FV_IGARSS11.ppt

Combining space-borne SAR data and digital camera images to monitor glacier flow by

remote and proximal sensing

R. Fallourd, F. Vernier,Y. Yan, D. Rosu, E. Trouvé, J.-M. Nicolas, J.-M. Friedt and L. Moreau

ANR-07-MDCO-04

2

Experimental Site: Mont Blanc

Mont blanc valley Argentière glacier

SAR LOS

3

Overview

The correlation algorithm− NCC (Normalized Cross Correlation)− Bases of fast correlation

Optical data set− Data & processing

− Results

SAR data set− Data & processing− Results

Data fusion Conclusion & Perspectives

4

Fast Correlation Technique Texture tracking, looking for

the maximum of a similarity function.

Use the Normalized Cross Correlation.

− Classical function− Optimized

implementation− Parallel implementation − Optimization and

parallelization can be extended to others similarity functions

5

Fast Correlation Technique The main objective is to

reuse already computed values.

A master window centered at the position (k,l)

6

Fast Correlation Technique The main objective is to

reuse already computed values.

Due to a dense correlation, the overlapping of the computation is important.

A master window centered at the position (k,l+1)

Hatchure part is already computed. It is not recomputed for this new

position of the master window. Sliding vectors or matrices are used

to manage the computed data.

7

Proximal sensor: Camera

Lognan serac falls

6 images every day, with 2-hour intervals, from 8:00 AM to 6:00 PM.

16:9 High Resolution images of 10 Mega pixels (4224 x 2376 pixels)

Automated digital camera installed near the Argentière glacier.

6 months without supervision.

8

Proximal sensor: Camera Processing:

− RGB JPEG images are converted in gray-scale images:

Luminance = 0.3 x Red + 0.59 x Green + 0.11 x Blue

− An initial co-registration between the images is made on the motion-free part of the images.

− The fast correlation is applied with: 31 x 31 pixels master window 51 x 51 pixels slave window

9

Proximal sensor: Camera

Highlight:− Displacement− Fallen serac − Serac that accelerates (will

fall)

Magnitude of 2D displacement(2008-10-9 / 2008-10-10).

Orientation of 2D displacement.(2008-10-9 / 2008-10-10).

Lognan serac falls, 2008-10-09.

10

Remote sensor: SAR 35 stripmap TerraSAR-X images on the

Mont-Blanc test site:− Many Ascending/Descending

temporal series.− In polarization HH, HH/VV or HH/HV.− Incidence angle of 37°.− 1.36 m per pixel in range and 2.04 m

per pixel in azimuth.− Large HR scene (about 30x50 km²).− 380 Mega pixels per image.

SAR LOSTS-X amplitude strip-mapimage 2008-09-29.

11

Remote sensor: SAR

Processing:− An initial co-registration by a simple translation (without

resampling).− The fast correlation is applied with:

61 x 61 pixels master window 77 x 77 pixels slave window ~16m max.

− A post-processing step can be necessary to deduce the offsets only due to the glacier movement.

12

Remote sensor: SAR

Dense correlation of a whole SAR image

Each alpine glacier of the area appears

Others particular structures are enlightenedSAR

LOS

TS-X image 2008-09-29. Magnitude of 2D displacement(2008-09-29 / 2008-10-10).

13

Remote sensor: SAR

Good results in textured areas Dis-correlation due to:

− Too many changes ( snowfall, too large movement...).− Not enough texture.

SARLOS

TS-X image 2008-09-29.TS-X image 2008-09-29. 2D displacement magnitude(2008-09-29 / 2008-10-10).

2D displacement orientation(2008-09-29 / 2008-10-10).

14

Computation time

Images Optimisation 1 cpu 8 cpu

Opticwithout 12 days 36 hours

with 80 min 10 min

Whole SAR

without - 18 days

with 5 days 15 hours

SAR partwithout 120 hours 12 hours

with 4 hours 30 min

octo-core Intel(R) Core(TM) i7 3GHz 24GB memory

15

Data Fusion

16

Data Fusion Problem to solve:

R = PU

− R: 2D displacement vector mesured on each projection.

− P: matrix of projection vectors.

− U: unknown vector of 3D displacement

WLS solution:

− With the input covariance matrix

U=(PT∑R

−1P)

−1PT∑R

−1R

∑R

17

Data Fusion Fusion of Optic and

SAR displacement Displacement speed in

meter. Small area due to

orthogonal point of view between SAR and Optic measurement.

3D displacement magnitude

18

Conclusion & perspectives

The computation on each point of the image can be achieved in a reasonable time.

The optimization deals with optical and SAR images. The experiments highlight the problems and results obtained

by fusion of these results. The software is available in the “EFIDIR Tools” (GPL)

www.efidir.fr. ''Fast Correlation Technique for Glacier Flow Monitoring by

Digital Camera and Space-borne SAR Images''accepted in Journal Image and Video Processing.

19

Conclusion & perspectives

Install two cameras higher to:− See the top of glacier with the camera.− Have larger overlapping area observed by satellite and

cameras.− Use the stereo effect to compute 3D displacement with these

new cameras.− Update DEM for the use of SAR images.

20

Questions...

21

22

Experimental Data

6 images per day from an automated digital camera installed in the front of ice falls of “Argentière glacier”.

− Data size: 6x10 Mega pixels per day. 1 TerraSAR-X image every 11 days.

− Data size: more than 380 Mega pixels per image. Objectives:

− Compute displacement for both data sets.− Combine the results to compute a 3D displacement.

23

Fast Correlation Technique

The main technique is based on sliding vector and matrix

Vector of size 4, 1st step:

1 2 3 4 ? ? ?

The first line

11 12 13 21 22 23 31 32 33 ? ?

The second line

The third line

11 12 13

21 22 23

31 32 33 ? ?

Matrix of size 3x3, 1st step:

Head

24

Fast Correlation Technique

The main technique is based on sliding vector and matrix

Vector of size 4, 2nd step:

1 2 3 4 ? ? ?

Matrix of size 3x3, 2nd step:

11 12 13 21 22 23 31 32 33 ? ?

11 12 13 21

22 23 31

32 33 ? ?

The first line The second

lineThe third line

Head

25

Fast Correlation Technique

The main technique is based on sliding vector and matrix

Vector of size 4, 2nd step:

Matrix of size 3x3, 2nd step:

The first line

11 12 13 14 22 23 24 32 33 34 ?

The second line

The third line

11 12 13 14

22 23 24

32 33 34 ?

Head

1 2 3 4 5 ? ?