1. 3D High Resolution Tracking of Ice Flow using Multi-Temporal Stereo Imagery: Franz Josef Glacier, New Zealand Sébastien Leprince 1, Jiao Lin 1, Francois Ayoub 1, B. Conejo 1, F. Herman 2, and Jean-Philippe Avouac 1 1 California Institute of Technology 2 Universite de Lausanne, Switzerland AGU - December 2013

2. COSI-Corr: Co-registration of Optically Sensed Images and Correlation Time  Satellite imagery acquired at different times, any resolution, possibly by different sensors http://www.tectonics.caltech.edu/slip_history/spot_coseis/ Leprince et al., IEEE TGRS, 2007  Automatic registration with accuracy of 1/10 of the pixel size (sub-pixel)  Automatic comparison of images to measure motion Vector field Correlation Measuring surface processes from optical imagery  Ground deformation COSI-Corr

3. Extension: processing multi-angle pushbroom images to measure 3D ground deformation In this presentation: (More) Detailed processing chain and possible extensions, Application to ice flow monitoring using Worldview images above Franz Josef Glacier, New Zealand.

4. 3D and 4D Processing Flow 1.Optimize Viewing Parameters

5. Optimize Viewing Parameters Jointly optimize external parameters: roll, pitch, yaw angles: r(t), p(t), y(t) Spacecraft position in time x(t), y(t), z(t) 2 nd order polynomials approximation If no GCP, regularized solution to stay within instrument uncertainties.

6. 3D and 4D Processing Flow 1.Optimize Viewing Parameters -Pairwise image matching between all images, -Only keep tie-points on stable surfaces (e.g., bedrock), -Optimize external viewing parameters of all images jointly using regularized bundle adjustment. 2.Produce Disparity Maps

7. Produce Disparity Maps Image Matching Framework with Regularization Similarity criteria (ZNCC) d constant on patch Piecewise constant prior Weighs the prior Neighbors of pixel x Given a stereo-pair of images (I S,I T ) how to retrieve the disparity map d?

8. Produce Disparity Maps Regularization: Semi-Global Matching (SGM) – current implementation Use of multi-scale pyramidal approach to lower the complexity Non convex problem => variational approaches won’t work => Restrict d(x) to a finite set of value and rewrite as first order Markov Random Field Only regularize along 1D directions and aggregate costs. Reduces complexity. Hirschmuller, CVPR, 2005

9. Produce Disparity Maps Regularization: Global Matching (GM) – under implementation => Restrict d(x) to a finite set of value and rewrite as first order Markov Random Field Complex problem (NP- Hard) but several effective techniques exist, e.g. Fast Primal- Dual from Komodakis, Tziritas and Paragios See B. Conejo poster G33A-0971 on Wed afternoon.

10. 3D and 4D Processing Flow 1.Optimize Viewing Parameters -Pairwise image matching between all images, -Only keep tie-points on stable surfaces (e.g., bedrock), -Optimize external viewing parameters of all images jointly using regularized bundle adjustment. 2.Produce Disparity Maps -Project all images on reference surface (e.g. low res DTM, GTOPO or smoothed GDEM), -Cross-correlate image pairs using multi-scale, regularized image correlation. 3.Produce Point and Vector clouds (3D, 4D)

11. 1.Produce Point and Vector clouds (3D, 4D) Triangulate multiple disparity maps to retrieve 3D topography and displacement fields

12. 3D and 4D Processing Flow 1.Optimize Viewing Parameters -Pairwise image matching between all images, -Only keep tie-points on stable surfaces (e.g., bedrock), -Optimize external viewing parameters of all images jointly using regularized bundle adjustment. 2.Produce Disparity Maps -Project all images on reference surface (e.g. low res DTM, GTOPO or smoothed GDEM), -Cross-correlate image pairs using multi-scale, regularized image correlation. 3.Produce Point and Vector clouds (3D, 4D) -Triangulate disparity maps, -(x 1, y 1, z 1 ) -(x 2, y 2, z 2 ) -(x 1, y 1, z 1, D x, D y, D z ) -Output surface models at all times.

13. 3D and 4D Processing Flow 1.Optimize Viewing Parameters -Pairwise image matching between all images, -Only keep tie-points on stable surfaces (e.g., bedrock), -Optimize external viewing parameters of all images jointly using regularized bundle adjustment. 2.Produce Disparity Maps -Project all images on reference surface (e.g. low res DTM, GTOPO or smoothed GDEM), -Cross-correlate image pairs using multi-scale, regularized image correlation. 3.Produce Point and Vector clouds (3D, 4D) -Triangulate disparity maps, -(x 1, y 1, z 1 ) -(x 2, y 2, z 2 ) -(x 1, y 1, z 1, D x, D y, D z ) -Output surface models at all times. 4.Grid Point Clouds and Vector Clouds -Use standard gridding libraries on each components (only external processing).

14. Application: Monitoring Glacier Flow in New Zealand Multi-temporal Stereo Acquisitions using Worldview GSD 50 cm: -January 30, 2013 (x2) -February 9, 2013 (x2) -February 28, 2013 (x2) 2 km -Bundle adjustment between all images, -Multi-scale image matching due to large disparities (up to 1000 pixels), -Regularized matching because of occlusions Franz Josef Glacier Fox Glacier Tasman Glacier

15. 2 km Application: Monitoring Glacier Flow in New Zealand East-WestNorth-SouthVertical Measurements with intersection errors larger than 2m (20cm/day) have been removed (white areas). 3D motion between January 30 and February 9, 2013 -1.5 m/day 1.5 -1.5 m/day 1 -0.8 m/day 2

16. 2 km Application: Monitoring Glacier Flow in New Zealand East-WestNorth-SouthVertical 3D motion between January 30 and February 9, 2013 -1.5 m/day 1.5 -1.5 m/day 1 -0.8 m/day 2 Measurements with intersection errors larger than 2m (20cm/day) have been removed (white areas).

17. 1m GSD Shaded Elevation Model generated from stereo pair: January 30, 2013 2 km Application: Monitoring Glacier Flow in New Zealand

18. 1m GSD Shaded Elevation Model generated from stereo pair: February 9, 2013 2 km Application: Monitoring Glacier Flow in New Zealand

19. 1m GSD Shaded Elevation Model generated from stereo pair: February 28, 2013 2 km Application: Monitoring Glacier Flow in New Zealand

20. 2 km Velocity changes Application: Monitoring Glacier Flow in New Zealand 0 m/day >2 Jan 30 – Feb 9, 2013

21. 2 km Velocity changes Application: Monitoring Glacier Flow in New Zealand 0 m/day >2 Feb 9 – Feb 28, 2013

22. Velocity changes (1) (2) Velocity (GPS), Precipitations, Temperatures B. Anderson (pers. Com.) Univ. Wellington, NZ Glacial dynamics strongly affected by hydrology Application: Monitoring Glacier Flow in New Zealand

23. Average daily velocity of up to 4 m/day at the Fox Glacier Application: Monitoring Glacier Flow in New Zealand 0 m/day >2

24. Conclusions Rigorous method to measure 3D ground deformation using multi-temporal optical stereo acquisitions, All software and algorithms developed in house, except for the gridding that uses standard libraries (but can be improved), Generic methods to monitor a variety of surface processes (fault rupture, landslides, sand dune migration, glaciers, etc.), Implementation on cluster computing environment for fast processing, Next generation of matching algorithm using Global regularization is coming soon (see B. Conejo poster on Wed), Successfully demonstrated on earthquake last year, and now on glaciers. Working on making the workflow more robust and more automatic. Would like to propose a cloud application.