ANALYSIS OF RULE-BASES IN HYBRID NEURAL NETWORKS FOR GEO-REFERENCING
EARTH OBSERVATION IMAGERY
Martin J Smith and Mark Dumville
Institute of Engineering Surveying and Space Geodesy
The University of Nottingham, University Park
NOTTINGHAM NG72RD UK
Tel 0115 951 3880 Fax 0115 951 3881
email: martin.smith@nottingham.ac.uk
Commission IV Working Group 2
KEY WORDS: processing, neural networks, image rectification, geo-referencing
ABSTRACT
The paper evaluates the performance of using both a neural network and a rule-base in developing a fully integrated
geo-referencing routine. Aspects of hybrid neural network
times, improved geo-referencing precisions and overcomi
network processing models.
1. INTRODUCTION
A hybrid neural network is a network whose architecture
consists of two or more separate processing structures.
Hybrid neural networks consist of a neural network
augmented with another processing structure (rule-base)
which can be included within the system by running
parallel to it, or in series with it. In addition to providing a
more precise geo-referencing system, the hybrid
architecture is highly parallel and ideally suited to parallel
processing producing a highly effective and
computationally efficient system.
1.1 Hybrid Neural Networks
The hybrid network technique adopted for this study is
similar to that presented by Bengio et al. (1992). Hidden
Markov Models have a proven success in the modelling
of the temporal structure of speech, whereas the artificial
neural network (and in particular the multi-layer
perceptron) has a proven success in continuous function
approximation. The system used by Bengio et al. (ibid)
combines the advantages of the two independent
techniques. A similar approach was used by Bumiston
(1994), consisting of a simple rule-based model to
approximate the speech pattern. The multi-layer
perceptron neural network was then used to identify and
model the peculiarities and fine detail of the speech. This
form of network construction removes the requirement for
the network to learn the complete function. This allows
the network to focus its ability on recognising the patterns
which exist in the difference between the rule-based
estimate and the true geographic coordinates of the
training patterns. This is achieved in a learning process
similar to that used in conventional neural network
training algorithms (e.g, back-error propagation). The
integration method adopted was through the use of a
performance are reviewed, focussing on improved training
ng convergence problems associated with stand-alone neural
813
rule-base to assist the removal of the majority of the
systematic errors within the geo-referencing process
(Dumville, 1995).
1.2 The Rule Base
The Platform Trajectory Model combines satellite
ephemeral information with ground control to create the
geo-referencing model. Unlike popular geometric
rectification algorithms the Platform Trajectory Model only
requires a single ground control point (GCP). This point is
used to anchor the image to the cartesian reference
system to be used within the geo-referencing routine. The
satellite ephemeris acts as the control for the orientation
of the image and the timing information is used for the
scaling of the image pixels. The ephemeral and timing
information is made available within the header files of
the satellite image.
2. THE TEST IMAGE
A Synthetic Aperture Radar (SAR) image of the North of
Scotland from the European Remote Sensing Satellite,
ERS-1, was used in this study. The image contains 8000
by 8000 pixels corresponding to a ground area of 100 km
by 100 km. This area was selected as a test site due to
the availability of the appropriate satellite imagery and the
mapping of the region.
A relative GPS ground survey was carried to obtain
ground control for geo-referencing the image. A set of
eleven control points were observed after by identification
of suitable features on the image (Putter, 1993).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
