See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/266178336

Parameter Sensitivity in Calibration and Validation of an Annualized

Agricultural Non-Point Source Model

Article · January 2003

DOI: 10.1029/WS006p0331

CITATIONS READS

6 42

2 authors:

Barbara Baginska William milne-home

State of California University of Technology Sydney
12 PUBLICATIONS 321 CITATIONS 16 PUBLICATIONS 170 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Barbara Baginska on 29 July 2015.

The user has requested enhancement of the downloaded file.

 ParameterSensitivityin Calibration and Validation of an
AnnualizedAgriculturalNon-Point SourceModel
BarbaraBaginska

NSW EnvironmentProtectionAuthority, Sydney,Australia

William A. Milne-Home

Universityof Technology,Sydney,Australia

The capabilityof the new, continuousmodel AnnualizedAGricultural Non-

Point Source(AnnAGNPS version2) for simulatingflow events,peak discharge,
and generationof nitrogenand phosphorusloads was testedon extensivefield
data. The hydrologicand water quality data were obtainedfrom an intensively
monitored,smallrural watershedwithin the Hawkesbury-Nepeanriver systemof
New South Wales, Australia. AnnAGNPS is a large environmentalsimulation
modelfor whichpredictionuncertaintyis inherentbothin the modelstructureand
in parameteridentificationdependingon how well watershedheterogeneityis
represented. In this studyAnnAGNPS was coupledwith the model independent,
nonlinearparameterestimationcode,PEST, for calibrationandsensitivitytesting.
This approachprovided insight into the sensitivitiesof outputpredictionswith
respectto the variationof parametersfrom a basevalue. The basevaluescan be
definedin relationto the calibratedmodel outputsand field measurements.As all
AnnAGNPS input parametersrepresentmeasurablepropertiesand conditions,
rangesof the parametersneed to be specifiedand violationsof the rangelimits
monitoredto minimisepredictionerrorsand problemsof non-uniqueness in the
parameterselection.PEST calibrationand sensitivityroutinescan be used sys-
tematicallywithin theseconstraintsfor parameteroptimizationand identifiabili-
ty. The measured event streamflows were matched satisfactorily by
AnnAGNPS/PEST,but modellingof daily generationof particulatenitrogenand
phosphorus achievedonly moderateaccuracy.The lattermay reflectfactorsinher-
ent in watershedprocesses as well as their representation
by the model.

1. INTRODUCTION southeastern Australia. These problemscan be attributedto

land usechangeandincreasingnutrientloadsin runoff from
Persistentalgal outbreaks,low dissolvedoxygenand ele- rural and urban land. Assessingcontributionsof nitrogen
vatedlevelsof nutrientsarejust a few symptomsof exces- and phosphorusfrom nonpoint sourcespresentsa constant
sive eutrophicationresulting in deteriorationof aquatic challengeto researchersand water quality managers.The
habitat and water quality in major watershedsthroughout majordifficultywith quantifyingnutrientloadsin runoffcan
be attributedto the fact that runoff eventsare highly unpre-
Calibration of Watershed Models dictableand rainfall has been long recognisedas one of the
Water ScienceandApplicationVolume6 major factorscontrollingnutrientmovementin Australian
Copyright2003 by theAmericanGeophysicalUnion watersheds[Eyre, 1995; McKee et al., 2000]. Furthermore,
10/1029/006WS 24 a simpleaggregationof singlesourcesdistributedacrossthe

331
332 PARAMETER SENSITIVITY IN CALIBRATION OF AnnAGNPS MODEL

watersheddoesnot reflect a tributaryload enteringa water- ment generatedby sheetand rill erosion.As part of the
way and rapidly changingland use patternsand manage- deliveryprocess,the overlanddepositionof the erodedsed-
ment practicescontributeeven furtherto the complexityof iment ratherthan a completedelivery of the materialto the
the problem. streamsystem,is simulated.The generationof phosphorus
An intensive,field monitoringstudywas conductedin a (P) was improvedin Version2. Processbasedmodelssuch
small subwatershed of CurrencyCreek to quantifynitrogen as AnnAGNPS can often be applied by simply adjusting
andphosphorus contributionsfrom nonpointsourcesandto parametersfrom the initial input of physicallyrealisticval-
provide improved long-term estimatesof nutrient runoff uesuntil an acceptablefit is obtainedto the observedfield
from agriculture.The studyareais situatedon the southern data.An effect of this procedureis the non-uniqueness of
slopes of the creek valley, 90 kilometres northwestof parameterestimatesresultingfrom the over parameterisa-
Sydney,Australia (Figure 1). It is a subwatershedof an tion inherentin large complexmodels.This problemmay
unnamed,ephemeralstreamdraining255 hectaresof inten- be overcome partly through the sensitivity analysis of
sively used agricultural and rural residential land. The parametervalues.
monitored area representsapproximately 7.5% of the Sensitivityanalysishasbeenapproachedpreviouslyon a
CurrencyCreek watershedand 0.01% of the Hawkesbury- large scale within the parameterspace [Hornbergerand
Nepean watershedwhich, with an area of approximately Spear, 1981; Spear, 1997] or on a restrictedscalewithin a
22,000km2,is oneof the largestandmostdiversecoastal more localised region [Pastres et al., 1997]. Brun and
watershedsin New SouthWales.We extendedthe studyby Reichert [2001] point out that the bestresultsare obtained
attemptingto simulate the generationand transportof from a combinationof both methodsin casesof a high
nitrogen and phosphorousthrough the Currency Creek dimensionalparameterspace,with local parametersbeing
watershed with the Annualized Agricultural Nonpoint used to indicate those areas which result in the best fits
Source Pollution (AnnAGNPS version 2) watershedmod- amongthe model outputs.Our approachto the problem
eling package. was to couple the model-independent Parameter
AnnAGNPS packageis a large,environmentalsimulation ESTimation software, PEST, with AnnAGNPS. PEST
model, which can suffer from the problemsof parameter allows for the optimizationof an initial set of parameter
identifiability and sensitivity common to such models. valuesto obtain the best fit. A sensitivityanalysisroutine
Thosemodelsmightbe more applicableto rural watersheds (SENSAN) is included in the package.Previousapplica-
with limited monitoringdata, as they, in principle,do not tions of PEST have been with MODFLOW and HSPF mod-
require calibration. However, they require extensive elling packages[Dohertyand Johnston,2002]. Our linked
amounts of information on watershed characteristics,which use of AnnAGNPS version2 and PEST appearsto be the
may or may not be readily available.Furthermore,Jamieson first attemptin Australiato testthe performanceand appli-
and Clausen [1988] maintain that all models must be care- cability of theselinked modelingpackagesfor simulation
fully calibratedor verified for site specificconditionseven andpredictionof nutrienttransport.
if no calibration is claimed to be necessaryin general.
Nevertheless,AGNPS has been used extensivelyto model 2. MODEL STRUCTURE AND DATA INPUTS
nonpointsourcepollution and to assistwith the manage-
ment of runoff, erosion and nutrient movement in rural AnnAGNPS [Cronsheyand Theurer, 1998] is a daily
landscapes[Summeret al., 1990; Tim and Jolly, 1994]. In time-stepmodel for the continuoussimulationof pollutant
Australia, Foerster and Milne-Home [1995] describedthe loadingon the scaleof a watershed.The watershedis divid-
applicationof AGNPS to simulationsof nutrientgeneration ed into homogeneousareas(cells) on the basisof soils,cli-
andmovementunderdifferentfarmingpracticesin northern mate and land use. Runoff, sediment and nutrients are rout-
New SouthWales. It was necessaryin this caseto calibrate ed througheachcell via a networkof channelsto the water-
the modelfor simulatingpeak flows by adjustingthe runoff shed outlet. The movement of contaminants from within
curve numbers.The calibratedmodel was then capableof their cell of origin can be trackedthroughthe channelnet-
simulatingthe effect of proposedbest managementprac- work in the watershed so that the relative contribution of
tices on nutrientmovementin agriculturalwatersheds. point andnon-pointsourcescanbe estimated.
The conversion of AGNPS to the annualized runoff and The key feature of the package is the Input Data
nutrientsimulator,AnnAGNPS, lifted the capabilityof the PreparationModel into whichthe datarequiredby the two
packagefrom modelingindividual stormeventsto contin- input files, AnnAGNPS input and Daily Climate Data, are
uoussimulation.This allowsfor betterrepresentationof the entered.Up to 33 sectionsof datamay be neededincluding
processesinvolved in transportand depositionof the sedi- soil type, land use, crop characteristics,
pesticideand fer-
 BAGINSKA AND MILNE-HOME 333

Watershed Delineation

..... ..........
. •E '• •'•% • '":'• •::;"""::
...........
?:
.........
?•"•
.................
,:
::::t••:•• •

Amorphous...........
•rai•.i• ?•
......
•.......
:•'•";•
•11 (L:=•
............
E5- firdrol•i• •oil Group / Curv•

Land Use

Unimproved
pasture
TuffFarm
Irigatcddairypasture
Marketgarden
Poultry.
Poultryshed
Residcn.ti:al
Semi-improved
p•Lsturet
hobbyfarms

Monitoring
station
Farm.dams
Watercourse

Station 5 - Rainfall-runoff records

: o

Rainfall
Discharge 25
280

50

160 116
February •[
120 Event
rain: 165 •1 75

80
40
;542'50Event
discharge:
214026/k
...........................................
l .......................................................................... l _•, •m=._.
100

27/]/97 I ,Od97 6/Zt97 i 1/2/•97 16/2/97

Time

Figure 1. Map of thestudyarea,watershed

segmentation,
landusesandtypicalrainfall-runoff
records.
334 PARAMETER SENSITIVITY IN CALIBRATION OF AnnAGNPS MODEL

tiliser application,irrigation and land managementprac- (Table 1). Spatialand temporaldistributionof rainfall ero-
tices.The terrain-based parametersare automaticallygener- sivepowerdiffersthroughoutAustraliaandduringthe year.
ated by Flownet GeneratorModule which evaluatesthe In general,R increasesduring summermonthswhen high
topographyandresultantdrainagenetworkof the watershed intensitystormsare mostcommon.The averagevalue of R
throughthe TopAGNPS,AGFlow and VBFlonet modules. for the studyareawasinterpolatedfrom a map showingthe
Raster-type digital elevation model (DEM) data are distribution of the R factor. A cumulative value of the R
requiredas input to TopAGNPS to delineatethe modeled index basedon a 15-day period formed part of the input
area into upstreamand lateral subwatersheds and to setup data.The maximumrainfall intensity(I) for an eventwith a
the runoff and drainagechannelnetworkfor the flow, sedi- recurrenceperiodof 10 yearswas determinedfrom the IFD
ment yield and pollutantsimulations.Intuitively this flow (intensity-frequency-duration)
data for Richmondsupplied
drivendiscretization accountsbetterfor spatialvariabilityin by the Bureau of Meteorology, and the storm energy E
hydrologiccontrols.AGFlow generatesthe reachand cell (J/m2 mm) was estimatedfrom the formula developedfor
topographiccharacteristics which controlthe flow from the easternAustraliaby Rosewell,[1993]:
outputof TopAGNPS.VBFlonet is a modulefor the graph-
ics displayof the generatednetworks.Outputfrom all these E = 29.0 (1 - 0.596*10 -0.041)
modulesare passedto the PollutantLoadingModel for the
actualsimulationsfollowedby the OutputProcessor Model. where I is rainfall intensity (mm/h).
The surfacerunoffPollutantLoadingModel predictsnon-
point sourcepollutant generationand performsrisk and The TR-55 method [USDA, 1986] is used in AnnAGNPS
cost/benefitanalysis.It can simulatethe chemicaltransport to generaterunoff, computerunoff volumesand peak dis-
of particulateandsolubleformsof phosphorus andnitrogen, chargesand to route the resulting excess precipitation
organic carbon and pesticidesusing modified routines throughthe watershed.The methodappliesthe unit hydro-
derived from the CREAMS model [Knisel, 1980]. graphtheoryanddependson traveltime for peakdischarge
computationand watershedrouting. Routing procedures
2.1. Model Data Input describethe lagging and attenuationof water flow that
occursin the watershed.The simplified Manning's kine-
The boundariesof the modeledarea and hydrologicseg- matic solutionis usedto computetravel time for generated
mentationof the watershedinto amorphouscellscontribut- sheet flow.
ing flow to channellinks and the corresponding drainage A 24-hour syntheticrainfall distributionprovidesmeans
divides required by AnnAGNPS were approximated for estimationof peak discharges for a given watershedby
throughthe analysisof the DEM obtainedfor this project specifyingthe lengthof the mostintenserainfall duration
from the Land andPropertyInformationCentre,NSW. As a contributing to the peak runoff. Each distribution is
resultof DEM data processingthe modeledarea of 264.9 expressedas a masscurve of maximumrainfall intensities
hectareswas discretizedinto 13 drainageareas(amorphous arrangedin a sequencethat is criticalfor producingrunoff
cells) and 6 reaches(Figure 1). Terrain-basedgeomorphic andis relatedto the time of concentration. The Type-II syn-
parameterssuchas slope,aspect,elevationandreachlength thetic rainfall distribution was selectedfor the Currency
were alsodeterminedas a resultof DEM interpretation. Creekwatershed.The selectionwasbasedon experimental
The AnnAGNPS Input Editor was usedto developand studies[Browne,1999] showingthat it was the mostrepre-
modify the input data to the pollutant-loadingmodel.Most sentativehyetographfor areaswhereshort-duration summer
of the input parameterswere sourcedfrom measureddata thunderstorms dominate.
and where measureddata were not available,the parame-
ters were estimated based on the literature and the refer- Table 1. Selected Parameters for Runoff and Sediment Generation
encedataprovidedwith the modelingsystem.The simula-
Parameter Value Unit
tion periodfor the CurrencyCreekwatershedextendsfrom
01/01/95 to 31/12/99. The key datainputsare the groupsof MJmm/ha-hr-annum
Rainfall Erosivity (R) 2500
parameters controlling rainfall, streamflow and related
nutrienttransport. Energy
Intensity
(EI) 1888 MJmm/ha-hr
(1O-yearARI)
Rainfall dependentparameters,which reflect the ability
of a stormto causeerosion,are expressedby averageannu- 30-min
rainfall
intensity 65.2 mm/hr
(1O-yearARI)
al rainfall erosivity(R) and rainfall energy-intensityfactor
(EI30) for a 10-year average recurrenceinterval (ARI) ARI- averagerecurrenceinterval.
 BAGINSKA AND MILNE-HOME 335

The runoff volumes are predictedusing the SCS curve structingthe plot of rainfall (P) againstdirectrunoff (Q), a
number (CN) method, which uses commonly available visual comparisonof plotted data with the USDA curve
informationsuchas soil type, cover and hydrologiccondi- numberplotswasconductedto selectthe appropriatemedi-
tions to estimaterunoff. The methodhas been appliedto a an curvenumberfor the CurrencyCreek watershed(Figure
wide rangeof watersheds and climaticconditionsfor esti- 2). Although the curve numbersshouldbe constantfor a
mationof runoff volumesfor ungaugedareasin the United particularwatershed,the comparison in Figure2 showscon-
States[Knisel, 1980; Rallison, 1980]. The applicationof the siderablevariationsin the measuredwatershedresponses.
method is aided by numeroustables and graphsgiving Three distinct groupsof storm runoff curve numbersare
examplesof relevantcurvenumbersfor differentconditions noticeable,namely50- 55, 75 - 80 and 90- 95 showinga
includingsoil type, permeability,percentof impervious very high runoff potential.The variationscan be linked to
area, land cover,land use and vegetation. soil characteristicsand the high intensity and sporadic
The processof selectingrunoff curve numbersfor the natureof the stormeventsrecordedin the studywatershed,
purposeof the CurrencyCreek modelingis describedhere which emphasises the importanceof the soil moisturecon-
becausesimulatedrunoff volumes and nutrient transport ditions to watershedresponses.The CN plot represents
provedto be sensitiveto the valuesof theseparameters. solutionsto the runoff equationfor the averageantecedent
A comprehensive evaluationof theapplicabilityof theCN runoff conditions. Further adjustmentsto the CN are
methodfor Australianconditionsis providedby Boughton requiredto accountfor soil cover,land use and conditions
[ 1989]. Dilshad and Peel [ 1994] testedthe performanceof precedingthe storms,in order to fully describeCurrency
the CN method for Australiansemi-aridtropics.Although Creek watershed.
theusefulness of themethodis acknowledged, theAustralian Duringanalysisof therainfalldatafor thisstudyit became
resultsshowlarge variationsin calculatedrunoff volumes apparentthatthe highestdaily precipitationrecordedcorre-
and the importanceof antecedentmoistureconditionsin spondedwith the lowest CN of 52 (Figure 2). Boughton
determiningthe appropriate CN. Furthermore,the estimated [1989] has noted that curve numbershave the tendencyto
runoff volumesare very sensitiveto the selectionof the decreaseas the rainfall depthincreasesdue to the empirical
curvenumber,suchthat a relativelysmallchangeof 15% to natureof the methodandnonlinearityof therunoffequation.
20% in the selectedCN may resultin morethan 100% dif- As a result, the remainingtwo groupsof curve numbers
ference in the estimated runoff volume. which accountfor different hydrologicconditionsin the
In thisstudythe initial curvenumberswereselectedusing watershed,were usedin the calibrationprocess.The initial-
field measurements of rainfall and runoff. A method of CN ly selectedcurvenumbersare documented in Table2.
curve fitting by graphicalplotting of daily rainfall and AnnAGNPS also requiresthe input of the terrain-based
runoff volumes was used [Boughton, 1989]. After con- parameters for eachcell derivedfrom theDEM datatogether
6

0 I 2 3 4 5 6 7 8 9 I0 I1 12

Rainlhll {P), inch-•

Figure 2. Graphicalcomparison
of runofffor the CurrencyCreekwatershedandthe CurveNumberplot.
336 PARAMETER SENSITIVITY IN CALIBRATION OF AnnAGNPS MODEL

Table 2. Selection of curve numbers

Land Use CN Range HydrologicSoil Group Moisture Conditions Selected CN

Pasture 75 - 80 B - moderate infiltration rate Average 75

Market Garden 90 - 95
C - slow infiltration rate,
Wet 92
surfacesealing

with detailedlandmanagementandfertiliserapplicationdata. tion rangedfrom 3 to 13 dependingon the durationof the

Informationon croptypesandlanduse,soilsandclimateare monitoringperiod,the watershedarea and performanceof
also needed. the samplingequipment.In general,only 13 out of 34 rain-
fall events were significant enough to result in runoff.
2.2. Field Data Stations1, 4 and 5 recordedthe mostcomprehensive dis-
chargeand water quality data.
Five monitoringstationswere established in the Currency Altogether 420 water sampleswere collectedover 23
Creek watershedto measurerunoff and water quality in monthsof monitoringand analysedfor suspended solids,
order to assessthe contributionsof different agricultural solubleandparticulateformsof nitrogenandphosphorus as
land usesto water pollution.The event-basedwater quality well as for total nitrogen(TN) and total phosphorus (TP).
monitoringwas conductedfrom May 1995 to March 1997. Dischargeand concentrationdata were used to determine
Fully automatedsampling and logging equipment was nutrientexportsat eachmonitoringstationfor everyevent
deployedto collectthe dataandrunoff samples.In addition, monitored.The period-weightedmethodwas usedto com-
a weatherstationwasinstalledat the monitoringstationSt.3 pute event nutrientloadswhich were then summedover 12
(Figure 1) for validationof rainfall recordsandmonitoring monthsto determineannualwatershedloads.The pollutant
of climatic data. A continuous record of water level and load was computedas a productof averageconcentrations
streamflow and discretewater quality sampleswere col- measuredin samplestaken at the beginningand at the end
lectedduringflow events. of an intervalandthe volumeof waterleavingthe monitor-
The variabledischarge-increment approachwas selected ing pointduringthatinterval.The intervallengthdepended
as the most suitable for water quality sampling. This on the monitoringlocationand magnitudeof the monitored
approachenabledthe adjustabledistributionof sampling stormand, in general,rangedfrom 5 minutesto 2 hours.
throughan event which preventedfilling the autosampler The detectedconcentrations of differentspeciesof N and
bottles too soon. The dataloggercontinuouslyestimated P varied significantlybetween the monitoring sites and
streamflowand demonstratedthe progressiveincreasein events. Nitrate was found as the dominant form of soluble N
dischargeincrementsfrom low to high flows. A good in irrigationrunoff from differentland usesin the watershed
record of flow data togetherwith the selectivesampling while the elevated concentrations of soluble P dominated in
approachallowed for a relatively high accuracyof assess- the uppermostpart of the watershedwhere it could be
ment of watershednutrient exports. Table 3 shows the attributedto grazing of improved pasturesirrigatedwith
hydrologiccharacteristics
of the observedmajoreventsand dairy effluent. Typically a reductionin concentrationand
correspondingmeasuredloads of soluble and particulate load of TP betweensites1 and 5 occurreddespitethe pres-
nitrogen(N) and phosphorus(P) measuredat the outlet of enceof vegetablefarmscontributinglargeamountsof phos-
the watershed.The numberof eventssampledby eachsta- phorusto runoff [Baginskaet al., 1998]. The concentrations

Table 3. Characteristics
of theEventsMonitoredat theOutletof Currency
Creek

Event Duration Rainfall Discharge PeakDischarge SolubleN SedimentN SolubleP SedimentP

(mm) (m3) (m3/s) (kg) (kg) (kg) (kg)
24-25 / 09 / 1995 95.0 70225 2.82 554.3 36.6 29.3 10.0

30-31 / 08 / 1996 85.8 74572 1.69 598.7 84.6 86.7 37.1

29 / 09 / 1996 54.1 8773 0.29 30.2 16.2 6.9 5.1

28-31 /O1 / 1997 115.8 54246 0.75 361.6 57.2 41.5 29.2

11-12 / 02 / 1997 164.6 214026 5.15 798.5 162.3 257.8 101.7

 BAGINSKA AND MILNE-HOME 337

of nitrogenusuallyincreasedduringrunoff.Also, nitrogen from user specifiedmodel outputs.This is particularlyuse-

dominated the nutrient loads at the outlet of the watershed. ful for interpretinglarge parameterspacemodels suchas
AnnAGNPS as it allows testingof a large numberof input
3. SENSITIVITY ANALYSIS AND PARAMETER variablesat a time. The three sensitivityoutput files from
ESTIMATION WITH PEST SENSAN contain: 1) predictionsread from the model out-
come file for the range of valuesspecifiedby the user,2)
Performingsensitivityanalysisof AnnAGNPS respons- relative differencesbetweenthe model generatedresponses
es to varyingthe initial input valuesis critical to identify- for a given/testedrange of input parametersand a user
ing key model parametersand guiding the calibration definedbasesetof inputdata,and 3) modeloutcomesensi-
process.It is particularly important in the case of the tivities defined with respectto parametervariationsfrom
AnnAGNPS model, which has extensive data require- their basevalues.Sensitivityis calculatedas the difference
mentsandmay often sufferfrom poorparameteridentifia- between the model outcomefor a particular set of input
bility in respectof availableobservationdata. Despitethe variablesand the pertinentoutcomesfor the base values,
fact that all input parameters required by the model divided by the differencebetweenthe currentparameterset
attemptto representmeasurablepropertiesand conditions, and the baseparameterset.For a simplifiedscenario,when
we are only able to assignparametervalueswithin physi- only a singleparameterp variesfrom the baseset,the rela-
cally acceptableranges based on usually sparsepoint tive sensitivityis definedas:
measurements within a watershed. Initial runs of the
model clearly indicatedthat modeleddischargesandnutri- Sr = (O - O0)/(p - P0)
ent loadsboth displayedcomplexand often contradictory
responsesto changinginput variables.We used PEST to where Ob andpb are the modeloutcomeandthe parameter
examine the non-linear model responses,optimize select- base values and O and p are the model outcomeand the
ed model inputs and to assessmodel sensitivity over a parametervaluespertainingto a particularmodelrun.
large region of the parameterspace. SENSAN usesthe samemodel interfaceprotocolas the
PEST offersa uniquecombinationof model-independent PEST optimizationroutinesandthe samestructureand for-
calibrationtools. The tools includeparameteroptimization mat of the control files.
routinesfor typical model calibration,predictiveanalysis
moduleanda sensitivitymodule(SENSAN). PEST usesthe 3.1. Initial Conditionsfor SensitivityTestingand
Gauss-Marquardt-Levenberg algorithmin the modelcalibra- Parameter Optimization
tion/parameteroptimizationprocess.This procedurecom-
binesthe advantages of the inverseHessianmethodand the Approximately770 data fields had to be assignedin the
steepestdescentmethod[Presset al., 1989]. It attemptsto model applicationto the study watershed.These fields
minimizethe weightedsum of squareddifferencesbetween rangedfrom simple vegetationcodesor links betweenthe
the model-generated valuesand thosemeasuredin the field modelcellsanddatasectionsto detailedtopographic, hydro-
by varyinguserdefinedmodelinputs.The goodness of fit is logic, geomorphicand agronomicparameters.Topographic
apparentfrom the value of the optimizedobjectivefunction featuresof the watershed,suchas slope,overlandflow seg-
and is alsoprovidedby a computedcorrelationcoefficient. mentlength,drainagedimensions andthe subsequenttime of
The coefficientis independent of the numberof observations concentration may substantially influencethemagnitudeand
and levelsof uncertaintyassociatedwith thoseobservations, the dynamicsof runoff. However,theseparameterswere not
thus allowing for direct comparisonof differentparameter optimizedasthey werecomputedfrom the DEM data.
estimation runs. A user can observe the results of iterative Sensitivityanalysiswas conductedonly for thoseinput
runsin tabularandgraphicalformswhile PEST optimization variables,which exhibitedlarge physicalvariationsdue to
processis in progressand intervenein the executionof the naturalheterogeneity within the watershed,or variablesfor
modelat anytime.PESToffersmanyadditionaloptions,such which data was not routinely collected.Due to the com-
asparameterscalingandweightingandfreezingof sensitive plexity of the model, sensitivitytestingand parameteropti-
parameters to supportthe identificationof an optimalobjec- mization were carriedout in two steps.This approachmin-
tive functionand to avoid local minima [Doherty,2001]. imized interactionsbetweenthe calibratedparametersand
SENSAN is a command-lineprogram,which facilitates enabledclear determinationof the input parametersaffect-
sensitivityanalysisby allowing a userto initiate numerous ing model-simulateddischargesand respective nutrient
model runsand post-processing the resultsof thoseruns.It exports.Table4 showsthe groupsof parametersincludedin
subsequentlygeneratesa range of formatted output files the sensitivityanalysis.
338 PARAMETER SENSITIVITY IN CALIBRATION OF AnnAGNPS MODEL

Better optimizationof input parameterswas necessaryto of weightedmodel-generated outcomes.R is independent

compensate for the possible underperformance of from the numberof observedmodeloutputsincludedin the
AnnAGNPS becausethe degreeof variation exhibitedby optimizationprocessand from the absolutelevels of uncer-
the results of sensitivity analysis was quite large. tainty associated with thoseoutputs.It thereforeallowsfor
Calibrationwas undertakenusingPEST, as the usualstep- directcomparison of differentparameterestimationrunsin
wise adjustmentof input parametersuntil model outputs the contextof goodnessof fit. Valuesof R above0.9 indicate
matched observations was ineffective due to model nonlin- goodagreementbetweenthe observedandsimulatedresults.
earltiesandinterdependencies betweeninputs. Sensitivityof modelresponses was basedon the analysis
Optimizationfocusedon the inputvariablesdisplayingthe of outputsgeneratedat the oufiet of the CurrencyCreek
highestsensitivities.
Parameters representing
soilandfertiliz- watershed. A 4-day rainfall eventobservedin January1997,
er propertieswere optimizedwith PEST suchthatdiscrepan- plottedin Figure 1, was selectedfor comparisonof simulat-
ciesbetweenAnnAGNPS generatedoutputsandfield meas- ed andobservedvaluesof daily flows and solubleandpar-
urementswere minimized.The optimizationruns were car- ticulateloadsof nitrogenand phosphorus. This event was
ried out separatelyfor hydrologicresponsesand for daily usedbecauseit was observedat all samplingstationsin the
nutrientloadsof solubleandparticulateN andP. Scalingof catchment.The subsequent optimizationwasfirst undertak-
parameters andweightingof measuredloadsof N andP were en with referenceto the January1997 event,then all major
used to minimize the impact of ouflierson the computed eventswere includedin the optimizationprocessto detect
objectivefunction.The bestinitial valuesfor all parameters temporalpatternsin modelresponses.
andlimits for all physicallybasedparameters hadto be sup-
plied. Performanceassessment of the optimizationprocess 3.2. Sensitivityof Model Inputs
andthe subsequent adjustments of theparameters controlling
calibrationwerebaseduponachievedreductions in theobjec- Sensitivitytestingshoweda complexmatrix of responses
tive function and the values of correlation coefficient as dependingon the observedoutputsandthe rangeof the ini-
definedby Cooley and Naff [1990]. The correlationcoeffi- tially selectedbaseparametervaluesin relationto whichthe
cient, R, is calculated as follows: relativesensitivitieswere calculated(Table5). The analyzed
patterns were further obscuredby correlationsbetween
Z - - mo) parametersandby inherentdifferencesin processes govern-
[Z -too) -too ing generationand delivery of solubleand sedimentassoci-
atedformsof N andP. Nevertheless, sensitivitytestingaided
For the range of testedparameters ci represents ith by SENSAN, enabledevaluationof a largeparameterspace
observed value,Coiis therelevant modelgenerated value,m and resultedin identifyingthe key model inputparameters,
isthemeanof weighted observed valuesandmois themean andin determinationof uncertaintyand degreeof influence
of parameterperturbationon the modeloutcomes.
Table 4. Parameters Testedin the SensitivityAnalysis Table 5 shows a relative comparisonof the sensitive
model parametersand their implicationsfor the modeled
TestingScenarios SelectedParameters results.As canbe expected,curvenumbers(CN), soilmois-
ture properties(FC1, FC2) and hydraulicproperties(SC1,
Discharge Curve numbers (CN), Field capacity, SC2) are the sole factorsdeterminingthe capacityof the
Saturatedconductivity
watershedto generaterunoff. They also have a visible
Nutrients pHa, Field capacity a, Saturated impact on the computedemissionsof soluble N and P.
conductivity b,Organicandinorganic N Althoughcalculatedsensitivitiesfor runoff and solubleN
ratio in soilc, Organicand inorganicP indicategradualand steadychangewithin the testedparam-
ratio in soilc Fraction of organic and
eter spaceof the representativecurve numbers(Figure 3),
inorganic N andP in femhzer,
.. d Annual
massrootfor pasture, they resultin the generationof opposingresponses. That is,
increasingsurface runoff due to raising curve numbers
aTested for two soiltypesandsubsequent two soillayers
bTested fortwosoiltypes, topsoillayers only resultsin higher flow velocitiesand less contactwith soil,
CTestedfor the soilrepresentative for thecroppingarea.Organic which, subsequently, reducesthe amountsof solublenutri-
andinorganicratiosrepresent theinitial amountsof nitrogen(N) and entsgeneratedby the model.
phosphorus (P) at the startof the simulation The responseof soluble P to parameter perturbation
dFertilizer fraction whichis organic N andP andmineralizable exhibitsa large local variability in the computedsensitivi-
(inorganic)N andP
ties (Figure3) andis the mostsusceptibleto changesof pH
 BAGINSKA AND MILNE-HOME 339

Table 5. SensitiveParameters
andthe ExpectedLevel of Changein Model Predictions

Parameter
CN SC1 SC2 FC1 pill pH3 Norg NinorgPinorgRM FNinorg
Discharge
Particulate N

Soluble N

Particulate P L

Soluble P ':H .H:.

CN- CurveNumber;
SC1,SC2
- saturated
conductivity,
FC1,FC2- fieldcapacity;
Norg,
Ninorg,
Pinorg'-
organic
and
inorganic
ratios
ofNand
Pinsoil,
RM- root
mass,
FNinor
•-inorganic
ratio
ofNinfertilizers
L- Low change- up to 5% changein the output
M - Medium change- up to 25% changein the output
H- High change- oftenmorethan25% changein the output

in the top layer of soil(Figure4). Maximumchangesinduced The intrinsiccorrelationbetweenparametersand a local

by pH perturbations in the top soil layercanbe two to three minimumis evidentin the responsesurfacesgeneratedfor
foldin magnitude. The increase of pH from4 to 5 couldresult runoff (Figure5). While attemptingto minimizedifferences
in reductionin particulateandsolublephosphorus generation betweenthe observedand modelgenerateddaily discharges
by 12- 25% and9 - 34%, respectively. The importanceof pH it was noted that an incrementalchangein curve numbers
valuesfor the modeloutcomeswas examinedby comparing from 70 to 75 could have a significanteffect on the values
the extent of differences between the observed and model of contributingparameters.Althoughthe generalshapeof
generateddaily phosphorus loadsfor thefull parameterspace the responsesurfaceremainedunchanged,the best match
of pH andfield capacity.The resultsgivenin Figure4 demon- between the observed and predicted discharges was
stratevery steepgradientsresultingalmostexclusivelyfrom achievedfor reversedvaluesof the saturatedconductivity
thechangein pH andidentifytwo matchingregionsof sensi- and field capacitydata.
tivity in thetestedparameterspace.For thegivenstructure of The observedchangesin solubleand particulateN and P
the modelfor the CurrencyCreekwatershed,the bestmatch due to variationsin field capacityprovide a snapshotof
betweenthe observedand modeledphosphorus loadscan be model sensitivitiesand indicate links to land management
achievedby applyingtwo entirelydifferentsetsof pH values practices.Field capacitywas testedfor two soil groupsand
rangingfrom 4.5 to 5.2 or from 7 to 8. Large changesin pH
measuredin the CurrencyCreek watershed,which varied
from 4.4 to 8.1, andtheoccurrence of identicalregionsin the
objectivefunction,may be the sourceof significantdiscrep-
anciesin modelpredictions.

. •..• • • . • :•- -.•.

0.001-

.i•' 0.000-
.M- :. • :•-. :.•.. •...-•.• /" .::
.:- ..
.... ,a. - •, • • '•+'-•.
, •.g
,. ..... :-:..:,.y•,
..., .
• "• • :--•• ..: -*•,,. •. • .>••.. :......•......::-
• -0.001-
-60 .......
..... • ..... '•/•.•'
..... 0.6.=
. .-,O,,- Discharge
ß - .•- Soluble N
-0.002-
• .SolubleP
......... • ............................ • ............................ i- -- '""' - ......... •
4
00
8• 90 92 94
Curve Numl.•
•i•m 4. Difference betweenthe obse•ed and predicteddaily
Figure 3. Sensitivitiesof dischargeand solubleN and P to curve loads of total phosphorus(TP) with unit changeof pH in the
numbers. topsoil.
340 PARAMETER SENSITIVITY IN CALIBRATION OF AnnAGNPS MODEL

for pertinentland usesrepresenting predominantlyunim- tion. However, Ndiritu and Daniell [1997] claim that cali-
provedpasture(FC1) and a mixtureof intensiveagricultur- brationis likely to remainoneof the mostimportantsteps
al activitiesdominatedby vegetablegrowing (FC2). The in modelapplication,includingand especiallyfor process
resultsindicatedthat undernaturalconditionsonly soluble basedmodels.Moreover,for large domainmodels,manual
N was susceptible to changesin field capacity,as the total stepwise calibrationmayoftenresultin underestimating
or
availablepool of nutrientswould determinethe outerlimits even entirelyomittingkey parametersand, therefore,auto-
of the extentof changein modelpredictions.In addition, matic optimizationprocedures
are more likely to warrant
sensitivitytestingidentifieda distinctgroupof parameters, better results with less effort.
which exhibited exceptionallyhigh sensitivitiesand for In this studythe aim of calibrationwas to optimizethe
which no measureddatawere available.Theseparameters model inputsso the differencesbetweenthe simulatedand
werefertilizerproperties,asa proportionof inorganicnitro- observeddatacouldbe minimizedand betteraccuracyof
gen, land usereferencedata, suchas annualroot mass,and model predictionsaccomplished.In addition, calibration
soilpropertieslistedin Table5. The selectionof appropriate allowedfor basicverificationof the initial assumptions
of
valuesfor theseparameters remainsdifficultandmayintro- watershedparameterization and providedmeansfor assess-
duce considerableuncertainty in the estimatednutrient mentof how well the modelinputparameters described
the
loads.The difficultiesstemfrom the likely heterogeneity of relevantcharacteristics
of the CurrencyCreek watershed.
theseparameterson the field scaleas valuesmay range Detailed optimizationof Anr•GNPS using PEST started
across most of their recommended domain. with 4 parameters
controllingdailyrunoffvolumes(Table6).
Thiswasfollowedby theoptimization of 11parametershav-
3.3. OptimizationResults ing majorimpactson thesimulatedsolubleandparticulate N
andP, suchaspH, soil moisture,annualrootmassandratios
It is sometimesclaimedthat agriculturalnonpointpollu- of soilandfertilizerN andP.Thevaluesof optimizedparam-
tion modelsare not to be calibratedas they containnumer- etersareshownin Table6. Optimization of discharge
related
ous interdependent variablesresultingin complexinterac- parameters for theJanuary1997eventrequiredordy4 itera-
tions between them [Shepherd and Geter, 1995]. tionsanda veryhighcorrelationcoefficientR, exceeding
0.9,
Furthermore,processbasedmodels,suchas AnnAGNPS, wasachieved.As expected,theadditionof extraeventsadded
are designed to characterizewatershedprocesseswell inherentvariabilityin watershedresponses,
andsubsequently
enoughto enablethe useof measurable propertiesandcon- causedalterations in the optimizedparametervaluesmainly
ditionsand, therefore,they do not requireformal calibra- by increasingthe curvenumbers.Nevertheless, a relatively
70

?...• CN
=75
60

= 50

• 4o

30

20 ,
0.1 03 0,4 0.5 0,6 0,7

Ficld Capacit3•

Figure5. Contourgraphsof differences

betweentheobserved andpredicted runoffin theparameter
spaceof saturated
conductivity,
field capacityandcurvenumbers.Contourvaluesrepresent differencesbetweenobserved andpredicted
dailyrunoffin cubicmetersx1000.Regionsoutlinedby a dottedlineindicatetherangeof parameters
for whichthemin-
imum discrepancywas observed.
 BAGINSKA and MILNE-HOME 341

high value of R, reaching0.7, was maintained,and a good Table 6. OptimizedParameters

matchbetweentheobservedandsimulateddaily dischargeis OptimizationScenarios
Parameter
clearlyvisiblein Figure6. JanuaryEvent All Events
Despitegoodcorrelationachievedfor simulationof event
Field capacity 0.17 0.1
anddaily discharges,in the subsequent
optimizationrunsfor
nutrientemissionsR valuesdid not improvebeyond0.54 - Saturatedconductivity 21.90 27.2
0.61. The objectivefunctioncomputedby PEST duringthe Curve Number 1 73 79

entire estimationprocesswas dominatedby contributions Curve Number 2 80 90

from soluble N and P. Sediment-associated P contributions
exhibitedtwo ordersof magnitudelower impact(Figure7) thosemeasuredin the field. This verificationphasefocused
andshowedinsignificantvariabilitydespitethe magnitudeof on how well the modelcouldsimulaterunoffin the upland
the eventstested.The mainimprovementin the optimization sectionof the watershed(Figure 1, St.1) and at the outletof
resultscamefrom assigningweightsto the observations of the watershed.The qualityof predictionsfor nutrientgener-
particulateP and from scalingparameters,such as annual ationwas thentestedby comparingsimulatedandobserved
rootmassfor which therangeof representative valueswas at loadson daily, eventand annualbasis.
least two orders of magnitudelarger than for any other Despite a few simplifying assumptionsmade in the
parameterusedin the optimization.Althoughhigh sensitivi- processof runoff simulation,acceptablegoodness of fit was
ties to pH were observed,in orderto satisfythe criteriafor achieved for runoff volumes. The level of calibration was
minimizing the differencebetweenthe observedand simu- quantifiedwith the coefficientof efficiency(E) [Nash and
lated nutrientloadsthe optimizedpH valueshad to remain Sutcliffe,1970] andthe mean,usedasmeasuresof degreeof
closeto thelowerendof thepH rangemeasured in field sur- modelaccuracyanddistributionof centraltendency,respec-
veys(4.5 - 5.0). The very strongpH dependency andappar- tively. Generally, acceptancecriteria for rainfall-runoff
ent inabilityof the modelto adequatelysimulatedaily varia- modelingare still very much subjectiveand may vary sig-
tionsin particulateN andP mayintroducelargeuncertainties nificantlyfrom applicationto application.The quality of
to modelpredictionsfor ungaugedwatersheds. AnnAGNPS hydrologicpredictionswas assessedwith the
criteriasuggested by Chiew et al. [1993], which were based
4. MODEL PREDICTIONS AND PERFORMANCE on 112 monthlystreamflowsimulations conducted through-
outAustralia.Accordingto theirfindingsflow estimatescan
Evaluationof the AnnAGNPS ability to predictflow and be classifiedas acceptableif they have coefficientof effi-
nutrient rates in the ephemeralstream subwatershedof ciency (E) greater than 0.6 and mean simulated flow is
Currency Creek involved sensitivitytesting, systematic always within 15% of mean recordedflow. The E criterion
optimizationof the key input parameterswith PEST and for eventflows was met spatiallyfor the studywatershed,as
verificationof N andP loadsgenerated
by themodelagainst thecoefficientconsistentlyexceeded0.8 for theuppergaug-

1oooooo 1000000

',• 1ooooo g I00000

.[,,•

--• lOOOO
. .
• !0000

!ooo
I000 10000 100000 ! 000000 1000 10000 100000 1000000

Measured
Daily
Flow Mcasurcd
Evcm
Flow(m3)
Figure 6. Correlationbetweenthe observedandpredicteddaily andeventflows.
342 PARAMETER SENSITIVITY IN CALIBRATION OF AnnAGNPS MODEL

100(
- Particulate P

<>-
ßSoluble
P ,I• TN-predictcd
':•TN-mettsured
I
,.•75t
'= 0.01

¸
251
0

0 2 4 6 8 Scp-95 Aug-96 Sop-96 Jim-97 Feb-97

Iteration Nmnber

Figure 7. Objectivefunctionversusiterationnumber. Figure 8. Measuredandpredictedeventloadsof totalnitrogenat

the watershed outlet.
ing site (St.1, Figure 1) and at the outlet of the watershed
(Figure 6). However, the model did not perform equally nitrogenloads at the outlet of the watershedwere mostly
well for daily andevent-based assessments, representingthe underestimatedwhile the oppositeoccurredfor the total
two temporalscalestested. phosphorusloads. It seemsevident that the currentmodel
The simulateddaily flows showeda wider scatter,which formulation underestimatesparticulateN, which subse-
often causedthe coefficientof efficiency (E) to be lower quentlyresultsin lower thanexpectedability of the model
thanthethresholdlimit for an acceptablesimulation.For the to simulate TN loads.

majorityof the events,the modelpredictionsoverestimated The modelfailed in its ability to predictparticulatephos-

the recordeddaily flows at the upper gaugingsite (St.1). phorusandnitrogenloadson all temporalandspatiallevels
The optimizationof hydrologicparametershelpedreduce tested.Those predictionswere consistentlylow, exhibited
previouslyobservedsignificantvariationsaroundthe 1:1 smallsensitivityto the eventmagnitudeandunderestimated
line for daily flows at the outletof the watershed,causing the observeddataby at leastan orderof magnitude.Despite
low E valuesrangebetween0.26 and 0.44. extensiveoptimizationand sensitivitytestingthe resultsof
In order to determinethe predictive capabilitiesof the particulateN and P simulationswere not changingmuch
model for simulationof nitrogenand phosphorus loadsin unlessunrealisticallyhigh erosionrateswere allowed.This
semi-arid conditions, considerable effort was directed indicatesthat the model cannotadequatelysimulatetrans-
towardsnot only optimizationof input parametersbut also port, immobilizationand re-suspension of particle-bound
to revisions of the structure of the model. As recommended pollutantsin the CurrencyCreek watershed.The somewhat
by Novotnyand Olem [1994], in order to avoid temporal low overallqualityof the modelpredictivecapacityindicat-
and spatialerrorpropagation,hydrologyand sedimentneed ed by this studymay resultfrom a combinationof factors
to be calibratedbeforeany attemptsto modelwaterquality and may be specificfor the watershedand not necessarily
are made.Despiteachievingacceptablelevelsof optimiza- for the model.Internalmodeldeficienciesmay be relatedto
tion for flow, predictionsof nitrogenandphosphorus loads the representationof watershedprocesses and the selection
with the calibratedmodel still differed significantlyfrom of relevant assessment methods,while external problems
the observedvalues(Figure 8, 9). may be relatedto the conditionswithin the modeledwater-
The best fit between the observed and the simulated nitro- shedandthe qualityof observeddata.
gen load was achievedwhile comparingthe resultson an The descriptionof physicalprocessesof runoff genera-
eventbasis.Despitethe meanvaluesfor the predictedand tion in the modelmay not be adequate.As mentionedearli-
measuredloadsmatchingclosely,the calculatedcoefficient er, calculation of runoff is based on the SCS Curve Number
of efficiency(E) was usuallynegativeindicatinghigh devi- techniquewhich was designedto predictthe total channel
ationsof the predictednitrogenexportsfrom the measured flow at the watershed outlet, for which considerationsof
ones. Large deviations from the measured data were watershedflow pathsand runoff generationareaswere not
observedfor daily simulationsregardlessof calibration essential[Garen et al., 1999]. Although the methodis still
efforts.Notwithstandingthe uncertaintyin absolutepredic- quiteuseful,its originaldesignand applicationsare extend-
tionsof nutrientexports,relativelyclosepatternsbetween ed in the model to accountfor runoff occurringon the land
the simulated and the observed data could be seen for total surfaceandnot in the streamchannel.A highnumberof input
nitrogenand phosphorus
(Figure 8, 9). The simulatedtotal parameterscan also contributeto difficultiesin calibration
 BAGINSKA AND MILNE-HOME 343

Creek subwatershed is characterizedby a mixture of agri-

culturaland rural-residentialland usesand is representative
of managementpracticesand land use patternscommonin
the Hawkesbury-Nepean watershed.The modeledsubwater-
shedalso experienceswidespreadenvironmentaland water
quality problemsoften related to agriculture.These prob-
lems suchas extensivesoil erosionandhigh lossesof nutri-
entsin runoff, in turn, may contributeto the deteriorationof
waterqualityin the entireHawkesbury-Nepean river system.
The quantificationof nutrientloadsfrom nonpointsources
Sep-95 Aug-96 Sep-96 Jan-97 Feb-97 is the primaryfocusof many watershedstudies.Estimatesof
nutrientloadsform the basicprerequisitefor the subsequent
assessment of how agriculturemay influencethe long-term
[g•
TP-predicted
• TP-measured quality of surfaceand ground water. On the other hand,
direct measurements of nonpointnutrientloads are always
difficult, costly and even impracticalin somemanagement
applications.Consequently,new generationwater quality
modelscapableof simulatingerosionratesandwaterquality
on a watershedscale are highly desirablein the light of
increasingneedfor suchmodelsin land managementplan-
ning andimplementationof conservation measures.
Sep-95 Aug-96 Sep-96 Jan-97 Feb-97 Evaluation of the model predictionsundertakenin this
studydemonstrates thatAnnAGNPS producesresultsof sat-
Figure 9. Measuredandpredictedeventloadsof total phosphorus isfactoryquality when simulatingevent flows but a high
at the gaugingsiteSt. 1 (top) andat the watershedoutlet(bottom). degreeof uncertaintyis associated with predictionsof nutri-
ent loadings.The ability of the modelto adequatelysimulate
resultingfrom interdependence of variousparametersand phosphorus loadsin watershedswith no permanentflow and
discontinuitiesin the responsesurfaceof the model [Chiewet multi-peakrunoff eventsis, at this stagequestionable.
al., 1993].In everymodelapplication,includingthisone,the This deficiencydoes not discreditthe quality of model
assessment of the model's ability to simulate watershed predictionsother than particulatenitrogenand phosphorus
responses is basedon a fundamentalassumptionof absolute loads and does not prevent a more successfuluse of the
qualityof the measureddata.However,the datais errorprone model in Australian conditions. There is indication that the
becauseof uncertaintyusuallyassociated with the estimation model predictive capacity increasesin perennial streams.
high-magnitude An improvementin waterqualitypredictionscanbe noticed
of eventflow, in particularfor short-lasting,
events. Furthermore, simulation of stream flow and water at the watershed outlet where a stream channel is better
quality in the CurrencyCreek watershedis complicatedby defined and baseflow occursdue to limited groundwater
the fact that flow occursonly as a resultof significantrain- rechargewhichmakesthe flow andsoil moistureconditions
falls andwatershedresponses to a particularstormcanbe to morepredictable.
a high degreemodifiedby the antecedent moistureconditions In addition,the resultssuggestthat the modelmay be bet-
whicharehighlyunpredictable. The factorsmentionedabove ter suitedfor studieson a largerregionalscalethanfor small
can influence the predictive capacity of the AnnAGNPS subwatersheds. Local conditionsmay prevail in the latter
model and reducesignificantlyits ability to simulatephos- and over-parameterization is likely, causingadverseeffects
phorusemissionsin ephemeralstreamwatersheds. on model predictivecapacity.It shouldalso be noted that
annualand eventbasedpredictionsare betterthan thosefor
5. CONCLUSIONS shorter time increments, which is not uncommon in contin-
uous simulation models. Despite the fact that all model
The objectiveof this studywas to model nitrogenand inputshave physicalmeaning and can be measuredin the
phosphorus loadingsfrom nonpointsourcesfor the subwa- field, calibrationis alwaysrecommendedasit allowstuning
tershedof CurrencyCreek in orderto examineapplicabili- of the parameterscontrollingmajor deliveryprocessesand,
ty, predictivepower and implementationeffort of the new subsequently,may improve the quality of the results.
continuoussimulationAnnAGNPS model. The Currency Although the popularityof modelssimilar to AnnAGNPS
344 PARAMETER SENSITIVITY IN CALIBRATION OF AnnAGNPS MODEL

comesto someextentfrom the fact thattheycanbe applied Chiew,EH.S.,M.J. Stewardson,

andT.A.McMahon,Comparison
in data-poorwatersheds,the needfor calibrationshouldbe of six rainfall-runoffapproaches,
J. Hydrol.,147, 1-36, 1993.
the uncertainty Cooley,R.L., andR.L. Naff, Regression
recognized,as it alsohelpsto understand modelingof groundwater
associated
with the results.Nevertheless,
as interdepend- flow: US GeologicalSurveyTechniques in Water Resources
ence of model parametersis evident, calibrationcan be a Investigations,
Book3, ChapterB4, pp232, 1990.
Cronshey,R.G., and F. D. Theurer, AnnAGNPS - Non-Point
difficultprocess.The issuesof parameteridentificationand
Pollutant
LoadingModel,in Proceedings
1stFederalInteragency
sensitivitycan be addressed by includingan optimization ModelingConference, 19-23April,LasVegas,NV, 1998.
processas part of the modeling.This studyhas shownthat Dilshad, M., and L.J. Peel, Evaluation of the USDA Curve
couplingAnnAGNPSandPESTprovideda semi-quantita- Number Method for agriculturalwatershedsin the Australian
tive estimationof parametersensitivity.
The SENSAN fea- semi-aridtropics.Aust.J. Soil Res.,32, 673-685, 1994.
ture of PEST allows the modeler to track and control the Doherty, J., PEST-ASP User's Manual. Watermark Numerical
incrementalchangein parametervalues. Computing,2001.
Runoffgenerationandsediment predictionsaresimulated Doherty,J., andJ.M. Johnston,
Methodologies
for calibrationand
in themodelwith separate
functions,butnitrogenandphos- predictiveanalysisof a watershedmodel,J. Am. WaterResour.
phorustransportis flow dependent.Therefore,particular Assoc.,in press2002.
attentionis neededduringthe verificationprocessso the Eyre, B., A first-ordernutrientbudgetfor the tropicalMoresby
Estuaryand watershed,North Queensland,Australia,J. Coastal
predictedflow volumesmatchthoseat thegaugingstations, Res., 11(33), 717-732, 1995.
if available.Otherwise,any inconsistencies
originating Foerster,J., and W.A. Milne-Home, Applicationof AGNPS to
from inadequatepredictionsof the flow volumesand event modelnutrientgeneration
ratesunderdifferentfromting
man-
patternsare likely to be transferredand amplifiedin the agementpracticesat the GunnedahResearchCentrewatershed,
water quality simulationswhich follow. Aust.J. Exp. Agric., 35, 961-967, 1995.
The modelinputrequirements canbe very extensiveand Garen,D., D. Woodward,andF. Geter,A useragency'sview of
a considerable amount of time should be allowed for assess- hydrologic, soilerosionandwaterqualitymodelling,Catena,
mentof the initial inputdataandwatershed conceptualiza- 37, 277-289, 1999.

tion.The structure of themodelinputfile permitsa reason- Hornberger, G.M.,andR.C. Spear,An approach to thepreliminary
analysisof environmental systems, J. Environ.Manage.,12, 7-
ablelevelof flexibilityin selection of datasections to rep- 18, 1981.
resentthe desiredwatershedcomplexitydependingon the Jamieson,C.A., and J.C. Clausen, Test of the CREAMS model on
aim of themodelingandtheexpected prediction accuracy. agriculturalfieldsin Vermont,WaterResources Bulletin,24(6),
A highlevelof empiricalknowledge and,in particular, prior 1219-1226,1998.
knowledgeof the watershed, agriculturalactivities,soiland Knisel,W.G.(ed.),A Field-ScaleModel for Chemicals,Runoff and
climaticconditions is a big advantage duringall phasesof Erosion from Agricultural Management Systems. US
modeling,from watersheddiscrefization to optimization Departmentof Agriculture.Conservation, ResearchReportNo.
andinterpretationof the results. 26, 640pp,WashingtonDC, 1980.
McKee, L., B. Eyre, and S. Hossian, Intra- and interannual
export of nitrogen and phosphorusin the subtropical
Acknowledgments.
The authorswould like to thank Prof. Peter RichmondRiver catchment,Australia,Hydrol. Process.,14,
Cornishfrom Universityof WesternSydney-Hawkesbury for 1787-1809, 2000.
makingtheCurrencyCreekdataavailablefor thisproject. Nash,J.E.,andJ.V.Sutcliffe,Riverflow forcasting throughcon-
ceptualmodels:PartI. A discussionof principles,J. Hydrol.,10,
REFERENCES 282-290, 1970.
Ndiritu,J.G.,andT.M. Daniell,An improvedgeneticalgorithmfor
Baginska,B., P.S.Cornish,E. Hollinger,G. Kuczera,andD. Jones, rainfall-runoffmodelcalibrationandfunctionoptimization,in
Nutrient exportfrom rural land in the Hawkesbury-Nepean Proceedingsof the InternationalCongresson Modellingand
watershed,in Proceedings loth Australian Agronomy SimulationMODSIM'97, 4, 1683-1688, 1997
Conference, July 1998.WaggaWagga,753-756, 1998. Novotny, V., and H. Olem, Water Quality: Prevention,
Boughton,W.C., A review of the USDA SCS Curve Number Identification,and Managementof Diffuse Pollution, Van
method,Aust. J. Soil Res., 27, 511-523, 1989. Nostrand Reinhold. New York, 1994.
Browne,F.X., Stormwater
management, in StandardHandbookof Pastres,
R., D.Franco,G. Pecenik,C. Solidoro,andC. Dejak,Local
EnvironmentalEngineering,editedby R.A. Corbitt,pp. 7.1- sensitivityanalysisof a distributedparameterswater quality
7.127, McGraw-Hill, New York; 1999. model,Reliab. Eng. Syst.Safety,57(1), 21-30, 1997.
Brun,R., andP.Reichert,Practicalidentifiability
analysis
of large Press,W.H., B.P. Flannery,S.A. Teukolsky,andW.T. Vetterling,
environmentalsimulationmodels,WaterResour.Res., 37(4), Numerical Recipes, 702pp, Cambridge University Press,
1015-1030, 2001. Cambridge,1989.
 BAGINSKA AND MILNE-HOME 345

Rallison, R.E., Origin and evolutionof the SCS runoff equation, Summer,R.M., R.A. Alonso,. and R.A. Young,Modeling linked
Am. Soc. Civ. Eng. Symposiumon WatershedManagement, watershedandlake processesfor waterqualitymanagemntdeci-
Boise, Idaho, USA, 1980. sions,J. Environ. Qual., 19(3), 421-427, 1990.
Rosewell,C.J., SOILOSS 5.0 User's Manual, A Programto Assist Tim, U.S., andR. Jolly,Evaluatingagricultural
nonpointsourcepol-
in the Selectionof ManagementPracticesto ReduceErosion, lutionusingintegrated geographicinformationsystems andhydro-
NSW Department of Conservationand Land Management. logic/waterqualitymodel,J. Environ.Qual., 23, 25-35, 1994.
Gunnedah Research Centre, 1993. USDA, Urban Hydrology for Small Watersheds - TR-55,
Shepherd,R.G., and W.F. Geter, Verification,calibration,valida- www.ncg.usda.gov/tech_tools.html, 1986.
tion, simulation:Protocolsin groundwaterandAGNPS model-
ing, in Proceedingsof the International Symposium:Water
QualityModeling,April 2-5, Orlando,Florida, 87-91, American
Societyof AgriculturalEngineers,1995.
Spear,R.C., Large simulationmodels:Calibration,uniqueness and B. Baginska,NSW EPA, Water ScienceSection,PO Box A290,
goodness of fit, Environ.Modell.Software,12(2-3),219-228, 1997. SydneySouth,NSW 1232, Australia
Spear,R. C., T.M. Grieb,andN. Shang,Parameteruncertaintyand W. Milne-Home,NationalCentrefor GroundwaterManagement,
interactionin complex environmentalmodels, Water Resour. Universityof TechnologySydney,PO Box 123, Broadway,NSW
Res., 30(11), 3159-3169, 1994. 2007, Australia

View publication stats