PERFORMANCE AND APPLICATION OF DIFFERENT IMAGE MATCHING ALGORITHMS
FOR INVESTIGATING GLACIER AND ICE-SHELF FLOW, PERMAFROST CREEP AND
LANDSLIDES
Torborg Haug, Misganu Debella-Gilo, Jonas Karstensen, Andreas Kääb
University of Oslo
Department of Geosciences
P.O. Box 1047 Blindern, 0316 Oslo, Norway
ABSTRACT
This study focuses on evaluating image correlation methods
for deriving surface velocity fields of glaciers, ice shelves,
rockglaciers and landslides from both repeat optical and synthetic aperture radar (SAR) images. Firstly we show that low
resolution MODIS images (250 m spatial resolution) can in
fact be used for determining displacements on Antarctic ice
shelves. In this particular case, orientation correlation operated in the frequency domain and using only the phase part of
the signal performs better than normalized cross-correlation
operated in the spatial domain. Secondly we demonstrate that
for determining subpixel displacement when using normalized cross- correlation, image intensity interpolation before
the matching process outperforms interpolation of the correlation peak after the matching. We also study methods for
adaptively varying the matching template size in order to create the most reliable and accurate matches. The last part of
the work shows some application studies of glacier flow from
optical and SAR images, rockglacier creep and land sliding.
Index Terms— Image matching, displacement measurements, glacier and ice shelf, rockglacier, landslide
1. INTRODUCTION
Correlation of repeat image data and radar interferometry are
key technologies for investigating surface velocity fields of
glaciers, ice shelves, permafrost creep and landslides, and
their changes over time. Due to loss of phase coherence,
radar interferometry over snow and ice usually requires tandem missions or short temporal baselines with a few days
time base [1]. No such mission is currently or in the near
future available, so that image correlation, both of optical and
SAR images, is the only present way to measure terrain movements under snow and ice conditions. Further development
and evaluation of image correlation techniques is therefore
needed to select the optimal techniques for individual movement processes and image types, and to ensure optimal exploitation of the vast archives of repeat earth observation images. In this contribution we focus on the comparison of dif-
978-1-4244-9566-5/10/$26.00 ©2010 IEEE
3636
ferent image matching algorithms, on displacement determination by subpixel accuracy, and on adaptive matching methods that take account of the spatial variability of the image
content. As a representative set of image types within our assessment studies we select aerial photos, Landsat and ASTER
type medium resolution optical satellite data, low resolution
optical data such as MODIS, and medium and high resolution
spaceborne SAR data. Application areas of our assessments
are ice shelves, glaciers, rockglaciers and landslides.
2. ICE-SHELF VELOCITIES FROM LOW
RESOLUTION SATELLITE IMAGES
Ice shelf velocity is an important indicator of the stability of
Antarctic ice shelves. Several of the ice shelves that have disintegrated so far have shown a speed-up prior to the collapse
[2, 3, 4]. The size of the ice shelves and the amount of clouds
in Antarctica make the ice shelves difficult to monitor, and
up until now there has been only few studies focusing on velocity of these ice shelves. However, low resolution images
have short repeat times, a fact that increases the likelihood of
obtaining cloud free images. One image covers easily an entire ice shelf, which makes it possible to perform large-scale
monitoring. Due to the spatial resolution, subpixel-accuracy
measurements are needed in order to obtain the accuracy necessary to identify velocity changes. In this study [5], we use
the Larsen C ice shelf, Antarctic Peninsula, as a test area. The
velocity field is derived using MODIS images (250 m spatial resolution) over the periods 2002–2006 (not shown) and
2006–2009 (Fig. 1). The measurements are validated for two
ice shelf sections against repeat medium-resolution Landsat 7
ETM+ pan data (15 m spatial resolution).
Horizontal surface velocities are obtained through image
matching using orientation correlation (OC) operated in the
frequency domain and using only the phase part of the signal (Fig. 1) and normalized cross-correlation (NCC) operated
in the spatial domain (Fig. 2). Using the same window size
and the same position of the matching windows, OC produces
332 correct matches out of 471 possible (70%), whereas NCC
IGARSS 2010
Fig. 1. Average annual velocity between 2006 and 2009 measured with orientation correlation operated in the frequency
domain on MODIS images.
Fig. 2. Average annual velocity between 2006 and 2009 measured with normalized cross-correlation operated in the spatial domain on MODIS images.
produces only 129 correct matches (27%). OC operated in the
frequency domain using only the phase seems therefore better
suited for this purpose compared to NCC operated in the spatial domain because it produces fewer mismatches, is able to
match images with regular noise and data voids, and is faster.
Since it can match images with regular voids it is possible to
match Landsat 7 ETM+ images even after the 2003 failure of
the Scan Line Corrector (SLC off) that leaves significant image stripes with no data. The accuracy of the two matching
methods is comparable because the root mean square error
(RMS) of the measurements over stable ground is 27.8 m in x
direction and 29.5 m in y direction for NCC and 21.2 m and
35.9 m, respectively, for OC.
The uncertainty in the velocities turns out to be about 70 m
for the MODIS derived data, and about 15 m for the Landsatderived ones. The difference between MODIS and Landsatbased speeds is -15.4 m a−1 and 13.0 m a−1 for the first period
for the two different validation sections on the ice shelf, and 26.7 m a−1 and 27.9 m a−1 for the second period for the same
sections. This is within the uncertainty level, and leads us to
conclude that repeat MODIS images are well suited to measure ice shelf velocity fields and monitor their changes over
time. We also applied our method successfully to ten other
ice shelves around Antarctica.
3637
3. SUBPIXEL ACCURACY USING NORMALIZED
CROSS-CORRELATION
In a second study [6], subpixel-accuracy measurements, influence of spatial sensor resolution on the matching results, and
adaptive methods for optimizing the matching template sizes
are tested using repeat aerial photographs of a glacier, a rockglacier and a landslide in the European Alps. Thereby, we
focus on the use of the normalized cross-correlation (NCC),
one of the most widely used image matching techniques. Our
results reveal that image intensity interpolation using the bicubic scheme beforehand the matching outperforms methods
of defining the correlation peak at subpixel accuracy after the
matching process (Fig. 3). By increasing the spatial resolution
of the matched images by intensity interpolation with factors
of 2 to 16, 40% to 80% reduction in mean error in reference
to the original image with same resolution could be achieved.
Based on an image resolution pyramid starting from the
high-resolution aerial photographs, the study also quanti-
Fig. 3. Comparison of the different sub-pixel precision approaches for the real mass movements averaged from the three mass
movement types investigated.
Fig. 5. Displacements between 1981 and 1984 for Muragl
rockglacier, Swiss Alps, measured using adaptive window
sizes shown in Fig. 4.
Fig. 4. Size of matching templates determined based on the
contrast in the images. Muragl rockglacier, Swiss Alps.
fies how the mean error, the random error, the proportion
of mismatches and the proportion of undetected movements
increase with increasing pixel size (i.e. decreasing spatial
resolution) for all of the three mass movement types investigated.
As a third step in this study we demonstrate methods for
adaptively varying the size of matching templates in order to
create the most reliable and accurate matches. Fig. 4 shows
the different sizes of the matching templates when these are
3638
determined based on the contrast in the images, and Fig. 5
shows the outcome of this matching.
4. APPLICATION STUDIES
Finally, we assess a subset of the above matching techniques for a number of additional application studies, mainly
measuring glacier flow based on repeat optical satellite
data (ASTER, Landsat) and SAR data (ERS SAR, ASAR,
Radarsat-1, Radarsat-2 and TerraSAR-X). Fig. 6 shows the
displacement of Kronebreen on Svalbard, Norwegian high
Arctic, from 5 April 1996 to 10 May 1996 measured using
NCC on ERS-1 scenes. Fig. 7 shows the displacement of
Aletsch landslide in the Swiss Alps, measured using NCC on
aerial photographs obtained 6 September 1976 and 5 September 2006.
Fig. 6. Displacements of Kronebreen, Svalbard, from 5 April
1996 to 10 May 1996 measured using normalized crosscorrelation on ERS-1 scenes.
Fig. 7. Displacements of the Aletsch landslide from 6
September 1976 to 5 September 2006 measured using normalized cross-correlation on aerial photographs.
7. REFERENCES
5. CONCLUSIONS
Systematic exploitation of the vast existing and near-future
archives of repeat satellite images would certainly provide
fundamentally new insights about Earth surface mass movements such as glaciers, ice shelves and ice streams, rockglaciers, or landslides, as demonstrated by a number of examples given here. In order to cope with the huge amount
of data potentially useful for displacement extraction, their
analysis requires highly automated processes of image retrieval, image selection, image matching, and postprocessing
of displacement fields. Our set of studies suggests that there
seems not to be simply one optimal image correlation method
and parametrization for all kinds of image types and mass
movement processes. Also, selecting suitable image templates beforehand of the matching and adapting the correlation parametrization to the individual template characteristics
seems promising.
6. ACKNOWLEDGEMENTS
This study is funded by The Research Council of Norway
(NFR) through the CORRIA project (no. 185906/V30). The
work presented here benefited from the fruitful discussions
within the study ”Monitoring Earth Surface Changes from
Space” by the Keck Institute of Space Studies at Caltech and
JPL (KISS). The method development for ice flow measurements serves also the ESA DUE GlobGlacier project.
3639
[1] A. Kääb, Remote sensing of mountain glaciers and
permafrost creep, Schriftenreihe Physiche Geographie
Glaziologie und Geomorphodynamik. Geographisches
Institut der Universitt Zrich. ISBN 3 85543 244 9, 2005.
[2] R. A. Bindschadler, M. A. Fahnestock, P. Skvarca, and
T. A. Scambos, “Surface-velocity field of the northern
larsen ice shelf, antarctica,” Annals of Glaciology, vol.
20, pp. 319–326, 1994.
[3] P. Skvarca, W. Rack, and H. Rott, “34 year satellite
time series to monitor characteristics, extent and dynamics of Larsen B Ice Shelf, Antarctic Peninsula,” Annals of
Glaciology, vol. 29, pp. 255–260, 1999.
[4] P. Skvarca, H. De Angelis, and A.F. Zakrajsek, “Climatic
conditions, mass balance and dynamics of Larsen B ice
shelf, Antarctic Peninsula, prior to collapse,” Annals of
Glaciology, vol. 39, pp. 557–562, 2004.
[5] T. Haug, A. Kääb, and P. Skvarca, “Monitoring ice
shelf velocities from repeat MODIS and Landsat data - a
method study on the Larsen C ice shelf, Antarctic Peninsula, and 10 other ice shelves around Antarctica,” The
Cryosphere, vol. 4, pp. 161–178, 2010.
[6] M. Debella-Gilo and A. Kääb, “Sub-pixel precision algorithms for normalized cross-correlation based image
matching of mass movements,” Remote Sensing of Environment, in review.