In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). 1APRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3, 2010 
Note that, this scheme determines the bias term by projection 
from image-1 to image-2. One can also perform bias term 
determination in the inverse direction. Although the average of 
the two results may be used as the result, in this study, only 
one-directional bias compensation is performed. 
Moreover, if the biases for both images are in the same 
direction, the benefits of this method are questionable. Still, the 
SRTM data will behave as anchors, similar to the GCPs. 
3. EXPERIMENTS AND RESULTS 
3.1.SRTM-Single IKONOS Image Registration 
For single-image case, the registration accuracy is evaluated on 
123 Ground Control Points (GCPs). The results are as presented 
in Table 1. 
Value GCP- Value r cg 
Mean error 
Gerror 
RMS error 
Latitude 
2.13xl0' 5 ° 
1.54x1 O' 50 
2.624x1 O' 5 ° 
Longitude 
-5.56x10 5 ° 
1.46x1 O' 50 
5 751xl0' 5 ° 
Height 
-5.3 m 
4.28 m 
6.80 m 
Plannimetric 
5.28 m 
1.64 m 
5.28 m 
Along UTMX 
-4.54 m 
1.15 m 
4.68 m 
Along UTM Y 
2.43 m 
1.65 m 
2.93 m 
Table 1. Error figures for SRTM-image registration, for 121 
GCPs. IKONOS, (Hobart region, Australia) 
The errors presented in Table 1 are based on: 
i) SRTM plannimetric and height errors, 7.2m and 6.0m, 
respectively [Rodriguez 2005]. 
ii) RPC projection bias (horizontal: -4.1 pixels, vertical: 
-5.59 pixels) 
iii) Imperfection of quadratic interpolation 
As it can be observed from Table 1, geolocation errors are 
mainly systematic shifts, effective in the entire image. At least, 
they can be corrected by some systematic shifts. When the error 
vector is investigated, it can be observed that the error is 
consistent (e.g., in [-4m -5m] range for UTM X) over all GCP 
points. The only exceptions are the ones that are near the lower 
and upper lines of the image. In fact, this is a known 
phenomenon. 
RPC projection bias along u (vertical image axis) is 1.5 pixels 
(corresponding to ~1.5m) more than the bias along v (horizontal 
image axis). Axes u and v are almost aligned with UTM X and 
UTM Y, respectively. This difference is directly reflected to the 
error figures in Table 1. On the other hand, it is also noticeable 
that the errors are smaller than RPC bias. This result may be 
caused by another cancelling bias in SRTM errors. If SRTM 
data were error-free, the error figures would have been higher. 
This fact brings the conclusion that for another region, the bias 
terms may add up to result in higher error figures. 
The relations of geolocation error with terrain height, image 
row number and image line number are also examined to 
investigate the uniformity of the error behaviour. The results are 
given in Figures 2, 3 and 4. Note that, UTM X, UTM Y and 
Height are not correlated for our GCPs. As it is observed from 
these graphs, the errors in UTM X and UTM Y are not related 
to the position in the image, or terrain height, thus the 
behaviour is close to uniform. On the other hand, the errors 
along UTM X and UTM Y are quite (and inversely) correlated. 
The reason for this phenomenon is unknown and to be 
investigated. 
Geolocation error vs Height 
Height (m) 
Figure 2. The effect of height in UTM Y (blue) and UTM X 
(green) errors. 
Geolocaùon error vs image row Index 
Figure 3. The effect of image row in UTM Y (blue) and UTM X 
(green) errors. 
Geolocaon итог vs image column index 
Figure 4. The effect of image column in UTM Y (blue) and 
UTM X (green) errors. 
206