Imageodesy on MPI

1
Imageodesy on MPI GRID for Co-seismic Shift
Study Using Satellite Optical Imagery
Knowledge discovery from massive data processing
for earthquake study

• Jian Guo Liu and Jinming Ma
• Department of Earth Science and Engineering
• Imperial College LondonSouth Kensington campus,
  London SW7 2AZ, UK
• j.g.liu_at_imperial.ac.uk

2
Outline of the presentation

1. Introduction data mining software development
   for geohazard study
2. Imageodesy the principle and implementation on
   MPI and GRID
3. Case study Co-seismic shift of the Ms 8.1 Kunlun
   earthquake
4. Conclusions

3
Acknowledgement
This research is part of DiscoveryNet project
(GR/R67750/01) supported by EPSRC e-science pilot
project grant. Computing Centre of Imperial
College London provided parallel processing
facilities and technical support. MIT Phase
Correlation website provided free access and
technical support for software development. ER
Mapper image processing software has been used
for data visualisation and analysis. Xinjiang
Bureau of Seismology is acknowledged for
providing field photos and some reference
materials of Kunlun earthquake.
4
1. Introduction
Geohazard study in DiscoveryNet project Geohazard
monitoring and assessment is one of the major
application areas of DiscoveryNet project aiming
to scientific knowledge discovery. Remote
sensing A typical characteristic of geohazards is
that movement and displacement will be produced.
For instance, an earthquake produces co-seismic
displacement while a landslide is characterised
by slope failure and mess movement. Remotely
sensed imagery data enable detection and
measurement of these changes and thus provide
vital information for hazard assessment and
prevention. Algorithm and software
development Fast imageodesy algorithms and
software for high accuracy change detection have
been developed and integrated into DiscoveryNet
workbench. The massive data processing is
possible only with advanced algorithms using
powerful MPI/GRID computing.
5
2. Imageodesy the principle
The Imageodesy technique (Crippen 1992 Crippen
Blom 1996) is capable of measuring the
horizontal shift of image features, at a
sub-pixel level accuracy, through normalized
cross-correlation (NCC) between pre- and
post-event images. This technique is a
complementary to the well-established radar
interferometry technique that is sensitive to
vertical deformation. The major technical
challenge of high accuracy imageodesy is the huge
demand on computing to handle with massive data
processing. We have developed software to
implement the imageodesy on MPI parallel
processor based on the conventional template
normalised cross-correlation (NCC) algorithm.
Furthermore we developed new algorithm and
software based on phase correlation algorithm.
6
2. Imageodesy FNCC algorithm
The NCC is defined as
The Fast NCC algorithm (FNCC) reduces large
quantity of repeated operations and further
improvement including conditional jumps and smart
sampling avoids unnecessary operations over
homogeneous image features. Thus the processing
can be speed up by 5-10 times depending on the
size of processing windows and the image
characteristics.
7
2. Imageodesy FNCC MPI implementation
With an improved FNCC algorithm, operating on a
MPI UNIX parallel computer with 24 processors, it
takes 10 hours to complete imageodesy processing
for one pair of cross-event Landsat-7 ETM Pan
imagery data. The image size is 3.75GB after
interpolating to 3m pixel size.
8
2. Imageodesy FNCC GRID implementation

• FNCC is a neighbourhood processing, each image
  line (the minimal data unit) must carry its
  neighbour image lines with it in order to conduct
  the processing. This wipe off 50 efficiency of
  FNCC and increases data communication by m times,
  where m is the search window size.
• The experiments on GRID using a small image
  (512?512) completed much slower than local
  processing using a single PC (2GHz processor).
  Submission larger images of a few thousands lines
  and columns to the GRID simply blocked the
  processing pipe line and failed to complete the
  task.
• The current status of GRID is not sufficient for
  the massive neighbourhood processing of FNCC
  imageodesy. The future of GRID for dealing with
  the type of processing of imageodesy lies on very
  fast high throughput network.

9
2. Imageodesy Phase Correlation algorithm
We have implemented phase correlation algorithm
for imageodesy operating on single UNIX and PC.
By transforming the image data within a matching
window into frequency domain via FFT, the phase
correlation can pinpoint the best matching
position directly as the peak of the overlap
between the frequency distribution of the two
images, without the time consuming searching .
Phase Correlation Imageodesy algorithm scheme
10
3. Case study Co-seismic shift of Ms 8.1 Kunlun
earthquake
Kunlun Earthquake On 14 November 2001, at
092618 UTC, an Ms 8.1 earthquake occurred in
the East Kunlun Mountains, along the Kusai Lake
segment of Kunlun fault. A 400 km surface rupture
zone of E-W to WNW-ESE orientation was produced
and, according to the field observations of
Chinese scientists, the fault displacement was as
large as 16.3 m (Lin et al. 2002, 2003). Remote
sensing study Landsat TM and SPOT images have
already been used to locate and map the visible
surface rupture features (Fu Lin 2003). SAR
interferometry (InSAR) would be an ideal
technique to reveal the stress field of the
earthquake and to provide two-dimensional
quantitative measurements of the fault movement.
The lack of high quality cross-event ERS SAR
fringe pairs for this region and in this
particular time unfortunately hindered the use of
interferometry.
11
3. Case study data ETM imagery
Landsat ETM imagery data was chosen because of
its large coverage, improved 15 m resolution
panchromatic band, and the availability of
suitable pre- and post-earthquake, almost
cloud-free scenes. The output images of
imageodesy are X-shift, Y-shift and R (the NCC
coefficient).
Scene Names (Path/Row) Scene Names (Path/Row) Upper-Left corner coordinate (Lat/Lon, Dec. Deg.) Upper-Left corner coordinate (Lat/Lon, Dec. Deg.) Acquisition Date
Kusai Lake (KL) scene (138/035) Pre 36.9984856N 91.2616043E 3 Oct. 2001
Kusai Lake (KL) scene (138/035) Post 37.0025253N 91.2769623E 15 May 2002
Buka Daban (BD) scene (139/035) Pre 36.9999619N 89.7614136E 26 Oct. 2001
Buka Daban (BD) scene (139/035) Post 36.9977531N 89.7337036E 26 Aug. 2002
12
3. Case study results the left-lateral movement
of Kunlun fault
13
3. Case study results Kusai Lake scene


Histograms of smoothing filtered X-shift image of
KL scene with 0.7 correlation threshold. Left
The whole scene the high peak on the left is at
0.7 and the shoulder on the right is centred at
2.5. Middle The south side of the Kunlun fault
zone, the peak is at 2.5. Right The north side
of the Kunlun fault zone, the peak is at 0.7.
The range of the left-lateral shift is 1.58.1
m, while shift is most commonly 5.4 m. The
maximum net left-lateral displacement can be as
great as 13 m.

14
3. Case study results Kusai Lake scene
The co-seismic shift vectors of the Kusai Lake
area overlaid on the post earthquake ETM Pan
image. The vectors were derived from X and
Y-shift images, with 371?371 window averaging,
20 cut-off for elimination of extreme values,
and a 0.8 NCC coefficient criterion.



15
3. Case study results Kusai Lake scene
The co-seismic shift vectors of the Kusai Lake
area overlaid on the post earthquake ETM Pan
image. The vectors were derived with X-shift
compensated by -3.6 m.



16
The earthquake fault zone from space
17
3. Case study results Buka Daban scene




Histograms of smoothing filtered X-shift image of
BD scene with 0.7 correlation threshold. Left
The whole scene the high peak on the left is at
0.35 and the low peak on the right is centred at
1.7. Middle The south side of the southern
branch, the peak is at 1.7. Right The north side
of the northern branch, the peak is at 0.3. The
range of the left-lateral shift is 1.0 to 8.2 m,
while the most representative figure is 4.2 m.

18
3. Case study summary
Kusai Lake scene Fault is consistently in W-E to
WNW-ESE direction. Left-lateral shift range 1.5
m8.1 m, average 4.8 m, the maximum 13 m. Buka
Daban scene With the splayed nature of the fault
in this section, the displacement patterns become
complicated and stepwise. The average
left-lateral shift over the broad fault zone is
4.6 m and ranges from 1.0 m to 8.2 m. Both sides
of the faults moved toward the east. the south
side of the fault has been displaced
significantly to the right (east) relative to the
largely stable, or slightly right-shifting,
northern block. The relative movement of the
fault is left-lateral and the south side of the
fault is the active block.
19
4. Conclusions

• As an essential function of remote sensing data
  mining for geohazard study in DiscoveryNet
  project, FNCC imageodesy technique has been
  implemented on the workbench operating on MPI.
• The FNCC implementation on GRID yields a
  disappointing performance because the demand for
  data communication increases dramatically when
  the neighbourhood processing of imageodesy is
  distributed to many nodes.
• The phase correlation based algorithm is
  currently not operational on MPI or GRID
  processing mode. It is only more efficient when
  the forward and inverse FFT operations in phase
  correlation take less time than searching in
  FNCC.
• Our imageodesy results present the first regional
  2-D picture of the co-seismic displacement of
  Kunlun fault as the result of the Ms 8.1 Kunlun
  earthquake, in a vast area, of about 320 km W-E
  and 180 km N-S. It is an important scientific
  knowledge discovery.

20
Thank You!
Picture provided by Xinjiang Bureau of Seismology