DIGITAL SURFACE MODELLING FROM SPOT 5 HRS IMAGERY USING THE AFFINE
PROJECTIVE MODEL
C. S. Fraser ?, P.M. Dare ?, T. Yamakawa *
a : S ^A. . : . ” ; 3 s
Department of Geomatics, University of Melbourne, Victoria 3010, AUSTRALIA
c.fraser@unimelb.edu.au; yamakawa@sunrise.sli.unimelb.edu.au
^ Airborne Research Australia, Flinders University, PO Box 335, Salisbury South, SA 5016, AUSTRALIA
Paul.Dare@AirborneResearch.com.au
Commission I, WG 1/2
KEYWORDS: SPOT, DEM/DTM, high resolution, matching, accuracy
ABSTRACT:
This paper presents the results of work carried out under the auspices of the joint ISPRS CNES Scientific Assessment Program of
DEM generation from SPOT 5 high resolution stereoscopic (HRS) image data. À stereo pair of HRS images of Melbourne, Australia,
with a ground sample distance of 5m in the along track direction, and 10m in the across track direction, was used to create digital
surface models of varying grid spacings of the urban area surrounding the city of Melbourne. Conjugate points from the SPOT
stereopair were located using intensity based image matching, and triangulated using the affine projective model. The resultant
digital surface models were compared to precisely located ground control points and a reference digital terrain model. The results
showed that high quality digital surface models can be produced from HRS image data using the algorithms described. The
comparison between the surface model and ground control points revealed an RMS height error of 3-5m or 0.6-1 pixel.
1. INTRODUCTION
The purpose of the SPOT 5 Scientific Assessment Program
(SAP) is to evaluate the utility of SPOT 5 high resolution
stereoscopic (HRS) imagery for generating digital terrain
models (DTMs) (Baudoin et al., 2004). Under this program,
tests have been carried out by many different research teams on
SPOT 5 HRS data for many different regions of the world. The
study described in this paper has made use of data acquired
over the city of Melbourne, Australia. Since the research teams
involved in the SAP are independent, a range of different
processing techniques have been brought to bear on the issue of
extracting digital surface models (DSMs) from SPOT 5 HRS
imagerv.
Previous research carried out at the University of Melbourne on
the subject of geometric modelling of high-resolution
spaceborne sensors has focused on the use of ‘alternative’
mathematical models, such as the rational polynomial
coefficient (RPC) model (Fraser et al., 2002; Hanley et al.,
2002) and the affine projective model (Yamakawa et al., 2002,
Fraser ‘and Yamakawa, 2004). The success of these studies has
led to these models being incorporated into the digital surface
modelling algorithms created specifically for high-resolution
satellite imaging sensors, such as IKONOS and QuickBird.
Results of studies utilising stereopairs of IKONOS images have
shown that both the RPC model and affine projective model can
be used to create high quality DSMs (Dare, 2004; Dare and
Fraser, 2004).
Since SPOT 5 HRS data is not supplied with RPCs, the
research presented in this paper concentrates on the use of the
affine projective model. Note that traditional photogrammetric
techniques can also be used to generate surface models from
SPOT 5 data, but such investigations have been left to other
members of the SAP.
2. SPOT 5 DATA
The data used in this study covered an area of roughly 140km
by 50knf around the city of Melbourne, Australia (Figure 1).
The image was acquired at 10:20am local time on February
19^. 2003. The land cover types in the scene include water,
urban areas, agricultural land and forests. The elevation varies
from sea level to around 600m above mean sea level. A portion
of the western side of the region was partially obscured by
cloud cover during image acquisition, but this made little
difference to the generation of the surface models in this study,
since these were centred on other areas within the scene.
The pixel size of HRS imagery is 5m in the along track
direction, and 10m across track. Since the stereo images were
acquired in the same orbit in a fore and aft configuration, the
parallax differences occur in the along track direction.
Therefore, the along-track pixel size (5m) is relevant when
discussing the accuracy of surface models with respect to the
pixel size of the original images.
3. AFFINE PROJECTIVE MODEL
Since the launch of the first high-resolution satellites, and even
before, much attention has been focused on the use of
alternative geometric models, especially RPCs, for restitution,
orthorectification and terrain modelling (Dowman and Dolloff.
2000; Dial, 2000; Grodecki, 2001; Grodecki and Dial, 2001).
Significant research has also been carried out on other models,
such as the direct linear transform (DLT) model and the affine
projective (AP) model. An inter-comparison of these three
models (RPC, DLT and AP) with IKONOS data showed that
although the degree of complexity of the models differs widely,
differences in the results can be expected to be small (Fraser et
al., 2002, Hanley et al., 2002).