In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010 
204 
SRTM REGISTRATION FOR ELECTRO-OPTIC SATELLITE IMAGES WITHOUT GCP 
Yoldaç Ataseven ^ b *, A. Aydin Alatan“ 
Electrical and Electronics Engineering, M.E.T.U., 06531 Cankaya Ankara, Turkey (yoldas, alatan)@eee.metu.edu.tr 
b ASELS AN Inc., P.K. 1 Yenimahalle, Ankara, Turkey - yataseven@aselsan.com.tr 
Commission III, WG III/2 
KEY WORDS: Registration, SRTM, Stereo, interpolation, RPC , IKONOS 
ABSTRACT: 
This paper presents a simple methodology for accurate registration of the available elevation data sources, such as SRTM and 
ASTER DEM, to high resolution satellite images. For this purpose, widely used rational polynomial functions are utilized to project 
a point in the elevation data onto the satellite image. These projections are then interpolated in the image domain by using quadratic 
surface fitting. Possible hole points (empty pixels) are eliminated by using overlapping patches. The errors in SRTM data and 
satellite image geo-location are reduced by the help of tie points. The experimental results indicate better geo-location accuracy 
compared to the original satellite RPC's could provide. The proposed approach provides a good initial position estimate for each 
pixel in both stereo satellite images with no ground control points (GCPs), possibly resulting in fast convergence for expensive 
surface generation iterations and fast stereo match point detection. 
1. INTRODUCTION 
Modem Earth Observation Satellites (EOS) provide sub-meter 
resolution and images are in hundred mega-pixels range. The 
utilization of along-track stereo imaging provides massive 
amounts of data, which result in very expensive computations, 
especially for DTM/DSM generation. 
Sub-meter geo-location and elevation accuracies require 
rigorous satellite models (which are not made available by data 
providers), or non-linear projection information, namely 
rational polynomial coefficients (RPCs), that are provided by 
the images, and ground control points (GCPs) collected from 
the field of interest. 
Terrain reconstruction is often based on a convex optimization 
scheme, which optimizes the geodetic coordinates by 
minimizing the projection errors with respect to the stereo 
correspondences [Di 2001]. The common approach is to start 
the iterations with very coarse initial estimates, such as “average 
altitude of the region of interest”, which is reported to be 
sufficient from both convergence speed and accuracy points of 
view. However, this conclusion is empirical, since the 
projective relation defined by the RPC's is non-linear. Although 
the derivatives can be computed analytically, the error surfaces 
are quadratic only in the image coordinates, not in geodetic 
coordinates. 
Hence, an acceptable initial estimate should provide faster 
convergence and avoid possible trapping in false local minimas 
in RPC reconstruction. 
It is known that RPC coefficients of modem satellites are biased 
to yield a few meters of mislocation [Dial 2002a], The usage of 
GCPs is required for bias elimination. The geolocation error 
power is concentrated on the constant bias term; thus, even a 
single GCP might provide significant improvement in the 
geolocation accuracy [Dial 2002a, 2002b, Jacobsen 2008]. 
However, GCP collection is a time consuming and expensive 
task, which requires advanced devices for accurate 
measurements. 
On the other hand, there are widely known and freely available 
sources of terrain data, such as SRTM [Farr 2007] and ASTER 
DEM. Surprisingly, efforts to utilize these sources are limited 
[Arevaloa 2008, Gongalvez 2007, Papasaika 2008]. The 
common approach for such exploitation is to generate a sparse 
point cloud from the stereo images and then find the 
transformation that registers the point cloud to SRTM. When no 
GCPs are used, the errors in the projection functions remain and 
these errors are propagated to the sparse reconstruction, with 
possible amplification during the triangulation. A recent study 
[Gongalvez 2008] presented an approach to register SRTM with 
ALOS images without DEM generation. However, level 1B2 
images (which are already geo-referenced) are used for that 
purpose [Gongalvez 2008]. 
This paper presents a method to obtain good initial estimates for 
each pixel of a low level satellite image (only with radiometric 
corrections), by using freely available SRTM data. The method 
is then extended to utilize stereo images for an improvement in 
geolocation accuracy. 
The scope of this text is limited to registration (methodology 
and accuracy) only. The effect on reconstruction speed is left to 
another study. 
2. METHODOLOGY 
2.1.0bject-to-Image SRTM Registration 
The proposed algorithm uses the data vendor provided RPCs to 
perform the registration from the object domain to the image 
Corresponding author.