AIRS
(Atmospheric Infrared Sounder)
Regression Retrieval (Level 2)
Level 0 to Level 2
Level 0: raw data
Level 1A: geolocated radiance in counts
Level 1B: calibrated radiance in physical units
Level 2: retrieved physical variables
(temperature, humidity and ozone profiles, surface skin
temperature, total precipitable water, total ozone content,
cloud top height . . .)
Regression Model
1. Regression Model
X = C YT
2. Least squares regression solution
C = X Y (YT Y)-1
Y…measurements [nprofs x nchannels]
C...Regression coefficients [nlevels x nchannels]
X... Atmospheric variables [nlevels x nprofs]
Regression Retrieval (1)
1. Calculate Regression Coefficients
C = Xtr Ytr (YtrT Ytr)-1
2. Perform Retrieval (RTV)
X = C YT
Y… Measurements [nprofs x nchannels]
C ... Regression coefficients [nlevels x nchannels]
X ... Atmospheric variables [nlevels x nprofs]
Subscript tr refers to trainingset
Regression Retrieval (2)
1. Calculate Regression Coefficients
C = Xtr Ytr (YtrT Ytr)-1
with
Xtr = Xtr – mean(Xtr)
Ytr = Ytr – mean(Ytr)
2. Perform Retrieval (RTV)
X = C YT or X = mean(Xtr) + C YT
with
X = X – mean(Xtr)
Y = Y – mean(Ytr)
Principal Components (PC) Regression Retrieval
1. Calculate Regression Coefficients
M = Cov(Ytr )
U = eig(M)
Atr = Ytr U
C = Xtr Atr (AtrT Atr)-1
2. Perform Retrieval (RTV)
X = mean(Xtr) + C AT
with
A= Y U,
Y = Y – mean(Ytr)
M… covariance matrix [nchannels x nchannels]
U ... First few eigenvectors of M [nchannels x npc]
A ... Projection Coefficients (or amplitudes) [nsamples x npc]
The Trainingset
Xtr . . . Representative set of atmospheric
variables including temperature, moisture, ozone,
surface pressure, surface skin temperature and
surface skin emissivities
Ytr . . . Corresponding set of simulated
radiances, calculated by a fast RT (radiative
transfer) forward model
Radiance received by AIRS
R   s   s  B (TS )
0
  B (T ( p )) d  ( p )
ps
0
 Upwelling IR radiation from surface
 Upwelling IR radiation from atm. layers
  s  rs   B (T ( p )) d  ( p )
 Reflected downwelling IR radiation
sun
 Rsun  cos( )  ssun
(
p
)

r

s

 Reflected solar radiation
ps
R…radiance, …wavenumber, s…surface, p…pressure, sun…solar,
T…temperature, B…Planck function,  …emissivity,
…level to space transmittance, ...local solar zenith angle
r…reflectivity, with r = (1- )/,
*…level to surface (downwelling) transmittance [*= 2(ps)/ (p)]
Fast Radiative Transfer Forward Model
• Fast Model Regression :
- Computation of line-by-line Transmittance  for FM training data set
- Convolve with AIRS SRF (spectral response function)
- Solve regression scheme  = AC for coefficients C using predictors A
(predictors are functions of T, p, absorber amount, scanang …)
• Calculate transmittance  for any other profile
• Solve RTE to get radiance R
IMAPP AIRS Regression Retrieval Results:
Comparison with co-located
radiosonde observations (RAOBs)
Co-located RAOB / AIRS single profile retrieval (RTV)
residual = observed minus calculated spectrum
RAOB – RTV
RAOB Residual (Obs – Calc BT)
Retrieval Residual (Obs – Calc BT)
Temperature (pixel 1259)
Brightness Temperature (BT) residual
Co-located RAOB / AIRS single profile retrieval (RTV)
residual = observed minus calculated spectrum
RAOB – RTV
RAOB Residual (Obs – Calc BT)
Retrieval Residual (Obs – Calc BT)
Humidity (pixel 1516)
Brightness Temperature (BT) residual
Co-located RAOB / AIRS retrieval statistics (1899 profiles)
RMS of RAOB – RTV
STDEV of RAOB – RTV
Temperature
Humidity
Co-located RAOB / AIRS retrieval statistics (1899 profiles)
RMS of Residual (Obs – Calc BT)
Stdev of Residual (Obs – Calc BT)
Mean of Residual (Obs – Calc BT)
Radiosonde Observations
AIRS Retrieval
IMAPP AIRS Regression Retrieval Results:
Global (240 granules) retrievals and retrievals
over the CIMSS direct broadcast area
Ascending Grans
Descending Grans
Globally Retrieved Temperature at 850 mbar
Without Cloudmask
With Cloudmask
Ascending Grans
Descending Grans
Globally Retrieved Moisture at 850 mbar
Without Cloudmask
With Cloudmask
CIMSS Direct Broadcast area:
AIRS measurements (10-23-2003)
Brightness Temperature [K] at
1000 cm-1 (no cloudmask)
CIMSS Direct Broadcast area:
IMAPP AIRS Retrieval (10-23-2003)
Temperature [K] at 700 mbar
(no cloudmask)
Humidity [g/kg] at 700 mbar
(no cloudmask)
CIMSS Direct Broadcast area:
IMAPP AIRS Retrieval (10-23-2003)
Surface Skin Temperature [K]
(no cloudmask)
Total Precipitable Water [cm]
(no cloudmask)
CIMSS Direct Broadcast area:
IMAPP AIRS Retrieval (10-23-2003)
Ozone [ppmv] at 9.5 mbar
(no cloudmask)
Total Ozone [Dobson Units]
(no cloudmask)
CIMSS Direct Broadcast area:
IMAPP AIRS Retrieval (10-23-2003)
Surface Emissivity at 908 cm-1
(no cloudmask)
Surface Reflectivity at 2641 cm-1
(no cloudmask)
IMAPP AIRS Regression Retrieval Results:
Comparison with ECMWF analysis fields
ECMWF ANL
AIRS RTV
AIRS RTV vs. ECMWF Analysis:
Temperature at 700 mbar (G192, 09-02-2003)
Without Cloudmask
With Cloudmask
AIRS RTV vs. ECMWF Analysis:
Temperature at 500 mbar and at selected pixel
ECMWF ANL
AIRS RTV
ECMWF ANL
IMAPP RTV
Pixel 28/49
ECMWF – IMAPP RTV
AIRS RTV vs. ECMWF Analysis:
Humidity at 750 mbar and at selected pixel
ECMWF ANL
AIRS RTV
ECMWF ANL
IMAPP RTV
Pixel 78/53
ECMWF – IMAPP RTV
AIRS RTV vs. ECMWF Analysis:
Humidity along scanline 65 (without cloudmask)
Scanline 65
IMAPP AIRS Retrieval
ECMWF Analysis
AIRS RTV vs. ECMWF Analysis:
Humidity along scanline 65 (with cloudmask)
Scanline 65
IMAPP AIRS Retrieval
ECMWF Analysis
RMS and STDEV of ECMWF minus AIRS RTV
(G192, 09-02-2003, 4758 clear pixels)
Root Mean Square (RMS) Error
Standard Deviation
Surface Skin Temperature [K]
Total Precipitable Water [cm]
Temperature
Humidity
AIRS RTV vs. ECMWF Analysis:
Spatial mean of Brightness Temperature (BT) residual
RMS of Residual (Obs – Calc BT)
Stdev of Residual (Obs – Calc BT)
Mean of Residual (Obs – Calc BT)
IMAPP AIRS Retrieval
ECMWF Analysis
ECMWF Analysis
AIRS RTV
AIRS RTV vs. ECMWF Analysis:
Spectral mean of Brightness Temperature (BT) residual
Mean Long Wave
Mean Mid Wave
Mean Short Wave
IMAPP AIRS Regression Retrieval Results:
Comparison with ECMWF analysis fields
and L2 operational product
AIRS RTV vs. ECMWF Analysis vs. Operational Product:
Temperature at 500 mbar (G192, 04-04-2004)
IMAPP AIRS
RTV
ECMWF ANL
Operational AIRS RTV
• available on AMSU footprint
(=3x3 AIRS FOVs) only
• missing areas  retrieval
not successful or not
validated yet
AIRS RTV vs. ECMWF Analysis vs. Operational Product:
Temperature at selected pixel
ECMWF ANL
IMAPP RTV
OP RTV
Pixel 22/6
ECMWF – IMAPP RTV
ECMWF – OP RTV
AIRS RTV vs. ECMWF Analysis vs. Operational Product:
Humidity at 600 mbar (G192, 04-04-2004)
IMAPP AIRS
RTV
ECMWF ANL
Operational AIRS RTV
• available on AMSU footprint
(=3x3 AIRS FOVs) only
• missing areas  retrieval
not successful or not
validated yet
AIRS RTV vs. ECMWF Analysis vs. Operational Product:
Humidity at selected pixel
ECMWF ANL
IMAPP RTV
OP RTV
Pixel 14/15
ECMWF – IMAPP RTV
ECMWF – OP RTV
IMAPP AIRS Regression Retrieval Results:
Comparison with MODIS and GOES retrievals
Humidity @620
Temperature @620
MODIS RTV vs. AIRS RTV vs. GOES RTV:
Temperature and Humidity at 620 mbar (G192, 09-02-2003)
MODIS RTV
AIRS RTV
GOES RTV
Improved Profile and Cloud Top Height
Retrieval by Using Dual Regression
Elisabeth Weisz & others
Accurate atmospheric sounding and cloud variability observations
provide an understanding of atmospheric dynamic and cloud-radiation
processes that need to be included in climate models. Furthermore
they are critical for initializing storm scale models for the prediction of
severe weather events.
This requires an algorithm which
•fully utilizes the information about the atmospheric state inherent in
single field-of-view hyperspectral radiances (e.g. AIRS, IASI, CrIS)
•is fast to be useful for real-time applications
•provides T, H2O, O3 profiles, as well as surface temperature, surface
emissivity, cloud top pressure, cloud optical thickness, CO2
concentration simultaneously under all weather conditions at singleFOV resolution
•accounts for the non-linearity between radiances and cloud
parameters and atmospheric moisture
•accounts for the annual change of CO2
Dual Regression Retrieval Algorithm – Part 1
CLEAR
CLOUDY
Global cloudy training data set
19948 profiles
Global training data set (SeeBor V5)
15704 profiles
Clear radiance calculations
Cloudy radiance calculations
SARTA v1.7 (UMBC)
Fast RT cloud model (Wei, Yang)
BT, Scanning angle and
CO2 Classification
Additional Predictors
(spres, solzen)
EOF regression
T -1
C = dXA (AA )
T
Clear Regression Coefficients
Scanning angle, Cloud height and
CO2 Classification
X…atmospheric state, C… regression coefficients,
A…compressed radiances (projection coefficients)
Cloudy Regression Coefficients
Regression RTV
X = Xtr + CAobs
T, H2O, O3, Tskin, Emissivity, CO2
at single FOV
Clear trained EOF regression results
T, H2O, O3, Tskin, CO2, CTOP, COT
at single FOV
Cloud trained EOF regression results
•Cloud Height classification (for the cloudy set only):
8 overlapping classes between 100 and 1000 hPa (200 hPa range
each) i.e. 100-300 hPa, 200-400 hPa, ….800-1000 hPa.
CTOP is determined iteratively until the optimal cloud category is
selected.
•CO2 classification: 5 sets defined by overlapping periods
from 2002 to 2013. Annual mean values from
ftp:://ftp.cmdl.noaa.gov/ccg/co2/trends/co2_annmean_gl.txt
or via 370+2*(year-2000).
years
Ann mean CO2
[ppmv]
1
2002-2005
375.7
2
2004-2007
379.8
3
2006-2009
383.7
4
2008-2011
387.9
5
2010-2013
392.6
Dual Regression Retrieval Algorithm – Part 1
CLEAR
CLOUDY
Global cloudy training data set
19948 profiles
Global training data set (SeeBor V5)
15704 profiles
Clear radiance calculations
Cloudy radiance calculations
SARTA v1.7 (UMBC)
Fast RT cloud model (Wei, Yang)
BT, Scanning angle and
CO2 Classification
Additional Predictors
(spres, solzen)
EOF regression
T -1
C = dXA (AA )
T
Clear Regression Coefficients
Scanning angle, Cloud height and
CO2 Classification
X…atmospheric state, C… regression coefficients,
A…compressed radiances (projection coefficients)
Cloudy Regression Coefficients
Regression RTV
X = Xtr + CAobs
T, H2O, O3, Tskin, Emissivity, CO2
at single FOV
Clear trained EOF regression results
T, H2O, O3, Tskin, CO2, CTOP, COT
at single FOV
Cloud trained EOF regression results
Dual Regression Retrieval Algorithm – Part 2
General Idea
Clear trained EOF regression results
Cloud trained EOF regression results
Cloud Top Altitude
Level where Tcloudy-Tclear > 3 K
at all levels below that level
Final Profile
(clear-trained above,
cloud-trained below cloud level)
Dual Regression Retrieval Algorithm – Part 2
Some details
1. A retrieval (incl. CO2 amount) is generated by using the regression set associated with
current year. The CO2 classification is refined if necessary and the retrieval is run again.
2. A first estimate of the CTOP from the clear-trained and cloudy-trained retrieved T profile is
obtained. The cloud classification is refined if necessary (i.e. if T derived cloud class differs
from cloud-trained solution).
3. If CTOP is close to the surface and the maximum difference between clear and cloudy
retrieved Tskin<2.5K the clear-trained solution is used as the as the final at all levels,
otherwise the cloud-trained profiles is used.
4. If CTOP≤300mbar the cloud-trained profile is used as the final at all levels.
5. A second (upper level cirrus) cloud top pressure is found if relative humidity exceeds 70%
at a level above the cloud altitude derived by the T profile.
6. If maximum difference between clear-trained and cloud-trained T retrieval exceeds 25K the
final retrieval below the cloud top is set to missing.
7. Effective cloud emissivity, i.e. (Tskin-Tskin_gdas)/(Tskin_ctop-Tskin_gdas), the cloud condition, and the
vertical mean and standard deviation of differences between final T retrieval and GDAS T
above and below the cloud top are used to assign quality flags to the final retrieval product.
Retrieval error statistics (simulated)
Unstratified (orig) vs. stratified (DR)
SurfSkinTemp [K]
CTOP [mbar]
COTH
#/Bias/Stdev/RMSE
(orig)
16359/0.45/2.21/2.26
16009/-25.9/177.3/79.2
16009/1.6/2.9/3.3
#/Bias/Stdev/RMSE
(DR)
16359/-0.07/1.57/1.57
16009/1.1/156.1/156.1
16009/-0.1/1.7/1.7
CTH comparison with CloudSat and Calipso
tropical
polar
T and H2O Error Statistics
(comparison with radiosondes over CONUS, Dec 2009)
Hurricane Isabel (Sept-13, 2003)
MODIS 1km images (1705, 1710)
Eye
High Dense Cirrus
Cirrus Outflow
AIRS Temperature (at 850 hPa) and Humidity (at 700 hPa)
.
.
AIRS Temperature and Humidity (at 300 hPa)
Isabel Eye Sounding
Location: 22.6N, 62W
Pressure [mbar]
Pressure [mbar]
Location: 22.6N, 62W
Temperature [K]
Relative Humidity [%]
Joplin, MO Tornado
(May 22, 2011)
MODIS – 19:46 UTC 22 May 2011
Joplin Tornado
 Joplin
 Joplin
 Joplin
 Joplin
 Joplin
GOES Visible 22:30 UTC
IASI (at 15:40 UTC) and AIRS (at 19:45 UTC)
H2O and Temperature Change
Temp 500 hPa
IASI (153855Z)
May 22, 2011
1540 UTC
AIRS (G197)
May 22, 2011
1945 UTC
RH 700 hPa
RH 850 hPa
☐ Joplin
☐ Joplin
☐ Joplin
☐ Joplin
☐ Joplin
☐ Joplin
5-km Temp.
Decreased
3.0-km RH
Decreased
1.5-km RH
Increased
IASI and AIRS Temperature, Dew Point Temperature and
Humidity at Joplin, MO.
Solid lines: temperature
Dashed lines: dew point temperature
Lat/Lon: 37.0/-94.6
Lat/Lon: 37.1/-94.5
Temp
Dew Point
Temperature [K]
IASI (1540 UTC)
Temperature [K]
AIRS (1945 UTC)
AIRS at 19:40 vs. RAOB at 18:00 for Springfield
(75 mi east of Joplin, MO)
Pressure [mbar]
Solid lines: temperature
Dashed lines: dew point temperature
Temperature [K]
IASI/AIRS Stability Change (Total Totals)
Improved Profile and Cloud Top Height
Retrieval by Using Dual Regression
Summary
•
The cloud height stratified dual (EOF) regression
retrieval provides all-sky condition vertical temperature
and moisture profiles with profiles being retrieved down
to cloud level and below thin and/or scattered to broken
clouds.
•
The single FOV algorithm is capable to capture the
spatial detail of the atmospheric state which will benefit
various applications, in particular severe weather
forecasting.
•
Algorithm will be integrated in the UW/CIMSS IMAPP
AIRS software package. Retrieval products will include
TPW, TOC, Lifted Index and CAPE.