Title: Anomaly Compensation and Cloud Clearing of AIRS Hyperspectral Data
1Anomaly Compensation and Cloud Clearing of AIRS
Hyperspectral Data
- Presented at IEEE GRSS
- Boston Section Meeting
- August 24, 2005
- Choongyeun (Chuck) Cho
2Overview I
- Problem statement
- Definition of anomaly
- Background
- Atmospheric InfraRed Sounder (AIRS), Advanced
Microwave Souding Unit (AMSU), and Humidity
Sounder for Brazil (HSB) instruments - Signal characterization and reduction of
artifacts - Principal component analysis (PCA) and its
variants - Artifacts in AIRS data
3Overview II
- Stochastic cloud-clearing
- Background and prior work
- Description of stochastic-clearing (SC) algorithm
- Validation of stochastic-clearing algorithm
- ECMWF
- Physical clearing
- Conclusion
- Summary / contributions
- Future work
4Where Are We?
- Problem statement
- Definition of anomaly
- Background
- Signal characterization and reduction of
artifacts - Stochastic cloud-clearing
- Validation of SC algorithm
- Conclusion
5Problem Statement
- What is hyperspectral data?
- Hundreds or thousands of contiguous channels
- What is anomaly?
- Defined as an unwanted spatial or spectral
signature, statistically distinct from its
surrounding - Given X and a priori information about ?, what is
best estimate of ? or X? - A priori info about anomaly can have different
forms - Spectral statistical description, or usually
ensembles - Spatial structure or texture
- Joint spatial/spectral description
6Examples of Anomalies
- Anomalies of AIRS data discussed in this talk
- Instrumental noise
- Noisy channels
- AIRS data exhibits consistently noisy channels
- Scan-line miscalibration
- Resulting in striping patterns
- Cloud contamination
- Clouds generally make IR observations colder
- Compensation for cloud impact is critical for
accurate retrieval
7Where Are We?
- Problem statement
- Background
- AIRS/AMSU/HSB instruments
- Signal characterization and reduction of
artifacts - Stochastic cloud-clearing
- Validation of SC algorithm
- Conclusion
8AIRS/AMSU/HSB Instruments on Aqua
AIRS/HSB
- Atmospheric InfraRed Sounder (AIRS)
- 2378-channel infrared spectrometer covering
3.7-15.4µm - 1.1 FOV (13.5km at nadir)
- Advanced Microwave Sounding Unit (AMSU)
- 15 microwave channels
- 3.3 FOV (40.5km at nadir)
- Humidity Sounder for Brazil (HSB)
- 1.1 FOV (13.5km at nadir, same as AIRS)
- 4 microwave channels (150,190 GHz)
- Scan motor failure since Feb 2003
AMSU golfball
9Sample AIRS Brightness Temperatures
AIRS sample spectra for 3-by-3 FOVs
AIRS sample image
Data August 21, 2003 near Great Lakes
10Where Are We?
- Problem statement
- Background
- Signal characterization and reduction of
artifacts - Principal component analysis (PCA) and its
variants - Artifacts in AIRS data
- Stochastic cloud-clearing
- Validation of SC algorithm
- Conclusion
11Principal Component Analysis
- Useful to characterize multivariate signal, and
reduce dimensionality - Can be defined recursively (x m-dim multivariate
signal) - Solution y WTx where W w1w2wn, wi are
eigenvectors of CXX which is often estimated - To reduce dimension, n ltlt m
- The resulting reconstructed signal
(PC filter)
12Noise-Adjusted Principal Component
- Principal components are sensitive to arbitrary
scaling - Noise-adjusted PC (NAPC)
- Normalize data before applying PCA
- Guarantees maximum SNR
- What if noise variances are not known?
- Need to be estimated using blind signal
separation (BSS) technique such as Iterative
Order and Noise (ION) estimation
13Noise-Adjusted Principal Component
- Typical variances versus PC index plot
(truncated at 400) - NAPC shows sharper break than PC
- 6 NAPCs explain 99.8 of variances
Cumulative explained variance
Scree plot for AIRS data
Variance
14Artifacts in AIRS Data
- Measurements from a sensor can have different
sources of artifacts - AIRS data has different types of unwanted
artifacts - Instrumental noise
- Noisy channels
- Scan-line miscalibration
- Each dot in the image is
15Instrumental Noise
- Instrumental white noise is unavoidable
- NAPC filtering provides adaptive noise filtering
- NAPC filtering is extensively used in our
stochastic clearing algorithm and noisy channel
detector.
Before
After
Histogram of difference
16Noisy Channels I
- Some channels are consistently noisy
- These channels need to be excluded for further
analysis and retrieval of physical parameters - Noisy channel detection is done with NAPC
filtering such that a channel having 5s even once
is flagged noisy
Block diagram of noisy channel detector
17Noisy Channels II
- AIRS science team has their own list of bad
channels. - Detection rates for noisy and popping channels
Type of bad channel Number of NASA compiled bad channels Number detected by proposed algorithm detected
High noise 5 5 100
Detector response exhibits unexpected steps (Popping) 2 2 100
Total 7 7 100
18Scan-line Artifacts I
- Miscalibrated scan-line results
- in stripes
- A simple low-pass filter (in along-track
direction) in NAPC domain corrects this artifact
efficiently
Block diagram of removing-stripe algorithm
19Scan-line Artifacts II
Original Images
Processed Images
20Where Are We?
- Problem statement
- Background
- Signal characterization and reduction of
artifacts - Stochastic cloud-clearing
- Background and prior work
- Description of stochastic-clearing (SC) algorithm
- Validation of SC algorithm
- Conclusion
21Background and Prior Work
- What is cloud-clearing?
- Cloudy radiances (or TB) cause inaccurate
retrieval - Cloud-cleared radiances radiances which would
have been observed if FOV contains no clouds - Prior work on cloud-clearing
- Ignore cloudy FOVs only 5 of AIRS FOVs are
clear! - Physical cloud-clearing iterate between
estimation of physical parameters and calculation
of observed radiance - Adjacent-pair clearing use adjacent FOVs which
have different fractional cloud cover - Purely spatial processing restore 2-D
temperature field from sparse clear data
22Stochastic Clearing (SC)
- How does stochastic clearing (SC) work?
- SC estimates cloud contaminations solely based on
statistics without using any physical models - Hyperspectral measurements may contain sufficient
information about clouds in an obscured manner - Robust and stable training is necessary
- Nonlinearity is accommodated using stratification
(sea/land, latitude, day/night), multiplicative
scan angle correction, etc. - Advantages of SC approach
- Simple SC does not require physical models
(retrieval or radiative transfer). - Fast Based on matrix addition and multiplication
23Block Diagram of SC Algorithm
D-cloud 2 TBs
1 PC
1
Linear Operator A
Linear Operator B
?
1 PC
N
Select/average FOVs
3x3 AIRS TBs
Less cloudy
Cloudy Test
7
5 microwave ls Land fraction Secant q
More cloudy
?
Linear Operator D
Linear Operator C
Cleared AIRS TBs
N
N 314 channels
24Operators for SC Algorithm
ECMWF SARTA (v1.05)clear TBs
Trained with gt1000 golfballs
N AIRS TBs
Trained with gt1000 golfballs
Select avg FOVs
N
Operator A
Operator B
Find warmest among 9 pixels
Noise-Adjusted PCs
7
Operators C, D
E S T I M A T O R
L I N E A R
Noise- Adjusted PCs
3
N
Find coldestamong 9 pixels
AIRS D-cloud PCs
-
?TB
AMSU ch. 5,6,8,9,10
5
Warmest/coldest based on 11 4-mm channels
peaking 1-3 km Average 4 warmest pixels for
5-10 km WF, 9 for WF gt10 km
4
N
PC-1
?TBs
secant q
N
Land fraction
AIRS Cleared TBs
25Clearing Corrections vs ECMWF Ds
Weighting function peak 2.7 km, 2231.5 cm-1
-10 0 10 20 30 40
50 60 70oK
26AIRS-ECMWF for 827 Channels
Best 22 percent of golfballs (operator
C) Thresholds were 0.8K and 3K Nighttime ocean,
all scan angles Includes all 4- and 15-mm
channels plus one-fifth of the rest
314 good channels
314 good channels
Weighting function peak height (km)
27Where Are We?
- Problem statement
- Background
- Signal characterization and reduction of
artifacts - Stochastic cloud-clearing
- Validation of SC algorithm
- ECMWF
- Physical clearing
- Conclusion
28ECMWF Data Set Used
- ECMWF profiles are used to simulate cloud-cleared
TBs via SARTA v1.05 radiative transfer, for all
scan angles, 314 channels - Global, 3 days 8/21, 9/3, 10/12/03 (L1B v3.0.8)
- AIRS instrument noise reduced by averaging 1, 4,
or 9 of the warmest 15-km pixels for weighting
function peaks 0-5, 5-10, and gt10 km,
respectively - Used ?1000 golf balls to regress each of 10
categories based on day/night, land/sea, and
latlt40 or 40ltlatlt70
29RMS Difference (AIRS-ECMWF) for Ocean
- For best 28 sea latlt40º day (best for
sea)
30RMS Difference (AIRS-ECMWF) for Land
- For best 28 land latlt40º night (best for
land) -
31Cloud-cleared AIRS vs. ECMWF (best 28)
- RMS difference (oK) between cloud-cleared AIRS
and ECMWF/SARTA, 10 different estimators - RMS gt 0.7 K are boxed
- Excellent agreement for ocean and equatorial
regions - Degradation over daytime land near surface
Weighting function peak height (km) Ocean Latlt40 30ltLatlt70 Day Night Day Night Land Latlt40 30ltLatlt70 Day Night All Day Night All
0 1 1 2 4 5 6 7 10 11
0.38 0.4 0.86 0.91 1.68 0.77 1.36 1.48 0.78 1.19 0
.27 0.29 0.54 0.57 0.94 0.38 0.75 0.84 0.44 0.70 0
.28 0.30 0.45 0.45 0.34 0.29 0.33 0.41 0.33 0.39 0
.23 0.27 0.34 0.36 0.25 0.24 0.28 0.34 0.26 0.31 0
.24 0.27 0.33 0.35 0.23 0.25 0.26 0.24 0.28 0.27
32Global Cloud-Clearing Images (2392.1cm-1)
- SC applied to 8/21/2003 descending orbits (L1B
v4.0.9) - 2392.1cm-1, WF peak0.23 km
Observed AIRS
Cloud-cleared (best 78)
33Global Cloud-Clearing Images (2392.1cm-1)
- CC Residual Spatially high-pass filtered
version of CC - WF peaks at 0.23 km
34Cloud-Clearing Images and Sea Surf. Temperature
- Angle-corrected TB images at window channels
- Clearing works well even if there is no hole
(clear FOV)
AIRS 2399.9cm-1 near SW Indian Ocean
35Validation with Physical Clearing Visible vs.
AIRS 8.15-?m Data
Visible Ch 3
1227.7 cm-1
270
265
260
255
250
Granule 91(9/6/02) (solar reflection)
Channel 1284(H2O)
36Determination of AIRS Cloud-Cleared Brightness
Temperature Baselines
Stochastic CC
Clear Mask
Fitted
Channel 1284 (H2O) (1227.7 cm-1) Granule 91
(9/6/02) Indian Ocean, LAT -26.3, LON 70.2
Clear mask determined from Visible Ch 3, Brown
being clear Polynomial fit 4th order in scan
angle, 3rd order downtrack
37Stochastic versus Physical Cloud-Cleared8.15-?m
Brightness Temperatures
-Visible Ch 3
Stochastic Clearing
Physical Clearing
Channel 1284 (H2O) (1227.7 cm-1) Granule 91
(9/6/02) Indian Ocean, LAT -26.3, LON 70.2 Filter
subtracts cloud-cleared baseline and convolves
with 3?3 boxcar
38Correlation Coefficients Visible (v3) vs.
Cloud-Cleared 8.15-?m AIRS TBs
AIRS channel 1284(1227.7 cm-1)(water
vapor)September 6, 2002
Lon -26.3/Lat 70.2 Indian Ocean
Lon -12.5/Lat -106.1 Eastern Southern Pacific
39Stochastic and Physics-BasedCloud-Cleared Window
Channel TB
Stochastic Clearing
Physics-Based Clearing
Channel 2121 (2399 cm-1), WF peak 200m 9/6/02
North of Bermuda Note that the physical
clearing is most recent version (PGE v4.0.0),
calculated this week.
40Where Are We?
- Problem statement
- Background
- Signal characterization and reduction of
artifacts - Stochastic cloud-clearing
- Validation of SC algorithm
- Conclusion
- Summary / contributions
- Future work
41Summary
- Anomaly compensation techniques are discussed
based on signal processing techniques (NAPC and
ION) and nonlinear estimators - Anomalies Gaussian instrument noise, noisy
channels, scan-line miscalibration, and cloud
contamination - Stochastic clearing (SC) algorithm tested to be
successful using different validation schemes - ECMWF
- Sea surface temperature (SST)
- Physical algorithm
42Contributions Methodology
- We developed architected nonlinear estimators,
taking advantage of simplicity of linear
estimator and robustness of general nonlinear
estimator - NAPCs are used to reduce dimension and suppress
noise, making more robust and stable estimation - Prior knowledge (either spectral or spatial)
about anomaly is utilized to meaningfully
structure a nonlinear estimator
Architected Nonlinear Estimator
Linear Estimator
General Nonlinear Estimator
Dimension reduction
Prior info about anomaly
- May be unstable
- Complex/Slow
- Most powerful
- Stable
- Simple/Fast
- Less powerful
- Stable
- Simple/Fast
- Sufficiently powerful
43Contributions Stochastic Clearing
- SC algorithm developed based on anomaly
compensation and nonlinear estimation techniques - Enjoys excellent agreement with numerical weather
prediction model (ECMWF) - Performs superior to physical clearing
- Very fast depends on matrix addition and
multiplication consists of 664 lines of Matlab
code, 20 minutes to cloud clear an entire day of
AIRS data on ordinary PC - Clearing at extreme scan angles is good
hole-hunting using high spatial resolution may
not be essential for cloud-clearing
44Future Work
- Anomaly compensation theory
- Optimization of combining spectral and spatial
processing - More extensive spectral processing techniques
- SC algorithm
- Joint cloud-clearing and retrieval
- Better model for nonlinearities
- Optimum architecture for SC algorithm using
efficient design-of-experiment approach
45Where Are We?
- Problem statement
- Background
- Signal characterization and reduction of
artifacts - Stochastic cloud-clearing
- Validation of SC algorithm
- Conclusion
Thats all, folks. Any questions?
46Back-up Slides
47Global Cloud-Clearing Images (2392.1cm-1)
- Zoom-in images in Southeastern Pacific
- WF peaks at 0.23 km
48Determination of AIRS Cloud-Cleared Brightness
Temperature Baselines
Stochastic CC
Cloud Mask
Fitted
Channel 1284 (H2O) (1227.7 cm-1) Granule 91
(9/6/02) Indian Ocean, LAT -26.3, LON 70.2
Brown being clear (accepted) Polynomial fit
4th order in scan angle, 3rd order downtrack
49SC Validation w.r.t. SST
- NCEP sea surface temperature
50AIRS Stratified Stochastic Cloud Clearing
First-pass results Multiplicative
Scan-angle Stratified results
46 golfballs good 54 rejected
0.8K threshold
(AIRS cleared radiance) (ECMWF/SARTA) (oK)
46 golfballs good 54 rejected
3K threshold
Estimate of cloud-clearing radiance correction
(oK)
51AIRS Stratified Stochastic Cloud Clearing
First pass, all golfballs, all scan angles Second
pass, multiplicative scan angle Third pass, best
46 golfballs (all scan angle) Third pass, best
46 golfballs using ?lt16?
RMS cleared AIRS radiances vs. ECMWF/SARTA
0.5K
Altitude of weighting function peak (km)
52AMSU Contributions for Land
- For land 30ºltlatlt70º night
53Iterative Order and Noise (ION) Estimation
- To apply NAPC, the noise variances need to be
estimated - Signal model x Ap Gn, where A is a mixing
matrix, p is signal of unknown dimension k, G is
diagonal noise covariance matrix, n is
unit-variance white Gaussian noise. - ION iteratively estimates signal order (k),
mixing matrix (A) and noise variances (G) - Estimate order (k) using scree plot
- Estimate A and G using Expectation-Maximization
algorithm
54Blind Signal Separation
- Observation model x Ap G1/2w
- - A, p, G, w and k (dimension of p) are unknown.
- - A mixing matrix (nk), p source signal
vector (k1) - - G diagonal cov matrix (nn), w white
Gaussian noise vector (n1) - - Use matrix X, P and W to denote concatenated
samples of x, p and w - X AP G1/2 W
- Given X, how to estimate A, G and k.
- Previous signal separation techniques assume
either k or G is known - Iterative Order and Noise (ION) estimation
- - First, estimate signal order, k, using
eigenvector decomposition - - Estimate A and G using Expectation-Maximization
algorithm
55Blind Signal Separation II
X AP G1/2W
- Estimation of signal order, k
- - Selected based on scree plot, sorted
eigenvalue vs. - eigenvalue index
- Knee point separates
- signal from noise
56Blind Signal Separation III
X AP G1/2W
- Expectation-Maximization (EM) algorithm
iteratively finds maximum likelihood (ML)
estimate of parameters where model depends on
hidden (latent) variable - Expectation step estimate unobserved data (P and
PTP) using estimate of A and G - Maximization step compute ML estimate of A and G
using estimates of P and PTP
57ION Algorithm Block Diagram
X, optionally normalize rows to zero mean and
unit variance
Set imax, e.g. imax 10
Noise Normalization
Order Estimation
EM Algorithm
Scree Plot
Expectation
Maximization
SVD
Yes
iltimax
No
58Physics of Radiative Transfer
- Radiative transfer links environmental parameters
to hyperspectral data - For a black body, spectral brightness is defined
as - For microwave channels (hf ltlt KT) , this is
linear with temperature
59Physics of Radiative Transfer
- Upwelling radiation received at a sensor at
altitude L has four contributions - Can be rewritten as
60Physics of Radiative Transfer
AMSU Weighting functions
- Weighting functions for AIRS/AMSU
AIRS Weighting function peaks
61Physics of Radiative Transfer
- Microwave/IR atmospheric absorption spectrum
- MW Water vapor absorption lines at 22, 183, 325
GHz, O2 absorption at 118, 368 GHz, etc. - IR distinct water vapor, O3, CO2 absorptions
Microwave
Infrared