In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
4 
For the real mass movements, there is very little accuracy gain 
by interpolating to lower than 0.1 pixels. As the result of the 
control shows, when the movement is only translation, the 
magnitude of the deviation is very low. Besides, it seems that, 
for the control set, interpolation to more detail level improves 
the accuracy further. 
Mass movement 
Mean 
displa 
cemen 
t(m) 
Maxi 
mum 
displa 
cemen 
t (m) 
Standard 
deviation 
of 
displacem 
ent (m) 
Maximu 
m 
velocity 
(ma') 
Aletsch Rock 
slide 
1.5 
4.2 
0.45 
0.14 
Muragl 
Rockglacier 
2.4 
5.8 
1.20 
0.45 
Ghiacciaio del 
Belvedere 
Glacier 
12.22 
18.83 
5.0 
226 
Control 
7.50 
7.50 
0 
7.50 
Table 2. Summary statistics for the displacement magnitudes 
and average velocity of the mass movements and the 
translation-only control image as estimated from the matching 
of the high-resolution original ortho-images 
Figure 1. Displacement vectors on the Ghiacciaio del Belvedere 
(Sept - Oct 2001) 
Figure 2. Displacement vectors on the Muragl rockglacier 
(1981-1994) and Aletsch rockslides (1976 - 2006) from left to 
right respectively. 
3,3 Relative performance of the sub-pixel precisions 
Figure 5 shows the mean deviation between the matching 
position of the interpolated image pairs and that of the same 
resolution (but original) reference image pairs plotted against 
sub-pixel precision for the control set. When the difference 
between the images is only the here-applied translation, sub 
pixel interpolation of the image intensities up to l/8 th of the 
original pixel size prior to matching can perfectly substitute 
images of comparable original resolution. This perfect 
substitution can be achieved by using bi-cubic interpolation of 
the correlation surface only up to l/4 th of the original pixel size. 
For example, a 16m resolution image interpolated to 2m using 
bi-cubic interpolation before matching performs exactly as a 2m 
resolution image pair as long as there is no other source of 
difference between the image pairs than rigid translation. But 
when the level of detail goes beyond l/8 lh , there appears 
deviation between the two. The magnitude of these numbers 
depends, of course, on the translation magnitude applied in the 
control set. However, the test shows the better performance of 
bi-cubic intensity interpolation over the other sub-pixel 
algorithms tested. 
For all the real mass movement types (Figure 6), as the 
difference in pixel size between the coarse resolution and the 
reference resolution increases, the deviation of the sub-pixel 
matching position from the matching position of the same (but 
original) image resolution increases regardless of the algorithm. 
This means, not surprisingly, that the sub-pixel algorithm 
resembles images of comparable resolution less and less as the 
sub-pixel detail increases. At every resolution, the mean 
deviation is the lowest when intensity interpolation is used 
before matching followed by the bi-cubic interpolation of the 
correlation surface. The parabola- and Gaussian-based peak 
localisations perform poorer and alike. This confirms the above 
results. 
Remarkably, at a certain level of sub-pixel detail (about 1/16 th ), 
the deviation between the sub-pixel algorithm and same 
resolution original image gets so high that interpolating beyond 
that level has no meaningful advantage although the control set 
gives less deviation even at greater level of detail. 
Bi-cubic 
(correlation) 
Bi-cubic 
(intensity) 
Gaussian 
Parabolic 
Sub-pixel precision (pixel) 
Figure 3 Accuracy of the different sub-pixel precision 
approaches for the control set expressed as the mean deviation 
of the matching positions from that of the reference high- 
resolution original ortho-images 
Sub-pixel precision (pixel) 
—■— Bi-cubic 
(correlation) 
—»— Bi-cubic 
—•— Parabolic 
—*— Gaussian 
No 
interpolation 
Figure 4 Accuracy of the different sub-pixel precision 
approaches expressed as the mean deviation of the matching 
positions from that of the reference high-resolution original 
ortho-images (averaged for the three mass movement types)