Changes to the NCEP Operational Eta Analysis
Eric Rogers David Parrish Geoffrey DiMego
Mesoscale Modeling Branch, Environmental Modeling Center, National Centers for
Environmental Prediction
I. INTRODUCTION
In an effort to improve the timeliness of NCEP's mesoscale forecast guidance, a series of
changes to the operational "Early" Eta system were made in 1998. These changes were
implemented in two stages :
9 FEBRUARY 1998 (detailed in Rogers et.al, 1997)
- Resolution increased from 48 km / 38 levels (Eta-48) to 32 km / 45 levels (Eta-32)
- Replacement of the Eta Optimum Interpolation analysis with a regional 3-dimensional variational analysis (3DVAR, Parrish et.al 1996).
- Use of a "partial" continuous cycle in the Eta Data Assimilation System (EDAS), in which soil
parameters, turbulent kinetic energy (TKE), and cloud water are obtained from the previous
EDAS cycle instead of the NCEP Global Data Assimilation System (GDAS), accompanied by an
increase in the number of soil layers from two to four.
3 JUNE 1998:
- The EDAS was converted to "full" cycling mode, with atmospheric variables (temperature, wind, moisture) cycled from the previous EDAS cycled in addition to soil/cloud/TKE parameters.
- The 0300 UTC Eta-29 (so-called "Meso" Eta), which had continued unchanged after the 9
February 1998 implementation, was replaced by the Eta-32, initialized from the fully cycled
EDAS. The 1500 UTC Eta-29 was replaced by an 1800 UTC run of the Eta-32, also initialized
from the fully cycled EDAS.
Examination of Eta-32 3DVAR analyses by NCEP scientists and NWS field forecasters revealed
that the greatest deficiency in the 3DVAR was its degraded analysis of surface and lower
tropospheric data, particularly moisture. This problem is illustrated in Fig. 1, which shows the
observed and analyzed sounding at Rapid City, SD at 0000 UTC 13 July 1998. The analysis is
from a cycled 80-km EDAS which runs the same 3DVAR analysis as the operational Eta-32. The
80 km dew point temperature is clearly deficient below 700 mb, with a 5 deg C error at 850 mb.
A similar signal was seen in the operational Eta-32 (not shown), After careful examination of the
of this case, it was determined that the vertical correlation error length scale for moisture and the
background error covariances were both too low. These problems would cause the 3DVAR
analysis to give more weight to the first guess and would prevent it from accurately analyzing
details in the low-level moisture field. Additionally, an error was found in the 3DVAR which
caused it to essentially exclude all surface data. Accordingly, the 0000 UTC 13 July 1998
analysis was rerun with modifications made to alleviate the above problems. The new analysis at
Rapid City is shown in Fig. 2. The low-level moisture analysis, while still too dry, is a clear
improvement over the 80 km analysis using the original operational 3DVAR analysis shown in
Fig. 1.
These changes were tested for a 3 week period in July 1998 using the 80 km Eta parallel system
(limited computing resources at NCEP prohibited any extensive testing at 32 km resolution).
The results (not shown) revealed improved analysis fits to surface and rawinsonde data
(especially moisture) with little impact on forecast performance over the contiguous U.S. Based
on these findings, these changes were implemented into the operational Eta-32 on 3 November
1998.
II. DEFICIENCIES IN THE 3 NOVEMBER 1998 VERSION OF 3DVAR
During the winter of 1998-99 it became increasingly apparent to NWS field forecasters, EMC
personal, and academic observers that the Eta model forecasts had decreased skill from the
previous winter in comparison to the Nested Grid Model (NGM) and the NCEP Global Aviation
Model (AVN). Fig. 3 shows the 24-h forecast precipitation equitable threat score for December
1997 - February 1998 for the Eta, NGM, and AVN models, while Fig. 4 shows the same skill
scores for December 1998 - February 1999. Comparison of these charts show that during 1998-99 all three models had a drop in skill at the lower thresholds compared to the previous winter.
The Eta model's skill dropped at all thresholds, with a decreases of ~20% at the 0.75 inch and
1.00 inch thresholds. The AVN model had a slight increase in skill at these same thresholds,
undoubtedly due in part to model/analysis changes made during 1998 (Derber et.al, 1998).
The most persistent feature observed in Eta model forecasts during the 1998-99 time frame was
degraded analyses and forecasts in the eastern Pacific Ocean. Junker (personal communication,
1999) observed that the Eta model moved amplifying mid-tropospheric short wave troughs faster
than the AVN and NGM. Mass (personal communication, 1999) observed that the AVN analysis
of cyclones in the eastern Pacific during this period was consistently better, with the AVN
producing better forecasts than the Eta model. An example of such behavior is depicted in Fig. 5,
which shows the 48-h eastern Pacific Eta-32 and AVN sea level pressure forecasts valid at 1200
UTC 17 March 1999. The offshore cyclone was predicted to explosively deepen to < 980 mb by
the AVN model, while the Eta model predicted a much weaker storm. The GDAS analysis valid
at this time (Fig. 6) shows that although the AVN forecast position about 300 km too far south,
the central pressure forecast was nearly perfect. The NGM 48-h forecast (not shown) position
was closer to the observed center than the AVN but with a central pressure forecast of 990 mb.
Although there are many differences between the EDAS/Eta and GDAS/AVN initialization
procedures, including a different observation mix (e.g., later data cutoff time for the AVN vs. Eta
on-time analysis; use of satellite radiance data in GDAS vs. satellite thickness retrievals in the
EDAS over water), a preliminary examination of the Eta vs. AVN observation differences in the
15 March 1999 case and others (not shown) revealed that there was no obvious data source issue
that would explain the poor performance of the Eta model in the eastern Pacific. Thus, it became
apparent that the November 1998 changes to the Eta 3DVAR should be revisited.
. In a well-behaved multivariate analysis scheme (such as optimum interpolation or 3DVAR)
wind-only observations will modify the mass field through a balance condition, which for the Eta
3DVAR is a thermal wind constraint (Parrish et.al., 1996). Conversely, mass-only observations
will modify the wind field through the same balance condition. Where mass and wind are both
observed together (e.g., rawinsondes) the observation itself should define the balance. Balance
problems can arise with an incorrectly tuned analysis in regions where there are predominately
mass or wind-only observations (such as satellite winds over oceans) or single-level
observations.
Upon closer examination of the Eta analyses in the eastern Pacific, it was obvious that there was
an overall lack of geostrophic balance between mass and wind analysis corrections away from the
reference level (500 mb) in the 3DVAR analysis. An example at 850 mb from a cycled 80 km
EDAS run valid at 0000 UTC 15 March 1999, using the current operational 3DVAR analysis is
shown in Fig. 7. This chart shows the analysis-first guess height difference for wind and height,
along with the rawinsonde height/wind-first guess difference. If the balance constraint is well
behaved, then a positive height correction (such as the one associated with the St. Paul Island,
Alaska rawinsonde near 57N 170W) will be accompanied by a anticyclonic wind correction
coincident with the height correction center. As one can see from Fig. 7, the wind correction is
displaced about 10 degrees longitude west of the height correction. A similar (though less severe)
displacement of the wind correction is associated with the negative height correction near 42N
152W.
Since the 3DVAR (and most objective analysis schemes) can not simultaneously perform a
perfect fit to observations and have exact adherence to mass-wind balance, it is apparent that the
November 1998 tuning of the 3DVAR to improved the analysis fit to the rawinsonde data
weakened the ability of the analysis to created balanced mass-wind analysis corrections, leading
to analyses which were essentially univariate. This problem would be most severe in regions and
at analysis times without widespread rawinsonde data.
III. MODIFICATIONS TO THE 3DVAR ANALYSIS - SINGLE OBSERVATION TESTS
The observed 3DVAR analysis problems described above has led to an extensive examination of
the 3-dimensional structure of the analysis increments in the current operational system. In
addition to the poor mass-wind balance, it was determined that the horizontal and vertical
correlation length scales were too short. These length scales will determine the horizontal and
vertical extent to which an observation will modify the first guess away from its location. Thus,
adjustments were made to both improve geostrophic coupling of mass/wind increments and to
lengthen the horizontal and vertical correlation length scales. To illustrate the impact of these
changes, single observation tests of the 3DVAR analysis at 32 km and 80 km were performed.
Fig. 8a shows a vertical cross-section of the analysis height and wind correction caused by a
single observation with height observation-first guess difference of 10 m at 200 mb. Fig. 8b
shows the same feature from the modified 80-km 3DVAR analysis. The modified 3DVAR
analysis not only produced a larger wind increment with a greater vertical extent, it also
increased the height correction. Fig. 9a and 9b show the same experiment using the 32 km
3DVAR analysis.
These single observation tests were also performed at 900 mb and 500 mb and
these tests showed a similar response to those done at 200 mb.
IV. STATISTICAL SUMMARY OF NEW 3DVAR ANALYSIS PERFORMANCE :
PRECIPITATION SKILL SCORES AND GRID-OBSERVATION VERIFICATION
The modified 3DVAR analysis tested in a retrospective 80 km parallel system for the period 3
December 1998 - 16 January 1999. A fully cycled EDAS was run to initialize both the control
and experimental Eta-80 forecasts. The equitable threat score for 24-h accumulated precipitation
from control and experimental Eta-80 forecast is shown in Fig. 10. A 10-20% increase in skill is
seen at all thresholds. The similar bias scores observed for the same period (Fig. 11) indicated
that the improved skill scores are not the result of an increase in forecast precipitation.
To assess the impact of the new 3DVAR analysis on the mass, wind, and moisture forecasts, the
00-h, 24-h, and 48-h forecast vs. rawinsonde root-mean-square (RMS) error of temperature (Fig.
12), vector wind (Fig. 13), geopotential height (Fig. 14) and relative humidity (Fig. 15) are
presented. On these figures, the solid lines are the 80-km control run using the operational
3DVAR analysis (designated ETAV) and the dashed lines are the 80-km forecasts with the
modified 3DVAR analysis (designated ETAY).
More details on this statistical summary and additional charts can be found at:
http://sgi62.wwb.noaa.gov:8080/3DVAR_STATS/index.html
A) TEMPERATURE
Since the modified analysis was tuned to improve the mass-wind balance, one would expect the
00-h forecast fit to observations to worsen. This is confirmed by the 00-h temperature trace in
Fig.12, which shows a higher RMS at all levels. However, by 24-h the errors between the control
and experimental forecasts are similar, and the experimental run has slightly lower temperature
errors at most levels by 48-h.
B) VECTOR WIND
Fig. 13 shows that the improved mass-wind balance in the modified 3DVAR analysis had a
dramatic positive impact on the RMS vector wind errors. The forecasts from the modified
analysis had lower wind errors at all levels and at all times, with the greatest impact at jet stream
level. At 200 mb the 24-h RMS vector wind error from the experimental run is very close to the
00-h RMS vector wind error from the control run
C) GEOPOTENTIAL HEIGHT
The RMS errors for geopotential height in Fig. 14 mirror the temperature errors described above,
with a slightly degraded fit to the height data by the modified 3DVAR analysis. There is a
tendency for the difference in control-experimental RMS height error difference to increase with
forecast hour.
D) RELATIVE HUMIDITY
The RMS error of relative humidity (Fig. 15) from the experimental Eta-80 forecasts is very
similar to the control run. This indicates that modified 3DVAR had the desired effect of retaining
the improved fit to moisture data implemented in the November 1998 version of 3DVAR.
V. IMPACT OF ANALYSIS CHANGES :15 MARCH 1999 CASE
In addition to the December 1998-January 1999 retrospective test, a real-time 80 km test of a
fully cycled EDAS with the modified 3DVAR analysis was done from 13-31 March 1999. The
impact on quantitative skill scores (not shown) mirrors the results seen from the December 1998
- January 1999 retrospective test.
Fig. 16 shows the wind and height increments from a cycled EDAS analysis valid at 0000 UTC
15 March 1999 the modified 3DVAR system. Compared to the control 80 km run shown in Fig.
7, there is better balance between the mass and wind increments, as can be seen in the better
coherence associated with the positive height correction north of St. Paul Island, Alaska than was
seen in the control run. Fig. 17 shows the control and experimental 48-h Eta-80 forecasts valid at
1200 UTC 17 March 1999. The forecast using the modified 3DVAR produced forecast the
cyclone to move more slowly eastward and predicted a central pressure of 994 mb, 8 mb deeper
then the control Eta-80 forecast.
VI. CONCLUSIONS
By improving the mass-wind balance and the correlation length scales in the 80-km 3DVAR
analysis, forecasts using this modified version improve Eta-80 forecasts and reduce the observed
biases seen in the operational Eta system. An appendix which will soon be attached to this TPB
will present results from a limited number of Eta-32 forecasts using the modified 3DVAR
analysis.
VII. REFERENCES
Derber, J., and coauthors, 1998: Further changes to the 1997 NCEP operational MRF Model
Analysis/Forecast System : The use of TOVS level 1-b radiances and increased vertical diffusion. NWS Technical Procedures Bulletin No. 449, NOAA/NWS, Washington, DC.
Parrish, D. and coauthors, 1996: The regional 3D-variational analysis for the Eta model.
Preprints, 11th AMS Conference on Numerical Weather Prediction, 19-23 August 1996, Norfolk,
VA.
Rogers, E., and coauthors, 1997: Changes to the NCEP Operational "Early" Eta Analysis / Forecast System. NWS Technical Procedures Bulletin No. 447, NOAA/NWS, Washington, DC. [ Available at http://www.nws.noaa.gov/om/tpb/447.htm ]
