gps 2004

1
Assimilation of RO data using NCAR Data
Assimilation Research Testbed Liu H. and J.
Anderson Data Assimilation Initiative,
NCAR Consultants B. Kuo and S. Sokolovsky,
COSMIC/UCAR

2
Background
As so far, most studies of assimilation of RO
data were done by variational approach,
e.g., use of MM5/4DVAR for refractivity
NCEP/3DVAR for bending
angle One major limitation of the
variational approach The forecast-error
covariances is constant in time and isotropic
It may not be so accurate for daily weather
conditions In some variational systems, no
correlations btw Q and T in B.
3
Background (cont.)
Now, ensemble filter is becoming an
alternative and popular assimilation approach.
Advantages 1. Flow dependent forecast
error covariances estimated from
ensemble forecasts 2. Linearization and
adjoint of forecast model and observation
operators unnecessary Complex NWP and
climate models can be easily implemented.
4
NCAR Data Assimilation Research Testbed (DART)
It is designed for sharing and testing new
ideas and techniques in data assimilation
community It is based on an ensemble filter by
Anderson (2004, etc) A number of models are
implemented for various purposes Lorenz
models (L63, L96) GFDL B-grid model
NCAR WRF (meso-scale model) NCAR CAM
(climate model) Other models can be added
relatively easily
5
Assimilation of RO data with DART

• Recently, we began assimilation of RO
  refractivity using DART
• The purpose is to examine value of RO data to
  improve global analyses and forecasts by using
  ensemble filter approach.
• The current available global model in DART,
  CAM, is used as assimilation model.

6
Whats been done so far
Developed non_local refractivity/excess
phase path observation operators for DART/CAM,
which may also be used for WRF. Examined the
necessity of using non_local refractivity with
CAM. Assessed the discretization
sensitivities of the non_local refractivity
operator in realistic assimilation setting for
CAM. Completed an assimilation test of CHAMP
RO refractivity using the form of excess phase
path and initial results are encouraging.
7
Non-local refractivity (Sokolovsky
et al, 2004)
It consists of 2 steps 1. First, get excess
phase path from horizontal integration of model N
along a straight line below model top 2.
Then non_local refractivity is obtained by
8
Modeled local refractivity (
)
It is obtained simply by a linearly
interpolation (vertical and horizontal) of the
3-D log of the modeled refractivity on CAM model
grid (Nmod) to any RO observation locations.
9
Comparison of Nnon_local with Nlocal
To ensure high accuracy, the
discretizations of the integrations of Nnon_local
are set to very high resolutions (not feasible in
realistic assimilation case of CAM). 1.
Horizontal step-size is 1km. 2. Vertical
step-size is 20 m in low troposphere and 150m
at height of 30km. 3. 158 CHAMP profiles
were simulated. 4. CAM T42 forecast was
used as background.
10
RMS Max difference between
and
Average difference is small ( 0.2) (N obs.
error is 1) Max difference is significant
below 6km (2.5).
11
Latitude distribution of Max diff. between
and
The large difference is within tropics and
mid-latitudes
12
1st conclusion Even for CAM T42, it is
helpful to use below 6km within
tropics and mid-latitudes as model counterpart to
assimilate RO refractivity. may be
good in other areas. For T83 or higher
resolution, may be necessary for
larger areas.
13
Difficulties with using Nnon_local
1. It is found that the non_local refractivity
is very sensitive to its discretization and
dealing with its modeling error in realistic
situation for CAM is not easy.
2. The horizontal and vertical integrations
make the observation operator relatively more
complicated.
14
Excess phase path as observable ( proposed by
Sokolovsky et al, 2004)
Advantages 1. Only horizontal integration is
involved and easy for implementation with
ensemble filter. 2. The impact of
discretization error of S may be reduced or
cancelled when Sobs and Smod use the same
discretization..
15
Discretization sensitivity of excess phase path

• (Sobs Smod) for 3 discretization resolutions
  1km, 5km, and 25km
• The differences is 0.4 and less than the
  observation error (1)
• The 25km resolution is feasible for CAM

16
An assimilation test with real data
1. 158 CHAMP profiles were used and excess
phase path was used as an observable. (25km in
discretization) 2. A temporary constant 1
observation error was used. (a better
estimate is being considered) 3.
Assimilation was limited within troposphere
(lt12km) at the moment. 4. 20 ensemble
forecasts were used to estimate the forecast
error co-variances in the filter. 5. CAM
climatology is used as first guess.
17
An assimilation with real data Bias error
Evident bias exists in troposphere in guess and
it is much reduced in analysis
18
An assimilation with real data RMS error
The RMS error is reduced significantly to the obs
error level (1)
19
2nd conclusion
The initial version of RO refractivity
assimilation system using DART/CAM is working and
the RO data has promise to reduce the errors in
global forecast of CAM.
20
What to do next

• Further look at RO datas impacts on temperature
  and moisture analyses by synthetic data
  assimilation. And assess the ability of ensemble
  filter to better separate T and Q contribution
  from assimilation of RO refractivity.
• Examine Real RO datas impacts on global analyses
  with the improved forecast error covariance
  estimates from the ensemble filter.