Académique Documents
Professionnel Documents
Culture Documents
=
1
2
, ,
) (
Equation /-/
From the >*8(, '>*S, and >*S model fits, an Anova comparison was tested, loo)ing closel!
at the A-C and 98S( values as a measure of comparison3
Table /-/ Ano%a and ;:&E comparisons of different modelling fitting met#ods on t#e <e$a dataset
Plot .spha0 8odel df A-C +-C Test p-value 98S(
1,4
'>*S 4 %H34G G%3%4 #3J$F$G"J
>*8( H G$3GJ GJ3G% $ vs " #3$4FI #3I%$GH44
>*S % 4J3#H HF34$ " vs F T3###$ #3J$"H#J$
,4@
'>*S 4 $#"3#G $$F3%# #344H$4J4
>*8( H JJ3I4 $$F3$# $ vs " #3#%#4 #344H"#4J
>*S % $#43F4 $$F3JF " vs F #3##GI #34G4HJ$H
/@1
'>*S G $II3$# $JJ3HG #3I#FH%#$
>*8( H $%I3G" $4%3IF $ vs " T3###$ #3I$G$%J%
>*S % $J#3"$ $JJ3GF " vs F T3###$ #3I#"F4#J
434
'>*S 4 "%$3JJ "G43JF $3$GGFGG
>*8( H $4G3$" $I"3G% $ vs " T3###$ #3JGI$4%
>*S % "I43$$ "J43#4 " vs F T3###$ $3$G%"%J
@41
'>*S 4 FIJ3$H %#43H% $3#%IG$J
>*8( H "4G3GI "I43#H $ vs " T3###$ #34HIH4F"
>*S % %$"34J %"%3%# " vs F T3###$ $3#%I%H"
344
'>*S G %4"3G4 %HI3$G $3#44G#H
>*8( H F"$34H F%F3G# $ vs " T3###$ #3HJFFIHJ
>*S % G#"3%G G$%3J" " vs F T3###$ $3#4FFHJ
144/
'>*S 4 GIJ3H$ 4$#3F4 #3JJ#$$JH
>*8( H %II3IG G$"3JG $ vs " T3###$ $3"%G$"H
>*S % 4F#3H# 4%%3%H " vs F T3###$ #3JF$FG4J
,365
'>*S 4 IH"3H# IJ43$I #3I%HI4I%
>*8( H GH%3IF 4#"3"" $ vs " T3###$ #3GIJ444H
>*S % JFG3"# JG#3IG " vs F T3###$ #3I%H#HG4
According to the A-C values, &esides one plot, the >*8( fitted model outperformed the
other model t!pes with p-values all &elow #3#G, suggesting a significantl! improved fit on all
of the plots3 Furthermore, it most resulted in lower 98S( values, indicating an overall
improved fit3
#.2.9 3:*7 final results 0model selection1
The summar! of the >*8( models for all of the four sites is represented in Table "#$= with
illustrated curves n Figure /-@3 hat is noticea&le is that the site inde/ is not strictl!
sensitive to the stems per hectare .as opposed to mean height0, there is a general tendenc!
Stellenbosch University http://scholar.sun.ac.za
F-FI
of a densit! effect on site inde/ .Figures /-@ and /-1-03 This tendenc! must &e ta)en into
consideration when site inde/ is used for prediction of growth3 However, as previousl!
stated, each plot was treated as an individual with its corresponding potential height
.calculated in step two0, which are e/plained in the third section of this chapter results3
The >*8( model was finall! chosen for model fitting as it provided a &etter fit for the
relevant data range .G# !ears03 However, with the associated pro&lems using the as!mptote
as a random effect, to achieve an improved fit for the entire data range, e/trapolation
&e!ond G# !ears ma! &e compromised as the as!mptote seems to &e lower than the other
fitting methods used3 -t was decided that, seeing as it is unli)el! that plantations will &e
grown for such long periods for commercial purposes, this would &e an accepta&le
compromise3
The overestimation of the model at !oung ages, which is most li)el! an effect of the
infle/i&le Chapman-9ichards e2uation was not solved &! the >*8(3 However, to remain
compati&le to S-*EA the compromise of appl!ing the Chapman-9ichards e2uation was
made3
Stellenbosch University http://scholar.sun.ac.za
FJ
Table /-4 Fitted &0-age models for t#e 77T trials
<e$a :ac :ac Dukuduku Awambonambi
Densit" 2alue
&td.
Error
DF t-%alue
p-
%alue
2alue
&td.
Error
DF t-%alue
p-
%alue
2alue
&td.
Error
DF t-%alue
p-
%alue
2alue
&td.
Error
DF t-%alue
p-
%alue
1,4
a $$H34G H%3IH $H $3GH #3$F%G %G3GF $3FJ I# F"3IG # "G3$$ $3F# %G $J3"4 # "I3GF $3JF %H $%3HI #
b -#3#$ #3#$ $H -$3#" #3F"$4 -#3#% #3## I# -$"3HI # -#3#4 #3#$ %G -43%H # -#3#4 #3#$ %H -4344 #
c #3IF #3#4 $H $F34" # $3$I #3#F I# FH3$J # $3#$ #3#J %G $$34J # $3#J #3#H %H $43F% #
,4@
a %"3FI F3I" %$ $$3#I # FH34% $3G# J$ "G3#H # "I3"F $3"$ I# "F3"I # %I3JI G3#$ H% J3HI #
b -#3#F #3#$ %$ -%3GI # -#3#H #3#$ J$ -$$3GI # -#3#4 #3#$ I# -I34H # -#3#" #3## H% -G3$G #
c #3J4 #3#I %$ $$3G4 # $3%F #3#G J$ "43#J # $3$" #3#H I# $43J% # #3IJ #3#% H% "%3H% #
/@1
a %$3$# "34$ 4" $G3H% # FI3FH #3IJ $$F %F3"% # "J3%4 $3"F $$% "%3## # %"3G# $3J4 J$ "$34% #
b -#3#F #3#$ 4" -43#% # -#3#4 #3## $$F -$43JJ # -#3#% #3## $$% -J3$4 # -#3#F #3## J$ -J3F% #
c $3#" #3#I 4" $F3"% # $3FH #3#% $$F FG3IH # #3J% #3#% $$% "$3$4 # #3IJ #3#F J$ F#3#4 #
434
a F%34" $3GH HF ""3#4 # %$3%4 $3I# $"H "F3#G # "I3HG $3$$ $F4 "G3JJ # FJ3IG $3G$ $$H "43%# #
b -#3#G #3#$ HF -I3I% # -#3#G #3## $"H -$$3GJ # -#3#% #3## $F4 -$#3#F # -#3#F #3## $$H -$$3#G #
c $3$G #3#H HF $H34F # $3"F #3#% $"H F"3#$ # #3JG #3#% $F4 "F3$I # #3J" #3#F $$H F$3%# #
@41
a FI3#I $3%J $$H "G34% # %"3"4 "3$F $G$ $J3IH # F#344 $3$G $GJ "43GI # G#3FJ F3F$ $GG $G3"" #
b -#3#% #3## $$H -$#3"G # -#3#G #3## $G$ -$$3$% # -#3#F #3## $GJ -J3"$ # -#3#" #3## $GG -H3"% #
c $3#" #3#% $$H ""3JG # $3"4 #3#% $G$ FF3$H # #3IF #3#% $GJ ""3FH # #3I% #3#F $GG F#3%J #
344
a %G3F4 F3F% $%F $F3GI # %G3$H "3FG $I% $J3"F # "J3$I #3H" ""I %#34H # %H3"H "3"4 $I$ "#3J# #
b -#3#F #3## $%F -434# # -#3#G #3## $I% -$$3"% # -#3#% #3## ""I -$F3#% # -#3#" #3## $I$ -J3GG #
c #3JF #3#% $%F "%3"G # $3"% #3#% $I% F%3HI # #3J4 #3#% ""I "F3$% # #3IG #3#" $I$ F43$" #
144/
a F#3H% #3HH $I" FJ3H4 # %$3J% $3F4 "4I F#3II # F"3%F $3FJ "I$ "F3%$ # %"3JF $3HG "%4 "%3G% #
b -#3#4 #3## $I" -"$3HI # -#3#G #3## "4I -$43$4 # -#3#" #3## "I$ -I3IG # -#3#" #3## "%4 -$#3IJ #
c $3%4 #3#F $I" %J3$J # $3$J #3#F "4I %F3HG # #3HJ #3#F "I$ "H3"I # #3I4 #3#" "%4 FI3"F #
,365
a "I3%FH #3GIGJ4 F4" %I3GF #
%"3FH #3JH %J$ %F3IG # "H3"I #34$ G$I %%3I" # F%34F $3$G FII F#3#4 #
b -#3#GJJ #3##"$H F4" -"H34F #
-#3#G #3## %J$ -"%3I$ # -#3#G #3## G$I -$H3F4 # -#3#% #3## FII -$F3I" #
c $3F4J$H #3#$I$ F4" HG344 #
$3$J #3#" %J$ HF3$# # $3#" #3#F G$I F43"4 # #3J$ #3#" FII %43GI #
Stellenbosch University http://scholar.sun.ac.za
%#
Figure /-@ Fitted dominant #eig#t cur%es for t#e 77T trials
Stellenbosch University http://scholar.sun.ac.za
%$
#.# tep 2: Potential height modelling
>onlinear 2uantile regression is one possi&le method to 2uantif! potential height growth3
Kuantile regression was introduced &! :oen)er and +assett .$JHI0 as a statistical set of
methods for defining conditional means in data anal!sis3 Another option would &e to su&set
the data into 2uantiles and perform some t!pe of least s2uares estimation to achieve a
nonlinear fit for that su&set= however 2uantile regression was preferred as dividing the data
into su&sets and achieving a mean, instead of median response delivers differing values=
more elo2uentl! descri&ed in :oen)er and Halloc) ."##$03
#.#.1 .escription of the 3:/; procedure
Kuantile regression is a generalisation of the median regression3 hile the latter was
introduced as a ro&ust method for dealing with outliers, which used the median instead of
the arithmetic mean to fit a regression curve, 2uantile regression is a&le to use an! 2uantile
.denominated &! the AtauD value0 of the distri&ution to fit a regression curve to it .Cade and
>oon "##F03 Thus it is an appropriate techni2ue to fit potentials3 >onlinear 2uantile
regression was fit using the nlrq function in the quantreg pac)age in 9 .:oen)er "##403
-n this stud!, tau vectors of #3J, #3JG, #3JHG and #3JJ 2uantiles .which represent the
conditional 2uantile fits, e3g3 #3G represents the G#N - or median fit0 were used to determine
the sensitivit! of the nonlinear 2uantile regression on the parameterisation data3 As a
starting point for the potential modifier approach, the ma/imum possi&le definition must
first &e determined as the modifier onl! reduces the potential height value .Pretzsch "##J03
After careful consideration and visual inspection, it was decided to use the #3JHG 2uantile
for the potential height definition3 hile it does not achieve the ma/imum values for height
growth, it seems more fle/i&le than the #3JJ tau vector and does not e/clude too much of
the ma/imum potential height series3 The effects of these choices are discussed further on3
>onlinear 2uantile regression proved to &e a ro&ust and eas! method to o&tain potential
height-age growth series3 +! specif!ing the #3JHG 2uantile the fits were more consistent3
The potential height fits are shown in Figure /-4 &elow, fitted over their respective spacing
trial series for each plot .spha03 The results per plot are summarised in Table /-5.
Stellenbosch University http://scholar.sun.ac.za
%"
Dukuduku Awambonambi
<e$a :ac :ac
Figure /-4 Fitted potential #eig#t cur%es for t#e 77T dataset for different stems per #ectare.
From simpl! o&serving the growth curve series < a trend appears .seen within each trial
series0 that the #3JHG 2uantiles seem to &e affected &! densit! onl! to a small degree,
although the e/treme high and low ranges of stand densit! .e3g3 $"% and "J4G spha plots0
tend to dominate the upper and lower &ounds respectivel!3 However the difference
&etween sites does not seem to overtl! densit! dependent, &ut ma! &e more strongl!
dependent on the site 2ualit! or location of the plots, suggesting that a site 2ualit! measure
ma! &e useful for predicting potential height3
Stellenbosch University http://scholar.sun.ac.za
%F
-n order to see whether these potential heights are representative, visual assessment was
done on the PSP dataset3 Figure /-3 &elow shows that for the &est site in the CCT trials .the
8ac 8ac lowest densit! plot0= the potential captures the upper &ound of the PSP height
series 2uite well, &esides a few outliers, suggesting that the potential models fitted on the
spacing trials represent the upper &oundar! of o&served potentials3
Figure /-3 )otential #eig#t of t#e #ig#est 77T trial %alue (:ac+ sp#a B 1,4. plotted o%er )&) data of a wide range of
sites+ t#e blue line is t#e potential of t#e 77T plot+ t#e red line represents t#e mean #eig#t of t#e )&) dataset
Stellenbosch University http://scholar.sun.ac.za
%%
Table /-5 )otential #eig#t fits b" plot for t#e four spacing trials
<e$a Dukuduku :ac :ac Awambonambi
Densit" 2alue
&td.
Error t %alue )r(CDtD. 2alue
&td.
Error t %alue )r(CDtD. 2alue
&td.
Error t %alue )r(CDtD. 2alue
&td.
Error t %alue )r(CDtD.
,365
a %$3"4 "3I$ $%34H # FJ3"H "34J $%34$ # GJ3"F G34$ $#3GG # %43$4 F3JG $$34I #
b -#3#F #3## -43%F # -#3#" #3## -43F$ # -#3#F #3## -43"H # -#3#" #3## -G3GJ #
c $3#% #3#4 $H3$# # #3H% #3#" F#3IH # $3#" #3#F F"3%# # #3I" #3#F "J3G$ #
144/
a %%3HH F3%% $F3#F # FF3$F "3$$ $G34I # HJ3J$ $J3G4 %3#I #3####G %I3$G %3HJ $#3#G #
b -#3#F #3## -G3FH # -#3#F #3## -G3I% # -#3#" #3#$ -"34% #3##I%I -#3#" #3## -%34I #
c #3J4 #3#H $%3$G # #3H% #3#% "#34J # #3J# #3#% "F34H # #3HJ #3#F "F3I" #
344
a 4J3"G $"3IH G3FI # F#34" $3$H "43"$ # 4H3FI $I3"# F3H# #3###"" G#3JH F3"# $G3J# #
b -#3#$ #3## -F3#G #3##"F" -#3#% #3## -I3#4 # -#3#" #3#$ -"3"H #3#"F$ -#3#" #3## -43H$ #
c #3IH #3#% "#3$J # #3IF #3#G $H3F" # #3JI #3#H $%3#$ # #3HH #3#F "I3GI #
@41
a %"3G# F3%F $"3FI # FJ3H% F3FF $$3JG # 4%3HI H3#F J3"$ # GF3G% F3H4 $%3"4 #
b -#3#% #3#$ -G3"" # -#3#" #3## -%344 # -#3#" #3## -G3$% # -#3#" #3## -G3FJ #
c $3#% #3#I $"34# # #3H# #3#% $I3FI # #3JH #3#% "F3JJ # #3HG #3#F ""3"F #
434
a FI3FF #3I$ %H3%4 # F#3I" $3JI $G3GF # HF3$F $I3"" %3#$ #3####H %#34$ F3FJ $$3JI #
b -#3#G #3## -"#3%4 # -#3#% #3#$ -%3HI # -#3#" #3#$ -"3%" #3#$GH4 -#3#F #3#$ -%3I% #
c $3"% #3#F FH3%J # #3J" #3#J $#344 # #3J" #3#G $43JJ # #3I4 #3#4 $G3FF #
/@1
a %#3#J $3"G F"3$G # F"34G $3G4 "#3JI # %F34G $3%$ F#3IJ # %"3H" "3IF $G3$# #
b -#3#% #3## -$43"" # -#3#F #3#$ -G3JH # -#3#4 #3#$ -$#3HJ # -#3#F #3#$ -G3GJ #
c $3$4 #3#F FF3H4 # #3HJ #3#4 $%3#H # $3"% #3#4 "#34" # #3HJ #3#G $43I$ #
,4@
a FJ3FI $3#G FH34% # FI3#I "3"4 $43I% # GI3"I 43"G J3F" # %"3H" F3$F $F344 #
b -#3#G #3## -"#3$4 # -#3#F #3#$ -G3J4 # -#3#F #3#$ -%3%H #3####$ -#3#F #3#$ -G3"" #
c $3"4 #3#F G#3GI # #3I" #3#G $43J" # #3JJ #3#H $G3$I # #3IG #3#4 $%3FH #
1,4
a %I3"" F3F" $%3G$ # F#3#" $3"" "%3GH # GG3HI 43"G I3JF # F%3J# $3#F FF3I" #
b -#3#% #3## -J3GI # -#3#% #3## -J3#G # -#3#F #3#$ -F3$J #3##$G$ -#3#G #3## -$#3F$ #
c $3"$ #3#F FH3#G # #3I4 #3#% "#34% # $3#F #3$$ J3F# # #3J% #3#% "$3"" #
Stellenbosch University http://scholar.sun.ac.za
F-%G
#.$ tep #: Prediction of potential height from ite Inde4
-t was decided to use S-
"#
, the dominant stand height at the &ase age of "# !ears, as a
reference inde/3 First the relationship of the potential height and site inde/ curves was
plotted and anal!sed after which the corresponding dominant height age "# is used as the
site inde/3
The e2uation used to descri&e this relationship is a Chapman 9ichards height-age function
with site inde/ as a factor .Pa!andeh $JH%0:
pot
= o (SI
20
)
b
(1 c
cugc
)
d(SI
20
)c
Equation /-4
h
pot
is the potential height .m0, SI
,-
the e/pected site inde/, age is measured in !ears, and
a1, a2 and a3 are regression parameters3 This model was chosen due to its parsimonious
structure3
The model incorporates an effect of the site 2ualit! in S%
"#
, which was o&tained from the
predicted site inde/ model of the different sites .Step $03 Thus the potential height
development of the stand is modelled according to their respective S-
"#
values3 This was
done for each site first and finall! for all of the sites com&ined3 -n order to validate the final
model, first the potential height series predicted from dominant height was compared to
the actual potential height in the CCT plots3
#.$.1 7ffect of stand densit%
-n the dataset used, stand densit! affected site inde/3 Figure /-1- clearl! illustrates this
point with linear trendlines applied for each of the four sites used in the stud!, although
some variations around the trend e/isted3 This was possi&l! due to microsite and
management differences of the plots3 This was not significant to the overall performance of
the predictive model, as each site inde/ plot was used as an input for the parameterisation3
-t is included here as a possi&le pitfall of using measured site inde/ that does not ta)e into
account changing the possi&le changing densit! in a following rotation= although the
gradient is 2uite low for the range of commonl! planted stand densities3
Stellenbosch University http://scholar.sun.ac.za
F-%4
Figure /-1- &ite 0nde' %alues on different stand densities for t#e 77T dataset
3
The JGN confidence &ands are e/pectedl! wide due to the low num&er of points3 hat is
noticea&le from Fig F < $# is that each site has a different S--densit! gradient and that the
lines shift upwards or downwards dependent on the inherent site 2ualit!3 This further
emphasises the need for a site 2ualit! predictor &ased on edaphic conditions as proposed &!
.(sler "#$", *ouw and Scholes "##"03
#.$.2 /elationship between potential height and dominant height
-t is important to see what the relationship &etween dominant height and the fitted
potential height on the same site over age is3 As an e/ample, Figure /-11 illustrates this for
the eza dataset for each of the plots3
Stellenbosch University http://scholar.sun.ac.za
F-%H
Figure /-11 E'ample of potential and dominant #eig#t-age cur%es for t#e <e$a trial
This shows how the difference &etween potential and dominant height increases over age3
-n order to assess this relationship potential height is plotted over dominant height for all of
the four CCT trials in Figure /-1, &elow3
Stellenbosch University http://scholar.sun.ac.za
F-%I
Figure /-1, )otential #eig#t plotted o%er dominant #eig#t for all of t#e t#e 77T dataset plots
Figure /-1, shows a generall! linear relationship for the dominant and potential heights
over age, although the values &egin to var! as age, and conse2uentl! dominant height,
increase3 This gradient was modelled &! a least s2uares linear fit, forcing the intercept
through zero, which suggests a&out a $#N difference in potential and dominant height
values over age .Table /-60, i3e3 the difference &etween potential and dominant height
increases over age, as e/pected3
Table /-6 )otential #eig#t - dominant #eig#t gradient
&ite Eradient &td. Error t %alue )r(CDtD. ;
,
Du)udu)u $3#J%4"F #3##$4FG 44J3F T"e-$4 #3JJJ$
:wam&onam&i $3$"%JJF #3##"%44 %G43" T"e-$4 #3JJI$
eza $3#I#I4I #3##$J$J G4F3$ T"e-$4 #3JJIH
8ac 8ac $3$"%JJF #3##"%44 %G43" T"e-$4 #3JJI$
This linear relationship and the relativel! similar gradients suggest that a reference S-
"#
could &e introduced as an added effect, as specified in Equation /-43
Stellenbosch University http://scholar.sun.ac.za
F-%J
#.$.# Predictive e6uation
Site inde/ will change over two gradients: the site and, to a smaller degree, the densit! .a
potential source of error in the model03 The o&7ective here is to o&tain potential height
models for a given site inde/ .in this case, S-
"#
03 The potential height over site inde/
relationship is defined for all sites com&ined3 -t must &e noted that the effect of densit! on
site inde/, which was calculated on the South African top height definition, is not ideal, and
a great improvement could &e made &! &asing the site inde/ on site factors mentioned in
the previous section3
"inal Combined *odel
The four CCT spacing trials were pooled into the final model for potential height-age
prediction from site inde/ using Equation /-43 Pooling all of the sites together will prevent
the occurrence of e/treme cases and provide a &alanced model that is a&le to fit most cases
sufficientl!3 Figure /-1/ represents the final proposed model for prediction of potential
height from site inde/3
Figure /-1/ Final parameterised model of potential #eig#t-age using &0 as a predictor
Stellenbosch University http://scholar.sun.ac.za
F-G#
Table /-@ Final model parameterised on t#e pooled 77T trial datasets from Equation /-4
)arameter 2alue &td. Error t %alue )r(CDtD.
a #3IFFI #3#%"FH $J34HH"H #
& $3F$F#4 #3#$JI" 443"4%H4 #
c -#3#"J4G #3##$#$ -"J3FH"J #
d #3$44$F #3##J%I $H3G"#%4 #
e #3GGG4F #3#$HJG F#3J%I44 #
#.$.$ 'alidation
,bserved vs. Predicted Potential height
-t was decided to see how the potential height compared with the potential height
predicted from the site inde/ .in the pooled model03 This was done &! calculating the
difference in predicted potential height compared to the o&served potential height fitted for
each of the spacing trial plots .Fig /-4, Table /-503The models fits tested on the individual
sites a relativel! high error margin .over and under prediction0 &efore age $# .Fig /-15.3
The! &egin to converge age "# and then fan out again representing an increased error
predicted for higher ages .Fig /-140, although the values are small considering the actual
height of the trees at those ages3
Stellenbosch University http://scholar.sun.ac.za
F-G$
Figure /-14 De%iation from t#e obser%ed potential #eig#t-age compared to t#e potential #eig#t predicted equation
Stellenbosch University http://scholar.sun.ac.za
F-G"
Figure /-15 :edian de%iation in percentage of t#e obser%ed and predicted potential #eig#t cur%es
;verall, considering the vastl! differing site conditions, growing conditions and the different
shape of the height-age growth curves, the com&ined model reacts well, and although
seems to underpredict slightl! at after age "# .Fig /-150, it seems to sta&ilise to some
degree3 -t seems that site inde/ can indeed &e used as a predictor for potential height
modelling, although the values will range at ver! !oung and ver! old ages3 The high error
&efore age $# could also mean that this Chapman 9ichards &ased model ma! &e unsuita&le
for pulpwood rotations, although as these are onl! potential height values it would first
need to &e determined how the final predicted model performs once a modifier is applied3
/eferencing against independent PP data
The a&ove validation used sites in which the model was parameterised= however it is
necessar! to see how this model wor)s for independent data3 For this reason a few sites
were selected from the PSP dataset which had a suita&le num&er of re-measurements3 From
this an inspection of how the o&served vs3 the predicted potentials compare for sites of a
given site inde/ can &e seen .Fig /-,103
Stellenbosch University http://scholar.sun.ac.za
F-GF
Figure /-16 )otential #eig#t cur%es plotted o%er )&) data of different classes. T#e red lines represent t#e predicted
potential #eig#t for t#e upper bound of t#e &0 classes presented abo%e. T#e top left image presents all of t#e data wit#
15+ ,-+ ,5 and /- &0 predicted potentials
-n Figure F-"$, the predicted potential from site inde/ seems to capture the upper &ound of
the PSP dataset, with the top left hand graph representing all of the sites, with lines of the
predicted potentials for S-Ls of $G, "#, "G and F#3 (ach of the other graphs represents S-
classes within this dataset with the respective potential height curves predicted from S- of
"#, "G, and F# respectivel!3
To see how this performs on a few selected sites, five sites were selected from the second
PSP series, where measured S-Ls were not availa&le, and thus had to &e calculated from the
methodolog! proposed in this chapter3 The S- values would then predict a potential curve
which would &e superimposed on the plot height-age data for visual inspection .Fig /-1@03
Stellenbosch University http://scholar.sun.ac.za
F-G%
Figure /-1@ Fitted predicted potential #eig#t cur%es fitted on selected independent )&) data
The predicted potential seems to capture the potential heights of each site 2uite well, with
onl! a few trees which fall a&ove the predicted potential curves3
Stellenbosch University http://scholar.sun.ac.za
F-GG
#.& Chapter conclusion
The main o&7ective for this chapter, split onto three steps, was to model potential height
&ased on measured site inde/3
Site inde/ models and potential height models were developed in order to develop a
methodolog! to predict potential height from site inde/3 The relationship &etween site
inde/ and potential height was shown to &e a linear relationship, and a reference site inde/
of S-
"#
was used to predict potential height from site inde/3
The use of nonlinear mi/ed effects modelling has proved a superior method for modelling
dominant height-age curves with the associated clustered data structures and
heteroscedasticit! and it also represents a potential improvement for other height-age
models3 However, care must &e ta)en to ensure that including the as!mptote as a random
effect does not inflate or deflate the as!mptote implausi&l! < as it tends to change the
as!mptote to fit the o&served data, which ma! affect e/trapolation &e!ond the measured
!ears in the parameterisation training set3
>onlinear 2uantile regression proved to &e a ro&ust method of fitting o&served height
potentials3 The model training set .CCT trials0 covered the range of heights o&served PSP
data3
Potential height seems to &e well correlated to a standLs site inde/3 Eisual validation of the
results seem to suggest that site inde/ can indeed &e used to model the potential height,
which covers the first step of the potential modifier method and simulation initialisation &!
deriving a height potential from an industr! standard site inde/ information3
Stellenbosch University http://scholar.sun.ac.za
%-G4
Chapter $: *odelling diameter increment in response
to resource limitations and site classification
$.1 Introduction
The main purpose of this chapter is to create an age-independent diameter increment
model, which predicts diameter increment from D+H and the competitive conditions of the
tree at an! point in time3 Figure 4-1 illustrates this relationship for all of the plots in the 8ac
8ac spacing trial, where the gradient for the increment-diameter relation flattens out
.decreases0 over time, with each scatter cloud representing a different measurement
activit! at a certain age3 -n the proposed approach, which is following the S-*EA
methodolog!, the age effect is su&stituted &! tree size .D+H0, while the competition
interaction is supposed to cater for the change in increment-diameter relations over age3
Figure 4-1 D*! 0ncrement - D*! scatterplot on t#e :ac :ac 77T trial+ s#owing decreasing linear gradient o%er age
(mplo!ing a potential-modifier methodolog!, two steps were thus re2uired3 First the
potential increment for an! given tree diameter had to &e found3 Second, the modifier,
&ased on a competition inde/, had to reduce the increment and displa! a structure similar
Stellenbosch University http://scholar.sun.ac.za
%-GH
to Figure 4-1 a&ove, where&! the gradient of diameter increment for a given D+H
decreases3
Different competition indices are developed and descri&e growth differentl! according to
their mathematical structure3 For instance the ::* inde/ .Pretzsch et al. "##"0, with its
search crown is intrinsicall! more suited to descri&ing overtopping and thus competition in
light limited environments3 ;ther indices, &ased solel! on distance &etween individuals or
groups of individuals in an area, will &e intrinsicall! more size s!mmetric, such as the *ocal
+asal Area .*+A0 inde/, descri&ed later .Seifert et al3 in press03
Thus the sensitivit! of diameter increment to an inde/ can change according to site location
and the resource limitation which is e/perienced as demonstrated &! Seifert et al. .in press03
For instance, competition inde/ A might descri&e more of the variation in the predicted
model than inde/ + in a wet environment3 This can also change according to the 2ualit! of
the site3 The performance of competition indices ma! also change according to age, densit!
and longitudinal climate changes .wet and dr! spells03
$.1.1 .ataset
The >elder spacing design and the SSS-CCT design at Tweefontein were used for these
purposes3 -t must &e noted upfront that these are two ver! different designs < thus when
ma)ing comparisons &etween the &ehaviour of the indices, some differences could &e due
to the different spacing design, o&servation, etc3 However, these were the onl! trials where
tree positions could &e o&tained and most of these pro&lems should &e negated &! the
nature of the competition indices3
$.1.2 Chapter outline
ith onl! two sites from which to test competition with different water availa&ilit!, this
Chapter should &e strictl! seen as a methodological guideline for predicting diameter
increment, instead of an investigation on changing competition mode with different
resource limitations, although an indication of shift in importance of competition indices on
different sites is seen .%3G3F03 The chapter presents possi&le methodologies for future testing
of changing competition modes and their effect on diameter -ncrement3 The methodolog!
was developed according to the following steps:
Step $: Classif! sites according to water availa&ilit!
Step ": Determine the potential &ased on site conditions
Step F: Fit multiple competition indices
Step %: Select competition indices
Step G: 1se C-Ls in a deterministic potential modifier e2uation
Step 4: Create a stochastic model incorporating natural varia&ilit!
Stellenbosch University http://scholar.sun.ac.za
%-GI
Figure 4-, Flowc#art of t#e c#apter outline s#owing working steps of t#e met#odological approac# used in t#e stud"
Stellenbosch University http://scholar.sun.ac.za
%-GJ
$.2 tep 1: ite classification according water availabilit% and
site inde4
ith the h!pothesis in mind that the mode of competition shifts with the availa&ilit! of
edaphic factors, a simple categorisation of the sites according to water availa&ilit! was
sought, su&7ect to the availa&ilit! of data3 The FA; < 1>(P inde/ .1>(P $JJ"0 where aridit!
is calculated &! dividing the precipitation &! the potential evapotranspiration, as calculated
&! the Thornthwaite method .Thornthwaite $J%I03
-n this instance, the aridit! inde/ was fi/ed &! using monthl! weather data supplied &! the
Agricultural 9esearch Council .A9C0 climate data&ase3 This provides an added static
classification for the site3
Defining the water inde/ longitudinall! .changing over time0, either using this inde/ or other
candidate indices .8c:ee, Does)en, and :leist $JJG, Palmer $J4G0, would improve the
interaction with diameter increment3 1nfortunatel! measurements in the trials were not
ta)en at annual intervals= often at "-G !ears, which would mas) correlations3
Table 4-1 7lassification of t#e sites according to t#e FAO-89E) classification and &0. T#e Tweefontein and :ac :ac trials
used t#e same weat#er station.
1ocation )recipitation Aridit" inde' 7lassification (FAO-
89E).
A%erage &0
*ottering >elder JG# $3$#4 Humid "$3"
UTweefontein SSS-CCT $""" #3I4 Humid "F3#
:wam&onam&i $"#I $3" Humid "#3F
Du)udu)u J4# #3H4 Humid $H
eza J"H #34J Humid "#3F
U8ac 8ac $""" #3I4 Humid "%3#
Table 4-1 shows this inde/ for all of the spacing trials used in this thesis3 For the aridit!
inde/, onl! two sites were relevant, the >elder trial and the Tweefontein SSS-CCT trial, as
these were the onl! availa&le trials with tree positions necessar! for modelling the effects of
competition in this case3 However the average site inde/ .averaged over the S- values
presented in Chapter F0 was used for the potential increment estimation in Section %3F and
for the testing of the model in Sections %34 and %3H3
ith this is mind, a tentative linear model was drawn &etween the water inde/ values of the
two sites < accepting that more wor) would need to &e done to full! incorporate water
availa&ilit!3 This simple methodolog! would then &e developed and tested for future
e/periments, using this simple water inde/ as a pointer to determine whether there are an!
Stellenbosch University http://scholar.sun.ac.za
%-4#
functional relationships &etween the water availa&ilit! and the competition indices in this
model3
$.# tep 2: .etermine the potential based on site conditions
For the purpose of this stud!, the methodolog! proposed &! Pretzsch and +i&er ."#$#0 was
used to determine a wor)a&le simulation routine for diameter increment modelling3 -n this
procedure the potential diameter growthVincrement was first defined using the nlrq
pac)age in 9 for nonlinear 2uantile regression application .as done in Chapter F0, using the
#3JJ tau values, where:
Equation 4-1
here di
pot
is the predicted potential increment, d is the diameter .D+H0 of the tree and a
0
,
a
1
and a
2
are coefficient to &e determined in the model3 This represents the ma/imum
potential diameter increment for a given D+H < while not a true potential, it is an ade2uate
appro/imation for the purpose of this thesis3
$.#.1 Comparison of sites
Four sites were considered for potential increment parameterisation, the 8ac 8ac,
:wam&onam&i and Du)udu)u CCT trials and the >elder spacing trial3 The eza trial was not
included as there was a large period of measurements gaps= the Tweefontein SSS-CCT trial
was not included as it was too !oung3 The potential curves fitted for the considered sites are
presented in Figure 4-/ &elow3
Stellenbosch University http://scholar.sun.ac.za
%-4$
:ac :ac Awambonambi
Dukuduku 9elder
Figure 4-/ Fitted potential increment cur%es for t#e four sites considered for increment potential estimation
-ncluding the >elder trial to the model results in a much flatter curve which gives poor fit to
the data upon visual inspection3 This was &ecause the >elder trial was onl! measured until
"4 !ears, which did not include the flattening of the -ncrementVD+H relationship, resulting
in an interference with the gradient3 Thus the results of the potential increment curves of
the three CCT trials according to Equation 4-1 are represented &elow3
Stellenbosch University http://scholar.sun.ac.za
%-4"
Table 4-, )otential increment model coefficients according to Equation 4-1.
S- $H3#
S- "#3F
:wam&onam&i
S- "F3J
Du)udu)u 8ac 8ac
Parameter (stimate S( (stimate S( (stimate S(
a #3#GF" #3#$#F #3I##" #3""%$ F3JH$% #3GIGI
& "3"FIF #3#JHF #3HJ4H #3$FFH #3""I" #3#HJJ
c #3$"I #3##%" #3#GFJ #3##G" #3#F$4 #3##FJ
From Figure 4-/ and Table 4-, it can &e seen that site inde/ had an effect on the
determination of potential increment .as seen in the a&ove figures03 Thus dominant height <
or #
dom
- can &e included as a predictor varia&le to Equation 4-1, where a site 2ualit! effect
can &e included .Pretzsch and +i&er "#$#0 as was done in in Chapter F, resulting in Equation
4-, &elow
Equation 4-,
The resulting fit of the a&ove model is ta&ulated in Table 4-/ and the curves are illustrated
in Figure 4-43
Table 4-/ 7oefficients of t#e potential increment model (Equation ,. using site inde' and db# as predictor %ariables
7ombined 0ncrementF D*! model (Equation 4-,.
2alue &td. Error t %alue )r(CDtD.
a
1
#3GFH" #3"HFHH $3J4""$ #3#%JHG
a
,
#3$$%F% #3#$G4% H3F#II4 #
a
/
#3F%%"$ #3#I4"I F3JIJF #3####H
a
4
#3#FH"% #3##%$F J3#$"HH #
Stellenbosch University http://scholar.sun.ac.za
%-4F
Figure 4-4 )otential increment o%er D*! cur%es parameterised from t#e 77T trial dataG t#e different cur%es represent
different &0 %alues
These potentials are the JJN 2uantiles representing the ma/imum annual diameter
increment that a tree can o&tain for a given D+H and site inde/3 The estimated increment
was then modelled &! including a suita&le competition inde/, discussed in the following
sections, which should mimic the structure presented in Figure 4-13
$.$ tep #: "it various competition models
$.$.1 .istance dependent competition indices
-n this stud! five distance-dependent competition indices were used3 These indices were
chosen to capture a gradient, which would descri&e overtopping or local crowding, with the
o&7ective that two indices will &e selected which &oth ade2uatel! descri&e competition and
capture different modes of competition .Seifert et al. in review03 These, summarised in
Table 4-4 and descri&ed &elow, are the ::* .which is used in S-*EA0, the local &asal area
.*+A0, &asal area of larger trees .+A*0, Heg!i and Eoronoi pol!gons3
Stellenbosch University http://scholar.sun.ac.za
%-4%
Table 4-4 0llustration of t#e different competition indices used in t#is stud" (&eifert et al. in re%iew.+ i refers to t#e
central tree+ j refers to t#e competitor trees. :odels s#own below are discussed in more detail in te't below.
The ::* inde/ .Pretzsch "##J0 uses a search for competitors &ased on a search cone at an
angle of 4#W which starts at a point of 4#N of the tree height3 As most competition indices it
wor)s in a com&ination of competitor selection and the 2uantification of the competition
effect of those identified competitors3 >eigh&ouring trees, which fall inside the search cone,
are included as competitors and the angle from the tip of the neigh&ouring tree and its
ma/imum crown e/tension to the cone mantle &angle ' in Table 4-4( is calculated3 This is
used as a measure of competition, which is then multiplied &! the cross-sectional crown
area of the competitor in relation to the su&7ect tree .CCA
i
VCCA
7
03 Species specific values for
crown dimensions for P. elliottii were not availa&le, for this reason the S-*EA model was
used to calculate the ::* for the stands using Pinus sylvestris crown model as a pro/!3 The
::* is &! its design strongl! focussing on competition induced &! crown competition and
overtopping3
Stellenbosch University http://scholar.sun.ac.za
%-4G
The local &asal area .*+A0 is simpl! the sum of the &asal area of all trees within a
competition search radius and the reference tree located in the centre of the circle
.Stene)er and Qarvis $J4F03 -t has &een suggested as an effective measure of local crowding
and is more sensitive to measuring edaphic limitations as it does not discriminate &etween
the sizes of the trees included in the influence zone, e/cept for their &asal area contri&ution
.Seifert et al. in review03
The &asal area of larger trees .+A*0 is identical to the *+A e/cept that it onl! includes trees
which are larger in D+H than the reference tree, which would then provide a good
indication of overtopping and radiation e/clusion3
The Heg!i -nde/ .Heg!i $JH%0 calculates size ratios of the reference tree with its competitors
multiplied &! an inverse distance weighting3 These are then all summed up for each
reference tree within a zone of competition3
The Eoronoi inde/ is a simple growing space inde/ where the space in a given stand is
divided e2uall! among trees, first &! drawing distance lines &etween each tree and its
neigh&ours, these lines are then &isected and lines are drawn connecting the &isected lines,
resulting in a pol!gon for each tree &ased on its growing area with regards to its neigh&our
.Figure 4-503 This is thus simpl! a refined measure of a treeLs growing area and the
relationships would change with age and competition3 -t is included here as a &enchmar) <
to illustrate the performance of the other indices in relation to a simplistic growing area
inde/3 -t must &e noted that refinements of the Eoronoi e/ist, where the size of the
pol!gons is changed proportional to the size ratio of the tree compared to its neigh&ours or
even to measured crown dimensions .Seifert and 1tschig "##"03 This was not e/amined in
this stud!, &ut would &e a potential point of interest for future comparison3
Stellenbosch University http://scholar.sun.ac.za
%-44
Figure 4-5 2oronoi pol"gons calculated for t#e 9elder spacing trial as an e'ample
$.$.2 7dge effects
-n the Tweefontein SSS-CCT trial, although &order trees were present, the! were not
measured3 ith the small size of each stand, "G trees in a stand, man! of the indices will
have search zones which fall outside of the &oundar!3 This is noted as a possi&le limitation
of the stud!3
Pretzsch ."##J0 mentions three methods to simulate or e/trapolate stand structure outside
of the edge of the measured stand: plot mirroring, shifting and linear e/pansion3 -n this
stud!, the plot shifting techni2ue was used for the SSS-CCT trial, where the plot is shifted
along the sides and corners creating eight edge plots, with the measured plot in the centre3
This was not necessar! for the >elder trial, the first and last rings were simpl! repeated
along the gradient for positions3
$.$.# Competition search radius 0influence <one1
The definition of the size of the search radius is critical to the performance of the
competition indices < if the zone is too &ig or too small, information on competition will &e
Stellenbosch University http://scholar.sun.ac.za
%-4H
lost3 The first decision to &e made is whether the search radius should &e fi/ed or relative
.to stems per hectare for instance03 -n this stud! a fi/ed radius for the different stand
densities was chosen, however on which would change relative to the dominant height of a
stand at a certain age, the h!pothesis &eing that the size of the influence zone would change
as trees &ecome larger in a stand3 The Heg!i, *+A and the +A* indices use fi/ed search radii
and it was decided to standardise the radius for all three indices in order to draw reasona&le
comparisons3
The *+A was used as a reference to determine how large the search radius should &e as it is
the simplest of the three indices and is not complicated &! additional factors such as size of
competitors, etc3 Furthermore, the *+A inde/ has a much clearl! linear correlation with
diameter increment .Figure 4-40, with a much higher lin) to linear correlation than the other
indices, where more sophisticated nonlinear or transformation applications would have
&een necessar! to determine the ma/imum correlation with the search radius3
To determine the optimum search radius, different sizes of search radii were tested on
ever! measurement age of the >elder trial and the correlation of the *+A with diameter
increment determined .the minimum of each correlation was chosen since the size of the
competition has a negative effect on diameter increment03 This is represented in Figure 4-6
and summarised in Table 4-5 &elow3
Figure 4-6 9egati%e 7orrelation of t#e 1*A competition inde' wit# diameter increment at different measurement "ears
(ages. using different competition searc# radii+ t#e correlation was multiplied b" -1 for illustrati%e purposes.
Stellenbosch University http://scholar.sun.ac.za
%-4I
Table 4-5 7orrelation of t#e 1*A competition inde' at different ages of t#e 9elder trial wit# t#e resulting searc# radius
si$e included
Age
Dominant
#eig#t
:inimum correlation
Optimum searc#
radius
%3GI %3F$%HFH%I% -#3F4$I4 "34""$HF
4 G3J"J%HGH4G -#344HGF G3I4$G%J
$$3#I $$3HH4F%H#4 -#3IG"I %3F$"F#I
$F3FF $%3"IIH$F"G -#3HJ$H$ H3JH#44$
$G3GI $43H$%IJ""H -#3IF##F H3JH#44$
$J3$H "#3FHH$4#JJ -#3I"4F" H3JH#44$
"#3"G "$3%"GF#"#4 -#3GHHG" H3JH#44$
"F3IF "%3H$H%%"$ -#3I$JG$ H3JH#44$
"4 "43GH4J#$4F -#3I%HFJ $#3$$%%%
A simple linear model was then fitted for the values from Ages %-"4 < with their respective
dominant heights in order to e/clude age from the calculations illustrated in Figure 4-@, with
the resulting model in Table 4-63
Stellenbosch University http://scholar.sun.ac.za
%-4J
Figure 4-@ 1inear regression of t#e optimum searc# radius o%er dominant #eig#t
Table 4-6 1inear regression coefficients of t#e optimum searc# radius based on dominant #eig#t
1inear (O1&. competition $one model
)arameter
Estimate &td. Error t %alue )r(CDtD.
(stimate Std3 (rror t value Pr.XYtY0
(0ntercept. "3J#IJG $3##HI$ "3II4 #3#"F%% U
!dom #3"G#F4 #3#G4%4 %3%F% #3##F#F UU
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
9esidual standard error: $3"G4 on H degrees of freedom
8ultiple 9-s2uared: #3HFHG, Ad7usted 9-s2uared: #3H
Thus the competition search zone will increase with the given dominant height, which was
used for all three competition indices that re2uired a search radius .*+A, +A* and Heg!i03
$.$.$ Performance of competition indices
(ach of the competition indices introduced in Section %3%3$ was applied to the *ottering and
Tweefontein spacing trials resulting in the following diameter increment in relation to the
competition inde/ plots &elow .Figure 4-403
Stellenbosch University http://scholar.sun.ac.za
%-H#
1ottering 9elder Tweefontein &&&-&77T
::*
*+A
+A*
Heg!i
Figure 4-4 ;elations#ip of diameter increment and t#e competition indices fitted in t#e two spacing trials used for t#is
stud"
Stellenbosch University http://scholar.sun.ac.za
%-H$
The Heg!i and ::* indices &oth had a clear negative e/ponential trend of increment over the
C- value, the +A* inde/ also showed a nonlinear trend and the *+A inde/ shows a linear
response3 Significant variation e/ists in the plots, especiall! when competition indices
approach zero .state of no competition03 The Eoronoi inde/ did not show much correlation
.Appendi/ +0= however this could &e &ecause the correlation changes at different ages or
stages of development of the stand, which this investigation is tr!ing to avoid3
$.& tep $: electing competition indices
As seen in the previous section, each competition inde/, &! virtue of the different aspects
that the! measure, covers different e/planator! aspects of competition and thus would
potentiall! contri&ute e/planator! value3 Thus, com&ining two or more C-Ls to the modifier
e2uation &! creating a multiple regression e2uation, more aspects of competition might &e
covered and a &etter fit ma! &e o&tained with respect to diameter increment3
However, overlap of the e/planation &etween competition indices or collinearit! ma! distort
the e/planation value of each inde/ used3 Collinearit! .or multicollinearit!, when more than
two varia&les are included in a model0 occurs when predictor varia&les in a multiple
regression are highl! correlated .8!ers $JI403 Collinearit! does not alwa!s affect the
predictive power of a model on the data that it is fit3 However, collinearit! significantl!
affects the sta&ilit! of the model and hinders the a&ilit! to draw valid conclusions a&out
individual regressor varia&les= small changes in !-values can significantl! alter model
coefficients even though measures of fit .e3g3 9
"
values0 remain relativel! unchanged .8!ers
$JI403 This would compromise one of the o&7ectives of this chapter < to determine how the
importance of different varia&les changes under different edaphic and site 2ualit!
conditions3 Additionall!, although it is tempting to add regressor varia&les .the C-Ls0 to
o&tain a &etter fit, the concept of ;c)hamLs razor should &e followed, where parsimonious
models are sought3 Furthermore, tedious calculation of ever! competition inde/ would not
&e ideal for an! future model application3
-n view of the o&7ectives of this chapter, the set of chosen competition indices was thus
tested individuall! as well as in com&inations to identif! inde/ com&inations that were not
significantl! collinear3 -deall!, a com&ination of two C-Ls was to &e o&tained that ade2uatel!
descri&e diameter increment3 The resulting model should provide a good e/planator! power
with regards to overtopping .light limitations0 and local crowding .edaphic limitation0 < as in
the e/ample of Figure 1-,3
This re2uired a varia&le selection e/ercise, where correlation .collinearit!0 and &ehaviour of
the respective indices was sought through principle component anal!sis .PCA0 and variance
inflation factors .E-FLs03 Different com&inations of C-Ls varia&les were tested se2uentiall! to
determine the &est com&ination of C-Ls3 9elative importance measures were then carried
out with the two final selected C-Ls on different sites3
Stellenbosch University http://scholar.sun.ac.za
%-H"
Figures 4-3 and 4-1- show the pairwise matri/ scatterplots of the different competition
indices against each other in the >elder and SSS-CCT trials, indicating correlation &etween
certain varia&les3
Figure 4-3 )airwise scatterplot matri' of t#e %arious competition indices against eac# ot#er in t#e 9elder Trial
Stellenbosch University http://scholar.sun.ac.za
%-HF
Figure 4-1- )airwise scatterplot matri' of t#e %arious competition indices against eac# ot#er in t#e Tweefontein Trial
$.&.1 'ariable selection
As seen in Figure 4-4 most of the competition indices were nonlinear in their effect on id3 A
simple s2uare root transformation was found to &e ade2uate to linearise the response=
where f(x) represents a linear com&ination of one, several, or all of the competition indices3
S2uare root transformations have the advantage over logarithmic transformations that, for
diameter increment where # values are prevalent, no infinite values are produced3 The
transformation produced visuall! linear fits .ordinar! least s2uares0 for all of the models,
e/cept for the ::* and Eoronoi3
Stellenbosch University http://scholar.sun.ac.za
%-H%
Table 4-@ 1inear models coefficients of t#e square root transformation of diameter increment using t#e respecti%e
competition indices
:odel ;
,
A07
s2rt.D&h-ncrement0 Z +A*
#34##" "#II3%#J
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 $3%#%44I #3##%G#F F$" T"e-$4 UUU
+A* -#3#$I4% #3###$HH -$#G3$ T"e-$4 UUU
s2rt.D&h-ncrement0 Z *+A
#3GIF% "FJ"3$"F
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 $3%J4$$4 #3##G"J4 "I"3G T"e-$4 UUU
*+A -#3#$%%% #3###$%" -$#$3G T"e-$4 UUU
s2rt.D&h-ncrement0 Z HgC-
#3%4#4 %"J$3H#H
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 $3FG%FI #3##G$%4 "4F3$H T"e-$4 UUU
HgC- -#3#H##G #3###II% -HJ3"% T"e-$4 UUU
s2rt.D&h-ncrement0 Z ))l
#3%G"% %%#F3F""
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 $3"%#H%G #3##%F%G "IG3GJ T"e-$4 UUU
))l -#3"GH#H #3##F"JJ -HH3JF T"e-$4 UUU
lm.formula M s2rt.D&h-ncrement0 Z Eoronoi, data M sss0
#3$JJG H$JG3#"%
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 #3IJF"HH #3##4"J4 $%$3II T"e-$4 UUU
Eoronoi #3##I%IG #3###$JI %"3I$ T"e-$4 UUU
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
Table 4-@ shows that the *+A and +A* indices e/plain the variance well, .showing the
highest 9
"
values0, although it must &e noted that the ::* could &e &etter descri&ed with an
alternative transformation3
Com&ining all of the indices into a full model to descri&e diameter increment according to
Equation 4-/ shows how a much improved fit can &e o&tained &! adding varia&les together,
indicated &! their respective names in the e2uation, with parameters a-e to &e determined
.Table 4-403
_BE
I
=o kkl +b Ecgyi +c IBA+J Ioronoi +c
Equation 4-/
However, due to interactions and collinearit! &etween varia&les, a decision has to &e made
of which varia&les should &e included in the model that represents a good compromise
&etween model fit and reduced correlation &etween the varia&les3 Eariance inflation is one
method used to detect collinearit! .8!ers $JI40:
Stellenbosch University http://scholar.sun.ac.za
%-HG
Equation 4-4
where the VI for a varia&le is the reciprocal of the inverse of 9
"
from the regression3
Eariance inflation is calculated for each of the e/planator! varia&les3 As a guideline a E-F of
G, or a s2uare root of the E-F larger than two, indicate collinearit! in the model3
Table 4-4 Full linear model of t#e square root diameter increment transformation using all of t#e 70Hs as regressor
%ariables
8odel: s2rt.Diameter -ncrement0 M *+A [ +A* [ HgC- [ ::* [ Eoronoi
Parameter (stimate Std3 (rror t value Pr.XYtY0 E-F s2rt.E-F0
.-ntercept0 $3G#%4JI" #3##IJHH4 $4H34#4 T "e-$4 UUU
*+A -#3##H"FIJ #3###"F%4 -F#3I4F T "e-$4 UUU F3%J444G $3I4JJFH
+A* -#3##4G"$$ #3###F%HI -$I3H%J T "e-$4 UUU %3HG#$II "3$HJ%JF
HgC- -#3#$4G%H4 #3##$%IGH -$$3$FI T "e-$4 UUU %3H$#"J$ "3$H#F"$
))l -#3#%F%$#H #3##G"J%$ -I3" "3I"(-$4 UUU %3F4$$J$ "3#IIF%4
Eoronoi -#3###%GIH #3###$HJF -"3GGI #3#$#4 U "3#"4%I" $3%"FG%4
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
8ultiple 9-s2uared: #34H4H, Ad7usted 9-s2uared: #34H4G
F-statistic: F#HH on G and HF%J DF, p-value: T "3"e-$4
Table 4-4 shows a significant fit for all of the parameters, however E-F is 2uite high for most
of the varia&les indicating that the model needed to &e reduced and dependencies &etween
varia&les &etter understood3
$.&.2 Principle component anal%sis
Principal component anal!sis .PCA0, .Hotelling $JFF0, can &e used to reveal similarities in
e/planation value and thus collinear structures &etween varia&les3 -t is a multivariate
statistical feature e/traction method that wor)s &! pro7ecting the data onto orthogonal
vectors, called principal components, which e/plain the ma/imum variation of the data3
These principal components are the features of the data3
Detailed information of PCA can &e found in multiple te/ts3 +riefl! e/plained though: -f I is a
matri/ with the columns representing varia&les and the rows representing sample
o&servations, the first principal component p
$
is the linear com&ination where the
pro7ection onto p
$
, given &! + !ields the ma/imum variance su&7ect to 3
The pro7ection onto the second principal component p
"
is the com&ination that
gives the second highest variance su&7ect to and p
"
&eing orthogonal to p
$
3 This is
e/tended to all further principal components3 The orthogonalit! criterion ensures that each
component is completel! uncorrelated with all other components, there&! ensuring that the
components e/plain the ma/imum variation3 The principal component loading vectors, pi .i
Stellenbosch University http://scholar.sun.ac.za
%-H4
M $,\, m0, are the eigenvectors of the covariance matri/ of the data set, therefore the! can
&e o&tained from spectral decomposition of the covariance matri/3
;nce the principal components have &een determined a &iplot can &e used to visualise
which varia&les contri&ute most to each principal component, it also shows similarities or
collinearit! &ehaviour of the varia&les in relation to each other in the transformed space3
hen a &iplot is applied to PCA results, the a/es are a pair of principal components3 The
points represent the principal component scores of the o&servation, and the vectors
represent the coefficients of each varia&le on the principal components3 So in the &iplot the
longest vector will represent the varia&le that contri&utes most to the variation in the data
and therefore provides the most important information for the model3
9 has a function called princomp in its stats pac)age for PCA3 A data matri/ consisting of the
competition indices is passed into it and it performs PCA on it3
Table 4-3 0mportance measures of t#e principle components
0mportance of components
:easure 7omp.1 7omp., 7omp./ 7omp.4 7omp.5
&tandard de%iation $34%4#"$ $3#F%#$" #3IHJ%4F #3G4#H%I #3F4G%"J
)roportion of 2ariance #3G%$IHH #3"$FIF4 #3$G%4J$ #3#4"III #3#"4H#I
7umulati%e )roportion #3G%$IHH #3HGGH$F #3J$#%#G #3JHF"J" $
Table 4-3 a&ove shows that Component $ accounts for G%N of the variation in the data: The
loadings of each varia&le .C-0 of the principle components a&ove can &e shown numericall!
in Table 4-1-3
Table 4-1- )rinciple component loadings of t#e respecti%e principle components
)rinciple 7omponent 1oadings
0nde' 7omp.1 7omp., 7omp./ 7omp.4 7omp.5
AA1 -#3%GG #3%"% -#3%"$ -#3"IG -#3GJ4
*A&A1 -#3%HJ #3"I4 #3F#F -#3%J4 #3GJ"
1*A -#3%GG #34IH #3%%H -#3F%I
!g70 -#3%H4 -#3G#F #34 #3FJI
2or #3F4$ #3IGI #3FF4 #3$""
Eisuall! this is shown in the &iplots for &oth sites .Figures 4-11 and 4-1,0, where the length
of the arrow for each varia&le indicates its contri&ution to the component loading3
Stellenbosch University http://scholar.sun.ac.za
%-HH
Figure 4-11 )7A biplot of t#e %ariable plotted in t#e transformed space of t#e first two principle components in t#e
9elder trial
Figure 4-1, )7A biplot of t#e %ariable plotted in t#e transformed space of t#e first two principle components in t#e
Tweefontein trial
Stellenbosch University http://scholar.sun.ac.za
%-HI
>ot much can &e determined from the length of the vectors .or their loadings0 as the! load
relativel! e2uall! on the two main components3 However, Figures 4-11 and 4-1, show
consistent performance on the two sites with regards to &ehaviour .grouping0 of the C-
varia&les3 The Eoronoi inde/ is negativel! correlated to the other potential predictors < a
high Eoronoi area will mean that there is more growing space and thus a higher increment,
as opposed to the other varia&les which increase in values with increasing competition3
However, what is interesting is the correlation &etween the other four varia&les, which will
&e necessar! for stratification of the varia&les to ma)e deductions a&out their e/pected
similarit!3 The ::* and the +A* are ver! similar, as the! characterise competition strongl! &!
overtopping3 The Heg!i and *+A, which might e/plain edaphic competition .local crowding0
to a higher degree, are closel! grouped compared to the other indices3
As e/plained earlier, collinearit! will &e a pro&lem for creating a com&ined model3 The PCA
loadings a&ove give a 2ualitative idea of which varia&les to include3 Clearl!, the two groups
can &e seen, with the ::* and +A* closel! related, and the Heg!i and *+A closel! related as
well3 -t was decided to include one varia&le from each grouping .e3g3 the ::* and Heg!i0 for
multiple collinearit! pro&lems and for parsimon! in the model3 The Eoronoi inde/ was not
considered for further modelling purposes due to its low e/planation value .Table 4-@ and
Table 4-603 Table 4-11 &elow shows the results for the various pairings of indices:
Stellenbosch University http://scholar.sun.ac.za
%-HJ
Table 4-11 1inear model coefficients for t#e combinations of t#e competition indices
:odel 20F &qrt(20F. ; , A07
sqrt(Db#0ncrement. J kkl K 1*A
$3G4F $3"G# #34GG $##G34GH
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 $3%4H #3##G F##3JG T"e-$4 UUU
))l -#3$"I #3##F -FJ3#I T"e-$4 UUU
*+A -#3#$$ #3### -4G3HF T"e-$4 UUU
sqrt(Db#0ncrement. J kkl K !g70
"3J$G $3H#H #3G#% F4H$3HHG
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 $3F""JJ% #3##G#IG "4#3$H T"e-$4 UUU
))l -#3$F4%H4 #3##GFGI -"G3%H T"e-$4 UUU
HgC- -#3#%#$HJ #3##$%%H -"H3HH T"e-$4 UUU
sqrt(Db#0ncrement. J 1*A K *A1
F3#FF $3H%" #34G$ $#IJ3H4F
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 $3%I%$%J4 #3##%IGH F#G3GG T"e-$4 UUU
*+A -#3##H%"I #3###""H -F"3HF T"e-$4 UUU
+A* -#3#$#J#% #3###"IJ -FH3HH T"e-$4 UUU
sqrt(Db#0ncrement. J !g70 K *A1
"3#"% $3%"F #34FF $%GH3FI#
(stimate Std3 (rror t value Pr.XYtY0
.-ntercept0 $3%F"%IJG #3##%%%H F""3$" T"e-$4 UUU
HgC- -#3#"444$ #3##$#FH -"G3H$ T"e-$4 UUU
+A* -#3#$%""$ #3###"%" -GI3I" T"e-$4 UUU
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
The com&ination of the ::* and *+A indices had the &est 9
"
values, lowest A-C values and
had lowest E-F < suggesting that the! were the least collinear of the a&ove com&inations
and represented the &est fit3 A com&ination of *+A and +A* also resulted in a good fit=
however the E-F was the highest as the! are &ased on the same competition inde/3
The com&ination of ::* and *+A intuitivel! represent different modes of competition, which
is 2uantified in the &etter fit o&tained and the low collinearit! &etween the two C-Ls3 These
two were thus selected as a candidate com&ination to test the changing mode of
competition and com&ination of C-Ls for the modifier function3
$.&.# /elative importance
>ow that a candidate com&ination was o&tained, where collinear effects would &e minimal
to the e/planation value of the two C-Ls - a comparison &etween the importance of varia&les
in each site is given3 The full model and the suggested reduced model .::* and *+A0 are
introduced to see how the! change over the two sites < assuming that the water inde/ is
representative of a gradient3
Stellenbosch University http://scholar.sun.ac.za
%-I#
The relaimpo pac)age developed &! 'rPmping ."##40 in 9 was used for this purpose3 The
pac)age provides different methods for decomposing the contri&ution of 9
"
with different
predictor varia&les provided that the regressors are not correlated3 'rPmping ."##40
recommends that two particular methods AlmgD method and Apm%dD method, with lmg
more appropriate for causal anal!sis and pm%d more appropriate for predictive anal!sis
.'rPmping "##403 The other methods were included here for a complete picture to anal!se
whether the relative importance changes consistentl! with other variance decomposition
methods3
The results for the >elder and Tweefontein trials are summarised in Table 4-1, &elow, using
different variance decomposition methods for illustration and to see consistenc! with other
methods, although the lmg and pmvd methods are of importance3
Table 4-1, ;elati%e importance proportions (e'planation contribution. of t#e AA1 and 1*A indices respecti%el" using
different importance measures
9elder Tweefontein
:et#od AA1 1*A AA1 1*A
lmg #3%F$ #3G4J #3%#H #3GJF
pm%d #3F"J #34H$ #3"GG #3H%G
betasq #3F"J #34H$ #3"GG #3H%G
car #3%$F #3GIH #3FHJ #34"$
last #3F"J #34H$ #3"GG #3H%G
first #3%GH #3G%F #3%%F #3GGH
Total ;, 4$3H$N 4J3$4N
The *+A was attri&uted consistentl! a higher importance to the variance e/planation of the
model on ever! method t!pe3 -n the lmg method, this was small, suggesting a small
difference &etween sites, however the pm%d method showed a much larger difference3 The
results are illustrated in Figure 4-1/ and Figure 4-14 for the Tweefontein and >elder sites
respectivel!, with JGN &ootstrapped confidence intervals for the respective importance
values for the different C-Ls3
Stellenbosch University http://scholar.sun.ac.za
%-I$
Figure 4-1/ ;elati%e importance grap#s using different met#ods for t#e Tweefontein spacing trial
Stellenbosch University http://scholar.sun.ac.za
%-I"
Figure 4-14 ;elati%e importance grap#s using different met#ods for t#e 9elder spacing trial s#owing impro%ed AA1
importance as compared to t#e Tweefontein site
9egardless of the method it is clear the *+A has the largest contri&ution to the 9
"
value in
&oth sites, with the ::* increasing in importance in the >elder site, which was classified as a
wetter site according to the aridit! inde/ .Table 4-10 for &oth the lmg and pm%d methods
consistentl!3 -t can &e tentativel! h!pothesised that this could &e due to the &etter moisture
conditions on the site3 This is an encouraging result for changing importance of competition
indices &ased on resource limitations and changing mode of competition3 However more
sites are necessar! for proper validation of such a h!pothesis3 For the purpose of this thesis
water inde/ was used as a considered addition to the modifier function3
$.( tep &: =se CI>s in a deterministic potential modifier
e6uation
Step " defined the potentials, which can &e modelled for an! given diameter on a certain
site3 1sing the information gained in the previous section, the competition indices can now
&e incorporated into a modifier3 Two o&7ectives are represented here: to see whether the
Stellenbosch University http://scholar.sun.ac.za
%-IF
incorporation of an additional competition inde/, which would &alance the light oriented
::* with an inde/ more focussed on local crowding and edaphic limitations, would improve
the model3 The second o&7ective is to see if the incorporation of a water inde/, which can
scale the C-Ls according to the availa&le water, improves the model fit3
$.(.1 ,b-ective 1: Incorporating the :B! competition inde4
1sing the ::* and the *+A in a linear model within the modifier of the potential modifier
approach results in Equation 4-5:
Equation 4-5
here is the mean annual diameter increment, is the potential diameter
increment for that specific site with its given site inde/ calculated in Equation 4-, and
illustrated in Figure 4-4, and the e/ponential function with the linear e2uation representing
the modifier3 For the >elder and Tweefontein sites together this was represented &!:
Table 4-1/ 7oefficients of t#e diameter increment potential modifier formula (Equation 4-5.
Formula Equation 4-5
)arameter Estimate &td. Error t %alue )r(CDtD.
a -#3%""%J4" #3##4HF"" -4"3H4 T "e-$4 UUU
b -$3$JJ$JF4 #3#%#I%%% -"J3F4 T "e-$4 UUU
c -#3##F4$$F #3###GFI" -43H$ "3#J(-$$ UUU
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
9esidual standard error: #34H#F on H%JF degrees of freedom
The e/ponent of the linear e2uation scaled the linear e2uation &etween # .a&solute
theoretical competition: no growth0 and $ .no competition: ma/imum growth03 The
intercept presented a potential pro&lem as it will alwa!s reduce the potential .i3e3 the
modifier will never reach $ .the potential growth03 However, e/cluding this intercept
resulted in an over scaled, unrealistic model where too man! trees are growing at the
potential or not growing at all, which would not &e ideal for the simulation3 For this reason,
it was decided not to e/clude the intercept from the model in order to o&tain a good
average, with a stochastic component to &e added later .Step 403
As stated, the first o&7ective is to see whether including the local &asal area in addition to
the ::* improved the model3 This was done &! onl! including either the *+A or ::* in
Equation 4-5, comparing each of them in addition to the com&ined model3
Stellenbosch University http://scholar.sun.ac.za
%-I%
Table 4-14 Ano%a comparison if Equation 4-5 wit# AA1+ 1*A and bot# indices included in :odel 1+ , and / respecti%el"
Anal"sis of 2ariance Table
:odel ;es.Df ;es.&um &q Df &um &q A07 ;:&E F %alue )r(CF.
1 H%J% FFIG3G $GF"#3F% #34H"#%
, H%J% %$FJ3J # # $4I"I3FI #3H%F$GH
/ H%JF FF443G $ HHF3FH $G"I#3"J #34H#$GI $H"$3F T "3"#(-$4 UUU
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
The Anova comparison shows a significant improvement to the A-C, 98S( and F-statistic3
-ncluding &oth indices seems to indicate that including the *+A in con7unction with ::* in
Equation 4-5 improves the predictive fit for diameter increment modelling3
$.(.2 ,b-ective 2: Incorporating a water inde4
-n order to include the water availa&ilit! with the aridit! inde/ presented in Table 4-1, a new
linear model inside of the modifier can &e incorporated:
Equation 4-6
here ! represents the FA; < 1>(P .$JJ"0 water inde/, ""# and #$% are the C-Ls and a is
the intercept3 This can &e incorporated as the modifier into Equation 4 -@:
Equation 4-@
Table 4-15 7oefficients of t#e diameter increment potential modifier formula (Equation 4-@. incorporating t#e water
inde'
Formula Equation @
)arameter Estimate &td. Error t %alue )r(CDtD.
a -#3%F#"G #3##4I4I -4"34%H T "e-$4 UUU
b -$3$G4"H #3#%#""" -"I3H%H T "e-$4 UUU
c -#3##"JI #3###GF -G34"" $3J4(-#I UUU
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
9esidual standard error: #34HI" on H%JF degrees of freedom
Section %3G3F .9elative importance0 showed that for the two sites &eing used for
parameterisation, the importance of the indices changed for the s2uare root transformed
model3 Thus the ne/t o&7ective was to see whether the water inde/ does indeed improve
the model &! comparing Equation 4-@ and Equation 4-5
Stellenbosch University http://scholar.sun.ac.za
%-IG
Table 4-16 Ano%a comparison of t#e diameter increment model wit# (:odel1. and wit#out (:odel ,. t#e water inde'
Anal!sis of Eariance Ta&le
8odel 9es3Df 9es3Sum S2 Df Sum S2 A-C 98S( F value Pr.XF0
$ H%JF F%%43J $G%GH3#" #34HI$#%
" H%JF FF443G # # $G"I#3"J #34H#$GI
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
Table 4-16 shows no significant difference &etween the two models &ased on the F-statistic3
This could mean that with onl! two sites containing num&ers of o&servations and onl! two
water inde/ values, not enough information was availa&le to test whether this should &e
included in the full model, despite the plausi&le relative changes of the varia&le importance
o&served .%3G3F03 Furthermore, as the water inde/ was a static value for each site, it ma! &e
improved &! a water availa&ilit! inde/ in order to introduce d!namic effects3 1nfortunatel!,
such an inde/ was not availa&le in this stud!3 -n order to test whether the model
performance was due to the relative size of the datasets < random samples from the >elder
e/periment were ta)en to match the num&er of o&servations of the Tweefontein site and a
comparison tested .Table 4-1@03
Table 4-1@G Ano%a comparison of t#e diameter increment model wit# (:odel1. and wit#out (:odel ,. t#e water inde'
b" randoml" subsetting data from t#e 9elder trial to matc# t#e number of obser%ations in t#e Tweefontein trial
Anal"sis of 2ariance Table
:odel 9es3Df 9es3Sum S2 Df &um &q A07 ;:&E F %alue )r(CF.
1 "%FJ $#JF3F %JHG3HF" #344J$$I
, "%FJ $#%I3I # # %IH%3"F% #34GGFG4
Signif3 codes: # @UUUL #3##$ @UUL #3#$ @UL #3#G @3L #3$ @ L $
This result was consistent with multiple randomised samples, suggesting that the water
gradient provided here &etween the two sites ma! not have &een large enough to cover
changes in C- &ehaviour with water availa&ilit!3 For further modelling it was decided to use
the model .Equation 4-@0 with the water inde/ on the different sites nonetheless since it
showed a logical &ehaviour and was statisticall! not different in the e/planation value than
Equation 4-5 without the water inde/3
$.(.# *odel behaviour
Predictions from (2uation H were used on the two datasets to anal!se the structure of the
predicted values compared to the o&served increment values3 Figure 4-15 shows the
procedure where the potentials for each given site are calculated .a0, after which a modifier
.&0 reduces the increment3 The o&served and predicted increments for each site are shown
se2uentiall! .c-f03
Stellenbosch University http://scholar.sun.ac.za
%-I4
a. )otential increment cur%es b. :odifier
c. 9elder obser%ed d. 9elder predicted
e. Tweefontein obser%ed f. Tweefontein predicted
Figure 4-15 0llustration of t#e potential modifier approac# on t#e two spacing trials+ wit# t#e top rig#t #and diagram
representing t#e c#ange of t#e modifier o%er t#e si$e of competition
Stellenbosch University http://scholar.sun.ac.za
%-IH
The predicted increment follows a pattern that mimics the trend of competition at different
D+H values .Figure 4-10= however the predicted values show much less varia&ilit! than the
o&served increments shown a&ove3
hat is also evident is that the predicted increments are constrained &oth &! the potential
and &! the ma/imum value of the modifier .which was #34G in this case0, which means that
the increment cannot increase a&ove #34G times the increment potential for a given D+H3
The depression of the modifier was due to the inclusion of an intercept in the modifier
e2uation .a in Equation 4-@0, which was nonetheless included &ecause it includes variation
from factors une/plained in the model and prevents e/treme values in the modifier3 -t was
decided to test the performance of this model, and then to incorporate random varia&ilit!
as a stochastic component for further improvement in the ne/t section3
$.(.$ *odel validation
The chosen model .Equation 4-@0 was tested on the >elder spacing trial, with the first
measurement age .%3GI !ears in the >elder trial0 used as a starting point3 The simulation
time step period was not initiated .in S-*EA it is G !ears03 -n this case the simulation is
performed &etween measurement periods to test how the model reacts in the different
time steps3 This means that the effect of competition is constant &etween two
measurement ages < thus the simulated increment will remain the same for that period3
The simulated series of the deterministic model .without variance component0 produces a
series where the diameter increment and thus D+H is overpredicted from age 4 to "F, after
which point it &egins to converge again .Appendi/ C03 The full range of simulated periods is
shown in Appendi/ D3 Figure 4-16 &elow shows the end of the simulation at "4 !ears of the
>elder trial .the final measured age of the trial03
Stellenbosch University http://scholar.sun.ac.za
%-II
Figure 4-16 7omparison between t#e obser%ed (red bars. and t#e simulated diameter (D*!. represented as a
distribution of ,.5cm diameter classes. T#e mean D*! for t#e site is represented in t#e notc#ed bo'plot. T#is s#owed a
skewed prediction grouped around a narrower D*! band
As can &e seen from the simulation end .Figure 4-160 and Appendi/ D the diameter
simulated D+H distri&ution shifts towards higher D+H classes, clustered around a narrow
range of D+H3 hile this is still a significant improvement from a simple mean model, this
could &e pro&lematic in the simulation as it will reduce the varia&ilit! of the diameter
classes in the model and reduce the competitive comple/it!3 Thus with each simulation step
the model continues to shift in this trend, there&! compounding the error3 For this reason it
was necessar! to see whether the increment prediction with varia&ilit! improved the model
< and finall! in which rings .and corresponding SPHA0 of the >elder trials it overpredicted3
$.+ tep (: Create a stochastic model incorporating natural
variabilit%
$.+.1 Incorporating natural variation
Due to the comple/ nature of &iological s!stems, and our ina&ilit! .or difficult!0 to capture
this comple/it! in a model with distinct varia&les, it is often necessar! to account for this
varia&ilit! around the mean value with the help of stochastic models .Pretzsch "##J03 -n this
case it is desired to appro/imate .or mimic0 the high degree of varia&ilit! of diameter
increment, which could &e caused &! factors not included in the model, e3g3 micro-site
variation in water and nutrient suppl!, weather, genetics, etc3 As is seen &! the scatter plot
of Figure 4-1@, which represent a residual plot of the predicted value deviation, while much
of the increment patterns can &e discerned from stand ageVcompetition and the diameter
of the tree, there remains a need to incorporate random variance into the model in order to
mimic natural varia&ilit!3 Furthermore, the application of the model led to a mismatched
Stellenbosch University http://scholar.sun.ac.za
%-IJ
diameter distri&ution, s)ewed to the right and highl! concentrated around a narrow &and of
diameter classes .Figure 4-1603 For this reason a stochastic model was tested that
incorporates varia&ilit! &! including a random deviate in the increment function3
-t was decided to incorporate natural variance &! anal!sing the residual scatter plots of the
predicted values for the two parameterisation sites, with the residual standard deviation of
the o&served minus the predicted values plotted over the predicted values .Figure 4-1@03
*inear 2uantile regression models of the residual plots were created to represent the upper
and lower &ounds and a new random deviate would then &e generated &etween these two
&ounds &ased on a normal distri&ution3
Figure 4-1@ ;esidual plot of t#e predicted %alues of t#e spacing trials
The scatter plot shows that the variance seems to increase slightl! with increasing predicted
increment values3 This is more clear in the negative tail of the distri&ution, which is
e/pected since the increment is &ound &! the lower zero value .diameter does not usuall!
&ecome negative03 To model this relationship the #3JG and #3#G tau value .GN and JGN0
linear 2uantile regression lines were used to represent the upper and lower &ounds
respectivel! of the residual plot .Figure 4-1403 The &enefit of the 2uantile line is that it
allows continuous variance &ounds for an! given predicted value3
Stellenbosch University http://scholar.sun.ac.za
%-J#
Figure 4-14 5L and 35L linear quantile regression lines w#ic# represents t#e upper and lower bounds for t#e truncated
normal distribution
This significant positive .tauM#3JG0 and negative .tauM#3#G0 slope values for the 2uantile
lines illustrate an increasing varia&ilit! for higher predicted values, with strong deviation in
the negative direction, while the variance in the positive residual direction onl! increases
slightl!3 The values of the 2uantile lines are shown in Table 4-14 &elow3 For instance this
means that &ounds of the variance for a diameter increment of $3G would &e &etween -#3J"
and $3F#3
Table 4-14 1inear quantile regression coefficients of -.-5 and -.35 tau %alues representing t#e upper and lower bounds
of simulated residual prediction
rq(formula B 0ncrement residuals J )redicted 0ncrement+ tau B c(-.-5+ -.35..
tau %alue )arameter 2alue &td. Error t %alue )r(CDtD.
tau -.-5
(0ntercept. -#3$I4%I #3#""GJ -I3"G%$4 #
0ncrement -#3%IHI% #3#$%4% -FF3FF"F #
tau -.35
(0ntercept. $3$G#44 #3#FH"J F#3IG%FH #
0ncrement #3#J%%H #3#"G% F3H$III #3###"
hile the 2uantiles do not capture all of the data points, the! proved to &e the most ro&ust
with some o&served variation sacrificed3
Stellenbosch University http://scholar.sun.ac.za
%-J$
As is seen from Figure 4-14 the predicted &ounds from the 2uantile regression is 2uite
large, indicating a high degree of varia&ilit!, however, the points are still strongl! clustered
around the mean or predicted value and the distance of the &ounds increases with
predicted increment3 For this reason and &ecause the distance &etween the &ounds changes
over the size of the predicted increment0, a weighting factor was added to the standard
deviation value of the normal distri&ution, where the standard deviation is defined &!:
Stc: = u.S (uppcr lowcr)
Equation 4-4
where St&e' is the standard deviation of the normal distri&ution, (pper and lo)er are the
truncation &ounds for an! predicted value3
The results of the predicted diameter increment &ased on the complete stochastic model
produced a scatter plot illustrated &elow overlain on the o&served residual plot .Figure 4-
130, with the full visual procedure illustrated in Figure 4-,-3
Figure 4-13 )redicted residual generated from a random de%iate between two bounds (Figure 14. based on a normal
distribution
Stellenbosch University http://scholar.sun.ac.za
%-J"
Figure 4-,- &equential process of generating random de%iation+ from plotting residuals (top left.+ predicting bounds
(top rig#t.+ generating random de%iation (bottom left. superimposed on t#e residual plot (bottom rig#t.
The new predicted scatterplot showed a similarit! to the o&served natural deviation3 -t
remained to &e seen how these deviate result in an improvement of the increment model3
The new predicted icrement values with the random deviates added had to &e compared to
the increment values3 This was done visuall! initiall! to see if it matches the structure of the
increment-D+H scatter plots, with its distinctive pattern3 Figure 4-,1 and Figure 4-,, &elow
shows the improvement of the stucture from the o&served increment for the >elder and
the Tweefontein sites respectivel!, with the predicted plots without deviation and the
predicted plots with deviation3
Stellenbosch University http://scholar.sun.ac.za
%-JF
Figure 4-,1 Obser%ed and predicted scatterplots for t#e Tweefontein &&-77T spacing trial+ wit# black points
representing t#e obser%ed and green representing t#e predicted %alues. T#e left #and image represents t#e scatter of
t#e a%erage model M wit# t#e linear quantile in t#e top rig#t+ t#e bottom left and rig#t images represent t#e scatter of
t#e model wit# added modelled random %ariance
Stellenbosch University http://scholar.sun.ac.za
%-J%
Figure 4-,, Obser%ed and predicted scatterplots for t#e 1ottering 9elder spacing trial+ wit# black points representing
t#e obser%ed and green representing t#e predicted %alues. T#e top rig#t image s#ows t#e deterministic model+ t#e
bottom left and rig#t images represent t#e scatter of t#e model wit# added modelled random %ariance
Stellenbosch University http://scholar.sun.ac.za
%-JG
Figure 4-,/ ;esidual scatter of t#e added random %ariance model for t#e parameterised sites
Stellenbosch University http://scholar.sun.ac.za
%-J4
Figure 4-,4 ;esidual scatter o%erlain on t#e obser%ed increment scatter of t#e added random %ariance model for t#e
parameterised sites. T#e black points represent t#e obser%ed and t#e green points represent t#e predicted %alues from
t#e stoc#astic model.
The new plots .Figure 4-,/ and 4-,40 seemed to mimic the o&served increment 2uite well,
matching the point clouds o&tained from the competitive situation at different
measurement ages, e/cept for a depressed tail of points at higher D+H values .with lower
predicted increment curves03 -t is now necessar! to see how the new random deviates
&ehave in a simulation over time3 This is done in the validation section &elow3
*odel 'alidation
Adding varia&ilit! should prevent model from collecting or converging towards a mean
around one D+H class, as was seen form the simulated model .Appendi/ D03 The same
simulation was done as in Section %343%, this time with the random variance added to the
prediction, shown in Appendi/ ( with the end point at "4 !ears shown in Figure 4-,53
Diameter increment is clearl! overestimated at most ages, for &oth the model including
varia&ilit! and without3
Stellenbosch University http://scholar.sun.ac.za
%-JH
Figure 4-,5 Obser%ed (red bars. %s. simulated (blue bars. for t#e 9elder trial wit# random de%iation added at eac#
point. T#e model s#ows a clear o%erall o%erpredicition of D*!+ #owe%er wit# an impro%ed distribution compared to t#e
model wit#out de%iation (Figure 4-16..
From a diameter distri&ution point of view, the predicted model is clearl! superior <
although shifted to the right due to the overpredicted d&h originating from the
deterministic model part3 The model overall showed an overprediction for the >elder trial,
however with the e/treme planted densities that e/ist .from 4H##- $"4 spha0 it is necessar!
to see at which densities the model is not predicting well3 For the end of the simulation run
the following Table 4-13 shows the average o&served vs3 predicted D+H values for the
different >elder rings .and corresponding stems per hectare0 at "4 !ears:
Stellenbosch University http://scholar.sun.ac.za
%-JI
Table 4-13 A%erage of obser%ed and predicted %alues for t#e different 9elder plots+ s#owing clearl" t#at t#e model
underperforms at e'treme densities
)lot
(ring.
&)!A
()lanted.
A%erage of predicted
D*!(cm.
A%erage of obser%ed D*!
(cm.
Difference
% $"4 %"3%4 GF3$G -$#34J
G $4H %F3$I G$3%# -I3""
4 """ %"34% %I34I -43#%
H "JG FJ3G4 %%3JF -G3FH
I FJ$ %#3$% %"3## -$3IG
J G"# FH3G4 F43%I $3#I
$# 4J$ F43FH F%34G $3H$
$$ J$H F%3$4 F$3J% "3""
$" $"$J F"3"" "I3II F3F%
$F $4$J "J3JJ "F3I4 43$F
$% "$G$ "H3%# ""3G# %3J#
$G "IGH "G3HJ $J34H 43$"
$4 FHJ4 "F3%$ $I3F$ G3$#
$H G#%F ""3J% $43%# 43G%
$I 4H## ""3FG $%3HI H3GH
Table 4-13 shows that the model overpredicts at high densities and underpredicts at ver!
low densities= however, at moderate densities at which most stands would &e planted and
managed under in plantation stands .highlighted in &lue0, the difference is relativel! small3
This difference is illustrated as a percentage in Figure 4-,6 &elow3
Stellenbosch University http://scholar.sun.ac.za
%-JJ
Figure 4-,6 De%iation of predicted a%erage diameter increment for a more realistic of planted stand densities for a
plantation industr" setup in t#e 9elder trial
$.8 Chapter conclusion
The main o&7ective of this chapter was to model diameter increment in as a function of
resource limitations and site classification for application of a methodological &ased
approach for a specific dataset to outline pro&lems and future wor)3 This was split into the 4
su&-o&7ectives which formed the outline of the wor)ing steps of the chapter3 This main
o&7ective was achieved, however some )e! issues arose which will &e discussed according to
the si/ wor)ing stepsVo&7ectives of the Chapter3
The water inde/ introduced in the first step proved not to &e ade2uate to capture a large
enough gradient &etween the two sites for modelling purposes3 However, it did result in a
shift &etween importance of the competition indices discussed later3 Potential increment
curves were applied &ased on a given D+H on different stands, which showed clear
differences &etween sites3 -ncorporating site inde/ as a site 2ualit! measure into the
potential increment model resulted in a gradient of predicted potentials from site 2ualit!3
-n order to include a modifier multiple distance-dependent competition indices were tested
on two sites with availa&le tree positions with the o&7ective to identif! a com&ination of
Stellenbosch University http://scholar.sun.ac.za
%-$##
indices that showed small collinearit! while e/plaining diameter increment well3 The )e!
idea &ehind this was to capture a gradient &etween overtopping .light limitation0 and local
crowding .edaphic limitations0 modes of competition as highlighted &! Seifert et al. .in
press03 The chosen competition indices reacted differentl!, &ut in order to identif! and
2uantif! the differences the &ehaviour and correlation &etween the indices were e/amined3
Differences in PCA &iplots and improved fit with low collinearit! resulted in the decision to
choose the *+A and ::* indices in a com&ined model3 -ncorporating a competition inde/
which captures more size s!mmetric competition .*+A0 seemed, under the fitted model
constructed a&ove, to improve prediction compared to model rel!ing purel! on the
standard S-*EA competition inde/ ::*3 The relative importance of the ::* and *+A to the
transformed model changed under different site conditions on the two sites, with the ::*
increasing in importance in the less drought affected site3 -ncorporating &oth indices in one
model has potential to capture changing competition s!mmetr! under changing water
availa&ilit! - although this needs to &e investigated under a much wider gradient of sites
with var!ing water availa&ilit! and perhaps of different !ears of the same site in a
longitudinal stud! approach3
The results have shown that including the aridit! inde/ in the modifier e2uation presents
potential for inclusion of a water inde/ into the diameter increment prediction3 Although
the chosen approach was ver! simple, restricted &! site and data availa&ilit!, and relia&le
conclusions are difficult to ma)e with regards to changing importance of C- models, it could
&e used for further parameterisation of S-*EA3 The ideal situation would &e to com&ine a
water inde/ and weather data with a soil water &alance model to estimate water
availa&ilit!, which was outside the scope of this thesis3 -n this stud! the longitudinal data
aspect was not considered due to time and resource limitations, however, it must &e noted
that this ma! have a significant effect, which should &e anal!sed in future studies3
The presented model matched competitive stages, achieving the o&7ective of o&taining an
age-independent model3 The potential modifier model with the competition indices were
still not a&le to capture the wide range of varia&ilit! o&served in diameter increment
predictions= this lead to a s)ewed, narrow prediction of D+H distri&ution of predicted
compared to o&served distri&utions3
The use of a linear 2uantile regression for the truncation of a normal distri&ution &etween
residuals, and a random prediction &ased on a normal distri&ution proved to &e a ro&ust
methodolog! to incorporate natural variation3 However, the captured varia&ilit! was not
large enough to mimic the variance in o&served stands3 -t is the opinion of the author that
the pro&lem originates from the deterministic model, where the competition indices were
still not sensitive enough to model diameter &ased solel! on a potential and competition in
the modifier3 -t is thought that including allometric relationships, such as diameterVheight or
crown dimensions into the model would greatl! improve the situation3 The structure of the
::* inde/ ma! also not currentl! &e suited to capture the competitive situation in the South
Stellenbosch University http://scholar.sun.ac.za
%-$#$
African conte/t= changes could &e made to the angle of the search cone which detects and
counts potential competitors3 This could &e tested in a sensitivit! anal!sis3 The pro&lem
could also &e related to the size of the trees < larger trees ma! &e e/pected to deviate in the
upper regions of the random spectrum for instance3 This was not addressed in the model
and represents a )e! issue for future wor)3
Further refinements to the stochastic model could &e made &! testing the normalit!
assumption and &! using a ei&ull distri&ution for instance3 The variation predicted also did
not ta)e into account densit! effects in the o&served residual plot, which ma! affect the size
of the variation of diameter increment under different degrees of competition .a higher
variation at low competition for instance0= this could &e a consideration for an investigation
for future studies, which ma! ma)e the stochastic model more sensitive to different stand
densit! ranges3
Eisual anal!sis of the new predicted increments showed a good li)eness to o&served
-ncrement-D+H scatterplots3 -ncorporating a varia&ilit! to move into stochastic modelling
drasticall! improved the D+H distri&ution pro&lem, matching o&served D+H values
reasona&l! well3 The model seemed to capture the average situation 2uite well3
This Chapter serves as an e/ample of a methodological approach to simulate diameter
increment com&inations of distance dependent competition indices and stochastic
modelling3 Future studies can greatl! improve the process &! considering further sites with
different water indices and site 2ualit! and &! stud!ing the shifting importance of
competition indices under different water availa&ilit! in order to increase the sensitivit! of
the model to changing climatic conditions3
Stellenbosch University http://scholar.sun.ac.za
G-$#"
Chapter &: Conclusion and recommendations
This chapter provides a &rief overview of what was achieved and recommendations for
future wor), as most of the issues have &een discussed in the individual chapters alread!3
This is divided into two sections: the wor) which was applied in this thesis .the potential
height and diameter increment models in Chapters F and %0 and steps for further
parameterisation for S-*EA, which were not applied in this thesis3
&.1 Potential height modelling
The overall o&7ective of this chapter was to model potential height and to see whether site
inde/ can &e used as a predictor varia&le for potential height3 -n this sense the overall
o&7ective was achieved as it seemed to &e a via&le approach3
>onlinear 2uantile regression proved to &e a ro&ust method for conditional median
modelling, in this case of the potential height3 The algorithm used seems to &e less affected
&! outliers than other least s2uares estimation techni2ues3 There were some limitations
worth noting: for one, as with man! models, it is &ound &! the o&served values of the data
represented, and ma! thus &e una&le to 2uantif! the true potential3 This limits the potential
height model as it is then &! definition una&le to further model e/ceptional trees, which
have a large effect on competition and neigh&ours3 This can &e remedied &! the inclusion of
stochastic modelling techni2ues, as applied in Chapter % where&! the allowance of a
random deviation procedure could allow some trees to var! a&ove the allocated potential
after the application of a modifier3
8odelling site inde/ using different fitting techni2ues provided potential new and
interesting techni2ues for site inde/ modelling3 >onlinear mi/ed effects modelling proved
superior in predictive 2ualit! compared to the >*S and '>*S techni2ues for the data range
ac2uired3 -ncluding the as!mptote as the random effect provided the most ro&ust and
accurate inclusion of random effects into the >*8( methodolog!3
1sing site inde/ as a predictor, the overall o&7ective of the chapter, proved to &e possi&le3
Some validation and refinements to the process could &e made, however it is thought that
site inde/ is a&le to model potential height 2uite well3 A ma7or concern would &e the effect
of densit! on site inde/, while this thesis used the South African site inde/ definition as is, it
could lead to su&stantial errors at e/treme .ver! high or low0 stand densities and is an issue
that should &e addressed in future wor)3
Future wor) should focus on a site 2ualit! predictor &ased on edaphic conditions .(sler
"#$", *ouw and Scholes "##"0, which should improve the sensitivit! of the model to site
conditions and move the model into a h!&rid modelling approach3
Stellenbosch University http://scholar.sun.ac.za
G-$#F
&.2 .iameter increment modelling
The overall o&7ective of the chapter was, under data limitations, to present a methodolog!
for future parameterisation and to highlight pro&lems associated with this3 -n this sense, the
o&7ective was achieved= however improvements to this methodolog! can still &e made3
Stratif!ing the indices &ased on the collinear relationships in order to o&tain models with a
com&ination of good predictive power to diameter increment and low collinearit! to each
other using PCA and E-FLs was a novel and powerful approach3
As noted earlier, the discussion on the &ehaviour of competition indices on different site
t!pes must &e ta)en ver! tentativel!, as two sites does not represent a good gradient3
However, some trends can &e noted and the validation produced some interesting results,
showing clearl! that the importance of the different indices3
-ncluding the *ocal +asal Area inde/, which captures a more growth s!mmetric competition
structure representing edaphic factors, improved the model and often performed &etter
that the ::* for P. elliottii under the water limited South African conditions3
-t is necessar! though to see how this changes for different growth periods, ta)ing into
account rainfall and water availa&ilit! over different seasons and !ears to see whether the
competitive nature of the trees can switch &etween s!mmetric and as!mmetric
competition3 -t is strongl! suggested that a stud! &e underta)en to understand the cross-
sectional .site stratification0 and temporal shifts in competition mode3 This could for
e/ample &e done in com&ination with a dendrochronological stud!3
&.# !dditions for model completion
As this thesis did not cover the entire modelling and simulation process done in Silva,
further wor) must &e done in the following areas, although no strong deductions can &e
made until the! are tested in the South African growing conte/t3
The omission of the species specific crown model and light transmission factor ma! have
&een a ma7or factor in the &ehaviour of the ::* model3 A model should &e developed for
South Africa to model the crown structure of P. elliottii and other species to &e used in
future parameterisations, which could &e used in the competition indices and allometric
e2uations to improve growth prediction3 Height increment modelling &ased on potential
height and the stand initialisation should &e studied3 Although height increment is not as
strongl! affected &! competition as diameter increment, competition in addition to
allometric models and site 2ualit! could &e investigated3 8ortalit! was not parameterised in
this stud!3 Some surrogates could &e used, however a more size and competition &ased
approach should &e used for parameterisation into S-*EA3
Stellenbosch University http://scholar.sun.ac.za
G-$#%
&.$ ,verall thoughts
hile the two components studied are ver! different in their model approaches and are not
com&ined into a full model !et, the overall o&7ectives for each were o&tained3 The a&ove
process in the thesis gives a good indication of the comple/it! of constructing single-tree
growth models3 However, the thesis resulted in a strong indication that this t!pe of
modelling process would &e feasi&le, with the main focus to shift on completing the
simulation structure of the entire growth modelling structure and to focus on shifts in
competition inde/ and competition modes under changing water availa&ilit!3 Pro&lems
encountered, and the solutions suggested, with the models tested could &e of use for
further research3
Stellenbosch University http://scholar.sun.ac.za
4-$#G
Chapter (: /eferences
A)ai)e, H3 .$JH"03 Statistical predictor identification3 Ann. %nst. Math., )), "#F<"#H3
Ac)erman, S3A3, Ac)erman, P3A3, Seifert, T3 ."#$F03 (ffects of irregular stand structure on
tree growth, crown e/tension and &ranchiness of plantation grown Pinus patula3 Southern
Forests* +,&$(* "%H<"G43
+ot)in, D3 +3, Qana), Q3 F3, ] allis, Q3 93 .$JH"03 Some (cological Conse2uences of a
Computer 8odel of Forest 'rowth3 -ournal of cology, ./.F0, I%J<IH"3
+ragg, D3 C3 ."##$03 Potential relative increment . P9- 0: a new method to empiricall! derive
optimal tree diameter growth3 cological Modelling, 0"+, HH<J"3
+reden)amp, +3E3 .$JI%03 The C3C3T3 concept in spacing research < a review3 Proceedings:
S!mposium on site and productivit! of fast growing plantations3 Eol $:F$F-FF"3
+reden)amp, +3 E3 .$JJ#03 The Triple-S CCT design3 -n Proceedings of a Symposium arranged
by the Forest Mensuration and Modelling 1or2ing 3roup in collaboration 4ith the Southern
African %nstitute of Forestry and the ucalyptus grandis !esearch 5et4or2 on 6Management
of ucalyptus grandis in South Africa7 .pp3 $JI<"#G03
+reden)amp, +3E3 .$JJF03 Top height: a definition for use in South Africa3 Southern African
Forestry -ournal* $4H: GG3
+rown, '3 .$J4G03 Point density in stems per acre .p3 $"03
Cade, +3, ] >oon, +3 ."##F03 A gentle introduction to 2uantile regression for ecologists3 Front
col nviron, 0.I0, %$"< %"#3
Chertov, ;3 '3 .$JJ#03 SP(C;8 - A Signle Tree 8odel of Pine StandV9aw Hums Soil
(cos!stem3 cological Modelling, ,/, $#H<$F"3
Corral-9ivas, Q3, Sanchez ;rois, S3, :otze, H3 and 'adow von, :3 ."##J03 Testing the suita&ilit!
of the >epal-Somers stand ta&le pro7ection method for (ucal!ptus grandis plantations in
South Africa3 Southern Forests "##J, H$ .F0: "#H-"$%3
Dalgaard, P3 ."##I03 %ntroductory Statistics 4ith !3 >ew 6or): Springer3
Du Toit, +3 ."#$"03 8atching site, species and silvicultural regime to optimise the
productivit! of commercial softwood species in Southern Africa3 -n South African Forestry
8andboo2 .Gth ed3, pp3 %F<G#03 Southern African -nstitute of Forestr!3
Stellenbosch University http://scholar.sun.ac.za
4-$#4
D!e, P3 3, Qaco&s, S3, ] Drew, D3 ."##%03 Eerification of F-P' growth and water-use
predictions in twelve (ucal!ptus plantation stands in ^ululand, South Africa3 Forest cology
and Management, 09" .$-"0, $JH<"$I3 doi:$#3$#$4V73foreco3"##%3#$3#F#
Dzierson H, 8ason ('3 ."##403 Towards a nationwide growth and !ield model for radiata
plantations in >ew ^ealand3 Can Q For 9es3 F4: "GFF-"G%F3
(), A3, ] Dude), A3 .$JI#03 Development of individual tree &ased stand growth simulators:
preogress and applications3
(sler, +3 ."#$"03 :n the development and application of indirect site indices based on
edaphoclimatic variables for commercial forestry in South Africa. Stellen&osch universit!3
Fair&an)s, D3H3:3, Scholes, 93Q3 .$JJJ03 South African Countr! Stud! on Climate Change:
Eulnera&ilit! and Adaptation Assessment for Plantation Forestr!3 Prepared for: >ational
9esearch Facilit!3
Fang, ^3, ] +aile!, 93 *3 ."##$03 >onlinear mi/ed effects modeling for slash pine dominant
height growth following intensive silvicultural treatments3 Forest Science, $+.F0, "IH<F##3
Fe)edulegn, D3, Siurtain, 83 P3 8ac, Col&ert, Q3 Q3, Siurtain, 83, ] Parameter, Q3 Q3 .$JJJ03
Parameter (stimation of >onlinear 'rowth 8odels in Forestr!3 Silva Fennica, "".%0, F"H<
FF43
'ea--z2uierdo, '3, CaSellas, -3, ] 8ontero, '3 ."##I03 Site inde/ in agroforestr! s!stems: age-
dependent and age-independent d!namic diameter growth models for Kuercus ile/ in
-&erian open oa) woodlands3 ;anadian -ournal of Forest !esearch, "<.$0, $#$<$$F3
'rPmping, 13 ."##403 9elative -mportance for *inear 9egression in 9: The Pac)age relaimpo3
-ournal of Statistical Soft4are, 0+.$03
Heg!i, F3 .$JH%03 A simulation model for managing Qac)-pine stands3 -n Q3 Fries .(d30, 3ro4th
models for tree and stand simulation .pp3 H%<J#03 9o!al College of Forest, Stoc)holm,
Sweden3
Hotelling, H3 .$JFF03 Anal!sis of a comple/ of statistical varia&les into principal components3
-ournal of ducational Psychology, )$.40, %$H<%%$3
:assier, H3 3 .$JJF03 =ynamics of diameter and height distributions in even#aged pine
plantations3 1niversit! of Stellen&osch3
:oen)er, 93 ."##403 Kuantile regression in r: a vignette3 9etrieved Qul! $4, "#$", from
http:VVwww3econ3uiuc3eduVZrogerVresearchVr2Vvig3pdf
:oen)er, 93, ] +assett, '3 .$JHI03 >o Title3 conometrica, $..$0, FF<G#3
Stellenbosch University http://scholar.sun.ac.za
4-$#H
:oen)er, 93, ] Halloc), :3 F3 ."##$03 Kuantile 9egression3 -ournal of conomic Perspectives,
0,.%0, $%F<$G43 doi:$#3$"GHV7ep3$G3%3$%F
:omarov, A3, Chertov, ;3, ^udin, S3, >adporozhs)a!a, 83, 8i)hailov, A3, +!)hovets, S3, \
^ou&)ova, (3 ."##F03 (F-8;D "`a model of growth and c!cling of elements in &oreal forest
ecos!stems3 cological Modelling, 0+/."-F0, FHF<FJ"3
:otze, H3, :assier, H3 3, Fletcher, 63, ] 8orle!, T3 ."#$"03 'rowth 8odelling and 6ield
Ta&les3 -n +3 +reden)amp ] S3 1pfold .(ds30, South African Forestry 8andboo2 .Gth ed3, pp3
$HG<"$J03 Southern African -nstitute of Forestr!3
*ands&erg, Q3 Q3, ] aring, 93 H3 .$JJH03 A generalised model of forest productivit! using
simplified concepts of radiation-use efficienc!, car&on &alance and partitioning3 Forest
cology and Management, 9,.F0, "#J<""I3
*ee, 3:3 and 'adow, :3v3 .$JJH03 -terative +estimmung der :on)urrenz&aume in Pinus
densiflora +estanden3 AF-> $4I.FV%0: %$-%%3
*e)wadi, S3 ;3, >emesova, a3, *!nch, T3, Phillips, H3, Hunter, a3, ] 8ac Sibrtcin, 83 ."#$"03
Site classification and growth models for Sit)a spruce plantations in -reland3 Forest cology
and Management, )<", G4<4G3
*ouw, Q3 H3, ] Scholes, 83 ."##"03 Forest site classification and evaluation: a South African
perspective3 Forest cology and Management, 0+0, $GF<$4I3
8c:ee, T3, Does)en, >3, ] :leist, Q3 .$JJG03 Drought monitoring with multiple time scales3 -n
9th ;onference on Applied ;limatology .pp3 "FF<"F403 Dalas3
8unro, D3 .$JH%03 Forest growth models - a prognosis3 -n Q3 Fries .(d30, 3ro4th models for
tree and stand simulation .9esearch >3, pp3 H<"$03 Stoc)holm, Sweden3
8!ers, 93 H3 .$JI403 ;lassical and Modern !egression 4ith Applications3 +oston: PS
Pu&lishers3
>elder, Q3 A3 .$J4"03 >ew )inds of s!stematic designs for spacing e/periments3 ?ometrics, 0<,
"IF<F#J3
;LConnor, A3Q3 .$JFG03 Forest research with special reference to planting distances and
thinning3 +ritish (mpire Forestr! Conference, South Africa3 F# pp3
;dhiam&o, +3, 8einc)en, 83, ] Seifert, T3 ."#$%03 The protective role of &ar) against fire
damage: a comparative stud! on selected introduced and indigenous tree species in the
estern Cape, South Africa3 Trees, D;- $#3$##HVs##%4I-#$F-#JH$-#
Palmer, 3 .$J4G03 Meteorological =rought. !esearch Paper $,3 ashington DC3
Stellenbosch University http://scholar.sun.ac.za
4-$#I
Pa!andeh, +3 .$JH%03 >onlinear site inde/ e2uations for several ma7or Canadian forest
tim&er species3 For. ;hron ,/, $J%-$J43
Pinheiro, Q3 C3, ] +ates, D3 83 ."###03 Mi@ed#ffects Models in S and S#PAUS3 >ew 6or):
Springer3
Porte, A3, ] +artelin), H3 H3 ."##"03 8odelling mi/ed forest growth: a review of models for
forest management3 cological Modelling, 0,/, $%$<$II3
Po!nton, 93 Q3 .$JHJ03 Tree Planting in Southern Africa, Eolume $, The Pines3 Pretoria:
Department of Forestr!3
Pretzsch, H, +i&er, P3, ] Durs)!, Q3 ."##"03 The single tree-&ased stand simulator S-*EA:
construction , application and evaluation3 Forest cology and Management, 0.), F<"$3
Pretzsch, Hans3 ."##J03 Forest =ynamics* 3ro4th and Bield .p3 44%03 +erlin and Heidel&erg:
Springer3
Pretzsch, Hans, ] +i&er, P3 ."#$#03 Size-s!mmetric versus size-as!mmetric competition and
growth partitioning among trees in forest stands along an ecological gradient in central
(urope3 ;anadian -ournal of Forest !esearch, $/."0, FH#<FI%3
9eed, D3 D3, Qones, (3 A3, Tome, 83, ] Arau7o, 83 C3 ."##F03 8odels of potential height and
diameter for (ucal!ptus glo&ulus in Portugal3 Forest cology and Management, 0+), $J$<
$JI3
9itchie, 83 3, ] Hann, D3 3 .$JJ#03 (2uations for Predicting the G-6ear in Southwest
;regon Height 'rowth of Si/ Conifer Species, .8arch03
9Ptzer T, Seifert T, Pretzsch H3 ."##J03 A&ove and &elow ground car&on d!namics in a mi/ed
&eech and spruce stand influenced &! environmental changes3 (uropean Qournal of Forest
9esearch $"I."0: $H$ < $I"3
9Ptzer T, Seifert T, 'a!ler S, Priesac) (, Pretzsch H3 ."#$"03 (ffects of stress and defence
allocation defence on tree growth: Simulation results at the tree and stand level3 -n:
8at!sse) 9, Schn!der H, (rnst D, 8unch Q-C, ;dwald , Pretzsch H .eds0 'rowth and
Defence in Plants: 9esource Allocation at 8ultiple Scales3 (cological Studies ""#, Springer3
%#$-%F"3
Schwinning, S3, ] einer, Q3 .$JJI03 8echanisms determining the degree of size as!mmetr!
in competition among plants3 :ecologia, 00".%0, %%H<%GG3
Seifert, T3, ] Seifert, S3 ."#$%03 8odelling and simulation of tree &iomass3 -n T3 Seifert .(d30,
?ioenergy from 1oodC Sustainable Production in the Tropics .8anaging F3, pp3 %"<4G03
Springer3
Stellenbosch University http://scholar.sun.ac.za
4-$#J
Seifert, T3, Seifert, S3, Se!dac), A3, Durrheim, '3, ] von 'adow, :3 .n3d303 Comeptition effect
in an afrotemperate fores3 Forest (cos!stems .in revie40
Seifert, T3 ."##H03 Simulating the e/tent of deca! caused &! Hetero&asidion annosum s3 l3 in
stems of >orwa! spruce3 Forest cology and Management, )$<.$-"0, JG<$#43
Seifert, T, ] Pretzsch, H3 ."##%03 8odeling growth and 2ualit! of >orwa! spruce .Picea a&ies
:arst30 with the growth simulator S-*EA3 -n Proceedings of the $th 1or2shop 6;onnection
bet4een Forest !esources and 1ood DualityC Modelling Approaches and Simulation
Soft4are7 .pp3 G4"<GH%03 -1F9; or)ing Part! SG3#$3#%, Harrison Hotsprings, +ritish
Colum&ia, Canada3
Seifert T, du Toit +, Qooste 'H, Sei&old S, Durrheim ', Se!dac) A .in press0, Towards an
individual tree forest growth model for natural forests in South Africa: Testing competition
indices to model individual tree diameter growth in Afro-temperate forests3 G
th
>atural
Forests and oodlands S!mposium, 9ichards +a!, South Africa, $#3<$%3 April "#$$3
Seifert S, 1tschig, H3 ."##"03 Standraum und achstum - Eoronoi-Diagramme zur
+estimmung des achstums in +estanden3 +ericht des Forschungszentrums
aldP)os!steme, 1niversitat 'Pttingen, 9eihe +. 4I0, $$$-$$"3
Stene)er, '3 A3, ] Qarvis, Q3 83 .$J4F03 A preliminar! stud! to assess competition in a white
spruce < trem&ling aspen stand3 Forestry ;hronicles, "9, FF%<FF43
Stoll, P3, einer, Q3, 8uller-*andau, H3, 8Rller, (3, ] Hara, T3 ."##"03 Size s!mmetr! of
competition alters &iomass-densit! relationships3 Proceedings. ?iological sciences E The
!oyal Society, ).9.$G#40, "$J$<G3
Team, 93 C3 ."#$F03 9: A *anguage and (nvironment for Statistical Computing3
Thornthwaite, C3 .$J%I03 An approach towards the rational classification of climate3
3eographical !evie4, "<, GG<J%3
1>(P3 .$JJ"03 1orld Atlas of =esertification3 *ondon: (dward Arnold3
1hl (, 8etzger H-', Seifert T ."##40 Dimension und achstum von solitaren +uchen und
(ichen3 Tagungs&and der Qahrestagung der Se)tion (rtrags)unde im Deutschen Eer&and
Forstlicher Forschungsanstalten, %H < GF3
Eancla!, Q3 :3 .$JJG03 'rowth models for tropical forests: A s!nthesis of models and
methods3 Forest Science, $0.$0, %<%"3
van Staden, Eida, (rasmus, +3F3>3, 9ou/, Q3, ingfield, 83Q3, van Qaarsveld, A3S3 ."##%03
8odelling the spatial distri&ution of two important South African plantation forestr!
pathogens3 Forest cology and Management $IH: 4$<HF3
Stellenbosch University http://scholar.sun.ac.za
4-$$#
Eon 'adow, :3, ] +reden)amp, +3 .$JJ"03 Forest management .p3 $G$03 Preotria: Academia3
ar&urton, 83, Schulze 93 ."##403 Climate change and the South African commercial
forestr! sector: an initial stud!3 9eport to Forestr! SA3 A;!Ucons !eport ,$* Decem&er "##43
ichmann, *3 ."##$03 Annual Eariations in Competition S!mmetr! in (ven-aged Sit)a
Spruce3 Annals of ?otany, <<.$0, $%G<$G$3 doi:$#3$##4Van&o3"##$3$%%G
!)off, 393, Croo)ston, >3*3 and Stage, A393 .$JI"03 1serLs guide to the stand prognosis
model3 1SDA For3 Serv3 'en3 Tech3 9ep3 >o3 ->T-$FF3
Stellenbosch University http://scholar.sun.ac.za
H-$$$
Chapter +: !ppendices
!ppendi4 !: 3:*7 random effects anova table
Densit"F)lot :odel )arameter 7on%ergence (?+9. df A07 *07 log1ik Test 1.;atio p-%alue
1,4
$ a,&,c >
" a,& >
F a,c 6 J GG3H""#" 4G3J%$%4 -$I3I4$#$
% a 6 H G$3GII%4 GJ3GF4J" -$I3HJ%"F
G &,c 6 J GG3H"F4G 4G3J%F#J -$I3I4$I"
4 & >
H c 6 H G$3H"F4I GJ34H"$% -$I3I4$I% G vs H F3$$(-#G $
,4@
$ a,&,c 6 $" $#J3FHG $F"3#H4I -%"34IHG
" a,& >
F a,c 6 J $#F3IH#GG $"#3IJ4J -%"3JFG"I $ vs F #3%JGGGGG #3J$JJ
% a >
G &,c 6 J $#F3FH%JI $"#3%#$% -%"34IH%J
4 & 6 H JJ3IGG4F $$F3#JI% -%"3J"HI$ G vs 4 #3%I#4%J% #3HI4%
H c 6 H JJ3IGG4F $$F3#JI% -%"3J"HI"
/@1
$ a,&,c >
" a,& >
F a,c >
% a 6 H $%I3G$H $4%3IF"" -4H3"GIG"
G &,c 6 J $G$3FG#" $H"3F"4I -4434HG#I
4 & 6 H $%J3$J%J $4G3G$ -4H3GJH%F G vs 4 $3I%%H #3FJH4
H c 6 H $%I3F$F% $4%34"IG -4H3$G44J
434
$ a,&,c 6 $" $H%34$JI "#%3%IF% -HG3F#JII
" a,& 6 J $4I34$I% $J$3#$4$ -HG3F#J$J $ vs " #3##$FHF $
F a,c 6 J $H%3"FGG $J434FFF -HI3$$HHH
% a 6 H $4G3$"%F $I"3G%%H -HG3G4"$%
G &,c >
4 & 6 H $4F3IFG4 $I$3"G4$ -H%3J$HI 4 vs F 43FJJJFG #3#%#I
H c 6 H $H#3"F4$ $IH34G44 -HI3$$I#G
@41
$ a,&,c >
" a,& 6 J "4J3GH4I "JG3J"" -$"G3HII%
F a,c 6 J "4J3GH44 "JG3J"$J -$"G3HIIF
% a 6 H "4G3GHGG "I43#44" -$"G3HIHH
G &,c 6 J "H"3#4HG "JI3%$"I -$"H3#FFI
4 & 6 H "4I3#4HI "II3GGI4 -$"H3#FFJ 4 vs G #3###"%G""" #3JJJJ
H c 6 H "4J3$G#$ "IJ34%#J -$"H3GHG$
Stellenbosch University http://scholar.sun.ac.za
H-$$"
Densit"F)lot :odel )arameter 7on%ergence (?+9. df A07 *07 log1ik Test 1.;atio p-%alue
344
$ a,&,c 6 $" FF$344FG F4J3#HJ% -$GF3IF$H
" a,& 6 J F"G34I"$ FGF3H%%$ -$GF3I%$$ $ vs " #3#$I4FF$HH #3JJJF
F a,c 6 J F"G34HG FGF3HF4J -$GF3IFHG
% a 6 H F"$34H$% F%F3%JH% -$GF3IFGH F vs % #3##FGF#J%J #3JJI"
G &,c >
4 & >
H c >
144/
$ a,&,c 6 $" %JI3IG%I G%#3$4FI -"FH3%"H%
" a,& 6 J %J"3IG%J G"F3IF44 -"FH3%"H%
$ vs
" "3"H(-#G $
F a,c >
% a 6 H %II3IG%I G$"3JG$I -"FH3%"H%
G &,c >
4 & 6 H %II3IG%I G$"3JG$I -"FH3%"H% 4 vs " "3%H(-#G $
H c 6 H %II3"IG G$"3FI" -"FH3$%"G
,365
$ a,&,c nlme3fit$ $" H4F3H$$% I$"3GII -F4J3IGGH
" a,& nlme3fit" J HGH3H"FH HJ%3FI$$ -F4J3I4$I $ vs " #3#$""4J4F4 #3JJJ4
F a,c nlme3fitF J HGH3H$#H HJ%3F4I$ -F4J3IGGF
% a nlme3fitH H HGF3H#HI HI"3"$J$ -F4J3IGFJ
G &,c nlme3fit% J HGI3"FF" HJ%3IJ#4 -FH#3$$44
4 & nlme3fitG H HG%3"F"I HI"3H%%$ -FH#3$$4% G vs 4 #3###%#F#G$ #3JJJI
H c nlme3fit4 H H%%3G$FH HHF3#"G -F4G3"G4J
Stellenbosch University http://scholar.sun.ac.za
H-$$F
!ppendi4 B: 'oronoi pol%gon increment relationship for the
3elder and Tweefontein sites respectivel%
Stellenbosch University http://scholar.sun.ac.za
H-$$%
Stellenbosch University http://scholar.sun.ac.za
H-$$G
!ppendi4 C: ,bserved and predicted .B5 for the
deterministic model
Stellenbosch University http://scholar.sun.ac.za
H-$$4
!ppendi4 .: imulation steps for the deterministic model
Period 1: 6 years
Period 2: 7 years
Period 3: 11.08 years
Stellenbosch University http://scholar.sun.ac.za
H-$$H
Period 4: 13.33 years
Period 5:15.58
Period6: 19.17
Stellenbosch University http://scholar.sun.ac.za
H-$$I
Period 7:20.25
Period 8:23.83
Period 9:26
Stellenbosch University http://scholar.sun.ac.za
H-$$J
Stellenbosch University http://scholar.sun.ac.za
H-$"#
!ppendi4 7: imulation steps for the stochastic model
Period 1: 6 years
Period 2: 7 years
Period 3: 11.08 years
Stellenbosch University http://scholar.sun.ac.za
H-$"$
Period 4: 13.33 years
Period 5:15.58
Stellenbosch University http://scholar.sun.ac.za
H-$""
Period6: 19.17
Period 7:20.25
Stellenbosch University http://scholar.sun.ac.za
H-$"F
Period 8:23.83
Period 9:26
Stellenbosch University http://scholar.sun.ac.za
H-$"%
Stellenbosch University http://scholar.sun.ac.za
H-$"G
!ppendi4 ": elect e4amples of /2 Code
#########################################################
####### Nonlinear quantile regression example ###########
#########################################################
Allccttrials <- subset(Allccttrials, Height > 0)
str(Allccttrials)
head(Allccttrials)
###1.2 Subset the data into the different trials
Weza <- subset(Allccttrials, Location == "Weza")
Mac <- subset(Allccttrials, Location == "Mac Mac")
Kwam <- subset(Allccttrials, Location == "Kwambonambi")
Duku <- subset(Allccttrials, Location == "Dukuduku")
###2. Deriv For functions
#2.1.1 Chapman Richards Three Parameter
height.chapman <-
deriv(~ a * (1 - exp(b*x))^c,
c("a","b","c"),
function(x, a, b, c){},
hessian = TRUE)
Weza2965 <- subset(Weza, Spha =="2965")
a <- max(Weza2965$Height)
b1 <- max(Weza2965$Height) - min(Weza2965$Height)
b2 <- max(Weza2965$Age) - min(Weza2965$Age)
b3 <- (b1/b2)
b <- b3/(max(Weza2965$Height))
c <- 0.66
a
b
c
Weza.0.9 <- nlrq(Height ~ height.chapman(Age, a, b, c),
start = list(a = a, b = -b, c = c),
data = Weza2965, tau=0.9, trace=TRUE)
Weza.0.95 <- nlrq(Height ~ height.chapman(Age, a, b, c),
start = list(a = a, b = -b, c = c),
data = Weza2965, tau=0.95, trace=TRUE)
Weza.0.975 <- nlrq(Height ~ height.chapman(Age, a, b, c),
start = list(a = a, b = -b, c = c),
data = Weza2965, tau=0.975, trace=TRUE)
Stellenbosch University http://scholar.sun.ac.za
H-$"4
#Test0.9
handy.Height.hat <- function(Age)
predict(Weza.0.9, newdata = data.frame(Age = Age))
#Call
par(las = 1)
Plotnlrq <- plot(Height ~ Age, data = Weza2965,
xlim = c(0, max(Weza2965$Age, na.rm=TRUE)),
ylim = c(0, max(45)),
ylab = "Height(m)", xlab = "Age (y)", main = "0.9 Quantile")
curve(handy.Height.hat, add = TRUE)
summary(Weza.0.9)
#Test0.95
handy.Height.hat2 <- function(Age)
predict(Weza.0.95, newdata = data.frame(Age = Age))
#Call
par(las = 1)
Plotnlrq <- plot(Height ~ Age, data = Weza2965,
xlim = c(0, max(Weza2965$Age, na.rm=TRUE)),
ylim = c(0, max(45)),
ylab = "Height(m)", xlab = "Age (y)", main = "0.95 Quantile")
curve(handy.Height.hat2, add = TRUE)
summary(Weza.0.95)
#Test0.975
handy.Height.hat3 <- function(Age)
predict(Weza.0.975, newdata = data.frame(Age = Age))
#Call
par(las = 1)
Plotnlrq <- plot(Height ~ Age, data = Weza2965,
xlim = c(0, max(Weza2965$Age, na.rm=TRUE)),
ylim = c(0, max(45)),
ylab = "Height(m)", xlab = "Age (y)", main = "0.975 Quantile")
+ curve(handy.Height.hat, add = TRUE, lty = 2, lwd = 2, col =
"Green")
+ curve(handy.Height.hat2, add = TRUE, lty = 2, lwd = 2, col =
"Blue")
+ curve(handy.Height.hat3, add = TRUE, lty = 2, lwd = 2, col =
"Red")
legend(1, 45, c("0.90","0.95","0.975"), lty=c(2,2,2),
lwd=c(2,2,2),col=c("Green","Blue","Red"))
summary(Weza.0.975)
#########################################################
###### Site index modelling example ############
#########################################################
Stellenbosch University http://scholar.sun.ac.za
H-$"H
Wezadom1 <- subset(dominantheightclass, Location == "Weza")
Wezadom247 <- subset(Wezadom1, Spha == "247")
Wezadom247 <- subset(Wezadom247, Height > 0)
a <- max(Wezadom247$Height)
b1 <- max(Wezadom247$Height) - min(Wezadom247$Height)
b2 <- max(Wezadom247$Age) - min(Wezadom247$Age)
b3 <- (b1/b2)
b <- b3/(max(Wezadom247$Height))
c <- 0.66
a
c
b
nlc <- nls.control(maxiter = 200, tol = 1e-05, minFactor = 1/1024,
printEval = FALSE, warnOnly = FALSE)
handy.nls2 <-
nls(Height ~ Height.growth(Age, a, b, c),
start = list(a = a, b = -b, c = c),
data = Wezadom247, control = nlc)
#Test
handy.Height.hat <- function(Age)
predict(handy.nls2, newdata = data.frame(Age = Age))
#Call
par(las = 1)
Plotted2 <- plot(Height ~ Age, data = Wezadom247,
xlim = c(0, max(Wezadom247$Age, na.rm=TRUE)),
ylim = c(0, max(45)),
ylab = "Height(m)", xlab = "Age (y)", main = "Average Height")
curve(handy.Height.hat, col = "red", add = TRUE)
summ <- summary(handy.nls2)
require(nlstools)
resid <- nlsResiduals(handy.nls2)
plotresid <- plot(resid, type = 0)
library(nlme)
Weza247group <- groupedData(Height~Age|Number, Wezadom247)
plot(Weza247group)
head(Weza247group)
cor(coef(handy.nls2))
####Nlme
a <- max(Weza247group$Height)
b1 <- max((Weza247group$Height) - min(Weza247group$Height))
b2 <- max(Weza247group$Age) - min(Weza247group$Age)
b3 <- (b1/b2)
b <- b3/(max(Weza247group$Height))
c <- 0.66
Stellenbosch University http://scholar.sun.ac.za
H-$"I
a
b
c
nlc1 <- nlmeControl(maxiter = 100000, tol = 1e-05, minFactor =
1/1024,
printEval = FALSE, warnOnly = FALSE)
nlc2 <- nls.control(maxiter = 500, tol = 1e-05, minFactor = 1/1024,
printEval = FALSE, warnOnly = FALSE)
Weza247nlslist <-
nlsList(Height ~ Height.growth(Age, a, b, c), data = Weza247group,
start = list(a = 36.74088, b = -0.03314, c = 1.05124), control =
nlc2)
summary(Weza247nlslist)
pairs(Weza247nlslist, id = 0.1, na.action = na.exclude)
plot(intervals(Weza247nlslist), layout = c(3,1))
start <- c(a = a,b = -b, c = c) # starting value
nlme.fit <- nlme(Height ~ Height.growth(Age, a, b, c),
fixed = a + b + c ~ 1, random = a+ b + c ~ 1,
data = Weza247group, start=c(start),
control=nlmeControl(opt='nlm'),
weights = varPower(form = ~ Age), corr =
corAR1(0.296739967))
summary(nlme.fit)
pairs(nlme.fit)
anova(nlme.fit, handy.nls2)
###3 - Substantial correlation between b and c random effects -
makes it difficult to converge
intervals(nlme.fit, which = "var-cov")
nlme.fit.a <- update(nlme.fit, random = a ~ 1)
nlme.fit.b <- update(nlme.fit, random = b ~ 1)
nlme.fit.c <- update(nlme.fit, random = c ~ 1)
nlme.fit.bc <- update(nlme.fit, random = b + c ~ 1)
anova(nlme.fit.a, nlme.fit.b, nlme.fit.c, nlme.fit.bc)
anova(nlme.fit.a, nlme.fit.bc)
summary(nlme.fit.a)
summ
##Fit diagonal variance covariance matrix assuming random effects
are independent
nlme.fit.diag <- update(nlme.fit, random = pdDiag(a + b + c ~ 1),
control=nlmeControl(opt='nlm'))
#doesnt converge, thus try other random effects
nlme.fit.diagbc <- update(nlme.fit, random = pdDiag(b + c ~ 1))
nlme.fit.diagac <- update(nlme.fit, random = pdDiag(a + c ~ 1),
control=nlmeControl(opt='nlm'))
Stellenbosch University http://scholar.sun.ac.za
H-$"J
nlme.fit.diagab <- update(nlme.fit, random = pdDiag(a + b ~ 1),
control=nlmeControl(opt='nlm'))
nlme.fit.diaga <- update(nlme.fit, random = pdDiag(a ~ 1),
control=nlmeControl(opt='nlm'))
nlme.fit.diagb <- update(nlme.fit, random = pdDiag(b ~ 1),
control=nlmeControl(opt='nlm'))
nlme.fit.diagc <- update(nlme.fit, random = pdDiag(c ~ 1),
control=nlmeControl(opt='nlm'))
anova(nlme.fit.diagac, nlme.fit.diagab, nlme.fit.diaga,
nlme.fit.diagb, nlme.fit.diagc)
####With a, b and c all substantially corellated, it may not be
necessary to include them all in the model
# Just using a seems to be the best model
summary(nlme.fit.diaga)
summary(handy.nls2)
anova(nlme.fit.diaga, handy.nls2)
plot(handy.nls2)
plot(nlme.fit.diaga)
plot(nlme.fit.diagb)
plot(nlme.fit.diagbc)
#Now use the weights
nlme.fit.diaga <- update(nlme.fit.diaga, weights = varPower(form = ~
Age) )
summary(nlme.fit.diaga)
plot(nlme.fit.diaga)
anova(nlme.fit.diaga, nlme.fit)
###gnls
library(nlme)
Weza247gnls <- gnls(Height ~ Height.growth(Age, a, b, c), data =
Weza247group,
start = list(a = 46.86677, b = -0.04293, c = 1.45860))
summary(Weza247gnls)
summ
Nelder126gnls.Power3 <- update(Weza247gnls, weights = varPower(form
= ~ Age), corr = corAR1(0.296739967))
summary(Nelder126gnls.Power3)
plot(Nelder126gnls.Power3)
gnlspowerwightedresiduals <- Nelder126gnls.Power3
nlmepowerweightedresiduals <- nlme.fit.diaga
nlme <- nlme.fit
nls <- handy.nls2
anova(Nelder126gnls.Power3, nlme.fit.a, handy.nls2)
Stellenbosch University http://scholar.sun.ac.za
H-$F#
####rmse
RMSE(Nelder126gnls.Power3)
RMSE(nlme.fit.a)
RMSE(handy.nls2)
anova(gnlspowerwightedresiduals, nlmepowerweightedresiduals, nlme,
nls)
plot(nls)
plot(nlme)
plot(gnlspowerwightedresiduals)
plot(nlmepowerweightedresiduals)
##Test the autocorrelation regression improvements for the best nlme
(weighted and unweighted)
#Unweighted:
plot(ACF(nlme.fit, maxLag = 10), alpha = 0.05)
ACF(nlme.fit)
nlme.fit.acf <- update(nlme.fit, corr = corAR1(0.11717144))
plot(ACF(nlme.fit.acf , maxLag = 10), alpha = 0.05)
summary(nlme.fit.acf)
summary(nlme.fit)
anova(nlme.fit, nlme.fit.acf)
plot(nlme.fit.acf)
# Weighted
plot(ACF(nlme.fit.diaga, maxLag = 10), alpha = 0.05)
ACF(nlme.fit.diaga)
nlme.fit.diaga.acf <- update(nlme.fit.diaga, corr =
corAR1(0.29005147))
plot(ACF(nlme.fit.diaga.acf, maxLag = 10), alpha = 0.05)
ACF(nlme.fit.diaga.acf)
summary(nlme.fit.diaga)
summary(nlme.fit.diaga.acf)
anova(nlme.fit.diaga.acf, nlme.fit.diaga)
plot(nlme.fit.diaga.acf)
plot(nlme.fit.diaga)
anova(nlme.fit, nlme.fit.acf, nlme.fit.diaga, nlme.fit.diaga.acf)
##See if the Nlmeweighted and unweighted acf's compare
summary(nlme.fit.acf)
summary(nlme.fit.diaga.acf)
anova(nlme.fit.acf, nlme.fit.diaga.acf)
##See if changing the parameters used for random effects has and
effect
nlme.fit.diagbc <- update(nlme.fit, random = pdDiag(b + c ~ 1))
plot(nlme.fit.diagbc)
Stellenbosch University http://scholar.sun.ac.za
H-$F$
nlme.fit.diag.bc <- update(nlme.fit.diagbc , weights = varPower(form
= ~ Age) )
ACF(nlme.fit.diag.bc)
nlme.fit.diaga.bc.acf <- update(nlme.fit.diag.bc, corr =
corAR1(0.29005147))
plot(nlme.fit.diaga.bc.acf)
#unweighted
nlme.fit.diaga.unweight <- update(nlme.fit.diagbc, corr =
corAR1(0.29005147))
plot(nlme.fit.diaga.unweight)
summary(nlme.fit.diaga.bc.acf)
summary(nlme.fit.diaga.unweight)
anova(nlme.fit.diaga.bc.acf, nlme.fit.diaga.unweight)
anova(gnlspowerwightedresiduals, nlmepowerweightedresiduals, nlme,
nls, nlme.fit.diaga.acf)
#########################################################
##### Competition indices exapmle #############
#########################################################
####BALIndex
#df <- read.csv("C:/Users/Gerard/Documents/Masters/r voronoi/Nelder
Year by year/1998.csv")
df <- read.csv("F:/Tweefontein BAL/Tweefonteinnew.csv")
#df<- subset(df, TPH0 =="245")
#df<- subset(df, TPH0 =="403")
#df<- subset(df, TPH0 =="665")
df<- subset(df, TPH0 =="1097")
#df<- subset(df, TPH0 =="1808")
#df<- subset(df, TPH0 =="2981")
head(df)
# computes the BAL competition index as follows:
# all neighbour trees inside the competion zone radius (CZR)
# contibute to the total sum of the Basal area, if their
# DBH is thicker than the central tree.
#
# alle trees are assumed to be on a axias allinged rectangular plot
#
# stefan seifert 2013-1
require(RANN)
df<- subset(df, Age =="1.83")
Stellenbosch University http://scholar.sun.ac.za
H-$F"
df<- subset(df, Age =="2.5")
df<- subset(df, Age =="3.58")
df<- subset(df, Age =="6.25")
df<- subset(df, Age =="7.83")
df<- subset(df, Age =="10.5")
df<- subset(df, Age =="12.67")
df<- subset(df, Age =="14.58")
df<- subset(df, Age =="16.58")
df<- subset(df, Age =="18.33")
df<- subset(df, Age =="20.42")
d <- df
head(d)
createtorus = function(d, dx=NA, dy=NA) {
if (is.na(dx))
dx=diff(range(d$x))
if (is.na(dy))
dy=diff(range(d$y))
dnew = d
dnew$outside=F
# centerrow
dc=rbind(dnew,transform(d, x=x-dx, outside = T))
dc=rbind(dc,transform(d, x=x+dx, outside = T))
#+top
dnew = rbind(dc,transform(dc, y=y-dy, outside=T))
#+bottom row
dnew = rbind(dnew, transform(dc, y=y+dy, outside=T))
return(dnew)
}
head(createtorus(d))
neighbours.BAL = function(x,y,r=plotsize,k=100) {
# returns all neighbours in a distance of r
# neighbour number is limited to k
# warning will be given if there might be more neighbous than k
require(RANN)
k = min(k,nrow(x))
neigh = nn2(x[,c("x","y")], x[,c("x","y")], searchtype="radius",
radius = r-1e-6, k=k)
if(max(neigh$nn.idx[,k])>0 & k<nrow(x)) {
warning("neighbours.BAL: k might be to less to find all
potential neighbours")
}
return(neigh)
}
p = createtorus(d)
head(p)
str(p)
BAL = function(p,CZR=plotsize) {
# compute the BAL index, BASAL AREA, and DIAM for all trees in p
Stellenbosch University http://scholar.sun.ac.za
H-$FF
# p is a data.frame with the columns x, y for the tree position
# and D for the DBH of the tree
# a column newID with unique IDs
# CZR is teh radius to which neighbours are counted
# if CZR is NA then CZR will be estimated as 2 times the
# squareroot of the specific tree area
# bounding
BAL = rep(NA,nrow(p))
BASAL = BAL
CR13 = BAL
mimax = range(p$x)
mimay = range(p$y)
if (is.na(CZR)) {
a = diff(mimax)*diff(mimay)
spa = a/nrow(p) # 1/intensity : specific area per tree
CZR = sqrt(spa) * 2 # k-facor 2 : radius is 2-times the
specific tree area
}
border = with(p, x<=mimax[2]-CZR & x>=mimax[1]+CZR &
y<=mimay[2]-CZR & y>=mimay[1]+CZR)
sqha = CZR^2*pi/10000 # circle ha
ne = neighbours.BAL(p,CZR, k=100) # all neighbours for all trees!
# sum up all valid neighbours squared DBHs
# we do it in a loop
for (ct in 1:nrow(p)) {
nonu = ne$nn.idx[ct,-1]
nonu = nonu[nonu>0]
d = p[nonu,"D"]
d1 = d[d>p$D[ct]]
d2 = d[d>10 & d <=30]
BAL[ct] = sum( d1^2 )*pi/40000 / sqha # [ basal area per ha]
BASAL[ct] = sum( d^2 )*pi/40000 / sqha
CR13[ct] = sum(d2) / sqha
}
return(data.frame(newID=p$newID, BAL=BAL,BASAL=BASAL, Age =
p$Age, DBH = p$D, DIncrement = p$Di,HtIncrement = p$Hti,
Ht = p$Ht, x = p$x, y = p$y, SPHA = p$TPH0,
flagBALOk=border, CR13=CR13, CZR=rep(CZR,nrow(p)) ))
}
BALCALL <- BAL(p)
head(BALCALL)
plot(BALCALL$BAL, BALCALL$DIncrement)
plot(BALCALL$BASAL, BALCALL$DIncrement)
#################Hegyi Index
p <- BALCALL
head(p)
names(p)[names(p) == "newID"] <- "ID"
Stellenbosch University http://scholar.sun.ac.za
H-$F%
p$newID <- 1:nrow(p)
p$D <- p$DBH
CZR <- plotsize
neighbours.heg = function(x,r,k=100) {
# returns all neighbours in a distance of r
# neighbour number is limited to k
# warning will be given if there might be more neighbous than k
require(RANN)
k = min(k,nrow(x))
neigh = nn2(x[,c("x","y")], x[,c("x","y")], searchtype="radius",
radius = r-1e-6, k=k)
if(max(neigh$nn.idx[,k])>0 & k<nrow(x)) {
warning("neighbours.heg: k might be to less to find all
potential neighbours")
}
return(neigh)
}
it.heg.stand = function(p,CZR, k=100) {
# p are all trees in the plot
# this data.frame must contain x,y as coordinates
# and d as diameter [cm]
# the column newID should contain a unique ID like the row
number
# if some trees ar excluded this ID helps to reconstruct the
# original data
require(plyr)
ne = neighbours.heg(p,CZR, k) # all neighbours for all trees!
HgCI = rep(NA,nrow(p))
newID = HgCI
flagBorder = rep(NA,nrow(p))
mimax = range(p$x)
mimay = range(p$y)
flagBorder = with(p, x<=mimax[2]-CZR & x>=mimax[1]+CZR &
y<=mimay[2]-CZR & y>=mimay[1]+CZR)
for (ct in 1:nrow(ne$nn.idx)) {
#print(ct)
# test if current tree ct is to near the maximum extent of
the plot (B:think this comment is meant for line 53)
z = c()
# number of neighbours
nonna = (ne$nn.idx[ct,]>0)
nc = length(na.omit(ne$nn.idx[ct,nonna])) # B:na.omit just
removes incomplete cases, i.e. more than number k, or out of CZR
neighb = ne$nn.idx[ct,nonna][-1] # all valid neighbour
indexes
neighbd = ne$nn.dist[ct,nonna][-1] # all valid neighbours
distances
HgCI[ct]=sum( (p$D[neighb]/p$D[ct])/neighbd )
newID[ct]=p$newID[ct]
Stellenbosch University http://scholar.sun.ac.za
H-$FG
}
return(data.frame(HgCI=HgCI,flagInside=flagBorder,newID=newID))
}
source("J:/B Fury/ci-Hegyi-iterNocea.R")#specifiy which script has
the functions being used
head(p)
str(p)
heg=it.heg.stand(p,CZR,35)
plot(p$x,p$y)
head(heg)
str(heg)
heg.1=data.frame(ID = p$ID, newID=p$newID,
BAL=p$BAL,BASAL=p$BASAL,HgCI=heg$HgCI, Age = p$Age,
DBH = p$DBH, DbhIncrement = p$DIncrement,
HtIncrement = p$HtIncrement,
Ht = p$Ht, x = p$x, y = p$y, SPHA = p$SPHA,
CZR=rep(CZR,nrow(p)))
plot(heg.1$HgCI, heg.1$DbhIncrement)
head(heg.1)
str(heg.1)
###########################Voronoi
p <- heg.1
head(p)
str(p)
treecoords <- subset(p, select=c(newID,x,y))
require(deldir)
x<-c(treecoords$x)
y<-c(treecoords$y)
plot(x,y)
delresult <- deldir(x,y)
summ<-delresult[["summary"]]
summ
summ8<-summ[8]
summ8merge <- merge(summ8, heg.1, all=TRUE, by="row.names")
plot(delresult)
plot(delresult,add=FALSE,wlines=c("tess"),
wpoints=c("both","real","dummy","none"),
number=FALSE,cex=1,nex=1,col=NULL,lty=NULL,
pch=NULL,xlim=NULL,ylim=NULL,xlab='x',ylab='y',
showrect=FALSE)
plot(delresult,add=FALSE,wlines=c("tess"),
Stellenbosch University http://scholar.sun.ac.za
H-$F4
wpoints=c("none"),
number=FALSE,cex=1,nex=1,col=NULL,lty=1,
pch=NULL,xlim=NULL,ylim=NULL,xlab='x',ylab='y',
showrect=FALSE)
head(summ8merge)
summ8merge$conc <- paste(summ8merge$ID,summ8merge$newID, sep = '.')
str(summ8merge)
plot(summ8merge$dir.area, summ8merge$DbhIncrement)
setwd("F:/B Fury/Tweefontein/hegyi and cvor")
write.csv(summ8merge, "2981.20.42.csv")
Stellenbosch University http://scholar.sun.ac.za