ue xX ox V1 WD
al
ability and its merits compared to performing similar tasks
using a stand-alone neural network model.
Table 1 presents results demonstrating the effect of
altering the number of control points used in the hybrid
networks training phase, on the geo-referencing
precision.
GCPs Control Points (m) Check Points (m)
Used “dE Z dN RMSE | XZ dE 2 dN | RMSE
1 0 0 0 20491 | 23805 | 2861
2 -1 0 3 -965 1338 247
3 18 -14 39 -276 -216 104
4 37 46 61 -50 50 71
5 24 42 61 132 67 73
6 55 150 67 -94 65 69
7 -121 -68 64 -63 8 73
8 -180 -129 65 -140 -13 77
9 81 2 59 -152 11 77
10 70 -232 61 -31 -10 32
11 212 -71 58
Table 1 Hybrid Network Geo-referencing Results
Table 1 indicates that using only one GCP, for the geo-
referencing process, there is insufficient information
present in the single training pattern for the network to
establish a link between the rule-base estimate and the
true location of the training pattern. The introduction of a
second GCP into the training process produces a
significant improvement in the network's performance.
This extra GCP enables the link to be identified and the
hybrid network begins to function as an integrated
system. Through the addition of more control information,
the geo-referencing precision improves but reaches a
threshold when using between 4 and 9 control points
(RMSE ranges from 59 m to 67 m for the Control Points
and ranges from 69 m to 77 m for the Check Points). The
process does not improve or severely degrade when
using 4, 5, 6, 7, 8 or 9, demonstrating that the geo-
referencing function can be using fewer control points
than would be required by a stand alone neural network.
The RMSE fit to the control points in the final two tests
(i.e., using 10 and 11 GCPs) are in the same threshold
region (^60 m) as in the previous tests. However, as little
check point information is available for analysis these
results should not be considered for performance
evaluation, though they can be used in assessing the
hybrid network's learning ability.
5. SUMMARY OF RESULTS
There follows a summary of the results achieved using
the neural network approach (Figure 4), the rule-base
approach (Figure 5) and the integrated hybrid network
815
approach (Figure 6). The rule-base used for the tests in
the hybrid approach was that of the platform trajectory
model. Figure 4 presents the direction and magnitude of
the geo-referencing RMSE residual vectors within the
neural network process from a test that used 5 GCPs for
training 6 check points for recall. The direction of the
residual vectors associated with the GCPs appear
unrelated to one another, with no distinguishable pattern.
What is apparent, however, is the trend which exists in
the residual vectors associated with the 6 check points.
200000 220000 240000 260000 280000 300000 320000 340000
Eastings (m)
Figure 4 Residuals (neural network)
residual vectors @ scale 1:50
Figure 5 shows an example of the geo-referencing
residuals present in the Platform Trajectory Model (PTM)
approach to geo-referencing Earth Observation imagery.
The image was geo-referenced using a single GCP.
There is a clear trend in the directions of the residuals, in
the Easterly direction. This could be attributed to; pixel
dimensioning, reduction to the ellipsoid, Earth rotation or
atmospheric effects, all of which effect the image in an
along-track (Easterly) direction. Another distinctive
feature within the figure is the size of the residuals in the
bottom-right of the image as compared to those towards
the top of the image. The larger residuals can be
attributed to the propagation effects of the Platform
Trajectory Model.
OSGB Northings (m)
890000
200000 220000 240000 260000 280 000 320000 340000
Eastings (m)
Figure 5 Residuals (rule base)
residual vectors @ scale 1:50
Figure 6 illustrates the performance of the hybrid network
in the geo-referencing role. The plot contains the results
of using 5 GCPs and 6 check points. It can be seen from
Figure 6 that the resident systematic trends which were
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
