1. GENOA: Generic Ensemble generation for Ocean Analysis
Eric de Boisséson, Hao Zuo, Magdalena Balmaseda and Patricia de Rosnay
CMEMS Service Evolution 2 mid-term review, 19 March 2019

2. Context CMEMS Service Evolution 2 Lot 4
Section 4.7 of the CMEMS service evolution strategy: coupled ocean-atmosphere models with assimilative capability
Advanced assimilation methods targeted to provide improved estimations of upper ocean properties consistent with sea-surface observations and air-sea fluxes

3. Context What we propose:
A way to improve estimates of upper ocean properties is to improve the uncertainty estimate of the ocean state
Develop a generic ensemble generation scheme for ocean analysis applicable in both ocean-only and coupled models.
Optimize the use of the information from the ocean observing system
Account for uncertainties of the surface variables and air-sea fluxes

4. Context
Optimize the use of the information from the ocean observing system
Major source of uncertainty: observation representativeness errors
Mismatch between obs. and model resolutions
Standard method: super-observation and thinning techniques
But loss of information from the observing system
Alternative: random perturbation + ensemble approach
Better exploit the information from the obs.
Uncertainty estimate

5. Context
Account for uncertainties of the surface variables and air-sea fluxes
Partially taken into account in current ocean analysis systems
Example: current ECMWF ocean analysis uses monthly random perturbations on wind stress and SST based on differences between analyses
How can we extend it?
Other sources of uncertainty: bulk formulation parametrisations, sea-ice constraints and heat and freshwater fluxes
A wider range of temporal scales (submonthly to monthly)
Multivariate relationships
In practice
Use a wider variety of analysis datasets to provide structural and analysis uncertainties

6. Working plan WP1: development of the new perturbation scheme
Task 1: Perturbations of ocean observations
Task 2: Evaluation of the perturbation scheme for assimilated observations
Task 3: Building a perturbation repository for surface forcing and surface variables
Task 4: Evaluation of the perturbation scheme for surface forcing and surface variables
WP 2: Application and evaluation in both ocean-only and coupled analysis systems
Task 5: Application in both ocean-only and coupled assimilation systems
Task 6: Evaluation in both ocean-only and coupled assimilation systems

7. Working plan System and input data: ORAS5 (CMEMS MFC GLO-RAN)
Ocean only forced by atmospheric reanalysis
NEMO at 1/4˚ degree and 75 vertical levels
LIM2 sea-ice model
NEMOVAR DA
CERA-SAT
Coupled reanalysis system
Ocean component similar to ORAS5
Atmosphere ECMWF IFS 60km resolution and 137 levels
4D-Var DA and 10-member EDA system
Input datasets
In-situ T/S profiles
SLA from CMEMS SL TAC
SIC from CMEMS OSTIA
SST (HadISST2, OSTIA)

8. Working plan ORAS5 CERA-SAT NEMO Coupled NEMO-IFS NEMO
Observations
Surface forcing
Observations
Obs operator
Bulk formula
Obs operator
NEMO
Coupled NEMO-IFS
Model departures
Model departures
NEMOVAR
3DVAR FGAT
IFS 4DVAR
NEMOVAR
3DVAR FGAT
Increment
Increment
NEMO
Coupled NEMO-IFS
Next cycle
Next cycle
ORAS5 and CERA-SAT: perturbations + ensemble approach = uncertainty estimate
CERA-SAT: Uncertainty on surface forcing coming from atmos. EDA, flow dependent

9. The team
Eric de Boisséson (PI), ECMWF, coupled data assimilation (CDA) team
Hao Zuo, ECMWF, CDA team
Magdalena Balmaseda, ECMWF, Head of Predictability section
Patricia de Rosnay, ECMWF, CDA team leader

10. Summary of achievements
Specific Objectives
Status
Scientific results
Impact on CMEMS
WP1 task 1: Implementation of a new perturbation scheme for ocean observations
Complete
The perturbation scheme is working as intended exploiting better the variety of the observing system and accounting for uncertainties
None at the moment but the perturbation code could be provided to MFCs for implementation in other systems
WP1 task 2: Evaluation of the perturbation scheme for assimilated observations
On-going
The new perturbation scheme produces an ensemble of ocean analyses that better fits observations but remains underdispersive. Results suggests that the current specification for background errors is not always ideal
Same as above
WP1 task 3: Building a perturbation repository for surface forcing and surface variables
The building of the repository is on-going
WP1 task 4: Evaluation of the perturbation scheme for surface forcing and surface variables
To be done in the second half of GENOA
WP2 task 1: Application in both ocean-only and coupled assimilation systems
WP2 task 2: Evaluation in both ocean-only and coupled assimilation systems

11. WP1 task 1: implementation of a new perturbation scheme for ocean observations
System and input data:
ORAS5 (CMEMS MFC GLO-RAN)
Ocean only forced by atmospheric reanalysis
NEMO at 1/4˚ degree and 75 vertical levels
LIM2 sea-ice model
NEMOVAR DA
CERA-SAT
Coupled reanalysis system
Ocean component similar to ORA-S5
Atmosphere ECMWF IFS 60km resolution and 137 levels
4D-Var DA and 10-member EDA system
Input datasets
In-situ T/S profiles
SLA from CMEMS SL TAC
SIC from CMEMS OSTIA
SST (HadISST2, OSTIA)

12. WP1 task 1: implementation of a new perturbation scheme for ocean observations
Perturbation of in situ T/S profiles
ORAS5: OE stdev for T/S specified by empirical analytical function that provides approximate fit to the vertical profiles of T/S stdev estimated from observations by Ingleby and Huddleston (2007)
Uneven sampling on horizontal/vertical -> RE not fully taken into account
To account for the RE of profiles:
Perturbations of the horizontal locations of the in-situ observations
Stratified random thinning on the vertical.

13. WP1.1: Perturbation of horizontal location of T/S profiles
5 possible strategies are implemented
P1: perturbs a profile with random distance (0 to PD) and random angle (0-360 degree)
P2: perturbs a profile with constant distance (as PD) and random angle (0-360 degree)
P3: perturbs latitude of a profile with random distance (0 to PD)
P4: perturbs longitude of a profile with random distance (0 to PD)
P5: perturbs latitude and longitude of a profile with the same random distance (0 to PD)
Example: P2 for XBTs and P5 for moorings
The perturbations + ensemble approach provide estimate of uncertainties accounting for RE

14. WP1.1: Perturbation of vertical location of T/S profiles
Vertical resolution of T/S profiles much higher than model grid
Standard method: thinning
Alternative: random selection within the thinning grid
Ensemble approach
The random selection + ensemble approach provide estimate of the uncertainty and allow to further exploit the information on the vertical axis

15. WP1.1: Perturbation of sea-surface observations
1. Perturbation of sea-ice observations:
ORAS5 assimilates SIC from CMEMS OSTIA (1/4 degree)
Too many observations -> regular interval thinning (1/2 degree)
Information lost -> new method with stratified random thinning: random selection within thinning grid
The random selection + ensemble approach to better exploit the info from SIC data

16. WP1.1: Perturbation of sea-surface observations
1. Perturbation of sea-ice observations:
Standard: 0.5x0.5 thinning
ORAS5 assimilates SIC from CMEMS OSTIA (1/4 degree)
Too many observations -> regular interval thinning (1/2 degree)
Information lost -> new method with stratified random thinning: random selection within thinning grid
Alternative: random selection within the thinning grid
The random selection + ensemble approach to better exploit the info from SIC data

17. WP1.1: Perturbation of sea-surface observations
Raw data
2. Perturbation of sea-level observations:
ORAS5 assimilates along-track SLA from CMEMS SL TAC
Too many observations ->super-obbing (average) according to mission, area and time. Take into account part of RE
Information lost -> new method using random selection within the superob grid
Superobbing
Random thinning
The random selection + ensemble approach to better exploit the info from SIC data

18. WP1 task 2: evaluation of a new perturbation scheme for ocean observations
Evaluation of perturbation of in situ T/S profiles
ORAS5 low-res (ORCA1_Z42) with 5 ensemble members over
DA: T/S profiles from EN4 and SIC from OSTIA. SST nudging and SLA assimilation are deactivated
Experiment
Horizontal Perturbation (P1)
Vertical Perturbation
Hpert50
PD=50km -
Hpert100
PD=100km
Hpert200
PD=200km
HpertS15
PD=200km; 15 members
Vpert
PD=0km
N=3 observations per model level
HVpert100
Same as above
HVpert200

19. WP1.2: ensemble diagnostics
Ensemble spread diagnostics: method
Ensemble spread produced from the background fields of the experiments.
Compare ensemble spread to the specified BGE stdev in observation space
Specified BGE cov: analytical formulation (Mogensen et al, 2012)
T: based on dT/dz, max in the thermocline
S: based on dS/dT, max in the mixed layer
Desroziers’ diagnostic of a-posteriori BGE covariances based on innovations and analysis increments (Weaver et al, 2013) -> reference for the statistical consistency of specified BGE covariance.

20. WP1.2: ensemble diagnostics
Horizontal distribution 100m depth in Hpert 200
Horizontal distribution of the spread
Temperature spread shows right features but underdispersive, esp. in the Tropics
Salinity spread very different from specified BGE stdev. More similar to Desroziers diagnostic.
Specified BGE stdev for S overestimated? Need to be revised? Ensemble approach for BGE specification?

21. WP1.2: ensemble diagnostics
Vertical profiles - area average
Vertical distribution of the spread
Spread comes mainly from horizontal perturbations. Vertical perturb have some impact in poorly-sampled deep ocean
Spread lower than specified BGE stdev in the thermocline for T and mixed layer for S
Underdispersive in Tropical thermocline -> increase ensemble, inflate in thermocline
Spread closer to diagnosed BGE stdev in extratropics: energetic areas (sharp fronts, eddies …), perturbations more impactful
Temperature
Salinity
N ExtraTropics
S ExtraTropics
Tropics

22. WP1.2: ensemble diagnostics
Temporal evolution of the spread
OHC in model space: ensemble mean and spread
Global OHC – Ens mean
Global OHC detrended
Global OHC – Ens spread
Ens mean: similar values and trend in pertub experiments.
Larger PD = larger spread and interannual variability
Most spread comes from horizontal perturb

23. WP1.2: ensemble diagnostics
Comparisons to independent observations – WOA13
Ensemble mean temperature at 800m
Largest biases in WBC
Increasing the PD reduces bias in Gulf Stream
Ens mean from Hpert exp shows lower bias wrt to WO13 compared to the control member (not shown)

24. WP1.2: ensemble diagnostics
Comparisons to independent observations – OSTIA SST
Verification against OSTIA SST
Ensemble variance
(Model departure)^2 – (Obs error)^2
Hpert50
Largest departures in WBC and ACC
Larger PD, larger ensemble variance
Ideally ens variance comparable to departures
Under dispersive ensemble
Hpert100
Hpert200

25. WP1 task 2: evaluation of a new perturbation scheme for ocean observations
Evaluation of perturbation of sea-ice concentration
Experiment: ORCA1_Z42 with 5 ensemble members over with 100km thinning scale
Large spread close at ice edge and next to coast (Canadian archipelago, Greenland …)
Large seasonal variability in spread (forming/retreating ice)
Impact on other ocean processes such as transport across Arctic straits
RMSE wrt OSTIA SIC
Ens spread

26. WP1 task 2: evaluation of a new perturbation scheme for ocean observations
Evaluation of perturbation of sea-ice concentration
Experiment: ORCA1_Z42 with 5 ensemble members over with 100km thinning scale
Large spread close at ice edge and next to coast (Canadian archipelago, Greenland …)
Large seasonal variability in spread (forming/retreating ice)
Impact on other ocean processes such as transport across Arctic straits
RMSE wrt OSTIA SIC
Ens spread July

27. WP1 task 2: evaluation of a new perturbation scheme for ocean observations
Evaluation of perturbation of sea-level anomalies
Ensemble variance
Experiment: ORCA025_Z75 with 5 ensemble members over with 1-degree thinning scale
Large spread/departures in energetic areas such as WBC and ACC -> not fully resolved -> large REs
Ensemble slightly underdispersive
(Model departure)^2 – (Obs error)^2

28. Conclusions Next tasks and WP:
WP1 tasks 3 and 4 on perturbations of sea surface constraints
Build the perturbation repository
Variables selected: wind stress, SST, SIC, P-E and solar radiation
Structural errors: wind stress from different sources (ERA-Int, NWP, NCEP, ERA40) + different settings (wave effects, bulk) + SST and SIC from ESA-CCI and HadISST (v2.1 and v2.0)
Analysis errors: wind stress, SST, P-E and solar radiation from ERA-20C (monthly) + SST/SIC from HadISSTv2 (pentad) -> multiple timescales
Test and evaluate the new perturbation repository with the ORAS5 system

29. Conclusions Next tasks and WP:
WP2 on application and evaluation of the perturbation scheme in ocean-only and coupled analysis systems
run the full ORAS5 and CERA-SAT systems using the scheme from WP1
At the air-sea interface, ORAS5 will use uncertainty information from the perturbation repository while CERA-SAT will use the atmospheric EDA -> comparison EDA – perturbation rep

30. Conclusions
Issues and risks: so far everything is going according to plan
Deliverables:
Quarterly reports: up to date
Mid-term report: delivered
Final report: to be delivered KO+23
Outcomes: status
Generic ensemble generation scheme for ocean analysis applicable in both ocean- only and coupled models
Optimize the use of the ocean observing system (WP1 tasks 1-2)
Account for the uncertainties from air-sea fluxes (WP1 tasks 3-4, to do)
Potential impact of coupling on the upper ocean properties and their uncertainty (WP2, to do)

31. Thank you for your attention
Any question?