1287
PRECISE GEOREFERENCING OF CARTOSAT IMAGERY VIA DIFFERENT
ORIENTATION MODELS
J. Willneff, T. Weser, F. Rottensteiner, C. S. Fraser *
Cooperative Research Centre for Spatial Information, Department of Geomatics, The University of Melbourne,
VIC ЗОЮ, Australia-jwillneff@web.de, (tweser, franzr, c.fraser)@unimelb.edu.au
Commission I, SS-11
KEY WORDS: CARTOSAT, high-resolution, satellite, orientation, comparison
ABSTRACT:
The CARTOSAT 1 satellite, launched by the Indian Space Research Organisation (ISRO) in 2005, can provide panchromatic along-
track stereo imagery with a ground resolution of 2.5 m. Along with the imagery, encrypted files with rational polynomial
coefficients (RPCs) and meta-data are distributed by ISRO. The RPCs allow direct georeferencing within certain limits depending
on the on-board systems for registering the orbit path and attitudes of the satellite. At the Cooperative Research Centre for Spatial
Information at the University of Melbourne (Australia), the software package Barista for the processing of high-resolution satellite
images is being developed. Barista offers three techniques for precise georeferencing of such image data, namely the 3D affine
model, bias correction for RPCs, and a generic pushbroom sensor model. The 3D affine model can only be applied when ground
control points (GCPs) are available. The RPC model can be improved beyond the limits of direct georeferencing by correcting for
the biases contained in the original RPCs. This process requires at least one well-defined GCP per image. Whereas the meta-data for
CARTOSAT 1 imagery do not contain all the information required for using the generic pushbroom sensor model for direct
georeferencing, they provide initial values for such a sensor model to be determined if enough GCPs are available. In this paper, the
authors compare the geopositioning accuracy achievable with CARTOSAT 1 imagery via the 3D affine, bias-corrected RPC and
generic pushbroom sensor models. A stereo pair of images covering Hobart, Australia, was processed using Barista. In addition to
the imagery, an object point array of altogether 69 3D GPS-surveyed points was utilised. They were distributed all over Hobart and
covered about one quarter of the scene. In order to assess the georeferencing accuracy that can be achieved using CARTOSAT 1
images, bundle adjustment was carried out using all three sensor models and nine well-distributed GCPs. The absolute accuracy was
then assessed via the remaining 60 points, which served as independent checkpoints. The georeferencing results obtained for
CARTOSAT 1 in the Hobart test field are very encouraging. Whereas direct georeferencing using the RPCs provided by ISRO
yielded sub-optimal results, the provision of a small number of GCPs is enough to boost the positioning accuracy to subpixel level in
planimetry and to make it slightly better than 1 pixel in height, independent from the sensor model used.
1. INTRODUCTION
The number of commercial high-resolution satellites has
steadily increased over recent years. The available image
products from the current operational satellites have a ground
resolution from a few metres to half a metre and are used for the
extraction of spatial information for a variety of mapping and
GIS applications. For the extraction of metric information from
images, suitable sensor orientation models describing the
relationship between image space and object space are
necessary. This paper describes the application of three
different sensor models and the assessment of their accuracy
from processing a data set acquired with the CARTOSAT 1
satellite. Launched in 2005 by the Indian Space Research
Organisation (ISRO), CARTOSAT 1 can provide panchromatic
along-track stereo imagery with a ground resolution of 2.5 m.
The radiometric resolution is 10 bit. One option for sensor
orientation of CARTOSAT 1 images is a camera replacement
model using rational polynomial coefficients (RPCs),
comprehensively described in Grodecki and Dial (2001) and
Tao and Hu (2002). Along with the imagery, encrypted files
with RPCs and meta-data are distributed by ISRO. Certified
software distributers will additionally be provided with
decryption software for these RPCs so that they can be accessed
and used for direct georeferencing of the imagery within certain
limits. These limits depend on the accuracy of the on-board
systems for registering the orbit path and attitudes of the
satellite in order to generate the RPCs from a generic
pushbroom scanner model. In the case of CARTOSAT 1, the
RPCs provided by ISRO should allow direct georeferencing
with an accuracy of about 30 pixels. As the provided metadata
contain neither a precise camera calibration nor information
about orbit path and attitude angles, direct georeferencing is not
possible with a generic pushbroom sensor model.
At the Cooperative Research Centre for Spatial Information
(CRC-SI) at the University of Melbourne, Australia, the
software package Barista for the processing of high-resolution
satellite images is being developed to support various research
initiatives dealing with the extraction of spatial information.
Barista offers three techniques for precise georeferencing of
satellite images and all require ground control points (GCPs).
First, the 3D affine model has been shown in the past to deliver
pixel-level results for high-resolution satellite imagery in scenes
of limited extent. In this case, no additional information is
required from the vendor of the satellite images. Second, bias
corrections can be applied to the RPCs provided by ISRO if
they can be accessed. These bias corrections can be modelled
Corresponding author.