Near-Surface Data Assimilation in the NCEP Regional GSI system: Use of Mesonet Data

1
Near-Surface Data Assimilation in the NCEP
Regional GSI systemUse of Mesonet Data a New
Forward Operator

• Seung-Jae Lee

2
Introduction

• The NCEP analysis system- many changes since
  1978- 3DVAR-Gridpoint Statistical-Interpolation
  (GSI) System - can assimilate diverse kinds of
  obs.
• Importance of and demand for Real-Time Mesoscale
  Analysis (RTMA)- near-surface data assimilation
  is one of the challenges to conquer
• U.S. meso-network systems measure and provide
  information on the environment at the size and
  duration of mesoscale wx events.

3
Purposes

• assimilation of the sfc mesonet data in the NCEP
  regional GSI system- test of regional background
  error are also considered
• understanding the characteristics of innovations
  of the sfc mesonet data

4
Mesonet data QC

• We at NCEP currently do not perform any automated
  platform-specific QC on any surface data and
  simply honor quality markers provided by the FSL-
  MADIS.
• NCEP could (but not currently doesnt) place
  manual quality markers on the mesonet data. This
  would be done by either putting reports on the
  rejectlist which might be updated monthly or on a
  report-by-report basis by the NCEP Senior Duty
  Meteorologists.
• We perform some gross checks and flag data that
  are outside reasonable limits, data with mising
  lat/lon etc.

5
Mesonet data PrepBUFR file

• all mesonet data are included but the observation
  error for all mass and wind observations is set
  to missing? In this study, we modify the
  observation error file to assign the mesonet
  observations the same observation error as for
  METAR mass and wind observations
• The time window over which data collection is
  performed is /- 1.5 hours, and the 3DVAR
  analysis experiment is valid for t0h.
• Near-surface data usedMesonet data, METAR data,
  Synoptic land data, Synoptic sea data minor
  sfc data

6
Modification to background error statistics

• Until now, the background error for regional
  assimilation has been a downscaled version from
  the background error derived from the global
  model (? Default)
• In the existing approach, horizontal scales are
  estimated from derivatives, but in the new
  approach, these are estimated using
  auto-covariances (W.-S. Wu 2005, personal
  communication)

7
First Guess

• WRF-NMM 8 km
• There are three domains available for the WRF-NMM
  8 km model Initial fields for the western,
  central, eastern domains are at 06, 12 and 18
  UTC, respectively.
• Experiments are carried out for each model
  domain Single time analyses

8
experimental design
analysis use of mesonet data background error Note
AN1 No global operation (control run)
AN2 Yes global impact of mesonet data
AN3 Yes regional (WRF-NMM) impact of new regional background error statistics
9
Case 1 western USA, 0600 UTC 14 Feb 2005
Nighttime (around midnight)
Clear sky large-scale nighttime sfc inversion
10
analysis increments (case 1)
A -G lt 0 over much of the domain
11
Mean values of (Oi Gi) at each category of dP
The obs are consistently colder than the model,
leading to the large negative anal inc. at the
sfc. Part of the reason this is undesirable is
that T inc are coupled with W incin approximate
geos. balanceand large wind inc are created at
middlelevels in the troposphere.
12
Case 2 central USA, 1200 UTC 10 Mar 2005
Early morning (around 6-7 am)
- sfc inversion by clear weather- A low pressure
system in the northern-central region
13
Analysis Increments (case 2)
6th-7th 920-900hPa
- Smaller and detailed str- Positive in the
N-easternregion where the L is located
14
(No Transcript)
15
Case 3 eastern USA,1800 UTC 23 Mar 2005
Daytime (2 pm)
A large low-pressure system over the eastern coast
16
Analysis increments (case 3)
Unstable sfc
17
Small bias
Even in the east coast case,the A-G is large.It
just has smaller scales.
18
Accumulated Statistics of (O-G)

• Understanding the characteristics of innovations
  (observedguess) of the near-surface data
• Long-term statistics 1 month
• Scatter diagram of (O-G) during May 2005 for
  nighttime (0600 UTC), early morning western
  (1200 UTC),daytime eastern (1800 UTC) domains.
• Linear regression, bias, rmse etc.

19

• Western domain (0600 UTC)
• Sfc mesonet T data have a considerable amount of
  outliers compared with other land sfc T data.
• Bad obs or local effects
• Some stations can be seen to produce the same
  values regardlessof model fcsts.

20
Central domain (1200 UTC)
Very few outliers in METAR? possibility of local
effect
Many outliers This can mean quality markers
placed on the data by the FSL-MADIS QC are of
little valuenot only for wind but also T
data. ? Proper QC is required.
21
Eastern domain (1800 UTC) Slope and correlation
coeff. of land sfc T data are good and similar
in all three domains. However,in the case
of synoptic sea sfc data, they show a peculiar
pattern in the western domain a steep slope
andvery low correlation coeff. (lt0.5) ? Ocean
wave effect etc.
22
The nighttime western and central domains
indicate a model warm bias. The western domain,
in particular, shows a model warm bias of about
2.2 C at nighttime when compared with the
eastern domain which did not show any obvious
bias. The o-g statistics as a function
of surface pressure difference in the eastern
domain seems to indicate that the mesonet
observations have small temperature bias (about
0.5 C).
23
Horizontal distribution of stations with large
(O-G)
Unlike the other two types of sfc stations,
surface mesonet stations are very dense,
particularly around large cities. Stations
with large innovations are distributed
uniformly in the nighttime western and central
domains, while are mainly located in the large
cities in the daytime eastern domain.
In the case of the eastern domain, it is 14 LST
and the synoptic situation is characterized by
many local and unstable situations with
small-scale variation in daytime. In the case of
the western (central) domain, it is midnight
(dawn) and frequent large-scale inversion
situation prevail.
24
The number of stations as a function of surface
pressure difference
Western (0600 UTC)
Central (1200 UTC)
Eastern (1800 UTC)
The western domain revealed an asymmetric
distribution that implies that there are many
stations where the model surface is higher than
the observation surface. This could partly
account for the model warm bias in the western
domain at nighttime. In the other two domains,
the distributions are comparatively axisymmetric.
25
A complex forward operator
26
Motivation
Cost function
x an N-component vector of analysis increment B
the N by N previous forecast-error covariance
matrix O the M by M observational
(instrumental)-error covariance matrix F the M
by M representativeness (forward operator)-error
covariance matrix H a linear or nonlinear
transformation operator y an M-component vector
of observational residuals that is, M the
number of degrees of freedom in the analysis
and N the number of observations.
27
H, the forward model

• Converts the analysis variables to the
  observation type and location (May include
  variable transformations in case of, for example,
  radar or satellite data assimilations).
• The quality of simulated observation, the model
  state converted to observation units, depends
  on the accuracy of the forward operator
  Inaccurate pseudo-observations give rise to
  errors in observation increments leading to bad
  analyses.

28

• The accuracy of anl is dependent on the
  effectiveness of algorithm used to match obs with
  the bg values.
• Currently, there does NOT exist any forward model
  for near-surface variables in the NCEP GSI
  system Just involves simple interpolation of the
  background value to the location of the
  observation.

29
the new fwd operator

• Uses a similarity theory based on Dyer and Hicks
  formula which has been adopted in model PBL
  parameterization such as surface layer process in
  MRF PBL scheme (Hong and Pan 1996).
• realized in the MM5-3DVAR system (Barker et al.
  2004). It has been used for surface data
  assimilation in operational analysis systems and
  has proven to be encouraging in many contexts
  (e.g., Shin et al. 2002 Hwang 2005).

30
assumptions

• All the surface observation sites are assumed to
  be located at the model surface, regardless of
  the actual difference in elevation between the
  surface measurement and the model surface.
• The different types of surface observations are
  directly assimilated without any modification.
  The observed surface pressure is still reduced to
  the models lowest level.
• Near-surface wind and mass variables are obtained
  at different heights using the surface layer
  similarity theory Near-surface wind is obtained
  at 10 m, while temperature and specific humidity
  are obtained at 2 m.

31
Forward model description

• Monin-Obukhov similarity theory

Input output
Atmosphere PBL(SLEL) FA
32
near-surface variables
wind
temperature
specific humidity
Z 2 or 10 meter
33
stability functions
Turbulence regime determines function
type
34
turbulence regimes
(2) Damped mechanical
(4) Free convection
(3) Forced convection
35
Stability-dependent functions
(1) Stable
(2) Damped
(3) Forced convection
where
36
(4) Free convection
37
Implementation and tests

• Incorporated into the latest GSI version with
  some modification ? Jun 2005
• Application - eastern domain, 1800 UTC 23 Jun
  2005
• Parallel runs (during early July 2005)- western
  domain Jul 7, 11-13, 18 - central Jul 2-14,
  17-18- eastern Jul 6-12, 14,15,17

38
Case eastern domain,1800 UTC 23 Jun 2005
39
Large innovation sites
Newfwdmodel
Oldsimpleinterpol.
40
Innovations as a function of dp
Worse over sea, why? - Because of setting z2 m
(not the height of buoy) over water also.- Ocean
wave effect -gt roughness is uncertain. - Physics
etc. ? Despite this, the total bias was reduced
from 2.2 to 0.9.
41
Parallel run results Early July, 2005
old
new
Western, 0600 UTC
42
old
new
Central, 1200 UTC
43
old
Eastern, 1800 UTC
new
44
old
Western, 0600 UTC
new
45
old
Central, 1200 UTC
new
46
old
new
Eastern, 1800 UTC
47
Summary and Conclusions

• We have assimilated sfc mesonet data in the NCEP
  regional GSI using the same observation error as
  that adopted for METAR data within the WRF-NMM (8
  km) 3DVAR system.
• In the single-time anl experiments, the anl field
  was shown to contain mesoscale (smaller and
  detailed) structures as mesonet data are added.
• When the regional bg err statistics are used, the
  overall pattern is similar to the downscaled
  global version but the amplitude is somewhat
  intensified. It is believed to be the results of
  smaller (shallower) vertical structures in the
  regional bg err covar.

48

• sfc mesonet T data were found to have a
  considerable amount of outliers compared with
  other land sfc T data.
• The nighttime western and central domains
  indicated a model warm bias. The western domain,
  in particular, showed a model warm bias of about
  2.2 C at nighttime when compared with the
  eastern domain which did not show any obvious
  bias.
• Stations with large innovations are distributed
  uniformly in the nighttime western and central
  domains, while they are mainly located in the
  large cities in the daytime eastern domain. ?
  These differences could also be the result of
  urban heat island effects (not contemplated in
  the model) or erroneous station groups.

49

• The current near-surface observation operator in
  the NCEP GSI system was improved from a simple
  linear interpolation to a more complex similarity
  model.
• In the intercomparison of the old and new forward
  operators using case experiments and long term
  runs, the complex forward model is shown to
  improve the innovation statistics substantially.
  ? This is due to the realistic considerations of
  surface characteristics (e.g., roughness length)
  and atmospheric stability within surface layer.

50

• Sources of disagreement between observations and
  the model include observation errors and errors
  related to the model. It is shown that the latter
  was reduced by introducing a more complex forward
  model in the NCEP regional GSI analysis system.
• The new operator has significant and practical
  potential to assimilation of diverse surface data
  in the NCEP regional GSI system. It can be
  applied to not only regional but also global GSI
  system.
• It has diverse applications to atmospheric and
  oceanic data assimilation and can be used for not
  only surface data over land and but also
  satellite data, such as QuikSCAT sea surface
  wind.

51
On-gong or possible future works

• Tangent-linear and adjoint versions are under
  implementation.
• Application of new fwd model to GSI-2DVAR
• Connection with non-linear quality control to the
  mesonet data
• Link with anisotropic background error covariances

52
Thank you !