International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
from the images and unwanted information is removed before 
processing takes place. This method gives root mean square 
errors in the order of 1 - 2 pixels for the Istres data. Tests were 
carried out with various edge detectors with little variation 
shown to the final results. Between 1000 and 1700 points were 
extracted from the Istres scene. The Sobel operator was the most 
convenient because it is the only one tested which gives both the 
edge location and edge strength which is needed for the 
dynamic programming algorithm. The points extracted were 
used to refine the registration and better results were obtained. 
Figure 8. Matched patches from SAR and SPOT images. 
6. AUTOMATIC REGISTRATION OF FULL SCENE 
IMAGES 
The tests described above have been carried out on small scenes 
of about 500 x 500 pixels. These have been chosen to include a 
number of suitable features for matching and to be flat areas so 
that relief distortion does not need to be taken into account. In 
order to demonstrate that the method is more widely applicable, 
a test was carried out on registration of full SAR and SPOT 
scenes. The area used was the same as the one for the subscenes. 
The technique of using image pyramids was tested but found to 
give rather poor results. An alternative method is proposed 
which is based on using image tiles. The full scene images are 
approximately aligned using either ephemeris data, or with 3 or 
4 manually selected tie points and then split into tiles. The tiles 
can be selected automatically to give a full distribution over the 
image, or they can be selected manually to ensure that tiles with 
good features for matching are chosen. Each selected tile is 
processed in the manner described above using patch matching 
and edge matching and the tie points which are generated are 
used to transform the whole scene. 
The method was tested on the SAR and SPOT scenes of SW 
France. Manually tie pointing gave a good initial registration. 
Twelve 512 x 512 tiles were selected for the refined registration. 
Of the 12 ten produced correctly matched patches and these 
resulted in 39 matches across the whole image. These were used 
to register the images using an affine transformation with a 
resultant root mean square error of 16 pixels and 13 pixels 
derived from two sets of tie points split between control points 
and check points and then reversed. The error was much greater 
in the x direction than in the y direction, probably reflecting the 
inadequacy of the affine transformation for this operation. The 
tie points were nearly all located on water features and most 
points had an elevation of less than 50m. 
Edge matching was used to refine the matching. 3488 tie points 
were generated and the root mean square errors on the two 
groups of tie points were 11 pixels in both cases. This is clearly 
an improvement but the larger residuals in the x direction 
remain. 
The test has clearly shown that near automatic registration of 
whole scene images is possible using the techniques described 
in this paper. Since the paper is mainly concerned with 
techniques for tie point generation, no attention has been paid to 
selecting the most suitable model for transformation when two 
different types of image are used. Selection of a suitable model 
will clearly improve the results of the registration. 
7. CONCLUSIONS 
The work described in this paper has indicated that registration 
of data of different types is possible using automatic extraction 
and matching of polygons and that edge matching can be used to 
refine the process. 
The ARCHANGEL project, which is not discussed in detail 
here, has shown what can be done with images and vector data 
and results are presented which demonstrate the potential of this 
method. 
Similar techniques for patch extraction from images have been 
discussed in the context of matching SAR and SPOT data and 
have been shown to be effective both on subscenes and on full 
scene images. Data of only one area has been used but 
experience suggest that the requirement for only a few well 
distributed points over a scene should be possible in many areas 
and conditions in the world. 
This work has been done by the second author towards his PhD 
thesis (Dare, 1999) and he has identified many improvements 
which could be made to the methods discussed in the paper and 
also additional techniques which could be used. As regards the 
completion of an accurate end-to-end system, the most 
important of these is probably the use of a suitable 
transformation model. If a DEM is available, then geocoding of 
both images would be possible using the initial ephemeris of 
ERS for example to establish a reference system and then to 
register the SPOT data to the geocoded SAR data. Renouard and 
Perlant (1993) have described the basics of such a method. 
Alternatively, the ARCHANGEL method could be used, if 
maps and a DEM are available. 
ACKNOWLEDGEMENTS 
The work on ARCHANGEL was carried out as an EC Fourth 
Framework project Environment and Climate Programme, 
Theme 3 Grant No. ENV-CT96-0306 (DG12-DTEE). The work 
on registration of SAR and SPOT was carried out as a PhD 
project supported by NERC, Grant No. GT4/95/207D. 
SAR data was provided by ESA, and SPOT data was provided 
by SPOT Image for an OEEPE project on Aerial Triangulation 
of SPOT data.