A Variational Approach
to NWP Preprocessing
and Quality Control
K. Garrett1, S.-A. Boukabara2, Q. Liu3
18th International TOVS Study Conference
March 23, 2012
Toulouse, France
1. Riverside Technology, Inc 2. Joint Center for Satellite Data Assimilation 3. University of Maryland, College Park
Motivation
Motivation
 Assimilation of satellite data is fundamental to forecast initialization/accuracy
 Best representation of analysis fields requires:
 Knowledge of how the guess, conventional and satellite data can best be
utilized
 The types of scenes (sfc/atm) which have the ability of being assimilated
and ensuring unwanted observations are filtered out
?
 Satellite
radiance
observations contain information
to 6-hr
help Forecast
characterize the
GDAS
Analysis
GFS
surface and atmosphere before assimilation
 An assimilation radiance preprocessor can be utilized in this context
GDAS Analysis 23 GHz Emissivity and Tskin
2
Contents
Outline
1
The 1DVAR Algorithm
2
Preprocessor – QC/Geophysical Fields
3
Radiance Assimilation Impact
4
Summary
3
Overview of Overview
the 1DVAR Algorithm
 Physical algorithm (1DVAR) for microwave sensors (MiRS)
 MiRS applies to imagers, sounders, combination
 Cost to extend to new sensors greatly reduced
 MiRS uses the CRTM as forward operator (leverage)
 Applicable on all surfaces and in all-weather conditions
 Operational for N18, N19, Metop-A and F16/F18 SSMI/S
 On-going / Future:
 Extension operations to Metop-B, NPP/ATMS and Megha-Tropiques
(MADRAS and SAPHIR)
 Get ready for the JPSS and GPM sensors.
 Extend MiRS to Infrared Remote Sensing (CRTM is already valid)
4
1D-Variational
Retrieval/Assimilation
The 1DVAR
Algorithm
Measured Radiances
Vertical Integration and Post-Processing
MiRS Algorithm
Temp.
ProfileProfile
Temp.
TPW
Yes
Solution
RWP
Simulated Radiances
Contribution
Function
Within
Noise
Level
?
Humidity Profile
Reached
IWP
Vertical
-1
-1 -1 T
T
G
= (K Se K + S
CLW
Integration
a ) K Se
No
Liq.Humidity
Amount Prof Profile
Jacobians
1DVAR
Outputs
Initial State Vector
Comparison: Fit
Ice. Amount Prof
Update
State Vector
-Sea Ice Concentration
Measurement
-Snow
Water
Equivalent
Averaging Kernel
Rain Amount Prof
& RTM
-Snow Pack Properties
Uncertainty
-Land Moisture/Wetness
Post
ForwardEmissivity
Operator
Spectrum
New State
Vector
-Rain Rate
Matrix E
Processing
(CRTM)
-Snow Fall Rate
(Algorithms)
Emissivity
Spectrum
-Wind Speed/Vector
Skin Temperature
-Cloud Top
-Cloud Thickness
Geophysical
-Cloud phase
Core Products
A = (GK)
Mean
Background
Climatology (Retrieval Mode)
Geophysical
Covariance
Matrix B
5
Contents
1
The 1DVAR Algorithm
2
Preprocessor – QC/Geophysical Fields
3
Radiance Assimilation Impact
4
Summary
6
Preprocessor QC
 Convergence is reached everywhere: all surfaces, all weather
conditions including precipitating, icy conditions
 A radiometric solution (whole state vector) is found even when
precip/ice present. With CRTM physical constraints.
T −1  m
m
2


ϕ =  Y − Y (X ) ×E ×  Y − Y (X )




Previous version (1 attempt)
Current version (2 attempts)
(assume non-scattering atmosphere)
(assume scattering from precip)
7
Geophysical
Fields
Preprocessor-CLW
MiRS
Cloud Liquid Water
MiRS - GFS
GFS
8
Preprocessor-Emissivity
MiRS N18 retrieved emissivity at 31 GHz ascending node
Evolution
of emiss
before,
during
and after
rain
2010-10-20
2010-10-21
2010-10-22
CPC real-time 24-hour precipitation
CPC Figures courtesy http://www.cpc.necp.noaa.gov
9
Contents
1
The 1DVAR Algorithm
2
Preprocessor – QC/Geophysical Fields
3
Radiance Assimilation Impact
4
Summary
10
Radiance
Assimilation
Impact
Assimilation
Impact
Comparison of O-B using MiRS retrieved CLW
versus NWP CLW for cloud screening
~6000 pts after filtering out cloudy scenes
 1 day
 Ocean only
11
Assimilation Impact (O-B)
AMSU 23 GHz
AMSU 50 GHz
AMSU 31 GHz
AMSU 89 GHz
12
Assimilation Impact (O-B)
MHS 89 GHz
MHS 183 ±1 GHz
MHS 157 GHz
AMSU 183 ±3 GHz
13
Contents
1
The 1DVAR Algorithm
2
Preprocessor – QC/Geophysical Fields
4
Impact on Radiance Assimilation
5
Summary
14
Summary
 MiRS is a generic retrieval/assimilation system (N18, N19, Metop-A, DMSP
F16/18 SSMIS)
 In retrieval mode, MiRS can be used as an NWP (assimilation) preprocessor
 MiRS LWP has shown to produce reasonable O-B (improvement over using
guess fields)
 Future work will include investigating the use of other QC metrics and
retrieved fields to filter/parameterize DA
 For more detailed information about the MiRS project, visit:
mirs.nesdis.noaa.gov (more validation data, publication list and software
package)
15
BACKUP SLIDES
16
And…
ATMS Impact
Experiment
GDAS (Hybrid ENKF)
high resolution
ATMS experiment
Acknowledgements:
A. Collard
J. Jung
E. M. Devaliere
17
Mathematical Basis:
The 1DVAR
Algorithm
Cost Function
Minimization
Cost Function to Minimize:

1
  1Jacobians
T
−1
m
& TRadiance
J( X) =  (X − X 0 ) × B × (X − X 0 ) +  (Y − Y( X)) × E −1 × (Y mSimulation
− Y( X))
from Forward Operator: CRTM
2
 2

∂J( X) = J(' X) = 0
To find the optimal solution, solve for:
∂X


y( x) = y( x )+ K x − x 
Assuming Linearity
0
0 

This leads to iterative solution:

−1


Δ
= B−1+ KnTE−1Kn KnTE−1 Ym − Y( X n ) + KΔ
nX n 


X n+1






Δ
X
More efficient
(1 inversion)






−1



T K BK T + E   Ym − Y(
= BK n
 n


n
n+1 

 


X n ) + KΔ
n

X

n 
Preferred when nChan << nParams (MW)
18