1. Sea ice activities at the Climate Forecasting Unit
Virginie Guemas, Neven Fuckar, Francois Massonnet, Danila Volpi, Francisco Doblas-Reyes

2. Outline I - Introduction: Arctic sea ice predictability
II - Arctic sea ice variability modes
III - Sea ice initialization methods
IV - Added-value of sea ice initialization
V - Multi-model sea ice prediction

3. Outline I - Introduction: Arctic sea ice predictability
II - Arctic sea ice variability modes
III - Sea ice initialization methods
IV - Added-value of sea ice initialization
V - Multi-model sea ice prediction

4. I - Introduction: Arctic sea ice predictability
Persistence
Sea Ice Thickness
Sea Ice Thickness
Guemas et al (2014) QJRMS

5. I - Introduction: Arctic sea ice predictability
Ice thickness preconditioning
Thick ice = best predictor of September extent, thin ice good predictor for March extent
Chevallier and Salas y Mélia, JCLIM, 2012
Total volume
Total area
Area of ice h > 0.9m
Area of ice h > 1.5m
Total area
Area of ice h < 0.9m
Area of ice h < 1.5m
Total volume
Chevallier and Salas-Melia (2012) Journal of Climate

6. I - Introduction: Arctic sea ice predictability
Advection
Guemas et al (2014) QJRMS

7. I - Introduction: Arctic sea ice predictability
Ocean heat transport
Correlation detrended annual-mean ORAS4 60N heat transport and sea ice concentration
Atlantic OHT lag -3 years
Pacific OHT lag 0 year
0.8
0.6
0.4
0.2
-0.2
-0.4
-0.6
-0.8
Guemas et al (2014) QJRMS

8. I - Introduction: Arctic sea ice predictability
Reemergence
SST persistence
Local sea ice area anomalies associated with SST anomalies
Local SST anomalies impact sea ice area anomalies
Melting season
Growing season
Local sea ice thickness anomalies impact sea ice area anomalies
Local sea ice area anomalies with sea ice thickness anomalies
SIT persistence
Blanchard-Wrigglesworth al (2011) Journal of Climate

9. I - Introduction: Arctic sea ice predictability
Long-term trend
Observed (NSIDC) September sea ice extent
7.04 million km2
3.61 million km2
England + Scotland + Netherlands + Belgium + France + Germany + Poland + Switzerland + Austria + Italy + Spain + Portugal + Czech Republic

10. Outline I - Introduction: Arctic sea ice predictability
II - Arctic sea ice variability modes
III - Sea ice initialization methods
IV - Added-value of sea ice initialization
V - Multi-model sea ice prediction

11. II – Arctic sea ice variability modes
k-means cluster analysis of sea ice thickness (SIT)
Clustering methods more robust than EOF analysis + account for nonlinearities.
Tools available in s2dverification R package
Fučkar et al (in preparation)

12. II – Arctic sea ice variability modes
k-means cluster analysis of sea ice thickness (SIT)
● quadratic detrending
● Clusters (k=3) for all seasons/months
Fučkar et al (in preparation)

13. Outline I - Introduction: Arctic sea ice predictability
II - Arctic sea ice variability modes
III - Sea ice initialization methods
IV - Added-value of sea ice initialization
V - Multi-model sea ice prediction

14. III – Sea Ice Initialization: in-house reconstructions
NEMO3.2 ocean model + LIM2 sea ice model
Forcings : DFS4.3 or ERA-interim
Nudging : T and S toward ORAS4, timescales = 360 days below 800m, and 10 days above except in the mixed layer, except at the equator (1°S-1°N), SST & SSS restoring (-40W/m2, -150 mm/day/psu)
Wind perturbations + 5-member ORAS > 5 members for sea ice reconstruction
5 member sea ice reconstruction for 1958-present consistent with ocean and atmosphere states used for initialization
Guemas et al (2014) Climate Dynamics

15. III – Sea Ice Initialization: in-house reconstructions
October-November Arctic sea ice thickness
Reconstruction
IceSat
Too much ice in central Arctic, too few in the Chukchi and East Siberian Seas
Guemas et al (2014) Climate Dynamics

16. III – Sea Ice Initialization: in-house reconstructions
March and September Arctic sea ice
Sea ice area
Sea ice volume
HadISST
NSIDC
ERAint + Nudging
UCL
DFS4.3 + Nudging
DFS4.3 Free
UCL
Thousands km3
Millions km2
ERAint + Nudging
DFS4.3 Free
DFS4.3 + Nudging
PIOMAS
Bias but reasonable agreement in terms of interannual variability
Guemas et al (2014) Climate Dynamics

17. III – Sea Ice Initialization: sea ice data assimilation
The ensemble Kalman filter: a multivariate data assimilation method for smoother initialization
1. Model forecasts
2. Analysis
There may be several drawbacks of initializing separately the ice and the ocean, namely shocks between the two components at initialization time. One way to overcome this is to use so-called multivariate data assimilation methods, and the ensemble Kalman filter is one of them. The method works in two steps: first, an ensemble of simulations is forwarded in time so that a structure emerges in the ensemble. For instance, members that created more ice than others will naturally have higher salinities, and conversely. At this stage, an observation is injected in the system. This observation may be limited in space and restricted to one variable, for example sea ice concentration. The second and final step is the « analysis », where not only ice concentration but also salinity is updated in the model, based on the relationship between these two variables. This cycle is reproduced as many times as needed.
Observations (e.g., ice concentration only)
Massonnet et al (2014) Ocean Modelling

18. III – Sea Ice Initialization: sea ice data assimilation
Importance of multivariate initialization for seasonal sea ice prediction
OBS
The ensemble Kalman filter has been used extensively with the NEMO and LIM models, forced by atmospheric reanalysis, by François Massonnet during this thesis. The figure shows the initial conditions of sea ice thickness (left), and surface temperature (middle) in March 2007 for a run without data assimilation of sea ice concentration (top), and one initialized with the EnKF. It is important to remind here that only observations of ice concentration were fed in the model, but that because of its multivariate structure, the Ensemble Kalman filter automatically updates all other variables (here thickness and temperature). Thanks to the update of these non-observed variables, it is still possible to improve the spatial distribution of ice concentration in the next summer, because these variables hold a memory that can reemerge six months later.
OBS
Massonnet et al (2014) Ocean Modelling

19. III – Sea Ice Initialization: sea ice data assimilation
Fully-coupled sea ice data assimilation in EC-Earth: the next challenge
What are the perturbations required to generate adequate spread in EC-Earth during the forecast steps of the assimilation run ?
Should the atmosphere be updated when sea ice observations are assimilated?
Can we afford to run the EnKF with less members (CPU time is limited) ?
The next step is now to port the EnKF to a fully-coupled system (EC-Earth). But when the model changes, new questions emerge. It is uncertain at this stage how to perturb the model as to generate the appropriate error structure in the sea ice model. In fact, perturbing coupled systems is far from clear since the time scales of the ocean and the atmosphere are very different. Another question pertains to the update phase: when ice concentration is updated, should near-surface air temperatures be updated as well? What about the whole vertical column? Finally, it is clear that these simulations will be highly demanding. So far the filter has been run with 25 members, which seems a lot for climate modellers but ridiculous for data assimilation people. How to optimize the use of resources without sacrificing the statistics of our ensemble will be a very important question to address.
Massonnet et al (2014) Ocean Modelling

20. III – Sea Ice Initialization: different techniques
Melanie
IC3 CFU
S2d probabilistic forecasting using GCM ensemble approach
FP7 projects – Environment, call for European s2d forecasts towards climate services.
Climate services term from WMO GFCS.
My role
My background

21. III – Sea Ice Initialization: different techniques
Issue: Distinction between climate drift and climate signal
Hypothesis: If the model climate is stable (no drift), the simulated variability is independent of the model mean state within the range of current model biases and closer to the observed variability than when mixed with the drift
Testing the hypothesis: Allowing the climate model biases but constraining the phase of the simulated variability toward the contemporaneous observed one at the initialization time: Anomaly Initialization (AI)
Melanie
IC3 CFU
S2d probabilistic forecasting using GCM ensemble approach
FP7 projects – Environment, call for European s2d forecasts towards climate services.
Climate services term from WMO GFCS.
My role
My background

22. III – Sea Ice Initialization: different techniques
Melanie
IC3 CFU
S2d probabilistic forecasting using GCM ensemble approach
FP7 projects – Environment, call for European s2d forecasts towards climate services.
Climate services term from WMO GFCS.
My role
My background

23. III – Sea Ice Initialization: different techniques
Anomaly versus Full Field Initialization
EC-Earth2.3, 5 members, start dates every 2 years from 1960 to 2004
FFI : Full-field initialization from ORAS4 + ERA
OSI-AI : Ocean and sea ice anomaly initialization with corrections to ensure consistency
rho-OSI-wAI : Ocean and sea ice weighted anomaly initialization to account for the different model and observed amplitudes of variability + (density, temperature) Instead of (temperature, salinity) anomaly initialisation
Melanie
IC3 CFU
S2d probabilistic forecasting using GCM ensemble approach
FP7 projects – Environment, call for European s2d forecasts towards climate services.
Climate services term from WMO GFCS.
My role
My background
NOINI : historical simulation

24. III – Sea Ice Initialization: different techniques
Anomaly versus Full Field Initialization
RMSE sea ice area.
Ref: Guemas et al (2014)
RMSE sea ice volume.
Ref: Guemas et al (2014)
FFI
OSI-AI
NOINI
NOINI
FFI
OSI-AI
rho-OSI-wAI
rho-OSI-wAI
Lower forecast error when using ocean and sea ice anomaly initialization, even lower with weighted anomalies

25. Outline I - Introduction: Arctic sea ice predictability
II - Arctic sea ice variability modes
III - Sea ice initialization methods
IV - Added-value of sea ice initialization
V - Multi-model sea ice prediction

26. IV – Added-value of sea ice initialization
Clim
CNRM-CM and EC-Earth2.3
Seasonal forecasts initialized every 1st November and every 1st May from 1979 to 2012
7 month long
Initialized from ERA-interim for the atmosphere + in-home sea ice reconstructions + ORAS4 for the ocean
5/12 members using initial perturbations applied to the atmosphere, ocean and sea ice components for EC-Earth2.3, only the atmosphere and ocean for CNRM-CM
Sensitivity experiment : climatological sea ice initialization
Guemas et al (2015) in preparation

27. IV – Added-value of sea ice initialization Arctic predictions from May
Clim
95% confidence interval
Skill increase in the Arctic with sea ice initialisation, which is not only due to the long-term trend
Guemas et al (2015) in preparation

28. IV – Added-value of sea ice initialization
Antarctic predictions from May
Init
Clim
Negligible impact of sea ice initialization in May in the Antarctic
Guemas et al (2015) in preparation

29. Outline I - Introduction: Arctic sea ice predictability
II - Arctic sea ice variability modes
III - Sea ice initialization methods
IV - Added-value of sea ice initialization
V - Multi-model sea ice prediction

30. V – Multi-model regional sea ice prediction
CNRM-CM5.1 and EC-Earth2.3
Seasonal forecasts initialized every 1st November and every 1st May from 1990 to 2008
5 month long
Initialized from ERA-interim for the atmosphere + in-home sea ice reconstructions + ORAS4 (Ec-Earth2.3) or ocean reconstruction (CNRM-CM5.1) for the ocean
5 members using initial perturbations applied to all components for EC-Earth2.3, only the atmosphere for CNRM-CM5
Chevallier and Guemas (2015) in preparation

31. V – Multi-model regional sea ice prediction
Okhotsk
Bering
Summer prediction
Hudson
Kara
Baffin
Winter prediction
Barents
Greenland
Chevallier and Guemas (2015) in preparation

32. V – Multi-model regional sea ice prediction
Linearly Detrended Anomaly Correlation for Sea Ice Extent (SIE)
Total Arctic
Baffin Bay
95% confidence interval
EC-Earth 2.3
CNRM-CM5.1
correlation
EC-Earth CNRM-CM 5.1
Predictability until March Similar performance between CNRM-CM5.1 and EC-Earth 2.3
Chevallier and Guemas (2015) in preparation

33. V – Multi-model regional sea ice prediction
Linearly Detrended Anomaly Correlation for Sea Ice Extent (SIE)
Greenland Sea
Barents and Kara Seas
95% confidence interval
EC-Earth 2.3
CNRM-CM5.1
correlation
EC-Earth CNRM-CM 5.1
Predictability for the first month only in Greenland, until March in Barents and Kara Better performance from CNRM-CM5.1
Chevallier and Guemas (2015) in preparation

34. V – Multi-model regional sea ice prediction
Linearly Detrended Anomaly Correlation for Sea Ice Extent (SIE)
Okhotsk Sea
Bering Sea
95% confidence interval
EC-Earth 2.3
CNRM-CM5.1
correlation
EC-Earth CNRM-CM 5.1
Better initialisation in Okhotsk Sea with CNRM-CM5: ocean ? Multi-model allows to make the most of individual performance
Chevallier and Guemas (2015) in preparation

35. V – Multi-model extreme extent prediction
RMSE
RMSE
Prediction error overall depends on used bias correction method
(for SIE in these forecasting systems the best is trend b.c.m.)

36. V – Multi-model extreme extent prediction
⬆ deviation of Arctic sea ice from obs trend can ⬆ model errors in summer

37. Conclusions
Sources of sea ice predictability: persistence, ice thickness preconditioning, advection, ocean heat transport, reemergence, long-term trend
Initialization: need for ensemble reconstruction/reanalyses consistent with the atmospheric and oceanic state, making the best of available sea ice observations, anomaly initialisation seems promising for the sea ice component, clear impact in the Arctic, not in the Antarctic
Prediction: winter predictability until March in the Baffin Bay, Kara, Barents and Okhotsk seas