In: Wagner W„ Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Voi. XXXVIII, Part 7B 
182 
financial and technological limitations. The second option is to 
resample the image intensity to a higher spatial resolution 
through interpolation. The third option is to interpolate the 
cross-correlation surface after the matching process to a higher 
spatial resolution in order to locate the correlation peak with 
sub-pixel precision. Since the intensity interpolation approach 
and the correlation interpolation approach are both generic, and 
independent of image resolution, they are subject to this study. 
A number of investigations have explored these two approaches 
applied to medicine (Althof et al. 1997), mechanics 
(Westerweel 1993; Willert and Gharib 1991; Zhou and Goodson 
2001), and stereo matching and motion tracking (Karybali et al. 
2008; Yamaguchi et al. 2003). However, there is no study 
available that compares the relative performance of the two 
approaches when measuring the displacement of earth surface 
mass movements from repeat images. 
Many earth surface mass movements such as landslides, glacier 
flow, and rockglacier creep are characterized by displacement 
rates of the order of magnitude of cma' 1 or ma' 1 which is often 
less than the spatial resolution of the space-bome or air-borne 
imagery typically available for their measurement. Sub-pixel 
accuracy of image matching algorithms, here NCC, has 
therefore a large potential to improve the signal-to-noise ratio of 
the measurements. Using NCC as an example, this study 
compares the performance of two fundamentally different 
approaches to reaching sub-pixel precision in mass movement 
detection and measurement from repeat remotely sensed 
images. 
In the method section of this contribution we describe the 
dataset used, ways of reaching sub-pixel precision, 
quantification of matching accuracy, and our experimental set 
up. In further sections the matching results and their accuracy 
are presented and discussed. Short conclusions terminate our 
contribution. 
2. METHODS 
2.1 Image data and pyramid 
For this study, three different types of mass movements were 
selected: land sliding, glacier flow, and rockglacier creep. The 
selection of these mass movement types was made based on 
their frequency in high mountain areas. Three temporal pairs of 
images each covering one of these types of earth surface mass 
movements were used (Table 1). These images were accurately 
orthorectified prior to displacement matching. Additionally, one 
image pair was created from one of the original glacier images 
after artificially inducing a two-dimensional translation of 15 
pixels (9 pixels in the X direction and 12 pixels in the Y 
direction). Since this pair was made from just one original 
image and the movement applied was only translation the pair 
serves as a control data set as it is free of noise from temporal 
surface changes, changes in imaging condition, registration 
errors and geometric distortions. 
Better understanding the influence of spatial resolution on the 
accuracy of image matching requires images of the same area 
taken at the same time, under the same flight and ground 
conditions, but using sensors with different spatial resolutions. 
Such conditions are not easily met. Instead, resampling of the 
original images was used here. In our study, different optical 
satellites were simulated by down-sampling the original high- 
resolution aerial ortho-images to five levels lowering the 
resolution by factors of 2, 4, 8, 16 and 32. One image pyramid 
with six levels each was finally obtained for each of the image 
pairs. The down-sampling was performed using the MATLAB 
module ‘imresize ’ with the relatively most efficient and reliable 
algorithm for this purpose, bi-cubic convolution. The algorithm 
assigns the weighted average of pixel values in the nearest 4 by 
4 neighbourhood (Keys 1981). Although this resampling 
process is slightly different from the pure signal averaging 
happening in the instantaneous field of view of a sensors 
detector cell, we decided to choose bi-cubic convolution 
because most images used for matching will in practice have 
undergone such interpolation during image correction and pre 
processing steps, such as orthorectification (Toutin 2004). 
Type 
Location 
Pixel 
size 
Older 
Recent 
Rockglacier 
Muragl (Swiss) 
0.2m 
1981 
1994 
Glacier 
Ghiacciaio del 
Belvedere (Italian) 
0.5m 
Sep. 
2006 
Oct. 
2006 
Rock slide 
Aletsch (Swiss) 
0.2m 
1976 
2006 
Control 
(Glacier) 
Manually 
translated motion 
0.2m 
Table 1. Brief description of the image data used 
2.2 Matching and displacement measurement at 
different pixel sizes 
First, the original high-resolution aerial images were matched 
using the pixel-precision NCC algorithm to determine the 
matching positions and compute the horizontal displacement 
magnitude and direction. These results were considered as 
reference for the accuracy assessment. Mismatches that were 
characterized by low peak correlation coefficients, very large 
displacements in relation to their neighbouring templates, or 
displacements showing distinct upslope movement were 
removed manually. Additionally, displacements less than the 
mean orthorectification error were removed as they are not 
reliably distinct from the error. The orthorectification error 
(offset between the images) was computed by matching stable 
grounds. The computation revealed that a maximum of 1 pixel 
offset exists in each dimension. The positions of the templates 
with valid matches in the original resolution were then used in 
the matching of the coarser resolution images. 
Matching and displacement measurement were in a next step 
performed on all resolution levels of the image pyramid pairs 
for all those locations saved from the reference matching. The 
absolute sizes and positions of the reference templates and the 
search windows were kept constant metrically throughout the 
image pyramid by adjusting the number of pixels according to 
the resolution. In other words, the ground area covered by the 
templates remained the same, the respective image resolution 
changed. This was done in order to avoid variations in signal 
content as a result of inclusion or exclusion of ground features. 
The area covered by the images range from 0.25km 2 to just over 
3km 2 . The size of the template was kept at around 26m and 65m 
for the originally 0.2m and 0.5m resolution images, 
respectively. The size of the search window was kept at around 
102m and 265m for the originally 0.2m and 0.5m resolution 
images, respectively, so that it certainly included the expected 
maximum surface displacement.