stephanie guedj florence rabier vincent guidard benjamin ménétrier observation error estimation in...

20
Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Upload: osborne-newman

Post on 01-Jan-2016

222 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Stephanie GuedjFlorence RabierVincent Guidard

Benjamin Ménétrier

Observation error estimation in a convective-scale NWP system

Page 2: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Introduction

1. SEVIRI assimilation experiments (various observation densities)

2. Diagnosis of error correlationsSEVIRIIASI

Conclusions and Future work

Outline

Page 3: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

IRS : horizontal resolution of SEVERI and spectral resolution ~ IASI

AROME WMED (Fourrié et al., 2014)•Aims to support HyMeX campaignsto improve our understanding of the water cycle, with emphases on thepredictability and evolution of intense events

•Is inherited from the operational AROME/FRANCE model (Seity et al., 2011 and Brousseau et al., 2008)

Resolutions : 60 vertical levels, Horizontal 2.5 km

Assimilation : 3D-Var assimilation system used to produce 8 daily analysis usingconventional data, reflectivity, radar Doppler, GEO winds, GEO/LEO radiances …

ContextPotential of MTG for Convective scale NWP models

AROME WMED domain

Page 4: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

SEVIRI WV 6.2 observations Assimilated vs Rejected

1. SEVIRI assimilation experiments Overview of SEVIRI assimilation in AROME-WMED

Horizontal R. 4kmThinning 70km

Repeat cycle 15 min

Analysis every 3h

InformationHumidity

(±400 hPa)

17/10/2011 - 0UTC

Page 5: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

1. SEVIRI assimilation experiments High-density assimilation experiments : configurations (No-cycled)

BACKGROUND (with all observations previously assimilated)

70 km40 km20 km10 km5 km 100 km

Assimilation of SEVIRI WV only

Thinningdistances

Evaluation :1) Analysis Increments2) Forecast verification using independent observations (IASI, radiosondes …)

Ana-70Ana-40Ana-20Ana-10Ana-5 Ana-100

F3h-70F3h-40F3h-20F3h-10F3h-5 F3h-100

Current OPER

Page 6: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

1. SEVIRI assimilation experiments Analysis increments

Analysis – Backgroundspecific humidity (630 hPa)

17/10/2011, 0UTC

70 km

10 km

moisturemoisture

• Increments show similar but sharper structures in EXP10 than EXP70.

Page 7: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

1. SEVIRI assimilation experiments Analysis increments

Analysis – Backgroundspecific humidity (Cross-section)

17/10/2011, 0UTC

70 km

10 km

moisturemoisture

• Increments show similar but sharper structures in EXP10 than EXP70.

• Wrong propagation toward the surface ?

Page 8: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

1. SEVIRI assimilation experiments Forecast Verification

F3h vs IASI radiancesFg-departures (8 days)

(from 17/10 – 0h to 24/10 – 21h 2011)

• The bias in FGd to IASI high-peaking WV channels is significantly improved.

Page 9: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

1. SEVIRI assimilation experiments Forecast Verification

Scores for 2 WV IASI channels Fg-departures as a function of thinning distances for SEVIRI assimilation

F3h vs IASI radiancesFg-departures (8 days)

(from 17/10 – 0h to 24/10 – 21h 2011)

The RMS indicates a degradation of the F3h if SEVIRI is assimilated at very high density (5 and 10 km)

RMS STD

Page 10: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

1. SEVIRI assimilation experiments Forecast Verification

F3h vs radiosondesFg-departures (8 days)

(from 17/10 – 0h to 24/10 – 21h 2011)

• Scores for the fit to IASI observations :NEGATIVE >> POSITIVE

• Bias reduction in FGd to radiosonde humidity• But, large degradations close to the surface.

Seem to confirm the wrong propagation of humidity increments toward the surface ?

Page 11: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

1. SEVIRI assimilation experiments Comments

• Increasing the observation density :

• produce sharper analysis increments structures • Main results over the first-guess :

• Large impacts over the humidity fields (radiosondes & IASI WV channels) Indication of a bias in the model ?• The First-guess fit to independent observations can be slightly improved

when SEVIRI WV observations are assimilated every 20 km.

Page 12: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Liu and Rabier (2002) and Desroziers (2011) :

Separation distance (km)

For observations with spatially uncorrelated error, increasing the observation density always significantly improve the analysis accuracy.

The analysis quality decreases, if the density of the observational data set is too large and error correlations are neglected.

Current approach : 1)data thinning Reduce the amount of used obs

1)inflated diagonal R matrixReduce the weigth of obs in the analysis (Dando et al., 2007; Collard and McNally, 2009)

UncorrelatedSub-optimaloptimal

2. Diagnosis of error correlations Motivations

Page 13: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

DATA :First-Guess or analysis departures from pair of SEVIRI WV6.2 observationsBinning interval =20 km Period : 30 October (8 cycles – 32000 radiances)

METHODS :A priori• Hollingsworth/Lönnberg (1986)• Background ensemble method (Bormann and Bauer, 2010)A posterioriDesroziers diagnostic (Desroziers et al., 2003)

2. Diagnosis of error correlations Data & Methods

Error sources : Measurement, Forward model, Representativeness, Quality control error

For each data type, observation error are determined from random Gaussian distribution that may be horizontally, vertically or channel-correlated or uncorrelated.

Page 14: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Hollingsworth/Lönnberg

Assumption : errors in the observations are spatially uncorrelated and the spatially correlated part of the background departures (FGd) is due to errors in FG.

Sigma O² = 0.13Sigma B² = 1.32

Cov(FGd) = HBHT+R

Separation distance (km)

2. Diagnosis of error correlations Estimate of observation errors

Page 15: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Desroziers diagnostic

Assumption: since DA follows linear estimation theory, the weigth given to the observations in the analysis is in agreement with true error covariances

Separation distance (km)

2. Diagnosis of error correlations Estimate of observation errors

Sigma O² = 0.10Sigma B² = 1.33

Page 16: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Hollingsworth/Lönnberg and Desroziers diagnostic

Obs. error estimates :

PROBLEM : Radiometric error estimate = 0.75K

1)H/L limitation : « The presence of any spatially correlated observation error will lead to an underestimation of the observation error, as such spatial correlation are neglected. » (Bormann and Bauer, 2010)

2) Desroziers limitation : « The method have the capability of retrieving error structures as long as the true background error and the true observation error have sufficiently different correlation structures » (Desroziers, personal communication)

2. Diagnosis of error correlations Estimate of observation errors

H/L Desroziers

Sigma O 0.36 K 0.31 K

Sigma B 1.14 K 1.15 K

Page 17: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Estimation from IASI observationsObservation error amplitude (sigma O)

55 T channels 96 + 20 Q channels

• Good agreement between the 2 methods for T channels but large differences for Q channels.• Estimated errors usually close to instrument noise (Desroziers Method)• Estimated errors lower than errors IASI spec system

DATA: IASI clear radiances15 days (01/09-15/09)Domain: AROME-WMED

Page 18: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Estimation from IASI observations

T channelsDesroziers Q channels

Inter-channel observation error correlationsDesroziers

• Several elements in the first off-diagonal are correlated due to opodisation effects

• Tropospheric sounding humidity channels exhibit blocs of strong inter-channel error correlations

Page 19: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Temperature

Surface

Humidity

L0.2 Humidity ~ 25 km

Estimation from SEVIRI observations

Horizontal observation error correlationsDATA: SEVIRI clear radiances (full resolution)15 days (01/09-15/09)Domain: AROME-WMED

Page 20: Stephanie Guedj Florence Rabier Vincent Guidard Benjamin Ménétrier Observation error estimation in a convective-scale NWP system

Conclusion & Future work•Taking observation error correlations into account in the assimilation system is an area of active research at Météo-France and at various NWP centres.

•SEVIRI WV6.2 observations were assimiled at several density Thinning distance from 70km to full resolution (5km)No significant impacts were shown on 3h-forecast skills (except humidity bias)

•Estimation of observation errors and their correlation for SEVIRI/IASI data (with 3 methods) :• Following Bormann and Bauer (2010), observation error and their correlations have been

estimated.

• Desroziers diagnostic demonstrated misleading results for these data (obs error lower than the instrumental noise, low horizontal correlation°

• Realistic observation error correlations were estimated using the background error method.

• No/small inter-channel error correlations for temperature sounding channels

• Strong inter-channel error correlations for tropospheric humidity sounding channels

• No horizontal error correlation are considered because they appear small and are otherwise difficult to tune in conjonction with the channel correlation.

•Focus on channel correlation (to be implemented in AROME)