OIL SPILL DETECTION USING RBF NEURAL NETWORKS AND SAR DATA
K. Topouzelis * *, V. Karathanassi ^, P. Pavlakis ^, D. Rokos à
? Laboratory of Remote Sensing, S
School of Rural and Surveying Engineering, National Technical University of Athens,
Heroon Polytechniou 9, GR-15780, Greece, ktopo@mail.ntua.gr
b Laboratory of Remote Sensing, National Technical University of Athens, (karathan, rslab)@survey.ntua.gr
¢ National Centre for Marine Research (Greece), ppavla@ncmr.gr
PS WG VILVS
KEY WORDS: Neural, Networks, SAR, Environment, Pollution,
ABSTRACT:
Illegal oil spill discharges cause seri
for the detection of oil spills in the marine env
However, radar backscatter values for oil spills are
phenomena because dampen capillary a
spill detection have been conducted. Most of these stu
Bayesian probability of being oil-spills. The drawback of these m
involved. The use of Neural Networks (NNs) in remote sensing
linear data of a multidimensional input space. Furt
and output as they determine their own rel
Multilayer Perceptron (MLP) neural network and different training algorithms for o
paper another approach of NN use in oil spill detection is presented.
in order to be compared with the Multilayer Perceptron. For both ne
Sea, Extraction, Detection
ous damage to marine ecosystems. Synthetic Aperture Radar (SAR) images are extensively used
ironment, as they are not affected by local weather conditions and cloudiness.
very similar to backscatter values for very calm sea areas and other ocean
nd short gravity waves is caused by the presence of an oil spill. Several studies aiming at oil
dies rely on the detection of dark areas, which are objects with a high
ethods is a complex process, because there are many non linearities
has increased significantly as NN can simultaneously handle non-
hermore, NN do not require an explicitly well-defined relationship between input
ationships based on input/output values. In a previous study, the potential of the
il spill classification were investigated. In this
The Radial Basis Function (RBF) neural network is investigated
tworks, several topologies are examined and their performance
is evaluated. MLPs appear to be superior than RBFs in detecting oil spills on SAR images.
1. INTRODUCTION
Oil spills are seriously affecting the marine ecosystem and
cause political and scientific concern since they have serious
affect on fragile marine and coastal ecosystem. The amount of
pollutant discharges and associated effects on the marine
environment are important parameters in evaluating sea water
quality. Satellite images can improve the possibilities for the
detection of oil spills as they cover large areas and offer an
economical and easier way of continuous coast areas patrolling.
Synthetic Aperture Radar images have been widely used for oil
spill detection. The presence of oil film on the sea surface
damps the small waves and drastically reduces the measured
backscatter energy, resulting in darker areas in SAR imagery
(Martinez and Moreno, 1996; Pavlakis et al, 1996; Kubat et al,
1998; Anne et al, 1999; Frate et al, 2000; Gade et al, 2000).
However, dark areas may be also caused by other phenomena,
like locally low winds, currents or natural sea slicks called
‘lookalikes’.
Several studies aiming at semi-automatic or automatic oil spill
detection can be found in literature (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). These studies first detect
manually or with threshold filtering dark areas on the image
which could be oil spills. If not supported by visual inspection
(Lu et al, 1999: Frate et al, 2000), dark areas detection
prerequires a threshold wind speed (Anne, 1999; Gade et al,
* Corresponding author.
2000) sufficient to generate the sea state (Frate et al, 2000). The
extent of the sea state conditions is consequently included in the
estimation of the strength of the contrast signal that an oil spill
yields. When dark areas are detected, statistical classification
methods (e.g. Bayesian) are applied to characterize the dark
areas as oil spills or ‘lookalike’ objects. For this purpose,
estimation of a number of spectral and spatial features of the
dark areas (geometric, surrounding, backscattering, etc.) is
prerequired. In relevant studies, classification methods are
usually applied only on the dark areas, considering them as
objects (Anne, 1999; Frate et al, 2000), whilst dark areas
detection methods are based on pixel-basis processing. The
transition from the detection step to the characterization one
needs user interference in terms of masking, coding, and
selecting the dark objects in order to proceed to classification
processing (Topouzelis et al, 2002).
Neural networks have been employed to process remote sensing
images and have achieved improved accuracy compared to
traditional statistical methods (Kanellopoulos, 1997; Kavzoglu
and Mather, 2003). This success derives from neural network
characteristics. A single neuron can be compared with a
multivariance linear regression model, which works without
any a priori assumptions concerning the statistical nature of the
data set. The massive parallel work of several neurons gives
further capabilities for solving complex problems in the remote
sensing area. Moreover, NN are able to learn from existing
examples, making the classification adaptive and objective
(Kanellopoulos, 1997).
724
Internc
Neural
aspects
model
with hi
smalle
networ
compli
assum[
explici
of the
NN m
respect
The ge
demon:
(NNs)
paper,
compat
Multila
neural
perforn
the tw:
relation
2003) a
perforn
functio:
are thr
combin
type) c
present
the alg«
topolog
specifyi
intercor
The paj
2) we s
Section
network
dataset
datasets
sections
In this s
from re
by defir
describe
the neur
Oil is oi
may be
offshore
depositi
The aim
monitor
ballast
discharg
SAR sy
spills in
local we
SAR sy:
