High-resolution ensemble forecasts of a polar low by non-hydrostatic downscaling

UNIVERSITY OF OSLO Department of

Geosciences MetOs Section

High-resolution ensemble

forecasts of a polar low by

non-hydrostatic downscaling

Master Thesis in Geosciences

Meteorology and Oceanography

Silje Lund Sørland

September 1, 2009

Ashort range,limitedarea ensemble predition system, LAMEPS, isurrently inoper-

ationaluse at the Norwegian Meteorologial Institute. It employs 3D-Var for 6 hourly

data assimilation yling for analysis of the ontrol foreast. Initial time and lateral

boundaries ensemble perturbationsare omputed fromthe 20 +1 memberTEPS (tar-

geted EPS at ECMWF). LAMEPS is run with the quasi-hydrostati model HIRLAM

version 7.1.4. ona 12km horizontal grid mesh. In this study we havedownsaled eah

LAMEPSmemberwiththe non-hydrostatiUKMetOeUniedModel(UM)version

6.1inordertostudythepreditabilityandthepreditionsofextremeweatherrelatedto

apolarlow observed in the Barents and Norwegian Seas between 3and 4 Marh2008.

Thiseventwasextensivelyoveredbytheobservationampaignofthe IPY-THORPEX

projet. UM is in this study ongured with 4 km horizontal grid mesh. The domain

size has been investigated by using two dierent domains, one with 390

×

with300

×

Regularobservation data,satellitedata, andIPY-THORPEXampaigndatahavebeen

used to ompare with the ensemble foreasts. Probabilities of dierent meteorologial

parameters and ourrene of extreme weather events have been studied along with

ensemble means, ensemble spread and ontrolruns. In addition, two new model diag-

nostis for omparing against observation data have been developed. These are loud

top temperatures and traking of the polar lows path. The ensemble foreast shows

lear improvements by inreasing horizontalresolution with non-hydrostati dynamis.

However, the size of the integration domain aets the predition substantially. The

improvements are greatest for the large domain. The foreasts are also sensitive to

the physial parameterization. The experiments with less vertial mixing in the stable

boundary layer redue the area of high probability for the large domain. The results

of the traking algorithm, whih nds the strongest mesosale trak in eah ensemble

member, showthat the loationof the strongest trak depends ondomain size and the

perturbationof the physis.

During this master thesis I have been assisted and inspired by several people. First of

all I would thank my adviser Trond Iversen for giving me this interesting and highly

relevant thesis. The help I got from my two o-advisers, Jørn Kristiansen and Morten

A. Ø. Køltzow, has been indispensable. A speial thanks to Jørn for being espeially

enouragingevery daythrough the wholeyear. I'm alsoveryhappy I gottopartiipate

onthe IPY-THORPEXonferene in Rømskogand the SRNWP-EPS workshop inEx-

eter. I'mgratefultoØyvindS÷traforprovidingmewith theobservation datafromthe

IPY-THORPEX ampaign, Marit Helene Jensen for supplying me with the sripts for

the standard analysis tools at met.no and Thor Erik Nordeng for valuable disussions.

A speial thanks to Kevin Hodges for sharing the traking algorithm whih was very

interesting and fun towork with, and I'm verythankful forhis quik email replies.

I will also aknowledge David Weir, Kelly MCusker and Maria Sand for proof read-

ingmy thesis, and supporting me when I had run out of faith. I'm also very thankful

to Haldis Berge for being my latex orale (and for suering through the summer at

CIENS). And I'm very grateful to my family and friends for their (fake?) enthusiasm

about polar lows and for their always reliable support. In the end, the days atCIENS

would not have been the same without my student fellows. The lunhes, oee breaks

and other(food related...) ativitiesmade this year!

Abstrat i

Aknowledgement iii

1 Introdution 3

2 Bakground 7

2.1 Preditability of weather . . . 7

2.2 Dynamial Downsaling . . . 9

2.2.1 LAMEPS . . . 10

2.2.2 UM. . . 10

2.2.3 Downsaling LAMEPS with UM. . . 12

2.3 Polarlows . . . 14

3 Veriation Methodology 21 3.1 Standard Methodology . . . 21

3.2 New Methods . . . 22

3.2.1 Pseudo satellite images . . . 22

3.2.2 Traking polarlows . . . 28

4 Results 37 4.1 MSLP and pseudo satelliteimages. . . 37

4.2 Ensemble mean and spread. . . 42

4.3 Foreast probability omparedwith ampaign data . . . 43

4.4 Perturbing the physis inthe stable boundary layer . . . 52

4.5 Traking Polarlows . . . 54

5 Summary and Disussions 57

6 Conlusions and ideas for further work 61

Appendix 62

A Parameterization shemes in UM 63

Bibliography 65

Introdution

Polar lows are intense, mesosale vorties that develop during old air outbreaks over

a warmer oean, usually poleward of the polar front (Rassmussen and Turner, 2003).

Thesefeaturesoftenprodueheavypreipitationandstrongwinds,andourfrequently

along the oast of Norway during winter. Sine polarlows are ommonly aompanied

by severe weatherwhihmayause great risk tohuman life and property, itis partiu-

larly importanttoforeast thesephenomena withahigh degreeof auray. Numerial

Weather Predition (NWP) models are important for polar low foreasting. However,

due to sparse observational data, their small sale struture, and their rapid develop-

ment, it is hallenging to predit polar lows. Further, beause of the haoti nature

of the atmosphere, small errors in the initial onditions will grow with lead time and

inthis way gradually deteriorate the quality of a single deterministiforeast (Lorenz,

1963). Model deienies are also a soure of foreast error, whih redues the skill of

the deterministi foreast. In addition, for limited area modeling (LAM), the lateral

boundary data alsointrodue errors in the foreasts (Gustafssonet al.,1998). As a re-

sultof thesesoures offoreast error,deterministiforeasts are generallynotsuient

whenprediting polar lows and assoiated extreme weather.

An ensemble predition system (EPS) should take into aount all these soures of

foreast errors, and in this way foreast the atual preditability of the atmosphere.

An EPS onsist of a range of individual foreasts, i.e. ensemble members, where eah

member uses slightlydierentinitialonditions. The error growthin aforeast is ow-

dependent, whih means that in a regime of high preditability, the error growth is

muh smaller than when the weather is very unpreditable. The spread between the

dierent ensemble members gives an indiation of the atual preditability of the at-

mosphere. The dierent initial onditions are onstruted from the analysis whih has

been perturbed, exept the ontrolrun whih isunperturbed. At the European Centre

forMedium-RangeWeatherForeasts (ECMWF)the perturbationsaddedtotheinitial

onditionsare baseduponamathematialmethodalledsingularvetor deomposition

(Buizza and Palmer, 1995; Molteni et al., 1996). The model deienies have been in-

luded in the ensemble system by stohastially perturbing the model physis (Buizza

et al., 1999). By onstraining the perturbation norm to a spei area, the singular

vetors will seek the perturbations with largest norm at nal time (e.g. 48 h). This

methodof onstruting targeted EPS (TEPS) with targetover NorthernEurope for 48

hsingularvetors hasshowed improvementinforeastskill(FrognerandIversen, 2001).

Atthe Norwegian MeteorologialInstitute(met.no), the TEPSis usedtoonstrut ini-

tialand lateralperturbationstoaHighResolutionLimitedAreaModel, HIRLAM, and

in this manner a high resolution Limited Area Modeling Ensemble Predition System

(LAMEPS) is obtained. HIRLAM is a quasi-hydrostati model and its LAMEPS on-

guration is a 12 km horizontalgrid mesh. LAMEPS has, to a large extent, improved

the foreasting of high impat weather, but the resolution is still too oarse to resolve

many importantmesosale features (Jensen etal., 2006).

Itisthoughtthatwithinreasedhorizontalresolution,thepreditabilityofthemesosales

willbeenhaned (e.g. Mass etal.,2002). Hene tobetter investigate the smallersales

over an area of interest, dynamial downsaling is performed. In dynamial downsal-

ing a global or regional oarse resolution model provides initial and lateral boundary

onditions(LBCs)toamodel with higherresolution. The higherresolution model does

not produe itsown analysis. It isintended that the high resolution model should pro-

duerealisti,ne-sale detailsoveraregion,and inpartiularwheresurfae strutures

have ne details. However, dynamial downsaling is not straightforward, and several

studiesindiate that domain size, loation,horizontal resolution, and lateral boundary

onditions, in addition to the models' representation of topography, vegetation, and

physial desriptions, all aet the model results (e.g. Laprise et al., 2000; Xue et al.,

2007;Brankovi¢etal.,2008). Forinstane,Xueetal.(2007)found thatforeastresults

were ruially dependent on the domain size, LBCs, and grid spaing, and emphasized

the point that a small domain may hamper the improvements in the foreast. If the

domainissmall enoughitwillbe tooontrolledby the lateralboundaries. The domain

should be large enough to be able to spin-up small sale features not present in the

initialorlateralonditions. Xueetal.(2007)alsoinvestigatedthe eet of theloation

ofthe domains,and itwas observed thatinformationlostatthe lateralboundarieswas

hardto reprodue in the simulationswithin the new domain.

In this study, the 21 members of 12 km resolution LAMEPS are downsaled by the

UK Met Oe Unied Model (UM) with 4 km resolution. In this way a very high-

resolutionensemble preditionsystem, UM-EPS,isobtained. Thesensitivityofthe size

ofthe domainhas been investigated byemployingtwodierentdomainswith

300 × 300

and

390 × 490

The main goal of this study is to see how muh new informationis added with a very

high resolution ensemble predition system, as well as how the skill in prediting high

impatweather is inuened. UM-EPS has therefore been tested on a polar low event

thatwas extensivelyobserved during the IPY-THORPEXampaign (IPY-THORPEX,

2009),oneofthe projetsduringtheInternationalPolarYear(IPY,2009). Also,amore

realistidynamialorewillbeinvestigated,sineHIRLAMisaquasi-hydrostatimodel

wherethe onvetion is parameterized while UM is a non-hydrostati model, whih al-

lows for high vertial veloities and aelerations,and the onvetion is partly resolved

expliitlyand partly parameterized (Lean etal.,2008). Ensemble mean and the spread

between the members, together with foreast probabilities, have been ompared with

regularobservation data aswellasobservation datafrom the ampaign. In additionto

thesemoreonventionalveriationmethods,twonewmethodstoevaluatetheforeasts

have been inluded: (1)Pseudo satelliteimagesalulated frommodel prognostields

and, (2) a traking methodology to trak polar lows. The pseudo-satellite images are

adopted from a method originally developed for HIRLAM (Tijm, 2004). The traking

shemeofHodges(1994, 1995,1999)has been usedwith the aimtotrakthe mesosale

ylonesin UM-EPS.

This thesis is divided as follows: The bakground for this study is given in Chapter

2, where rst a review of the preditability of weather will be given. Then the dy-

namial downsaling performed in this study will be desribed. Chapter 2 ends with

an introdution to polar lows where there will be a loser desription of the polar low

observed during the IPY-THORPEX ampaign. The veriation methodology will be

presentedinChapter3,followedbytheresultsinChapter 4and adisussioninChapter

5. We end with some onluding remarks and ideas for future work.

Bakground

2.1 Preditability of weather

Numerial weather predition (NWP) is an initial value problem, where the ability to

makeskillfulforeastsdepends ontherealismofthe atmosphereandthe boundaryon-

ditions in NWP models, and to know the initial onditions aurately (Kalnay, 2003).

Thesetwo soures of inauraiesmay ontribute to amplifying the foreast errors and

deteriorate the quality of a single deterministi foreast with lead time. When the er-

rorsare saturatedthe errors grownofurther and the preditabilitylimitisreahed. At

that stage the foreast will not add any new information ompared to limate statis-

tis. The preditability of the atmosphere varies from day to day, beause the error

growth depends on the atual weather onditions. A omplete foreast thus also fore-

astthe preditabilityofthe atmosphere. Ensemble predition isamethod tointegrate

ensemblesofdeterministiforeastsinordertoestimatetheprobabilitydensityfuntion

(PDFs) of foreasted states (Buizza, 2002). In this way an ensemble predition system

(EPS) willforeast for how long time the weather an be predited. This setion gives

an introdution to preditability of weather and EPSs, with emphasis on the EPS ap-

pliedat European Centre forMedium Range WeatherForeasts (ECMWF).

Critial dependene on the init ial onditions

In1904the NorwegianmeteorologistVilhelmBjerknesproposedthat theatmosphere is

adeterministisystem, where allstates atagiven time an besolved with the physial

laws if only the initial state is known. But unfortunately the initial state annot be

of innite auray, and willalways ontain a tiny error. Lorenz (1963) disovered the

ritial dependen e on the initial onditions with his famous oee break, where he

did the same foreast twie just with a small initial round o error in dierene, and

surprisinglytheresultsgraduallydevelopedverydierently. Iftheowwasperiodi,the

smallerrorfromtheroundoshouldhavereturnedtotheinitialstate. Duetothisnon-

periodiity and sine there was a ritial dependen e on the initial onditions, Lorenz

realizedthathaos isprevalentforatmospheriows (Lorenz,1963). Inhisexperiments

Lorenz used a highlytrunated set of onvetion equations whih was represented in a

3D phasespae (Lorenz equations):

Figure2.1: TheLorenzattrator. Seetextforexplanations. FigureadaptedfromPalmer

et al. (2006) .

dx

dt = σ(y − x)

dy

dt = rx − y − xz

dz

dt = xy − bz

Theparameters

σ

σ

The two wings of the attrator an be onsidered astwo dierent ow regimes. Figure

2.1shows theowobtained by integrating theequations several times withslightly dif-

ferent initial onditions,and Figure2.1a, band showthree dierent sets ofthe initial

onditions and the evolution with time. In 2.1a the system is in a highly preditable

initialstate,asallpointsstaylosetogetherwithtime. Aless preditablestateisshown

in2.1b,wherethe pointsstaylose inthe beginning, butaftera whilebegin todiverge.

In2.1thereisveryshortpreditability,sineallthepointsdiverge earlyintheforeast

andend up farfromeahother. Figure2.1 illustrates thebasis forensemblepredition,

where the preditability of the atmosphere is dependent on the initial state (Palmer

etal.,2006).

Ensemble predition systems

Lorenz (1963) disovered that the foreast skill is ritially dependent on the initial

onditions as a onsequene of instabilities in the atmosphere. The unstable proesses

in the atmosphere determine how fast the small initial errors will grow and how far

into the future before the preditability limit is reahed. Lorenz (1969) estimated the

weatherpreditability limit toabout two weeks. However, the error growth isfaster at

the smallersales thanthe larger sales and reah the preditability limitrst (Lorenz,

1969). Furthermore, sine the preditability is ow dependent, the value of a foreast

byaddingsmallperturbationstotheanalysis. Byintegratingthedierentinitialstates,

arange ofdierent foreastsare obtained, and the spread amongthe foreasts givesan

indiationofthepreditabilityoftheday. ECMWFhasroutinelyemployedanensemble

predition system sine 1992, and isusing singularvetors for generating the ensemble

members (see Buizza and Palmer, 1995; Molteni etal., 1996). The singularvetors are

omputedwith the tangent linearmodel (TLM) and its adjointwith totalenergy inner

produt, thus the perturbation with fastest linear growth overan optimization time of

48hours are hosen. The singularvetors are targeted toseek the maximum perturba-

tions poleward of

30 ^o

in the foreast. And to get more slowly growing large sale perturbations the evolved

singularvetors from the previous 48 hours are alsoalulated. ByGaussian sampling

25perturbationswhihare thenaddedandsubtratedtotheanalysis,50perturbed ini-

tialonditions are obtained. The system now onsist of 50 perturbed foreast and one

unperturbed foreast(the ontrolrun), and itisdenoted 50+1 members. Sinethe al-

ulationof the singularvetorsis quiteostly,they are run with ahorizontalresolution

of T42 and 31 vertial levels and with simplied physis, and they are omputed sep-

aratelyoverthenorthern andsouthernhemispheres,aswellasoverpartsoftheTropis.

The appliation of singular vetors only aount for unertainty in the initial ondi-

tions. But to simulate model errors due to parameterized physial proesses, the EPS

atECMWF alsoemploysstohasti physis. Thisis doneby addinganextra stohasti

foring term to the parameterized physial proesses on all the members, exept the

ontrol run (Buizza etal., 1999).

The EPS at ECMWF has sine 2006 been run with T399 (50km) and 62 levels, and

provides a foreast for up to 10 days. The main purpose with the EPS is to bring ad-

ditionalvalue tothe deterministiforeast,foreast the preditability ofthe day and in

additionforeastthe probabilityof dierentweatherevents. The latterisimportantfor

extremeweatherevents,whihmayausedamagesandriskstohumanlifeandproperty.

Sineextreme weather is rare by nature, aprobabilisti approahis more appropriate.

2.2 Dynamial Downsaling

Globalmodels have toooarse resolution for resolvingmesosale and ner features. To

beable tostudy sub-synopti salesover anareaof interests, asmalldomain withhigh

resolutionisnestedinsideaoarserglobal/regionalmodel(e.g.Lapriseetal.,2000;Xue

etal.,2007;Brankovi¢etal.,2008). The nemodelisimposedbylateralboundaryon-

ditionsand initial onditionsfrom the oarsermodel. This nestingapproah isreferred

toas dynamialdownsaling.

In this study the dynamial downsaling of LAMEPS has been done with the UK Met

OeUniedModel(UKMOUM)andwillbepresented intheendofthissetion. First

LAMEPSwillbedesribed followed bya loserdesription of UM.The dierentmodel

2.2.1 LAMEPS

Sinethe introdutionof the EPS at ECMWF ithas shown inreasinglyskillin proba-

bilistiforeasting on the synopti sale 1

. Nevertheless, the global ensemble predition

systemstillemploysatoooarsegrid meshtorepresentmesoandsmallersalefeatures.

It is suggested that a ner model resolution willimprove this (e.g. Mass et al., 2002).

However, with inreasing model resolution, there will be a derease in error-doubling

time (Lorenz, 1969; Hohenegger and Shär, 2007), and the preditability limit will be

reahed earlier inthe foreast than forthe synopti sale integrations.

A shortrange High ResolutionLimited Area Model Ensemble Predition System

(LAMEPS) is thought to enhane the predition quality on the mesosale. At met.no

LAMEPShas beenrun inoperationallyweatherforeastingsine2005, andshowsgreat

skillinapturing dierentweathersituations(Frognerand Iversen,2002;Frogneretal.,

2006; Jensen et al., 2006). It is run with the Norwegian onguration of the quasi -

hydrostati limited area model HIRLAM version 7.1.4 2

with a horizontal grid mesh of

0.108 ^o × 0.108 ^o

3D-Var 6-hourly data assimilationyling. The perturbations to onstrut the ensem-

ble members are taken from targeted EPS at ECMWF (TEPS) (Frogner and Iversen,

2001). TEPS uses the same model version and set up asEPS desribed in setion 2.1.

However,insteadofusing25singularvetors(SVs)targetedtothenorthernhemisphere

northof

30 ^o

within northern Europe and adjaent areas at the nal optimization time of 48 hours

(Frogner and Iversen, 2001). Then the perturbation of eah TEPS member relative to

the TEPS ontrol, are added and subtrated tothe HIRLAM analysis, and inthis way

20perturbed ensemble members are obtained together with the unperturbed HIRLAM

analysis. Totakeintoaounterrorswhihmaypropagateand develop frominaurate

lateralboundary onditions, both the initial state and the lateral boundary onditions

are perturbed (Frogner et al., 2006). To inlude model unertainties, the members in

LAMEPSare run with alteringloud physi shemes.

Thesystems TEPS and LAMEPS are alsoombinedtoform amultimodel EPS, NOR-

LAMEPS, whih onsists of a system with 40+2 members (40 perturbed runs and 2

ontrol runs). This is a feasible method to obtain a new EPS without performing any

newruns, and it isanother way of inludingmodel unertainties. Even thoughthe two

systems are not ompletely independent from eah other, the ensemble spread islarger

thanfor both systems alone (Frogneret al.,2006).

2.2.2 UM

The Unied Model (UM) was introdued into operational weather foreasting at UK

Met Oe (UKMO) in 1991 and has sine been under ontinuously development. In

2004met.nostartedto routinelyemploy UM.In this study,version 6.1of UM has been

1

SeeECMWFveriationsores:

http : //www.ecmwf.int/publications/library/ecpublications/

pdf /tm/ 501 − 600 /tm 578 .pdf

2

See

https : //hirlam.org/trac/wiki/ReleaseN otes 7 . 1 . 4

Releasenotesof HIRLAM 7 . 1 . 4

usedandwasonguredwitha4kmhorizontalgridspaing,38levelsandanintegration

time step of 100s. These ongurations are the same as used for operational weather

foreastingatmet.no.

When utilizing UM for limited area modeling (LAM), the model runs on a rotated

latitude longitude horizontalgrid, where the omputationalNorth Pole is moved away

from the geographial North Pole. This allows the domain to take the advantage of

the even grid spaing overequatorial regions. Here the domainsare run with the same

rotationused in HIRLAM, where the rotated spherial pole is loated over Greenland

withthe oordinates

68 ^o

320 ^o

In UM the non-hydrostati equations are solved for the motion on a rotated almost

spherial planet, whih takes into aount the urvature of the earth, and desribe the

time evolution of the atmosphere. Sine the governing equations are non-hydrostati,

wherevertial aelerationisallowed, UM isdynamiallywellsuitedfor very highhori-

zontalresolutionmodeling(UKMetOe,2004). Inaddition,theequationsdepend on

thedeep-atmospheredynamis,whihrequiresthatdeep-onvetionisexpliitlyresolved

and shallow onvetion parameterized. The denition of deep and shallow onvetion

depend on the amount of the onvetive availablepotential energy (CAPE: an expres-

sion for the energy availabletoform deep onvetion), ineah grid box.

The variables used in UM are omputed every time-step and in eah grid point, and

the primary prognosti variables inlude the horizontal wind (u and v), vertial wind

omponent(w), potentialtemperature(

θ

Π

ρ

nents of moisture (vapor, loud water and loud ie) (UK Met Oe, 2004). To solve

these equations, a semi-impliit, semi-Lagrangian, preditor-orretor 3

sheme is used.

Inthe horizontaltheequationsaresolved withArakawa-Cgridpointsheme,and inthe

vertialthe Charney-Philps gridsheme isused (Staniforthetal.,2002). The Charney-

Philps grid sheme follows the terrain near the surfae and turns onstant higher up.

The 13rst levelsare below 3km, where level 1 isapproximately at20 m and level 38

is 65 km up in the atmosphere. The grids have a staggered struture in all diretions,

and the partiulargrid type has either integralor halfintegral values, i.e. P orP

±

whereP iseither i, j, k in the models physialspae. Figure2.2a shows the horizontal

arrangement of the primary variables u, v, and

Π

k = K ± 1/2

whereuandvare onthesamevertiallevelas

Π

Figure2.2b shows the arrangementof the vertialgrid struturerelativetothe top and

bottomboundaries. The horizontalveloity is onthe same level as

Π

θ

ture variablesare onthe same levelas vertial veloity.

Atmospheri proesses that operate on a smaller sale than the horizontal grid mesh

annot be resolved and are therefore parameterized. These physial proesses may in-

lude boundary layer turbulene, onvetion, large sale loud sheme, radiation and

subsurfae, surfae and layer proesses. In UM all of these proesses inlude a om-

3

Thepreditorstepapproximatethenon-lineartermsinallproesses,andintheendtheorretion

(a) (b)

Figure 2.2: (a) shows the horizontal grid struture, and (b) shows the vertial grid

struture in UM. Figure adapted from Staniforth et al. (2002)

prehensive set of parametrization, and they are listed in Appendix A. The standard

onguration at met.no inludes an enhaned vertial mixing in the stable boundary

layer (SBL), as a result of the unresolved heterogeneity (i.e. orography, land use) at

thesurfae. Inthis study wewanted tofurtherinvestigatethe sensitivitytothe physis

representingtheSBL.Thereforeadditionalexperimentswasperformedwithlessvertial

mixinginthestable boundarylayer. Aloser desriptionofthe parameterizationofthe

SBL isalsofound inAppendix A.

2.2.3 Downsaling LAMEPS with UM

Inthisexperimentwehavedownsaled LAMEPSwithUM.The reasonfordownsaling

LAMEPSinsteadofNORLAMEPSisbasedonthefatthatitwasatoobigjumptogo

fromT399(50km)to4kminsteadof12kmto4km. EahLAMEPSmemberprovided

initialonditions and lateral boundary onditionsto UM, and this gavean ensemble of

UM runs. The LBCs were imposedeah hour during the foreast. The rim width 4

was

set to8grid points. Thiswasused asreommended fromUKMO, and inthis study we

have not investigated the sensitivity to the rimwidth.

The downsaled ensemble system onsists of 20 + 1 member (one being the ontrol

run), alledUM-EPS. For this purpose we have set up two new domains, one with 300

4

Rim widthis thewith ofthe regionaroundthe edgeof thedomain thatwill undergoaweighted

relaxationbak to thevaluesin theLBC elds. If there isa largerim width, there will bea smooth

Figure 2.3: The dierent domains used in the experiments. The outer domain is the

HIRLAM domain used in LAMEPS. The two domains inside are the new domains set

up for the UM-EPS, one with 390

×

×

300 grid points. They are alled UM-EPS-big and UM-EPS-small, respetively. The

HIRLAM domain has a horizontal grid mesh of 12 km, and the two UM domains both

have a 4 km horizontal resolution.

×

×

showed in Figure 2.3. The largest domain is alled UM-EPS-big and the smaller one

UM-EPS-small. Theloationofthe domainswasdeterminedbythe developmentofthe

polarlowfromtheIPY-THORPEXampaign. Themainpurposeofperformingthefore-

astontwointegrationdomainswastoseehowtheintegrationsizeaetthepredition.

The foreasts LAMEPS, UM-EPS-small and big are all initialized 18 UTC 02.03.08

andarerun for60hours. Also,sinetherewasanenhanedobservation network during

the IPY-THORPEX ampaign, these observations are assimilated into the LAMEPS

foreast. The extra ampaign data onsists of the drop sondes from the airraft, ra-

diosondes from the dierent oastguard ships, Bear Island, Novaja Semlja, Murmansk

and Franz Josefs.

2.3 Polar lows

Arapidhangeofthewinddiretion,aninreaseofthewindstrength,andheavypreip-

itationanbeawarningoftheapproahofapolarlow. Throughouthistorymanyshing

boatshaveshipwreked duetothe sudden approahof the strongwind, whihforeasts

have failed to predit. The rapid development over oeans where the observations are

sparse,and theirsmallsalerenderpolarlowsdiulttoforeast,and thereforelifeand

property have been lost. Even though people were aware of these small storms, they

were not known to be aommonphenomenon until the introdutionof satellite images

in the 1960s, where the signiant loud struture was deteted in the images (Rass-

mussenandTurner,2003). Itwasrealizedthat thesephenomenamainlydevelop during

the winter months and over high latitudes. With growing awareness of these weather

phenomena,therehasbeenagreatinterestinthem. Thishasresultedinvariousstudies.

Toassess polarlows' temporaland spatialdistributionthere have been dierent lima-

tologialstudies (e.g. Harold et al., 1999; Noer and Ovhed, 2003; Kolstad, 2006; Zahn

andStorh,2008;Blehshmidt,2008). Toahievebetter understandingofthe dierent

physial strutures and foring mehanisms, several ase-studies and more theoretial

studies have also been performed (e.g. Rasmussen, 1979; Emanuel and Rotunno, 1989;

MontgomeryandFarrell,1992;NordengandRasmussen,1992;YanaseandNiino,2006).

Polar lows tend to develop during old air outbreaks. This is when old arti air

ows from the ie sheet over the oean. Under these onditions there are large tem-

peraturedierenes between the warmer oean and the old over-sweeping air,and the

lowest atmospheri layer will be destabilized and yield enhaned onvetion. The rea-

son for the relatively warm oean is a Western Boundary Current whih brings warm,

tropial,saline water up to higher latitudes (Hartmann, 1994). In the Northern Hemi-

sphere (NH) there are two main urrents like these, the North Atlanti urrent (more

ommonly known as the Gulf Stream), and the Kuroshio urrent. As a result of these

urrents the sea surfae temperatures (SST) in these areas have a higher temperature

when ompared to other regions at the same latitude throughout the year. Polar lows

are most ommonlyfound inthe areas aroundSvalbard, the Norwegian Sea and in the

Barents Sea, but there are frequently observed developments around Greenland and

east ofCanada, the Beaufort Sea, the BeringSea, the Northwest Pai and the Sea of

Japan. Polar lows an be found in high latitudes of the Southern Hemisphere (SH) as

well,but(possibly due tothe olderoean) they are notas intenseasthe ones found in

NH.

Thestudy of polarlows is stillrelativelynew and therefore the theoretial understand-

ingis not omplete. Butitis realizedthatthere are manydierentforing mehanisms

whih trigger polar low developments, giving a polar low spetrum. They may ap-

pear as almost purely barolini oralmostpurely onvetive systems (Rassmussen and

Turner, 2003). However, a ombination of these two instabilities is most ommonly

seen. A polar low is thought to develop in a barolini atmosphere through an inter-

ation between an upper-level positive potential vortiity (PV) anomaly whih moves

over a region of strong temperature gradients. The ylone is growing by onverting

warm air and desending old air (Rassmussen and Turner, 2003). The ylone may

ontinue developing as a barolini disturbane, or it may be intensied through ther-

malinstabilitiessuhasthe ConditionalInstabilityofSeond Kind (CISK)mehanism,

orthewind indued surfae heatexhange (WISHE)theory. TheCISK theory isbased

on a reservoir of CAPE, and through a ooperative feedbak between deep onvetion

andthe large sale ow, there maybeaslowintensiationof the ylone (Rasmussen,

1979;Bratseth,1985). TheWISHEtheory(formerlyreferredtoasairseainterationin-

stability;ASII) donotrequire theambientCAPE tointensifytheylone, butthe high

wind speeds indue sensible and latent heat uxes from the sea surfae whih is then

transportedupwardbyturbulentmotionsandonvetion(EmanuelandRotunno,1989).

There are dierent loud signatures assoiated with the dierent foring mehanisms,

wherethe barolinisystems are haraterizedwith a omma loudand the onvetive

systems have a more spiral form shape, often with a loud free eye. A barolini polar

low may develop as reverse shear systems where the horizontal wind speed dereased

with height, and the thermal wind is opposite in diretion to the mean ow. Reverse

shearonditionsarethoughttobeimportantformanypolarlowdevelopments(Kolstad,

2006). Further, polar lows dissipate very quikly after making landfall, as a result of

loosingtheir energy soure. Polarlows are haraterizedwith theirrelative smallsale,

andin the NordiSea areas the horizontalextent ismost ommonly from200 - 500km

(Noer and Ovhed, 2003). Polar lows may bring high impat weather, where there is

heavy preipitationand the windsoften exeedgale fore (14 -17m/s). Polarlows are

awintertimephenomenon, with the high frequeny season from Otoberto Marh.

The observed polar low during the IPY-THORPEX ampaign

The International Polar Year (IPY, 2009) lasted from Marh 2007 to Marh 2009 and

wasa ollaborative sienti eort amongseveral ountries where the fous was onthe

Arti and the Antarti. It was organized through the International Counil for Si-

ene (ICSU) and the World Meteorologial Organization (WMO), and the aimwas to

improve the understanding in polar regions through enhaned researh ativity. The

IPY-THORPEXampaign (IPY-THORPEX, 2009) wasone of the projetsduring the

IPY, and itlasted 3 weeks in February-Marh 2008. It was founded by the Norwegian

ResearhCounilandthemaininterestwastoimproveweatherforeastingofhazardous

weather in the arti region. During the ampaign there were several researhers sta-

tioned onAndøya, an island in Northern Norway, and together they analyzed weather

harts and satellite images to nd where polar lows might develop. A speial airraft

equipped with in-situsensors for basi meteorology and turbulene measurements, and

onewatervaporandonewindLidarsystem inadditiontoadropsondesystem,ewthe

routesthat the sientists planned the day before, and inthis way the observation data

were obtained. Along with the measurements from the ights, there was an enhaned

observationnetworkduringtheampaign;whereseveralradiosondesweredeployedfrom

dierent oastguard ships, the Bear Island and also from Novaja Semlja, Murmansks

and Franz Josefs where some Russian sientists were also partiipating. All of these

(a)

(b)

()

Figure2.4: Satellite images from 11.37 UTC 03.03.08(a) 17.21 UTC 03.03.08 (b) and

(a)

(b)

()

Figure 2.5: MSLP and Z500 from HIRLAM20 analysis at 12 UTC 03.03.08 (a), 18

UTC 03.03.08 and 12 UTC 04.03.08 (). The blue squares indiates the domain set

up for this study. The isobar interval is 2hPa and 40m for Z500. The line N-S is the

this ampaign, the researhers managed to observe and measure a polar low in a way

whih has never been done before.

From 3 through 4 Marh 2008 a polar low developed in the Barents and Norwegian

Seas, and there were 3 ightsin 2 days whih measured the entire lifeyle, two ights

03.03 and one ight 04.03. The satellite images from approximately the same time as

the ights an be seen in Figure 2.4, where 2.4a shows the rst stage of the polarlow

assoiated with the old air outbreak (11.37 UTC 03.03.08), 2.4b the early phase of

the ylone (17.21 UTC 03.03.08), and the seluded ylone just before it made land

fallan be seen in 2.4 (from11.28 UTC 04.03.08). Figure2.4 should be seen together

with Figure2.5 whih shows the HIRLAM20 (20 kmresolution) analysis of MSLP and

the geopotential height at 500 hPa, Z500, from 12 UTC 03.03.08, 18 UTC 03.03.08

and 12UTC 04.03.08, respetively. This polarlowwas assoiated with a synopti low,

whih was loated o the west oast of Norway a ouple of days before the polar low

development. The synopti lowtriggered the old air outbreak,whih isevident inFig

2.4a,wheretheold artiairows overtherelativewarmeroeanforminglongrowsof

stratoumulus (loud streets). The frontalzone, whih separated the shallow, low-level

Arti airmassesfrom the warmer, maritime airoverthe sea, is alsoseen inthe gure,

where ithas a north south orientation onthe west side of Svalbard. Flight1 ew over

thefrontalzoneandreleasedseveraldropsondes. Fromtheobservationdata(notshown

here)astronglow-level,horizontalwindshear arossthe frontalzoneisseen. Thewind

at 925 hPa was observed to be up to 26.2 m/s. The observation data also show very

strongtemperaturegradients aross the frontal zone.

Thenext stagein the developmentseen inFigures2.4b and 2.5b shows the earlyphase

of the polar low. There is still old advetion in the ow from the north and now it

is starting to bring the old air south of the synopti low and in its initial phase of

wrapping in the warm air. There are stillstrong surfae winds on the western ank of

the synopti low, at the same plae wherethe temperature gradientsare largest.

Around 00 UTC 04.03 (not shown) a mesosale vortex uts o the synopti low on

itswest ank, andontinues topropagatetowardsthe oastof Norway. Figure2.4and

2.5shows theseluded ylonejust beforeitmade landfall. Atthis timethe polarlow

had a diameter of approximately 500 km and the HIRLAM20 analysis shows a entral

pressureof996 hPa. The loudbandsarespiralingaroundthelowenter. Forthistime

the ampaign airraft ew immediately above the ore, and from the observation data

it is seen that the low level jet has a wind speed up to 28m/s, and it is evident how

the oldair has been adveted tothe north side of the warm ore. The polarlowmade

landfall around 18 UTC and died out as a result of the lak of energy from the warm

oean.

One of the great paradoxes during this ampaign was the use of foreasts whih the

sientists believed to have bad skill in prediting polar low events. They still had to

use them to analyze when and where a polar low might develop. Sunday 02.03 the

researhers realized they missed a polar low whih hit the middle of the Norwegian

oast the same day. Fortunately the same day they predited a polar low to develop

during the next day and to hit the middle of Norway on Tuesday morning. The de-

terministi HIRLAM20 foreast initialized 18 UTC 02.03.08 predited the large sale

owwithgoodaurayduringthe wholeforeast, butthe smallersale, andespeially

the observed polar, low was not aptured. The rst foreast whih had the polar low

was the HIRLAM20 foreast initialized 00 UTC 03.03.08 (Monday), but it predited

the polar low to make landfall a few hours earlier than atually ourred. The skill in

prediting the polar low inreased with initializing time loser to the polar low event,

and the foreasts initialized on 12 UTC 03.03.08 and beyond have the right strength

and loationof the observed polarlow.

Veriation Methodology

3.1 Standard Methodology

The observation data from the ampaign used in this study onsists of wind speed at

925hPa, inadditiontovertialross setionsof windspeed and potentialtemperature.

Thevertialross setionsareinterpolatedfromthe observationsfromthe drop sondes.

Duringthe threeightstherewere several dropsondesreleased, butforthis purposewe

have only hosen one ross setion from ight 3 to ompare with and the geographial

position an be seen in Figure 2.5. Regular satellite images (Fig. 2.4) and radar

reetivity(Fig. 4.9aand b)willalsobeused toompare withthe foreastsinaddition

to the HIRLAM20 analysis shown in Figure 2.5 and QuikSCAT 1

(not shown). The

standardanalysistoolsatmet.nohavebeenusedtoalulatethemeansealevelpressure

(MSLP) ensemble mean and the spread (

σ

foreast probabilities of wind, preipitation and potential temperature. The ensemble

meanis given by:

x = 1 n

n

X

i=1

x i ,

wherenisnumberofensemble members (inludingthe ontrol) and

x i

value of x for ensemble member i. A measure of the spread between the members is

given as the root meansquare (RMS) deviation from

x

σ =

v u u t 1 n

n

X

i=1

(x i − x) ² .

Bytaking the mean we are lteringout the unpreditable parts in the ow. The small

sale is the most unpreditable part in the ow where the errors are rst saturated,

therefore with the ensemble mean the small sales are rst ltered out. As a result of

inreasingleadtimetheensemblemeanwillgraduallybeomesmootherandonlyretain

thelargesale whihismore preditable. Theensemble meanisexpete dtobeasgood

asthe ontrolrunintheearlyrangeoftheforeast, butbeomemoreskillfulthereafter.

σ

σ

1

QuikSCATdenition: High-resolution satellite-derivedoeansurfaewind.

between the members, and the ensemble mean is expete d to be skillful. Large spread

between the members is an indiation of a less preditable regime, and the ensemble

mean(or the ontrolrun) isnot expete d tobeskillful (but an be luky).

The probability thresholds have been hosen with respet tothe observations, and the

probability isgiven inperentage,where100%(0%) meansthatallthemembers (none)

exeeds the threshold. As disussed in setion 2.1, the foreast probability of dierent

weather events is highly important espeially when it omes to high impat weather.

Probabilistiforeastinganinrease thewarningaheadofaninidenttoalargerextent

than what a deterministi foreast is apable of. In addition, probabilisti foreasting

isa more onsistent way of foreastingthan a deterministiforeast.

In this study we have hosen 3 veriation times where the model results are om-

paredagainsttheobservations. Sinethe ightslastedoverseveral hours,the leadtime

hasbeen hosenwith the goal tobeas loseas possibletothe ighttime. The veria-

tion times are listed inTable 3.1.

Table 3.1: The veriation times of the modelresults against the observation data.

Flighttime Veriation time Lead time

Flight1: 10.09- 13.58 UTC

→

Flight2: 14.56- 18.26 UTC

→

Flight3: 10.14- 13.28 UTC

→

Furthermore, in this experiment two new methods to analyze the foreasts have been

taken into use; pseudo satellite images and traking of polarlows. These two methods

willbe desribed in the following.

3.2 New Methods

3.2.1 Pseudo satellite images

Satelliteimagesofloud top temperatures are of greatimportanefor visualinspetion

andunderstandingofthe evolutionofweathersystems. Itisanimportantanalysistool,

espeiallyindata-sparseareas,forshortrangeweatherforeasting,helpingtodenethe

initial onditions to initialize numerial weather preditions models (NWP), and also

monitoring NWP model performane in the early stage of the foreasts (Bader et al.,

1995). Moreover, satellite images are also important for polar low foreasting, sine

these features are easily deteted in the images. In addition, many studies related to

polarlows havebenetedfromthe satelliteobservations (e.g.Haroldetal.,1999;Bleh-

shmidt,2008). Sine asatelliteimagerygivessuhgoodunderstanding ofthe dierent

parametersand the3D struture ofthe weathersystem, itis desirabletoobtain similar

imagesfrommodel foreasts. Also,the qualityof various model elds an beestimated

an only be made on the basis of satellite images (Tijm, 2004). Therefore a alulated

satellite image from model foreasts simplies the assessment of the dierent parame-

ters, espeially indata-sparse areas.

Tijm (2004) proposed a simple and quik method that estimates both infrared (IR)

andwatervapor(WV)pseudosatelliteimagesbasedonHIRLAMmodel foreasts. The

term pseudo is used to indiate that these images are based on model derived elds

rather than remotely sensed radiation. These pseudo satellite images are analogs to

onventional IR and WV images. The onventional IR imagery is derived from ter-

restrial radiation emitted in the 10 - 12

µ

derivedfromtheradiationemittedbywatervaporatwavelengthsinthe6-7

µ

(Bader etal., 1995).

We have here in this study adapted the method developed by Tijm (2004) suh that

pseudo satellite images an be obtained from UM foreasts as well as HIRLAM. We

have mainly been fousing on the pseudo satellite IR images, therefore only these will

be presented here. In the followingrst a brief presentation of the pseudosatellite im-

ages method and assoiated model variables used in the algorithm will be given. The

sensitivity to some of the parameters have been investigated, and the algorithm has

been veried using the ontrol run of UM-EPS-big in addition to a UM4 (4 km) op-

erationalforeast from met.no. This was done for two purposes: (1) The method was

developed beforeallthe runsinthis studywasdone, and(2) itgivesalargerondene

in the method if it is veried with more than one ase. The operational domain is

mainly loated over land in ontradition to the experimental domain whih is mainly

over oean. The foreast from the operational domain is initialized 12 UTC 16 Marh

2009. The initial and lateral boundary onditions are taken from HIRLAM8 (8 km)

foreast. Besides from this, the model ongurations are the same for both domains,

and isdesribed insetion 2.2.2.

The method

Whereas the satellite retrieved loud top temperatures (CTTs) are inverted from the

remotelymeasuredupwelling radiationatthe top ofthe atmosphere,the model derived

pseudosatelliteimagesestimatethe loudtop temperatureby using temperature, pres-

sureand loudand iewater ontent(Tijm, 2004). Weintegratestartingatthe surfae

usingthesurfae radiationtemperature. Inlear skyonditions thistemperatureorre-

spondstotheCTTinthepseudoimage. Inloudyonditionstheradiationtemperature

ofeahmodel layeris set equal tothe assoiatedmodel temperature. However, itson-

tribution to CTT is dependent on the amount of loud ondensate (liquid and ie) in

the layer. If the amount exeeds a ertainthreshold (see below), the model loud layer

radiatesasa blakbody.

The loudtop temperature inthe IR wavelengthband is alulated with the equation:

T cld = T cld,prev 1 − M IN n

1, Q l ∆P Q dp

o

!

+ T a M IN n

1, Q l ∆P Q dp

o

!

,

where

T cld

T cld,prev

the loud top temperatureat the previous model level.

Q l

box loud and ie water ontent at the urrent level (kg of ondensate per kg of air),

∆

of the model layers,

T a

Q dp

aforementionedthreshold value. Thethreshold

Q dp

kg

m ²

inTijm (2004). From the hydrostatiequation it an easily be seen that

Q l ∆P

portionalto the loud ondensate ontent persquare meter inthe olumnwith height.

Notethat thismethodisderivedforHIRLAM, whihisaquasi-hydrostatimodel. UM

is a non-hydrostati model, but by adapting this method even though it is based on

hydrostatiassumptions has not aetedthe results(see the following).

Startingatthesurfaeand movingupwards,Eq. 3.1issolvedforeahlayer. Ifthereare

noloud oriewater ontent inthe layer, i.e.

Q l

adjusted to the temperature of that level, where the ratio

Q _l ∆P

Q _dp

the CTT should be adjusted to the temperatureof the urrent layer. For values larger

than1, the loud layerbehaves asa blakbody.

The grid box mean loud and ie water ontent are prognosti variables readily avail-

able onthe full-levelsin UM. Potential temperature, fromwhih the temperatureused

in Eq. 3.1 is derived, is also given on the full-levels. In order to save disk spae and

avoidhavingtoolarge outputles, pressure isurrentlyonlyarhivedatthe half-levels,

and therefore is interpolated to the full levels for both the alulation of temperature

and pressure dierene inEq. 3.1.

Note that CTT of Eq. 3.1 will most likely depend on the number of vertial levels

employed in the model. In this study we have not investigated the sensitivity to the

numberof vertial levels.

Results

Pseudosatellite imagesare in operationaluse at met.nofor onlyHIRLAM12 foreasts,

and hene only these are used here for omparison to the UM4 foreasts. HIRLAM12

has a 12 km horizontal grid mesh and 60 vertial levels. In addition, satellite images

losestin time are used for veriation.

Both the satellite and model derived pseudo satellite images are displayed here using

the graphial visualization tool DIANA. Similar to the onventional satellite images,

the whiter (darker) area in the pseudo satellite image, the lower (higher) is the CTT.

However, the blak to white saling is dynami, i.e. the highest (lowest) CTTs within

the model domain will be displayed darkest (whitest). Sine we are mainly interested

inthe spatialgradientsinCTT, this dynamisaling issuient fordierentiatingbe-

tween high and lowlouds inthe model. However, for a more detailedomparison it is

(a) (b) ()

Figure 3.1: Pseudo satellite image (a) ompared against the UM total loud over (b)

andinverted total loud over (). Thehigh oluds are seen to the south and north-east

in the domains, and loud free regions are found in the middle of the domains. Lower

andmiddlelevel louds are seen to thenorth westand west. See textfor further details.

Thiswillmainly aetthe CTTswhere thereare louds,sine the CTToveroeanand

land where there are no louds and, to some extent, oinide well with the observed

darkolor (see the following).

Sensitivity tests

We have investigated the sensitivity of some of the parameters used in the algorithm.

Sinethe pressure eld is interpolatedfrom halfto fulllevels, we wanted to investigate

the sensitivity of this interpolation. The CTT image was alulated using pressure at

both half and full levels with the ontrol run from UM-EPS-big (not shown). There

were nodierenes detetedbyvisualinspetions,thereforewe onlude thatour inter-

polationof pressure fromhalf -tofull -levelsissuiently aurate forthe alulation

ofCTT in Eq. 3.1.

Thesurfae radiationtemperatureinHIRLAM isthe near surfae temperature (T2m).

InUM there is adiagnosti surfae temperature, Ts, whih isat 0mand willtherefore

better represent the surfae radiation temperature. To see if using T2m instead of Ts

gives signiant dierenes, we also estimated CTT by using Ts with the operational

UM4foreast (not shown). As expete d the sensitivity tothe hoie of lowerboundary

temperature (T2m orTs) is not very large. Hene, we deided toreplae T2m with Ts

inEq. 3.1.

Comparingwith the diagnosti total loud over in UM

Havingdeided onusinginterpolatedpressureeldsandTsinEq. 3.1, wenext ompare

the pseudo images to the modeled diagnosti total loud over, seen inFigure 3.1. We

use the operationalUM4 foreast, sameas inthe abovesubsetion. Forthe totalloud

over image (Figure 3.1b), white areas (i.e. areas without any shading) are indiative

of loud free grid regions, whereas the darker the grid box the larger the loud over.

Obviously, this is opposite to how it is seen in the satellite images, and an as a on-

sequen e easily be onfused. We have therefore alsoinverted the shading, Figure 3.1.

The advantage of using pseudo satellite images instead of total loud over is evident.

For instane, in the pseudo satellite image one an distinguish between high and low

louds. Nevertheless, the louds in the pseudo satellite image are o-loated with the

total loud over. Hene, we are ondent that Eq. 3.1 is able to detet the modeled

louds.

Comparingwith satellite IR images and HIRLAM pseudosatellite images

Pseudoimages are alulated from the operationalUM4 and HIRLAM12 foreasts ini-

tialized at12UTC 16Marh2009. Two foreast lead times,T +6h(afternoon) and T

+18h (early morning), are shown inFigure3.2. The orresponding satellite IR images

seen inFigs. 3.2 a and bare used for veriation.

First we ompare the UM4 images with the satellite IR images. At lead time T +

6h, the same features seen in the observed image also appear in the pseudo satellite

image. The high louds south - east and north - east in the UM4 domain, orrespond

to what is observed. There are more dierenes omparing these two images west in

the UM4domain,along the Norwegian oast. There are middlelevellouds foreasted,

onsistent with IR-images. Thus the amountof louds foreasted is less than observed.

BothimageshavelearskyonditionseastinNorthern-NorwayandSweden. However,

the surfae radiation temperature over land where there are no louds in UM appears

darker then to the observed, and this is also the ase for the radiation temperature in

loudfree areas overoean. At lead timeT +18h, the same haraters inboth images

are still seen. It is striking how well the louds in the pseudo satellite image are o-

loated with the satelliteimage. Though, there are fewerlouds in the pseudo satellite

image than what is observed. But overall the pseudo satellite images in UM oinide

wellwith the observed satelliteimages.

Having seen that the UM4 pseudo images ompare well with the satellite images, we

next ompare them to the HIRLAM12 images. Remember that the pseudo method

was originally developed for HIRLAM (Tijm, 2004). Due to the oarse grid mesh in

HIRLAM12the CTTeld is smoother and the range inthe CTT values is not aslarge

as in the satelliteIR imageor the UM4 image. Also, UM4 shows more distint dier-

enesbetweenoeanandland,andmoredetailedstrutures appear. Thisisdueboth to

thehigherhorizontalresolutionandthefatthatHIRLAMusesT2m. Generally,pseudo

satellite images from UM resemble the observed satellite images more than HIRLAM

(a) (b)

() (d)

(e) (f)

Figure3.2: IRsatelliteimages and pseudo satellite images from UM4 and HIRLAM12.

The foreasts were initialized at 12 UTC on 16 Marh 2009. The IR satellite images

are valid at 1808 UTC 16 Marh 2009 (a) and 0600 UTC 17 Marh 2009 (b). Pseudo

satellite images at lead time T + 6h from UM4 () and HIRLAM12 (e), and at T +

18h from UM4 (d) and HIRLAM12 (f). For onveniene the UM4 domain is shown in

pseudosatellite images.

Theuse of pseudo satellite images in this study

Sineloudsareaprodutofallproessesintheatmosphere,thepseudosatelliteimages

gives a good understanding of the 3D struture of the weather systems in the model.

In addition, these images give a diret judgment of the quality of various model pa-

rameters, espeially in data - sparse areas. In this study we will alulate the pseudo

satelliteimageson the outputs from UM-EPS and ompare them with satellite images

(presented inChapter 4).

3.2.2 Trakin g polar lows

Forpolarlowforeasting, as wellas deteting the real world ylones, it isimportant

toexlude falselyidentiedphenomenainmodel outputs. Several previousstudieshave

applied a traking algorithm (e.g. Hodges, 1994, 1995, 1999) on dierent model elds

basedon anautomated methodto identify synopti systems and provide statistialin-

formation about their positions, intensities and the genesis and lysis (the spatial and

temporal distribution of the development and the ending of the ylone) (e.g. Hoskins

and Hodges, 2002; Froude et al., 2007a,b). Note that the term trak here refers to

the trajetory of an individual storm, rather than the average trak of many storms

(Froude et al., 2007a). The traking tehnique is an essential foreast validation tool,

and it gives diret information about the model's ability to predit polar lows (Zahn

and Storh, 2008).

Inthe samemanner astraking synoptisystems it would givevaluable informationto

trak polarlows. Forinstane, Zahn and Storh (2008) haveemployed a near isotropi

bandpass lter to extrat mesosale parts of the MSLP elds from a two year long

simulations with CLM (a limate version of the Loal Model of the German Weather

Servie), and a traking methodology has been applied on the elds, aiming to repro-

duethelimatologyofpolarlowsoveratwoyearlongsimulation. Afterperformingthe

traking algorithm, there were too many deteted traks. However, introduing several

objetive riteria,the number of the deteted polarlows dereased.

In this study we have employed and modied the traking algorithm developed by

Hodges (1994, 1995, 1999) for traking polar lows instead of synopti sale systems.

The model elds are taken from UM-EPS, aiming to reprodue the trak of the ob-

servedpolarlowthatdevelopedduringtheIPY-THORPEXampaign. Someadditional

onstraints have been introdued sine this method was originally made for deteting

synopti systems. First the method TRACK will be presented, and then the modi-

ations applied will be desribed. In addition, further objetive riteria adapted from

the study of Zahn and Storh (2008) are presented. It should be mentioned that with

thediagnosisfromTRACK, broadstatistialinformationabout the deteted traks are

obtained, and only a small part is investigated here. To fully utilize this analysis tool,

itshould be performed onseveral ases.

The traking algorithm

Historially,therehavebeen twobasiapproahestodiagnosestormtraks,anEulerian

approah and a Langrangian feature point traking. Sine the beginning of superom-

putersthe Eulerianapproah has been the onvenient way toompute simple statistis

from NWPs at a set of grid point with a frequeny band representative of synopti

timesales (Hodges, 1999; Hoskinsand Hodges, 2002; Anderson etal.,2003). The Lan-

grangianfeaturetrakingapproahhas been usedsine theend ofthe nineteenentury,

and the early studieswere based on manual analysis using dailysynopti harts. With

the introdution of NWP models the approah has been advaned further and obje-

tive,automatedmethodshavebeenadopted. Thisprovidesgoodstatistialinformation

that desribes the storm trak ativity of the synopti system (Hodges, 1994; Hoskins

and Hodges, 2002). A feature traking algorithm developed by Hodges (1994, 1995,

1999)has been usedextensivelyinseveralstudies(e.g. Hoskinsand Hodges,2002,2005;

Froude etal., 2007a,b), aimingto trak synopti systems, and to ompute their statis-

tial properties and limatology and assess foreast skill and preditability of dierent

models. The traking algorithm has been adapted into this study to trak polar lows,

whih are of sub-synopti sales.

The basis of the method is to searh for maxima or minima in meteorologial elds,

and a range of elds an be used; MSLP, geopotential at pressure surfae e.g. 500

hPa (Z500), meridional wind (v), temperature (T), potentialtemperature (

θ

veloity (

w

ζ

algorithm has been performed on the MSLP and vortiity elds, and only these will

be onsidered here. The hoie of the eld should be done on the basis of what sale

is to be traked. MSLP is distintly inuened by strong bakground ow, and large

spatial sales and relative slower moving systems dominate. This yields MSLP eld a

better hoie to trak larger sales.

ζ

andtherefore tendtobeabetter eld foridentifyingsmallersales. However, there are

some disadvantages with employingMSLP and

ζ

tohow the extrapolation isperformed,and alsoto the representation of the orography

inthe model (Hoskins and Hodges, 2002).

Assumewe have hosen one ofthe meteorologialeldsfor traking. Beforeperforming

the algorithm, the bakground owis removed to only retain the mesosale part. This

isdone byrst performingaspetralspatialltering. The eldisthenrepresented by a

spherial harmoni expansion and the smallest and large spatial sales an beremoved

(Andersonetal.,2003). Theresult isalteredeld withspatialsalesrepresentativeof

polarlows. Thenthetrakingalgorithmisperformedonthelteredeld. Therststep

in the algorithm is the determination of feature points, whih are the positions of the

extremainthe hoseneld. Thenext step inthemethodistodeterminethe orrespon-

denebetweenthefeaturepoints,andthe aimistondtheset oftraksthatmaximizes

thesmoothnessofthe trajetories. This isdoneby minimizingaost funtion(Hodges,

1994,1995,1999). The detetedfeaturepointsare thenonstrainedtohaveaminimum

for the vortiity eld a given threshold needs to be exeeded. The feature points that

fulllthe requirements are linked together and give a trajetory. Now the traks with

theirrespetivestatistialinformationare storedinareord, andit isup tothe user to

interpret the trak diagnosis. The algorithm makes use of these riteria related to the

minimum horizontal displaement, the lifetime and the vortiity threshold and an be

hosen. Thesewillbedisussed inthe following.

The traking algorithm adopted to this study

Thetrakingalgorithmdeveloped byHodges(1994, 1995,1999) hasbeenperformedon

the foreasts made by UM-EPS, and the results will be given in Chapter 4. Here the

dierent aspets investigated before deiding on the nal riteria to detet and trak

the polar low desribed in setion 2.3 will be presented. All the tests have been per-

formed on the ontrol run of UM-EPS-small. In this study we have mainly hosen to

perform the traking algorithm on the vortiity eld, but a few tests were also done

on the MSLP eld. A summary is given in Table 3.2. When performing the traking

algorithm,asearhof maximuminthevortiity eldandaminimumintheMSLPeld

is done. This will result in too many deteted points due to many small sale loal

maximum and minimum. To pik out the trak of the features representative of polar

lows, additional onstraints are needed. This is rst done by a minimum horizontal

displaement distane (from the rst deteted point to the last deteted point) over a

minimum lifetime. At rst the minimum displaement was set to 10

o

(approximately

1000 km). Polarlows an be stationary orhave a very slow southward motion,so this

riterionwas swithed o. The minimum lifetime was originally set to 24 hours. This

onstraint is a bit more omplex, espeially for limited area foreasting. A polar low

whihalready has exisedfor awhileoutsidethe domainand isenteringthe domainon

the lateral boundaries, will be exluded with a too long life time onstrain. Polar lows

tend toexist forat least 12hours, therefore we deided to set the minimum lifetime to

12hours. All the tests in Table 3.2 have zero minimum horizontal displaement and a

minimumlifetimeof12hours. Belowisgivenapresentationofthedierenttestesdone.

In previous studies when the traking algorithm has been used to detet synopti sys-

tems, planetary sales with total wave number less than or equal to 5 have been l-

tered out (Froude etal., 2007a;Hoskins and Hodges, 2002, 2005). Sine polarlows are

mesosalefeatures, and an haveadiameter up to 1000km, weinitially removed sales

between 200 and 1000 km, but this gave too many traks. In the study of Zahn and

Storh (2008)a lteringof 200 - 600 km was used, and we deided to use the same.

The removal of the bakground state was investigated in Anderson et al. (2003)where

thesensitivity tothe spetral spatiallter wasexplored. They inreased the numberof

totalwavenumbers removed ontheMSLPeld, from5to7to10. Withtheremovalof

10wave numbers, the nature of the synopti feature started todeteriorate. From Test

14(200- 1000km) and 15(200- 600 km)inTable 3.2 we see that whendereasing the

lteringintervalonthe MSLPeldthereisaninreaseinnumberof traks, from6(200

Table 3.2: The dierent test performed on the hosen eld with the traking algorithm.

All the tests have 0 in minimum horizontal displaement and a minimum lifetime of 12

h. The deteted traks are the traks after the vortiity threshold and ltering interval

onstraint.

Test nr Field Threshold Filtering Deteted Traks

[

s ⁻¹

1 VOR850

2 × 10 ⁻⁵

2 VOR850

1 × 10 ⁻⁵

3 VOR850

1 × 10 ⁻⁶

4 VOR850

1 × 10 ⁻⁴

5 VOR850

1 × 10 ⁻⁴

6 VOR850

1 × 10 ⁻⁵

7 VOR850

2 × 10 ⁻⁵

8 VOR925

2 × 10 ⁻⁵

9 VOR925

1 × 10 ⁻⁵

10 VOR925

1 × 10 ⁻⁴

11 VOR925

1 × 10 ⁻⁴

12 VOR925

1 × 10 ⁻⁵

13 VOR925

2 × 10 ⁻⁵

14 MSLP 200-1000 6

15 MSLP 200-600 9

sameis seen for the vortiity eld when there isa low vortiity threshold (the vortiity

threshold will be disussed later). Dereasing the ltering interval from 200-1000 km

(Test 1,2,8,9)to200-600 km(Test 6,7,12,13) thenumberoftraksinreases. When

there is a high vortiity threshold, reduing the ltering interval from 200 - 1000 km

(Test 4,10) to200- 600 km(Test 5,11) the numberof traksdereases. This indiates

that with a low vortiity threshold, the eld is more inuened by larger sales, and is

therefore more sensitive to the ltering interval. This gives ondene in hoosing the

vortiityeldwithahighvortiitythresholdandanarrowerlteringintervaltoperform

thetrakingalgorithmon,sineitislesssensitivetotheremovalofthebakgroundstate

and alsobetter suited for deteting smallersale system than MSLP. Therefore in this

study we willperform the spetral spatial lteringonthe vortiityeld with a ltering

of200 -600 km.

Initially, the traking was performed on vortiity at 850 hPa, whih is also used to

trak synopti systems. Polar lows have a smaller sale than synopti lows, and also

ome with very strong surfae wind, in addition they do not neessarily penetrate as

highupintheatmosphere. Therefore 925hPavortiityeldwould bemoreappropriate