Cloud Detection over Ice and Snow using MISR Data

1
Cloud Detection over Ice and Snow using MISR Data
Tao Shi Department of Statistics, U.C.
Berkeley Bin Yu Department of Statistics, U.C.
Berkeley Eugene Clothiaux Dept. of Meteorology,
Penn State Univ. Amy Braverman Jet Propulsion
Laboratory July 2004
2
Outline

• Introduction to MISR and MISR cloud detection
  algorithms.
• Linear Correlation Matching Classification
  algorithm (LCMC)
• Enhanced Linear Correlation Matching
  Classification algorithm (ELCMC)
• Application of Quadratic Discriminate Analysis
  (QDA) and Support Vector Machine (SVM) on the
  cloud detection problem.

3

• Multi-angle Imaging Spectre Radiometer (MISR) was
  Launched by NASA on December 18, 1999.
• It boarded TERRA platform on the EOS-AM1 orbit.
• Built and maintained for NASA by the Jet
  Propulsion Laboratory (JPL) in Pasadena,
  California.

4

• MISR has 9 Angles
• 0O( AN),
• 26.1O ( AF, AA),
• 45.6O ( BF, BA),
• 60O ( CF, CA),
• 70.5O ( DF, DA)
• 4 wavelengths in each angle. (443nm, 555nm,
  670nm, and 865nm Near Infrared Red)
• Spatial Resolution
• - 275 x 275 m2 in red bands
• and all bands in nadir angle
  
• - 1.1 x 1.1 km2 in other bands.

5
Clouds are registered at different ground
locations in different angles
Flight direction
Figure Orbit 12986, Block 52-54, Red channel
image showing from DF angle to DA angle.
6
The MISR algorithm retrieves the cloud height and
cloud movement by matching the same clouds in
three angles.
The radiance is registered on the earth surface
ellipsoid.
The algorithm works well over dark surfaces, such
as deep ocean and vegetation covered land
surface, but does not work well over snow and
ice covered surfaces.
7
Outline

• Introduction to MISR and MISR cloud detection
  algorithms.
• Linear Correlation Matching Classification
  algorithm (LCMC)
• Enhanced Linear Correlation Matching
  Classification algorithm (ELCMC)
• Application of Quadratic Discriminate Analysis
  (QDA) and Support Vector Machine (SVM) on the
  cloud detection problem.

8
Linear Correlation Matching Classification
Algorithm
Figure MISR data registration.

• Key ideas
• Detecting snow- and ice-covered surfaces, not
  clouds.
• For snow- and ice-covered surfaces, different
  angle radiances registered at the same ground
  location are highly linearly correlated.

9
Figure MISR red image (Green Land, June 20,
2001). Left AN. Right DF.
Correlations between angles are strong over
snow- and ice-covered surfaces, and weak in
areas covered by high cloud. LCMC algorithm
tests if the linear correlation is strong
between angles.
Correlation with AN
DF CF BF AF AA BA CA DA
Correlation of AN red radiance with red radiances
at other angles.
10
Problems with LCMC

• LCMC works well detecting high clouds.
• LCMC is computationally fast.

• Over high terrain correlations are not strong,
  because of mis-matching of angles.
• Over very smooth surfaces (e.g. frozen rivers)
  correlations between angles are weak because
  variations in radiances are mostly due to the
  instrument noise.
• Very low clouds are difficult to detect because
  they also exhibit high correlations between
  angles.

11
Outline

• Introduction to MISR and MISR cloud detection
  algorithms.
• Linear Correlation Matching Classification
  algorithm (LCMC)
• Enhanced Linear Correlation Matching
  Classification algorithm (ELCMC)
• Application of Quadratic Discriminate Analysis
  (QDA) and Support Vector Machine (SVM) on the
  cloud detection problem.

12
Remedies for the three problems of LCMC

• We now use terrain projected radiances to address
  the problem of the low correlation over the
  terrain.
• We use the standard deviation (SD) of the AN red
  radiances in a small window to measure smoothness
  of the surface.

Figure Upper AN RGB. Left bottom correlations
between AN angle and other cameras. Right bottom
scatter plot of AN red radiances in the window on
upper image.
13
3. We use forward scattering to detect the low
clouds clouds have stronger forward scattering
than ice- and snow-covered surfaces.
Figure Orbit 7898, path 26, blocks 15-20. Left
AN red. Right DF red.
14

• ELCMC uses 3 features based on 275m, terrain
  projected red radiances. Classification results
  are reported on 1.1km MISR grid.
• correlation between angles ( CORR
  (rAF-ANrBF-AN)/2 )
• 2. surface smoothness ( SDAN )
• the angular signature of the radiances from
  different
• angles Normalized Difference Angular
  Index (NDAI)
• see Nolin, Fetterer, and Scambos (2002)

15
Three dimensional plot of features
NDAI
Correlation
SD
Scatterplot no cloud , low cloud, high
cloud.
Df red image
Figure MISR orbit 7898 collected over Green Land
on June 22, 2001
16
ELCMC Algorithm

• To classify each 1.1km pixel, we use 275m red
  radiances in the 8 by 8 window centered at the
  pixel to be classified, and declare clear (snow
  and ice) when
• 1. SDAN lt thresholdsd or
• 2. CORR gt thresholdcorr and NDAI lt
  thresholdndai

17
Threshold choice
First, the thresholds are determined by
inspecting the histograms and comparing resulting
cloud masks from different thresholds .
thresholdcorr 0.8
thresholdsd 2
thresholdndai 0.2
After studying the histograms of data in 2002, we
proposed a method to automate the threshold
choice.
18
Automate threshold choice

• Thresholds of SD and CORR are fixed at 2 and 0.8
• Threshold on NDAI is chosen by fitting a mixture
  Gaussian density with two components and using
  the dip of the density as thresholds.

19
DF red image
ELCMC cloud mask
MISR SDCM mask
Figure Orbit 7898, blocks 15-20.
20
AN red image
DF red image
ELCMC cloud mask
Figure Orbit 7898, Block 21-26.
21
Outline

• Introduction to MISR and MISR cloud detection
  algorithms.
• Linear Correlation Matching Classification
  algorithm (LCMC)
• Enhanced Linear Correlation Matching
  Classification algorithm (ELCMC)
• Application of Quadratic Discriminate Analysis
  (QDA) and Support Vector Machine (SVM) on the
  cloud detection problem.

22
QDA and SVM classifier captures the curved
boundary
NDAI
Correlation
SD
Scatterplot no cloud , low cloud, high
cloud.
DF red image
Figure MISR orbit 7898 collected over Green Land
on June 22, 2001
The classification boundary is curved. We tested
QDA and Gaussian kernel SVMs to approximate the
curved boundary.
23
Training data and input features

• Expert labeled data
• applied on three physical features
• applied on all angle red radiances
• ELCMC results as training data
• applied on three physical features
• applied on all angle red radiances

24
QDA using expert labels
DF red image
Expert Labels
QDA on three features
QDA on radiances
25
QDA using ELCMC results
DF red image
ELCMC Results
QDA on three features
QDA on radiances
26
SVM using expert labels
DF red image
Expert Labels
SVM on three features
SVM on radiances
27
SVM using ELCMC results
DF red image
ELCMC Results
SVM on three features
SVM on radiances
28
Results From Another Data Set (July 17 2002)
DF red image
AN red image
Expert labels
ELCMC-QDA on features
ELCMC-QDA on radiances
ELCMC Results
29
Misclassification Rates of Tested Classifiers

• QDA and SVM provide non-trivial improvements
  over ELCMC.
• SVM is more robust than QDA when using the ELCMC
  results,
• but the computational cost of SVM is much
  higher than QDA.

30
Concluding Remarks

• The collaborations of geoscientists and
  statisticians are
• essential to the success of this research.
• ELCMC algorithm uses three physical features to
  detect
• snow- and ice-covered surfaces, and obtains
  the cloud
• mask by exclusion.
• ELCMC classified data can serve as the training
  data for
• QDA and SVM, which provide curved
  classification
• boundaries and a lower error rate than ELCMC.

31
Acknowledgements

• David Diner, Ralph Kahn, Roger Davies,
  Jan-Peter Muller, Larry Di Girolamo, Anne Nolin,
  Jia Zong, Dominic Mazzoni, Mike Garay, the
  project manager, and all the MISR science team
  members.

NSF, USA Miller Institute at UC Berkeley