In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
SUB-PIXEL PRECISION IMAGE MATCHING FOR DISPLACEMENT MEASUREMENT
OF MASS MOVEMENTS USING NORMALISED CROSS-CORRELATION
Misganu Debella-Gilo 1 and Andreas Kaab 2
'^Institute of Geosciences, University of Oslo, P. O. Box 1047, Oslo, Norway
1 m.d.gilo@geo.uio.no (corresponding author), 2 kaeab@geo.uio.no,
KEY WORDS: Normalised cross-correlation, Sub-pixel, Image matching, Displacement measurement, Rockglacier, Glacier, Rock
slide
ABSTRACT:
This study evaluates the performance of two fundamentally different approaches to achieve sub-pixel precision of normalised cross
correlation when measuring surface displacements on mass movements from repeat optical images. In the first approach, image
intensities are interpolated to a desired sub-pixel resolution using a bi-cubic interpolation scheme prior to the actual displacement
matching. In the second approach, the image pairs are correlated at the original image resolution and the peaks of the correlation
coefficient surface is then located at the desired sub-pixel resolution using three techniques, namely bi-cubic interpolation, parabola
fitting and Gaussian fitting. Both principal approaches are applied to three typical mass movement types: rockglacier creep, glacier
flow and rock sliding. Their performance is evaluated in terms of matching accuracy and in reference to the images of the resolution
they are expected to substitute. Our results show that intensity interpolation using bi-cubic interpolation (first approach) performs
best followed by bi-cubic interpolation of the correlation surface (second approach). Both Gaussian and parabolic peak locating
perform weaker. By increasing the spatial resolution of the matched images by intensity interpolation using factors of 2 to 16, 40% to
80% reduction in mean error could be achieved in reference to the same resolution original image.
1. INTRODUCTION
Present climatic change shifts geomorphodynamic equilibriums
and intensifies related mass movement processes such as
landslides and permafrost creep (Haeberli and Beniston 1998;
Rebetez et al. 1997). Extension and intensification of human
activities in areas affected by such mass movements increase
the probability of connected adverse impacts like natural
hazards or building stability problems. The consequently
growing needs for monitoring mass movements are
complemented by growing remote sensing opportunities for
doing so. The increasing number of available stacks of multi
temporal space-borne, air-borne and terrestrial images, and the
improvements in remote sensing and image processing in
general significantly enhance the potential for applying
matching techniques to detect and quantify earth surface mass
movements from repeat remotely sensed data. All the above
needs and developments call for continued efforts to improve
terrain displacement matching methods based on repeat images.
Image matching is a group of techniques of finding
corresponding features or image patches in two or more images
taken of the same scene from different viewing positions, at
different times and/or using different sensors (Zitova and
Flusser 2003). Image matching is among others used for a large
variety of applications such as image (co-) registration, stereo
parallax matching for generation of digital elevation models,
particle image velocimetry (PIV), or displacement
measurements.
The group of area-based matching techniques is the most widely
used method due to its relative simplicity (Zitova and Flusser
2003). A number of similarity criteria can be used for the
matching process. Cross-correlation, in particular its normalised
form which accounts for intensity variations in image
sequences, is the most widely used due to its reliability and
simplicity (Lewis 1995). The normalised cross-correlation
(NCC) algorithm has been used to investigate earth surface
mass movements such as glacier flow, rockglacier creep and
land sliding in many empirical studies (Haug et al. 2010; Kaab
and Vollmer 2000; Scambos et al. 1992).
Although NCC has been documented to be simple and reliable,
a number of drawbacks have been reported as well: Firstly, its
precision is, in principle, limited to one pixel, and thus varying
with the pixel size of the image data used. Secondly, NCC is
sensitive to noise in the images. Such noise may result in wrong
correlation maxima leading to mismatches. This problem is
often partly addressed through image transformation (e.g.
Fourier) in case of images with low signal-to-noise ratio (Lewis
1995). Thirdly, NCC is sensitive to significant scale, rotation or
shearing differences between the images to be correlated (Zhao
et al. 2006). Due to this limitation, NCC is recommended in
cases where the movement is mainly due to translation with
limited rotation, scaling or shearing. A way to partly overcome
this limitation of NCC is to orthorectify the images used before
the matching (Kaab and Vollmer 2000). Fourthly, for the
measurement to be reliable the displacement has to be greater
than the mean error of the image (co-)registration. Improving
NCC precision improves displacement accuracy twofold: firstly,
it reduces image registration error; secondly, it improves the
matching accuracy directly.
To achieve sub-pixel precision in NCC, three approaches can be
used. One is to improve the imaging system towards a higher
spatial resolution. This approach is complicated by a number of
