2012
rn and local
>s was done
while the
the look-up
of original
correspond
of objects
els, provide
rs extracted
tionship can
mote better
'sponses. In
others was
eCognition
omentation"
ept (OBIA)
factor was
tures in the
T-2). After
is done by
archy. For
lue of each
ing process
ammed in a
objects was
geometric
y), physical
n or max
f other dark
nd route of
t al., 2006,
ck textural
>nce matrix
ular second
ropy, mean
001).
hape of the
riant planar
1e standard
s belonging
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
to the region of interest, selected by the user surrounding
the object.
- ConLa (Local area contrast ratio) is the ratio between the
mean backscatter value of the object and the mean
backscatter value of a window centred at the region.
- Opm (Object power to mean ratio) is the ratio between
standard deviation of the object and the mean of the
object.
- Opm/Bpm (Power to mean ratio) is the ratio between the
object power to mean ratio and the background power to
mean ratio.
- OSd (Object standard deviation) is the standard deviation
of the object.
- P/A (Perimeter to area ratio) which is the ratio between the
perimeter (P) and the area (A) of the object.
- C (Object complexity) describes how simple (or complex)
the geometrical objects.
- THm (Mean Haralick texture) is computed based on the
average of the grey level cooccurrence matrices of the
sub-objects.
In this study SP2, P/A, and C were used for geometrical
characteristics; BSd, ConLa, Opm, Opm/Bpm, and OSd were
used for physical characteristics, and THm was used for
textural characteristic (Topouzelis et al., 2009, Ozkan et al.,
2011). Totally 7 oil objects from RADARSAT-2 and 8 oil
objects from ALOS PALSAR were extracted. All of these
objects are used for the testing purpose. The statistics of oil
objects are given in Table 2.
Table 2. Statistics of features obtained from oil objects.
Features E RADARSAT-2
Minimum | Maximum | Mean Std.
SP2 0.172 1 0.67 0.269
BSd 35.583 52.582 47.382 6.153
ConLa 0.334 0.544 0.455 0.085
Opm 0.37 0.697 0.55 0.136
Opm/Bpm 1.11 1.551 1.383 0.161
OSd 16.365 34.026 25.827 6.98
P/A 0.076 0.293 0.182 0.067
C 1.543 5.442 3.628 1.512
THm 5.622 23.371 13.056 6.922
ALOS PALSAR
Minimum | Maximum | Mean Std.
SP2 0.319 1 0.719 0.196
BSd 148.546 785.317 | 388.672 | 218.870
ConLa 0.590 0.769 0.688 0.066
Opm 0.124 0.212 0.148 0.030
Opm/Bpm 0.260 1.132 0.679 0.364
OSd 104.009 182.516 135.732 |. 34.523
P/A 0.080 0.573 0.367 0.198
C 1.377 2.524 2.167 0.358
THm 1.092 17.010 5.487 6.922
— Classification
Artificial neural networks are computational systems based
on the principles of biological neural systems, ie. it is a
mathematical model composed of many neurons operating in
parallel. These networks have the capacity to learn, memorize
and create relationships amongst data. They have some
advantages such as their non-parametric and non-linear
nature, arbitrary decision boundary capabilities, easy
adaptation to different types of data and input structures, and
good generalization capabilities over classical statistic and
analytic approaches. Although the network design as a
69
classifier is a hard task despite the increment in the
performance of classification, an approach for oil spill
detection based on a Multilayer Perceptron (MLP) neural
network are described in recent research studies (Del Frate et
al., 2000), (Topouzelis et al., 2005). The optimization process
to determine weight and bias parameters of ANN is called
learning. Backpropagation Delta rule (BP) is the most well
known learning algorithm. Besides to derivative based
conventional approaches such as Levenberg-Marquardt
(LM), some heuristic optimization methods such as Genetic
Search have been used in learning phase of ANN, as well.
Artificial Bee Colony (ABC) is such an optimization method
that can be used for ANN. Artificial Bee Colony algorithm is
a new meta-heuristic population based swarm intelligence
algorithm developed by Karaboga (2005). The ABC
algorithm mimics the intelligent foraging behaviour of
honeybee swarms. The first researches about ABC algorithm
focused into examining the effectiveness of ABC for
constrained and unconstrained problems against other well-
known modern heuristic algorithms such as Genetic
Algorithm (GA), Differential Evolution (DE), and Particle
Swarm Optimization (PSO) (Karaboga and Basturk, 2007;
Karaboga and Akay, 2009). Later on, ABC has been used for
ANN classifier training and clustering problem (Karaboga
and Ozturk, 2009) where some benchmark classification
problems were tested and the results were compared with
those of other widely-used techniques.
APPLICATION and RESULTS
The artificial neural network models used in this study are the
ones that were used in the earlier study (Ozkan et al., 2011)
related to oil spill detection in the Lebanon coast in 2007.
Since the main purpose of the paper is to examine the
generalizing capability of different ANN learning algorithms,
no training process was applied. The network topology used
is 9-6-2 consisting one hidden layer with 6 neurons. Input and
output layers are fixed by the dimension of input and output
patterns. Two classes (oil and look-alike) are represented by
2 neurons in the output layer and nine different features are
represented by 9 neurons in the input layer. The logarithmic
sigmoid transfer function is employed at hidden and output
layer neurons. In total, 74 parameters are used. The details of
the parameters of the optimization algorithms ABC, LM and
BP and other application features can be found in Ozkan et
al. (2011).
As in Ozkan et al. (2011), each algorithm has been run 30
times independently to reveal the robustness of the
algorithms used. For epoch numbers, two different
approaches were considered: (i) all algorithms were trained
1000 epochs; and (ii) different optimum epoch numbers for
each algorithm were used. The robustness can be defined as
the consistency of performances of multiple runs, i.e. the
narrower interval the error values of algorithms are clustered
in, the better robustness is. Good robustness means that the
algorithm is not sensitive to changing of the initial
conditions. Since the artificial intelligence techniques can
always produce some results that must not necessarily be
optimal, robustness is a useful measure in the comparison of
such type algorithms. Therefore, the test data results from
BP, LM, and ABC algorithms are compared to each other in
terms of mean and standard deviation descriptive statistics
obtained from 30 independent runs. The producer accuracies
and descriptive statistics of ABC, LM and BP algorithms
through both 1000 epochs and optimum epochs are given in
Table 3 and Table 4.
