Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
of predictions,althoughthe two-dimensionalrepresentation
sent channeland floodplainareasat different spatialresolu-
may improvehydraulicpredictionsat a point. This disparity
tions appropriateto local processgradientsand concentrate
betweenthe dimensionalityof hydraulicmodel structureand
computationalresourcesin regionsof complexflow. As the validation data must be addressed.
modelingof a dynamicflood shorelineis a movingboundary
Three techniquesfor the investigationof the validation
problem,recentdevelopments in algorithmsfor the treatment
problempresentthemselves. The modelcanbe testedagainst
of domain wetting and drying, and the hydraulicprocesses
analyticalsolutionsof the governingequations[Horritt,2000;
Copyright2000by the AmericanGeophysical Union. Batesand Hervouet,1999]. This teststhe numericalability of
Paper number2000WR900206. the model to deliverapproximatesolutionsof theseequations,
0043-1397/00/2000 WR900206509.00 while avoidingthe issueof processrepresentation.This ap-
3279
3280 HORRITT: CALIBRATION OF A 2-D FINITE ELEMENT FLOOD FLOW MODEL
Spuriousfree surfaceslope ents betweenwet and dry nodes (Figure 2a), but these are
eliminatedin the model and replacedwith a zero free surface
(b) h <0
slope(Figure2b). The heightderivativeterm in the continuity
equation(2) [Definaet al., 1994;Batesand Hervouet,1999]is
alsoscaledto reflectthe volumeof the partiallywet element
availablefor flow, but this will not affect the steadystate
solutionsdevelopedhere.
A finite elementmesh(Figure 3) for the modelhas been
developedusingthe "cheesymesh mesh generator"[Horritt,
1998]. A curvature-dependent discretizationstrategyis used
for the channel,which has been found to give good results
when comparedto analyticalsolutionsof meanderingflow
Computationalnodo [Horritt,2000].This useselongatedstructuredelementsin the
Figure 2. (a) Partiallywet elementsmaygeneratea spurious channel,which allow the model to representcross-channel
water free surfaceslopenear the shoreline.(b) This is cor- velocityand depth gradientswith fewer elementsthan would
rectedfor by cancelingthe Vh term in the momentumequa- be required using a more equilateral discretization.Shorter
tion, equivalentto fixinga horizontalfree surface. elementsare used in regions of high streamlinecurvature,
which promotesthe accuraterepresentationof the velocity
advection term with a minimum number of elements, but this
the simplestmethod of turbulenceclosure(this is discussed will still fail to deal with dispersionterms generatedby the
further below). The TELEMAC-2D model uses Galerkin's depth-averaging processoverthe deeperflowsin the channel.
method of weightedresidualsto solve(1) and (2) over an The generator aims to produce approximatelyequilateral
unstructuredmeshof triangularfinite elements.A streamline floodplainelements,whosesizeincreases linearlywith distance
upwindPetrovGalerkintechniqueis usedfor the advectionof from the channel. The variable mesh resolution raises issues of
flow depthto reducethe spuriousspatialoscillations in depth representationof the shorelinefeaturesat differentscalesand
that Galerkin'smethod is predisposedto and the method of the resolutionof the observeddata, but at this stageit is the
characteristics is used for the advectionof velocity.An ele- large-scalestructureof the inundatedarea that is of interest,
ment-by-elementsolveris usedto solvethe nonlinearsystem, and this shouldbe fairly well representedby the meshused
and the time developmentof the solutionsis dealtwith using here. An inflow control structureis added at the upstream
an implicitfinite differencescheme.For thisstudy,onlysteady boundaryto allow the flow to developfrom any instabilities
state solutionsare sought,so the time stepis usedas a surro- imposedby the boundaryconditionsbefore enteringthe main
gatefor the developmentof an iterativescheme,wherebythe bodyof the mesh.The main difficultyin meshdiscretization is
model is "woundup" to a stablestate.Dynamicsolutionsare a pragmaticone:The floodplainelementsizenearthe channel
not investigatedhere, giventhe slowlychanginginput hydro- is determinedby the channelelementlength,andgivena linear
graph.The movingboundarynatureof the problemis treated increasein elementsize awayfrom the channel,this fixesthe
with a simplewettinganddryingalgorithm(Figure2). Partially minimum total number of elements. This mesh contains 17,000
wet elementscan generatespuriouslyhigh free surfacegradi- elements and is the smallest feasible with the constraint of
4000
2000
Figure 3. Finite element mesh for the Thames model, with detail showingcurvature-dependentchannel
meshing.
HORRITI': CALIBRATION OF A 2-D FINITE ELEMENT FLOOD FLOW MODEL 3283
,-;.•,.::..•
,:.'-•.•.•
"'.... . Elevation
4000
•. /2
,./ .
,_...
'•'-•.•(m
above
80
sea
level)
.......
,•....,•....... -•":••:i•:
":'•.......• .....
•:,.. .........
••; .... .•
:•-• .•.....
.......
..•.• ;:•..
_5:'½
'%;.::.:'½"•
76
';:?" ': :';.':0
'• =:,.'.,
,..::,, ,,,., ,,,,,----,...
, .:;;:.:;;;;/
72
"• "•'•""
....
' ........
; :''•" '" ' ' ::"-'- '"" ß '; '" •'•'' 68
2000
ß
z'*
.• ....
.;:
..,,;;
...'"':'"•-..;;;'
......
':........
.'":,.
,7.-.'
;,½..
_,:..'.:•-
•
'z,,,..'....•
.....
-:.:;;..:..•.-•-•
....
;,..."..•,.
'.•
......
-...;•
•
:::,•,,. ,:'-.
'.',•'" , .. ,.....':•:.
-;..:.:.:
.....
'..' ....... .::•',;• 64
• 2m
contours
o .......................... / .......
o ,ooo ooo 6oooooo ,oo%,,ooo,,o m
68
67
66
65
64 .•½"
..,•.:,"
channelwidth, channelelementaspectratio, and a reasonable provided by a rated sectionat the top of the reach, which
transitionfrom smallelementsnear the channelto larger ones providesmeasurementsof flow at 3-hourlyintervals.This fig-
near the domainboundary. ure shouldbe interpretedwith caution,however,as no error
Topographicinformationis providedby a combinationof informationis available,and this error will probablyincrease
channelcrosssections(every 50 m) and a stereophotogram- for highflows. A constant valueof 73m3 s- • (thevalueat the
metric digital elevationmodel (DEM) with a resolutionof time of the ERS-1 overpass)is usedfor the steadystatesim-
50 m and a heightprecisionof _+25cm (Figure 4), sampled ulations.The free surfaceheight at the downstreamboundary
onto the computationalnodesof the mesh using a nearest- is determinedby projectingthe SAR shorelineonto the air
neighborsscheme.This topographicrepresentation,however, photographDEM. The effect on model predictionsof these
wasfoundto generatemodel stabilityproblems,and a there is two conditionswill be fundamentallydifferent.While the mag-
a trade-offto be madebetweenthe accuraterepresentationof nitudeof the upstreaminflowwill affectthe wholereach(since
the bathymetryand the promotionof stable(and physically massof water is conserved),we would expectthe effect of the
realistic)modelsolutions.Theseproblemsmanifestthemselves downstreamboundaryconditionto diminishfarther upstream.
as large cross-channel velocities,generatedby the high bed A simpleanalysisof the equationfor one-dimensionalsteady
slopeson the banksof the channeland alsogeneratedwhere flow vindicatesthis intuitiveresult.With flow depthh, depth-
the channelrunscloseby regionsof steeprelief (alsoshownin averagedvelocityv, andx the downstreamdistance,thiscanbe
Figure 4). This problem is solvedby first droppingall bank written [Chow,1988] as
nodesby 50 cm andby ensuringthat the channelisborderedby
at least one reasonablyflat element. This modification of
modeltopographyis fairly arbitraryand is basedon pragmatic
dx 5 v2 +g•-gSø+•-• =0' d(1) dh gv
2 (4)
arguments,but it can be viewed as sacrificingtopographic
So representsa uniform down-reachbed slope, and C is
accuracyfor the sakeof model solutionstability.While chan-
Chezy'sfriction coefficient.If the x derivativesare eliminated
nel/floodplaininteraction has been identified as important
(uniformflow), a flow depthof 2.2 m and a velocityof 1.66 m
[KnightandSh•ono,1996],we havealreadysacrificed
an accu-
s- zgivesa discharge
of 73m3 s- • overa channel
20m inwidth
rate topographicrepresentationof the channel region by
withChezy's C - 40 mm s-z. Addingsmallperturbations to
adoptinga relativelycoarsemesh(only three elementsacross
this uniform solution,so that h = ho(1 + K(x)) etc., while
the channel).How this affectsthe representationof hydraulic
keeping vh a constantand linearizing,we gain the following
processes in the channelbank region has yet to be assessed,
solutionsfor K(x):
and it is unclear how accuratea representationof channel/
floodplaininteractionis possibleat this resolution. 3gvo
Provisionof an imposedupstreamvolumetricflow rate and • =Ae•x, a = C2ho(gho
_ 2%2). (5)
an imposedfree surfaceheight at the downstreamend of the
reach are adequateto constrainthe flow accordingto the A is a constantgivenby the sizeof the perturbation(and the
methodof characteristics.The upstreamboundaryconditionis lengthof the reach), and 1/a is the characteristiclengthover
3284 HORRITT: CALIBRATION OF A 2-D FINITE ELEMENT FLOOD FLOW MODEL
......•,•j•
1.5
200-
'••0.5
0
100'
Flow Velocity
ß 1 m/s
o
o 1 oo 200 300 400 500 600 700 800
2OO
lOO
o
o 1 oo 200 300 400 500 600
which perturbationsin depth imposed at the downstream [Sellinand Willets,1996]and thereforeto someextentlumped
boundary will decay upstream, and for this situation a • into the bed frictionterm. Thisvaluemayhaveimplicationsfor
1000 m. We would therefore expectthe effect of the down- other termsin the momentumequationsuchas the advection
stream boundary condition to extend only a few kilometers term, which has been shownto be importantfor meandering
upstreamand not to significantlyaffectthe wholeof the reach.flows [Hordtt, 2000], and the issueof channelprocessrepre-
However, we should bear in mind that the accuracyof the sentationmay be cloudedby the value used.
solutionsnear the downstreamboundary may be a stronger Beforethe provisionof frictionvaluesis discussed in section
functionof thisboundaryconditionthan of model formulation 4 aspart of the calibrationproblem,we'first examinesomeof
or calibration.The effect of downstreamboundarycondition the generalpropertiesof the modelsolutionsand somepoten-
will, of course,be different for dynamicsituations,and its tial sourcesof error. One possiblesourceof the differences
influencemay extendfarther upstream,as wavesmay propa- betweensimulationscould be causedby varyingconvergence
gate throughoutthe reach as flowsare generallysubcritical. properties,rather than differencesin the steadystate flows
Provisionof model boundary conditionsand topography themselves.Since at steady state the outflow at the down-
leaves the bed friction coefficientsand the turbulent eddy streamendof the reachshouldbe equalto the imposedinflow,
viscosityas unconstrainedparametersin this case.The eddy the differencebetweeninflow and outflow(i.e., the rate of
viscositywasfixedat a valueof 1.0m2 s-• (themedianof the changeof volume in the domain,neglectingmassbalanceer-
rangeof valuesusedby Bateset al. [1998]),a (perhaps)artifi- rors) is a usefulcheckof model convergence. All simulations
ciallyhighvalue but one that tendsto promotemodel stability were run until the magnitudeof this differencewaslessthan 1
in caseswhere numerical oscillationsare likely. High bank ms s-2,andthiswasfoundto ensure adequate convergence of
slopescan generate instabilitiesin the channel region, and the solutions. There is one caveat: Some simulations exhibit
guidelinesfor acceptablevalues are discussedby Bates and dampedoscillationsin this observable,so the flow difference
Anderson[1993], a lateral slope of 0.03 being given as the mustbe lessthan1 m3 s- • andbe decreasing, andthe rateof
maximumacceptable.This value is exceededin this casebe- changemust also be decreasing.This thresholdis essentially
causeof the channelgeometryand meshresolution,but this is arbitrarybut is probablylessthanthe error in the imposedflow
to someextentcounteractedby the highvalueof eddyviscosity rate (takenfrom the rated sectionat the upstreamend of the
used.This is a fundamentalproblemwith somefinite element reach),andsoconvergence errorsare likelyto be smallerthan
codes,and the use of an artificiallyhigh eddy viscosityis a thoseinducedby uncertaintyin the upstreamboundarycondi-
simpleway of overcomingthe problem.Previousstudieshave tion.
shownlittle sensitivity
to thisparameter[Bateset al., 1996]in A visual check on predictedwater levelsand velocitiesis
comparisonto bed friction, and this has been confirmedby a usefulin detectingphysicallyunrealisticflowsgeneratedby the
preliminary sensitivityanalysis.The effect of turbulent mo- model (like the spuriousvelocitiesgeneratedby topography
mentumdiffusioncanalsobe modeledasenergylossprocesses mentionedin section3). Figure5 showspredictedwaterlevels
HORRITT: CALIBRATION OF A 2-D FINITE ELEMENT FLOOD FLOW MODEL 3285
and velocitiesfor a number of regionsof the model including someinterestingflow processes not apparentwith more tightly
severalfeatureswe might expectto causeproblems:a bifurca- constrainedfriction values.This simplisticassumptionof uni-
tion, a transition between structured and unstructured ele- form frictionsoverthe channelandthe floodplainis unrealistic
ments,and regionsof high channelcurvature.The water sur- but does make for a manageablecalibrationproblem. One
face generatedis relativelysmooth,and the velocitiesare all strategywouldbe to perform simulationsat a numberof points
downstreamand approximatelyparallelwith the channel.Wa- spanningthe completeparameterspace,but giventhe compu-
ter is also flowing from the channelonto the floodplain and tationallyintensivenatureof the modelingproblemon thissize
crossingover some meanders, indicating that the model is of mesh, a more informed approachis adopted.We first in-
representingover-bankmassand momentumflow in the near- vestigateparameterizationswhere channelfriction coefficient
channelregion.Meanderingflowsin simplechannelstypically is generallydoublethat for the floodplain,and then we inves-
exhibitlowervelocitieson the outsideof bends[Horritt,2000], tigate the individualroles of the two parameters.Some mea-
and there is someevidencefor this in Figure 5, althoughthe sureof fit is alsorequiredto comparethe resultsof the simu-
situationis complicatedby the coarse-mesh resolution,channel lations and the SAR-derived flood map, and the calibration
topography,and interactionwith floodplainflows. processusedhere is to searchthroughthe parameterspacefor
A measure of mass conservation is another useful check on a global extremumof some objectivefunction, basedon the
the validityof model resultsfor two reasons.First, it is easyto correspondencebetween the model predictionsand the ob-
verify (comparedwith conservation of momentum),and, sec- serveddata. The choiceof this objectivefunction is a crucial
ond, it is more likely to be satisfiedby the real-worldsystem part of the calibrationprocess,as its form maywell affect the
under consideration.Most of the errorsin model processrep- optimum calibrationresult.
resentationoccurin the momentumequation(friction,turbu- Given that the flood boundaryis extractedfrom the SAR
lence, numericaltreatment of advection,etc.), whereasthe imagery,someform of regionalor area-basedmeasureof fit
continuityequationis more likely to be correct(althoughin- would seem to be an obviouschoice, but the exact formulation
filtration processesmay represent a loss of water from the of sucha measureis problematic.If, for example,the fraction
surfacesystem,and errorsin the imposedflow rate may mean of the model domain for which the model predictsthe flood
we are trying to conservethe wrong quantity).A measureof state (flooded or unflooded)correctlyis used, bias may be
mass balance error as a function of time, rather than over a introduced.A verylargemodelcontaininga smallfloodedarea
whole simulation,is useful,asit enablesus to ignoreregionsof will generallyappearto be successful, as it is easyto predict
(understandably) high error, suchas at the start of the simu- correctlythe large dry regions,generatinga goodfit between
lation where the initial conditionsare highly unphysicalor model predictionsand observations.A similar bias will occur
when friction coefficients are changed suddenly. A time- for a flood that almostfills the model domain.Fortunately,in
variant measurecalculatedat eachtime stepis givenby thiscasethe floodoccupies48% of the modeldomain,sobias
introducedby theseconsiderations shouldbe minimal, but a
bias-free measure will be useful when results from different
Qe
....(t) -- Qin-Qout-A-•-' (6) reachesand eventsare compared.Even so, an estimateof the
fraction of the domain correctlyclassifiedwill still be usefulin
This is the discrepancybetween the net flow entering the
the contextof practicalflood management,even if the bias
domain(Q inandQ outbeingthe volumetricflowsin andout of introduced clouds the calibration issue.
the reach,respectively)and the rate of changeof volumeV in
Area-based statisticsfor the model simulationsare gener-
the domain over the time step zXt,which givesa flow error
ated (Table 1), where eachsimulationis associated with four
Q erroras a functionof time. For all simulationsused in this
areavalues,expressedas a fractionof the model domainarea.
study,IQe•o•lwasfoundtobelessthan1 m3s-l, whichisagain Classes1 and 2 representareas correctlyidentified by the
probablylessthan the error in the imposedflow rate, and this
model, whereas classes3 and 4 represent misclassifications.
indicatesthat the solutionsare adequatelymassconservative.
The modelhascorrectlyidentified-75% of the domain.Pre-
The continuityequationitselfis subjectto processes represen-
dictingthe domainto be either completelydry or completely
tation errors,omittingthe effectsof groundwaterflow,rainfall,
inundatedwould give a figure of around 50%. This value is
and runoff from boundinghillslopes,and is thereforeunlikely
relatively insensitiveto different friction parameterizations,
to be satisfiedby real floodplainflows.Mass balancefigures
but the areas of classes3 and 4 do vary considerablywith
can, however,be a useful indicationof numericalproblems,
changesin friction,with a steadyshift from overpredictionto
and this analysisshowsthat they are an insignificantsourceof
underpredictionas friction parametersare increased(bed
uncertaintyin the model solutionswhen comparedto uncer-
sheardecreased).We mightexpectsomeintermediateparam-
tainty in friction parameterizationand boundaryconditions.
eterizationto exhibit a strongmaximumin the area correctly
predicted,but thisis not evident;asin simulations3 and 6, for
example,thereis considerable randomerror despitethe lackof
4. Model Calibration and Validation
net overpredictionor underprediction.Figure 6 showsthe
The main unconstraincd
parametersin the modelingprob- shorelinesfor simulations3 and 6, demonstratingthat the two
lem are the friction coefficients, and these are now used as differentpredictedshorelinescorrespondequallywell with the
calibrationparameters.Simulationswere performedwith val- SAR shoreline.An obviousquestioniswhetherthe differences
uesof Chezy's
C in therange20-150m1/2s-• for thechannel in modelpredictionsare due chieflyto sensitivityto floodplain
and5-75 m•/2s-• for thefloodplain.
Whilethesevaluesspan friction, channel friction, bank friction, or a combination of
a greaterrangethan that usedin previousstudies[Bateset al., these(the simulations
of Table1 usedthe samevaluesfor bank
1998]and indicatedin manualsof hydraulicdesign[e.g.,Chow, and floodplainnode friction). The sensitivityto theseparam-
1959],they do representa full transitionbetweenoverpredic- eters is explored in Table 2, which showsthat the channel
tion and underpredictionof inundationextentand mayreveal friction parameterhasrelativelylittle effect on the predicted
3286 HORRITT: CALIBRATION OF A 2-D FINITE ELEMENT FLOOD FLOW MODEL
aThe classesrepresentthe possiblecombinations of flood stateas found in the syntheticapertureradar (SAR) imageryand by the model
(wet-wet(ww), dry-dry(dd), etc.). Valuesare expressed as a percentageof the modeldomain.
inundatedarea (simulations7 and 9), whereasfloodplainand be partly a functionof the modifiedbank topographyusedto
bank friction exert a strongerinfluenceon shorelinelocation. promotenumericalstability,and a more realisticrepresenta-
The responses due to perturbationsin floodplainfriction are tion of channel/floodplaininteraction may reveal a greater
roughlydoublein magnitudecomparedto thosedue to bank sensitivityof floodplainflowsto channelhydraulics.It is inter-
friction. This is not reflectedwell in the area correctlyidenti- estingto note that the model givesreasonableresultseven for
fied as floodedbut is more obviousin the changesin classes3 simulation6, which usesextremelylow valuesfor bed shear.
and 4. This is potentiallyuseful,as it effectivelyreducesthe This couldbe dueto the interactionbetweenthe relativelyhigh
dimensionalityof the parameterspaceinvolvedin the calibra- value of eddy viscosityand low bed shear, and simulations
tion problem:We haveto varythe floodplainandbankfriction, usinga lower value for the eddyviscositymay exhibitgreater
but the valueof channelfrictionis relativelyunimportant.The sensitivityto friction as the bed shear is decreasedat the
functionaldependenceof friction term on the value of C will expenseof numericalstability.
have a role in this sensitivity,but a normalizedsensitivity(1/ The area correctlyidentified can also be evaluated as a
F)(OF/OC) will be proportionalto 1/C (the casefor anypower function of along-reachdistanceto give a one-dimensionally
law dependence),and this is reflectedin the choiceof larger distributedmeasureof fit, whichis shownin Figure 7 for three
perturbationsfor larger valuesof C. Channel friction may of the simulations listed in Table 1. The surface shows a num-
becomemore important for different flowswhere hydraulic ber of interestingfeatures.The correctlyidentifiedareavaries
processes in the channelmay be more influential.It may also along the reach, as does the sensitivityto friction. The most
4OOO
2000
0
0 20'00 4600 60'00 80b0 ' ' '•'0600' ' '12600' ' 'i4600"
4OOO
2000
0
0 2600 4600 6600 80'o0 • 0600 126oo 14600'
100
1 '
<)--(>C•= 5.0m•/•s -• , A comparisonbetween model predictionsand the aerial
photographyis also useful, as it is unclearhow much of the
90•- A-• • C,,=57.5m'/•s-' _ discrepancy betweenmodelandSAR shorelinesis due to error
in the model or error in the flood mappingprocedure.Aerial
photography
is available
for approximately
halfof thereach,
8o andthiscanbe usedto generatemeasures of fit for the model
shorelineswhencomparedto both SAR and aerial data for the
[2 70 subdomaincoveredby the aerial photography.Forty-nineper-
cent of this subdomainis flooded, so again the correct pre-
dicted area should make for an unbiased measure of fit. The
60 •
resultsfor three model simulationsare givenin Table 4, along
with a directcomparison betweenSAR andaerialdata,andthe
50 ............ main results are similar for all simulations. The measures of fit
0 5 10 15
are approximately3.5% lessthan for the SAR imagery,which
Along-reachdistance (km) canbe partlyexplainedby the 2.6% overpredictionof flooded
Figure ?. •e•ce•ta•e of modeldomai• classifiedcorrectlyby area apparentwhenthe SAR floodmap is comparedwith the
the model,asa •ctio• of •oodplai• frictio• pammetedzatio• aerial data (also shownin Table 4). If we regard the area
a•d alo•-•eack distance.•ke •esuitso[ tk•ee simulationso•]y predictedincorrectlyby the modeland the SAR flood map (as
am skow• [o• cladS. comparedwith the aerialdata) asa measureof the error, then
3288 HORRITT: CALIBRATION OF A 2-D FINITE ELEMENT FLOOD FLOW MODEL
the error in the modelpredictionsEmod (29% for simulation3) resolutionand streamlinecurvaturecan have a large effect on
is approximately
givenbyEmod
2 ----ES2^R 2 SAR,
+ Emod, where predicteddepths,and if thiswasto be compensatedfor in the
Es^R(18%)andEmod, SAR(23%)aretheerrorsbetween
the calibration process,it would result in an optimum friction
SAR and aerial data and between the model and the SAR
coefficientwith a strong correlationwith channel sinuosity.
data, respectively.Two conclusionscan be drawn from this Although this is partially evidentin this study(a particularly
analysis.First, there is only a weak correlationbetween the low value for bed shearis required in the 7- to 9-km region
errorswhen the model is comparedwith SAR and aerial data, wherechannelsinuosity
is high),the uniformchannelelement
and so model errors and observation errors are uncorrelated.
curvaturemesh generationstrategyused here shouldensure
Second,it showsthat the total model error (betweenmodel that errors in predicteddepthsdue to this phenomenonare
and "ground" truth) can be (approximately)equally parti- <20%. This would require a changein Chezy'sC of around
tionedbetweenobservationerrors(betweenSAR and ground
30% to compensate,whereasthe optimum friction valuesare
truth) andmodelingerrors(betweenmodelandSAR), but the
seen here to vary by a factor of --•7 along the reach. This
modelingerrorsare larger than the observationerrors.
indicates(togetherwith the apparentinsensitivityto channel
friction) that althoughthis may be part of the explanation,
5. Discussion other sourcesof variation are present.These could be more
The resultsindicate that spatiallyvariable friction coeffi- numerical errors (poor solutionsof the shallowwater equa-
cients are required to improve model performance, but tions) or processrepresentationerrors (hydraulicprocesses
whetherthe variationssuggested by this calibrationstudyare a poorlyrepresentedby the shallowwater equations).For exam-
result of genuine heterogeneoushydraulicpropertiesor an ple, highchannelsinuositycouldinteractwith the highvalueof
artefactof the modelingprocessremainsunclear.Validation of turbulent eddy viscosityto increasewater levelsin the 7- to
the model usedin this studyagainstanalyticalsolutionsof the 9-km region,and thismay requirean artificiallylow bed shear
shallowwater equations[Hordtt, 2000] indicatesthat mesh to compensate.The variationsmay be a genuinereflectionof
4OOO
2000
0
0 20'00 40'00 60'00 80'00 10000 12600 14600'
4000
2000
0
0 ' 20'00 40'00 60'00 80'00 10600 12600 14600'
4000
2000
4000
2000
Figure
9. (top)
Distributed
friction
and(bottom)
simple
plan•r
freesurface
model.
floodplainroughness, and giventhe obviousheterogeneity of parts of the domainwould indicatethat the differencesin
floodplainvegetation,thiswouldseemto be the obviousline of shorelinelocationare due to genuinechangesin water levels
enquiryfor futureresearch.While the effectsof frictioncoef-rather than the random errors that would be expectedfrom
ficientshave been explored,the inundatedarea may alsobe uncorrelatedtopographicnoise. Again, this points to poor
affectedby the bed shear-velocity-depth
relationshipused.For friction parameterization(whichwill affect water levels) as
example,usingManning'sequation(frictiontermproportional causingmost of the model error. The influenceof the mor-
toh-4/3) leadsto bedshearthatfallswithdepthmorerapidly phologyof the near-channelregion is alsoworthy of future
than Chezy'slaw. There is also evidencethat bed shearis a investigation,as it will be expectedto exert an effect on chan-
functionof Reynoldsnumber[Falconerand Chen, 1996],and nel/floodplain interactions,especiallyi.nthe light of the mod-
thereforea more sophisticated bed shearformulamaybe use- ificationsof the bank topographymade here.
ful in the spatiallyheterogeneous
flowsoverthe floodplain.A The dynamicbehavior of these models also needs to be
less simplisticapproachmay also be required for dynamic investigated (thisstudyis limitedto steadystatesimulations).
simulations,where different hydraulicconditionsexistduring For example,if we use a dynamicflood hydrographas the
the wetting and dryingphasesas the flood wave propagates model upstreamboundarycondition,it is at presentunclear
downstream. how much of the hydrographhistorymust be modeled, and
Topographyalsoexertsa stronginfluenceon the inundated suchquestionsof temporalcorrelationsin modelpredictions
area, and errorsin the predictedshorelinemaybe (partially) must be addressed.Given likely errors in the provisionof
due to noisein the DEM and interactionsbetweentopography boundaryconditions(whichmaybe largefor a ratedsectionat
andthe finite elementmesh(relativelycrudenearest-neighbor overbank-fullflow),we alsoneedto investigate theireffectson
samplingwasusedto samplegroundheightsonto the mesh). modelpredictions.The passageof a floodwavedownthe reach
The systematicunderpredictionand overpredictionseen in wouldbe expectedto manifestitselfasa convexwater surface,
as opposedto the concavesurfaceseenhere. We would also observationarbitrarilywell usinga more heterogeneous fric-
expecthydraulicconditionsover the reach to be reasonably tion parameterization,but thiswould do little for the model's
well representedby the steadystatesolutionusedhere. Long- predictivecapabilityand providelittle informationon the im-
wavelengthwaveswill take approximately 1 hour to propagate portanthydraulicprocesses, which are the originalreasonsfor
alongthe reach(lesswhen advectionis takeninto account), modelingthe problem.By adoptingcalibrationtechniques, we
and for the rate of changeof the input hydrograph(of the are thereforenot makingthe mostof the informationpresent
orderof 10-4 m3 s-2) thisrepresents a change of only0.3m3 in the satelliteimagery.Introducingmore data (manyflood
s-• in flowbeforeeffectshavepropagated out of the reach. eventsover many reaches)into the calibrationprocessmay
The slowrate of changeof the hydraulicconditionsalsojusti- help in identifyingfeaturescommonto floodplainsystems and
fiesthe assumption that the air photographandSAR datawere henceincreasethe model'spredictivepower.However,given
effectivelysimultaneous, when,in fact, theywere separatedby the transient nature of flood events combined with the rela-
2 hours. Previousstudiesusing point hydrometricdata for tivelypoor temporalcoverageofferedby currentsatelliteSAR
validationand calibrationhave generallyperformeddynamic missions, thismaybe impossible.The alternativeis to abandon
simulationsin order to make use of the availabledata (zero- (to some extent) the calibrationmethodology,breakingthe
dimensional in spaceand one-dimensional (i-D) in time), but link betweenparameterizationand validationdata, and pro-
dynamiceffectsare unlikelyto have a large effect on model vide independentinformationon floodplainfrictionalproper-
predictionsfor this reach and event. Effectswill be greater ties (whetherthroughremotesensing,groundsurvey,labora-
when modelinga narrow hydrographover a long reach and tory measurements, etc.). In additionto modelimprovements,
maythen becomeimportant,and the effectsof massand mo- further work is also requiredto improvethe satellite-derived
mentum conservationerrors may be exacerbatedover long floodmapsif furthermodeldevelopments are to be adequately
dynamicsimulations. evaluated.
The problemof dealingwith uncertainty(in processrepre-
sentation,parameterizationdata, and validationdata) in the
modelingprocessis a difficultone. One of the simpleststrat- 6. Conclusions
egieshasbeen adoptedhere,whereit is assumedthat there is This work has servedto demonstratethe use of remotely
a singleoptimalfrictionparameterization (whetheruniformor sensed data for the calibration and validation of two-
distributed)whichminimizesthe discrepancy betweenmodel dimensionalflood flow models.The model has provedto be
predictionsand observations. Given the uncertaintiespresent relativelyrobustwith respectto friction parameterizationfor
in the data, it may be more usefulto work with a far greater this reach and event when an area-based measure of fit is used
number of simulationsand describeeach parameterization to compareits predictionswith the SAR imagery,but an op-
with a likelihoodfigurebasedon the fit betweenobservations timum friction value is still evident.Spatiallyvariablemodel
and predictions,rather than definingoptimumvaluesof each sensitivityand predictionaccuracypoint to a distributedfric-
parameter.This is the aim of the generalizedlikelihoodun- tion parameterization,whichperformsaswell asthe bestuni-
certaintyestimationprocedure[Bevenand Binley,1992], and form friction simulation in terms of the measure of fit, and also
the procedurehasbeenusedto assess the effectsof uncertainty capturessome other floodplain flow processes.Most finite
in flood flow models[Aronicaet al., 1998;Romanowicz et al., element simuiationsperform betterthana simple planar
1996].While this is certainlya rigorousand usefulway to go model of the water free surface,which tendsto greatlyover-
forward,the need for a largenumberof simulationsto covera predict inundationextent.
parameterspaceof high dimensionality is a disadvantage, es- This studyhas providedonly a very limited validationof
peciallywith complexfinite elementcodeswhich are compu- model performance,as is to be expectedwithout accessto a
tationallyintensive.The simplerapproachadoptedhere relies validationdata set independentof that used for calibration.
on a relatively small number of simulations,and while the The resultsdo indicate,however,that the modelis capableof
effectsof uncertaintyare not fully explored,it does allow a reproducingsignificantfeatures of the observedinundation
rapid assessment of sensitivityto importantparametersand extent,andfurthermorethesefeaturesare not predictedby the
givesa good indicationof parametervalueslikely to lead to simplerplanar water surfacemodel. Thus the finite element
realisticmodel predictions. approachhas provedusefulin making more accuratepredic-
The discussion above reveals three areas of future research:
tionsthan thoseof the simplerapproach.
distributedfriction parameterization,dynamiceffects,and in-
fluenceof topography.Of theseeffects,frictionis certainlystill
the leastwell understood.While the methodologyusedhere Acknowledgments.This work wasfundedby the U.K. Natural En-
has givensomeinsightinto the nature of floodplainfrictional vironmentResearchCouncil,grant GR3 CO/030. Thanksgo to Ken
Blythof the Instituteof Hydrology,Wallingford,U.K., for provisionof
properties,the success of calibrationtechniques
will alwaysbe satellite imageryand aerial photographyand to Paul Bates of the
a function of the uncertainties in the observed data. While Universityof Bristolfor commentson the manuscript.The authoralso
usingpointhydrometric
data,uniformfloodplainfrictionwas wishesto thank the refereesfor the commentsgivenon this paper.
usuallyfound to be adequate.Satellitedata now indicatethat
a distributedfrictionparameterizationis required(or wouldbe
advantageous), andthishasincreasedthe dimensionality of the References
parameterspacein whichthe optimumfrictionvalue mustbe Aronica,G., B. Hankin, and K. Beven,Uncertaintyandequifinalityin
found.The friction distributionusedhere is essentially1-D in calibratingdistributedroughnesscoefficients
in a floodpropagation
nature,but it maywell varyacrossthe reach(especiallyasthere modelwith limiteddata,Adv. WaterResour.,22(4), 349-365, 1998.
Bates, P. D., and M. G. Anderson, A two-dimensionalfinite element
maywell be a correlationbetween!and covertype and eleva- model for river flood inundation, Proc. R. Soc. London A, 440,
tion abovechannel),and againthe calibrationproblemis ill- 481-491, 1993.
posed. We could be able to match model predictionsand Bates,P. D., and J.-M. Hervouet, A new methodfor moving-boundary
HORRITF: CALIBRATION OF A 2-D FINITE ELEMENT FLOOD FLOW MODEL 3291