lifferences
activities
daries for
uld differ
£e. from a
the same
les.
'OSt-object
Insterdam
est result,
ased post
hand side)
likelihood
> previous
and rules
d control
IS stage as
d at same
3,2 Neural network in pixel-wise classification
The main advantages with the neural network as a
classifier, for this project was the ability to handle
complex data patterns as it is nonlinear. In this case the
complexity in the data patterns concerned the variation
between the image channels on one hand versus the slope
and aspect channels on the other hand. For some of the
subclasses it has been noticed that signatures for the near-
infrared channel and slope and aspect channels have
spread and complex patterns, which had not been solved
by the subclass division.
The same subclasses of the training fields that were used
with the maximum likelihood classifier were used to train
the neural network. The values for the 5 input channels
were normalized to a range between 0 and 1 in order to
speed convergence to the minimum error point in the
network. An equal number of pixels in each class would
have given the Neural network its recommended balanced
data sets. But because of the comparison of the methods
exactly the same data sets are used.
Several tests with different combinations of
learning-factors number of nodes in the hidden layer were
conducted in the standard back-propagation training
process. The settings of the variables are plausible and the
guidelines from other projects only partly gave the same
results. When it comes to the learning factor there are
different recommendations in the literature. We got best
convergence with a decreasing learning factor from
approximately 4 to 0.1 for net with moderate number
nodes in the hidden layer. But for larger net with around
100 hidden nodes a stable low (0.5) learning factor seems
to give the best training result. Among others Skidmore et
al. 1994 pointed out that even if the percentage of
correctly classified training patterns has a slight tendency
to grow as the number of hidden nodes is increased.
There is an almost opposite tendency for the percentage
of correctly classified test data that indicates an optimal
number of hidden nodes as being quite low.
À "program" called Stuttgart Neural Network Simulator
(SNNS) was used to conduct the neural network
classification . SNNS is distributed by the University of
Stuttgart as "Free Software’, for more information see the
User Manual [Zell et al. 1994]. A three-layer feed-forward
net with 5 input nodes, 100 nodes in the hidden layer and
10 output nodes was found to give the best training results
and thereby used in the classification.
32.1 Training results from Neural Networks. Dalen
1995 has done some tests with the same training data sets
4s used in this project. He got an overall training result of
86% and for the individual classes a result from 65-95%.
This best result was achieved with a network with 20
hidden nodes. He used 1200 iterations where the learning
factor decreased from 4.0 down to 0.1. For his limited test
545
data sets the overall results were 1246 and the total of all
stands in the individual classes ranges from 0-51%
correctly classified. In these data sets there is no
separation into areas with low and medium-high site
classes as been done in section 4.
In this project the same training data sets were used, but
for the test data all the coniferous stands in the control
area were used. For nets with moderate number of nodes
in the hidden layer the best result was for a net with 25
hidden nodes. The result for this net was 86% correctly
trained when only the highest "score" counts (winner
takes all). This was achieved after 1050 iterations with a
decreasing learning factor.
As complex signature patterns necessitated the
examination of neural networks, larger nets with more
hidden nodes were also tested. The best training results
was achieved after approximately 10000 iterations and 45
hours on a "Sun Sparc-station LX". The training-results
was over 91% correctly trained (winner takes all) or over
87% if the 40-20-40 method is used in the training
analysis.
4. RESULTS AND DISCUSSIONS
First the results from phase 1 with the maximum-
likelihood classifier in the post-object classification are
presented. The coniferous stands in the control area have
to be divided into two groups to get any meaningful
analysis at all. In one group that consists of areas with
mainly low-productive stands (low site classes), hardly
any of the stands were correctly classified. The spectral
reflections in these stands are totally dominated by the
background vegetation and the trees are too spread to
make a dominating influence on the reflection signals.
The other group consists of stands with medium-high site
classes and here the results are better. The total
classification accuracy in this area is approximately 50%
in the correct class, 30% in the correct super-class and
2096 wrongly classified after the post-object classification.
Among the stands that were not correctly classified are
almost all the small stands. The rest of the wrongly
classified stands are dominated by “spectral”
heterogeneous stands. Many of these have physical
variations in topography, crown coverage and/or amount
of broad-leaf trees.
The pixel-based classification results are given here for
two interesting groups of stands: homogeneous and
heterogenous stands in the area dominated by medium-
high site classes. For the homogeneous stands 70-95% are
correctly classified and for the problematic heterogeneous
stands they vary from 5 to 30%. These heterogeneous
stands with complex signature patterns (see Section 3.2)
were the direct reason for phase 2.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
