International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
AUTOMATED PROCEDURES FOR MULTISENSOR REGISTRATION AND ORTHORECTIFICATION OF
SATELLITE IMAGES
Ian Dowman* and Paul Daret
♦Department of Geomatic Engineering, University College London, Gower Street, London WC1E 6BT, idowman@ge.ucl.ac.uk
■¡■Formerly at UCL now at Department of Geomatics, University of Melbourne, Parkville, Victoria 3052, Australia
KEY WORDS: Automated registration, satellite data, image segmentation, polygon matching.
ABSTRACT
The need for automation in the registration of image to image and image to map is widely recognised and work has been going on for
some time in both the photogrammetric and remote sensing disciplines. The image to image registration problem is some way to
being solved for images from the same or similar sensors but there are considerable problems if the resolution of the images are
different or if optical and microwave sensors are involved. Strategies are needed which will identify common features in images with
different pixel size or radiometry and without dependency on initial orientation.
The basic techniques which are used are the extraction of line and area features from image and/or a map and then the matching of
these features. The polygon or patch is the main feature used. These are extracted from vector data by selecting features with an
appropriate attribute, for example lakes, forests or field boundaries. The polygons can be extracted from the images by thresholding,
homogeneous patch detection or segmentation. The paper describes studies to determine the best method for patch extraction for an
automatic system to match SAR and SPOT data and some new techniques for this extraction, including techniques for increasing
automation.
The basic techniques of matching polygons is adapted from Abbasi-Dezfouli and Freeman, (1994). Polygons are characterised by a
number of parameters such as shape and area. A revised approach is described here which does not use chain code and in which an
iterative approach is followed. Other developments from previous methods are also described and results show that registration of
complete satellite images can be achieved. Results from the image to map registration in the ARCHANGEL project are also
described.
1. INTRODUCTION
The term automated needs some clarification. The ultimate aim
of any digital system will often be to complete a process from
end to end without human interference. Such an aim cannot be
met at present and will not be met within the near future for
registration of images. However, human interaction can be
minimised and rationalised at a number of levels. We can make
individual algorithms or processes as automatic as possible and
only require human intervention to set parameters and check
output. We can also design strategies in such a way that some
decisions can be automatically taken, or at least set at the
beginning of the procedure. Both these aspects will be discussed
in the paper.
The paper will first review the requirements for image to image
registration and image to map registration, and set out the
overall strategies which have been adopted. The algorithms to
be used to extract features for matching will then be described
and choice of suitable algorithms discussed. The selected
algorithms will then be applied to polygon extraction and
techniques for this process will be described and discussed. The
techniques for polygon matching will then be presented and
finally some examples will be given.
2. STRATEGIES FOR IMAGE REGISTRATION
An image registration system needs to be as generic as possible
and we have worked towards a system which will accept any
type of image data as input. This includes data from optical
sensors, from microwave sensors and from vector data bases
(maps). There are differences in the types of feature which can
be extracted from data of different scale or resolution. Here, we
will concentrate on small scale data such as might be collected
from satellite sensors and be compatible with map scales of 1:10
000 to 1:50 000. Such an approach implies the setting of an
overall strategy and the provision of a toolbox of processes to
meet different requirements. The high level strategy is shown in
Figure 1.
Figure 1. A registration strategy.
Polygonal features are chosen as the prime feature for matching
because such features are found in all types of data. Obvious
examples are water bodies, forests and field boundaries. These
features can be found by segmenting images or will be extracted
