International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
topology for MLP. The process continued until a complicated
network with five input nodes, two hidden layers with five and
four neurons respectively and an output layer with one neuron
were created. The nature of RBF network requires a strictly
three-layer network. A two input initial network was selected
but without luck as the training algorithm had relatively good
performance with 14 RBFs, which makes the classification
procedure very complex. Three input topology was better suited
as initial network with the hidden layer of 3 RBFs and 2 output
nodes. From the process, a complicated network with 13 RBFs
was constructed. Network performances can be viewed to next
session of the paper.
A classified image identifying the presence of oil spill was
produced for every network topology. For each image produced
a comparison with a reference dataset was made. The reference
dataset was produced by photo-interpretation methods and
techniques (Topouzelis et al, 2002). A comparison was made
using confusion matrices. For each image produced the
confusion matrix and overall accuracy were calculated. Overall
accuracy was calculated by dividing the total number of the
pixels correctly classified by the total number of the pixels of a
sample.
5. EXPERIMENTAL RESULTS
Figures 6 and 7 show results of dif! ferent network topologies in
terms of input nodes for MLP and RBF respectively. For MLP
it can be seen that topologies with one (original SAR image) or
with two input nodes do not classify the image correctly. On the
contary, topologies using more than.two input nodes have much
better performance. If we concentrate in the latter, we can see
that there is a special performance to topologies containing 7
nodes. Also, if we investigate the performance of accuracy due
to neurons we can assume that the topologies having the better
performance are 3:3:1, 4:2:1, S:1:1 (Table 1).
Network - Accuracy
Topology Yo
MPL - 3:3:1 99.37
MLP - 4:2:1 99.43
MLP - 5:1:1 99.49
RBF - 3:2:2 98.45
RBF - 3:32 98.48
RBF - 3:4:2 98.19
RBF - 4:2:2 99.08
Table 1. Best network accuracies
Due to the fact that for each input node the size of the original
data increases by 2096 and the needs of computational time and
power are significantly increased, the topology that is proposed
is 4:2:1. The accuracy of this topology is strongly connected
with the specific inputs as these are presented in paragraph 4.2.
Four input nodes topology. where the inputs were the
predetermined images, was proved the appropriate topology for
classifying the oil spill very accurately. Looking at the five
input nodes topology in details, we can see that the information
in the borders of the oil spill was lost. Furthermore, more
computational power and training time were needed.
For RBF we observe that their performance is poorer than MPL.
It starts nearly from 4596 while only four topologies can be
compared with the majority of MLP, which are above 95%
(small window in Figure 8 — Table 1). Topology with 3 input
units is the only with relative stability close to 98.5 and is
bounded from 7 to 9 nodes (3:2:2, 3:3:2, 3:4:2, 4:2:2). Better
performance was observed for 4:2:2 topology with 99.08%
while the performance of all the other topologies is under 99%.
From the above it can be concluded that the MLP network has
better performance in oil spill detection than the RBF network.
MLP
|
100 Nodes-Accuracy |
|
99,5 |
X
99 |
|
s 985 a |
g . i
3 s9- linput |
S o5 e —a—2input |
€
97.5 e— Sinput |
—e— input |
97 eese Sinput |
96,5 : =
1 2 3 4 5 6 8 9 0 4 49 DQ49— 4 71848 |
Nodes
Figure 6. MLP overall accuracy according to nodes
RBF 10
100 Nodes-Accuracy 99.5
99 |
98.5 |
90 98 |
97.5 |
* 97 |
80 ^ 96.5
I ^ 7 8 9 10
g
370 °
< :
—4—- 2 input
eo —^—- 3 input
0 4 input
--a--5 input
40
6 7 8 9 160. M 2 13 14 45 16 qq queo
Nodes
Figure 7. RBF overall accuracy according to nodes
For evaluation reasons, the method was tested on a broader
area. Figure 8 contains a SAR image window with several oil
spills. Oil spills can easily be identified but it is extremely
difficult, even for an expert, to specify the border of oil spill
and sea. Classification was performed using the MLP — 4:2:1
topology. The classification total accuracy was 99,291%. There
was a very good discrimination between sea and oil spill but
some of the linearities were lost, especially in cases that oil spill
covers very thin areas.
6. CONCLUSIONS
In this study a neural network approach for oil spill
identification was investigated using SAR images. Two types of
neural network were used: the feed forward Multilayer
Perceptron and Radial Basis Function. Original images and
other images generated from them were used as inputs to a
neural network. The method was tested on SAR image
windows, containing oil spills and lookalikes.
In general, RBF networks work faster than LMP. RBF almost
guarantee convergence while MLP some times stick in local
minima. The use of a hybrid algorithm of backpropagation and
728
Interr
conju
MLP
and h
From
more
perforr
while
using t
classifi
specifi
above
classifi
99,609
and are
best pe
the mo
3:427
Further
differei
the pei
Vector
Hopfie
REFEI
Alpers,
N., Mo
waves
multi-fi
IGARS
Anne, |
1999, A
IEEE T
pp. 191
