In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part 3A - Saint-Mandé. France. Septentber 1-3. 2010 
3.2.SRTM Registration for Stereo IKONOS Images 
Bias compensation experiments are conducted in both image 
domain and the real world (object) domain. The results for 
object domain correction are better than pixel domain 
correction and in this paper, we provide only the results for the 
object domain correction algorithm. Since the errors (RPC bias) 
are mainly caused by satellite's own mislocation, the corrections 
in real world coordinates are more effective. 
In Table 2, geolocation and height errors are given after bias 
correction is achieved. For stereo image pair registration, the 
same region (IKONOS HOBART) is used during these 
simulations. 
Note that, the height error is larger than the SRTM height error. 
This is acceptable, since the registration error includes the joint 
effects of SRTM height error, RPC error and interpolation error. 
SRTM alone, does not provide any information about the 
satellite image without registration. 
Mean error 
terror 
RMS 
Latitude 
2.43x10' 5 ° 
1.53x1 O' 5 0 
2.873xl0‘ 5 0 
Longitude 
-1.76x10' 5 ° 
1.46x10’ 5 ° 
2.288xlO- 50 
Height 
-5.30 m 
4.27m 
6.8m 
Plannimetric 
error 
3.28 m 
1.71 m 
3.69 m 
Along UTMX 
-1.43m 
1.15m 
1.83 m 
Along UTM Y 
2.75m 
1.65 m 
3.21 m 
Table 2. Error figures for SRTM-image registration, after bias 
compensation using stereo pairs, for 121 GCPs. IKONOS, 
Hobart region, Australia 
The increase in the error along UTM Y hints the limits of the 
bias removal approach. 
It should be noted that the error figures in Table 2 are better 
than the RPC projection errors (~5m plannimetric error) . This 
result is remarkable, since our errors in SRTM registration 
include both RPC bias and SRTM geolocation error. The bias 
terms cancel each other for Hobart region. But they may add up 
to increase the error power and worsen the registration for some 
other region. Thus, a more thorough experimental study is 
required. 
4. CONCLUSION 
This paper presents a method to register high-resolution EOS 
images to SRTM data without GCPs. The method can be used 
for both stereo and single-image cases, providing better 
accuracy for the stereo case due to the utilization of stereo 
correspondences. Although the experiments presented in the 
paper cover only the IKONOS images, this methodology can be 
applied to other satellite images as long as object-to image 
projection information is provided. 
The SRTM sampling quantizes latitude and longitude, not the 
ground sample distance, and for the Hobart region, the 
longitude lines are closer to each other than the latitude lines 
are. On the other hand, IKONOS data is not in harmony with 
the geodetic coordinates. Since the IKONOS RPC coefficients 
are defined over the [latitude, longitude] grid and parse the 
earth surface from south to north while the Earth is rotating, the 
error figures for different regions may change. This change may 
be more severe along UTM X. 
An application of this method could be in stereo 
correspondence generation. Once the registration is performed, 
each pixel in one of the image pair can be mapped to its 
conjugate in the other image with only a few pixels of error. For 
the urban areas, this error may increase due to building heights 
or recent changes in the terrain. Even in this case, the region 
that the conjugate point lies is still bounded to a radius of at 
most a few ten pixels about the estimated point (for tall 
buildings) In such a small region, the search is fast. This 
provides significant processing time savings in stereo 
correspondence determination, which is a key and time 
consuming step of surface reconstruction. Our recent 
preliminary experiments agree with this conclusion, but are left 
to another text. 
Acceptable accuracy for registration is determined by the 
application. For example, if the registration output will be used 
for generating initial estimates for stereo (conjugate) pair points, 
the required accuracy is the radius of the search range of the 
stereo correspondence extraction algorithm. 
For image orthorectification, obviously, more accurate 
registration results in better orthorectification. 
If the registration output will be used as initial geodetic 
coordinate estimates for surface reconstruction, the registration 
output must reside inside the correct bowl of the error surface 
for the optimization cost function. At this point, it is difficult to 
determine what accuracy is required, since the complicated 
projection (rational polynomial model) functions results in very 
high order error surfaces and the projection function's 
coefficients (RPCs) are different for each scene. Thus, although 
very coarse initial estimates can provide sub-meter accuracies, it 
is preferable to generate the initial estimates as accurate as 
possible. Additionally, even if the accuracy of the initial 
estimates are unimportant for surface reconstruction, a good 
initial estimate will fasten the convergence of the reconstruction 
algorithm. 
This methodology may also be useful in applications that do not 
require sub-meter accuracies and recent changes in elevation 
(e.g., some cartography, civil engineering applications), without 
terrain reconstruction. 
This study should be considered as an initial step. A proper 
error analysis that accounts for the effects of SRTM and image 
accuracies is still required. 
REFERENCES 
Arevaloa,V., Gonzaleza, J., 2008, “An experimental evaluation 
of non-rigid registration techniques on Quickbird satellite 
imagery”, International Journal of Remote Sensing, Vol. 29, 
No. 2, pp.513-527 
Bouguet, J., 2000, "Pyramidal Implementation of the Lucas- 
Kanade Feature Tracker: Description of the Algorithm", 
207