Deep Belief Active Contours (DBAC) with Its
Application to Oil Spill Segmentation from
Remotely Sensed Sea Surface Imagery
Fatema A. Albalooshi, Paheding Sidike, Vasit Sagan, Yousif Albalooshi,and Vijayan K. Asari
Abstract
In this paper, we propose a machine learning-based oil spill
segmentation using aerial images. In detail, a novel deep neu-
ral network-based object segmentation, named Deep Believe
Active Contours (
DBAC
), is introduced, where a pre-trained
deep belief neural network is utilized to guide the moments
of active contours. Results show that (1) Unsupervised
pre-trained deep neural network can efficiently control the
evolution of active contour segmentation of oil spill regions;
and (2) When applying the proposed
DBAC
algorithm on the
test data from an oil spill image database, it produced a recall
rate of 66% and a precision rate of 60%, which outperformed
the state-of-the-art methods in the range of 4% ~ 18% and
1% ~ 10%, respectively. Moreover,
DBAC
produced a better
Hausdorff distance (an amount of 13.34) compared to the
competing methods. These results show the promises of
DBAC
for the task of oil spill segmentation in ocean environment.
Introduction
Oil spills has become one of the major factors that cause pol-
lution in the ocean environment and affects marine ecosys-
tems, damage marine life, and leads to substantial economic
impacts (Gill
et al
., 2012). Therefore, there is an urgent need to
identify and detect the spread of the oil spills in contaminated
areas rapidly to facilitate marine oil spill migration. However,
identification of surface or subsurface oil in the complicated
ocean environment is not an easy task. The most challenging
issue is to differentiate oil slicks from natural phenomena that
create dark patches such as natural films, grease, ice, wind-
shattering by land, internal waves, etc. (Espedal, 1999).
Continuous developments in remote sensing technologies
have automated oil spill inspection which is a cost-effective
yet more efficient way for environmental monitoring (Leifer
et al
., 2012; Salem and Kafatos, 2001; Alam
et al
., 2012).
Particularly, hyperspectral imaging techniques are capable of
identifying oil spills, monitoring mitigation, and preventing
unwanted incidents. In details, the rich spectral informa-
tion provided by hyperspectral images allows for assisting in
target (oil spills) detection in noisy backgrounds since objects
differ uniquely in absorbing and reflecting radiation at dif-
ferent wavelengths. Therefore, Hyperspectral Imagery (HSI)
has been widely used for oil spill measurements (Leifer
et al
.,
2012). Salem and Kafatos (2001) used spectral angle mapper
for identifying oil spills in water and shoreline. Richard
Geomez (Gomez, 2002) reviewed
HIS
based applications for
oil spill detection and tracking. Sanchez
et al
. (2003) used
supervised and unsupervised machine learning methods for
detection and monitoring oil spill using HSI. As known that
BP Deepwater Horizon (
DWH
) oil spill was a big disaster in
many aspects. Alam
et al
. (2012) analyzed HSI data that were
captured in
DWH
area, and provided techniques to identify
surface and subsurface oil-derived substances. Alam and
Sidike (2012) described and evaluated several approaches
for oil spill detection using HSI, whereas Sidike
et al
. (2012)
conducted a comparative study of several complete spectral
unmixing and partial spectral unmixing models for oil spill
detection in the ocean environment, and it showed that con-
strained energy minimization method yields the best results
among the other competitors. Optical pattern recognition
technique such as spectral fringe-adjusted joint transform
correlation (Alam and Ochilov, 2010) was also utilized and
combined with a logarithmic filtering operation for oil slicks
detection from remotely sensed HSI (Sidike and Alam, 2013).
Another type of popular remote sensing sensors for oil
spill inspection is Synthetic Aperture Radar (
SAR
). One of the
primary merits of
SAR
is that it can capture images regardless
of day and night, plus all-weather conditions (Brekke and
Solberg, 2005). In
SAR
imagery, oil spills can be differentiated
based on a number of features such as color, shape, contrast,
etc. (Chen and Lu, 2017). However, the noise from
SAR
im-
agery introduces challenges for extracting oil slick features,
such as geometric or edge detection features (Wang and Hu,
2015). Artificial Neural Network (
ANN
) based oil spill inspec-
tion has been attempted (Singha
et al
., 2013; Topouzelis,
2008; Del Frate
et al
., 2000) by researchers, and it can be of
the solutions to accommodate the noise effect in
SAR
imagery.
Singha
et al
. (2013) demonstrated that
ANN
-based oil spill
segmentation provides better accuracy than edge feature or
adaptive thresholding based methods. Lang
et al
. (2017) in-
troduced a multi-feature fusion based scheme where Support
Vector Machine (
SVM
) (Cortes and Vapnik, 1995) classifier was
used for oil spill classification from single-pol
SAR
images. In
contrast, Lupidi
et al
. (2017) introduces a new oil spill moni-
toring system, achieved by combining the significance param-
eter, wavelet correlator and a two-dimensional constant false
alarm rate for efficient processing of a large scale data sets.
Although
SAR
and hyperspectral images based oil spill
identification have provided promising results, improvements
in existing methods and generalization of algorithms over
Fatema A. Albalooshi and Yousif Albastaki are with the
University of Bahrain, Department of Computer Science,
University of Bahrain-Sakheer, P.O. Box 32038, Bahrain
(
).
Paheding Sidike and Vasit Sagen are with the Dept. of Earth &
Atmospheric Sciences, Saint Louis University, St. Louis, MO
63108.
Vijayan K. Asari is with the University of Dayton, Electrical
and Computer Engineering, Dayton, OH 45469.
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 7, July 2018, pp. 451–458.
0099-1112/18/451–458
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.7.451
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2018
451
