International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
iterative method, corrections to weight parameters are
computed and added to previous values as below:
GE
OW, ,
fod (Eq. 2)
Aw, ; =—n
A. (t T » = Aw, , * 4^w, , (r)
^
Where "ij is weight parameter between neurode i and j, 7] is
a positive constant that controls the amount of adjustment and is
called learning rate, C is à momentum factor that can take on
values between 0 and 1 and “t” denotes the iteration number.
The parameter & can be called smoothing or stabilizing factor
since it smoothes the rapid changes between the weights (Yang,
1995).
Recalling refers how the network can operate based on what it
has learned in the training stage. It is actually using the trained
network for interpolation and extrapolation and is called
generalizing either.
One of the most important advantages of neural networks
with respect to conventional statistical methods is that they are
distribution-free operators because the learning and recalling
depend on the linear combination of data pattern instead of the
statistical parameters of the input data (Civco and Waug, 1994).
Neural networks are also highly capable to deal with multi-
source data because they do not require explicit modelling of
the data from different sources and therefore there is no need to
treat them independently as in many statistical methods
(Benediktsson and Swain, 1990). They also avoid the problem
in statistical multi-source analysis of specifying how much
influence each data source should have in classification
(Benediktsson and Swain, 1990).
3. NETWORK DESIGN
Road detection from satellite images can be considered as a
classification process in which pixels are divided into road and
background classes. Recent researches have shown ANNs to be
capable of pattern recognition and classification of image data.
For using neural networks in road detection, input layer is
consisted of neurodes the same number as input parameters and
output layer is made up of just one neurode that shows the
belief of network whether the input parameters can represent a
road pixel or not. Usually one hidden layer is sufficient,
although the number of neurodes in the hidden layer is often not
readily determined (Richard, 1993). More neurodes in hidden
layer enables the network to learn more complicated problems,
but there would be an associated increase in training time
(Foody, 1995).
The most important factor in employing ANNs is to decide
what type of information should be extracted from input image
to be fed through the network as its input parameters. The
discrimination ability of the network is highly affected by
chosen input parameters. Roads are presented as linear features
in low resolution images while they are displayed as
homogeneous areas in high resolution satellite images like
QuickBird and Ikonos. For that reason, input parameters for
road detection from low resolution images should be able to
distinguish shape patterns while for high resolution images
homogeneity and spectral characteristics seem more important
to be emphasized.
The learning rate and momentum parameters have a major
influence on the success of learning process and they should be
defined by user in advance. This assignment is problem
dependent.
Another factor that affects network ability is the number of
iterations done in training stage. If the network is trained more
than what is needed, training samples will be memorized by the
network that reduces the discrimination ability of the network.
Therefore termination conditions should be assigned accurately
to avoid over-training problem.
Accordingly network design, consisted of defining hidden
: : : 7 ;
layer size, input parameters selection, 7 and o assignment
and termination conditions, is a crucial stage that must be
performed before applying neural networks.
4. METHODOLOGIES
As a case study a part of an RGB Ikonos image with the size of
550*550 pixels from Kish Island in Iran is chosen which is
enhanced with linear function. Figure 2 shows the original
image and its manually produced reference map which is used
in accuracy assessment.
(b)
Figure (2) a) RGB Ikonos Image from Kish Island in Iran. b)
Manually produced reference map.
Internal
LS
A loca
heights
A syst
system:
6. El
We in
terrain,
referen
charact
DEM.
the si
calcula
were ar
6.1 St
The fir:
include
relief f
east. T
contain
respons
6.2 El
ERS el
retrack:
elevati
30-m L
photogi
Spot, o
A root
quality
coordin
We ext
bilinear
ERS sir
we use
method
160
|
|
155
heigth (mi
6.3.1
Fig 7 sl
seem si
show a
