Assimilation of the satellite SST data in the 3D CEMBS model

(1)

ORIGINAL

RESEARCH

ARTICLE

Assimilation

of

the

satellite

SST

data

in

the

3D

CEMBS

model

§

Artur

Nowicki

a,

*

,

Lidia

Dzierzbicka-G

łowacka

a

,

Maciej

Janecki

a

,

Maciej

Ka

łas

b

a

InstituteofOceanology,PolishAcademyofSciences,Sopot,Poland

b_Marine_Institute_in_Gda_ńsk,_Gda_ńsk,_Poland

Received4July2014;accepted4July2014 Availableonline22October2014

KEYWORDS Satellitedata assimilation; Marineecosystem modelling; BalticSea; Operational oceanography

Summary The3DCEMBS(3DCoupledEcosystemModeloftheBalticSea)isacoupledecosystem modeloftheBalticSea.Inoperationalmodeitcomputes48-hforecastsofthehydrodynamicand biochemicalparametersdescribingtheBalticSeastate.TheCressmanassimilationschemewas implementedaspartofthesysteminordertoimproveoverallmodelaccuracy.Thesystemuses satellite-measuredsea surfacetemperature fromtheMODISAqua spectroradiometerforthe assimilationprocess.ThesatellitemeasuredSSTisobtainedfromapredeﬁnedserver,whichis partoftheSatelliteMonitoringoftheBalticSeaEnvironmentproject(SatBałtyk).

Tovalidatethemodelresultsandtheimpactofassimilationonthemodel'saccuracy,two separatetestrunswereperformedusinghistoricaldatacoveringtheyears2011and2012. Inde-pendentcomputationswereperformedforthemodelwithandwithoutsatelliteSST assimilation, respectively referred to in this paper as 3D CEMBS_A and 3D CEMBS. The results of the computations werethen comparedwith satellite and in situ measured data tovalidatethe modelandtheassimilationscheme'simplementation.

TheobjectiveofthispaperistodescribetheimplementationofthesatelliteSSTdataassimilation algorithmandtopresenttheresultsofthepreliminaryvalidationofthemodelswithobservations. #2014InstituteofOceanologyofPolishAcademyofSciences.ProductionandhostingbyElsevier Urban&PartnerSp.zo.o.

§_The_partial_support_for_this_study_was_also_provided_by_the_project_Satellite_Monitoring_of_the_Baltic_Sea_Environment_—_SatBa_łtyk_founded_by

EuropeanUnionthroughEuropeanRegionalDevelopmentFundcontractno.POIG01.01.02-22-011/09.

PeerreviewundertheresponsibilityofInstituteofOceanologyofthePolishAcademyofSciences.

* Correspondingauthorat:InstituteofOceanology,PolishAcademyofSciences,Powsta_ńcówWarszawy55,81-712Sopot,Poland. Tel.:+48587311915.

E-mailaddress:[email protected](A.Nowicki).

Availableonlineatwww.sciencedirect.com

ScienceDirect

jo u rn al ho m e p age :w w w. el s ev i er. co m / lo c a te /o c e an o

http://dx.doi.org/10.1016/j.oceano.2014.07.001

0078-3234/#2014InstituteofOceanologyofPolishAcademyofSciences.ProductionandhostingbyElsevierUrban&PartnerSp.zo.o.All Open access under CC BY-NC-ND license.

(2)

1. Introduction

NumericalmodellingoftheBalticSeabasinisacomplicated

problem.Manyfactorshavetobetakenintoaccount,suchas

the inﬂow of waters from the North Sea, as well as the

inﬂuenceofriversandatmosphericconditions.Thevertical

parameterization must be very accurate as the distinct

stratiﬁcationoftheBalticSeaisveryimportant.Atmospheric

datamustalsobeofthehighestqualityastheyarethemain

forcingﬁeldsofthemodel.Evenmeetingallthese

require-mentsdoesnotguaranteethatthemodelitselfwillbeableto

producegoodqualityresults,closetotherealstate,overa

longperiodoftime.Thisiswhysatellitedataassimilationisa

very important matter that needs to be implemented to

constrain the model with observations. There are many

differentmethodsofsatellitedataassimilationused

world-wide.TheCressmananalysisscheme(Cressman,1959)isone

ofthesimplestbutalsooneofthefastestmethods,whichis

important, as themain aim ofthe 3D CEMBS(3D Coupled

EcosystemModeloftheBalticSea)istoproduceforecastsin

operationalmode.Thiswasthemainargumentforchoosing

this method over other more complicated methods that

require much more computingpower and time. Following

itsvalidation,theassimilationprocedurewasimplemented

intotheoperationalmodeofthemodel.Thisversionofthe

model provides data for the operational system of the

SatBałtyk project, described in detail by Woźniak et al.

(2011a,b).Theresultsofthisworktogetherwiththe

opera-tionalsystem'sconﬁgurationarepresentedinthispaper.

2. Material

and

methods

2.1. Assimilation scheme

Dataassimilationisananalysisthatcombines

time-distrib-utedobservationsandadynamicmodel.Thiskindofanalysis

gives much better results than simpler methods like the

spatialinterpolationofobservations.Accordingto theway

inwhichtheupdatingisdoneintime,dataassimilationcan

bedividedintovariationalandsequentialdataassimilation.

Intheﬁrstapproach,pastobservationsuntilthepresenttime

areusedsimultaneously tocorrecttheinitialconditionsof

themodel.Insequentialassimilation,observeddataareused

assoonastheyappearinordertocorrectthemodelstate.

There are many different methods of introducing the

observeddataintothemodel,from theCressmanscheme,

throughOptimalInterpolation,3-Dand4-Dvariational

meth-ods,todifferentmodiﬁcationsoftheKalmanFilter.Asthe3D

CEMBSoperationalsystemusestheCressmanscheme,other

methodswillnotbepresentedingreaterdetailinthispaper.

TheCressmanmethodisasimpleandcomputationallyfast

assimilationscheme,whichmakesitagoodchoiceforadata

assimilationsystemused tocreateforecastsinoperational

mode. It is also very accurate in comparison to its low

complexity. Its main disadvantage is that it may produce

unrealisticextremainthegridvaluesneartheedgesofthe

spatial domain. It can be also unstable if the model grid

densityishigherthantheobservationgriddensity.However,

inthecaseofsatellitedatathisisnotanissue,asthespatial

resolutionofthesatellitedatausedishigherthanthemodel

gridresolution.

The Cressman method comesdownto few simplesteps

thatareperformedasfollows.Firstly,thebackgroundstate

xbis set equal to the previous forecastperformed by the

model.Thenthesatellitedatausedfortheassimilationare

storedinthematrixdenotedbyy.Datasuspectedofbeing

invalidbecauseofclouds,thepresenceoficeoranyother

reasonaremaskedout.Theresultoftheanalysisxaisthen

calculatedaccordingtothefollowingequation:

xaðjÞ¼xbðjÞþ

Pn

i¼1_Pwði;jÞfyðiÞxbðiÞg

n

i¼1wði;jÞþE 2 ;

whereiandjrepresentthesatelliteandmodeldata

grid-pointsrespectively,anddi,j isthedistancebetweenpoints

iandj.ThemainparametersoftheCressmanmethodthat

needtobechosenaretheinﬂuenceradiusRandtheshapeof

the weight functionw, which determinehow the satellite

datainﬂuencethemodel.Oneofthedisadvantagesofthis

methodisthattheinﬂuenceradiushastobedeterminedby

trialanderror;thismakesparameterizationofthismethod

laborious. Aftermany trials with different sets of the

pa-rameters,theonethatgavethebestresultswaschosen.The

radiusRoftheinﬂuencewassetto20grid-points. Beyond

that distance the satellite data weight equals zero. The

weightfunctioninthiscaseisequalto:

wði;jÞ¼max 0;R 2 d2 i;j R2þd2 i;j ! :

Inaddition,theparameterEusedinthesuccessivecorrection

methodwasintroduced.E2isanestimateoftheratioofthe

observationerrortotheﬁrstguessﬁelderror.Ewassetto

0.5 (E2=0.25), which means that the satellite data are

treated as more accurate than the model data. However,

theyneverhaveaweightequaltoone.Intheabsenceofthis

parameter(E2₌_0),_the_satellite_data,_if_present_at_a

parti-cularlocation,wouldbegivenaweightofone.Thismeans

that the model data at this point would beomitted. The

presenceofE2ensuresthatthemodeldataare takeninto

accounteverywhereandensuressmoothingof theanalysis

product,whichpreventspossibleinstabilities.Theproductof

assimilationisthenusedasthenewinitialstateofthemodel

fromwhichthenewforecastiscalculated.

2.2. Theoperational modelsystem

Thecurrentversionofthe3DCEMBS(3DCoupledEcosystem

ModeloftheBalticSea)isbasedontheCESM(CommunityEarth

SystemModel)developedat theNationalCenter for

Atmo-sphericResearch.ItwasadaptedfortheBalticSearegionasa

coupledsea-icemodelconsistingofPOP(TheParallelOcean

Program)andCICE (TheLos AlamosSea Ice Model).

Atmo-sphericﬁeldsfromtheICM(InterdisciplinaryCentrefor

Math-ematicalandComputationalModelling)ofWarsawUniversity

areusedtoforcethemodeltogetherwithhistoricaldataof

riverinﬂows.71mainriversaretakenintoaccount.Allthese

componentsarecoupledbyaCPL7(Coupler,version7),which

controlstimeanddataexchangebetweenthesecomponents.

The model isconﬁgured in a horizontalresolution of 1/48

degreesanditisdividedinto21verticallevels.Intheﬁrsthalf

of2013theCressmananalysisschemewasusedtoimplement

(3)

SatBałtykprojectwereusedasthesourceofsatellitedata.The

aimofthisimplementationwastoimprovethemodel's

accu-racy.Themodelandsatellitedataarecomplementarytoeach

otherasinthecaseofhighcloudcoverageovertheBalticSea

themodelisthemainsourceofdata.The3DCEMBS_Amodelis

currentlyrunninginoperationalmode.Thismodeissplitinto

two separate sub-modes. The regular mode produces 48-h

forecastsusingnewweatherforecastsfromtheICMasforcing

ﬁelds.Theforecastsareproducedonaregularbasisevery6h.

Thehydrodynamicpartofthemodelproducessea

tempera-ture, salinity, current speed and direction, sea surface

height,iceareacoverandicethickness(Dzierzbicka-Głowacka

etal.,2013a).Italsoprovidesseveralbiological,chemicaland

ecologicalparameters(Dzierzbicka-Głowackaetal.,2013b).

Resultsarethenstoredinthelocalarchiveandpostedtoa

modelwebsite.Theparametersfromthesurfaceare

inter-polatedto1kmresolution,uploadedontotheSatBałtykserver

andareavailablefromtheproject'swebsite.Thesecondmode

istheassimilationmode.Thisremainsidleuntilanewsetof

satellitedataisavailableon theSatBałtyk server,at which

timethesystemswitchestoassimilationmode.Itperforms

dataassimilation,setstheassimilateddataasthenewinitial

stateofthemodeland performsnew calculationsfromthe

timeofthesatellitedata'sappearanceuntilthecurrent

fore-castendingtime.Afterwardsthesystemuploadsnewresultsin

thesamewayasintheregularmode.Thenitswitchesbackto

regular mode.The Fig. 1 outlines the scheme of how the

systemoperates.

3. Results

and

discussion

Thetestrunofthemodelwas performedonthehistorical

datacoveringtheyears2011and2012.Independent

calcula-tionswereperformedforthemodelwithandwithout

satel-liteSSTassimilation,respectivelyreferredtointhispaperas

3DCEMBS_A and3D CEMBS.The resultsof bothruns were

comparedwitheachotheraswellaswithsatellitedataand

different in situmeasurements. Validation ofthe satellite

dataassimilationwiththe3DCEMBSmodelconsistedoftwo

parts.Firstly,theresultsofbothmodelswerecomparedwith

thesatellite data to check whetherthe assimilation

algo-rithmwasworkingproperlyandtoexaminetheimpactofthe

assimilationonthemodelresults.Then,theresultsfromboth

Figure1 3D CEMBS_Awithoperationalassimilation working scheme.

(4)

modeltestrunswerecomparedwithdifferentinsitudatato

checkwhethertheassimilationactuallyimprovedtheoverall

modelaccuracy.Forapreliminaryassessmentofthe

correct-nessoftheassimilationalgorithm,sampleimagesfromthe

satellitewerecomparedwiththeresultsofbothmodelsfrom

differentdays.Fig.2showsthesamplescenefromJanuary

1st, 2011. The ﬁgure consists of the model data before

assimilation, the satellite data used for assimilation and

themodeldataaftersatellitedataassimilation.Thepicture

at bottom right shows the difference between the two

models.Inthisexamplethesatellitemeasuredtemperature

ismostlylowerthantheonecalculatedbythemodelbefore

assimilation. Assimilation lowers the temperature in the

model surface layer, as expected. The same results were

obtainedforotherscenes,whichindicatesthatthe

assimila-tionalgorithmisworkingproperly.Ofcourse,visual

compar-isonisnotsufﬁcient,soadditionaltestswereperformed.In

ordertoassesstheaccuracyoftheassimilationalgorithmand

modelaccuracy,statisticalparameterssuchasthe

correla-tion coefﬁcient r, the mean systematic error hei and the

standarddeviation hsibetween both models and satellite

datawerecalculated foralldatafrom theyears2011 and

2012,aswerethemeanvaluesanddifferencesbetweenthe

models.Aftervalidationof theassimilationalgorithm,the

same methods were used to assess the model error with

respect to in situ data. The parameters were calculated

accordingtotheprinciplesofarithmeticstatisticspresented

byfollowingequations:

-absolutemeanerror(systematic):

hei¼

Pn

i¼1ðXiYiÞ

N ;

-standarddeviation(statisticalerror):

se¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

P_n

i¼1ðeiheiÞ2

N s

:

IntheaboveequationsXiandYiaremodelandinsitudata,

ei=XiYi. The results of the assessment are presented

below. Table 1 lists the calculated statistical parameters.

Thestatisticsprovideclearconﬁrmationthattheassimilation

algorithmdoesimproveaccuracy.Correlationofthemodel

resultswiththeassimilationandthesatellitedataisbetter

andthe errors are smaller.Figs. 3—7 illustrate calculated

meanvalues anddifferencesof surfacelayer temperature

fromboth models.Fig. 3shows themeanvalueover time

throughtheyears2011and2012.Theblackpointsrepresent

the 3D CEMBS model and the grey ones the 3D CEMBS_A

model.Onecan seethat the surfacelayerresponds more

slowlyto theweatherconditions inthemodel without

as-similation;hence,thetemperatureofthislayerislowerin

springandsummerandslightlyhigherinautumnandwinter

Table1 Statisticalcomparisonofbothmodelswithsatellite measuredSST.

Modeltype R hei[8C] hsi[8C]

3DCEMBS 0.949 1.36 2.01

3DCEMBS_A 0.976 0.72 1.35

Figure5 Meandifferencebetweenthesurfacelayer temper-atureofthe3DCEMBSand3DCEMBS_Amodelsovertheyears 2011_—2012.

Figure 3 Mean surface layer temperature over the model domainintheyears2011_—2012.

Figure4 Meansurfacelayertemperaturedifferenceoverthe modeldomainintheyears2011—2012.

(5)

thanthetemperaturecomputed usingassimilation.To

dis-tinguishthesetwoperiodsbetter,Fig.4showsthedifference

between both model temperatures. Positive values mean

thatthemodelwith theassimilationof satellitemeasured

SSTgiveshighertemperaturevalues.Thisdifferenceisdueto

thedeﬁnitionofthesurfacelayerinthemodel,namely,the

ﬁrstlayerofthemodelgrid.Thethicknessofthislayeris5m,

whichismuchmorethanthatoftheactualsurfacelayerfor

which the SST is measured from satellite instruments. Of

course,athickerlayerrespondsmoreslowlytoatmospheric

forcing,asithasahigherheatcapacity.Assimilationofthe

satellitedataintothemodeladjuststhesurfacelayer

tem-perature to the atmospheric conditions faster; hence, it

responds fasterto thechanges. Fig.4 showsthetemporal

differencesintemperaturebetweenthetwomodels.Onthe

otherhand,Fig.5showsthespatialdistributionofchangesin

thesurfacetemperaturedrivenbythesatellitedata

assimi-lation.Theﬁgurepresentsthemeantemperaturedifference

betweenthetwomodels.Theblacklinedividesareaswith

positiveandnegativetemperaturechanges.Positivevalues

mean that assimilation caused an increase of the model

temperature, while negative values mean a decrease of

temperatureas aresultofassimilation.Mostofthemodel

domainiscoveredwithpositivevalues.Thismeansthat,in

general, assimilation of satellite SSTcauses a rise in the

modeltemperature.Inotherwords,themodelbiasismostly

negative.To investigate theimpact of assimilationon the

resultsmoreprecisely,thetesteddataweredividedintotwo

periods,basedonFig.4.Theseperiodswilllaterbereferred

toasthewarmandcoldseasons.Theresultsarepresentedin

Figs.6and7.Assimilationalsoincreasedthecorrelationof

thedata,whichisreﬂectedinTable1andFig.8.Figs.6and7

showthatduringthewarmerseasonoftheyearthemodel

biasismuchgreaterthanduringthecoldseason,whenitis

closetozero.ThisisalsoconﬁrmedinFig.4.Thismaybedue

partiallytothefactthatthereismuchlessinformationfrom

the satellite during winter. Further information emerging

fromFigs.6and7isthatthereareregionswherethemodel

resultsare alwayslower than thesatellite measurements.

Theseare the regionswith positivevalues on both ﬁgures

suchastheGulfofFinlandandpartsoftheGulfofBothnia.

One can also see that the differences between the two

models are much lower near seashores. This is probably

duetothelowerdepthincoastalzones,asthetemperature

of shallower water can change faster. Fig. 9 presents a

correlation of model data from both cases with satellite

measuredSST.Thepointsselectedforthecomparisonwere

basedonthelocationsofdatafromtheICESdatabaseshown

inFig.8.

TheinﬂuenceoftheSSTassimilationontheother

param-eterssuchassalinity,currentsandseasurfaceheightwasalso

investigated.However,these parametersdidnotshowany

Figure6 Meandifferencebetweenthesurfacelayer temper-ature of the3D CEMBS and 3D CEMBS_Amodelsduringwarm seasons.

Figure7 Meandifferencebetweenthesurfacelayer temper-ature of the 3D CEMBS and 3D CEMBS_A models during cold seasons.

Figure8 MapoflocationsofinsitudatafromtheICES data-base.

(6)

meaningful differences. Even during the period with the

greatest differences between 3D CEMBS and 3D CEMBS_A

in the computed temperature, that is, in summer 2012,

theotherparametersvariedonlyslightly.

After positive validation of the assimilationalgorithm's

performance,bothmodelresultscouldbecomparedwitha

setof insitudata to estimatethe actual inﬂuenceofthe

assimilation.Theinsitudatausedforthecomparisonwere

obtainedfromtheICESdatabase.Thispartofthevalidation

alsocovereddatafromdifferentlocationsinallpartsofthe

BalticSeafrom2011to2012.Thelocationsoftheinsitudata

are marked in Fig. 8. Table 2 presents the results of the

statisticalanalysis ofthe data.The not-assimilatedmodel

resultshaveanegativebiaswithrespecttotheinsitudata,

butitissigniﬁcantly smallerin comparisontoresultsfrom

Table1.Thismeansthatthesatellitemeasurementsgivea

highertemperaturethanthatmeasuredinsitu.Thisis

con-ﬁrmedbythepositivebiasofthesatellitedatawithrespect

tothe insitumeasurements. Nevertheless, assimilationof

the satellite measured SST improves the accuracy of the

model,whichis conﬁrmedbythe resultspresentedinthe

lastrowofTable2.Figs.10and11presentacorrelationof

theinsituresultswiththeresultsfromremotesensingand

bothversionsofthemodel.

Thestatisticsshowtheaverageperformanceofthe

assim-ilationalgorithmover thewholeyear.Thismeansthatthe

dataaredominatedbythemainseasonalsignal.Removalof

this signal from the data reveals the model's accuracy in

greaterdetail.Table3liststhestatisticsofbothmodelsafter

removaloftheseasonalsignal.Thisshowsclearlythat

assim-ilationofthesatellitemeasuredSSThasapositiveimpacton

the model simulations. The correlation coefﬁcient, when

Figure11 Correlationofthesurfacetemperaturefromthe3D CEMBSand3DCEMBS_Amodelswithinsitumeasurements.

Table 3 Statistics of themodelresults withand without assimilationin comparison within situmeasuredSSTafter removalofthemainseasonalsignal.

Compareddata r _hei[_8C] _hsi[_8C] 3DCEMBSvs.insitu 0.64 0.24 1.81 3DCEMBS_Avs.insitu 0.73 0.09 1.59 Figure9 Correlationoftheseasurfacetemperaturefromthe

3DCEMBSand3DCEMBS_Amodelswithsatellitemeasurements.

Figure10 CorrelationofthesatellitemeasuredSSTwithinsitu data.

Table 2 Statistical analysis of bothmodels and satellite datawithinsitumeasuredSST.

Compareddata r hei[8C] hsi[8C] Satellitevs.insitu 0.958 0.59 1.77 3DCEMBSvs.insitu 0.957 0.21 1.90 3DCEMBS_Avs.insitu 0.973 0.06 1.53

(7)

notdominatedbytheseasonalsignal,changessigniﬁcantly

moreafterassimilationisimplemented.Thesystematicand

statisticalerrorsaresimilartothosepriortotheremovalof

thissignal.

Toprovidemoredetailedresultsshowingtheperformance

of the models in different months of the year, the main

statisticalparameterswerecalculatedforeachmonth

sepa-rately. This givesa betterinsight intothe model and the

assimilationresultsindifferentseasons.Figs.12and13and

Table4givetheresultsofthesecalculations.Asonecansee,

the systematic error after assimilation is closer to zero,

whichconﬁrmspreviousﬁndingsabouttheeffectivenessof

theassimilationalgorithm.The shapeoftheplotindicates

thatduringcolderseasonsoftheyearthemodelispositively

biasedandthatduringspringandsummeritsbiasisnegative.

Thiscorroboratesearlierconclusionsthatthesurface

tem-peraturecalculatedbythemodelchangesmoreslowlythan

the one measured in situ or by satellite. The standard

deviationdoesnotshow anycleardependenceon time of

year;nevertheless, SSTassimilationdecreasedits value in

mostmonths.

4. Conclusion

ApplicationoftheCressmanassimilationalgorithmintothe

3DCEMBS_Amodelimproveditsaccuracyandconformance

ofitsresultswithinsituandsatellitemeasuredSST.Analysis

oftheresultsgivesabetterviewofthespatialandtemporal

errordistributionintheinvestigatedperiodoftime.

Over-all,thestatisticsshowanincreaseinmodelcorrelationwith

thesatellitedatafromca0.95forthe3DCMEBSmodeltoca

0.98for3DCMEBS_A.Also,themeanarithmeticerrorand

standard deviation are smaller for the model with SST

assimilation, which conﬁrms the assimilation algorithm's

correctness.Similarresultsareobtainedwhenthemodels

arecomparedwithinsitudata.Thecorrelationcoefﬁcient

inthiscaseincreasedfrom0.957to0.973andthe

systema-ticerrordecreasedstronglyinvalue.Inaddition,the

stan-darddeviationdecreasedinvalueslightly.Afterremovalof

themainseasonalsignal,thestatisticsofthemodelresults

presentedin Table3 reveal aneven biggerdifference in

correlationbetweenthetwomodelsandtheinsitudata.

The simulations of SST are also better with respect to

monthly means, as shown in Table 4 and Figs. 12 and

13.Assimilationofsatellitedataintothe3DCEMBS_Amodel

isthereforereasonable,asisitsfurtherdevelopment.The

ongoingdevelopmentoftheSST assimilationsystemaswell

asotherparameterssuchaschlorophyllaisincludedinour

researchplans.

Table4 Statisticsofthemodelswithand withoutSSTassimilationincomparisonwithinsitumeasurementsforeachmonth separately.

Month Measurements 3DCEMBS 3DCEMBS_A

hei[8C] hsi[8C] hei[8C] hsi[8C]

1 297 1.62 1.44 1.02 1.07 2 531 0.17 1.69 0.30 0.85 3 688 0.21 1.16 0.09 0.17 4 366 0.49 2.22 0.66 0.80 5 848 0.46 1.92 0.10 1.80 6 544 1.44 1.83 0.67 2.02 7 514 1.26 2.06 0.39 1.75 8 1010 0.62 2.05 0.05 0.18 9 489 0.02 1.45 0.11 1.07 10 394 0.68 1.64 0.24 0.86 11 402 0.44 1.27 0.32 0.10 12 175 0.84 1.09 0.61 1.53

Figure13 Meanmonthlystandarddeviationoftheseasurface temperature calculated from the3D CEMBS and 3D CEMBS_A models.

Figure12 Meanmonthlysystematicerrorofthesea surface temperature calculated from the3D CEMBS and 3D CEMBS_A models.

(8)

Acknowledgements

Thecomputingpresentedinthispaperwascarriedoutonthe

GalerasupercomputerattheAcademicComputerCentrein

Gdansk(CITASK).

InsitudatausedforvalidationwereobtainedfromICES

DatasetonOceanHydrography.TheInternationalCouncilfor

theExplorationoftheSeaCopenhagen.2011,http://ocean.

ices.dk/helcom/Helcom.aspx?Mode=1.

References

Cressman, G.P., 1959. An operational objective analysis system. Mon.WeatherRev.87,367_—374.

Dzierzbicka-G_łowacka, L., Jakacki, J., Janecki, M., Nowicki, A., 2013a. Activation of the operational ecohydrodynamic model

(3D CEMBS) — the hydrodynamic part. Oceanologia 55 (3), 519—541.

Dzierzbicka-Głowacka, L., Jakacki, J., Janecki, M., Nowicki, A., 2013b. Activation of the operational ecohydrodynamic model (3DCEMBS)—theecosystemmodule.Oceanologia55(3),543—572.

Woźniak,B.,Bradtke,K.,Darecki,M.,Dera,J.,Dudzińska-Nowak,J., Dzierzbicka-G_łowacka, L., Ficek, D., Furma_ńczyk, K., Kowa-lewski,M.,Kr_ężel,A.,Majchrowski,R.,Ostrowska,M.,Paszkuta, M.,Sto_ń-Egiert,J.,Stramska,M.,Zapadka,T.,2011a.SatBa_łtyk_— a Baltic environmental satellite remote sensing system _— an ongoingprojectinPoland.Part1:Assumptions,scopeand oper-atingrange.Oceanologia53(4),897_—924.

Wo_źniak,B.,Bradtke,K.,Darecki,M.,Dera,J.,Dudzi_ńska-Nowak,J., Dzierzbicka-Głowacka, L., Ficek, D., Furmańczyk, K., Kowa-lewski,M.,Krężel,A.,Majchrowski,R.,Ostrowska,M.,Paszkuta, M.,Stoń-Egiert,J.,Stramska,M.,Zapadka,T.,2011b.SatBałtyk— a Baltic environmental satellite remote sensing system — an ongoing project in Poland.Part 2: Practical applicabilityand preliminaryresults.Oceanologia53(4),925—958.