AIRS Radiance and
Geophysical Products:
Methodology and Validation
Mitch Goldberg , Larry McMillin
NOAA/NESDIS
Walter Wolf, Lihang Zhou, Yanni Qu and M.
Divakarla
Science Activities









Data compression.
Validate and improve radiative transfer calculations.
Cloud detection and clearing.
Cloud products
Channel selection (super channels).
Validate and improve retrieval algorithms.
Trace gases
Surface emissivity
Use MODIS to improve AIRS cloud detection and cloud
clearing
 Radiance bias adjustments
 Forecast impact studies
TOPICS
 Use of principal components (a.k.a.
eigenvectors) for data compression.
 Surface emissivity
 Cloud detection
AIRS Geophysical Products
 Microwave-only retrieval of sfc emissivity, sfc
temperature, sfc type and profiles of temperature,
water vapor and cloud liquid water.
 AIRS retrieval of cloud amount and height, sfc
emissivity, sfc temperature, and profiles of
temperature, water vapor and ozone.
 AIRS has two retrieval steps – very fast
eigenvector regression followed by a physical
retrieval algorithm.
Data Compression
 Advanced IR sounder data are very large compared with current
sounders (1 orbit ~ 2GB vs. 8 MB) Much larger for GIFTS.
 Information is not independent. Principal component analysis
(PCA) is often used to reduce data vectors with many
components to a different set of data vectors with much fewer
components that still retains most of the variability and
information of the original data
 Data are rotated onto a new set of axes, such that the first few
axes have the most explained variance.
 Principal component scores are provided instead of the
individual channels.
 Individual channels can be reconstructed with minimal signal
loss with added benefit of noise reduction.
Generating AIRS eigenvectors
 Collect an ensemble of AIRS spectra (2378
channels).
 The radiances are normalized by expected
instrumental noise (signal to noise)
 Compute the covariance matrix S
 Compute the eigenvectors E and eigenvalues 
S = E ET
 E = matrix of orthonormal eigenvectors (2378x2378)
 = vector of eigenvalues (explained variance)
Training Ensemble
 Eigenvectors are generated from a spatial
subset of AIRS data (200 mbytes vs 30 GB full
data)
 Eigenvectors are generated daily.
 A static set of eigenvectors is used, but the
ensemble is occasionally updated with new
structures.
 When the ensemble is updated a new set of
eigenvectors is also updated.
Locations used in generating eigenvectors
Applying AIRS eigenvectors
 On independent data – compute principal component
scores.

P
=
ET R
; elements of R = (ri- ri ) /ni
 Invert equation and compute reconstructed
radiances R*.

R*
=
EP
 Reconstructed radiances are used for quality control.
 Reconstruction score = [ 1/N 3(R*i - Ri)2 ]1/2
i = 1 ….N channels
Square root of the eigenvalues
1 7497.60
2 1670.40
3 945.52
4 496.01
5 284.01
6 266.30
7 156.95
8 139.67
9 88.27
10 72.83
11 60.03
12 53.42
13 45.01
14 39.72
15 34.54
16 26.57
17 22.62
18 17.60
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
14.68
13.49
12.28
11.32
10.70
9.08
8.24
7.85
6.77
5.98
5.83
5.39
5.34
4.98
4.34
4.09
3.62
3.48
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
3.38
3.11
2.82
2.53
2.41
2.39
2.34
2.24
2.03
1.86
1.78
1.71
1.65
1.61
1.54
1.52
1.35
1.34
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
1.25
1.19
1.16
1.15
1.09
1.05
1.02
0.98
0.90
0.86
0.81
0.80
0.78
0.77
0.73
0.72
0.70
0.66
Reconstruction score = [ 1/N 3(R*i - Ri)2 ]1/2
i = 1 ….N channels
Reconstruction score = [ 1/N 3(R*i - Ri)2 ]1/2
i = 1 ….N channels
Monitoring Eigenvectors
 Monitoring eigenvectors is critical
 Eigenvectors may need to be updated due to
new structures that were not in the original
ensemble
12/4/00 reconstruction scores
Monitoring reconstruction score is important
July Aug
Sep
Oct
Nov
Days
Dec
Jan Feb
Noise
Noise free
75 PCS
Noise Reduction
Observed vs noise-free
reconstructed vs noise-free.
“ Observed”
Reconstructed
Observed vs. Reconstructed
New Plan
 Generate full spatial resolution AIRS principal
component score datasets
 Size ~ 5 MB instead of 150 MB per six
minute granule
Surface emissivity
Retrieval error based on 18 channels
Background Std dev.
Retrieval error
Clear detection
BACKGROUND
 NWP centers will assimilate clear radiances
 Need very good cloud detection algorithm
 Very important for radiance validation and to
initiate the testing of the level 2 retrieval code.
Cloud Detection over Ocean
 Use VIS/NIR channels during
day.
 Compare SST with 2616 cm-1
at Night.
 Predicting SST from 11 and 8
micron channels (works for day
and night)
 Predict 2616 from 8 micron
channels (night)
 11 micron window > 270 K
ONLY 0.5% residual clouds
Cloud detection – Non Sea
 Predict AIRS channel at 2390.9
cm-1 from AMSU
 FOV is labeled “mostly clear” if
predicted AIRS – observed
AIRS < 2
AND IF
 SW LW IR window test is
successful:
[ch(2558.224)-CH(900.562)] <
10 K
 Variability of 2390.910 radiance
within 3x3 < 0.0026
Clear Detected Fovs
Cloud cleared cases
Future Work –
Merge MODIS and AIRS
 High spatial resolution will improve determination
of clear AIRS fovs.
 High spatial resolution will greatly improve clear
estimate needed for cloud clearing.
MODIS Sounder Radiance Product
 MODIS has HIRS-like sounder channels – but at
high spatial resolution (1 km).
 Find a few clear MODIS fovs in a 50 x 50 km area
should provide a yield of 80% -- similar to AMSU
Summary
 Busy getting ready for real AIRS data
 Simulating AIRS in real-time has provided a means to
develop , test and validate the delivery of products to NWP
centers,
 AND created a platform to develop scientific tools to
analyze the data and test algorithms.
 Early releases of the data should be available 3 months after
launch
 Final radiance products ~ 7 months
 Retrievals ~ 12 months
 First activity will be to examine biases between measured
and computed radiances and validation of the clear
detection algorithm.
 “Day-2” Utilize MODIS