International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
called /raining the neural networks. Once the network is well
trained, it can be used to perform the image classification.
Input layer hidden layer Output layer
Figure 2. A backpropagation feed-forward multilayer neural
networks.
3. SIMULATED ANNEALING
Simulated Annealing (SA) [13, 14] will be used to enhance the
simple competitive learning neural networks. SA is an
optimization algorithm that is based on the process of annealing
metals. When a metal is heated up and slowly cooled down, it
is hardened into an optimal state. The analogy behind this is
that the algorithm begins the search for the optimal solution by
exploring many possible solutions [13, 14]. Slowly, it restricts
the search paths to only the most promising solutions as the
algorithm is said to be cooling down. The laws of
thermodynamics state that at temperature, /, the probability of
an increase in energy of magnitude, JE, is given by
P(3E) = exp(-SE /kf)
where & is a constant known as the Boltzmann's constant.
This equation is directly used in simulated annealing, although
it is usual to drop the Bolzmann's constant as this was only
introduced into the equation to cope with different materials.
Therefore, the probability of accepting a worse state is given by
the equation
P = exp(-c/t) » r
where
c = the change in the cost function
t = the current temperature
r = a random number between 0 and 1
The probability of accepting a worse move is a function of both
the temperature of the system and of the change in the cost
function. This approach allows SA to explore solutions that the
simple competitive learning networks might reject on its own.
Simulated annealing allows for some randomness in its search
for the optimal or near optimal solution.
Simulated annealing introduces some randomness into the
selection of clusters (categories). This releases the algorithm
from being trapped in a local optimum and allows it to venture
into a search for the global optimum.
4. EXPERIMENTS
The image shown in Figure 3 is one of tested images using
both competitive learning networks and backpropagation feed-
forward multilayer networks. The results from applying the
simple competitive learning networks, modified competitive
learning networks and backpropagation feed-forward multilayer
networks to the image, respectively and the results are shown in
Figures 4 — 6.
Figure 3. An original TM satellite image.
Figure 4. A classified image using the simple competitive
learning neural networks.
3
i
]
j|
E
|
|
dE
i
Ib
iL
p
T
|
Figure 5. A classified image using the modified competitive
learning neural networks.
Figure 6. A classified image using feed-forward multilayer
neural networks.
106
Internatio
Hte fn
The adva
priori kno
optimizat
by simul:
Like mos
be applie
classifica!
results of
results of
to classif
network 1
The preli
learning |
overall c
our exper
| Thoi
Clas
2 : Sche
Proc
Aca
Lille
and
4. Bish
Recc
S. Hert
the 1
6. Lipp
Net
7. Aha
o
Quai
8. Hun
Unsi
No.
9. Atlu
Arti]
Soil
Mar
‘99,
Mul
Usin
ASP
Nev
Pote
Usin
Eng:
129:
Som
an A
18,1
13. Alat
Opti
Algc
Unh
14. Batt
and
Task