Assessing Tree and Stand Biomass:

A Review with Examples and

Critical Comparisons
Bernard R. Parresol

ABSTRACT. There is considerableinterest today in estimatingthe biomass of trees and forests for
both practicalforestryissues and scientificpurposes.New techniquesand proceduresare brought
togetheralong with the more traditionalapproachesto estimatingwoodybiomass. General model
formsandweightedanalysisare reviewed,alongwithstatisticsfor evaluatingandcomparingbiomass
models.Additivityand harmonizationare addressed,andweight-ratioand density-integralapproaches
are discussed.Subsamplingmethodson trees to deriveunbiasedweightestimates are examined,and
ratioanddifferencesamplingestimatorsare consideredin detail. Errorcomponentsfor standbiomass
estimates are examined.This paper reviewsquantitativeprinciplesand gives specificexamplesfor
predictionof tree biomass.Theexamplesshouldproveusefulfor understandingthe principlesinvolved
and for instructionalpurposes.FOR.Scl.45(4): 573-593.

AdditionalKeyWords:Modelforms,weighting,selectioncriteria,subsampling,
errorcomponents.

the realizationof the shortcomings of traditionalvolume

T HERE
IS
CONsiDERABLE
IN'ITzREST
TODAY
in
estimating
the biomassof forestsfor bothpracticalforestry
issuesand scientificpurposes.
importantforcommercial
Forestbiomassis
uses(e.g.,fuelwoodandfiber)and
measurement, thatis, themyriadlog rulesin use,interestin
anduseof weightfor measurement
rapidlygrown(Guttenberg
andvaluationof treeshas
1973,Huschet al. 1982,Avery
for nationaldevelopment planning,as well as for scientific and Burkhart 1994). The use of end-productunits as a
studiesof ecosystem productivity,
energyandnutrientflows, measure of the amount of raw material is rare outside the
andfor assessing thecontributionof changesin forestlands forestproducts
industry.Rawcottonisnotboughtandsoldin
(especiallytropical)totheglobalcarboncycle.Thus,it isnot "shirt" or other similar units,nor is crudeoil marketedwith
surprisingthat duringthe pastfour decades,researchon liters of gasolineasthe measurement unit. Of course,a shirt
biomass production by forestshassteadilygrownin impor- cannotbeidentifiedoutin a cottonfield,buta veneerlogor
tance(Zeide 1987,WaringandRunning1998).As earlyas a sawlogcanbeidentifiedin a forest.Hencevolumemeasure-
1950weight
I asa measure
of woodquantity
wasused
by ment will continueto be essential.Nonetheless,the current
manyof thelargerCompanies
in NorthAmericaandnorthern trendis towarddecreasing
theusageof end-productunitsas
Europe(Taras1967).With theincreasing
valueof woodand expressionsof stemcontent.The interestin completetree
utilization(roots,stumps,branches,
etc.),theuseof residues
I Theterm"weight"
iscommonly
usedformass,
butstrictly
speaking
this from the manufactureof forestproducts,fuel quantityin
tsincorrect.
Massisthemeasure oftheamount ofmattcrprescnt ina body; relation to forest fire conditions, and other issueshas in-
whereas theweightof a bodyis theforceexertedonitsmassby gravity.
To knowwhethermassOrforceisbeingmeasured, theSI usestwounits: creasedthe use and importanceof biomassmeasurement
thekilogramfor massandthenewtonfor force. (Huschet al. 1982,Philip 1994).

BemardR. Parmsolis MathematicalStatistician,USDAForestService,SouthernReseamhStation,P.O.Box2680, Asheville,NC28802--Phone:

(828) 259-0500; Fax: (828) 257-4840; E-malhbparreso/srsL•fs.fed.us.
Acknowledgments: Theauthorwishesto thank Dr. TimothyGmgoireof Yale University
for his invaluablecommentswhichimprovedthe scopeand
clarityof thiswork.Thanksalsoto Dr. ClarkBaldwinwiththe USDAForestServicein Pineville,LAfor hisinsightful
comments.Gratitudeis extended
tothethreeanonymous reviewers.Thisworkis basedona sedesof lecturesgivenat NanjingForestry University,
Nanjing,People'sRepublic
ofChina.
Mythanksto the President,faculty,andespeciallyDeanandProfessorCaoFuliangforthe opportunity to experience
the beautyandrichcultureof
Chinaand for the chanceto sham my knowledgewith the studentsof the University.

ManuscriptreceivedAugust13, 1998. AcceptedMay 14, 1999. Copyright¸ 1999 bythe Societyof AmedcanForesters

ForestSctence
45(4)1999 573
 A reviewof pastpracticesby Cunia(1988) showedthat
in some instances estimates of biomass content were
obtained by ocular means based on intuition and past
experience.Later, thiswassupplemented by (1) measure-
mentsperformedonsubjectivelyselectedsamplesof trees
or plots and (2) resultsobtainedfrom subjectivelyde-
signedexperiments.Today, forestinventorymethodsare
basedon soundstatisticaldesigns(de Vries 1986). The
bias,if any, is largelyreduced,andthe error of estimates
canbe quantifiedin probabilisticterms.Indeed,research
forestersand statisticianshave come to recognize the
various error componentsof forest biomassinventory
estimatesandto developtechniquesto accountfor them.
Greatprogresshasbeenmadein thelastfew decadesin the
methodologyof selectionof sampletrees and plots and
estimationof forestparametersof interest.New andexcit-
ing developments in samplingtheory,suchas importance
and randomizedbranchsampling,have changedthe way
we view forestinventory(Schreuderet al. 1993). These
modern proceduresof error componentsand sampling
techniqueshaveprovidedconsiderable gainsin reliability
andefficiencyby improvingforecastsandcorresponding
inferencesand by reducing the number of samplesre-
quired and the costsinvolved.
Remotesensing,geographicinformationsystems,and
photogrammetry arepowerfulinterrelatedtoolsfor forest
resourceassessment, asevidencedby the scopeof presen-
tationsat the First InternationalConferenceon Geospatial
.Informationin AgricultureandForestry(Petoskey1998).
Biomassestimationby usingsuchtools is a fascinating
and intricatesubjectin itself and will not be considered
here. Statisticalmethodologies,suchas the expectation-
maximization or EM algorithm and its extensions,mul-
tiple imputation,and Markov chainMonte Carlo (Rubin
1987,Schafer1997),arestartingto beappliedto inventory
data as an alternativeto growth and yield modelsfor
forecasting(Van Deusen1997).Again,theserelatedmeth-
odologiesandtheir usein calculationof biomassconsti-
tutea topicneedingits ownreviewanddevelopment.This
article focuseson modeling and sampling procedures,
because these have been the main avenues of biometrical
researchand developmenton biomass.
The critiquestartswith generalmodelformsandstatis-
tics useful for comparing models. The issue of
heteroscedasticityis addressed,and the theory of esti-
mated generalizedleast squaresis presented.I elaborate
1 Biomass Estimation Techniques
The basicmanagement unit is the foreststand.However,
anystandis anaggregation
of trees,andthestandbiomass is
defined as the sum of the biomass of the individual trees that
comprise thestand.All methods forestimating
standbiomass
mustthereforeinvolve,atleastin theirdevelopmentalstages,
a predictionof individualtreebiomassandthesummation of
thesequantities to obtainper-hectarestandbiomass.
1.1 Regression Modeling
The mostcommonprocedure for estimatingtreebiomass
is throughtheuseof regression. Treesarechosenthroughan
appropriateselectionprocedurefor destructivesampling,
andthe weightsor massof the components of eachtreeare
determined andrelatedby regression to oneormoredimen-
sionsof thestandingtree.Thetreeis normallyseparated into
threeaboveground components: (1) boleor main stem,(2)
bole bark, and (3) crown(branchesandfoliage).Occasion-
ally,a fourthcomponent, belowground biomass,whichisthe
stumpandmajorrootswithina fixeddistance, is considered
SeeKarizumi(1977), LossaintandRapp(1978), Satooand
Sassa(1979), Deanset al. (1996), Kurz et al. (1996), and
Reedet al. (1996) for exampleson samplingandestimation
of belowground biomass.Othertreecomponent schemes are
possibleand are usuallydevisedbasedon the milling and
pulpingtechnologies of theusersfor thepopulationof trees
of interest.The freshweightof an individualtreemay be
determined by weighingall components usingfieldscalesor
by sampling. For largetrees,weighingof theentiretreecan
bequitetimeconsuming andlaborious. Samplingprocedures
as an alternativeto directweighingof an entirecomponent
will be considered later.The processof collectingdataand
developing biomassrelationships falls underthesubjectof
allometry,themeasureandstudyof growthor sizeof a part
in relationto an entireorganism.West et al. (1997) provide
a generaltheoryof allometricscalinglawsbasedon fractal
networksof branchingtubes,andBroad(1998)givesa theory
of multivariateallometry.
1.1.1 General Model Forms
Researchers haveuseda varietyof regression modelsfor
estimatingtotal-treeand tree-component biomass.Earlier
reviewsof biomassstudies(e.g.,Pard61980,Baldwin1987,
Clark 1987, Pelz 1987) indicatethat predictionequations
generallyhavebeendeveloped
utilizingoneof thefollowing
three forms:

on the threegeneralprocedures to handlethe additivity

problemand follow with specificillustrativeexamples. Linear
(additive
error):
Y= Do+ • X•+...+ •j Xj+ • (1)
The next three sections deal with bole biomass and the
techniquesof harmonization,theratio approach,andden-
sity integrals.The next part of the article deals with
Nonlinear
(additive
error):
Y=[5oX•X•...X?+
• (2)
sampling--ratio-typeestimators,randomizedbranchand
importancesampling,and differencesampling.Estima- Nonlinear
(multiplicafive):
Y=]5oX•X2
•...X?• (3)
tion with a ratio estimatorand difference sampling are
demonstrated. The article continues with a section on error
whereY= totalorcomponent
biomass,
)• = treedimension
of inventoryestimatesand concludeswith a look at past variable,
•j = modelparameter,
ande = errorterm.Some
and presentstudiesand generalthoughtson application commonly
usedtreedimension
variables
arediameter
atbreast
and future directions of research. height
(D),D2,total
height
(/-/),D2H,age,
andlivecrown
length

574 ForestSctence
45(4)1999
(LCL).Diameterat thebaseof thelive crownhasbeenproven andPienaar1981,YandleandWiant 1981,Sprugel1983).
tobeoneof thebestpredictor variables
forcrownweight(Clark There is some evidence that these corrections tend to overes-
1982).On thebasisof thepipemodeltheory(Shinozaki et al. timatethe truebias(MadgwickandSatoo1975,Heppand
1964a,1964b),manyresearchers haveusedsapwoodarea Brister1982).Snowdon(1985),workingwithPinusradiata
(activeconducting tissue)measured at variousheightsin the D. Don, showedthat the square-root transformation was a
stemasapredictor of foliageweightandsurface area(e.g.,Snell viablealternativeto thelogarithmictransformation if curvi-
andBrown 1978,RogersandHinckley1979,Kaufmannand linearitybetweentheuntransformed predictorsandbiomass
Troendle1981,Waringet al. 1982,Robichaud andMethven waslow. To correctforbiasunderthesquare-root transform,
1992).An innovativeapproach for predicting seedlingand add6 fromtheregression
to thebiomass
estimate
(Kilkki
saplingbiomass hasusedprojected areaof the seedling or 1979).A listof commonly
usedequationformsforbiomass
sapling(asmeasured bycomputer-based imageanalysis) asan estimationcanbe foundin Clutteret al. (1983, p. 8).
explanatoryvariable.Studieshaveshownthatprojected area
1.1.2 ComparingAlternativeModels
alonecanexplainmorethan97%of thevariation in seedlingor
Schlaegel(1982)recommends thereportingof a seriesof
saplingmass(SuhandMiles1988,Norgrenet al. 1995).Model
statisticsfor evaluatinggoodness-of-fit
andfor usein com-
(1) producesmultiplelinearregressions thatcanbe fittedby
paring
alternative
biomass
models.
Thefirst,anR2statistic,
standardleastsquares estimation procedures.Model (2) pro-
is calledthe fit index (F/). Kv•lseth (1985) examinedeight
duces nonlinear
regression equations
thatrequire useofiterative
proceduresfor parameter estimation.
alternative
R2statistics;
FI corresponds
tohisR•
2, which
is
theoneherecommended. Modelpredictions, if notalreadyin
Normally,biomass dataexhibitheteroscedasticity; thatis,
the error variance is not constant over all observations. If
originalunits,are transformed back to the originalunits,
correctingfor any biasif needed.The total sumof squares
Models (1) and (2) are fitted to suchdata, then weighted
(TSS)andtheresidualsumof squares (RSS)arecalculated as
analysis,
typicallyinvolvingadditional
parameters,isneces-
saryto achieveminimumvarianceparameterestimates(as-
sumingall other regressionassumptions are met: e.g., rss= (r,- )2,Rss=
i=1 i=1
-
uncorrelatederrors).A statistical
modelconsists
jointlyof a
partthatspecifiesthemeanX'• anda partdescribingvaria-
whereF = arithmetic
meanof Y(totalorcomponent
biom-
tionaroundthemean,andthelattermaywell needmorethan
ass)andn = numberof sampleobservations.
The fit indexis
oneparameter
(c•2) to beadequate.
A weighted
analysis
procedure,basedon modelingthe error structure,will be FI = 1- (RSS/ TSS) (5)
describedshortly.
The second statistic is the standard error of estimate in actual
Model (3) nonlinearregressionequationsare usually
transformedintolinear(additiveerror)regression
equations units(Se).It iscalculated
as
by takingthelogarithmof bothsidesof theequation. In this
form,theequationparameterscaneasilybeestimated byleast S,=JRSS
/ (n-p) (6)
squaresprocedures.Typically,the varianceof Y is not uni- wherep = numberof modelparameters. The thirdstatistic,
form
across
thedomain
ofoneormore
ofthe3•'s;however, usefulfor makingquickcomparisons betweenmodels,is the
when transformedto logarithms,Model (3) generallyhas coefficientof variation(CV) expressed
asa percent:
homoscedasticvariance.The logarithmicformis
CV= (S•/ F) x 100 (7)
lnY=lnf•o+f• lnXl+...+f•jlnXj+lne (4a)
The fourthstatisticthatSchlaegelrecommends is onepro-
whereIn is thenaturallogarithm.All commongoodness-of- posedby Fumival(1961)basedon normallikelihoodfunc-
fit statistics
relateto the transformedequationonly andare tions.The generalformulafor Fumival'sindex(/) is
not directlycomparable with the samestatistics
produced
throughuseof eitherModels(1) or (2). Whenthelogarithmic ] = [f,(y)]-IXRMSE (8)
transformation is used,it is usuallydesirableto express
estimatedvaluesof Y in arithmetic(i.e., untransformed) wheref'(Y) is thederivativeof thedependent variablewith
umts.However,the conversionof the unbiasedlogarithmic respect
tobiomass, thebracketssignifythegeometric
mean,and
estimates of the mean and variance to arithmetic units is not RMSEistherootmeansquare errorof thefittedequation.
The
d•rect.The antilogarithmof In Y yields the medianof the index reducesto the usual estimate of the standarderror about the
skewed
arithmetic
distribution
ratherthanthemean.If tl = • curvewhenthedependent variableisbiomass.Whenthedepen-
and62= sample
variance
ofthelogarithmic
equation,
then dentvariableis somefunctionof biomass,the indexmay be
regardedasan averagestandard errortransformed to unitsof
biomass. ThewayFumivalderivedtheindexputsit in inverse
P-'exp(+2/2) orderascompared to likelihood,
thatis,a largevalueindicates
(4b)
• a2_-exp(262
IJ + 2•)- exp(•2+ 2•) apoorfit andvice-versa.Thefifthstatistic,
suggestedbyMeyer
(1938)andrecommended by Schlaegel,isthepercentstandard
^2
where
• istheestimated
value
inarithmetic
unitsandc•ais error[S(%)].Knowledge canbe obtained aboutthemodelby
theestimatedvarianceof Y in arithmeticunits(Fiewelling calculating
the/th
residual's
size
relative
to•, allvalues
being
ForestSctence
45(4)1999 575
in actualbiomass
units.Foreachresidual,
thepercent
standard [weightedModel (1)] and weightednonlinear[weighted
erroris
S(%)i
=[IY/-•//I/•Ix100.
This
statistic
indicates
the Model (2)] equationswere acceptablealternativesto the
sizeoferror
asapercentofthe-meanofthedistribution
ofYi'The transformedallometricmodel.JacobsandMonteith(1981)
expectted
valueof S(%)i = 0 because theexpectedvalueof obtained similarresults.Themaximumlikelihoodapproach,
Y/- Y/=0.Thus,if allS(%)/sarenearly0,theequation
isvery orFurnival'sindex,reflectsnotonlythemagnitudeof residu-
precise.
Naturally,
theS(%)/susually fluctuate
widely.For alsbutalsopossible departures fromassumptions of normal-
reportingpurposes,
I recommend takingall residuals
intoac- ity andhomogeneityof variance.Thesefindingsleadto two
countto forma compositestatistic,
themeanpercentstandard conclusions: (1) Furnival'sindex can generallybe recom-
error
(•(%)) ofpredictions,
defined
as mendedasoneof themostusefulstatistics forevaluatingand
comparing biomass models,and(2) weightedregressionsare
importantandoftennecessary for developing
biomassmod-
(9) els of high precision.
i=l
1.1.3 WeightingBiomassModels
Thesixthstatistic
is thepercent
error(Pe)'It is a precision Forestmodelersare typicallyfacedwith multiplicat•ve
indexusingthepercentstandard
errorandthechi-square
test. heteroscedasticity
in theirdata(Parresol1993).It is oftenthe
Let Pe represent
therelativedifference
in percent
of the casethat the errorvariance(or disturbance)is functionally
estimateof treeor component
weightto its truevalue.This relatedto predictorvariablesin regression.Harvey (1976)
statistic
computes
thevalueofPethatwouldbenecessary to andJudgeet al. (1988) haveshownthatif theerrorvariance
assureanonsignificant
g2test.Thepercent
errorisdefined
as is a functionof a smallnumberof unknownparameters, and
if theseparameters canbeconsistently estimated,
thenesu-
r- 2 n ^ 2'nl/2 mated generalizedleast squares(EGLS) estimationwfil
/o96)x,(r provideasymptotically efficientestimates
of themodelpa-
Vek t-v i=1 X. '/ / (10) rmeters.

In thegenerallinearstatisticalmodely= XI• + •, X is
where
thetx= 0.05valueforg2withvdegrees
offreedom
is a (T x K)observablenonstochastic matrix,I• is a (K x 1)
approximated
by vector of parametersto be estimated,y is a (T x 1)
observablerandom vector, and the error vector, •, is a (T
g•v)
=0.853
+v+ 1.645(2v
- 1)1/2 x 1) unobservable
randomvectorwith propertiesE[• ] = 0
andE[•]=•=o27, where• is a (T x T) diagonal
For derivationof thisstatisticseeSchlaegel(1982). Finally, matrix. Heteroscedasticity
existswhen the diagonalele-
Schlaegel advocates
reportingthenecessary
information for ments of • are not all identical. In the general
the constructionof predictionconfidenceintervals.This heteroscedastic
specification •=diag(o•2,o22
.....0}). If
usually involvesreportingthe model mean squareerror we assumethat each o t2 isanexponential
function of P
(MSE)andthesums
ofsquares
andcr•s products
matrix,i.e., explanatoryvariablesthen
(X'X)-• ormoregenerally
theCov(l•). 2 2
To summarize, statisticsuseful for model evaluation and E[•t ] = ot = exp[z;ot] t = 1,2,...,r (11)
comparisonare: (1) fit index (F/), (2) standarderror of
estimate in biomass of variation whereZ[= (ZtlZt2...Ztp)is
units(Se),(3) coefficient a (1 XP) vectorcontaining
thetth
observationonP nonstochastic explanatoryvariablesand
based onSe(CV)_, (4)Furnival's
index(/), (5)meanpercent
standard error( S(%)),(6)percenterror(Pe)oftheresiduals, ot=(oqo•2...O•p)'is
a (P x 1) vectorof unknowncoeffi-
and (7) informationneededfor buildingpredictionconfi- cients.Thefirstelementinztistakenasunity( ztl m1),and
denceintervals.Anotherusefulmodelselectionprocedure-- the other z's could be identical to, or functions of, the x's.
prevalentin the statisticsliterature--istheAkaike Informa- The normalconventionis to parameterizethe scalefactor
tion Criterion(AIC). For a descriptionanddiscussion of the 02asexp(t•l),
ort•1= In02.Thismeans
theexpression
in
AIC, seeJudgeet al. (1988, p. 848). (11) can be written as
A numberof researchers have publishedaccountsof
comparisons of alternative biomass regression
models.Crow
Ot2=o exp[zt
ot] (12)
(1971) usedF1 as a meanswith whichto comparemodels.
Althoughthe transformed allometricequation[Model (4)] where
zt*' =(zt2...Ztp)
andor*----
(a2...ap)'.
Thecovariance
provedsuperior, Model(1) wasfoundto bealmostasreliable matrix can now be written as
whentherewasarelativelysmallrangeintreesizes.Schreuder
andSwank(1971, 1976)usedF1and1tocomparea weighted
linearmodelwith six othermodelsbasedon the family of
powertransformations definedby BoxandCox (1964).They [exp(z•'lx*)
foundthattheF1criterioncouldgivemisleading results,but exp(z2ot)
that Furnival's index was a usefultool in comparisonof
modelsfor estimating

576 Forest
Sctence
biomass.

45(4)1999
CrowandLaidly(1980)also
usedthe likelihoodapproachto showthat weightedlinear 02?
=02[ exp(zr ot
(13)
 In ordertoestimate
• wefirsttakelogarithms
of Equation Notethatthenumerator istheregression (orexplained) sum
(11) to obtain of squaresobtainedwhenestimating o[ andthatthistestis
asymptotically equivalentto the F testfor testingthat all
In•2 =z;• (14) coefficients,excepttheintercept,are0.
2 Gregoireand Dyer (1989) and Williams and Gregoire
Sincetheot arenotknown,weuseinstead thesquares ofthe (1993) advocatetheuseof maximumlikelihood(ML) with a
ordinary leastsquares (OLS)residuals.
2 Theseresiduals
specifiederror structurefor fitting weightedregressions.
(denoted
2.'
e,) are
likely
to reflect
the size
2.
of (h2,that
ß
is,large
2 CarrollandRuppert(1988) discuss theincreased efficiency
when0 t mlargeandsmallwhenot •ssmall.AddingIn et
of maximumlikelihood(undernormality)overgeneralized
to bothsidesof Equation(14) yields
leastsquares,
withincreases
ofabout8%beingcommon.'The
lne•2
+in• = zto[
ß + Ine•2 ML procedure requiressolvingfor bothfirst and second
partialderivativesandresultsin a simultaneoussystemof
or nonlinearequations.In contrast,
theEGLSestimator issimple
anddirect,requiresnospecialsoftwareto implement,andis
lne,= z;a+ (15) almostas efficientasML. If iterated,the EGLS procedure
converges to theML estimates undernormality.
wherevt = Inet2- In02
t = ln(et2
/ or2).In matrix
notation,
Model (15) canbe writtenasq = Z•x + v wherethe vectorq 1.1.4 BiomassAdditivity
= (lnq
2lne22...ine2r)
'. One
way
toesfimate•xis
toapply
OLS A desirablefeatureof treecomponent regression
equa-
toModel
(15)which
yields
(I = (Z'Z)-•Z'q.Harvey
(1976) tionsis thatthepredictions for thecomponents sumto the
showed thatif the•t's arenormally distributed
thenthe predictionfor the total tree. Kozak (1970), Chiyendaand
intercept
•xI will notbeconsistentlyestimated,
butthere- Kozak (1984), and Cunia and Briggs(1984, 1985a)have
mainingelements in & will beconsistent
or unbiased. discussed theproblemof forcingadditivityon a setof tree
Substituting
&* for•x*inexpression (13),weobtainthe biomassfunctions.The meansto forcingadditivitycan be
estimated
covariance matrix• = •2•. TheEGLSestimator groupedinto threedifferentprocedures dependingon how
is formed as theindividualcomponents areaggregated.
In procedure1, thetotalbiomasssampleregression func-
• =(X'•-•X)-•X'•-•y
=(X"•-•X)-•X"•-•y(16) tionis definedasthesumof theindividuallycalculated best
regression functionsof thebiomass of itsk components:
Fortunately,
• only
dependsontheconsistently
estimated
elements
of &,since
&•.canbefactored
o•ut
asaproportion-
ahtyconstant.
Thecovarianee
matrixof I• is 91= f(x0

(17) (20)
where
82=(y- X•)'•s-•(y
_X•)/ T- K)
Theusualhypothesis
testsandintervalestimates
arebasedon =h
thismatrix.
Forprediction
intervals
onsome
future
valueY0
thesampling
erroris estimated
by Reliability(i.e., confidenceintervals)of the total biomass
predictioncan be determinedfrom variancepropertiesof
linear combinations:
b2(bo+x6X"k xo
.... (18) k
where
•0isthescaler
exp(z0
•x) (21)
i=1 i<j
To test the hypothesisof homoscedastic
errorsversus
heteroscedasfic
errors
yousimply testHoaX*= 0 against where
Hl:•t*;•0.LetRbethematrix
( Z'Z)-l withitsfirstrowand
fn'stcolumn
removed.
If the• t'sarenormally
distributed
then&*- N[•x*,4.9348R]
(Harvey
1976)
andthefollowing COV(•i,
h)'=PYiYj
4Var(•i)Var(h
)
staUstic
(Judge
etal. 1988,p. 370),based
onthedistribution py•y•
=correlation
between
Y/and
Y•
of quadraticformsin normalvariables,teststhe abovenull
hypothesis: In procedure2, theadditivityof thecomponents
isensured
by using the same independentvariables(and the same

4.9348
_ (19) weightfunction)in the(weighted)
sionsof thebiomass
leastsquares
of eachcomponent
linearregres-
andthatof thetotal.
Underthismethod,onecancomputetheregression coeffi-
cientsof thetotalequation
simplybysummingtheregression
2 Some
statisticians,
such
asCarroll
andRupert
(1988,
p.79-82),suggest
thatbetterperformance
canbe obtained
usingabsolute
residuals
over coefficientsof the(assumedindependent)
component equa-
squaredresiduals. tions(thebi vectors),
thatis,

ForestSctence
45(4)1999 577
 permitsthe veryreal possibilityof multicollinearity.Thts
cancauseunstableparameterestimatesandinflated stan-
darderrors.In fact,applyingjoint-generalized leastsquares
to the setof equationsin (22) is of no benefitbecausethe
(22)
=x'b covariancesbetweenthe equationsget concentratedout
when each equationhas identical explanatoryvariables
9,o = x'[bl+b2
(SrivastavaandGiles 1987).Thus,it is asif theequations
areindependent, andthesameresultsareobtainedaswhen
Thisresultholdsonlyundertherestrictiveassumption that applying least squaresto each equationseparately.If
thek components Yi (i = 1..... k) areindependent,
which disturbances or errorsin the differentequationsarecorre-
impliesthattheßi (i = 1..... k)areuncorrelated.
Regression lated (contemporaneouscorrelation), then procedure 1
statistics
andreliabilityof estimatescanbecomputedforthe [formulationin (20)] is inferior to procedure3 [formula-
totalequation(seeChiyendaandKozak 1984).Underinde- tion in (23)] becauseSUR takesinto accountthe contem-
pendence, thevariance of •to• is simplythesumof the poraneouscorrelationsandresultsin lower variance.With
variances of the •i's, the covariance termsdropout of the ready availability of econometricsoftware,suchas
Equation(21), thus
SAS/ETS© (SAS InstituteInc., SAS CampusDrive, Cary,
k
NC 27513), complicatedstatisticalprocedureslike SUR
caneasilybeimplemented.A comprehensive referenceon
Var(•tøtal)
=Z Var(•i)'
i=1
SUR is Srivastava and Giles (1987).

Procedure 2 allowsnoflexibilityfor usingdifferentcompo- 1.1.5 Example

nentequationforms.ChiyendaandKozak(1984),however,
generalizedprocedure2 usingrestrictedleast squaresto
allow for differentequationforms.
Procedure3 is the mostgeneraland flexible methodand
themostdifficultto employ.Statisticaldependenciesamong
sampledataareaccounted forusinggeneralized leastsquares
regression with dummyvariablestechniques to calculatea
set of regression functionssuchthat: (1) eachcomponent
regressioncontainsits own independentvariables,and the
total-treeregressionisa functionof all independent variables
used;(2)eachregression canuseitsownweightfunction;and
(3) theadditivityisensured by settingconstraints(i.e.,linear
restrictions)on the regression coefficients.The Cuniaand
Briggs(1984, 1985a)procedure3 is the sameasusingjoint-
generalizedleastsquares, alsocalled"seeminglyunrelated
regressions" (SUR), for a setof contemporaneously corre-
latedlinearstatistical
modelswithcross-equation constraints.
The structuralequationsfor thesystemof modelsof biomass
additivitycanbe specifiedas
At thisjuncture,an exampleis in orderto demonstrate
equationselection,weightedanalysis,equationadditivtty,
and goodness-of-fit statistics.Considerthe sampleof 39
willow oak (QuercusphellosL.) treesin Table 1. Treesfor
destructive samplingwereselected from 10 naturalbottom-
land hardwoodstandsin Mississippi.Trees were felled,
separatedintocomponents ofbolewood,bolebark,andcrown,
andweighedin thefield. The 39 treesgivenin Table 1 area
subsetof a largerdatasetfroma biomassstudyby Schlaegel
(1981),usedhereforillustrativepurposes.Scatterplots
of the
data,a stepwiseregression procedure,andresidualanalyses
wereusedto selectthefollowingindividually"best"biomass
componentequations:
•wood
= bo+ blD2H
•bark
= bo'l-b102 (24)
D 2H x LCL
crown
=bO+bl lOOO
4-b2H

Yl '=fl(Xl )+ •l For total treebiomass,the bestindividualequationwas

y2 = f2(x2)+ •
•total
mboq-blD2t-I (25)
(23)
y• = ft (x•)+ • Scatterplots
oftheresiduals
overD2HforYwood
andYtotal
Ytotal
mftotal($1,$2
..... Sk)'t-ß'total (Figure
1A)andoverD2 forYbark
revealed
similar
fan
patterns of increasing error variance. This type of
heteroscedasticity
is commonand is usuallymodeledas a
andredundant
columns
inftotalareeliminated.
Whenthe power
function,
thatis,{•t
2= (j2xtk,
where
X isD2HorD2.
stochasticpropertiesof the error vectors are specified,
Hence,thefollowingvariancefunctionwasfittedto theOLS
alongwith the linearrestrictions,the structuralequations residualsfrom the bolewood, bolebark, and total biomass
becomea statisticalmodel for efficient parameteresti-
regressions:
matesandreliable predictionintervals.The procedure3,
or SUR, methodis preferableto procedures1 and 2 for
severalreasons.Procedure2 requiresthe assumptionof e2----
exp[a
1+ a2InX]
independence amongcomponents onthe sametree,which or

is unrealistic.Anotherconsiderationagainstprocedure2
is thatloadingthesamepredictorvariablesin all equations Ine2= a1+ a2InX (26)
578 ForestSctence
45(4)1999
 Table 1. Greenweight data for willow oak treesfrom the state of Mississippi,USA.
Greenweight
Tree Dbh(cm) Height LCL Age(yr) Wood Bark Crown Tree
.................(m)...................................................... (kg).......................................
I 73.2 29.0 16.2 93 4,463.4 572.9 186.4 5,222.7
2 30.5 18.3 10.4 40 550.7 83.9 44.9 679.5
3 48.3 22.9 11.9 69 1,689.2 225.0 93.4 2,007.6
4 69.6 27.4 18.3 74 3,441.5 435.5 178.3 4,055.3
5 28.7 19.8 11.6 38 482.2 75.3 45.4 602.9
6 53.1 32.0 14.9 78 2,281.6 307.5 57.2 2,646.3
7 45.7 32.0 17.7 79 1,771.3 230.4 24.9 2,026.6
8 46.5 30.5 17.4 83 1,611.6 228.2 17.2 1,857.0
9 33.8 25.9 10.1 68 861.8 122.5 20.0 1,004.3
10 55.4 30.5 13.1 70 2,952.9 367.4 65.3 3,385.6
11 30.5 24.4 9.1 75 679.9 125.2 17.7 822.8
12 70.1 30.5 13.4 81 3,867.8 546.1 79.4 4,493.3
13 41.9 24.4 12.2 79 1,289.1 185.5 36.7 1,511.3
14 42.9 25.9 13.7 76 1,495.5 201.4 71.7 1,768.6
15 66.3 35.1 17.7 81 4,091.0 413.2 99.3 4,603.5
16 40.6 25.9 11.6 64 1,264.2 175.5 12.7 1,452.4
17 28.7 21.3 9.8 75 485.8 67.6 19.1 572.5
18 80.5 32.0 17.4 93 5,782.0 657.7 186.9 6,626.6
19 66.5 33.8 16.8 106 4,085.1 524.4 112.0 4,721.5
20 61.2 29.0 17.1 71 2,621.4 225.4 82.6 2,929.4
21 21.8 25.9 11.3 35 292.6 51.7 5.0 349.3
22 29.0 24.4 8.8 35 616.4 94.3 21.8 732.5
23 49.8 27.4 16.2 35 1,757.2 220.0 63.5 2,040.7
24 34.3 27.4 10.4 37 902.2 144.7 21.8 1,068.7
25 43.2 27.4 15.5 41 1,251.5 212.3 20.9 1,484.7
26 25.1 24.4 8.8 38 437.7 63.5 6.4 507.6
27 29.7 25.9 13.7 40 704.4 89.4 10.0 803.8
28 34.8 25.9 13.1 40 906.7 117.5 23.1 1,047.3
29 38.6 27.4 13.4 41 1,309.5 148.8 26.8 1,485.1
30 49.8 22.9 14.3 40 1,497.3 160.1 49.0 1,706.4
31 36.1 25.9 16.5 34 794.3 116.6 20.9 931.8
32 37.1 21.3 11.0 67 846.9 120.7 26.8 994.4
33 33.0 21.3 9.8 67 635.5 89.8 27.7 753.0
34 57.7 25.9 18.6 84 2,545.1 371.0 70.3 2,986.4
35 53.8 25.9 12.2 87 2,275.7 359.3 32.7 2,667.7
36 57.9 27.4 10.1 89 2,822.3 379.2 61.2 3,262.7
37 75.4 27.4 13.7 91 3,782.1 579.2 61.7 4,423.0
38 57.2 25.9 12.2 86 2,055.7 362.0 45.4 2,463.1
39 69.1 27.4 14.6 87 3,618.4 498.1 87.1 4,203.6
No•: DBH is diameter breast height and LCLis live crown length.

For thecrownmodel,variancewasassumed
proportionalto standarderror is around 8% to 11% for the wood, bark, and
a powerof D2Hx (LCL/1000)
based
on a fanpattern
of totaltreeregressions butover32% for thecrownregression;
increasingvariance.With increasingtree height,however, and the fit index is lowestfor the crownregression,which
varianceappearedto expandthen decrease(Figure IB), alsohasthehighestcoefficientof variationandpercenterror.
suggesting
anegativequadratic
trend
orc•t
2= c•2exp[-kHt
2]. All in all thisshows(not surprisingly!)thatcrownbiomass
Combining these two heteroscedastictrends into one hasgreatervariabilitythanwood,bark,or totalbiomass.
mutiplicativeerrormodelresultsin Underprocedure1 for additivity,total tree biomassis
simplythe sumof the components. For example,usingthe
coefficients
inTable2 andthesetofequations
in(24• atree
Ine2=aI+a2InD2HxLCLa3H2 (27)
I000 withD = 30•cm,H = 18m,and•LCL = 10m,gives:Ywooa=
468.2kg, Yb•rk
= 91.8 kg, and Ycrown
= 43.4 kg; therefore,
Table 2 gives the Coefficients,weightfunctions,and
heteroscedasticity
tests[Equation
(19)] fromtheEGLSfit of •'total
=468.2
+ 91.8+ 43.4= 603.4
kg
the threewillow oak componentbiomassfunctionsandthe
total tree function.As readily seenin Table 2, all the Thesampling
errorforeachcompone•ntp•rediction
iscom-
heteroscedasticitytestsare significant,indicatingthe need
pu«ed^using
Equation (18),•giv•ing:
Var(Ywood
) = 2028.89,
Var(Ybark
) = 221.19,andVar(Yc•own)
= 139.45.Thecorrela-
for modelingthe error structure.The statistics[Equations tionsbetweenthe (weighted)biomasscomponents
are:
(5)-(10)]recommended by Schlaegel(1982) are shownin
Table3 foreachof thefourequations.
3 Themeanpercent
•yw•y•t= 0.26,byw•ay,xow
n= 0.31,
and•y•,yc,o•
= 0.14;
3 ASASprogram
isavailable
fromtheauthor
forcomputing
these
statistics. therefore,usingEquation(21) we obtain

Forest
Science
45(4)1999 579
 For an approximate95% predictionlimit, we will use
A +2x]'•16.84giving
603.4 kg + 111.7kg (28)
loo
From Equation(25), the bestindividualequationfor tree
biomass,we obtain an alternate value of
-loo

-8oo tal= 557.6kgwithVat ( Ytota]

) = 2644.31;
hencean approximate
95% predictioninterval
•is
' I ' I ' I ' I ' I ' I ' I ' I * I ' I '
557.6 kg + 102.8kg (29)
o

In considering
thepredictionintervalsin (28) and(29), one
D•H/1000 canseethatthepricefor additivityusingprocedure1 is an
expandedinterval(_+111.7vs._+102.8),indicatinga lossof
efficiency.
40
Supposewe wish to considera set of linear models
wherebywe allow statisticaldependenceamongcompo-
20 nentsand the total tree biomass.A setor systemof linear
models whose parametersare estimatedby SUR with
0 .
linear restrictions should result in efficient estimates and
additive predictions.Reasonableequationsand variance
functionsfor the willow oak sampledata are:

•wood
=blo+blID2H;
18
82= (D2H)1.95
Height (m)
Figure1. Scatterplotsof ordinaryleast squaresresidualsfrom (A)
total tree willow oak biomassregressionshowingfan pattern of •bark
=b20
+b21D2H;
increasing
variance,and(B)willowoakcrownbiomassregression
showing negative quadratic trend in variance. 82= (D2H)I.745

A ^

Var(Yto•)= 2028.89+ 221.19+ 139.45

+2-0.26. •/2028.89.221.19 •crown
=b3o
+b3]
D2H XLCL
1000
+b32H; (30)

+2.0.31.42028.89.139.45
+2.0.14. x/221.19-139.45
•2=[D2HXLCLI
L lOOO J
'.(•Xexp[_0.00406.2
]
=3116.84

Prediction confidence intervals are constructed as

D2H x LCL
•to•=b40
+b4]
D2H+b42 1000
•cb43H;
•+t(a/2
)V•(•). •2 = (D2H)i.s44
Table 2. Coefficients,weight functions,and hateroscedasticitytests from best individualcomponant and total tree
regressionsfor willow oak biomasssample data (n = 39).
Model* b0 b• b2 Weightfunction %2 p
Wood 25.749477 0.027310 (DZH)"9•ø 18.1 <0.0001
Bark -0.515317 0.102529 (D•)•'8•4 12.7 0.0004
Crown 117.195175 0.057502 -4.616870 (D•Hx LCId1,000)
Le'• 9.4 0.0091
x exp[-0.00406H
2]
Total 46.380555 0.031558 (D•/-/)
2'ø•4 20.7 <0.0001
* Model forms:

Ywoo•
= [•o+• D2H+•,Yba•k = [•o+[5•
D2+•,
Ycmwn
=•0 +[51D2 H x LCLI1000+[52 H+•,
Y•m•= •0 +• D2H+•
where D is dia•ter breast height, H is t• height, and LCLis live crown length.

580 ForestSctence
45(4)1999
 Table3. Goodnees-of-fitstatisticsfor the individuallybestwillow oakcomponentandtotal tree biomassequations.

Model FI $e CV I • (%) Pc
Wood 0.98 182.32 9.50 134.83 7.47 15.76
Bark 0.94 41.01 16.08 30.48 11.00 31.19
Crown 0.81 21.15 38.67 15.33 32.15 80.32
Total 0.98 217.37 9.76 279.40 7.80 16.34

NOT•: FIisfit index,Soisstandard errorofestimate

in actualbiomassunits,CViscoefficient
ofvariation
expressed
fromactual
biomass units,I isFurnival's
index,•(%) ismeanpercent standard
errorofpredictions,
andPeispercent
error.Seetextfor
definitions.

where b40= b10+ b20+ b30,b41= b11+ b21,b•42

= b31,and 1 ^ A
b43= b32.Forsystempaxsimony,I altered
the Yb•rt
equation
from that used in (24). Note that a separatevariance
8ij=(r_Ki)m(r_Kj)V2
e[A;Ajej
(32)
function is specifiedfor each equationin the set. The
coefficientsfor the variancefunctionswere determinedby where
thedegrees-of-freedom
corrections
Ki andKjare
regressingon the OLS residuals[Model (15)] from the
thenumber
ofcoefficients
perequation.
If wedefine
• asthe
four equations.A brief explanationof fitting theseequa- matrix
containing
theestimates
•ij from(32),thenthe
unrestricted
SUR estimatorfor • canbe writtenas
transby SUR follows.
Thesystemof fourequations in (30) canbewrittenin the
usualmatrixalgebranotationas •= [X'•'(• -I ©I)•XI-iX'•'(• -• ©I)•y (33)
Yl = J•'l•l+ E1 where © denotesthe Kronecker or direct product. The
Y2= X2 + % restricted SUR estimator is obtained by minimizing
(• y- • X [})'(•-I©I)(• y- • X [•)subject
tothelinear
restrictionsR[} = r. It is given by
Y4= X4 + E
Combiningall equationsinto onebig modelyields
•* = •. •'R'(R•'R')-t(r
- R•) (34)

where

• = [X'•i'(• -I • l)•X] -•
Y2= 0 X2 0 0 •2+ E2
o E3 andfiisfrom
(33).
Thelinear
restrictions
b40
=blo+b20
+
(31)
Y4 0 0 0 X4 [•4 E4 b30, b41= bll + b21
, b42= b31
, andb43= b32, canbewritten
(4TxI) (4TxI 1) (11x1) (4TxI) alternatively
asblo+ b20+ b30- b40= 0, bll + b21- b41= 0,
b31- b42= 0, andb32- b43= 0. Writingthese restrictions
in
or alternatively theformatR[}= r yields

y = f([5) = X[5+E
where T is numberof observations(39 for this data). The
matrixof weightscanbe writtenas

0 0 •2• [03

,e,
ooo]
0%0 o
'I'=o o % o
io1o1oo-1
ooo_l!1
1 o 1 o o o o -1
o o o o 1 o o o
o o o %
o o o o o 1 o o o
(4Tx4T)

where
•I•iisdefined
asinEquation
(13),
andletA=
Theimplicit.assumptions
forModel(31) are: EIEi]= 0 and
EIAiEi•j A•] = o•iI;i,j = 1,2,3,4,andI isardimensional LIhJ
identitymatrix.
The variancesand covariancesof Model (31) are un- The covariancematrixof therestrictedparameterestimates
known
andmust
beestimated.
Toestimate
theoij,wefirst is calculated as follows
estLmate
eachequation
byEGLS[Equation
(16)]andobtain
theresiduals
ei = Yi - Xi bi. Consistent
estimates
of the (35)
variancesandcovariances
arethengivenby

ForestScience
45(4)1999 581
Onecanconstructthebiomass tablesandtheassociated
(1 -
confidenceintervalsforthepredictedmeanvalueandfor a
S]=279.895,
82=0.832,
predictedvalue of an individual(new) outcomeby the 844=3.99x10'-5,and
formulas:
•/= 57855638,

for anapproximate
95% prediction
interval(t = 2) of
(36a)
= biomass
estimate
fromith systemequation 583.8 kg + 93.8 kg (37)
TheSUR
prediction
interval
on•otal
in(37)js
narrower
mean
value
confidence
limits
(36b) thantheleastsquares
prediction
intervalon Ytomin Q29)
Onemightexpecttheindividuallybestregression
on Ytom
to have the smallestvariance, becauseit is the best estima-

• +t(•,•)•S]
+828//•:confidence
limits tor thatis a linearunbiased
functionof Ytotal'
However,
(36c) becauseof theexistenceof contemporaneous correlations,
for anindividualprediction it is possibleto obtaina betterlinear unbiasedestimator
thatisa function
ofYwood,,
Ybark,
Ycrown'
andYtotal
ßThus,
wherexiis
the
vector
forthe
ithsystem
equation,
S•=x[•.xi evenundertheconstraintof additivity,theSUR estimator
=variance
of•, 82istheestimated
variance
ofModel
(31) can achieve lower variance and be a more efficient estima-

(i.e.,e'•'(cX-1
©l)•,e/ dferror
), and8ii• istheestimate
of tor. Clearly,procedure3, theSUR estimator,is themethod
theconditional
variance
oftheithsystem
equation
[ 8ii isthe of choicefor additivity.
i, ithelement
of•; reEquation
(32)and• istheestimated 1.1.6 Bole Biomass and Harmonization
weight].Table4 givescoefficientsandtheir standarderrors
Often the biomassof chief interestis just the tree bole,
fromtheunweighted(each•/-- 1) restrictedSUR fit (RSUR)
particularlyfor dryweightyield.Whatisfrequentlyneeded
andtheweighted
restricted
SURfit (WRSUR).
4 Notehow is a meansto predictbolebiomassfor differentmerchant-
weightingreducesthemajorityof standarderrors,dramati-
abilitylimits.For example,bolebiomasscomponents can
cally for some(threeof the coefficientshadstandarderrors
be defined in the following nested fashion. The first
reducedin excessof 50%), andhenceachievesmoreefficient
component,the entire bole, containsin its entirety the
parameterestimates.
secondcomponent, theboletoa 10cmtopdiameter,which
Usingthecoefficientsin Table4, ifD = 30 cm,H= 18 m,
in turncontainstheentirethirdcomponent, thetreeboleup
LCL = 10 m, andi = 4 (totalbiomass),we havefrom (36a):
to a 15 cm top, and so on. When calculatinga separate
X•=[00000001 16200 162 18] regression functionfor eachcomponent, theproblemthat
usuallyarisesis that the regressionlines may crosseach
and other; consequently,the estimate of the biomass of a
nestedcomponentmay exceedthat of the next larger
•=fi•0+16200•;1
+162fi•2
+18fi;3
=583.8
kg component.The processof forcingseveralsimultaneous
regressions to behavelogicallywith respectto eachother
Thepredictionlimitsforthispointestimatecanbecalculated
is known as harmonization.Jacobsand Cunia (1980) and
from (36c).For thisexamplewe have
CuniaandBriggs(1985b)solvedtheintersection problem
by (1) usingthe samemodelform for all components and
(2) makingall regressions parallelbyrestrictingtheslopes
to be identical.Further,they controlledthe spacingbe-
4 ASAS/IML
andaGAUSS
matrix
language
program
areavailable
fromthe tweenconsecutive regressions to follow a reasonablepat-
authorforfittingModel(31). tern.They reasonedthatthe differencebetweenthe inter-

Table 4. Resultsfrom f•ting the willow oak data with seeminglyunrelated regressions.
RSUR WRSUR Reduction in
Estimate SE Estimate SE SE' (%)
tiao 59.186526 48.243531 29.634908 19.354063 60
Bii 0.026598 0.000558 0.027190 0.000587 -5
B20 29.651015 10.383880 17.612412 4.406701 58
B2a 0.003225 0.000120 0.003435 0.000124 -3
B•0 133.106587 26.624556 106.065804 17.471978 34
fi•a 0.061543 0.005081 0.056544 0.004325 15
fin -5.336039 1.114125 -4.153695 0.730026 34
fl• 221.944128 62.183373 153.313124 28.802732 54
fi41 0.029824 0.000623 0.030624 0.000643 -3
B• 0.061543 0.005081 0.056544 0.004325 15
B• -5.336039 1.114125 -4.153695 0.730026 34
No'm: RSUR is restrictedseemingly unrelated regressions,and WRSUR is weighted restrictedseemingly unrelated regressions.
a Computed as [std err(RSUR)- std err(WRSUR)] /std err(RSUR) x 100.

582 ForestSctence
45(4)1999
cepts is a function of both squaredlog (that is, top)
dinmetersandthe lengthof thelog, whichis a functionof
top diameter.Hence,the interceptsof the variousregres-
sions should be related as
•iO= aO4'O•lZ
i 4'O•2Zi
2 (38)
wherezi is thetopdiameter
of theith component.
Cunia
and Briggs (1985b) recognizedthat the harmonizedre-
gressionfunctionswere serially correlatedbecausethe
variouscomponentswere not independent,being mea-
suredon the sametrees. As with the additivity problem,
Cunia and Briggsput forth a procedurethat allowedthe
estimationof thecovariancematrixof the samplebiomass
valuesand circumventedthe problemof storingand in-
verting large covariancematrices.They constructeda
giantsizeregression with dummyvariablesthatcontained
all of the individual componentregressions,and they
estimatedthe parameterswith generalizedleastsquares.
Again, as with the additivityproblem,their procedureis
equivalentto usingjoint-generalizedleast squareswith
cross-equation constraints.With thepropersoftware,such
asthe GAUSSTMmatrix language(AptechSystems,Inc.,
23804S.E.Kent-KangleyRoad,MapleValley,WA 98038),
SUR is easyto implement.
Thoughtheharmonization technique solvestheintersec-
tion problemof nestedcomponentbole regressions and
logicallyspaces theintercepts,it is basedonassumptions of
parallelismandprecisespacingbetweenconsecutive regres-
sions,assumptions which may or may not be true for any
particulartree population.Anotherdrawbackis that each
standard of utilizationrequiresanotherequation to beadded
to thesetof regressions. Two techniques thatdo notrequire
theseassumptions andminimizethenumberof equations are
(1) the weight-ratioapproachand (2) the density-integral
approach.Both approaches providea systemto calculate
total bole biomassand merchantablebole biomassto any
standardof utilizationexpressedas a functionof stump
height,andof topdiameteror sectionheight.
1.1.7 Weight-RatioApproach
To circumvent the equationcross-over problem,Honer
(1964)deviseda two-stepmethodto calculatemerchantable
volumeto any utilizationspecification.
First,he developed
an equationto predicttotal treevolume.Second,he devel-
opedan equationto predictthe proportionor ratioof mer-
chantablevolumeto totalvolumegiventhemerchantability
limits.Whenattentionshiftedto estimatingtreebiomass,it
wasnaturalto applytheratioapproach thusavoidingsimilar
problemsencountered in treevolumeestimation(Williams
1982).
The weight-ratioapproachusesthe followingrelation-
ship:
•, = kl• (39)
wherew is merchantableweight,Wis totalweight,andR is
w/W.Interestedpersonsmayreferto Honer(1964),Burkhart
(1977), andVan Deusenet al. (1981) for the early develop-
mentalwork on the ratio approach.Parresolet al. (1987)
revieweda numberof ratio modelsfrom the forestrylitera-
ture. Parresoland Thomas(1989) comparedthe density-
integralapproachagainstthe weight-ratioapproachand
concludedthatthedensity-integralapproachgavemorepre-
ciseestimatesof sectionalandtotalboleweight.
1.1.8 Density-IntegralApproach
ParresolandThomas(1989) introducedthedensity-inte-
gralmethodology.
The generalized density-integral
model
for stem biomass is
w= H fXup(x)f(x)dx+ •
•X l
(40)
whereH is total treeheight,x is relativeheight,p(x) is a
functiongivingdensityor stemspecificgravityatx,f(x) is an
equationexpressing stemprofilein cross-sectional
areaasa
function
ofx,wisboledrymass ofwoodbetweenlimitsxtand
xu,and• isstochastic
error.Fora specific
biomassmodel,
oneneedsto definep andf.SeeTasissa andBurkhart(1998)
for recentworkon modelingspecificgravityandMaguire
and Batista(1996) for a reviewof tapermodels.For the
derivationof the generalizeddensity-integralmodel and
examplesof its useseethearticlesby ParresolandThomas
(1989, 1996), andThomaset al. (1995).
Onecouldfit stemprofileIf(x)] anddensity[p(x)]inde-
pendently andplacethemintoModel(40) for prediction of
biomass. However,as with the additivityproblemandthe
harmonizationproblem,it is importantto recognizethatthe
datafor stemprofile(i.e., volume),density,andmassarenot
independent,comingfromthesametrees.Onewouldexpect
mass,density,and volumeto be correlatedat the same
measurement bolt on thetree.This contemporaneous corre-
lation, if not accountedfor, leadsto inefficient estimatesof
the parameters. In addition,observedstemmassshouldbe
incorporated intothefittingprocess. Joint-generalized
least
squares or SUR, aspreviouslyoutlined,takesinto account
contemporaneous correlationsand leadsto efficientesti-
mates.ParresolandThomas(1996) showedthatparameter
estimates(•i 'S) fromS[JRestimation of thesimultaneous
equationsfrom the densityintegralhad smallerstandard
errorsthanfromOLS estimationoff(x) andp(x).
1.2 Samplingon the Tree
The processof physicallycollectingbiomassdatacanbe
very laborintensive.In shortrotationwoodybiomasspro-
grams,treesusuallydo not attain large sizes,and field
weighingof the entiretree to measurefreshweightis not
overlydifficult.Thevarioustreecomponents, asdetermined
bythescheme used,canbemeasured directlyassoonasthey
areseparated fromthetree.The onlypossible errormaybe
due to faulty measurement instrumentsor methods.How-
ever,if biomassexpressed as dry weightis required,direct
measurement maybetooexpensive andtimeconsuming for
the largercomponents suchas the bole.The only practical
alternativeis subsampling.Smallsamples are selectedfrom
the tree componentby someusuallyrandomprocedure.
Greenandovendryweightsof thesesamples aredetermined
in thelaboratoryandtheresultsareusedtoestimate theentire
treecomponent. Notethatthe"measurement" of biomassis
Forest
Sctence
45(4)1999 583
definedastheprocessof directdetermination
of thebiomass freshanddry weight.For example,if the randomnumber
of theentiretreecomponent of interest,whereasthe"estima- was24 andthe sectionlengthwas5.0 m, thena disk would
tion"ofbiomass isdefinedastheprocess ofdetermination
of be located at 0.24 x 5.0 = 1.2 m from the base of that
thebiomass by subsampling. section.Eachof thethreebolesectionswascutinto logsof
variouslengthsand weighedon a 90 kg capacityfield
1.2.1Ratio-TypeEstimators
scale.Disks approximately5 cm in width were removed
Briggset al. (1987) describeda proceduretheyusedto
fromtheboleat therandomlyselectedlocations,weighed,
measurethegreenweightandto estimate,by subsampling,
and transportedto the laboratory.Foliage,branches,and
the dry weight of the aboveground components of ran-
diskswereplacedin forcedair kilns at 65øCuntil constant
domly selectedsugarmaple (Acer saccharumMarsh.)
weights were obtained.The ovendry weight was deter-
trees.Foliageandbranchdry weightsweredeterminedby
minedindividuallyfor eachpile of branchesandfoliage,
directmeasurement. Bolewoodandbolebarkdry weight
aswell asfor eachdiskof eachindividualsampletree.The
wereestimatedby stratifiedsubsampling andsubsequent bark was removed from each disk, dried at 65øC, and Its
applicationof ratio-typeestimators.A brief descriptionof
weight was recorded.
their procedurefollows. After measurementof diameter
The three bole sectionscan be considered as strata, and
andtotalheight,eachtreewasfelled,andtenplasticsheets three disks are selectedat random from each section,hence
wereplacedonthegroundsurrounding thetree.Beginning
the methodof disk selectionis stratifiedrandomsamphng.
at the baseof the crownandworkingtowardsthe top, the
tree branches with their leaves attached were removed and
Becausethe green weight of the entire bole, individual
sections,
anddisksareknown,andtheovendryweightsof the
separatedinto ten piles suchthat eachpile had a similar
sampledisksaremeasured, onecanestimateovendryweight
distributionof branchesandfoliagewith respectto weight
of the bole by a stratifiedratio estimator.Notation and
and point of origin from the crown. For each of the ten definitions for the ratio estimator are shown in Exhibit A.
piles,all of thefoliagewaspickedfrom thebranchesand
placedin paperbags.Foliageandbrancheswereweighed Because
theDh'sareindependent
random
variables,
the
ovendryweightof theboleandits errorcanbe estimatedby
for greenweight and then sentto a laboratoryfor oven
dryinganddirectmeasurement of dry weight.The bole of
eachsampletree wasdividedinto threesectionsof equal D = X;D
h = stratifiedratioestimator
of thedryweightof
length. For each section,three integerswere randomly wood and bark of the bole
selected from 1 to 100. Each of these numbers was multi-
plied (asa decimalnumber)by thesectionlengthto obtain B=ZB
h=estimator
ofthe
bias
ofD (42)
the locationof a sampledisk for the determinationof the S}=ES}•
=estimator
ofthevariance
ofD

Exhibit A

greenweightof sectionh
greenweightof woodandbarkof diskk in stratumh
ovendryweightof woodandbarkof diskk in stratumh
3 = numberof sampledisksperstrata
Zgn•/ mh
= •.dnk/ mn
Gn/ •n = conceptual
number
of disksof weight•n in section
h
finitepopulation
correction
factorof sectionh

•(dh•- •n)2/ (mn- 1)= sample

variance
ofthemndrydiskvalues
within
section
h
•(gh•--•n)2/ (mn--1)= sample
variance
ofthemngreen
diskvalues
within
section
h
Z(dh•- •n)(gh•--•n)/ (mn--1)= sample
covariance
•n/ •n= ratioestimator
ofovendry
weight
togreen
weight
ofsection
h
Ghrn= Mh•n= ratioestimator
ofovendry
weight
ofsection
h
(Mn- mn)(rnSg•
- Sa•g
•)/ (mn• n)=estimator
ofthebias
oldn
Mn(Mn- mh)(S•h- 2rnSahg
• +rsSg•
)/ mn=estimator
ofthevariance
ofDn

584 Forest
Sctence
45(4)1999
 Table 5. Morphologicaldata for the four examplesweetgumtrees.

Greenweightof bole(wood+ bark)

Tree Dbh(em) Totalheight Bolelength Bottom Middle Top Total
1 15.5 14.6 13.5 57.5 25.2 7.6 90.3
2 33.0 24.7 24.0 435.1 162.8 24.0 621.9
3 48.8 29.9 28.4 1,447.1 805.5 92.5 2,345.1
4 67.8 34.4 29.8 2,785.0 1,707.8 403.3 4,896.!
No'm: The strata (bottom, middle, top) are of equal length, being 1/3 x bole length.

If wedefine
dnk,
w= ovendry
weight
ofwoodofdiskk in was fitted, where W is branchweight,D is branchbase
stratum
h,anddnk,b
= ovendry
weight
ofbarkofdiskk in diameter,andL is branchlength.All brancheson the tree
stratum
h,andif these
values
aresubstituted
fordn&
in(41), weresubsequently measuredfor D andL, thenweightswere
thenonecandefinetheestimators
DwandDb,thestratified estimatedand summedfor total crown weight. Branches
ratioestimatorsof the ovendryweightof woodandbark, werechosenfor weighingasfollows.Withinthecrownof the
respectively,
of thetreebole,aswell asthecorresponding tree, five locationsalongthe main stemwere determined
estimatorsof their errors.Briggset al. (1987) give ex- randomlywith a probabilityproportional to stemdiameter.
amplesof calculationsfor threesugarmapletrees.Tables For eachlocationthe nearest(unselected)nodeof branches
5 and6 presentmorphologicaldataanddiskweightsfrom was selected,and from this node a branchwas randomly
four sweetgum(LiquidambarstyracifiuaL.) treesfrom a chosen for measurement. Error of estimates can be deter-
standin west-centralMississippi.Tables7 and8 showthe minedbasedon formulasgivenearlierfor ratioestimators
calculations for the stratified ratio estimator. These trees andregessionvariance.
arepartof a largerdatasetthatwasusedto developweight Valentineet al. (1984),aswell asCunia(1979),pointout
tablesfor sweetgum(Schlaegel1984). the well-known fact that ratio estimatorsare biased.Indeed,
Kleinn and Pelz (1987) in Germany estimatedboth Briggset al.(1987)acknowledge thisbut arguethatin their
greenanddry weightof theboleincludingbarkby simple procedure,bole biomassis basedon nine disks,so biasis
ratioestimatesof volume/green weightandgreenweight/ expectedto be negligible.Ratio-typeestimatorshave the
dry weight on the basisof five disks that were selected advantageof beingsimpleto understand andapply.How-
w•th a probabilityproportionalto estimatedvolume.That ever,efficient,unbiasedtechniquesareavailablewhichtypi-
is, random numbers between 0 and 1 were drawn, and cally involveonly two to four sampledisks.Thesewill be
proportionalcumulativevolumesup the stemwere esti- discussed next.

matedanddisksremovedfromthetreeat thesepoints.For 1.2.2Randomized BranchandImportanceSampling

example,say 0.333 is randomlydrawn, then a disk is Valentineet al. (1984) and Gregoireet al. (1995b) de-
removedat the pointonthe stemwhereit is estimatedone- scribetwo procedures, randomizedbranchsampling(RBS)
thirdcumulativevolumeoccurs.For crowngreenanddry and importancesampling,for selectingsamplepathsto
weight,a few brancheswere selectedand weighedand a obtain unbiased estimates of the biomass content of the tree.
regressionof the form A samplepath--from whichboledisks,crownbranches,and
foliageareselected---extends
from thebuttto a terminalbud
(43) andhasselection
probabilities
associated
withit. Thepathis

Teble6. Greenenddryweightsof the threerandomlyselecteddisksperstratumforthe fourexemplesweetgum

trees.

Disk 1 Disk 2 Disk 3

location * location location
Tree Stratum (m) g,, d,, (m) g•2 d,2 (m)
..........(kg)........................ (kg)........................ (kg)..........
1 Bottom 2.0 0.586 0.269 2.6 0.540 0.258 3.6 0.461 0.238
1 Middle 2.2 0.280 0.133 2.4 0.271 0.129 2.9 0.246 0.119
1 Top 0.0 0.176 0.095 2.4 0.081 0.028 4.3 0.012 0.009
2 Bottom 3.1 2.408 1.083 3.5 2.286 1.008 6.4 !.832 0.810
2 Middle 4.5 1.018 0.470 4.9 0.888 0.43 ! 5.7 0.720 0.34 !
2 Top 1.0 0.325 0.139 6.2 0.036 0.016 7.0 0.020 0.015
3 Bottom 3.0 8.157 3.444 3.2 8.092 3.426 9.2 6.000 2.670
3 Middle 2.7 5.744 2.754 4.4 5.287 2.395 5.6 4.160 1.938
3 Top 1.5 0.998 0.499 2.2 0.815 0.410 3.5 0.545 0.259
4 Bottom 4.7 13.254 6.162 7.5 12.554 6.003 8.0 12.251 5.828
4 Middle 2.0 9.995 4.693 6.0 8.192 3.902 7.4 7.149 3.544
4 Top ! .2 4.399 2.355 2.6 3.045 1.639 9.8 0.601 0.299

No•: ghkisthegreenweightof diskkin stratumh anddhkisthedryweightof diskkin stratumh.Eachdiskisapproximately

5
cm thick.
Withina stratum,the baseof a diskwasrandomlylocatedbygeneratinga uniformrandom(0,1)numberandmultiplying
it by
the stratum length.

Forest
$ctence
45(4)1999 585
 Tabla 7. Statisticsassociatedwith the estimationof the bole ovendryweight for the three sectionsof the four
example sweetgum trees.

2 S2 S2
Stratum •h •h Sdh Sdhgh gh rh Dh Bh Dh
............
(kg)...............................(kg2)................................. (kg)........... (kg2)
Tree 1
Bottom 0.255 0.529 0.00025 0.00099 0.00400 0.482 27.72 0.062 0.8346
Middle 0.127 0.266 0.00005 0.00013 0.00031 0.478 12.05 0.002 0.0043
Top 0.044 0.090 0.00204 0.00363 0.00678 0.491 3.73 -0.092 0.2567
Tree 2
Bottom 0.967 2.175 0.01989 0.04271 0.09213 0.445 193.41 -0.053 1.6051
Middle 0.414 0.875 0.00438 0.00977 0.02232 0.473 77.00 0.055 1.4357
Top 0.057 0.127 0.00508 0.01223 0.02947 0.446 10.71 0.448 0.4281
Tree 3
Bottom 3.180 7.416 0.19516 0.54204 1.50556 0.429 620.49 0.894 89.0315
Middle 2.362 5.064 0.16726 0.32861 0.66467 0.467 375.79 -0.190 44.0265
Top 0.389 0.786 0.01472 0.02763 0.05193 0.495 45.82 -0.093 0.4082
Tree 4
Bottom 5.998 12.686 0.02791 0.08322 0.26464 0.473 1,316.65 0.238 132.6241
Middle 4.046 8.445 0.34567 0.84494 2.07306 0.479 818.24 1.166 159.8811
Top 1.431 2.682 1.08923 2.00885 3.70521 0.534 215.21 -0.580 2.6900
NOTE:
•, ismean
drydisk
weight
ofstratum
h.• ismean
wetdisk
weight
ofstratum
h.Sd•= isthevariance
ofdh.Sd•;•isthe
covariance.
S2 isthevariance
ofgh.rhistheratio
estimator
ofdrytogreen
weight
ofstratum
h.Dhisratio
estimator
ofdry
weight
ofstratum
h.Bhisthebias
ofDh.andSo•isthevariance
ofDh.See
text
fordetails.
a seriesof connected
branchsegments
or internodes,where biomasscanbeestimated froma singlepath,buttwoormore
a branchis definedastheentirestemsystemthatdevelops pathsareneededto computea standarderrorof theestimate
from a singlebud.A segmentis a partof a branchbetween Estimationof the green weight of the tree involvesthe
two consecutive nodes.The butt,by definition,is the first weightsof eachof the n segmentsof the path. Denotethe
nodeandhasselection
probability
q] = 1.Thesecond
node weight
ofthekthsegment
asbk,thenanunbiased
estimate
of
occursat thepointof live treelimbs.To continuethepath,a tree weightis
selectionprobabilityis assigned to eachbranchemanating
from the secondnode,andoneis chosenat random.Valentine n bt•
et al. suggest
assigning
a selectionprobabilityastheproduct /•=E •
k=l (45)
of thesquared diameterandlengthfor a branch,dividedby
the sumof theseproductsfor all branchesat the node.The whereQkis definedin Equation^(44). For an unbiased
second
segment
ofthepathhasselection
probability
q2'The estimate
of greenfoliageweight,f, substitutefk forbt••n
pathcontinues
to thenextnode,wherea branchisselected
by Equation
(45),wherefk is theweightof thefoliageattached
RBSwithprobability
q3andsoonuntila terminal
shootis to the kth segment.
reached
withprobability
qn'Theqi'sareconditional
prob- Valentineet al. (1984) developeda procedurebasedon
abilities.The unconditional
probabilityof selection
for the importancesampling(a techniqueof MonteCarlo integra-
kth segmentin thepathis tion) for selectingdisksthatproducesunbiasedestimatesof
dry weight.To begin,eachsegmentin the selectedpathis
k
enlarged
bytheinflation
factor1/Qi
e sotheenlarged
stem
(44)
represents
theentiretree.Visualizetheinflatedpathasbeing
composedof thin disksof constantthicknessand known
volume. One of these disks is selected at random with
All materialthatis notpartof thepathcanbediscarded.
This is a big advantageof RBS; as a result,researchers
can probabilityproportional to its volume.If the dry weightof
significantlyreduceprojecttimeandlaborcosts.Aboveground that inflateddisk is measuredand dividedby its selection
probability,the resultis an unbiasedestimateof the dry
Table 8. Summary statistics for the four example sweetgum weightof thetree.
trees. In practice,Valentine et al. (1984) used a continuous
95%
(segmented linear)interpolation functiontopredictthecross-
Confidence limits sectionalarea(volumeperunitlength)of all pointsalongthe
2 path.Theymeasured diameterat numerous pointsalongthe
Tree D Bias SD Lower Upper pathfor thispurpose. Denotethediameterof the stemat a
............
(kg)........... ............(kg).............
1 43.49 -0.027 1.0957 41.40 45.58 distance
Ls fromthebuttasD(Ls),anddefinea quantity
2 281,12 0.449 3.4689 277.40 284.84 proportionalto the inflatedcross-sectional
areaas
3 1,042.10 0.611 133.4661 1,018.99 1,065.21
4 2,350.11 0.824 295.1952 2,315.75 2,384.47 A(Ls)= D(Ls)
21Qk (46)
NOTE: D is the stratified ratio estimate of the bole dry weight (wood + Now the interpolation
function,S(L), is fittedto the values
bark)andS• is theestimate
of thevariance
of D usedfor
constructing confidence intervals. A(Ls)andintegrated
overthelength,
•,,ofthepathtoapproxi-

$86 Forest
Sctence
45(4)1999
matetheinflatedwoodyvolumeof thepath,thatis,
kH-kJ

V(•,)
--f;$(L)dL (47)wherek isstumpheight.DifferentiatingEquation(52)gives
A point, 0, for cuttinga disk is randomlyselectedwith a(a)= aV(H- k) (53)
probabilityproportional to S(L). The pointis chosenwhich
saUsfies V(0) = uVO•),whereu is a randomnumberfroma Usingthesamefunctional
formfor massgives
uniform(0,1) distribution.Next, determinethe dry weight
•'(h) = •3W(H-h)•-l (H- k)-I! (54)
perunitthickness (Valentineetal.used10-cm-thick disks)of
the disk cut at L = 0 as B(0). The inflatedweightper unit DividingEquation(54) by Equation(53), as indicatedin
thickness of the disk is Equation(51),andsimplifying
results
in

B*(O)
= B(O)
/ (48) = (u- (55)
wherek is theindexof thepathsegmentin which0 occurs.
It shouldbepointedoutthatanyfunctionthatgivesvolume
Finally,theunbiased
estimateof thetruewoodydry weight
to someheight'h, asdoesEquation(52), canbe usedfor
of thetreeis computedas
approximating p(h).
•, = B*(0)V(•)/ S(0) (49) Importance
sampling is usedfor estimating
thevalueof
anydefiniteintegral.SinceEquation(50) (thedensity-inte-
If multiplepathsareselected
onthetreefromRBS,obtaina gralmodel)describes bolewoodymassasa definiteintegral,
diskfromeachpathanduseEquation(49) to computean a samplingschemethatutilizesthe aboveequations canbe
estimatefrom eachdisk,thenaveragethe estimatesto pro- developed.VanDeusenandBaldwin(1993) devisedascheme
duce one combined estimate. For further details and ex- to estimate the difference between the model and actual bole
amplesonRBS andimportance sampling,seethepapersby biomass.
Thisapproach,
called"differencesampling,"
com-
ValentineandHil ton(1977), Valentineet al. (1984), Gregoire binesimportance
sampling
andcontrolvariatemethods.
The
et al. (1986), de Gier (1989), andGregoireet al. (1995b). desired difference can be written as

1.2.3BoleMassbyDifferenceSampling
This sectiondescribesaninnovativetechniquefor ob-
taininganunbiased estimateof treebolebiomass.
Gregoire
w(h)
- •(h)= ' w'(x)
-•/(x)
f(x) f(x)
dx (56)
et al. (1986) showed how to unbiasedlyestimate bole
wheref(x)is a probabilitydensityfunction(PDF). Equa-
volumeby importancesampling.Van DeusenandBaldwin
tion(56)conveysthatwecandrawa height,Xi, fromthe
(1993) usedimportancesamplingin conjunctionwith the
density-integralconceptof ParresolandThomas(1989) to
PDF,f(x), andmeasurew'(Xi)tocontrast
withthemodel-
basedestimate3'(Xi) for an unbiased
estimateof the
obtain an unbiasedestimate of tree bole dry mass. The
difference, w(h)-•(h). Hence, a procedurebasedon
procedurerequires obtaining increment cores at breast
differencesamplingfor obtainingunbiasedestimatesof
heightandanotherrandomlyselectedheight.The specific bole mass would use
gravity of the coresand their associatedcross-sectional
areasare then usedto unbiasedlyestimatebole mass.
In thedensity-integral
modelthebolewoodymasstosome %(h)=
½(h)+ w(xi)-w(x0
heighth is ni:1 f(x) (57)
^ PDF should be chosen that will lead to most of the
w(h)
--I• p(x)a(x)
dx (50) measurementsbeinglow on the stem,wheremuchof the
wood mass occurs and where measurement cost is low.
where a(x) is the cross-sectional
area at heightx. If dx Define r = (h - x)/h, the relativedistancefrom the upper
representsdiskthickness,
thena(x)dxis volume,andvolume heightlimitof interestonthebole.NotethattotalheightH is
tamesdensityyieldsmass.Thusp(x) is theratioof massto just a specialcase.A simplecumulativedensityfunction
volume.Takingthederivativewithrespectto h andrearrang- (CDF) for r is
ing gives
F(r) = r'/, 0 -<r < 1 (58)
p(h) = w'(h) / a(h) (51)
Using the inversetransformmethod,a randomheightin
A reasonable functionfor p(h)depends on thepropertiesof terms ofx is drawn as
w(h)andthevolumefunction,v(h),whichis theintegralof
a(h). Both functionsincreasemonotonicallyup the stem Xi = n(l- u?) (59)
startingfrom0 atthebaseandgoingto totalwoodymassW,
andtotalvolumeV, at totalheightH. For volumeto heighth, where
ui isauniform
random
variate,
ui~U(0,1).
Differenti-
Van Deusenand Baldwin (1993) used ating(58) givesthePDF in termsofx as

Forest
Sctence
45(4)1999 587

f(x)=
yr•-I
=y/h•-)
•-I
Substituting
(60) into Equation(57) yieldsthe suggested
differencesamplingformula
(61) multiplied by the numberof trees (Attiwill and Ovington
To implementtheproceduredefinedby (61), firstrecall
thatEquation(54)provides a modelfor ½'(h),andEquation forestinventoryconsistof twoprincipalphases.
(52) givesa modelfor ½(h) afterchanging thev's to w's. phase,a relativelylargesampleof treesis selected,andthe
Second, generate
auniform randomvariate,ui,andsubstitute treesare measuredfor diameter,height,andpossiblyother
thisintoEquation(59)togeneratea measurement height,Xr characteristics.
Third,measurethecross-sectional
area,a(Xi),onthetreeand
takeacoretoobtainp(Xi ) sincewz(Xi) = p(Xi )a(Xi ).They- (Bitterlich)samplepoints.Thesetreesarenotmeasured
parameterin Equation(59) influencesthe probabilityof
whereXi occurs
on thebole.Basedon simulations,
Van
DeusenandBaldwin(1993)showedthata valueofT= 3 kept
measurement heightslow (nearly90% of the timelessthan
halftreeheight,and60%of thetimelessthanone-quarter tree
height) while minimizingpoor predictionson individual
trees.
The onlyremainingelementneededis anestimateof total
woodymassWfor usein Equations (52) and(54). Notethat
onecaneasilyobtainw' (1.3) by takinga coreatbreastheight
to determinedensityandthenby multiplyingthisdensityby
the measuredbasalarea a(1.3). The following equationis
derivedfromEquation(54) by lettingh = 1.3m andrearrang-
ing terms:
if/=w'(1.3)(H-
1.3)
•-•(H-k)•
If thisestimateof W is used,then ½'(h) is constrained
to
predictthemeasured valueat 1.3m regardless of
Differencesamplingcanbe usedto providean unbiased
estimateof thebiomassof a standof trees.The procedurecan
be appliedto eachtreeon a sampleplotto give anunbiased
estimateof theplotmass.Theseplot-massestimatescanthen
beusedin theusualwaytoproducesampleestimates of stand
biomass.To illustratethe aboveprocedure,consideragain
the 39 willow oak trees in Table 2. Let us estimate the total
bolewoodymass,excluding bark,of eachtree(abovestump),
henceh = H. Onerandomheightonlywill bedrawn(n = 1),
andwewill set¾=3, [5= 3, andk = 0.5 in theformulas.In the
willow oak dataset,a measureof specificgravity occurs
every 1.5 m alongthe stem,so the randomheightwill be
adjustedto thenearestheightwherecross-sectional areaand
specificgravitymeasurements wererecorded. Also, w' (1.5)
willbeused
instead
of w'(1.3)todetermine
if/ofEquation
(62). Resultsaregivenin Table9. Thetotalwoodymassfor
the 39 treesis 40,987kg, whereasthedifferencesampling
estimateis 39,099 kg, a differenceof only 4.6%. In this
example,onlyonerandomheightpertreewasdrawn,butin
practiceusuallytwo to four randomheightsmaybe drawn
andmeasurements takenat thosepointsalongthebole.This,
of course,shouldimprovethe accuracyand precisionof
WD(h).
588 ForestSctence
45(4)1999
2 The Error of Forest Biomass
(60)Inventory Estimates
Historically,attemptshavebeenmadeto estimateforest
biomassusing"meantree" techniques.For example,the
weightof thetreeof averagegirthwouldbedetermined and
1968).Thisgenerallyprovedunsatisfactory andtodaylarge-
scaleinventoriesbasedon soundstatisticaldesignsare •n
placein manypartsof theworld.Most samplingdesignsof
In thefirst
The sampletreesare usuallyin clustersde-
fined in termsof sampleplotsof fixed areaor horizontal
for
biomass.In the secondphase,a relativelysmallsampleof
treesis selected,and the treesare measuredfor biomassand
the samecharacteristics
asthe first phasetrees.The second
phasetreesare usedto estimatea relationshipbetweentree
characteristics
(diameter,height, age, etc.) and biomass,
usually,thoughnot always,expressed asa regression func-
tion.Thisrelationship
is thenappliedto thetreesof thefirst
phasesampleto calculateforestinventoryestimates of aver-
age biomassper unit area. When previouslyconstructed
biomass regressions
areavailable,thesecond phasesampleis
nolongernecessary. However,a criticalassumption is being
madethatthetreepopulationfor whichtheregression func-
tion wascalculatedandthetreepopulationscurrentlybeing
inventoriedare very similar.Somerecentstudiesdealing
with forestbiomassestimates from inventorydatainclude
(62) Brown et al. (1989) and Brown and Lugo (1992) in the
neotropics,
Brownet al. (1991) in SouthandSoutheast
Asia,
and Monserudet al. (1996) in Russia.Two excellentrefer-
enceson forestinventorymethodologyarede Vries (1986)
and Schreuderet al. (1993).
The error of the forestinventoryestimateshastwo mmn
components. First is the componentdue to the random
selectionof the sampleunitsof the first phase.Successive
applicationsof the sameselectionprocedureto the same
forestarearesultin differentsetsof sampletreesand,thus,
different setsof estimates.The size of this component•s
greatlyaffectedby (1) thesamplingdesignof thefirstphase,
(2) thesamplesize,(3) thetypeof estimatorused(for given
sampledata and requiredparameterto estimate,thereare
generallyseveralestimators, eachhavingitsownprecision),
and(4) theinherentvariationbetweenthesampleunits.The
secondcomponentis associatedwith the sampleof the
second phase,thatis,withtheerrorof thebiomass regression.
The size of this componentis affectedby (1) the sampling
designusedto selectthesetrees,(2) the samplesize,(3) the
estimationprocedure,and(4) the inherentvariationof the
treebiomass valuesabouttheregression function.Thesetwo
components constitutewhatis knownasthesamplingerror.
An approach proposedby Cunia(1965, 1987a)can be
usedto combinetheerrorfromthe firstphasesampleplots
with the error from the secondphasesampletrees.Thts
approach requiresthattheestimators be of theform
Table9. Comparison
of actualbolawooddrymasswithdifferanca
samplingestimataforwillowoaktraesfromt hastataof Mississippi,
USA.

True
Tree H mass p(l.5) a(1.5) I• X p(X) a(X) w'(X) ½'(X) WD(H
)
(m) (kg) (kg/m
3) (m2) (kg) (m) (kg/m
3) (m2) ......................
(kg).......................
1 29.0 2,493.4 584 0.361 2,152.5 18.0 603 0.092 55.69 33.75 2,203.3
2 18.3 309.4 596 0.061 243.4 3.0 599 0.052 30.96 30.31 243.7
3 22.9 941.7 613 0.173 869.7 3.0 601 0.143 85.95 91.93 867.0
4 27.4 1,772.7 536 0.364 1,886.8 3.0 530 0.273 144.55 173.13 1,874.8
5 19.8 277.1 608 0.053 229.4 9.0 613 0.028 17.01 11.16 235.9
6 32.0 1,301.4 585 0.177 1,160.9 6.0 605 0.125 75.56 75.33 1,161.1
7 32.0 1,000.6 554 0.161 996.1 4.5 565 0.114 64.42 72.31 992.6
8 30.5 885.4 544 0.146 852.5 19.5 567 0.037 21.25 11.46 877.6
9 25.9 464.9 573 0.073 383.6 6.0 590 0.056 32.96 27.81 386.5
10 30.5 1,558.1 564 0.234 1,413.7 10.5 582 0.143 83.23 62.83 1,429.5
11 24.4 376.5 574 0.064 316.6 12.0 575 0.031 17.73 10.70 325.7
12 30.5 2,117.9 580 0.372 2,309.8 4.5 575 0.230 132.19 173.49 2,290.8
13 24.4 721.7 580 0.112 566.2 1.5 580 0.112 65.25 65.25 566.2
14 25.9 802.9 551 0.133 672.3 9.0 557 0.091 50.68 35.15 684.4
15 35.1 2,290.2 593 0.327 2,370.9 1.5 593 0.327 193.86 193.86 2,370.9
16 25.9 662.7 588 0.123 665.3 4.5 582 0.095 55.35 55.77 665.0
17 21.3 272.2 578 0.058 256.6 6.0 607 0.037 22.75 20.02 258.3
18 32.0 3,029.1 582 0.471 3,072.5 1.5 582 0.471 274.33 274.33 3,072.5
19 33.8 2,147.3 568 0.268 1,796.3 1.5 568 0.268 152.25 152.25 1,796.3
20 29.0 1,438.8 560 0.261 1,491.9 15.0 603 0.071 42.54 37.90 1,498.6
21 25.9 171.0 600 0.032 178.5 3.0 600 0.026 15.33 17.14 177.8
22 24.4 365.6 617 0.061 328.3 9.0 647 0.031 19.95 17.11 330.6
23 27,4 1,009.7 607 0.189 1,108.2 7.5 692 0.112 77.85 67.63 1,114.6
24 27.4 523.5 603 0.087 507.2 16.5 630 0.031 19.42 9.29 528.5
25 27.4 667.2 564 0.131 716.5 9.0 591 0.064 37.57 37.39 716.6
26 24.4 240.4 589 0.042 214.5 6.0 596 0.030 17.91 15.96 215.6
27 25.9 394.2 606 0.062 347.1 13.5 640 0.024 15.44 9.77 355.4
28 25.9 510.8 580 0.090 477.0 0.5 590 0.152 89.48 56.34 488.5
29 27.4 732.1 591 0.104 592.3 4.5 602 0.080 48.43 47.87 592.6
30 22.9 836.9 606 0.177 878.4 4.5 612 0.130 79.39 79.38 878.4
31 25.9 440.0 584 0.084 451.8 3.0 584 0.067 39.14 43.38 450.0
32 21.3 468.6 526 0.094 377.2 1.5 526 0.094 49.30 49.30 377.2
33 21.3 356.5 580 0.074 329.2 3.0 588 0.060 35.40 36.76 328.6
34 25.9 1,385.7 561 0.270 1,391.7 4.5 560 0.236 132.39 116.68 1,399.4
35 25.9 1,240.6 587 0.205 1,102.5 1.5 587 0.205 120.17 120.17 1,102.5
36 27.4 1,530.4 567 0.245 1,345.0 4.5 562 0.193 108.28 108.71 1,344.8
37 27.4 2,056.6 560 0.359 1,942.0 12.0 543 0.191 103.55 70.98 1,976.4
38 25.9 1,208.4 464 0.243 1,034.6 12.0 576 0.131 75.65 36.59 1,079.8
39 27.4 1,984.5 567 0.329 1,807.0 9.0 585 0.241 140.87 94.29 1,841.4

SUM 40•986.8 39,099.3

NOTE: p(1.5) is ratio of mass to volume at height 1.5 m (i.e., specificgravity measuredat height 1.5 m x 1000), a(1.5) is bole cross-sectionalarea
at 1.5m, W is estimateof boledry massvia Equation(62),Xis a randomheightgeneratedfrom Equation(59), w'(X} is p(X}a(X),•'(X} is
a model-basedestimateof w'(X}via Equation(54),and WD(H)is the differencesamplingestimateof bolebiomassvia Equation(61).

• = b•z•+ b2z2+...+ bmZ

m= btz (63) where
SzzandSbbarethecovariance
matrices
ofz andb.
ThefirsttermofSww
isthevariance
component
associated
whereb is the coefficientvectorfrom the biomassregression with the error of the sampleplots,andthe secondtermis
function andz is a vector of statisticscalculated from the data of the variancecomponentassociatedwith the biomassre-
thesample pointsorplots.It is assumedthat(1) theregression gression.The definitionofz dependson (1) the sampling
oftreebiomass onXisofthelinearformy=X• + •, (2)thevector designby which the plots or pointsare chosen,(2) the
z isdefined sothat½ isanunbiased estimateof theparameter specificparameterg one wishesto estimate,and (3) the
of interestg, thatis, variablesx' usedin the biomassregressionfunction.
Cunia(1987a,b,c,d,e,f),in a seriesof papers,described in
Iz= II'lz== E[tl'E[z] detailthesteps
oftheaboveapproach forcombining thefirstand
second phase
errorcomponents whentheparameter g ofinterest
and(3) thevectorsb andz arestatisticallyindependent.
The is the averagebiomass perhectareandthe samplingdesigns
variance of w is calculated as were:(1) simplerandomsampling, (2) sa-atified
sampling, (3)
two-stage sampling,(4)double sampling, (5)Continuous Forest
= b'S=b+ z'Sbbz (64) Inventory(CFI) withoutSamplingwith PartialReplacement

Forest
Sctence
45(4)1999 589
(SPR),and(6) CFI withSPR.Detailsandexamples
onallthese to achieve efficient estimates and to construct valid standard
designsareprovidedin Cunia'spapers. errorsandconfidence regions.New research is showingthat
spatialandtemporalcorrelation is commonin all typesof
3 Discussion forestrydata,andmodelingthesecorrelations in treeandplot
datacanprovideconsiderable gainsin efficiencyandestima-
In earlierdraftsof thisarticle,I waschidedby reviewers tion (Gregoireet al. 1995a,GoelzandBurk 1996,Gregoire
forusingthetermsweightandmassinterchangeably. Weight andSchabenberger 1996).Futureeffortsin modelingtreeand
andmass,thoughrelated,are notthe same(seefootnote1). standbiomassandupdatinginventoryestimates shouldtake
I havemadeaneffortthroughout thisretrospection
toreferto intoaccountthesecorrelations. Further,normallydistributed
eitherweightor massas appropriate, andto usethe term errorsarealmostalwaysassumed andrarelyverified.Will-
weightwith thosetechniques, suchastheratio-typeestima- iams and Schreuder(1996) have looked at the normality
torsof 1.2.1,thatdealwith weightestimation, andto usethe assumption with volumemodelsandoffer alternativeerror
termmasswiththosetechniques, suchasthedensity-integral distributionsthatcouldwell be appliedto biomassmodels.
modelof 1.1.8,thattrulydealwith estimationof mass.This As alludedtoin theintroduction, remotesensing will play
isasit shouldbe,andresearchers inthefutureshould properly aneverincreasing rolein standbiomassestimationandforest
distinguish thetwo. productivityin general(de Gier andSakouhi1995,Gholzet
Researchon estimatingbiomasscomponentsof trees al. 1997), with much researchbeing neededin this area.
and forests has a long tradition. Information on many Large-scaleforestinventories,suchas thoseconductedby
speciesfor different sitesand standstructuresis available. stateandfederalagencies, arelookingto methodologies such
Of historical interest are the volumes put out by the asimputationfor updatingcurrentbiomass.Tree andstand
IUFRO workinggroupthatwasinitiatedby HaroldYoung: biomassmodelingandsamplingmaywell be supplanted by
Forest biomass studies 1971, ILlFRO biomass studies remotesensing andmultivariate statistical
analysisor,pref-
1973, and Oslo biomassstudies1976. These volumeswere erably,linkedwith themin the future.
publishedby the University of Maine at Orono. Two
importantbooks on biomassare Satoo (1982), dealing Literature Cited
primarily with Japanese effortsto systematizeforestbio- ATIIWlLL,P.M., A•a•J.D. OVlNGTON.
1968. Determinationof forestbiomass
massdata and estimation,and Madgwick (1994), a com- For. Sci. 14(1):13-15.
prehensivework dealing with the single speciesPinus Awg¾, T.E., • H.E. BUgrdqART.
1994.Forestmeasurements.
Ed. 4. McGraw-
radiataD. Don.A shortlist of currentarticlesdealingwith Hill, New York. 408 p.
biomassestimationis: Korsmo(1995) on sevenhardwood BncDWm,V.C., Jg.1987.A summary of equationsforpredicting
biomassof
speciesin Norway; Llsol'TevandVanclay (1995) onScots plantedsouthernpines.P. 157-171 in Estimatingtree biomassregres-
pine (Pinus sylvestrisL.) in Kazakhstan;Wang et al. sionsandtheirerror.Proc.of theWorkshop onTreebiomass regression
functions
andtheircontribution totheerrorof forestinventory
estimates,
(1995) on aspen(PopulustremuloidesMichx.) in British Wharton,E.H.,andT. Cunia(comps.). USDAFor.Serv.Gert.Tech.Rep
Columbia;Tahvanainen(1996) on seven$alix clonesin NE-117. 303 p.
Finland;Bartelink(1997) onbeech(FagussylvaticaL.) in BARtaUNt,,
H.H. 1997.Allometricrelationships
for biomass
andleafareaof
theNetherlands;andin thePeople'sRepublicof China,Li beech(FagussylvaticaL). Ann.Sci.For.54:39-50.
et al. (1996) on Japanese redpine(PinusdensifloraSieb. Box, G.E.P.,• D.R. Cox. 1964.An analysisof transformations.
J. Roy
and Zucc.), and Zhou et al. (1997) on Manglietia Stat. Soc. Series B 26:211-252.

hainanensisDandy. Bmc,
as, R.D., T. Cure^,E.H. WHrm,^tinH.W. Y^w•¾. 1987.Estimating
Thereare manywaysto determinetreebiomass.Ratio- sampletreebiomass bysubsampling:Someempiricalresults.
P. 119-127
in Estimatingtree biomassreg•ssionsand their error. Proc.of the
typeestimators, differencesamplingestimators, andothers Workshop onTreebiomass regression
functions
andtheircontribution
to
areappropriate if oneonlyneedsestimates of thetotalwoody the errorof forestinventoryestimates,Wharton,E.H., and T. Cuma
biomassof the tree (or bole). If, however, one wants to (comps.).
USDAFor.Serv.Gen.Tech.Rep.NE-117.303p.
developweight-ratioor density-integral typemodels,a more Bgo^D,L.R. 1998.Allometryandgrowth.For. Sci. 44(3):458-464.
intensivesamplingschemeon the bole (suchas systematic
BgowN, S., A.J.R. G•tJ•sPm, ^tin A.E. Luc,o. 1989. Biomass estimaaon
sampling)wouldbe moreappropriate. Undershort-rotation methodsfor tropicalforestswith applications
to forestinventorydata
woodybiomass programs,treestypicallydonotattainalarge For. Sci. 35(4):881-902.
size,sodevelopment ofweight-ratio
ordensity-integralmodels BgowN,
S.,A.J.R.GiLL•s•m,
^tinA.E.Luc,o. 1991.Biomass
oftropicalforests
isprobablynotlaborefficientorcosteffective.Olderplanta- of South and SoutheastAsia. Can. J. For. Res. 21:111-117.
tionsandtreesthatattaina largesizemightyield a mix of Bgow•, S., • A.E. Luc,o. 1992. Abovegroundbiomassestimatesfor
products in thebolesuchaspulpwood,fuelwood,andsmall tropicalmoistforests
of theBrazilianAmazon.Interciencia
17(1):8-18.
dimensional lumber.For theseplantations andtrees,devel- Bugrd-tngT,
H. E. 1977.Cubic-footvolumeof loblollypineto anymerchant-
opmentof predictionsystemsfor merchandizing treeboles abletoplimit. South.J. Appl.For. 1:7-9.
wouldbe advantageous, andintensivesamplingschemes to
CARROLL,
R.J., ANDD. RUPERT.
1988. Transformation
and weightingin
developsuchpredictionsystemswouldbe costeffectivein regression.
ChapmanandHall, New York. 249 p.
thelongrun.
CmYENr•^,
S.S.,ArmA. KOZAK.1984.Additivityof componentbiomass
Modelingtreebiomasshasbeena wide-ranging effortin regression
equations
whentheunderlying
modelis linear.Can.J. For
forestry.It is importantto becognizantof theerrorstructure Res. 14:441-446.

590 Forest
Sctence
45(4)1999
Ct.n•œ,
A. III. 1982.Predicting
biomass
production
in theSouth.P. 119-139 Cumn,T., Am)R.D.BRIGGS.
1985b.Harmonizing biomass
tablesbygeneral-
inPredicting
growth
andyieldintheMid-South.
31stAnnual
For.Symp., izedleastsquares.
Can.J. For.Res.15:331-340.
Hotvedt,
J.E.,andB.D.Jackson
(eds.).Louisiana
StateUniversity,
Baton
Rouge,LA. 160p. DEANS,
J.D.,J. MORAN,
ANDJ. GRACE. 1996.Biomass
relationships
for tree
species
inregenerating
semi-deciduons
tropical
moistforest
inCamesoon.
CLnaZ,A.III. 1987.Summaryof biomaasequations
availablefor softwood For.Ecol.Manage.88: 215-225.
and hardwoodspeciesin the southernUnitedStates.P. 173-188 in
Estimating treebiomass
regressions
andtheirerror.Proc.of theWork- r• G•, A. 1989.Woodybiomass for fuel--estimating
thesupplyin natural
shoponTreebiomass regressionfunctions
andtheircontributiontothe woodlandsandshrublands.
ITC publication 9. Intemat.Inst.for Aero-
errorofforest
inventoryestimates,
Wharton,
E.H.,andT.Cunia(comps.). spaceSurveyandEarthSci.(ITC), Enschede, TheNetherlands.
USDAFor.Serv.Gen.Tech.Rep.NE-117.303p. DEGrog,
A.,Am)A. SA•COUm.
1995.Woodybiomassmapping,
usingfielddata
CLtrrm•, J.L., J.C. FoR•ot•, L.V. Pl•'•nng,G.H. BPaS•R,ANOR.L. BAm•Y. and SPOT-satellite
imagery.in Proc.IUFRO XX World Congress,
Tampere,Finland.
1983.Timbermanagement:A
quantitative
approach.
Wiley,NewYork.
333 p. DEV•ES,P.G.1986.Samplingtheoryforforestinventory:
a teach-yourself
CRow,
T.R. 1971.Estimation
of biomass
in aneven-aged
stand--regression course.
Springer-Verlag,
Berlin.399p.
and"meantree"techniques.
P. 35-48 in Forestbiomassstudies.Proc.of Fr.EW•t•o, J.W.,Ah'O
L.V. Pt•nnR. 1981.Multiplicative
regression
with
IUFROSection25, Young,H.E.(ed.).Collegeof Life Sci.andAgric., lognormal
errors.For.Sci.27(2):281-289.
Univ.of Maineat Orono,ME. 240p.
FvR•lVnl.,
G.M. 1961.Anindexforcomparing
equations
usedinconstructing
CRow,
T.R.,Ah'O
P.R.LAIDLY.1980.Alternative
models
forestimating
woody volume tables. For. Sci. 7:337-341.
plantbiomass.
Can.J. For.Res.10:367-370.
GHOLZ,
H.L., K. NAgs^s,Am)H. Sm•!ODA
(EDS.).1997.The useof remote
CV•aA,
T. 1965.Sometheory
onthereliability
ofvolume
estimates
ina forest sensingin themodeling
of forestproductivity.
KluwerAcademicPub-
inventorysample.For. Sci. 11:115-128. lishers,
Dordrecht,
TheNetherlands. 323p.
CV•A, T. 1979.On sampling
treesfor biomass
tableconstruction:
some Gost.z,J.C.G., Am)T.E. BVRZ.1996. Measurementerror causesbiasin site
statisticalcomments.P. 643-664 in Forestresourceinventories,Vol. 2. indexequations.
Can.J. For.Res.26:1585-1593.
Frayer,
W.E.(ed.).Colorado
StateUniv.,FortCollins,
CO. 1030p.
GREoomE,
T.G.,Am)M.E.Dv•_a.
1989.Modelfittingunderpatterned
hetero-
Ctn•A,T. 1987a.
Errorofforest
inventory
estimates:
itsmaincomponents.
P. geneityof variance.For. Sci. 35:105-125.
1-13 in Estimating
treebiomassregressions
andtheirerror.Proc.of the
GP,
sooms, T.G., Am) O. SC,nSr.•SERa•R.1996. Nonlinear mixed-effects
Workshop onTreebiomass regression
functions
andtheircontributionto
theerrorof forestinventoryestimates,
Wharton,E.H., andT. Cunla modeling
ofcumulative
holevolume
withspatially
correlated
within-tree
data.J. Agric.Biol. Environ.Star.1(1):107-119.
(comps.).USDAFor.Serv.Gen.Tech.Rep.NE-117.303p.
GP,
soo•P,$,T.G., O. SC•nBn•B•Ra•R,ANOJ.P. BAgRgrr.1995a. Linear model-
CV•A, T. 1987b.On theerrorof forestinventory estimates:
stratified
samplinganddoublesamplingforstratification.
P.63-70 inEstimating ingofirregularly
spaced,
unbalanced,
longitudinal
datafrompermanant-
plot measurements.
Can.J. For. Res.25:137-156.
treebiomassregressions
andtheirerror.Proc.of theWorkshop ontree
biomassregression
functions
andtheircontribution
totheerrorof forest GaEaO•ar_,
T.G., H.T. Vnt.smn•ts,• G.M. FvRmVnr.1986. Estimationof
inventory
estimates,
Wharton,E.H.,andT. Cunia(comps.).USDAFor. holevolumeby importance
sampling.
Can.J. For.Res.16:554-557.
Serv.Gen.Tech.Rep.NE-117.303p.
Gasc,
o•P,S,
T.G., H.T. Vnrsn'nns,
ANoG.M. Fvgn•vnr.1995b.Sampling
Ctn•A,T. 1987c.On theerrorof forestinventory
estimates:
Two-stage methodstoestimatefoliageandothercharacteristics
of individualtrees.
sampling of plots.P. 71-77 in Estimatingtreebiomassregressions
and Ecology76(4):1181-1194.
theirerror.Psoc.oftheWorkshop onTreebiomass regression
functions
andtheircontribution totheerrorofforestinventory
estimates,
Wharton, GV"rr•BVRO,
S. 1973.Evolutionof weight-scaling.
South.Lumber.227:12.
E.H.,andT. Cunia(comps.). USDAFor.Serv.Gen.Tech.Rep.NE-117.
303p.
HnRv•v,A.C. 1976. Estimatingregression
modelswith multiplicative
heteroscedasticity.
Econometrica44:461-465.
CV•A,T. 1987d.
Ontheerrorofforest
inventory
estimates:
double
sampling l-l•PP,T.E., • G.H. BPaS•R.1982.Estimating crownbiomass in 1oblolly
withregression.
P.79-87inEstimating
treebiomass
regressions
andtheir pineplantations
in theCarolinafiatwoods.
For. Sci.28:115-127.
error.Proc.of theWorkshopon Treebiomass
regression
functionsand
theircontribution
to theerrorof forestinventory
estimates,
Wharton, HOn•R,T.G. 1964.The useof heightandsquared
diameterratiosfor the
E.H.,andT. Cuuia(compe.).USDAFor.Serv.Caen. Tech.Rep.NE-117. estimation of merchantable cubic foot volume. For. Chron. 40:324-331.
303p.
Hvscn, B., C.I. M•L•_X,• T.W. B•_as. 1982. Forestmensuration.Ed. 3.
CUBA,T. 1987e.Ontheerroroffore,
stinventory
estimates:continuousforest Wiley,New York.402 p.
mvantory withoutSPR.P.89-98 inEstimating treebiomassregressions
andtheirerror.Proc,of the Workshop on Tree biomass regression JACOBS,
M.W.,Am)T. CV•^. 1980.Useofdummyvariables
toharmonize
tree
functionsandtheircontribution
totheerrorof forestinventory
estimates, biomasstables. Can. •. For. Res. 10:483-490.
Wharlon,
E.H.,andT.Cunia
(comps.).
USDAFor.Serv.Gen.Tech.Rep. JACOBS,
M.W.,rim)D.B.Manturin.1981.Feasibility
ofdeveloping
regional
NE-117.303 p. weighttables.I. For.79:676-677.
CV•UA,
T. 1987f.Ontheerrorofforestinventory
estimates:
Continuousforest JvoaE,G.G., R.C. H•, W.E. GRin:tins,H. Lffrr,sronu,Am)T. Lss. 1988.
inventorywithSPR.P. 99-106in Estimating treebiomass regressions Introduction
to thetheoryandpracticeof econometrics.
Ed, 2. Wiley,
andtheirerror.Proe.of the Workshop on Tree biomass regression New York. 1024p.
functionsandtheircontribution
totheerrorofforest
inventoryestimates,
Wharton,
E.H.,andT.Cunia
(comps.).
USDAFor.Serv.Gen.Tech.Rep. IG•zvs•, N. 1977.Methods forestimatingrootbiomass.
P.25-29inPrimary
NE-117.303 p. productivityof Japaneseforests.
JapaneseIBPSynthesisSeries
Vol. 16,
Shidei,T., andT. Kira(eds.).Univ.of TokyoPress,Tokyo.
CVnlA,
T. 1988.Ontheerrorofbiomass
regressions
andthecorresponding
inventoryestimates.
P.93-109 in Proc.of the9thAnnualSouthern
Forest Kn•ANt•, M.R.,• C.A.TRor•tom. 1981.Therelationship
ofleafareaand
Biomass
Workshop,Daniels,R.A.,W.F.Watson,andI.W. Savelie
(eds.). foliagebiomasstosapwood conducting
areal^foursubalpine
foresttree
Mississippi
StateUniv.,Mississippi
State.216p. species.For.Sci.27:477-486.
CtnnA,
T., Am)R.D.Bvac,
as:1984.Forcing
additivity
ofbiomass
tables:
some YaLzra,
P. 1979.Outlinefora dataprocessing
system
in forestmensuration.
empiricalresults.Can.J. For. Res.14: 376-384. Silva Fenn. 13:368-384.

CV•A,T., A• R.D.BRg•3S.1985a.Forcing
addifivity
ofbiomasstables:
use KL•h'•,C., ^h'oD.R.P•LZ.1987.Subsampling
treesforbiomass. P.225-227
of thegeneralized
leastsquares
method.Can.J. For.Res.15:23-28. in Estimatingtree biomassregressionsand theirerror. Proc.of the

ForestScience
45(4)1999 591
 Workshop onTreebiomass regression
functionsandtheircontribution
to RoaE•s,R., AN• T.M. HINCKLEY.
1979. Foliar weightand area relatedto
the errorof forestinventoryestimates,Wharton,E.H., and T. Cunia currentsapwoodareain oak.For. Sei.25:298-303.
(comps.).USDA For.Serv.Gert.Tech.Rep.NE-117. 303p.
Runre,D.B. 1987.Multipleimputationfor nonresponse
in surveys.Wiley,
KORSMO,
H. 1995.Weightequationsfor determining
biomassfractionsof New York. 258 p.
younghardwoods
fromnaturalregeneratedstands.
Stand.J.For.Res10:
333--346. S^xoo,T. 1982.Forestbiomass.
MartinusNijhoff,TheHague.152p.
S^xoo,T., A•D T. S^ss^.1979.Recherehes
surla biomasse,I'aceroissement
KOZA•C,
A. 1970.Methodsof ensuring
additivityof biomass
components
by et la pertepardessiccation
desracinesdeAbiesfirmaet Tsugasieboldtt
regression
analysis.For. Citron.46(5):402--404. StationdeSylvicultureet deProduction,CentreNationaldeRecherches
Ku•, W.A., S.J.BEUg•_•^,• M.J. AI,I,S.1996.Estimationof rootbiomass Foresti•:resDocumentNo. 79/07, Nancy,France.
anddynamics
forthecarbon
budgetmodelof theCanadian
forestsector. SC•A•R,
J.L.1997.
Analysis
ofincomplete
multivariate
data.
Chapman
and
Can. J. For. Res. 26:1973-1979. Hall, New York. 430 p,
Kv•sgr•,T.O.1985.Cautionary
noteabout
R2.Am.Stat.39:279-285. Sc-•mn•a•,B.E. 1981.Willow oakvolumeandweighttablesfor theMissis-
sippiDelta.USDA For.Serv.Res.Pap.SO-173.14p.
L•,Y., E'rA•. 1996.Modelsof biomass
fornaturalJapanese
redpineforestin
Yanbianarea.J. Jilin For. Univ.12(3):142-146, 163 (in Chinesewith SCmA•aEr,B.E. 1982.Testing,reporting,andusingbiomassestimaUon
Englishsummary). models. P. 95-112 in Proc. of the 3rd Annual Southern Forest Biomass
Workshop, Gresham,C.A. (ed.).BelleW.BaruchFor.
Sci.Inst.,Clemson
LossA•rr,P., n• M. RAil'. 1978.La for•t m6diterran6enne
de chinesverts. Univ., Clemson,SC. 137p.
P. 129-185 in Probl•:mes d'6cologie:Structureet fonctionnementdes
6cosystatuesterrestres,Lamotte,M., and F. Boufii/•re(eds.).Masson, Sc-•mn•a•,B.E. 1984.Sweetgum volumeandweighttables.USDA For
Paris. Serv.Res.Pap.SO-204.14 p.

MAcaWlOt,
H.A.I. 1994.Pinusradiata-biomass,
formandgrowth.H.A.L SC•ODEa,H.T., T.G.G•c, oit•, • G.B.WooI•.1993.Sampling methods
Madgwick,36 SelwynRoad,Rotoma,New Zealand.428 p. for multiresource
forestinventory.Wiley, New York. 464 p.
Sc•ul•El•, H.T., ^•-• W.T. SwA•e.1971.A comparison
of severalstatisUcal
MA•vnc•c, H.A.I., A•) T. SA•OO.1975. On estimatingthe aboveground models in forest biomass and surface area estimation. P. 125-136 tn
weightsof treestands. Ecology56:1446-1450.
Forestbiomassstudies.Proc.of IUFRO Section25, Young,H.E. (ed)
MAalOX,D.A., A•) J.L.F.BA•S'rA.
1996.Sapwood
tapermodels andimplied Collegeof Life Sci.andAgile.,Univ. of Maineat Orono,ME. 240 p
sapwoodvolumeandfoliageprofilesforcoastalDouglas-fir.Can.J.For. Sc•ovE•, H.T., n•m W.T. SWAmC1976. Statisticalconsiderationsin sam-
Res. 26:849-863.
plingbiomass andsurface areaovertimefora Pinusstrobus
L. forestP
M•R, H.A. 1938. The standarderror of estimate of tree volume from 133-141in Oslobiomassstudies.Proc.of IUFRO S4.01,Young,H E
logarithmic
volumeequation.
J. For. 36:340-342. (ed.).Collegeof Life Sci.andAgile.,Univ.of MaineatOrono,ME. 302
p.
MO•SE•D, R.A., A.A. Oh'scm,, Ah• N.M. Tc•r•A•OVA. 1996. Needle,
crown,stem,androotphytomass
ofPinussylvestris
stands
inRussia.
For. Sm•ozAra,K., K. YOD^,K. Hozu•l, • T. Kin^. 1964a.A quantitative
Ecol.Manage.82:59-67. analysis
ofplantform-thepipemodeltheory.I. Basicanalyses.
JapanJ
Ecol. 14:97-105.
No•a•, O., B. Ec•vmo, A•mO. OCSSO•.1995. Non-destructivebiomass
estimationof tree seedlings
usingimageanalysis.Scand.J. For. Res. Sm•ozAra,K., K. YOl•^,K. Hozv•l, • T. Kn•. 1964b.A quantitaUve
10:347-352. analysis
of plantform-thepipemodeltheory.II. Furtherevidenceofthe
theoryanditsapplicationin forestecology.Japan.J.Ecol.14:133-139
PA•O•,J. 1980.Forestbiomass.
For. Abstr.41(8):343-362
S•,r•, J.A.,A• J.K•BROw•.1978.Comparison oftreebiomass
estimators-
PAInt,SOL,B.R. 1993.Modelingmultiplicative
errorvariance:An example DBH andsapwood area.For.Sci.24(4):455-457.
predicting treediameter
fromstumpdimensions inbaldcypress.
For.Sci. S•ow•o•, P. 1985.Alternativesampling
strategies
andregression
modelsfor
39:670-679.
estimatingforestbiomass. Aust.For.Res.15:353-366.
PA•SOL, B.R., J.E. Ho•vr•'r, A• Q.V. CAO.1987.A volumeandtaper Sl,l•O6•, D.G. 1983. Correctingfor bias in log-trausformed
allometric
prediction
system forbaldcypress.Can.J.For.Res.17:250-259. equations.
Ecology64:209-210.
PA•SOC, B.R., A•mC.E. THOMAS.1989.A density-integral
approach to Sg•VASTAVA,
V.K., • D.E.A. G•r•s. 1987.Seemingly
unrelatedregression
estimatingstembiomass.For. Ecol. Manage.26:285-297. Erratum equations
models:Estimation
andinference.MarcellDekker,NewYork
28:321-322.
374 p.
P^gaESOC,
B.R., • C.E. T•o•^s. 1996. A simultaneous
density-integral
So•, S.R.,• G.E.M•r•s. 1988.Measurementof morphological
propcrees
systemforestimating
stemprofileandbiomass:
Slashpineandwillow of treeseedlings
usingmachinevisionandimageprocessing.
Am. Soc
oak. Can. J. For. Res. 26:773-781.
Agric.Eng.Pap.88-1542.16p.
P•t.z,D.R. 1987.Biomassstudiesin Europe--anoverview.P. 213-224 in T^h'vte4ta•ms,
L. 1996.Allometricrelationships to estimateabove-ground
Estimatingtreebiomass regressions
andtheirerror.Proc.of the Work- dry-massandheightin Salix.Stand.J. For. Res.11:233-241.
shoponTreebiomass regression
functions
andtheircontribution to the
errorofforestinventory
estimates,
Wharton,E.H.,andT. Cunia(comps.). T^•^s, M.A. 1967.Weightscaling:Its past,presentandfuture.Faculty
USDA For. Serv.Gen.Tech.Rep.NE-117. 303 p. of For., Univ. of Toronto, Wood Measurement Conf. Proc. Tech
Rep. No. 7.
P'•rOSl•¾,B.J.(•D.). 1998.Proceedingsof thefirstinternational
conference
on geospatial information in agricultureand forestry.ERIM Interna- TAS•SSA,
G., ta• H.E. BURKHART. 1998.Modelingthinningeffectson ring
tional,AnnArbor,MI. 1348p. specific
gravityofloblollypine(PinustaedaL.).For.Sci.44(2):212-223
IMnal,,M.S. 1994.Measuringtreesandforests.Ed. 2. CAB International, T•O•AS,C.E., B.R. PAIntESOt,
Am•K.H.N. L•. 1995.Biomassandtaperfor
New York. 336 p. treesin thinnedandunthinned
longleafpineplantations.
South.J. Appl
For. 19:29-35.
PEnn,D.D., H.O. Lincre'v,E.A. Joi,•r_s,
• Y. Z•-u•a. 1996. Above- and
belowground
drymatteraccumulationpattern
derivedfromdimensional USoL'TEv,
V.A., • J.K.VANCLAY. 1995.Standbiomass
dynamics
ofpme
biomass
relationships.
For.Sci.42:236-241. plantations
andnaturalforests
ondrysteppe
inKazakhstan.
Stand.J.For
Res. 10:305-312.
RomCUA•V, E., aNDIV•T•V•, I.R. 1992.Theapplicabilityof thepipemodel
theory for the predictionof foliage biomassin trees from natural, VAtsre'mE,H.T., • S.J. Hm?on.1977. Samplingoak foliageby the
untreatedblacksprucestands. Can.J. For. Res.22:1118-1123. randomized-braoch method. Can. J. For. Res. 7:295-298.

592 ForestSctence
45(4)1999
V•h'ru•, H.T.,L.M.Tal'rlx•,• G.M.Fvgmvn•.1984.Subsampling
trees ForestBiomass
Workshop,
Baldwin,V.C., Jr.,andR.E.Lohrey(eds.).
for biomass,volume,or mineralcontent.For. Sci. 30:673-681. USDA For. Serv.,South.For. Exp. Sta.,New Orleans,LA. 131p.
Vnn D•vssn,P.C. 1997.Annualforestinventorystatistical
concepts
with WILLIAMS, M.S., A•DT.G. G•C,Oll•. 1993.Estimating
weightswhenfitting
emphasisonmultipleimputation.
Can.J. For.Res.27:379-384. linearregressionmodelsfortreevolume.Can.J.For.Res.23:1725-1731.
VnnD•Jssn,P.C.,• V.C.Bnt•wm,JR.1993.Sampling
andpredicting
tree WILLIAMS,
M.S., • H.T. SC•l•t•D•. 1996.Prediction
of grosstreevolume
dry weight.Can.J. For. Res.23:1826--1829. usingregression
models
withnon-normal
errordistributions.
For.Sci.
42(4):419-430.
Vnn D•vssn,P.C.,A.D. S•VAn, • T.G. Mn•,•Y. 1981.A prediction
systemforcubicfootvolumeofloblollypineapplicable
through
muchof YA•DL•,D.O.,A•DH.V. Wlat,rr. 1981.Biomassestimates
basedonthesimple
itsrange.South.J.Appl.For.5:186-189. powerfunction--somecalculation considerations.
P. 350-354 in Proc.
In-placeresourceinventories: Principleandpracticesworkshop. Soc.
Wnno, I.R., A.L. Z•ono, P. COmau, M. Tsz•, n• J.P. Fat•MmS.1995.
Am. For.,Bethesda,MD. 1101p.
Aboveground biomass
andnutrientaccumulation in anagesequence
of
aspen(Populustrernuloldes)
standsin theborealwhiteandblackspruce Z•m•,, B. 1987.Areasof biomassresearch. P. 193-196 in Estimatingtree
zone,BritishColumbia.For.Ecel.Manage.78:127-138. biomassregressions and their error. Proc.of the Workshopon Tree
biomass regression
functionsandtheircontribution to theerrorof forest
WARIN•,R.H., n• S.W. Rt•n•l•. 1998.Forestecosystems:
analysisat inventoryestimates,Wharton,E.H., andT. Cunia(comps.).USDA For.
multiple
scales.
Ed.2. Academic
Press,
SanDiego,CA.370p. Serv.Gert.Tech.Rep.NE-117. 303 p.
WAging,
R.H.,P.E.SCaRO•D•a, n• R. O•n. 1982.Application of thepipe Zaot•, G., •r •. 1997. Biomassand nutrientallocationin Manglietia
modeltheoryto predictcanopyleafarea.Can.J. For.Res.12:556-560. hainanensis plantation
ecosystem
at Jianfengling.
For.Res.10(5):453-
W•x, G.B.,J.H.BRown, • B.J.EnQ•sx.1997.A generalmodelforthe 457(inChinese
withEnglishsummary;
Published
byChinese
Academy
originof allometric
sealtag
lawsin biology.Science
276:122-126. of Forestry,Beijing).

W•in•s, J.G.,JR.1982.Modelingproblems
inpredicting
total-tree
andtree-
componentbiomass.P. 111-115 in Proc.of the 4th Annual Southern

ForestScience
45(4)1999 593