Channel selection for IASI
in clear-sky conditions
Florence Rabier and Nadia Fourrié
Météo-France
ITSC-XII February 2002
Rationale and objectives
 Advanced IR sounders
– large volume of data, prohibitive in operational
NWP
 Selection of individual channels
– Which channel selection method leads to the
best analysis accuracy?
– In practice, how can this method be applied
robustly to a large set of atmospheric
conditions?
 IASI = Infrared Atmospheric Sounding
Interferometer developed by CNES-EUMETSAT
Linear estimation theory
 Retrieval
 Covariance matrix
 Gain matrix
 Data Resolution Matrix
 Model Resolution Matrix
 Metric based Jacobian matrix
 Degree of freedom for signal
 Shannon entropy reduction
xa= xb + K(y-Hxb)
A-1=B-1+HTR-1H
K= A HTR-1
DRM=HK
MRM=KH
H’=R-1/2HB1/2
DFS = Tr (I-AB-1)
ER = -1/2 log2 |AB-1|
Linear estimation theory
 Resolution matrices
• xa- xb = K(y-Hxb) =KH (x- xb) = MRM (x- xb)
• ya- yb = H (xa- xb)=HK (y- yb) = DRM (y- yb)
– Link the analysis and the signal from the data
 Diagnostics of retrieval accuracy
• Standard-deviations of analysis errors
a(i)
• Vertical resolution (Purser and Huang)
Resol(i) = dz (i) /MRM(i,i)
Channel selection methods
 Methods based on the DRM
(Menke, Prunet)
 Method based on Jacobians
(Goldberg, Aires)
– Equation ya- yb = DRM (y- yb)
– Select the most useful data in the analysis
– Characteristics of H’=R-1/2HB1/2
– For each parameter to be retrieved, select the most
useful channel
 Iterative method
(Rodgers)
– Measures of improvement ER or DFS (AB-1)
– Iteratively, pick up the most useful channel to improve
on the current analysis. Update the analysis errors.
Methods based on the DRM
(Menke)
• Data resolution matrix: DRM=HK
• From ya-yb=DRM (y-yb), the diagonal
elements of DRM indicate how much weight
a datum has in its own analysis
• These diagonal elements measure the
« importance » of the various channels
• The method needs the computation of A
Methods based on the DRM
(Prunet)
• SVD of H, with metrics B and R
• G= R-1/2HB1/2 =UΛVT
• Truncation in Λ2 such that eigenvalues of
GTG= B1/2HT R-1HB1/2 , equivalent to σb2/ σo2
represent 10% of contribution of the
observations to the analysis
• G= R-1/2HB1/2 =>UpΛpVpT
• DRM = VpVpT . Its diagonal elements are
used as channel « importance »
Method based on the
Jacobians (Goldberg, Aires)
• Is it based on the shape of the weighting functions
• Normalisation of H: R-1/2HB1/2
• For each retrieved parameter, at each level in the
vertical, one selects channels
 Among those peaking next to the level
 With the largest ratio:
Amplitude of the peak/Width of the weighting function
Iterative Method (Rodgers)
• This method is a step by step selection scheme. At
each step, Bi=Ai-1 is updated by using the most
informative channel among those which have been
previously selected.
• After normalisation of the Jacobian by R
 Ai-1=Bi-1+hTh
 Where B0=B and h is a line of H
• The selection criterion is either DFS or ER
 DFS(h)i=Tr(I-ABi-1)=hTBih/(1+ hTBih )
 ER (h)i=-1/2 log2det(ABi-1)=1/2 log2(1+ hTBih )
Experimental context
 500 atmospheric situations
– Profiles (T,Q), various sites and dates
 IASI data simulated with RTIASI
(Matricardi and Saunders)
– 8461 radiances (645 cm-1 – 2760 cm-1)
 B based on a 60-level ECMWF matrix
 O from CNES, F=0.2K
 Removal of bands sensitive to trace gases
– (700-720, 1000-1080, 1267-1312, 2092-2355 cm-1)
Results on mid-lat profiles
 24 atmospheric situations
– Profiles (T,Q), one site at various dates
 4 channel selection methods tested
– For each profile, optimal selection performed
Profile i
Selection i
Retrieval i =
Profile i
+ Selection i
– Results averaged over all profiles
Results on mid-lat profiles
First Channels for T
Iterative method
First Channels for T
Iterative method
Non-optimal set of channels
Iterative method
 For a set of profiles, optimal selection performed
Selection i
Profile i
 « Constant » selection obtained by averaging the ranks of
the channels
Cst selection =Ave (Selection i)
 Non-optimal retrievals
Profile j
Retrieval j =
Profile j
+ Cst Selection
 Would allow to pre-compute a constant selection off-line,
and to apply it to new profiles in real time
« Constant » selection
Iterative method
(300 channels)
(492 profiles)
Results for analysis errors
Iterative and Jacobian methods
(300 channels, 492 profiles)
Results for analysis vert resolution
Iterative and Jacobian methods
(300 channels, 492 profiles)
Influence of number of channels
Results for analysis errors
Iterative method (24 profiles)
Influence of number of channels
Results for analysis vert resolution
Iterative method (24 profiles)
Conclusions
 Iterative method
– Among 4 channel selection methods tested, the
iterative method is giving the best results
 Main strength
– Update the error covariance matrix each time a channel
is selected
 Constant selection gives promising results
– Pre-selection based on a set of profiles, then applied to
all profiles
– Robustness: selection performed for 62 profiles out of
492 gave 84% of channels in common with the one
computed on all 492 profiles
Perspectives
 Method can be applied to other sounders
– Thépaut and Fourrié
 Study to be extended
– Inclusion of different scan angles, surface types, cloud
conditions
 Possible operational channel selection
– Pre-selection based on monitoring statistics
– Use several sets of channels for various configurations
of scan angles, surface types, cloud conditions and also
air-mass