1

Satellite Winds Superobbing

Howard BergerMary Forsythe

John EyreSean Healy

ImageCourtesyof UW - CIMSS

Hurricane OpalOctober 1995

2

Outline

•Background/Problem

•Superob Methodology•Method•Observation Error

•Results

•Conclusions/Future Work

Problem:• High - Resolution satellite wind data sets showed negative impact (Butterworth and Ingleby, 2000) Why?

•Suspected that observations errors were spatially correlated

•To account for this negative impact, wind data were/are thinned to 2º x 2º x 100 hPa boxes

4

•Bormann et al. (2002) compared wind data to co-located radiosondes showing statistically significant spatial error correlations up to 800 km.

Cor

rela

tion

Met-7 W V NH Correlations

Graphic fromBormann et al.2002

5

Question:

Can we lower the data volume to reduce the effect of correlated error while making some use of the high-resolution data?

6

Proposed Solution:

Average the observation - background

(innovations) within a prescribed 3-d box to create

a superobservation.

7

Advantages:

•Data volume is reduced to same resolution that resulted from thinning.

•Averaging removes some of the random, uncorrelated error within the data.

8

SuperobbingMethod:

9

1) Sort observations into 2º x 2º x 100 hPa boxes.

28 N

16 W

26 N

18 W

10

2) Within each box: Average u and v component innovations, latitude, longitude and pressures.

28 N

26 N

16 W18 W

11

3) Find observation that is closest to average position and add averaged innovation to thebackground value at that observation location.

26 N

28 N

16 W18 W

12

Superob Observation Error

13

•Superobbing removes some of the random observation error.

•This new error can be approximated by making a few assumptions about the errors within the background and the observation.

Superob Observation Error

14

Superob Observation Error

Assume that within a box:

•Observation and background errors not correlated with each other.

•Background errors fully correlated.

•Background errors have the same magnitude.

15

Assumptions (cont):

•All of the innovations weighted equally.

•Constant observation error correlation.

Superob Observation Error

Token Evil Math Slide

2 ( )Tse W DED W=

2se = Superob Observation Error

W =Vector of Weights (1xN)

D =Diagonal matrix of component observation errors (N x N)

E =Observation Error Correlation Matrix (NxN)

1

1

a a

aE

a

a a

=

L

O O M

M O O

L

Observation Correlation Matrix

a =Correlation within box. Value calculated from correlation function in Bormann et al., 2002

1 1 1 2 1

2 1

1

n

n n n

c c c c c c

c cE

c c c c

=

L

O O M

M O O M

L L

i jc c =Correlation of ith observation with jth observation

18

00z 10 June, 2003.(20 N - 40 N) (0E 30 E)

Old Observation Error Superob Error

19

Experimental Design

•Control Run:•Operational Set up plus GOES BUFR VIS/IR/WV winds•GOES-9 is still Satob format

•Thinning to 2˚ x 2˚ x 100 hPa boxes

•Superob Experiment•Same as control run, except winds are superobbed to 2˚ x 2˚ x 100 hPa boxes

20

• Trial Period: 24 Jan -17 Feb 2004

• 4 Analyses and 6-hr forecasts• 00z,06z,12z,18z

• 1 analysis and 5-day forecast (12z)

Experimental Design (cont)

21

Token Model Info Slide

• Grid – point model (288 E-W x 217 N-S)

• Staggered Arakawa C-Grid

• Approx 100 km horizontal resolution (one-half operational resolution)

• 38 levels hybrid-eta configuration

• 3D-Var Data Assimilation

22

Results

• % normalized root mean square (rms) error against control rms differences calculated for:

•Mean sea-level pressure (PMSL) •500 hPa height (H500) •850 hPa wind (W850) •250 hPa wind (W250)

• In regions:•Northern Hemisphere (NH) •Tropics (TR) •Southern Hemisphere (SH)

•For forecast periods of:•T+24, T+48, T+72 ,T+96 , T+120

Trial Statistics

24

-2

-1

0

1

2

3

4

PM

SL T

+2

4

PM

SL T

+4

8

PM

SL T

+7

2

PM

SL T

+9

6

PM

SL T

+1

20

H500 T

+24

H500 T

+48

H500 T

+72

W250 T

+24

W850 T

+24

W850 T

+48

W850 T

+72

W250 T

+24

PM

SL T

+2

4

PM

SL T

+4

8

PM

SL T

+7

2

PM

SL T

+9

6

PM

SL T

+1

20

H500 T

+24

H500 T

+48

H500 T

+72

W250 T

+24

Exp

erim

ent –

Con

trol

RM

S E

rror

(%

)

TP – ObservationsTP – Analysis

NH – ObservationsNH – Analysis

SH – ObservationsSH – Analysis

25

Anomaly Correlations

Vs. Forecast RangeCompared to

Analysis500 hPa Height

NH

SH

TR

26

T+24 Forecast – Sonde

RMS Vector Error 250 hPa Wind

NH

TR

SH

27

250 hPa u-componentAnalysis Increments

28

250 hPa u-componentAnalysis Increments

29

Results Summary

•Superobbing experiment results are small and mixed

•Generally more positive in the northern hemisphere than in the southern hemisphere or tropics

•Time series results are mixed: Some forecasts better than control, some worse

30

Implications

•Mixed results suggest either:•Random Error not most significant error component of AMVs

•Superobbing set up not ideal to treat random error

31

Future Work

•Back to basics approach•Re-calculate observation errors from innovation statistics

•Experiment with “model independent” quality indicators and “model independent” components in Bufr file

32

• Stripped down impact experiment (i.e no ATOVS radiances)

• Experiment using simulated AMV’s in Met Office System

• Ideas from IWW!!!!!

Future Work (cont)