poisson dis- variances D12(Q |V = v0) and D22(Q |V = v0) are linear functions of the
volume index v0, which means that for shoots from a homo-
tribution the problem is poor fit of the distribution of the number
geneous population both the conditional expected value
of shoots in the plant.
Most of the theoretical calculations and derivations of for- E (Q |V = v0) and the conditional variance D2(Q |V = v0) are linear
functions.
mulae, including those with the use of the Laplace transform, has
The two-dimensional distribution presented above (1) was
been transferred to the appended Annex.
used for the estimation of the distribution of observed pairs
(qi , vi )i = 1, … ,533 – mass and volume index of individual shoots. The
2.2. Two-dimensional distribution of mass and volume index of
data were obtained from measurements taken in the vegetation
basket willow shoot
periods of the years 2011–2013. The estimation of the unknown
In the study we analysed pairs of random variables parameters of the distribution (1) was conducted in three steps. In
(Q j, Vj ) , j = 1, … , n. Basing on the results of the research [39], it step one, as in reference [42], it was assumed that the scale
parameters β1, β2 are known and their values are equal to the
was assumed that the pairs constitute a sequence of independent
random variables with identical distribution. That distribution is a means of the estimators determined for the one-dimensional
mixture of two-dimensional gamma distributions with the density distributions. In step two the remaining parameters
function: (mqi , mvi , ri, i = 1, 2) were estimated with the maximum likelihood
method. The final parameter a – showing the proportion of se-
h(q, v) = ah1(q, v) + (1 − a)h2(q, v) (1) lection of the distributions, was obtained with the use of the ex-
q v pectation-maximisation algorithm [43].
and distribution function H (q , v ) = ∫ ∫ h(u, w )dwdu, where
0 0 The fit of the observations (qi , vi )i = 1, … ,533 to the two-dimensional
a (0 < a < 1) is a parameter of the mixture, and the density func- mixture (1) was tested with the goodness-of-fit test χ2. The test
tions (i¼1,2) was conducted by dividing the observations into classes according
⎡ ⎤ to the algorithm proposed by [44]. Table 1 presents the calculated
⎢ βi ⎛ q v ⎞⎥ p-values of test χ2 for the two-dimensional distribution and, for
hi (q, v) = exp⎢ − ⎜⎜ + ⎟⎟
(1 − ri ) ⎝ mqi mvi ⎠⎥ comparison, also for the marginal ones. In each case there were no
⎣ ⎦
grounds for the rejection of the goodness-of-fit hypothesis.
q βi − 1v βi − 1βi2βi
βi
f β (ciqv)
i
( (1 − ri)mqimvi) Γ (βi ) (2) Table 1
χ2 goodness-of-fit test of the observed shoot vo-
lume index and mass to two-dimensional and
are two-dimensional gamma distributions [41,42] with para-
marginal mixed gamma distributions (1).
βi2ri
meters mqi , mvi , ri, βi and ci = , while function fβ (z ) is
(1 − ri )2mqimvi i
Gamma distribution p-Value
defined as follows:
∞ Two-dimensional 0.13
zk 1 − βi
Marginal – volume index 0.11
f β (z ) = ∑ = z 2 I βi − 1(2 z ), i = 1, 2
Marginal – mass 0.17
i
k=0
k!Γ (βi + k ) (3)
W. Jakubowski et al. / Renewable and Sustainable Energy Reviews 66 (2016) 843–851 847

Graphical illustration of the fit is presented in Figs. 1–4.


Fig. 1 presents the estimated two-dimensional density function
(1). The arguments of the function are the volume index and the
mass of an individual shoot. The uni-modal area with a distinct
ridge, describes the variation of probability of basket willow shoot
features. In the graph two curves traversing the bounding area are
plotted that illustrate the variation of the estimated distribution
for preset values of volume index v0 = 40 cm3 and q0 = 50 g .
Projection of the probability density bounding area onto the
plane of shoot mass and volume index is shown in a graph (Fig. 2).
Asterisks are used to indicate the observed features of individual
shoots, and solid lines represent the lines of constant probability
(contour lines). They delineate the most probable areas Dp, such
that the probability of inclusion within each of the areas is p. On
the curves bounding areas D70% , D80% there is an observable change
Fig. 1. Two-dimensional estimated density function (1) of shoot mass and volume
in the shape of the contour lines. That is an effect of the mixture of
indicator, observations from 2011 to 2013.
distributions. Up to the inflection point shorter shoots with larger
diameter dominated, beyond that point thinner and longer shoots
were dominant. Table 2 presents also what percentage of the ob-
served masses and volume indices is included within the esti-
mated area. The results obtained confirm the very good fit of the
observed features of basket willow shoots to the proposed theo-
retical distribution (1).
Subsequent graphs (Figs. 3 and 4) present the estimated curves
of marginal densities. They are very similar to the curves obtained
during the estimation of each of the features separately. The
goodness of their fit is not much worse, which results from the fact
that the high correlation between both features is taken into ac-
count in the two-dimensional distribution (Table 3). In the case of
a mixture of two-dimensional gamma distributions defined in
such a manner upon transition to the marginal distribution the
information on the strength of the correlation is lost.
The form of both regression equations also carries important
information. With shoot mass of the order of several hundred
Fig. 2. Contour line of probability density bounding area with 0.6, 0.7, 0.8 and
grams in both equations the free argument of the order of a unity
0.9 probability, observations from 2011 to 2013.
is insignificant. It is the slope of the line that is of primary im-
portance, and here the variation is large. This supports the as-
sumption that we should not model the features of basket willow
shoots as observations from a homogeneous population. This
problem is noticeable in the study by [15], where the authors
adopted the simplifying assumption that only the longest shoots
from a bush are analysed.
Fig. 5 presents the observed shoot masses and volume indices
together with the regression curve (5) determined for the dis-
tribution of the mixture. The regression line (marked double) can
be divided into three sections: lower – up to shoot volume below
100 cm3, central – for shoot volumes within the range from 100 to
200 cm3, and upper – over 200 cm3. That division reflects the
domination of shorter shoots (lower section), and the domination
Fig. 3. Marginal distributions of basket willow shoot mass, observations from 2011 of long shoots (upper section). The centre interval corresponds to a
to 2013.
mixed situation.
In addition, using the distribution function of the mixed

Table 2
Theoretical probability and observed inclusion of basket willow shoot features to
the most probable areas Dp marked in Fig. 2.

Probability
Area Bounding area Dp [%]

Theoretical Observed

D60% 60 59.7
D70% 70 69.7
D80% 80 82.7
D90% 90 89.7
Fig. 4. Marginal distributions of basket willow shoot volume index, observations
from 2011 to 2013.
848 W. Jakubowski et al. / Renewable and Sustainable Energy Reviews 66 (2016) 843–851

Table 3
Regressive correlation of shoot mass with its volume index.

Shoot type Regression line Correlation

Shorter with a bigger diameter q^ = 1.075v + 1.696 0.972


Longer with a smaller diameter q^ = 0.800v + 3.086 0.984

Fig. 6. Estimated conditional basket willow bush mass density function (years
2011, 2012), assuming a shoot volume index of 100 cm3.

homogeneity of individual shoots – in 2011 dominant were shorter


shoots, with larger diameters.
Following a similar procedure as in the case of an individual
shoot, the conditional expected value and the variance of basket
willow bush mass were determined (8). Those values are de-
termined by differentiating a Laplace transform (A.7) or making
use of the properties of a compound distribution (8) [45,46]
Fig. 5. Prediction area of basket willow shoot mass.
E(M|V = v0) = EW E(Q |V = v0)

density function (A.6) quantiles q10% , q90% were calculated, de- D2(M|V = v0) = EW D2(Q |V = v0) + D2W E2(Q |V = v0) (8)
pendent on the volume index v = v0 , such that 2
where EW is the mean value and D W is the variance of the
H (q10|v = v0) = 0.1 and H (q90|v = v0) = 0.9 (6) number of shoots in a bush.
As in the case of a single shoot (5), the regression line obtained
The lines obtained (solid ones on the Fig. 5) constitute 80% of the
is a function of the volume index V = v0 . The regression lines differ
conditional prediction area of the shoot mass. Thus, we can esti- only in the value of coefficient EW. The situation is different in the
mate the shoot mass, knowing the easily measurable volume case of the conditional variance. Formula (8) indicates that it is a
index. quadratic function of the condition V = v0 . This also means that the
Theoretically, 80% of the observations should fall between the error of prediction will increase with increase in the value of the
quantile lines, prediction lines. In the example under analysis it is volume index, in a manner close to linear. This is important during
78%. Changing the prediction error similar results are obtained. the application of the method proposed. With increase of the
This means that once again we have confirmation of the very good number of shoots in a bush the value of the term D2WE2(Q |V = v0)
fit of the mixed two-dimensional gamma distribution to the ob- in formula (8) grows rapidly.
served masses and volume indices of basket willow shoots. In Fig. 7, apart from the regression lines plotted separately for
each of the years 2011–2013, the 80% prediction areas are also
2.3. Conditional distribution of basket willow bush mass marked. Similarly to (6), the conditional probabilities were de-
termined:
For the determination of the conditional distribution of basket
willow bush mass a compound distribution was used, built on the Pr(M ≤ q10|v = v0) = 0.1 and Pr(M ≤ q90|v = v0) = 0.9 (9)
basis of the conditional distribution of single sheet mass (4) and
Unfortunately, the prediction area determined in this way is
the Pascal distribution describing the variation of the number of
shoots in a bush [39]. The condition was the volume index of a
single shoot. The probability density function of the random event
M|V = v0 has the form
∞ ⎛ −ν ⎞
m(q|v = v0) = ∑ ( − 1)n⎜⎜ ⎟⎟pn (1 − p)ν h(*(n + 1))(q|v)
n= 0 ⎝ n ⎠ (7)

where (*n) denotes n-th convolution, while ν and p are parameters


of Pascal distribution.
Sample functions of conditional bush mass density (7) for the
years 2011 and 2012, calculated with the assumption that the
shoot volume index equals 100 cm3, are presented in Fig. 6. The
successive maxima are related with the Pascal distribution of the
number of shoots in a basket willow bush. The density functions
differ notable in terms of shapes and the heights of the local
maxima. The shift of the density function plot to the left, towards
lower bush mass, results from the smaller number of shoots in a
bush in the second year of the plantation (2011). The higher am- Fig. 7. 80% prediction of basket willow bush mass in 2011–2013, with the as-
plitudes of local maxima result from greater statistical sumption that the shoot volume index is v.
W. Jakubowski et al. / Renewable and Sustainable Energy Reviews 66 (2016) 843–851 849

population with the use of a single group of distributions the


obtained correctness of fit is so low that it cannot be used for
subsequent estimations. Only the application of a mixture of two-
dimensional gamma distributions in the study allowed correct
statistical modelling. As a result, such a procedure not only per-
mits the estimation of the correlation between the geometry of a
shoot and its mass, but also – through the use of compound dis-
tributions – it allows the estimation of the mass of a whole bush of
basket willow on the basis of biometric measurements of one or of
a larger number of shoots.
The proposed innovative model of probabilistic variation of
Fig. 8. Conditional density function of (mean) mass of basket willow bush in 2013,
with the assumption that the (mean) shoot volume index is 100 cm3. mass of basket willow allows to estimate the current mean mass of
a plant on the plantation on the basis of non-destructive biometric
measurements. The results obtained permit very good estimation
Table 4
Intervals of 80% prediction of mean mass [g] of k basket willow bushes in 2013 – with as high as 80% probability – of the mass of a whole bush
determined with the assumption that the mean shoot volume index is v . with a relative error not exceeding 22%, irrespective of the age of
the plantation and of the weather conditions during the vegeta-
Number of measurements k v = 100 cm3 v = 150 cm3 v = 200 cm3 tion period. The only condition is the performance of at least 10
1 (532,2213) (713,2963) (885,3676) simple non-destructive biometric measurements on random-se-
3 (829,1822) (1110,2440) (1379,3028) lected shoots on the willow plantation. Numerical and probabil-
5 (929,1702) (1244,2280) (1546,2833) istic analyses demonstrated that increasing the number of bio-
10 (1034,1583) (1385,2125) (1721,2644)
metric measurements of willow shoots had no significant effect on
estimation improvement. This results from the observed variance
of the correlation of shoot mass with volume index, which causes
very broad. In the case of measurement of a single shoot with
that any increase in the number of measurements – even up to
volume index of 150 cm3, 80% prediction of bush mass in 2011
several hundred – has no significance. This observation is very
varied from about 380–1150 g, in 2012 from 790 to 2540 g, and in
important when using models of this type, as it allows to propose
2013 – within the range from 710 to 2960 g. This is too big an error
a procedure that is the least cost- and time-consuming with un-
for use in the estimation of biomass on a plantation.
In this situation the number of shoots measured was increased. changed accuracy of yield estimation. The very good results ob-
Already when the number of measurements was increased to tained confirm unquestionably that the method of biometric
three the range of variation of mean mass of three bushes was measurements of basket willow, proposed by these authors, is
narrowed notably. This is presented in Fig. 8. For data from 2013 correct and provides good statistical description, which allows it to
the function of probability density of bush mass or mean mass for be proposed as a standard in research of this type.
three bushes is shown with the measurement of one, three or ten Extrapolation analyses related with the age of the plantation
shoot volume indices, respectively. The variation of the mean mass permit the formulation of the hypothesis that the correlation of
of ten bushes was considerably reduced. It corresponds approxi- the willow bush mass with the shoot volume index and the re-
mately to that which can be achieved with the use of the Linde- lative error of estimation are of an asymptotic character. In the first
berg central limit theorem [47] when measuring 100 shoots. case, with increase of the plantation age there is an increase of the
Table 4 presents sample intervals of estimation of mean bush ratio of bush mass to shoot volume index, but it does not exceed
mass for various (mean) values of shoot volume indices and 15, while the relative error, at a minimum of 10 random mea-
numbers of their measurements. That estimation was performed surements of shoot volume index does not exceed 25%. For ex-
for the Pascal distribution [39], for the third year of the plantation. ample, in the eighth year of the plantation, for the mean shoot
With increase of the mean volume index the relative error in- volume index of 150 cm3 the extrapolated relative error of the
creases and at ten measurements it does not exceed the value of obtained estimate increases to 23.5%.
21.5%. The biometric measurements related with the shoot volume
index and mass were conducted at various moments of vegetation
in different years of plantation cultivation and at notable variation
3. Discussion of weather conditions. That confirms the permanent correlation
between shoot mass and volume index irrespective of those fac-
Allometric models developed so far, basing on frequently
tors. That last conclusion is of particular importance with regard to
questionable or subjectively adopted methods of biometric mea-
the application of the obtained results in practice. It will allow the
surements [5,28], that are not substantiated with statistical ana-
development of a simple software tool for the estimation of the
lysis, cannot provide a reliable answer to the question of the
amount of biomass on a plantation. That tool, on the basis of at
probabilistic accuracy of the estimates obtained with their use.
least 10 non-destructive measurements of random-selected
This results from the lack of statistical elaboration of the features
under analysis. The presented results are limited to giving the shoots, with a known number of bushes of basket willow, will
coefficient of determination comparing the developed model with permit the estimation of the amount of biomass on the plantation
direct measurement data. This may result from the fact, which has at any moment during its vegetation. The simple biometric mea-
been demonstrated by the authors of this study, that the entire surements can be made by the farmers themselves, with no fi-
population of the willow shoot features under analysis (volume nancial outlays and without causing any major disturbance on the
index and mass) can be divided into two homogeneous popula- plantation. And the information acquired in that manner can be
tions described with gamma distributions. Those features become used for the estimation of production capacity for the purpose of
apparent only in the third and fourth years of cultivation. In the meeting the requirements of specific recipients of biomass for
case of other authors' study of biometric features of an entire energy generation.
850 W. Jakubowski et al. / Renewable and Sustainable Energy Reviews 66 (2016) 843–851

4. Conclusions This means that density function (2) is infinitely divisible and that
its n-th convolution is also a gamma distribution (two-dimen-
The analyses of the range of applicability of the developed sional). Hence, the distribution of the sum of n shoots (of a single
model of yielding of basket willow, conducted within the scope of type), defined as (Q1 + Q 2 + ⋯ + Q n, V1 + V2 + ⋯ + Vn), also has a
the study, permit the formulation of the following conclusions: gamma distribution, but with an altered scale parameter equalling
nβ .
1. Due to the fact that weather condition have no effect on the The marginal distribution of distribution (A.2) are standard
correlation of the biometric features of shoots and bush of one-dimensional gamma distributions from the density function:
basket willow, the model developed is independent from them.
2. The two-dimensional distribution of the features, i.e. shoot vo- q β − 1 − pq v β − 1 − pv
fQ (q) = e 1 fV (v) = e 2.
lume index and shoot mass, permit the determination of the p1β Γ (β ) p2β Γ (β ) (A.4)
area of the most probable occurrence of correlation of the two
features. It is situated along the line of regression, with a dis- Let us take note that both of the marginal distributions do not
tinctly marked area of transition of short and thick shoots into depend on the coefficient of correlation r. Parameter p12, directly
long and thin ones. related with the coefficient of correlation r, becomes reduced as a
3. The use of compound distributions permits the determination result of integration. This means that the estimators of parameters
of the distribution of mass of basket willow bush in relation to of one-dimensional distributions will differ from the equivalent
the volume index of the individual shoots. estimators of parameters of two-dimensional ones.
4. The theoretical results obtained permit the construction of an Relatively simple calculations allow to determine the moments
easy to use calculator for the estimation of basket willow bio-
(expected values, variance and coefficient of correlation) of dis-
mass on a plantation on the basis of the known structure of its
tribution (2).
plantings. The only input data are standard measurements of
the length and diameter at mid-length of 10 random-selected EQ = mq = qp1 EV = mv = qp2
willow shoots on the plantation, that can be taken by the p1p2 − p12
farmers themselves. D2Q = qp12 D2V = qp22r (Q , V ) =
p1p2 (A.5)
5. The numerical and probabilistic analyses performed in the
scope of the study demonstrated that any increase of the The probability density function of shoot mass, with the as-
number of biometric measurements of willow shoots above 10 sumption that its volume index is v, is as follows:
had no significant effect on improving the quality of the
estimations. ⎛ p q+pv ⎞ β β− 1
v ⎟ p2 q
g (q|v) = exp⎜⎜ − 2 1
+ f β (cqv)
⎝ p12 p2 ⎟⎠ p12
β
(A.6)
Acknowledgements The Laplace transform of conditional density (A.6) is as follows:

Research financed from Polish budget funds for science in the p v


− p1 + pv p2β p12 cv
[g (q|v)](s ) = ω(s ) = e 12 2
β
e p12 s + p2
years 2011–2014 as research Project no. NN305 3835 39. ( p12s + p2) (A.7)

This means that distribution (A.7) is infinitely divisible. Its n-th


Appendix A. Some properties of two-dimensional gamma convolution is a distribution of the same type as (A.6) and has the
distribution of shoot mass and volume index (2) and condi- form:
tional bush mass (7)
⎛ p q + np v ⎞ nβ nβ − 1
nv ⎟ p2 q
g *n(q|v) = exp⎜⎜ − 2 1
+ fnβ (cqnv)
For simplicity, in the formulae below we will omit the index ⎝ p12 p2 ⎟⎠ p12nβ
(A.8)
i, i = 1, 2 numbering the components of the mixture (1). Density
function (2), a component of the mixture, can be written in a Differentiating the Laplace transform (A.7) the moments used
simpler form using the set of parameters {p1, p2 , p12 , β}. They are in formula (5) – conditional expected value and conditional var-
related with parameters mq , mv , β , r in the following way: iance – were determined.
mqmv mq
mq = p1β, mv = p2 β, p12 = (1 − r ) E(Q |V = v) = mq(1 − r ) + rv
β2 (A.1) mv
mq2(1 − r )2 2mq2
In the Annex below all calculations will be conducted in the con- D2(Q |V = v) = + r (1 − r )v
vention of parameters (A.1). β βm v (A.9)
Making use of (A.1), the density function (2) has the form
The conditional expected value informs about the regression
⎛ p q + p v ⎞ q β − 1v β − 1 relation, and the conditional variance about its variation. It is easy
g (q, v) = exp⎜⎜ − 2 1 ⎟
⎟ β f β (cqv) to notice that both values are linear. As the variation of shoot mass
⎝ p12 ⎠ p12 Γ (β ) (A.2)
is characterised by standard deviation (variance square root), it
p1p2 − p12
where c = 2
and fβ (z ) is defined in a similar manner to that means that with an increase of the volume index v the variation of
p12
in Eq. (3). mass and the error of estimation increase proportionally to that
Direct calculations, as well as the results contained in [41], root.
demonstrate that the Laplace transform of the density function The Laplace transform ψ (s ) of the mixture of distributions (1) is
(A.2) is a mixture of transforms (A.7). This means that the distribution of
the mixture is not infinitely divisible. The situation is similar in the
1
[g (q, v)](s , u) = case of the Laplace transform of the conditional distribution of
(1 + sp1 + up2 + sup12) β (A.3) basket willow mass; it has the following form:
W. Jakubowski et al. / Renewable and Sustainable Energy Reviews 66 (2016) 843–851 851

⎛ 1 − p ⎞ν ASAE 1989;32:497–511.
[m(x|y)] = ψ (s )⎜ ⎟ [21] de Cortázar VG, Nobel P. Worldwide environmental productivity indices and
⎝ 1 − pψ (s ) ⎠ (A.10) yield predictions for a cam plant, Opuntia ficus-indica, included effects of
doubled CO2 levels. Agric For Meteor 1990;49:261–79.
Hence, the distribution (A.10) is not infinitely divisible. It means [22] Landsberg J, Waring R. A generalized model of forest productivity using
that the analytical form of the mass density function of any two simplified concepts of radiation-use efficiency, carbon balance and partition-
ing. For Ecol Manag 1997;95:209–28.
bushes of basket willow is unknown. In the calculations presented [23] Evans S, Baldwin M, Henshall P, Matthews R, Morgan G, Poole J. et al., Yield
in Section 2.3 a numeric approximation of convolution (A.10) was models for energy: Coppice of poplar and willow. volume a – empirical
applied during the determination of basket willow mass predic- models. Tech. rep., Biometrics Surveys and Statistics Division Forest Research,
tion. The calculations were performed using the Talbot algorithm Farnham; 2007.
[24] Labrecque M, Teodorescu T. High biomass yield achieved by salix clones in
[48]. SRIC following two 3-year coppice rotations on abandoned farmland in
southern Quebec, Canada. Biomass–Bioenergy 2003;25(2):135–46.
[25] Bergante S, Facciotto G, Minotta G. Identification of the main site factors and
management intensity affecting the establishment of short-rotation-coppices
References
(SRC) in northern Italy through stepwise regression analysis. Cent. Eur. J. Biol.
2010:522–30.
[1] Kuzovkina YA, Volk T. The characterization of willow (salix l.) varieties for use [26] Nordh M, Verwijst T. Above-ground biomass assessments and first cutting
in ecological engineering applications: co-ordination of structure, function cycle production in willow (salix sp.) coppice – a comparison between de-
and autecology. Ecol Eng 2009:1178–89. structive and non-destructive methods. Biomass–Bioenergy 2004;27:1–8.
[2] Krasuska E, Rosenqvist H. Economics of energy crops in Poland today and in [27] Arevalo C, Volk T, Bevilacqua E, Abrahamson L. Development and validation of
the future. Biomass–Bioenergy 2012;38:23–33. aboveground biomass estimations for four salix clones in central New York.
[3] Mola-Yudegoa B, Aronsson P. Yield models for commercial willow biomass Biomass–Bioenergy 2007;31(1):1–12.
plantations in sweden. Biomass–Bioenergy 2008;32:829–37. [28] Lupi C, Larocque G, DesRochers A, Labrecque M, Mosseler A, Major J, et al.
[4] Stolarski M, Szczukowski S, Tworkowski J, Klasa A. Productivity of seven clone Evaluating sampling designs and deriving biomass equations for young
sof willow coppice in annual and quadrennial cutting cycles. Biomass–Bioe- plantations of poplar and willow clones. Biomass–Bioenergy 2015;83:196–
nergy 2008;27:1227–34. 205.
[5] Sevel L, Nord-Larsen T, Raulund-Rasmussen K. Biomass production of four [29] Ens JA, Farrell R, Bélanger N. Rapid biomass estimation using optical stem
willow clones grown as short rotation coppice on two soil types in Denmark. density of willow (salix spp.) grown in short rotation. Biomass–Bioenergy
Biomass–Bioenergy 2012;46:664–72. 2009;33:174–9.
[6] Liu B. Biomass production of willow short-rotation coppice across sites and [30] Hangs RD, Van Rees K, Schoenau J, Guo X. A simple technique for estimating
determinants of yields for SV1 and SX61 [Ph.D. thesis]. State University of New above-ground biomass in short-rotation willow plantations. Biomass–Bioe-
York; 2013. nergy 2011;35:2156–62.
[7] Sleight NJ, Volk T. Recently bred willow (salix spp.) biomass crops show stable [31] Zapata-Cuartas M, Sierra C, Alleman L. Probability distribution of allometric
yield trends over three rotations at two sites, [Online; posted 1-April-2016]; coefficients and bayesian estimation of aboveground tree biomass. For Ecol
April 2016. 〈http://link.springer.com/article/10.1007%2Fs12155-016-9726-2〉. Manag 2012;277:173–9.
[8] Sleight NJ, Volk T, Johnson G, Eisenbies M, Shi S, Fabio E, et al. Change in yield [32] Scott D, Bravington M. A Poisson-Gamma model for analysis of ecological non-
between first and second rotations in willow (salix spp.) biomass crops is negative data. Environ Ecol Stat 2013;20:533–52.
strongly related to the level of first rotation yield. Bioenergy Res 2016;9:270– [33] Hiroshi S. Application of Tweedie distribution to zero-catch data in CPUE
87.
analysis. Fish Res 2008:154–62.
[9] Aylott MJ, Casella E, Tubby I, Street N, Smith P, Taylor G. Yield and spatial
[34] Tascheri R, Saavedra-Nievas J, Roa-Ureta R. Statistical models to standardize
supply of bioenergy poplar and willow short-rotation coppice in the UK. New
catch rates in the multi-species trawl fishery for Patagonian grenadier (Ma-
Phytol 2008;178(2):358–70.
cruronus magellanicus) off Southern Chile. Fish Res 2010;105:200–14.
[10] Tallis MJ, Casella P, nad Henshall E, Aylott M, Randle T, Morison J, et al. De-
[35] Lecomte J, Benot M, nad Etienne HP, Bel L, Parent E. Modeling the habitat
velopment and evaluation of ForestGrowth-SRC a process-based model for
associations and spatial distribution of benthic macroinvertebrates: a hier-
short rotation coppice yield and spatial supply reveals poplar uses water more
archical Bayesian model for zero-infated biomass data. Ecol Model
efficiently than willow. GCB Bioenergy 2013;5(1):53–66.
2013;265:74–84.
[11] Kiniry J, Williams J, Gassman P, Debaeke P. A general, process-oriented model
[36] Chen J, Shiyomi M, Hori Y, Yamamura Y. Frequency distribution models for
for two competing plant species (A). Trans ASAE 1992;35:801–10.
[12] Clifton-Brown J, Neilson B, Lewandowski I, Jones M. The modeled productivity spatial patterns of vegetation abundance. Ecol Model 2008;211:403–10.
of Miscanthus x giganteus (GREEF et DEU) in Ireland. Ind Crop Prod 2000:97– [37] Verwijst T, Telenius B. Biomass estimation procedures in short rotation for-
109. estry. For Ecol Manag 1999;121:137–49.
[13] van den Broek R, Vleeshouwers L, Hoogwijka M, van Wijk C, Turkenburg W. [38] Mukti RS, Ram P. Allometric biomass models for bark of Cinnamomum tamala
The energy crop growth model SILVA: description and application to eu- in mid-hill of Nepal. Biomass– Bioenergy 2012;47:44–9.
calyptus plantations in Nicaragua. Biomass–Bioenergy 2001;21:335–49. [39] Żyromski A., Szulczewski W., Biniak-Pieróg M., Jakubowski W. The estimation
[14] Sannervik A, Eckersten H, Verwijst T, Kowalik P, Nordh N. Simulation of willow of basket willow (Salix viminalis) yield – New approach. Part I: Background
productivity based on radiation use efficiency, shoot mortality and shoot age. and statistical description. Renew Sustain Energy Rev 2016; 65: 1118 - 1126.
Eur J Agron 2006;24:156–64. [40] Balakrishnan N, Chin-Diew L. Continuous bivariate distributions. Springer;
[15] Amichew B, Hangs R, Van Rees K. A novel approach to simulate growth of 2009.
multi-stem willow in bioenergy production systems with a simple process- [41] Bernardoff P. Which multivariate gamma distributions are infinitely divisible?
based model (3PG). Biomass–Bioenergy 2011;35(1):473–88. Bernoulli 2006;12(1):169–89.
[16] Jing Q, Conijn S, Jongschaap R, Bindraban P. Modeling the productivity of [42] Chatelain F, Tourneret J, Inglada J, Ferrari A. Bivariate gamma distributions for
energy crops in different agro-ecological environments. Biomass–Bioenergy image registration and change detection. IEEE T Image Process 2007;16
2012;46:618–33. (7):1796–806.
[17] Nair SS, Kang S, Zhang X, Miguez F, Izaurralde R, Post W, et al. Bioenergy crop [43] McLachlan GJ, Krishnan T, Ng SK. The EM Algorithm, vol. 24. Humboldt:
models: descriptions, data requirements and future challenges. GCB Bioenergy Universität Berlin, Center for Applied Statistics and Economics (CASE); 2004.
2012;4:620–33. [44] Rosenblatt J. Note on multivariate goodness-of-fit tests. Ann Math Stat
[18] Di Vittorio AV, Andersen R, White J, Miller N, Running S. Development and 1962;33(2):807–10.
optimization of an Agro-BGC ecosystem model for C4 perennial grasses. Ecol [45] Schmidli H. Risk theory. University of Aarhus; 2006.
Model 2010;221:2038–53. [46] Wüthrich M. Non-life insurance: mathematics & statistics.Zurich: ETH; 2013.
[19] Inman-Bamber NG. A growth model for sugarcane based on a simple carbon [47] Resnick S. A probability path. Birkhauser; 1999.
balance and the ceres maize water balance. S Afr J Plant Soil 1991;8:93–9. [48] Murli A, Rizzardi M. ALGORITHM 682, Talbot's method for the laplace inver-
[20] Williams JR, Jones C, Kiniry J, Spanel D. The EPIC crop growth model. Trans sion problem. ACM Trans Math Softw 1990;16(2):158–68.

Das könnte Ihnen auch gefallen