Seasonal forecast skill of Arctic sea ice area

  • Slides: 27
Download presentation
Seasonal forecast skill of Arctic sea ice area Michael Sigmond (CCCma) Sigmond, M. ,

Seasonal forecast skill of Arctic sea ice area Michael Sigmond (CCCma) Sigmond, M. , J. Fyfe, G. Flato, V. Kharin, W. Merryfield, GRL, 2013 (Can. SIPS) Merryfield, W. Lee, W. Wang, M. Chen and A. Kumar, GRL, 2013 (Can. SIPS+CFSv 2) Can. SISE East meeting, CIS, 10 February 2014

Increased interest in seasonal predictions Sept. 1980 Sept. 2012 • Number of commercial vessels

Increased interest in seasonal predictions Sept. 1980 Sept. 2012 • Number of commercial vessels through NE passage: 2009: 2 2012: 46 • Sept. 2013: first commercial vessel through NW passage

Statistical models: • Until recently, forecasts were made exclusively with statistica models (MLR, etc)

Statistical models: • Until recently, forecasts were made exclusively with statistica models (MLR, etc) • Based on observed statistical relationships between: - T, circulation, SST, sea ice etc. in month X (predictor) - sea ice cover in month X+1, 2, 3, …. (predictand) • But: Relationships depend on the mean state of the climate Correlation AO winter and SIE in September Begin year: 1979 Holland Stroeve (2011)

Statistical models: • Until recently, forecasts were made exclusively with statistica models • Based

Statistical models: • Until recently, forecasts were made exclusively with statistica models • Based on observed statistical relationships between: - T, circulation, SST, sea ice etc. in month X (predictor) - sea ice cover in month X+1, 2, 3, …. (predictand) • But: Relationships depend on the mean state of the climate → statistical models may have large errors → Need to develop new tools

Dynamical models: • Models based on laws of physics (like climate models) • Require

Dynamical models: • Models based on laws of physics (like climate models) • Require substantially more computational power than statistical models • Have been used operationally to produce seasonal forecast of temperature, precipitation • But only a few operational seasonal forecast systems include an interactive sea ice component • Not yet clear how skillful forecasts of sea ice are Geophys. Res. Lett. , 2013

Canadian Seasonal to Inter-annual Prediction System (Can. SIPS) • Environment Canada’s seasonal forecasting system

Canadian Seasonal to Inter-annual Prediction System (Can. SIPS) • Environment Canada’s seasonal forecasting system • Based on two coupled climate models (Can. CM 3/Can. CM 4) • Initial conditions (including sea ice area) constrained to be close to observations (20 ensemble members) • But: Sea ice thickness not initialized (instead: climatology of previous model version) • Re-forecasts initialized in each month between January 1979 and December 2009 (12 month duration)

September forecasts:

September forecasts:

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1 p Se 1 g Au l 1 Ju 1 n Ju 1 ay M ov N 1 time

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1 p Se 1 g Au l 1 Ju 1 n Ju 1 ay M ov N 1 time

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1 p Se 1 g Au l 1 Ju 1 n Ju 1 ay M ov N 1 time

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1 p Se 1 g Au l 1 Ju 1 n Ju 1 ay M ov N 1 time ?

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1

September forecasts: Lead 11 (months) 4 3 2 1 0 1 ct O 1 p Se 1 g Au l 1 Ju 1 n Ju 1 ay M ov N 1 time ?

September forecasts (detrended): Lead 11 (months) 4 3 2 1 0 1 ct O

September forecasts (detrended): Lead 11 (months) 4 3 2 1 0 1 ct O 1 p Se 1 g Au l 1 Ju 1 n Ju 1 ay M ov N 1 time

September forecasts (detrended): • No skill for predicting deviations from trend when initialized prior

September forecasts (detrended): • No skill for predicting deviations from trend when initialized prior to June 1 • Several studies have shown that winter/spring sea ice thickness good predictor for September sea ice (‘preconditioning’) Skill may be enhanced by initializing sea ice thickness

Forecasts for other months:

Forecasts for other months:

Correlation Skill Score (TOTAL, not detrended): Sigmond et al. (2013)

Correlation Skill Score (TOTAL, not detrended): Sigmond et al. (2013)

Decomposition Correlation Skill Score TOTAL TREND DE-TRENDED + Sigmond et al. (2013)

Decomposition Correlation Skill Score TOTAL TREND DE-TRENDED + Sigmond et al. (2013)

Decomposition Correlation Skill Score TOTAL TREND DE-TRENDED + ? ? Sigmond et al. (2013)

Decomposition Correlation Skill Score TOTAL TREND DE-TRENDED + ? ? Sigmond et al. (2013)

Trend-independent skill: DE-TRENDED ● ● OBSERVED LAG COR Consistent with potential predictability studies (Holland

Trend-independent skill: DE-TRENDED ● ● OBSERVED LAG COR Consistent with potential predictability studies (Holland et al, 2010) Explanation: winter sea ice edge closely related convergence of ocean heat fluxes (predictable on longer timescales)

Trend-independent skill: DE-TRENDED OBSERVED LAG COR Sigmond et al. (2013) ● ● Good news:

Trend-independent skill: DE-TRENDED OBSERVED LAG COR Sigmond et al. (2013) ● ● Good news: we understand seasonal dependency of skill Potentially bad news: Similarity suggest that all skill is due to persistence Does our model outperform a persistence forecast?

Skill relative to persistence (detrended) ● ● Merryfield et al. (2013) Model outperforms persistence

Skill relative to persistence (detrended) ● ● Merryfield et al. (2013) Model outperforms persistence forecasts initialized in January and June Averaged over all months and lead times, enhancement is statistically significant (p<0. 01)

Skill relative to persistence (detrended) Can. SIPS performs slightly better than CFSv 2 for

Skill relative to persistence (detrended) Can. SIPS performs slightly better than CFSv 2 for detrended anomalies

Skill relative to persistence (Total anomalies) Can. SIPS performs substantially worse than CFSv 2

Skill relative to persistence (Total anomalies) Can. SIPS performs substantially worse than CFSv 2 because: ● Underestimation of trend: 1) SIC initialization: dataset used (Had. ISST) underestimates trend → Large skill increase expected just by changing initialization dataset 2) SIT not initialized: (does not decrease with time as in observations) → Further skill increase expected by initializing sea ice thickness

Does a multi-system ensemble outperform single systems?

Does a multi-system ensemble outperform single systems?

Skill relative to persistence Detrended: Total anomalies: Merryfield et al. (2013)

Skill relative to persistence Detrended: Total anomalies: Merryfield et al. (2013)

Conclusions: • Initial examination of forecast skill of sea ice area in Can. SIPS,

Conclusions: • Initial examination of forecast skill of sea ice area in Can. SIPS, which forms a baseline for improvements to be achieved by Can. SISE • Substantial skill, but most of the skill is due to strong downward trend in observations • Forecast skill of detrended anomalies for longer lead times is generally small except for January/February • Trend-independent forecast skill exceeds that of an anomaly persistence forecast • Forecast skill for sea ice can be increased by combining multiple forecasting systems

Future research: • Do we get skill on regional and local scales? • Will

Future research: • Do we get skill on regional and local scales? • Will model and initialization improvements lead to enhanced skill? • Multi-model study on impact of sea ice initialization on prediction of 2007, 2008, 2011 and 2012 September minima (SPECS)