| 2004
—
I) are
arning
of the
dy by
vith a
> data
ers of
m the
m the
pixels
pixels,
ded to
f the
set of
id 1/3
pixels
to an
The
es that
ereby
1997;
ss can
efined
iched
d and
atasets
, used,
pochs.
isation
et al.,
it, the
| fit of
ver it
ase of
et al.,
or the
CTOSS-
juared
ishop,
on and
aining
:d and
> rank
~ulates
1s and
ted in
on the
ANNS
ata to
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
3.1. Classification accuracy of training and validation data
The trained ANNs were applied to classify the training data and
validation data of the training site. For both data sets overall
accuracies of the same magnitude were achieved for all
networks, independent of the number of hidden nodes.
Additionally the overall accuracies were found to be
comparable to the accuracies of the Maximum Likelihood
classification. The overall accuracies ranged from 57% to
77.4% for the training data and from 55.2% to 76.1% for the
validation data of the training site of all ANNs. The highest
overall accuracy of the validation data of the training site was
76.1%. This showed that the ANNs had resulted in a high
generalisation ability, enabling them to classify unseen data of
the same area as the training data to a high accuracy. On the
other hand the choice of random weights influenced the
performance of the neural networks more strongly than the
choice of hidden nodes (Figure 2 and 3). Differences in
accuracy of up to 15% of both training and validation data
(ANN F vs. ANN G) were found between ANNs consisting of
different initialising weights (Figure 2). Therefore different sets
of random weights should be applied to each ANN within one
study to determine the optimal ANN for the application.
Overall accuracy of 5-11-8 ANNs
100.00%
90.00%
80.00%
70.00%
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
{
B C D E F G | K L
@ Training data i Validation data Set of random weights
Overall accuracy (95)
f T 1 1 1
YU UTE : | 1
i
A
Figure 2. Overall classification accuracy of training (blue) and
validation (red) data of the training site for the ANN consisting
of 11 hidden nodes.
3.2. Overall accuracy of the classification of the remote site
The trained ANNs were applied to pixels from the remote site to
investigate the ability of the ANNs to classify unseen data. The
data of the remote site consisted of 384 mixed and unmixed
pixels, being of similar vegetation cover, but had not been
integrated in the training process. The averaged overall
accuracy of the remote site data for the three different ANN
designs was 21.6% (11), 19.1% (28) and 21.7%% (12 hidden
nodes) (Figure 3). This showed a significant decrease in
classification accuracy. The highest overall accuracy of the
transferability data was 27.3% for a ANN (ANN G) consisting
28 hidden nodes. However even the highest accuracy of the
remote site data achieved was far below expectations, indicating
a poor classification performance. The generalisation and
knowledge of the ANN gained at the training site was not
sufficient to classify pixels of a remote part of the image.
Modifications and improvements of the ANN classification
were therefore required.
Overall accuracy of 8 class-ANNs
10000 re to
90.00%
80 00% reden i tel Min eim treo act at i e e e te A e m d Aq
70.00% =
60.00% |-
50.00%
40.00%
30.00% ff
20.00% | |
10.00%
0.00%
Overall accuracy (%)
5-11-8 ANN 5-12-8 ANN 5-28-8 ANN
ANN design
El Training data (average)
0 Remote data (average)
li Validation data (average)
D Remote data (Maximum)
Figure 3. Averaged overall classification accuracy of training
(blue), validation (brown), remote site (yellow) data and
maximum overall classification (turquoise) for the remote site
data (light blue), shown for all ANN designs.
3.3. Adding a geographical label to the data
The ANN training process was repeated but with modifications
of the training data. Some pixels of the remote site (156 pixels)
were integrated in the training process, but being much lower in
number than the original training data. ^ Additionally a
geographical label was added as input band referring to the test
site of the pixel. The label of 1 was given to pixels of the
training site and 2 to pixels of the remote site. All ANNs were
retrained, partly consisting of a new number of hidden nodes
depending on the 6 input bands, applying again early stopping.
The overall accuracy of the training and validation data of the
training site was similar to the accuracies of the original training
process. The averaged overall accuracy ranged between 31.6%
and 78% for the training data and between 33.3% and 74.9% for
the validation data (Figure 4).
Overall accuracy of 8 class-ANNs trained with labels of
geographical location
100.000%
90.000%
80.000%
70.000%
__ 60.00%
50.000% |
40.000% +
30.000% E
20.0004 | 15 A. : :
10.000% +4 | Dis t [-
0.0009, L—— E Nu E
61384NN — ANN design 6:208 ANN
ii Validation data orig. (average)
O Transferability data all (average)
© Transferability data labelled (max.)
i
—
Overall accuracy
%
@ Training data (average)
0 Transferability data orig. (average)
| Transferability data labelled (average)
Figure 4. Averaged overall accuracy of the training (blue),
validation (brown), original remote site (yellow) data and
averaged accuracy (red) and maximum overall accuracy (green)
of remote site data, when training data included geographical
label.
A big improvement in classification accuracy was however
found for pixels of the remote test site. The overall accuracy of
the remote site ranged between 25.1% and 65.3% for the ANNs
trained with a geographical label. The highest accuracy of all
three ANN designs, being 64.9% (13 hidden nodes) and 65.3%
(28 hidden nodes), which was of the same magnitude as the
overall accuracy of the validation data of the training site. The
overall classification accuracies achieved using a geographical
label were of the same magnitude as other remote sensing
