2012
atures of oil
fect. Journal
zy approach
nce and Re-
ntitative Re-
ional, Vol. 1,
onetic waves
'hilosophical
n by satellite
1), pp. 1-13.
rsen, O. and
ing of an oil
illetin 62(2),
'alabresi, G.,
ers-sar data.
ns on 38(5),
ing wind his-
ensing 20(1),
and Trivero,
zes. Interna-
3566.
/. and Lange,
r by oceanic
ir theoretical
), pp. 52-70.
of terrasar-x
and Remote
hine learning
es. Machine
and Garrett,
ity waves by
on with a mi-
c Technology
2009. On the
ion. Interna-
602.
fferent semi-
ase study in
ARSeL Sym-
ll detection in
note Sensing,
istent, hierar-
AL OF GEO-
3741-8743.
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
TESTING THE GENERALIZATION EFFICIENCY OF OIL SLICK CLASSIFICATION
ALGORITHM USING MULTIPLE SAR DATA FOR DEEPWATER HORIZON OIL SPILL
C. Ozkan!, B. Osmanoglu', F. Sunar’, G. Staples”, K. Kalkan’, F. Balik Sanl1°
! Erciyes University, Engineering Faculty, Geodesy and Photogrammetry Engineering Dept., 38039 Kayseri, Turkey,
cozkan@erciyes.edu.tr
? University of Alaska - Fairbanks, P.O. Box 757320, Fairbanks AK 99775, bosmanoglu@alaska.edu
? Istanbul Technical University, Civil Engineering Faculty, Geomatics Engineering Dept., 34469 Maslak Istanbul,
Turkey, fsunar@itu.edu.tr, kalkank@itu.edu.tr
^ MDA,13800 Commerce Parkway, Richmond, V7S 1L5, Canada, gstaples@mdacorporation.com
* Yildiz Technical University, Civil Engineering Faculty, Geodesy and Photogrammetry Engineering Dept., Davutpasa
Campus, 34220 Esenler, Istanbul, Turkey, fbalik@yildiz.edu.tr
Commission VII, WG VII/2
KEYWORDS: Marine pollution, Oil spill classification, Generalization, SAR
ABSTRACT:
Marine oil spills due to releases of crude oil from tankers, offshore platforms, drilling rigs and wells, etc. are seriously affecting the
fragile marine and coastal ecosystem and cause political and environmental concern. A catastrophic explosion and subsequent fire in
the Deepwater Horizon oil platform caused the platform to burn and sink, and oil leaked continuously between April 20th and July
15th of 2010, releasing about 780,000 m? of crude oil into the Gulf of Mexico. Today, space-borne SAR sensors are extensively used
for the detection of oil spills in the marine environment, as they are independent from sun light, not affected by cloudiness, and more
cost-effective than air patrolling due to covering large areas. In this study, generalization extent of an object based classification
algorithm was tested for oil spill detection using multiple SAR imagery data. Among many geometrical, physical and textural
features, some more distinctive ones were selected to distinguish oil and look alike objects from each others. The tested classifier was
constructed from a Multilayer Perception Artificial Neural Network trained by ABC, LM and BP optimization algorithms. The
training data to train the classifier were constituted from SAR data consisting of oil spill originated from Lebanon in 2007. The
classifier was then applied to the Deepwater Horizon oil spill data in the Gulf of Mexico on RADARSAT-2 and ALOS PALSAR
images to demonstrate the generalization efficiency of oil slick classification algorithm.
INTRODUCTION The backscatter energy level for oil-spilled areas received in
| E. ; SAR systems is too low since the oil dampens the capillary
Marine oil spills, a form of pollution, caused by releases of waves of the sea surface. However, there is a serious problem
crude oil from tankers, offshore platforms, drilling rigs and that the other natural phenomenas also dampen the short
wells, etc. can seriously affect the fragile marine and coastal waves and create dark areas on the sea surface due to
ecosystem and cause political and environmental concern. suspension of Bragg scattering mechanism depending to
The amount of oil spilled annually worldwide has been ocean and/or atmospheric conditions, (Solberg et al. 2007).
estimated at more than 4.5 million tons (Bava et al., 2002). Thus, the dark image regions of which the probabilities of
The Deepwater Horizon oil spill (also referred to as the Gulf being either oil spill or alike are high must be segmented.
of Mexico oil spill) happened on April 20, 2010, is the largest These segmented parts of original images are used to obtain
accidental marine oil spill in the history of the USA the features of shape, contrast and textural characteristics. In
petroleum industry. The resulting oil slick quickly expanded terms of these features, some classification algorithms based
to cover hundreds of square miles of ocean surface, being a on statistical, neural, fuzzy, rule based, boosting algorithms
threat to marine life and adjacent coastal wetlands and to the etc. are used for identification of the dark areas in a manner
Gulfs fishing and tourism industries. of binary classification, i.e. oil or alike (Fiscella et al., 2000;
; Del Frate et al., 2000; Solberg et al., 1999; Keramitsoglou et
For rapid emergency response activities, four basic issues; i) al., 2006; Ramalho and Medeiros, 2006).
prevention ii) alarm, iii) monitoring and iv) damage
quantification, should be planned carefully (Bava et al.,
2002). In order to lessen its effect, the improvement of its
detection and continuous monitoring are the most important
In this study, segmentation of dark objects was done by
eCognition software. Among many geometrical, physical and
C ; textural features, some more distinctive ones were selected to
issues to effectively plan countermeasures responses. Today, distinguish oil and look alike objects from each others. The
PURE ANR Being d am pr pow tested classifier was constructed from a Multilayer Perceptron
over He Past decade Symmenc Aperture Radar (SAR) Artificial Neural Network trained by Artificial Bee Colony
satellites are often preferred to optical sensors due to the all (ABC), Levenbera-Marguardt (LM). and. Backpropagntion
Weather and ell dav camaialifies aud Dens used tQ detect th e (BP) optimization algorithms. The training data to train the
oil spills discharged into the sea with sufficient accuracies classifier were constituted ‘from SAR data consisting of oil
Cabas à à 2007) Oir Spi doipotion procedures m SAR spill originated from Lebanon in 2007. The classifier was
data generally comprise segmentation, feature extraction and then applied to the Deepwater Horizon oil spill data in the
classification stages (Solberg et al, 2007; Brekke and Gulf of Mexico ‘on RADARSAT-2 and ALOS PALSAR
Solberg, 2005).
67
