s,
used
ness.
cean
at oil
high
rities
non-
input
f the
1 this
gated
1ance
). The
in the
1 spill
cation
> dark
rpose,
of the
tc.) is
is are
em as
areas
x, Ihe
on one
x. and
>’
ication
sensing
red to
vzoglu
etwork
with a
without
e of the
s gives
remote
>xisting
bjective
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
Neural networks differ from statistical approaches in four main
aspects (Bishop, 1995; Kanellopoulos 1997): i) problem and
model complexity: NN deal with large amounts of training data
with higher complexities whereas statistical methods use much
smaller training sets, ii) goal of modelling: when using neural
networks, the main objective is the representation of
complicated phenomena rather than explanation. iii) no
assumption about data distribution: NN do not make any
explicit a priori assumptions about the underlying distribution
of the data, iv) robustness and quality of prediction estimation:
NN methods appear more robust than statistical ones with
respect to parameter tuning.
The general objectives of this project have been to describe,
demonstrate and test the potential of artificial neural networks
(NNs) for oil spill detection using SAR satellite images. In this
paper, we investigate two different NN architectures and
compare their performances. Two well known NN models,
Multilayer Perceptron (MLP) and Radial Basis Function (RBF)
neural networks are examined in order to evaluate their
performance in oil spill detection. The main difference between
the two architectures lies in the nature of the input-output
relations of their nodes. In a previous study (Topouzelis et al,
2003) a first attempt to examine the efficiency of MLP-NN was
performed. MLP networks are based on nonlinear sigmoid
functions and on combinations between them. RBF networks
are three-layer networks, whose output nodes form a linear
combination of the basis functions (usually of the Gaussian
type) computed by the hidden layer nodes. The main aim of the
present work is to detect the best topology for our network and
the algorithm better fit to our classification problem. The term
topology refers to the structure of the network as a whole,
specifying how its input, output and hidden units are
interconnected.
The paper is organized in six sections. In next section (section
2) we state the problem of oil spill detection from SAR images.
Section 3 presents a brief summary of MLP and RBF neural
network architectures and training algorithms. In Section 4 a
dataset description is given, presenting SAR images and
datasets derived from them. Results and conclusions follow is
sections 5 and 6, respectively.
2. PROBLEM DESCRIPTION
In this section, we briefly state the problem of oil spill detection
from remote sensing data acquired by active sensors. We start
by defining the direct problem on oil slick detection. Then we
describe the general methodology used and we compare it with
the neural network approach.
Oil is one of the major pollutants of the marine environment. It
may be introduced in diverse ways, such as natural sources,
offshore production, sea traffic, tanker accidents, atmospheric
deposition, river run off and ocean dumping (Pavlakis 1996).
The aim of the present work is to describe a methodology for
monitoring illicit vessel discharges to the sea surface, including
ballast water, tank washings and engine room effluent
discharges.
SAR systems are extensively used for the determination of oil
spills in the marine environment, as they are not affected by
local weather condition and cloudiness and occupy day to night.
SAR systems detect spills on the sea surface indirectly, through
725
the modification spills cause on the wind generated short
gravity — capillary waves (Alpers et al, 1991). Spills damp
these waves which are the primary backscatter agents of the
radar signals. For this reason, an oil spill appears dark on SAR
imagery in contrast to the surrounding clean sea. Other
phenomena which could cause dampen of short gravity-
capillary waves are (Alpers, 1991): organic film, grease ice,
wind front areas, areas sheltered by land, rain cells, current
shear zones, internal waves and upwelling zones. The existence
of a light wind, sufficient to generate short gravity — capillary
waves (Alpers et al, 1991) is necessary in order to detect spills.
It is well known that oil spill detection by radar is limited by the
sea state. Too low sea states (z2m/sec), as mentioned above,
will not produce sufficient sea surface roughness in the
surrounding area to contrast to the oil, and very high sea states
(«12m/sec) will break up the oil spills, creating scatters
sufficient to block detection. In their vast majority, the ships
discharge their oily effluents en route, leaving back linear oil
spills. This linearity is the most targeted feature by SAR image
interpreters when they trace oil spills (Pavlakis, 2001).
Several studies aiming at oil spill detection have been
implemented (Martinez and Moreno, 1996; Ziemke, 1996;
Kubat et al, 1998; Anne et al, 1999; Frate et al, 2000 ; Gade et
al, 2000, Benelli and Garzelli 1999; Lu, 1999, Topouzelis et al,
2002). Most of these studies rely on the detection of dark areas,
which are objects with a high probability of being oil-spills.
Once the dark areas are detected, classification methods based
on Bayesian or other statistical methods are applied to
characterize dark areas as oil spills or ‘lookalike’ objects.
Characteristics (geometric, surrounding, backscattering, etc.) of
spectral and spatial features of the dark area are used in order to
feed the statistical model. The drawback of these methods is a
complex process not fully understood, as it contains several
nonlinear factors. The development of an inverse model to
estimate such parameters turns out to be very difficult.
Recent work has demonstrated that neural networks (NNs)
represent an efficient tool for modelling a variety of nonlinear
discriminant problems. NNs may be viewed as a mathematical
model composed of several non-linear computational elements
called neurons, operating in parallel and massively connected
by links characterized by different weights (Bishop, 1995;
Ziemke, 1996; Kanellopoulos et al, 1997; Frate et al, 2000).
NNs have been successfully used for remote sensing
applications (Bishop, 1995; Kanellopoulos et al, 1997; Frate et
al, 2000; Kavzoglou and Mather 2003; Uiu and Jensen, 2004)
For oil spill detection NNs have been used (Zimke and Athley,
1995; Ziemke, 1996; Frate et al, 2000) in different perspective
from the present work; one using airborne data (SLAR) and
another for dark object classification. The innovation of the
present study is the use of the original SAR image and some
features derived from it as inputs to NN. The network is called
to determine if the image contains an oil spill or not.
3. NEURAL NETWORK ARCHITECTURE AND
TRAINING ALGORITHMS
In the present study two different networks are tested in order to
evaluate the one most suitable for oil spill detection: Multilayer
Perceptron (MLP) and Radial Basis Function (RBF) networks.
Both of them belong to the feed-forward networks where there
is no feedback connection between layers and no connections
between units in the same layer. Moreover, both work in a
