Introduction to subpixel offset tracking

1
Introduction to sub-pixel offset tracking
Matt Pritchard Cornell
Eric Fielding JPL
What is pixel tracking? What can pixel tracking
be used for? How does pixel tracking compare
with InSAR? What special types of errors exist
when using optical data? How accurate and
precise are pixel tracking? How do I get started
using pixel tracking?
2
Example Measure temporal variations in glacier
velocities

• Who cares?
• Observed glacier acceleration
• Via satellite data, Helheim glacier Greenland
  observed to increase velocity by 50 between 2001
  and 2005 then return to normal in 2006 (Howat et
  al., 2007)
• Other glaciers observed
• to double their velocities
• Surprising result
• the cause is unknown
• Implications for sea
• level rise in 21st century
• Is it 30 cm? Or 200 cm?

ASTER data From Howat et al., 2005
3
What is pixel tracking?
Find a distinctive feature Determine the
distance it has moved between 2 different images
Reference image
Slave image
Algorithm uses a box of pixels in and is able to
detect motions that are a fraction of a pixel.
Resolution of ASTER and SAR scenes 15 m/pixel
Possible to see motion of cm/day
Reference window
Search area size
Distance between searches
Slave image
Reference image
4
Correlation Peaks
These figures show the correlation peak for
different offset values The sharper the peak the
more precise the match Some problems can occur
when the chip sizes are too small and the
amplitude pattern is repeated (e.g., on large
icesheets)
From Wuite, 2006
5
Pixel tracking flowchart
1) Select images How long of a temporal
separation? How large of a pointing
difference? 2) Orthorectification (for
optical, SAR georectification is step 5) Test
LPDAAC vs. manual process in ENVI 3)
Pre-filtering of images Test Principle
Component Analysis, High Pass Filter 4) Pixel
Tracking Algorithm Test COSI-CORR (Leprince et
al., 2007) and Ampcor (Rosen et al., 2004) 5)
Post-processing to remove unreliable values
6
What data can be used?
Satellite SPOT (commercial) Level 1A SPOT 1
(1986-1990, 10 m) SPOT 2 (1990-present, 10 m)
SPOT 3 (1993-1997, 10 m) SPOT 4 (1998-present,
10 m) SPOT 5 (2.5-10 m, 2002-present) QuickBird
(commercial) Need Level 1B (2001-present, 0.6 m,
DigitalGlobe) Worldview 1 (commercial)
(2007-present, 0.5 m, DigitalGlobe) Formosat 2
(commercial) (2004-present, 2 m, sold by
SpotImage) Landsat 7 (NASA) ETM has severe
sensor artifacts (e.g., Lu et al., 2006) ASTER
(NASA) Level 1A (3 VNIR bands, 15 m,
1999-present) -- available in product geocoded by
LPDAAC using DEM acquired on given satellite
overpass. The problem is that many of these DEMs
are corrupted SAR See slide from Fieldings
lecture Airborne USGS NAPP High resolution
images (0.25-1m) scanning starts this fall SAR
See slide from Fieldings lecture Under certain
circumstances, can combine different satellites
or aerial and satellite images
7
A confusing terminology!
Image registration Sub-pixel correlation of
images Imageodesy (Crippin Blom) Feature
tracking (also used in fluid dynamics, Moroni
Cenedese, 2006) Pixel tracking (also used for web
statistics) Speckle tracking (also used in
medicine -- ultrasound) Intensity tracking (also
used in spectroscopy and computer
science) Satellite Radar Feature Tracking
(SRFT) Offset tracking Pixel offsets Range/Azimuth
offsets Phase correlation method (refers to
Fourier method of McHugh, 1991) Coherence
tracking (a.k.a fringe visibility, Strozzi et
al., 2002 based on Derauw, 1999) Suggest new
name Subpixel Image Registration (SIR) either
Optical OpSIR or SARSIR(R or C) Particle Image
Velocimetry (Adrian, 1991, Ann. Rev. Fluid
Mech) Applications Medical imaging (compare
images for Fluid dynamics experiments Computer
vision pattern recognition (e.g., face
recognition and movement) Remote sensing
(satellite airborne) applications to oceans,
solid Earth, and atmosphere
8
Currently available software

• Several possible algorithms, but 2 seem to be
  most common for remote sensing
• Normalized cross correlation
• Imcorr (Scambos et al., 1992) Open source
  distributed by Bruce Rapp (NSIDC)
• Ampcor (Rosen et al., 2004) Open source part of
  ROI_PAC InSAR processing software
• Fournier shift theorem
• Cosi-Corr (LePrince et al., 2007) Open source
  software implemented in ENVI commercial package.
  Requires the scenes to be co-registered before
  tracking

9
Algorithms
Normalized cross-correlation (a.k.a. statistical
correlation) What is it? Take small region of
each image, take F.T. Correlation theorem says
F.T. of correlation is product of F.T. of one
image times complex conjugate of F.T. of other
sub-sample correlation surface to find peak pro
good for images with white noise con require
similar images and sensors, under variable
lighting and atmospheric conditions favor Fourier
methods Good only for images misaligned by small
rigid or affine transformation Fourier methods
(a.k.a. frequential correlation) What is it? Use
Fourier Shift Theorem take F.T. of 2 sub-images,
any displacement will be detectable by a phase
shift (in frequency domain) take inverse F.T.,
and a weighted average of several test subimages
to find correlation peak pro good for images
with frequency dependent noise (e.g., low
frequency) con images have to be co-registered
first
10
A (partial) history
1986-1991 Glacier displacements picked by hand
in Antarctica (Landsat, see references in Scambos
et al., 1992) 1989-1992 Crippen Blom
normalized cross-correlation of SPOT for dune
migration 1992 Scambos et al. use IMCORR
normalized cross-correlation applied to glacier
movements in optical (Landsat) 1996-1999
Normalized cross-correlation and Fournier methods
applied to SAR (glaciers) by 3 groups (see
citation in Joughin et al., 1999) 2000 Fournier
methods applied to optical images (SPOT) of
earthquakes (van Puymbroeck et al.) 2001, 2005
Combination of InSAR and SARSIR for an
unambiguous 3D velocity field for earthquake
(Fialko et al., 2001, using ROI_PAC) and
glaciers (Gray et al., 2005) 2004 Application to
airborne images combination of airborne
satellite images (Delacourt et al. -- QuickBird,
Landslide) 2008 Release of COSI-CORR
11
Compare InSAR subpixel image registration
When InSAR works, it usually has greater
precision (has a lower detection
threshold) Supports a greater horizontal
spacing Might work in areas where there arent
any optical features to track Sensitive to
vertical motions Can be effected by
ionosphere/troposphere Subpixel image
registration Can work in areas that InSAR phase
is incoherent because of too much deformation or
change in surface scattering properties No
unwrapping errors 2D instead of 1D
measurements Optical problem with clouds/night
but can be co-registered over years and between
different sensors SAR little backscatter on
sand changing backscatter in melt (optical might
be better)
1997 Mw 7.6 Manyi, Tibet earthquake (from Peltzer
et al., 1999)
12
InSAR at JIF
Oct. 1995
13
Seasonal differences in 1-day interferograms
May 1996
Jan. 1996
Sept. 1995
Dec. 1995
14
Hector Mine with SPOT
Leprince et al., 2007
15
Hector Mine Correct CCD distortions
Leprince et al., 2008
16
Courtesy Sebastien Leprince
Different types of optical sensors
17
Special issues with optical data
Special problems with optical Unknown or
unmodeled attitude variations (sometimes called
jitter) Distortion or misalignment of CCD
arrays Old film can have distortions in the film
or scanning process
Courtesy Sebastien Leprince
18
Courtesy Sebastien Leprince
Special issues with optical
19
Hector Mine air photos
Leprince et al., 2007 IGARRS
20
High resolution temperate glacier
SPOT5 2.5 m/pixel
21
MAI vs. Offset tracking
Technique development (Bechor Zebker, 2006,
sources of the figures) Siberian earthquake,
using a slightly different method (Barbot et al.,
2008) Find that baselines must be factor of 2
smaller than conventional interferometry to be
successful
22
Reconstructing the full 3D deformation field

• Use interferograms from different satellite look
  directions

• PLUS use the amplitude images to track pixels
  that moved

Observed interferograms
Observed pixel tracking
Inferred horizontal displacement
Inferred vertical displacement
Fault
Fialko et al., 2005
23
Construction of 3D velocity Taku glacier, Alaksa
Ascending Phase, Oct. 28-29, 1995
Descending Phase, Oct 29-30, 1995
Descending Azimuth Offsets
Ascending Azimuth Offsets
24
Invert for 3D flow Taku
E-W component
N-S component
U-D component
25
Who cares about Patagonia Alaska?
Temperate glaciers 3 of total glaciated
area Patagonia Alaska 20 of temperate
glaciers
Despite small size, Patagonia Alaska contribute
much to current sea level rise because more
responsive to climate change via large annual
throughput (2-11 m/yr precipitation) Meier,
1984 Arendt et al., 2002 Rignot et al.,
2003 Over the last 10 years, sea level rise
about Ramillien et al., 2006 3.1 /- 0.4
mm/yr 1.6 mm/y from thermal expansion 0.8 mm/y
from all temperate glaciers (0.4 mm/yr from
Alaska Patagonia) 0.5 mm/yr from Antarctica
Greenland combined
26
How to measure velocities?

• 1) Pixel tracking
• Optical ASTER Maybe SPOT
• Radar ERS-1/2 ALOS SIR-C Envisat Radarsat
• 2) InSAR (limited data available)
• 3) Ground measurements via GPS (mostly on Juneau
  Icefield)

Example of temporal data coverage from different
types of data over Juneau
27
1 Problem Too many clouds
We have submitted our first DARs this summer to
get repeat imaging of Juneau and Patagonia --
Hopefully some images will be cloud free!
28
2 Problem Not enough scenes
Most of Southern Patagonia has no available pairs!
Good pairs for Juneau northern Patagonia, but
need better temporal coverage
29
ASTER image of Taku Glacier
Local granule ID AST_L1A00305082005201124_051220
05105218.hdf
30
m/day
Blue
Red
31
m/day
Blue
Red
32
Different time periods SAR Pixel Tracking
Oct. 1995 Azimuth velocity in cm/day
33
Compare orthorectification products

• All examples use COSI-CORR
• Coregister Level 1A product to SRTM 1 arcsecond
• Start with LPDAAC Orthorectified image DEM.
  Coregister another L1A product
• Coregister 2 LPDAAC orthorectified images with
  individual DEMS

34
Compare DEM products
DEM errors translate into incorrect estimates of
glacier velocities
35
High pass filter Stretch Taku Glacier
3.1 km
3.1 km
High pass filter stretch
Raw
36
Different pre-filtering with ampcor
Meade Glacier 3 April 2007- 21 May 2007
Original
East-West Velocity (cm/day)
Highpass filtering before offsets
37
ASTER image of Pakistan
Local granule ID AST_L1A003_11142000060642_07202
001101528.hdf
38
cm
White
Black
39
Avouac et al (2006)?
40
Compare Ampcor COSI-CORR quality
41
Compare Ampcor prefiltering
42
Error analysisAre the rocks moving?
COSI-CORR does not seem superior to
Ampcor Standard deviation about 3-20 cm/day
43
Offset Tracking with Radarsat-1 March 2005 24-day
pairSeward and Malaspina Glaciers, Alaska
Rick Forster and Evan Burgess (AGU 2008 and in
prep)
44
Error analysisHow well can we measure pixel
offsets?
Pakistan example (5 years) 15 m pixels
COSI-CORR, Ampcor with filtering Ampcor without
filtering SIR-C 1 day L-band SAR Michel Rignot
(1999) 14 cm/d in ground range determined by
comparison with InSAR (1/30 of pixel) using 32x32
pixels Howat et al., 2005 Speckle tracking in
greenland accurate to 3 Rabus and Fatland,
2000 Black Rapids Glacier a principle error
source is rapid changes in glacier velocities 20
during 1995/1996 tandem campaign Luckman et al.,
2003 Greenland ERS feature tracking 1/20th of
pixel coverage only slightly reduced in 1 day
vs. 35 day (80 vs. 70), 20-40 cm/day vs. .5-3
cm/day Strozzi et al., 2007 JERS 1.5 of pixel
in 44 days (64x256) Pathier et al., 2006 0.35
m
45
Different pre-filtering with ampcor
Original
Comparison over Meade glacier, Alaska Principle
component analysis does not seem to provide a
benefit -- unlike with Landsat (Scambos et al.,
1992) High pass filter reduces scatter in areas
with no expected offset
46
Comparison with ground truth from JIRP
Profile 7a
Scott McGee making GPS measurement From
crevassezone.org
Used InSAR SAR pixel tracking
47
Some useful references/webpages
COSI-CORR Download the software, users manual,
publications www.tectonics.caltech.edu/slip_hi
story/spot_coseis/index.html IMCORR Software
instructions www-nsidc.colorado.edu/data/velmap/
imcorr.html ROI_PAC Download automatic offsets
software and instructions www.roipac.org/Offset_T
racking General overview of image registration
methods Brown, 1992, AMC Computing Surveys, vol.
24, p. 325-376
48
Practical considerations
Window size -- square vs. not for irregular pixel
sizes Allowed incidence angle difference? 8
degrees for Hector Mine Topographic effects (not
relevant when Baseline lt 300 m and topo up to 1
km Strozzi et al., 2007) Pathier et al., 2006
-- use radar.unw for range (how
scaled/normalized?) and a quadratic surface for
azimuth Removing CCD, satellite attitude,
ionospheric, and photograph/scanning
distortions For SAR use real or complex?
(complex is more strongly peaked in
low-correlation regions, but phase gradients can
reduce correlation peak in regions of high strain
or topo -- need larger window for real, Joughin,
2002)
49
Practical considerations
What data is available (if optical, is it cloud
free)? How expensive is it? Is there knowledge
of the camera properties, flight lines, and
orbital characteristics (only certain sensors
currently work with COSI-CORR)? What types of
data can be combined together? Is the viewing
geometry similar enough (how similar does it need
to be)? If optical images, what is the available
resolution of a DEM for orthorectification? How
big is the expected deformation signal? Is it
theoretically detectable given the resolution of
the available data? How big should the reference
window be? Trade-off between precision and
ability to measure spatial variations in
deformation.