Examining effect of TPW-classified a priori error and quality control on
atmospheric temperature and water vapor sounding retrieval
Eun-Han
1School
1
Kwon ,
Jun
2
Li ,
2
Li ,
Jinlong
B. J.
1
Sohn ,
of Earth and Environmental Sciences, Seoul National University, Seoul, Korea
2CIMSS,
University of Wisconsin-Madison, WI, USA
Introduction
Results
priori first-guess error
a Classification
of variance
a priori errors by total precipitable water (TPW)
100
TPW (kg m-2)
All
0-10
10-20
20-30
30-40
40-50
50-
200
Pressure (hPa)
300
400
500
600
700
800
900
1000
0
2
4
6
8
0.0
0.1
Bias
100
AIRS Granule 178 on July 20, 2009
Water vapor (g/kg) retrieval result at 904.87 hPa
200
300
300
400
400
500
500
600
600
700
700
800
800
900
900
Reg
Phy1
Phy2
1000
-1.5
-1.0
-0.5
0.0
0.5
-3
1.0
-2
▪ Constrained retrieval methods find an optimal solution which is consistent with
both satellite-measured radiances and a priori information (first-guess and a priori
first-guess error)
Cost function: [Y m − F( X)] E −1 [Y m − F( X)]+ [X − X 0 ]T γS −a1 [X − X 0 ]
[
Iterative solution: X n +1 = X 0 + K n E K n + γS
T
−1
]
−1 −1
a
[
200
300
300
400
400
500
500
600
600
700
700
800
800
900
900
1000
1000
1
2
]
Regression retrieval
First guess profile x0
AIRS TB calculation
using RTM
Convergence?
200
3
4
0
1
Temperature (K)
X0: first-guess of atmospheric state
K: Jacobian (weighting function) matrix of F
E: measurement error covariance matrix Sa: a priori first-guess error covariance matrix
Update profile xn
No minimizing cost function
1
100
K n E −1 Y m − F( X n ) + K n (X n − X 0 )
T
0
Water vapor (g/kg)
RMSE
0
T
-1
2
3
4
Water vapor (g/kg)
 Quality control tests
Retrievals rejected by QCs
QC1 Non-convergence or bad physical retrieval
QC2 Large residual (> 1K)
QC3 High terrain (Ps < 750 hPa)
QC4 Desert (using IGBP surface type)
QC5 Large temperature difference between first-guess (Tg)
and physical retrieval (Tr)
|Tg – Tr| > 5K (at any level below 100 hPa)
QC6 Large moisture difference between first-guess (Qg) and
physical retrieval (Qr)
|Qg – Qr| / Qg > α
Non Conv/Bad Ret (QC1) & Large Resi (QC2)
100
200
200
400
500
600
100
200
200
300
300
400
400
500
Reg
Phy
Reg_QC (66.6%)
Phy_QC (66.6%)
700
800
800
900
900
1000
1000
2
4
6
8
10
0
12
100
100
200
200
Reg (all)
Phy (all)
Reg_QC3(1.2%)
Phy_QC3(1.2%)
Reg_QC4(6.0%)
Phy_QC4(6.0%)
400
500
600
600
600
900
900
700
700
1000
1000
800
800
900
900
0
1
2
3
4
200
200
300
300
400
400
500
500
Reg
Phy
Reg_QC (88.8%)
Phy_QC (88.8%)
600
700
700
800
800
900
900
1000
1000
3
4
Temperature RMSE (K)
8
0
2
4
6
8
Water vapor RMSE (g/kg)
100
100
200
200
300
Pressure (hPa)
100
2
6
Water vapor RMSE (g/kg)
100
1
4
Large T (QC5) and Q(QC6) diff. from first-guess
0
Temperature RMSE (K)
0
2
Temperature RMSE (K)
1000
600
10
500
600
4
8
400
800
3
6
300
800
2
4
High terrain (QC3) & Desert (QC4)
500
1
2
Water vapor RMSE (g/kg)
700
0
Ocean
600
700
1000
Pressure (hPa)
 Radiative transfer model: Stand-alone AIRS Radiative Transfer Algorithm (SARTA)
 Simultaneous retrieval of temperature, water vapor, ozone, surface skin temperature,
and surface emissivity (over land)
 Retrieved vertical levels: 101 pressure levels
 Vertical atmospheric structure and surface emissivity spectrum are represented by a
few EOF coefficients (15 for temperature, 6 for water vapor, 3 for ozone, and 6 for
emissivity). 101*3+2378+1 = 2682 → 31 unknowns
 The maximum number of successful iterations is set to be six.
 Exclude SW channels >2240 cm-1 during day.
 Atmospheric clear-sky sounding retrievals are performed from AIRS radiance
measurements by applying a physical iterative approach (Li et al., 2000) using the
regression retrieval (Weisz et al., 2007) as the first-guess.
-Domain: East Asia (0 – 55°N, 90 – 150°E) -Period: July 20 ~July 24, 2009
 Collocated ECMWF analysis data is used as a reference for the evaluation of AIRS
retrievals
Pressure (hPa)
Land
100
500
Temperature RMSE (K)
Pressure (hPa)
Retrievals accepted by QC1-6
400
700
300
Final retrieval
300
Reg (all)
Phy (all)
Phy_QC1(4.9%)
Reg_QC2(0.8%)
Phy_QC2(0.8%)
0
where, α varies depending on the magnitude of Qg
and the sign of Qg – Qr
Yes
100
300
Pressure (hPa)
▪ Mathematically ill-posed problem → need a priori constraint
0.6
100
200
100
Pressure (hPa)
Y = F (X) + σ
0.5
ln Q
Temperature (K)
m
0.4
▪ Phy1: Physical retrieval using a fixed error ▪ Phy2: Physical retrieval using TPW-classified errors
1000
-2.0
Ym: satellite radiance measurement
X: atmospheric state {T(p), Q(p), O3(p), Ts, ε(ν)}
F: forward model
σ: measurement error
0.3
▪ Calculated from 15704 atmospheric profiles and
corresponding simulated AIRS radiances using the
difference between first guess (regression retrieval) and
true profile
▪ The first-guess errors coming from the regression
retrieval depend on the atmospheric moistness.
▪ Applying the a priori first-guess error values classified
by TPW in the physical inverse procedure will result in
more weight to the first guess for moist cases and less
weight
to
the
first
guess
for
dry
cases.
0.7
 The effect of TPW-classification of a priori errors
Methodology & Data
“Inverse problem”
0.2
T
Pressure (hPa)
In the typical atmospheric profile retrieval methods, a priori error is used to
specify the ranges within which the retrieved atmospheric state departs from the
first-guess to match the satellite radiance observations. An incorrect error can
constrain the solution inappropriately, and thus it could result in unstable or largely
biased solution. Considering the importance of a priori errors in the inverse
problem, we suggest to use a priori errors classified by atmospheric moistness as an
alternative to a fixed a priori error in the temperature and water vapor profile
retrievals.
It is also valuable to assign appropriate quality control (QC) flags in the retrieval
process. Under some conditions, such as an incorrect specification of cloud,
measurement error, surface information, etc., the inverse of satellite radiance
measurements to the atmospheric state cannot be performed well and thus can
result in divergence of the solution or convergence to an unrealistic solution. Also,
poor performance of atmospheric temperature and moisture retrievals was reported
over high terrain and desert. Since those poor retrievals can have a substantial
negative impact on their applications, assigning QC flags is necessary so that users
can appropriately select good retrievals depending on their purpose.
This study investigates the use of dynamic a priori error information according
to atmospheric moistness and the use of quality controls in temperature and water
vapor profile retrievals from hyperspectral infrared (IR) sounders. In doing so,
temperature and water vapor profiles are retrieved from AIRS radiance
measurements using the retrieval algorithm which is a physical iterative approach
using the regression retrieval as the first-guess.
AIRS measured
TB
and Elisabeth
2
Weisz
0
1
2
3
Water vapor RMSE (g/kg)
4
Reg (all)
Phy (all)
Reg_QC5(6.4%)
Phy_QC5(6.4%)
Reg_QC6(16.2%)
Phy_QC6(16.2%)
400
500
600
700
300
400
500
600
700
800
800
900
900
1000
1000
0
2
4
6
Temperature RMSE (K)
8
0
2
4
6
Water vapor RMSE (g/kg)
Conclusions
A priori first-guess error variances for each TPW class are significantly different indicating that the first-guess errors coming from the regression retrieval depend on the
atmospheric moistness.
In the physical retrieval results from AIRS measurements over East Asia, the use of a priori first-guess errors classified by TPW, rather than a fixed error, appears to improve
water vapor retrievals in the boundary layer.
The six QCs tested here can be appropriately used to rule out the retrievals with large errors. The use of QCs for the temperature and water vapor soundings has more impact
over land than over ocean.
The use of dynamic a priori error information according to atmospheric moistness, and the use of appropriate QCs dealing with the geographical information and the deviation
from the first-guess as well as the conventional inverse performance are suggested to improve temperature and moisture retrievals and their applications.
8