1. the Most Predictable Patterns in CFS over the Tropical Atlantic Ocean
Mean, Variability, and
the Most Predictable Patterns
in CFS over the Tropical Atlantic Ocean
Zeng-Zhen Hu1
Bohua Huang1,2
(Contributors: K. Pegion and B. Jha)
(Acknowledgments: S. Saha, S. Yang, K. Campana)
1Center for Ocean-Land-Atmosphere Studies
Calverton, Maryland, USA
2Department of Atmospheric, Oceanic, and Earth Sciences, College of Science
George Mason University, Fairfax, Virginia, USA
COLA
COLA & NCEP CTB Joint Seminar: NCEP, Dec. 10, 2007

2. Why study Atlantic climate of CFS?
Atlantic Ocean is very important to US and European (even global) climate. Hurricanes, Heat Waves, Meridional Overturning Circulation (MOC), North Atlantic Oscillation.
CFS is the operational model in US to do the seasonal climate prediction
We are the Atlantic research group at COLA and supported by NOAA Atlantic project.
There are less published works of NCEP CFS about Atlantic than about Pacific.

3. Main topics: Part 1: What is the ability of CFS in simulating the mean climate in the tropical Atlantic? What is the possible role of the cloud-radiation-SST feedback in the biases? Part 2: What is the ability of CFS in simulating the leading variability modes and the associated physical processes in the tropical Atlantic? Part 3: What is the predictive skill and most predictable patterns of CFS in the tropical Atlantic Ocean? What is their connection with ENSO ?

4. Data & CFS Observations CFS Hindcasts: CFS Free Run
NCEP/NCAR reanalysis I (Kalnay et al. 1996); ERA40 (Uppala et al. 2005)
Cloud coverage data by ISCCP on 2.5ox2.5o, Jul1983-Dec2004 (Rossow and Dueñas 2004); Corresponding radiation calculated by Zhang et al. (2004)
The SST dataset is the extended reconstruction (ER-v2) on 2ox2o (Reynolds et al. 2002)
NCEP Global Ocean Data Assimilation (GODAS) (Behringer et al. 1998; Behringer and Xue 2004)
CFS Free Run
AGCM: NCEP T62&L64
OGCM: MOM3, 1/3o(10oS-10oN), 1o(higher latitudes); 74oS-64oN, L40
Free run: Jan 1985-Dec 2036, conducted by Kathy Pegion at COLA
CFS Hindcasts:
All calender months: Jan Dec. 2003
15 predictions of 9 month length from lead month 0 to 8
Oceanic ICs: 11th, 21st of month 0, 1st of month 1 (Saha et al. 2006; Huang and Hu 2006).
The atmospheric IC: the NCEP reanalysis II (Kanamitsu et al., 2002). Same day & those within two days before and after
Data available from

5. Part 1: Mean Biases & the Role of Cloud-Radiation-SST Feedback
Hu, Z.-Z., B. Huang, and K. Pegion, 2008: Leading patterns of tropical Atlantic variability in a coupled GCM. Climate Dynamics, 30 (7-8), , DOI /s x. Hu, Z.-Z., B. Huang, and K. Pegion, 2007: Low cloud errors over the southeastern tropical Atlantic in the NCEP CFS and their association with lower-tropospheric stability and air-sea interaction. JGR-Atmo. 113, D12114, doi: /2007JD

6. Seasonal Mean and Variance of SST: (1) General pattern and zonal gradient well simulated (2) Warm bias (3) Too small variance in southeastern Ocean

7. CFS Simulation of Tropical Atlantic SST Climatology
Mean Field
Equatorial SST (Gradient)
CFS
CFS
OBS
OBS
OBS
DJF
MAM
CFS
OBS
Annual
CFS
CFS
CFS
OBS
JJA
SON
OBS

8. For majority of current CGCMs, it is still a challenge in simulating the zonal SST gradient along equatorial Atlantic (Davey et al. 2002)
No Flux Correction
With Flux Correction

9. Observed and CCSM Simulated JJA SST (Deser et al. 2006: J
Observed and CCSM Simulated JJA SST (Deser et al. 2006: J. Climate, 19, )

10. Observed and CCSM3 Simulated MAM SST (Deser et al. 2006: J
Observed and CCSM3 Simulated MAM SST (Deser et al. 2006: J. Climate, 19, )

11. Compared with other coupled models (CCSM, MPI, ECMWF, MO, and others), CFS did very good job in the tropical Atlantic, particularly in simulating the equatorial SST gradient. However, most of these CGCMs show a similar error pattern in the tropical Atlantic: warm bias in the southeastern Ocean.

12. CFS Systematic Biases MAM JJA SON DJF
(CFS-ISCCP//ERv2)
MAM
There are coherent biases in the southeastern Atlantic Ocean:
(1) Too few low clouds (2) Too much short-wave radiation reaching the sea surface
(3) Warm biases
JJA
SON
DJF
Low Clouds
Short-Wave
SST

13. Do the differences of cloud definition between the CFS and ISCCP affect the results of the comparison? The answer is “YES or NO”.

14. Difference of the Cloud Definition
ISCCP
(Wear, 1999; Rossow& Dueñas 2004)
CFS
(Xu& Randall 1996)
High clouds
< 440 hPa
< 350 hPa
Middle clouds
hPa
hPa
Low clouds
Surface-680 hPa (Shallower)
~ hPa (Deeper)
Boundary layer clouds no
Surface-900 hPa
(about 10% of atmosphere mass)
ISCCP: Low clouds is accounted only in regions without being covered by higher clouds (underestimated) CFS: Low clouds=BL+low clouds (overestimated) SO: CFS-ISCCP: underestimate the amplitudes of negative differences; overestimate the positive ones

15. JJA Low Clouds & Vertical T Gradient
T700-T850
St. Helena Radiosonde Station
Low Clouds B A
T850-T925
CFS: Region A
OBS: Region B

16. Low Clouds & Vertical T Gradient
Quasi-inversion layer
NO quasi-inversion layer
T700-T850
Inversion layer
NO inversion layer
T850-T925
CFS Region A
OBS Region B
Low Clouds

17. Region A Region B <5% <10% T700-T850 BLCloud Amount <30%
CFS Has More High Clouds than in ISCCP The Atmosphere is more Stable in Real World than in CFS
Region A
Region B
<5%
<10%
T700-T850
BLCloud Amount
<30%
<30%
T850-T925
CFS
ISCCP
CFS
ISCCP

18. UV & SST Regression onto Low Clouds in region B
Also low clouds may be associated with different climate modes: CFS: Subtropical South Atlantic Mode (SSA) OBS: Southern Tropical Atlantic Mode (STA) or Zonal Mode/Atlantic Nino
UV & SST Regression onto Low Clouds in region B
CFS
OBS

19. There are also some errors in the seasonal cycle: (1) Along Equator: 1 month delay of seasonal transition Overestimated variability (2) In the tropical South Atlantic: Without westward propagation Southerly wind errors

20. Monthly-Annual Mean of SST& Wind Stress along Equator
CFS: one month delayed transition
similar wind stress
Related to the Difference
of African Monsoon ?
Overestimated: STA mode

21. Monthly-Annual Mean of SST& Wind Stress along 15˚S
No propagation
Westward propagation
STA Mode
Warm Bias:
African Monsoon
&tropical transition?

22. Summary of Part 1: Mean & the Role of Cloud-Radiation-SST Feedback in the Biases
The seasonal mean climate is well simulated; SST zonal gradient along the equator is qualitatively correct in the CFS. However, there are warm SST biases and reduced variability in the southeastern Ocean. The annual cycle of SST along the equator is delayed by one month. Westward propagation of SST along 15oS is not seen in the CFS
CFS: Deficit low-cloud and excessive SW radiation can cause the warm SST bias in the southeastern Atlantic. In return, the warm bias does not favor inversion layer and also more low-clouds. CFS can not correctly simulate the vertical inversion layer of the temperature and the formation of low cloud. Low-cloud formation: Observed: (T850-T925); CFS: (T700-T850)
There are much more high clouds in CFS than in ISCCP. The atmosphere is more unstable in CFS than in the real world
The variability of wind stress and heat flux is too small in the southeastern Atlantic. There are double ITCZ in MAM and DJF, and too much precipitation in the western tropical Atlantic

23. Part 2: Leading Variability Patterns and the Associated Physical Processes in the Tropical Atlantic Ocean
Hu, Z.-Z., B. Huang, and K. Pegion, 2008: Leading patterns of tropical Atlantic variability in a coupled GCM. Climate Dynamics, 30 (7-8), , DOI /s x.

24. CFS Well Simulates Leading Modes
STA
NTA
SSA
ER-v2
1st
16%
2nd
15%
3rd
10%
CFS
3rd
10%
2nd
11%
1st
12%

25. Power Spectrum of Leading Modes
2.3, 1.5, 1.05 years
3.0, years
STA
4.0, years
2.4, 1.1, 0.5 years
NTA
5.5, 1.45, 1.0, 0.65 years
5.0, 1.2 years
SSA
ER-v2
CFS

26. Seasonality (Variance) of the Leading Modes
STA
NTA
SSA
ER-v2
CFS

27. Links with ENSO CFS/ER-v2 not similar STA similar NTA wrong SSA
ENSO Leading
ENSO Lagging

28. Associated downward net HF STA: thermodynamic damping & dynamic processes (similar to ENSO) NTA: thermodynamic processes &WES SSA: thermodynamic processes & WES & northwest shift
STA
NTA
SSA
NCEP/NCAR
CFS

29. STA: Lead-Lag Correlation of ATL3 (3oS-3oN, 0-20oW) with SST, HC & Wind Stress Along Equator (10 YR HF)
SST&UV
ATL3 Leading
SST&UV
ATL3 Lagging
ATL3 Leading HC HC
ATL3 Lagging
OBS:
CFS

30. NTA Patterns in MAM MAM NTA: Total (NAO/AO) OBS CFS
Asymmetric: Tropical Pacific
OBS
CFS

31. SSA pattern in DJF OBS CFS OBS SHIFT CFS DJF SSA: Total: AAO
Asymmetric: Tropical Pacific (Mo et al.)
NW shift
OBS
SHIFT
CFS

32. Summary of Part 2: Leading Variability Patterns and the Associated Physical Processes
The leading variability modes, their associated physical processes are reasonable well reproduced.
STA associated anomalies are too strong in the CFS
SSA and NTA are associated with similar physical processes in their hemispheres: zonal component (AO/NAO; AAO), zonal asymmetric variability (tropical Pacific Ocean)
The shift may be associated with teleconnection difference
The frequency feature of the leading variability modes and their links with ENSO, as well as the seasonality are similar/different between CFS and observations/reanalyses
Overall, the simulation of NTA is better than that of STA and SSA, which may be due to the fact that the ENSO influence is (not) correctly simulated in tropical N. (S.) Atlantic in CFS

33. Part 3: What are Predictive Skills, Most Predictable Patterns and the Influence of ENSO ?
Hu, Z.-Z. and B. Huang, 2007: The predictive skill and the most predictable pattern in the tropical Atlantic: The effect of ENSO. Mon. Wea. Rev., 135,

34. SST Predictive Skill: Anomalous Corr. Coefficient
Lead 1
Lead 3
Lead 5
Lead 7
CFS
Hindcast
Persistence
CFS
Minus
Persistence

35. TB is the best and Dipole is the worst, NB and SB are in between
*TB is the best and Dipole is the worst, NB and SB are in between *prediction with IC from JJA&SON is more accurate than that from MAM
JJA&SON
MAM

36. What are the most predictable patterns of SST, H200, and precipitation in March?
Why choose March as target month?
� High predictive skill in CFS (previous figure)
� Large variability (Nobre and Shukla 1996)
� Strong connection between TAV & ENSO (Tourre & White 2005)
� However, the general patterns are similar for different targeted months
Maximized signal-to-noise EOF is used to identify the most predictable pattern (MSN PC: orthogonal; EOF: not orthogonal)

37. MSN EOF1 of SST: A basin wide warming or cooling

38. MSN EOF1 SST & ENSO (the observation & predictions; Nov
MSN EOF1 SST & ENSO (the observation & predictions; Nov. IC & 4 month lead)
ER-v2 SST
CFS SST

39. MSN EOF1 of H200: A basin wide coherent variation symmetric about the equator
Nino3.4
& H200

40. MSN EOF1 of precipitation: Contrary variation of e
MSN EOF1 of precipitation: Contrary variation of e. tropical Pacific and the tropical land

41. Maximum Signal-to-Noise EOF1 (CFS Hindcast SST)
March Forecast:
at 4-months Lead
OBS
Correlation with MSN EOF1
CFS
SST
H200

42. 12-8OS temperature regression onto MSN EOF1 of precipitation
Theory of Chiang & Sobel (2002): The existence of atmosphere convection is crucial for temperature variations to propagate from troposphere to surface
OBS X
CFS
Warm Bias
12-8OS temperature regression onto MSN EOF1 of precipitation

43. Summary of Part 3: Predictive Skills, Most Predictable Patterns and the Influence of ENSO (1) There are reasonable predictive skills over the tropical Atlantic. The skill is higher in the tropical North Atlantic than in the tropical South Atlantic, with IC in JJA&SON than in MAM (2) The skills and the most predictable pattern are largely attributed to the influence of ENSO in the tropical Atlantic Ocean

44. Summary
MEAN: The seasonal mean climate is well simulated. There are warm SST biases and suppressed variability in the southeastern ocean. Deficit low-cloud and excessive SW radiation are associated with the warm SST bias in the southeastern Atlantic. In return, the warm biases do not favor inversion layer. CFS can not correctly simulate the vertical inversion layer of the temperature and the formation of low cloud. Low-cloud formation: Observed: (T850-T925); CFS: (T700-T850). There are much more high clouds in CFS than in ISCCP. The atmosphere in the tropical Atlantic is more stable in the real world than in CFS.
MODES: The leading variability modes, their associated physical processes, as well as the seasonality are well reproduced. The frequency feature and the association with ENSO are partly simulated. STA associated anomalies are too strong in the CFS; SSA and NTA are associated with similar physical processes in their hemispheres: zonal component (AO/NAO; AAO), zonal asymmetric variability (tropical Pacific Ocean). The shift of SSA may be due to the teleconnection difference. Overall, N. Atlantic is better simulated than S. Atlantic
PREDICTION: The predictive skill depends on region and initial condition month. The skill is higher in N. Atlantic than in S. Atlantic, and with IC in JJA&SON than in MAM. The predictive skill and most predictable pattern are associated with the influence of ENSO. The unrealistic part of the most predictable pattern of SST in the SE Atlantic is partly caused by the biases.

45. Related Publications
Hu, Z.-Z., B. Huang, and K. Pegion, 2008: Leading patterns of tropical Atlantic variability in a coupled GCM. Climate Dynamics, 30 (7-8), , DOI /s x.
Hu, Z.-Z., B. Huang, and K. Pegion, 2008: Low-cloud errors over the southeastern tropical Atlantic Ocean in the NCEP CFS and the association with lower tropospheric instability and air-sea interaction. JGR-Atmos, 113, D12114, doi: /2007JD
Hu, Z.-Z., B. Huang, and K. Pegion, 2009: Biases and the most predictable patterns in the NCEP CFS over the tropical Atlantic Ocean. The Atlantic Ocean: New Oceanographic Research, Nova Science Publishers, Inc., New York, USA.
Hu, Z.-Z. and B. Huang, 2007: The predictive skill and the most predictable pattern in the tropical Atlantic: The effect of ENSO. Mon. Wea. Rev. 135,
Comments or Ask reprints/PDF files, send to

46. Future Work: In cooperation with CFS experts, do some sensitive experiments by specifying low clouds in the model to explore the impact of low clouds on the CFS biases in the tropical oceans

48. What is the Maximized Signal-to-Noise EOF (I)
The MSN EOF is a method to derive patterns that optimize the signal-to-noise ratio from all ensemble members.
This approach was developed by Allen and Smith (1997) and has been used in Venzke et al. (1999), Sutton et al. (2000), Chang et al. (2000), and Huang (2004).
An ensemble mean is supposedly including a forced and a random part, which may be attributed respectively to the prescribed external boundary conditions and the unpredictable internal noise. In our case, the forced part is associated with the predictable signals which show certain consistency among different members of the ensemble predictions.
The leading MSN EOF mode is the one with the maximum ratio of the variance of the ensemble mean to the deviations among the ensemble members.
In this work, we only analyze the leading MSN EOF mode, which is defined as the most predictable pattern and is significant at the 95% level using a F-test.
Details of this method were documented in Allen and Smith (1997), Venzke et al. (1999), and Huang (2004).

49. MSN EOF (II) Variable: Xensemble =Xforced+ Xrandom
F is estimated from the first K weighted EOF patterns of the deviations X’i =Xi-Xensemble (i denotes the ith member within the ensemble).
The matrix of eigenvectors (E) of FT CensembleF contains a set of optimal noise filters, which can be restored into physical space by É=FE.
The optimal filter (the 1st column vector é= É) maximizes the ratio of the variances of the ensemble mean and within-ensemble deviations (Venzke et al., 1999).
The optimally filtered time series of Xensemble (i.e., its projection onto é) gives the 1st MSN principal component (PC).
In practice, one can simply first project Xensemble onto F to form the pre-whitened data in the noise EOF space and then conduct an SVD to get both E and all MSN PCs simultaneously (Huang 2004).
The 1st MSN EOF pattern, which represents the dominant spatial response, is derived by projecting Xensemble onto the 1st MSN PC.
Variable: Xensemble =Xforced+ Xrandom
Covariance: Censemble =Cforced+ Crandom (If Xforced and Xrandom are temporally uncorrelated)
Crandom = Cnoise/15 (Cnoise is the average noise covariance of the 15 ensemble members
To find the eigenvector of Cforced, the key procedures is to eliminate the spatial covariance of the noise. Which is equivalent to a transformation F such that
FT CrandomF= I
The transformation guarantees that FT CforcedF and FT CensembleF have identical eigenvalues.

51. CFS Error Growth with Season

52. DEMETER Error Patterns (Nov. Mean)

53. Compared with other world famous coupled models (CCSM, MPI, ECMWF, MO, and so on), CFS did very good job in the tropical Atlantic, particularly in simulating the W-E SST gradient. However, the error patterns in the tropical Atlantic are similar among these models.

54. Hu, Z.-Z. and B. Huang, 2007: The predictive skill and the most predictable pattern in the tropical Atlantic: The effect of ENSO. Mon. Wea. Rev. 135,

55. Definition of clouds in ISCCP & CFS
ISCCP (Weare 1999; Rossow&Dueñas 2004)
High
CFS (Xu& Randall 1996)
~350hPA Middle
~650hPA Low
~900hPA BL

56. Inversion layer in CFS, Reanalyses & Radiosonde

57. Seasonal Mean and STDV of Wind Stress (1) General pattern and meridional STDV gradient well simulated (2) Small STDV along Equator (3) Too small STDV in southeastern

58. Mean & STDV of Downward Surface HF (a) Following solar activity (b) Suppressed variability in southeastern

59. Mean & STDV of Precip: (1) Large biases in the west (2) Double ITCZ in MAM & DJF

60. HC & Wind Stress regression on to STA (all seasons) (Signals too strong in CFS)
OBS
CFS

61. Clouds in ISCCP & CFS ISCCP CFS ISCCP (Weare 1999; Rossow&Dueñas 2004)
CFS (Xu& Randall 1996)
ISCCP
CFS
High
High
Middle ( hPa)
Middle ( hPa)
Unstable & convective clouds
Stable & stratus clouds
Low ( hPa)
Low
(680-surface)
BL (>900hPa)

62. Connection of the Modes with SLP

63. Difference in the connection of SSA and SH

64. MSN EOF1 of H200 is also associated with ENSO

65. MSN EOF1 of precip is also associated with ENSO

66. CFS overestimates the contrary variation
Nino3.4 &
observed P

74. Mean Low-Clouds and differences
MAM
JJA
shift
SON
DJF
CFS
ISCCP
CFS-ISCCP

75. Mean SW (down-up) and differences
MAM
JJA
SON
DJF
CFS
ISCCP
CFS-ISCCP

76. Mean SST and differences
MAM
JJA
SON
DJF
CFS
ERv2
CFS-ERv2

77. Possible Influence of the Cloud Definition
MAM
JJA
SON
DJF
BL Clouds/CFS
BLC+LC/CFS
CFS(BLC+LC)-ISCCP (LC)

78. Spatial Coherence between Low Clouds and Atmospheric Stratification
CFS Simulation
OBS
Low Cloud
T( )
T( )

79. Temperature profile in lower troposphere
Inverse layer is stronger in observed than in CFS

80. Scatterplot: Low Cloud vs. T Stratification
CFS ∆T
OBS
Low Cloud Amount

81. BL Clouds in CFS & T Stratification
BL Cloud Amount ∆T
ISCCP
Low Cloud Amount

82. Mixed Relation of Stratification and Middle Clouds
Comparable Middle Cloud Amount in CFS & ISCCP (Region A; Region B)
T700-T850
T850-T925
T925-T100
CFS
ISCCP

83. Cloud-Radiation-SST feedback in observation and model (a) Observation: Warm SST in the southeastern Atlantic in early boreal spring suppress the formation of low-cloud. Lacking low-cloud results in excessive SW radiation reaching sea surface, the warm SST is further intensified and extend northwestward in boreal summer. (b) CFS: Deficit low-cloud and excessive SW radiation may be the main reason resulting in the warm SST bias in the southeastern Atlantic. The cloud-radiation-SST feedback is not well simulated. (c) CFS model can not correctly simulate the vertical inversion layer of the temperature and the formation of low cloud: Low-cloud formation: Observed: (T850-T925); CFS: (T700-T850)

84. To All NCEP Collegues