Best-member-selection as part of an ensemble based data ...

30
Eidgenössisches Departement des Innern EDI Bundesamt für Meteorologie und Klimatologie MeteoSchweiz Best-member-selection as part of an ensemble based data assimilation approach COSMO/CLM User Seminar, Langen, 09.-11.03.2009 Tanja Weusthoff 1 , Daniel Leuenberger 1 , Christian Keil 2 , George Craig 2 1 MeteoSwiss, 2 DLR

Transcript of Best-member-selection as part of an ensemble based data ...

Page 1: Best-member-selection as part of an ensemble based data ...

Eidgenössisches Departement des Innern EDIBundesamt für Meteorologie und Klimatologie MeteoSchweiz

Best-member-selectionas part of an ensemble baseddata assimilation approach

COSMO/CLM User Seminar, Langen, 09.-11.03.2009

Tanja Weusthoff1, Daniel Leuenberger1, Christian Keil2, George Craig2

1MeteoSwiss, 2DLR

Page 2: Best-member-selection as part of an ensemble based data ...

2Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

New data assimilation approach: SIRF

Driving EPS (SREPS)

COSMO-DE-EPS

Assimilation period time

Best-Member-Selection 1

Best-Member-Selection 2

Ensemble Enhancement/Resampling

Page 3: Best-member-selection as part of an ensemble based data ...

3Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Objectives

WP1: Evaluation of classical and spatial metrics for the determination of weighting schemes for ensemble members (DLR, MeteoSchweiz)

WP1.1 Implement spatial metrics and evaluation on selected COPS eventsFQM of Keil and Craig (2007), which is already available, SAL (Wernli et al. 2007)fuzzy verification package of Ebert (2007).

WP1.2 Implement classical metrics for conventional observations

WP1.3 Investigate correlation of metrics between models of different resolution

WP1.4 Compare object-oriented and classical metrics

WP1.5 Investigate persistence of skill in different metrics and different meteorological situations

Page 4: Best-member-selection as part of an ensemble based data ...

4Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Concept

)]([)]([ 1 xHyARxHym T −−= −

Apply classical quadratical error measure for conventional observations:

y = vector of observationsH(x) = vector of forecasts

H = forward operator, H(x) = observation equivalent, calculated from model state x

R = covarianz-matrix of the observation errors, DiagonalmatrixA = additional matrix for the weightings of parameters, Diagonalmatrix

)]([)]([ 1 xHyRxHy T −− −

In order to give more important observations a higher weighting („tuning“), an additional matrix A is multiplied to the error covariance matrix R containing the setting of the weights.

⎩⎨⎧ =

=else

jiA i

ij 0α

⎩⎨⎧ =

=−

elseji

R ijij 0

/1)(

21 σ

Page 5: Best-member-selection as part of an ensemble based data ...

5Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Quality measure m= distance of model member x(k) to observations

α⋅=Dm k)(

[ ] )(1 )()( ~~ ki

Tki

k dARdm −=

[ ]2

2 )(

1

)(~

ii

ki

i

Y

i

k dmσ

α=∑=

optimal weighting α to be determined

)]([)]([ 1 xHyARxHym T −−= −

Yi ,...1=Y = # variables * # stations * # levels

Kk ,...1=K = # member

)(~ )()( ki

ki xHyd −=

Page 6: Best-member-selection as part of an ensemble based data ...

6Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Quality measure m= distance of model member x(k) to observations

α⋅=Dm k)(

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎟⎟⎟⎟⎟⎟

⎜⎜⎜⎜⎜⎜

=

⎟⎟⎟⎟⎟⎟⎟

⎜⎜⎜⎜⎜⎜⎜

Y

KY

KK

Y

K ddd

dddddd

m

mm

α

ααα

...

...

............

...

...

3

2

1

)()(2

)(1

)2(2

)2(1

)1()1(3

)1(2

)1(1)2(

)1(

Page 7: Best-member-selection as part of an ensemble based data ...

7Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

• High resolution ensemble with 20 members, 00 UTC runs

• 4 driving models:• Member 1-5 EZMWF• Member 6-10 GME• Member 11-15 NCEP• Member 16-20 UKMO

• 5 perturbations of model physics• Members 1, 6,11,16 Entrainment rate (entrscv=0.002)• Members 2, 7,12,17 Cloud cover at saturation (clc_diag=0.5)• Members 3, 8,13,18 Laminar layer depth (rlam_heat=50)• Members 4, 9,14,19 Laminar layer depth (rlam_heat=0.1)• Members 5,10,15,20 Turbulent lengh scale (tur_len=150)

Forecast: COSMO-DE-EPS

Page 8: Best-member-selection as part of an ensemble based data ...

8Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Best-member-selection based on standard observations like Synop(hourly), Radiosondes (6 hourly or less), …

First investigation with radiosondes• Variables on pressure levels:

• temperature, • relative humidity, • u- and v- wind components• …

• Integrated variables: • CAPE• CIN• integrated water vapour (IWV)• …

Observations

Page 9: Best-member-selection as part of an ensemble based data ...

9Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Weighting …

Seven variablesTemperature T, relative humidity RH,Wind components U and VCAPEIWVCIN

T RH U V CAPE IWV CIN

1 1 1 1 1 1 1

Initial Weighting everything is set to the same weight 1

Three level regions1000 – 700 hPa650 – 400 hPa350 – 100 hPa

21 levels

singlelevels

1000-700

650-400

350-100

1 1 1

Weightings of variables and levels are multiplied!

Page 10: Best-member-selection as part of an ensemble based data ...

10Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

… & EvaluationQ: Do the target quality measure and our m identify the same best members?

1. Correlation between target quality r and m

2. Distance between the two quality measures:

Need m(k) and r(k) in the same range, e.g. [0,1] in order to determinethe optimal weighting α = [0,1] such that

Here: Target quality r = DAS

min=⋅−=− αDrmr

mr −

mr

mrcorσσ ⋅

= ,cov

Page 11: Best-member-selection as part of an ensemble based data ...

11Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Target Quality: DAS• The target quality r of the ensemble is provided by the quality measure with

respect to precipitation (radar) the Displacement and Amplitude Score DAS (Keil and Craig, 2009)

• DAS is based on an areal image matcher using classical optical flowtechnique & has 2 components: • displacement error field DIS • amplitude error AMP (RMS of observed and morphed forecast imagery)

Dmax maximum search distanceI0 characteristic intensity

Page 12: Best-member-selection as part of an ensemble based data ...

12Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

DAS,

hourly calculated

DAS,

6-hourly mean(forward averaged)

DAS,

every 6 h (at start time of averaging period)

Target Quality: DAS EZMWFGMENCEPUKMO

DAS at specifictime is highlycorrelated with 6-h mean

Page 13: Best-member-selection as part of an ensemble based data ...

13Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

First Results… investigated period: 08.-16.08.2007

10738 Stuttgart - Echterdingen

0

2

4

6

8

10

12

14

20070808 20070810 20070812 20070814 20070816

rain

sum

[mm

] and

tota

l sun

shin

e du

ratio

n [h

]

0

10

20

30

40

50

60

70

80

90

100

mea

n re

lativ

e hu

mid

ity [%

]

sunshine duratiom rain amountcloud coverage [1/8] relative humidity

Mesoscale forcingMesoscale forcing Convection

Page 14: Best-member-selection as part of an ensemble based data ...

14Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

1 DeBilt

2 Payerne

3 Nancy

4 Schleswig

5 Greifswald

6 Emden

7 Bergen

8 Lindenberg

9 Essen

10 Meiningen

11 Idar-Oberstein

12 Stuttgart

13 Kuemmersbruck

14 Muenchen

15 Hohenpeissenberg

16 Prag

1

2

3

4 5

67

8

9

10

11

1213

1415

16

First Results… investigated period: 08.-16.08.2007… 8 radiosondes within considered DAS area

Page 15: Best-member-selection as part of an ensemble based data ...

15Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Equally weighted variables and levels

EZMWFGMENCEPUKMO

m

das

Maximum in the afternoon, minimum in the morning(new run);

Partly clustering memberswith same driving models

Page 16: Best-member-selection as part of an ensemble based data ...

16Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Scatterplots (m vs DAS)

EZMWFGMENCEPUKMO

Page 17: Best-member-selection as part of an ensemble based data ...

17Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Scatterplots (m vs DAS),Individual members

EZMWF

GME

NCEP

UKMO

0.11 – 0.24

0.22 – 0.33

-0.01 - 0.08

0.11 – 0.19

Range of correlation…

Page 18: Best-member-selection as part of an ensemble based data ...

18Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

CorrelationIndividual time steps…

large variation of correlation with time

correlation apparentlyhigher for mesoscaledriven days…

Mesoscale forcingMesoscale forcing Convection

Page 19: Best-member-selection as part of an ensemble based data ...

19Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Frequency distribution

histogram of normalizedquality measures:

das/max(das) and m/(max(m))

same as above, but: das x=das2

PDF become more similar,

Decrease in |das-m|

Page 20: Best-member-selection as part of an ensemble based data ...

20Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Sensitivity studies1. Variables

T RH U V CAPE IWV CIN

0.5 0.5 0.5 0.5 0.5 0.5 0.5

1000-700

650-400

350-100

1 1 1

Procedure: Set all variables to 0.5, vary each variable from 0 to 1.0 in steps of 0.1. Level weightings stay 1.

RH and CAPE showstrongest (positive) response to weighting

Page 21: Best-member-selection as part of an ensemble based data ...

21Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Sensitivity studies2. Levels

T RH U V CAPE IWV CIN

1 1 1 1 1 1 1

1000-700

650-400

350-100

0.5 0.5 0.5

Procedure: Set all levels to 0.5, vary each level from 0 to 1.0 in steps of 0.1. Level weightings stay 1.

Levels between 650 and 400 hPa show strongest (positive) response to weighting

Page 22: Best-member-selection as part of an ensemble based data ...

22Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Combination of variables and levels …T RH U V CAPE IWV CIN

0.4 1 0.4 0.6 0.1 0.1 0.1

1000-700

650-400

350-100

0 1 0

EZMWFGMENCEPUKMO

m

das

Page 23: Best-member-selection as part of an ensemble based data ...

23Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

EZMWF

GME

NCEP

UKMO

0.39 – 0.53

0.32 – 0.40

Range of correlation…

(for equallyweighted)

Scatterplots (m vs DAS),Individual members

(0.22 – 0.33)

(0.11 – 0.19)

0.25 – 0.31

0.31 - 0.38

(0.11 – 0.24)

(-0.01 - 0.08)

Page 24: Best-member-selection as part of an ensemble based data ...

24Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Comparison

• Equally weighted • Combined weightingT RH U V CAPE IWV CIN

0.4 1 0.4 0.6 0.1 0.1 0.1

1000-700

650-400

350-100

0 1 0

T RH U V CAPE IWV CIN

1 1 1 1 1 1 1

1000-700

650-400

350-100

1 1 1

Corr(m,das)

TOTAL(|das-m|)

Corr(m,das²)

TOTAL(|das²-m|)

0.17

241.93

0.16

108.68

0.37

242.62

0.36

109.36

Rarely a change in distance |das - m|, but increase in correlation!

Page 25: Best-member-selection as part of an ensemble based data ...

25Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Preliminary Fazit

• Radiosondes do give information on best members in terms of precipitation forecast (as measured by DAS).

• Test with radiosoundings reveal a sensitivity of the best memberselection with respect to parameter and heights.

• The correlation with DAS could be improved by using radiosondesdata with different weightings for variables and heights.

• The correlation between the two quality measures DAS and m isapparently time-dependent.

• The correlations are in the order of 0.4, while DAS values at startingtime of averaging period are highly correlated with DAS averaged overnext 6 hours

the precipitation measure DAS provides a good basis for the best member selection. The use of standard observation as alternative quality measure seems to be appropriate.

Page 26: Best-member-selection as part of an ensemble based data ...

26Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Outlook

• find suitable combination of only a few parameters fromradiosondes and synop stations (e.g. total energy)

• determine optimal weighting of those parameters withrespect to target quality measure r

• expand best member selection to driving ensemble

Page 27: Best-member-selection as part of an ensemble based data ...

27Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Thank you for the attention!

Page 28: Best-member-selection as part of an ensemble based data ...

28Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Page 29: Best-member-selection as part of an ensemble based data ...

29Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

Reference: Keil, C. and G. C. Craig, 2009: A displacement and amplitude score employing an optical flow technique, submitted to Wea. and Forecasting.

Page 30: Best-member-selection as part of an ensemble based data ...

30Best-member-selection as part of an ensemble based data assimilation approach | COSMO/CLM User SeminarTanja Weusthoff et al.

∫∫ == dzqdzIWV w ρρ

q = spezifische Feuchte

CAPE, CIN…(max CAPE)

Calculation of variables

Integrated Water Vapour