You are on page 1of 13

Marine Pollution Bulletin 188 (2023) 114651

Contents lists available at ScienceDirect

Marine Pollution Bulletin


journal homepage: www.elsevier.com/locate/marpolbul

Ocean oil spill detection from SAR images based on multi-channel deep
learning semantic segmentation
Rogelio Hasimoto-Beltran a, *, Mario Canul-Ku a, Guillermo M. Díaz Méndez b,
Francisco J. Ocampo-Torres c, Bernardo Esquivel-Trava b
a
Centro de Investigación en Matemáticas (CIMAT), Jalisco S/N, Col. Valenciana, Guanajuato 36023, Guanajuato, Mexico
b
Centro de Investigación Científica y de Educación Superior de Ensenada (CICESE), Carretera Ensenada - Tijuana No. 3918, Zona Playitas, Ensenada 22860, Baja
California, Mexico
c
CEMIE-Océano A. C, Ciudad Universitaria, C.U., Coyoacán, Ciudad de México 04510, Mexico

A R T I C L E I N F O A B S T R A C T

Keywords: One of the major threats to marine ecosystems is pollution, particularly, that associated with the offshore oil and
Oil spill detection gas industry. Oil spills occur in the world's oceans every day, either as large-scale spews from drilling-rig or
Deep learning tanker accidents, or as smaller discharges from all sorts of sea-going vessels. In order to contribute to the timely
SAR images
detection and monitoring of oil spills over the oceans, we propose a new Multi-channel Deep Neural Network (M-
Spatial filtering
DNN) segmentation model and a new and effective Synthetic Aperture Radar (SAR) image dataset, that enable us
to emit forewarnings in a prompt and reliable manner. Our proposed M-DNN is a pixel-level segmentation model
intended to improve previous DNN oil-spill detection models, by taking into account multiple input channels,
complex oil shapes at different scales (dimensions) and evolution in time, and look-alikes from low wind speed
conditions. Our methodology consists of the following components: 1) New Multi-channel SAR Image Database
Development; 2) Multi-Channel DNN Model based on U-net and ResNet; and 3) Multi-channel DNN Training and
Transfer Learning. Due to the lack of public oil spill databases guaranteeing a correct learning process of the M-
DNN, we developed our own database consisting of 16 ENVISAT-ASAR images acquired over the Gulf of Mexico
during the Deepwater Horizon (DWH) blowout, off the west coast of South Korea during the Hebei Spirit oil
tanker collision, and over the Black Sea. These images were pre-processed to create a 3-channel input image IM =
{IO, IW, IV}, to feed in and train our M-DNN. The first channel IO represents the radiometric values of the original
SAR Images, the second and third channels are derived from IO; in particular, IW represents the output of the wind
speed estimation using CMOD5 algorithm (Hersbach et al., 2003) and IV represents the variance of IO that in­
corporates texture information and at the same time encapsulates oil spill transition regions. IM channels were
split and linearly transformed for data augmentation (rotation and reflection) to obtain a total of 80,772 sub-
images of 224 × 224 pixels. From the entire database, 80 % of the sub-images were used in the DNN training
process, the remaining (20 %) was used for testing our final architecture. Our experimental results show higher
pixel-level classification accuracy when 2 or 3 channels are used in the M-DNN, reaching an accuracy of 98.56 %
(the highest score reported in the literature for DNN models). Additionally, our M-DNN model provides fast
training convergence rate (about 14 times better on the average than previous works), which proves the effec­
tiveness of our proposed method. According to our knowledge, our work is the first multi-channel DNN based
scheme for the classification of oil spills at different scales.

1. Introduction crude oil have been (and still) dumped into the ocean by drilling plat­
forms, transport ships, and natural sources, causing severe damage to
Oil exploration and extraction are among the most important activ­ marine life and economic development (Solberg, 2012). Two of the most
ities in the industrialized world and, unfortunately, historically repre­ catastrophic oil spills over the oceans have occurred in the Gulf of
sent a significant hazard to the ocean and coastal ecosystems. Tons of Mexico. The Deepwater Horizon (DWH) is known as the largest

* Corresponding author.
E-mail address: hasimoto@cimat.mx (R. Hasimoto-Beltran).

https://doi.org/10.1016/j.marpolbul.2023.114651
Received 26 August 2022; Received in revised form 9 January 2023; Accepted 20 January 2023
Available online 1 February 2023
0025-326X/© 2023 Elsevier Ltd. All rights reserved.
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

accidental oil spill in history, wherein 594,000 tons of oil were dumped are the evolution of the Artificial Neural Networks (ANNs), which were
into the U.S. Gulf Coast (from Texas to Florida) in 2010 (United States. proposed in the 1950s. ANNs are based on the generalization of per­
Coast Guard and National Response Team (US), 2011). The world's ceptron concept, proposed by Rosenblatt (1958). Different than a human
second-worst oil spill occurred in Mexico's Ixtoc-1 well during June of brain structure wherein an external stimulus is processed by millions of
1979, located off the west coast of the Peninsula of Yucatan (south­ interconnected neurons (creating a unique processing unit) (Wang,
western the Gulf of Mexico). Ixtoc-1 rupture released 530,300 tons of oil 2003), an ANN is structured as stacked layers (one or two successive
into the Bay of Campeche, where residues were washed ashore on layers). Each layer comprises a set of neurons which process data in­
adjacent coasts, including U.S beaches in the northern part of the Gulf of formation as a mathematical operation or activation function (Chollet,
México (Jerneloev and Linden, 1981). More recently (October 2019) 2021).
and once again off the coast of Campeche (Southeast Gulf of México), the DNNs conversely, incorporate even hundreds of successive layers,
Cayo Arcas maritime terminal suffered a blowout accident, causing a being able to solve very complex problems. Currently, DNNs incorporate
discharge of ~89.87 tons of oil with an extension of 4484 ha (UNACAR, different types of layers and parameters that altogether define the model
2019; SEMAR, 2019). Chronic oil pollution in the ocean is more harmful to be trained for different purposes and applications, as for example,
in the long term than a sole accidental oil spill (Soto et al., 2014). YouTube Recommendations (Covington et al., 2016), Text ranking
Therefore, it is crucial to detect and monitor the extension of oil spills in (Severyn and Moschitti, 2015), 3D Archaeological artifacts classification
the early stage (whether by accidental or natural sources), in order to (Canul-Ku et al., Jan 2019), micro-fossils semantic segmentation (Hou
emit a warning to corresponding authorities and take prompt actions to et al., 2021), image recognition (Simonyan and Zisserman, 2014), oil
mitigate severe ecological damage, that otherwise may take years or spill segmentation (Zeng and Wang, 2020). One of the main problems is
decades for its recovery (Soto et al., 2014; Meyers, 2012). that DNNs require large databases (in the order of millions of objects) for
Remote sensor technologies have become a valuable tool for the training phase (Soekhoe et al., 2016), which most of the time is
detecting and monitoring the extension and dynamics of ocean oil spills difficult to obtain, incurring in what is called overfitting (the model only
at different scales, going from micro (natural sources, illegal ship works for the source of data used during the training phase) (Glorot and
discharge, etc.) to large-scale spills (DWH). Different types of sensors Bengio, 2010; LeCun et al., 2015).
have been used for this purpose, for example laser fluorescence, optical A good alternative to avoid overfitting is Transfer Learning (TL)
(visible spectrum), infrared (8–14 μm wavelengths), and microwave scheme (Deng and Yu, 2014; Ng et al., 2015), in which a pre-trained
sensors (Fingas and Brown, 2018). Laser fluorosensors are not for model from a (related) large-scale dataset is transferred to the new
deploying on satellites with sufficient resolution, and commonly they problem by just tuning or adjusting the model with the new and much
are mounted on airplanes or ships (Solberg, 2012), therefore, the sea smaller dataset in order to thousands of images, saving time and
surface coverage is limited. The optical and infrared sensors offer large computational resources. In our case, a general purpose DNN segmen­
coverage of the Earth surface. However, they are passive sensors that tation model is transferred to the problem of oil spill classification (se­
depend on the emitted electromagnetic radiation by the sun, so, they mantic segmentation) by tuning-up or re-training the pre-defined large-
cannot be used under cloudy and night conditions (Shamsudeen, 2020). scale model with our smaller SAR image database. Another technique to
Nevertheless, under ideal weather conditions, the infrared sensors can improve the generalization of DNN is data augmentation, this technique
estimate the oil thickness better than microwave sensors. consists in increase the number of images through linear trans­
The active microwave sensors, such as the Synthetic Aperture Radar formations, such as rotations and reflections (Zeng and Wang, 2020).
(SAR), have shown great ability to operate in all weather conditions, Unlike traditional schemes (image processing, computer vision, etc.),
both day and night, and at different spatial resolutions (meters), offering DNNs can learn the best characteristics that discriminate and segment
more extensive coverage of the ocean. The presence of winds over the oil slicks, independently of the complexity of the geometry and exten­
sea surface generates capillary waves producing a certain roughness on sion (Song et al., 2017). Because of this, DNNs have more robustness and
the surface that is responsible for the microwave pulse emitted by the generalization ability (Wang et al., 2021) than traditional methods.
SAR being reflected in the sensor direction (backscatter). This roughness The use of SAR for the study of oil slicks dates back to the 80's, with
is usually related, in the SAR images, to common wind waves without oil initial experiments on the interaction of microwaves and oil slicks
or uncontaminated areas (Krestenitis et al., 2019). Surfactant films over (Deloor, 1983; Johnson and Croswell, 1982; Kuilenburg, 1975). Subse­
the ocean surface, such as oil, dampen the capillary waves (Leifer et al., quently, more advanced methodologies were proposed for the segmen­
2012), causing the absence of backscattering, resulting in black spots in tation and characterization of dark-spots associated with oil spills
the SAR images. As easy as it may appear, oil detection by SAR is not an through contrast enhancement, filtering process, texture analysis, sta­
easy task, the oil layer on the sea surface may be dispersed by the dy­ tistical measures, etc. Some methods based on this approach are adap­
namics of the ocean and dissipated into the environment due to multiple tive thresholds (Brekke and Solberg, 2005), features based on single-
physical-chemical processes that modify its composition, such as evap­ polarization (Lang et al., 2017), dual-polarization (Velotto et al.,
oration, dispersion, emulsification, dissolution, oxidation, sedimenta­ 2011; Kim and Jung, 2018) and quad-polarization (Li et al., 2018).
tion, and bio-degradation (Kvočka et al., 2021). In this way, oil spills are Adaptive thresholds methods compute statistics of image intensity in
inherently dynamic, making their temporal observation rather difficult. order to obtain ideal thresholds to classify those pixels associated with
Fortunately, in most cases, the spill can be detected accurately in the oil spills. Brekke et al. (Brekke and Solberg, 2005) compute a local
early stages (close to the source), and monitor its evolution, such as adaptive threshold based on the ratio σμ for each window around the
coordinates location, extension, dispersion, mixing process, direction of pixels in the SAR image; where σ and μ correspond to the variance and
motion, etc. For this reason, it is important to continuously observe average of pixels intensities into the window. The ratios are used to
those well-known oil-production platforms and oil transportation define six homogeneity categories. Each category contains confidence
permanently, particularly if micro or small-scale oil spills are frequent. intervals and threshold value in decibels, which are used to compute the
There are several methods for oil spill detection based on image adaptive threshold and consequently the pixel segmentation process.
processing techniques (adaptive threshold), computer vision (segmen­ Since Brekke's segmentation results are not sufficient to characterize
tation), and very recently Deep Neural Networks (DNNs) methods. The the oil slick completely, Lang et al. (2017) proposed additional features
popularity of DNNs methods is related to several factors, among the to improve the segmentation performance. These features combine
most important are the increase in computing performance (high-end radiometric, geometric, and texture information to build a feature vec­
CPUs and GPUs) and the availability of larger datasets (in particular tor, which characterizes a pixel as either oil spill or sea-water. A support
marine oil spills) that allows a robust training of complex models, vector machine is applied after the segmentation step, where each
leading to high accuracy classification rates (Deng and Yu, 2014). DNNs

2
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

feature vector is classified as oil or not oil pixel. Shamsudeen (2020) and Al-Ruzouq et al. (2020).
Velotto et al. (2011) use TerraSAR-X images to compute the complex Different than previous DNNs approaches for oil spill detection in the
coherence measure based on the complex HH and VV scattering am­ literature, in this research we propose a Multichannel DNN (M-DNN)
plitudes. Complex coherence is then used to characterize pixels with the model based on 3 input images or channels: original SAR image (IO),
presence of oil. Similar to Velotto et al., in the research of Kim and Jung derived variance (IV), and derived wind field map (IW). These three-
(2018), complex coherence, co-polarized phase difference (CPD) of HH channel input improves significantly the performance of the oil-spill
and VV polarizations, and VV texture information are used to build a classification rates at different scales or dimensions, such as small, me­
feature vector for each pixel of TerraSAR-X images, where the vectors dium, and large-scale, including look-alikes. We are not aware of any
are classified through an ANN. Li et al. (2018) use quad-polarization previous DNN model using the proposed variance and/or wind field as
(HH, VV, HV, VH) of RadarSAT-2 to build ten SAR features to charac­ input channels so far. Complementing this model, we propose a new
terize the pixels, and then apply a stacked auto-encoder as classifier. relatively large and diverse (it considers different oil spills scales) multi-
Advanced methods based on DNN models have also been used for oil channel dataset that improves significantly the convergence speed of the
spill detection on SAR images for prompt decision making. Yu et al. M-DNN the training phase. Furthermore, the Ground Truth (GT) is
(2018) propose a Generative Adversarial Network (GAN) for oil spill carefully selected in which high mixture levels of oil-seawater are not
segmentation. The GAN models are represented by two DNNs which marked as oil in our labeling process, since they are most likely to be
compete during the training, the first DNN is known as generator, and confused with look-alikes (our interest is to detect the main oil-spill
the second as discriminator (Goodfellow et al., 2020). The generator stream).
corresponds to the DNN U-net (Ronneberger et al., 2015), and the The rest of the paper is organized as follows. The next section de­
discriminator is a convolutional neural network used as regressor clas­ scribes our DNN methodology including multichannel image formation,
sifier. During the training, each DNN has an associated cost or loss database creation, and DNN training phase. Section 3 presents the
function. The authors propose the f-divergence as loss function to pro­ experimental results and comparison against state-of-the-art methodol­
duce accurate oil spill pixel segmentation for SAR images, and make a ogies. Finally, conclusions of our work are presented in Section 4.
comparative analysis among five f-divergence functions. Their perfor­
mance results are based on a pair of independent images, and not over 2. Methodology
the entire oil spill database they are based on.
Krestenitis et al. (2019) make a comparative analysis between six Our proposed methodology is based on M-DNN and SAR-images of
DNNs to solve the task of semantic segmentation. The authors used five the sea surface for the extraction or segmentation of oil slicks. It consists
classes to label each pixel of the images, which are related to regions of the following main steps (see Fig. 1): 1) A new Multi-channel SAR
with presence of oil. The images were re-scaled to meet the resolution, Image Database Development; 2) M-DNN Model based on U-net and
and split on patches of 320 × 320 pixels. The best results correspond to ResNet; and 3) M-DNN Training and Transfer Learning. Broadly
the DNN U-net with DNN ResNet (He et al., 2015) as encoder or back­ speaking, we take the original radiometric SAR image and derive two
bone, and 600 epochs. additional information channels (for a total of three channels) to be
Wang et al. (2021) propose re-use the AlexNet DNN to classify the processed by the M-DNN. The M-DNN output is a binary image (one-
pixels as oil and not oil. The DNNs takes as input small cropped images output channel) wherein pixels related to continent, sea-water, drilling
centered into pixel of interest. The interest pixels are selected from platforms, ships, etc., have been filtered out, preserving only those pixels
Regions of Interest (RoIs) with presence of oil, and are labeled as oil and related to oil slicks. Detail of each step is described in the next
not oil. The training process employs 2500 epochs and 17,100 cropped subsections.
images in total.
Li et al. (2021) propose a novel method based on GAN-DNN, similar 2.1. Multi-channel SAR Image Database Development
to Yu et al. (2018). The authors use a set of GAN-DNN, trained with n
multiscale images. During training, each generator and discriminator We developed an effective and reliable oil spill database that in­
process a single multiscale image, and the nth outputs are passed to the corporates the following features: 1) a multi-channel SAR image infor­
next generator and discriminator, to be combined by a sum operation. mation in the presence of oil spills at different scales (small, medium,
The sum operation is equal to the sum of residual neural network (He and large-scale), geometries and textures; 2) SAR image global
et al., 2015), which improves the learning process of each generator and normalization; and 3) Ground Truth construction and post-processing
discriminator. The DNN was trained with cropped images with size of (SAR image splitting and data augmentation). These features are
944 × 912, 352 × 404, 400 × 420, and 398 × 398 pixels related to described next.
regions with presence of oil.
Zhu et al. (2022) proposed a novel DNN called Contextual and 2.1.1. Multi-channel SAR Image Formation
Boundary-supervised Detection Network (CBD-Net). CBD-Net is based Our new database was built using 16 full-resolution Envisat ASAR
on auto-encoder architecture, where the encoder component is a vari­ images from the DWH (2010) (Wan and Cheng, 2013), off the Black Sea
ation of ResNet (He et al., 2015), where each encoder block is followed coast (2005) (Akar et al., Dec 2011), and from the oil tanker Hebei Spirit
by an attention function. This function uses an average to compress the (2007) (Alpers et al., 2017), corresponding to 14-WSM and 2-IMP with
spatial dimension of the encoder output. The authors propose the joint C-band and VV polarization, courtesy of the European Space Agency
loss function defined as the sum of the binary cross-entropy, the struc­ (ESA) (see Table 1). SAR images store radiometric information of the
tural similarity index (SSIM) loss and the Dice coefficient loss. backscattered signal intensity, which is affected by the physics of the
Huang et al. (2022) propose a novel DNN method based on bounding reflectivity and the interference from the environment (such as speckle
box regression for oil spill detection, which is different to our oil spill noise). Using the Sentinel Application Platform (SNAP, https://step.esa.
segmentation task. In our case, we are able to classify the oil spill at a int/main/toolboxes/snap/), SAR images were calibrated, multi-looked,
pixel level and thus compute its extension area and precise location. georeferenced, and their Normalized Radar Cross Section (NRCS)
Another difference with our proposed scheme is that they do not directly σ 0 values converted to dB. From this set of pre-processed images, we
work with radiometric values of the SAR images, but with transformed derived the corresponding variances and wind field maps to get our
gray levels normalized between 0 and 255. Finally, a data augmentation proposed multi-channel input to the DNN IM = {IO, IW, IV}, where IO is
technique is applied to expand the number of samples to prevent over­ the original radiometric SAR image, IV and IW are the corresponding
fitting. A comprehensive search remote sensing and machine learning Variance Filter (VF) and wind field images of IO, as shown in Fig. 2 and
methods for oil spill detection and monitoring, is presented in explained next:

3
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

Fig. 1. Oil spill segmentation pipeline based on U-net and ResNet DNNs: 1) Multi-channel SAR Image Database Development; 2) M-DNN Model based on U-net and
ResNet; and 3) M-DNN Training and Transfer Learning.

Table 1 IV (x, y) = σ 2V
Envisat ASAR images reported oil spills.
Date Type Date Type

2005-04-25 at 07:52:25 h IMP 2010-05-15 at 16:01 h WSM


2005-04-25 at 07:52:42 h IMP 2010-05-24 at 03:59 h WSM 1 ∑(WV − 1)/2 ∑(WV − 1)/2
2007-12-11 at 01:40 h WSM 2010-05-28 at 15:52 h WSM σ 2V = (IO (x − iy − j) − μ )2
WV2 i=− (WV − 1)/2 j=− (WV − 1)/2
2010-04-29 at 03:45 h WSM 2010-05-31 at 15:58 h WSM
2010-05-02 at 03:51 h WSM 2010-06-03 at 03:44 h WSM
2010-05-09 at 15:48 h WSM 2010-06-06 at 03:49 h WSM
2010-05-12 at 15:55 h WSM 2010-06-13 at 15:49 h WSM
2010-05-15 at 03:41 h WSM 2010-07-11 at 03:49 h WSM
V − 1)/2
(W∑ V − 1)/2
(W∑
1
μ= IO (x − iy − j)
WV2
– Variance Channel (IV): Variance filter behaves as an edge and texture i=− (WV − 1)/2 j=− (WV − 1)/2

detection (Fabijaśka, 2011), in particular it helps in highlighting


transition zones (statistical changes) normally associated with
seawater-oil interaction. VF is the computed variance of a local
where WV is the side length of a square local window, σ 2 the local in­
square window around the pixel IO(x, y), according to the following
tensity variance, and μ the local intensity average. After a careful
equations:
compromise between computation performance and edge extraction

-
(a) IO original image (b) IV variance image (c) IW wind speed image (d) Ground Truth
Fig. 2. Pre-processed images obtained from Envisat SAR Image acquired in 2010-04-29 to 03:45 h TC, which is associated with Deepwater Horizon oil spill event.
NaN values in image (c) are set to transparent.

4
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

characteristics, the optimal WV is set to 9. A worth mentioning feature in variance and high wind speed edges (see Fig. 2b and d). Our interest is to
IV (Fig. 2b), is the capacity of the variance filter to emphasize high- detect the main oil-spill stream in its early stages, capable of inhibiting
frequency mixing zones (seawater-oil transition) surrounding the oil the formation of capillary waves (roughness) at wind speeds approxi­
slick, providing additional information to the DNN model. mately between 3 and 7 m s− 1 (Brekke and Solberg, 2005). Weathered
oil spills on the other side, have been spread and dispersed over the
– Wind Speed Channel (IW): Sea surface roughness is fundamentally ocean by the action of a physical phenomenon (wind, turbulence, waves,
driven by a complex interaction between the relative motion of the etc.), showing high mixture levels of oil-seawater and consequently
wind field and the underlying ocean dynamics (eumetsat, 2020). In higher values of variance. These mixtures zones are not taken into ac­
the presence of surfactants, oil slicks for example, this air-sea inter­ count as oil spills in the GT. In addition, our dataset was carefully
action is affected and may result in a significant reduction of the designed to include oil spills with different geometries and scales (large,
ocean roughness, depending on wind speed and oil thickness (Leifer medium, and small-scale including natural discharges), as shown in
et al., 2012). If the wind speed is low to moderate, slicks are clearly Fig. 3. This is to deliver the best classification of dark spots, either as oil
visible as dark areas against the otherwise homogeneous back­ or not oil (continent, seawater, look-alike, etc.).
ground. In contrast, if the wind speed is high, oil may be dispersed In order to fit the actual dimensions of the multi-channel components
and degraded into the ocean (Leifer et al., 2012). Due to this close into our DNN model, we split the original IM and GT set into sub-images
relationship between oil slicks and sea surface roughness, we pro­ or patches Pk ∈ IM and Bk ∈ GT respectively. A total of 42,892 patches Pk
posed the addition of the wind speed IW computed from the original were obtained with dimensions w × h, where w = 224 ≪ IMx and h = 224
SAR image Io, as an additional input channel to our DNN model ≪ IMy (see Fig. 4), where (IMx, IMy) represent the original size of the full
(Fig. 2c). From each SAR scene, the wind speed was estimated resolution SAR images (IO). Additionally, in order to increase the num­
through the CMOD5 Geophysical Model Function (GMF) (Hersbach ber of sub-images associated with oil spill and to balance to number of
et al., 2003), which relates the NRCS σ0 of a radiometrically cali­ oil and not-oil pixels in our database, we apply data augmentation to
brated image to the following variables: wind speed 10 m above the those Pk patches holding more than 10 % of pixels labeled as oil. Data
sea surface U10, the incidence angle of the radar beam θi, the azimuth augmentation includes random rotation, horizontal and vertical re­
angle between the SAR look-direction and the wind direction ϕ, for a flections according to Zeng and Wang (2020). For each selected patch,
given radar co-polarization HH or VV. we produce a total of 20 additional sub-images or patches (considering
the three channels patches Pk plus the GT patch Bk), in this way the
The wind direction Udir was computed directly from the SAR image IO resulting augmented dataset reaches 80,772 patches, which corresponds
by standard 2D Fourier analysis. Computed values were then compared to our final experimental dataset.
with the numerical results from the IOWAGA database (Rascle et al.,
2008; Rascle and Ardhuin, 2013). When the former did not match the
numerical results, the latter were preferred as input information to 2.2. Multi-Channel DNN model based on U-net and ResNet
derive the azimuth angle ϕ. In the case of the WSM images, with a pixel
size of 150 m by 150 m (after multilooking), the resolution of the wind Oil spills in the ocean have complex geometries and associated non-
field (U10, Udir) was 900 m. Similarly, in the case of the IMP images, the uniform sea-surface roughness, driven by the ocean-atmosphere inter­
wind field resolution was 150 m. For an extended description on the action. Because of this, DNNs are excellent tools for solving this kind of
theoretical background of the CMOD5, please refer to Hersbach et al. problems. The main disadvantage of DNN models is the large amount of
(2003) and Díaz Méndez et al. (2010). data required (in the order of million of objects or images) to be hand
labeled and passed through the training process. In order to overcome
2.1.2. SAR image global normalization this task and being able to use a relatively small oil-spill dataset, we
Since our DNN model requires image intensity values between [0,1] make use of a pre-trained network called U-net (Ronneberger et al.,
(it accelerates training convergence (Wang et al., 2019)), we perform a 2015).
global normalization of the IO dataset before the training phase, as The U-net is one of the most used DNN for different segmentation
opposed to image independent normalization applied in previous works problems offering high accuracy results. U-net is based on auto-encoder
using DNNs. Global normalization keeps the same point of reference for architecture which has two components: the Encoder and Decoder, each
all images in the dataset, that is observing the ocean surface (seawater, component is composed by four blocks (see Fig. 5). Each Encoder's block
oil spills, look-alikes, etc.) with the same eyes. Normalizing every single performs the feature extraction process of the input patch Pk, producing
(l− 1) l
image independently, may create artificial errors of false image contrast, two feature vectors r ∈ R n/(2 ) and x ∈ R n/(2 ) , where l ∈ {1, 2, 3, 4}
l l
that may confuse the DNN model. Under this case (independent is the block number and n (=224 * 224 * 3) is the dimension of the input
normalization), every image is stretched out in different degrees patch Pk for l = 1. xl goes through the remaining Encoder's blocks, and rl
creating “perhaps” look-alike regions or pixels, producing significant is used later in the Decoder process as shown in Fig. 5. These feature
classification errors. vectors represent the main characteristics that identify the multi-
For global image normalization, we compute the global minimum channel image, it is pretty much like a trained compression algorithm
(Min) of the IO dataset (usually negative in the presence of oil spills), and that learn how to encapsulate the main representation of the input image
add it up to the same dataset (Io − Min). The new dataset has now a (raw data). Each encoder's block consists of the following layers as
minimum value of 0. We now search for the global maximum (Max), and shown in Fig. 6: two convolutional layers, one REctified Linear Unit
compute Io/Max; now the range of IO dataset is [0,1]. This operation (ReLU) layer, and max pooling layer.
ensures that all images are stretched out in the same rate or point of The convolution layers process the input patch with 2m convolu­
reference. The image variance IV is computed after the normalization tional filters of dimension 3 × 3 × m, where m is the depth of the filter
process. applied for the extraction of prominent image features. The corre­
sponding output is then fed into the ReLU layer. The Relu layer elimi­
2.1.3. Ground Truth (GT) construction and post-processing nates those values close to cero to prevent potential numerical problems
Once images have been normalized, the oil spill GT was created in future processing, either training or classification process (Glorot
based on Wan and Cheng (2013), Akar et al. (2011) and Alpers et al. et al., 2011). The max pooling layer also known as down-sampling layer,
(2017) dataset information. We have been very careful on this process reduces the length of the primitive feature vector rl by a factor of 2. This
by hand-labeling (as GT) those areas confirmed as oil and well correlated process is repeated for the rest of the remaining blocks to have a final
with low variance and low wind speed magnitudes surrounded by a high down-sampling factor of 16 in both coordinate axes.

5
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

Original SAR images (Gulf of México)


Large Scale Mid-scale Small-scale
Image Area: 154,051 Km2 Image area: 39,964 Km2 Image area: 18,397 Km2

(a) Large (b) Middle (c) Small


Fig. 3. Different oil spills scales (dimensions) included in our dataset.

Fig. 4. Patches dataset creation from multi-channel image for oil spill segmentation.

U-net Encoder has been modified to include an additional process at process. It takes as input the Encoder's feature vector x4 and produces a
the end of each block, called residual block (He et al., 2015). The pur­ weighted image u1, wherein oil related features have been emphasized.
pose of this block is both to avoid the problem of vanishing gradient During this process, x4 is upsampled and concatenated with the corre­
(Pascanu et al., 2012) that occurs in the optimization process during sponding rl vector in each independent block (Fig. 5).
training, and to gain additional accuracy of deeper than deep neural Finally, u1 is transformed by a sigmoid activation function Fk = σ(u1)
networks. Thanks to this groundbreaking architecture called ResNet, it that assigns a probability measure to every element in u1, that guaran­
is possible to train even thousands of layers effectively. ResNet residual tees the output Fk is between 0 and 1. Finally, Fk is binarized by using a
block is aimed at learning the distribution error of the input feature predefined threshold (experimentally set to 0.6), in which oil pixels are
vector xl, which is modeled as follow: set 1 if B
̂k ≃ Bk .

zl = h(xl ) + R(xl , Wl ) 2.3. Multi-channel DNN training and transfer learning

Since a DNN is a supervised method, it requires both the input Pk ∈ IM


and corresponding Ground Truth (Bk ∈ GT) for the learning or training
process. Therefore, DNN is a mapping process that associates IM to GT,
xl+1 = max(0, zl )
such that GT
̂ ≃ GT, represented as DNN:IM → GT. ̂ The GT image corre­
sponds to a binary image where all pixels have been manually associated
to oil or not-oil (sea-water, continent, drilling platform, etc.) using
previous knowledge of the spill.
where h(xl) is the identity mapping function, R(xl, Wl) is the residual
The training phase consists of an iterative optimization process
function or residual block with parameter Wl, and xl, xl+1 are the input
aimed at finding the best DNN weights for an optimal data classification,
and output vector of l-th residual block (He et al., 2015).
that is the segmentation of oil spills at a pixel-level. The optimization
The Decoder on the other side, performs the Encoder's inverse
method is based on the gradient-descent Adam algorithm (Kingma and

6
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

Fig. 5. General diagram of U-net + ResNet DNN.

best model, we make use of Adam algorithm (Kingma and Ba, 2014) as
the optimizer to update the weights using 100 iterations, α = 0.0001 as
learning rate, and batch size of 128 to the training process. We randomly
split the experimental dataset into three subsets: a training set (80 %), a
test set (20 %), and a cross-validation (20 %) set extracted from the
training set itself to evaluate if the network is learning correctly without
overfitting. From the test set we get global statistics such as the mean,
variance and confidence intervals to evaluate the effectiveness of our
proposed method. Binary cross-entropy function is used as loss function
Fig. 6. U-net encoder block. in the training step, defined as follows:

Ba, 2014) and a cost function evaluated at the end of each iteration, that
work close together to find the best convolutional filters (in the Encoder 1∑n ∑ ( )
loss = − B(xy)k log F(xy)k
and Decoder) and residual blocks parameters Wl that conform the DNN n k=1 (xy)∈{1,…,w}×{1,..,h}
layers (see Section 2.2). The cost function computes the error between
the GT and GT̂ and based on this result all DNN weights are updated
until a predefine number of iterations (epochs) is met.
The original U-net Encoder weights to be used in our DNN model where n is the total number of patches in training dataset, B(x, y) is the
have been pre-trained with ImageNet Large-scale Visual Recognition Ground Truth, and F(x, y) is the predicted pixel of the DNN (U-net).
Challenge (Russakovsky et al., 2015), which consists of 1.4 million of Our DNN code implementation is written in Python programming
RGB images with dimension 224 × 224 splitted into 1000 classes. On the language, using a pre-trained ResNet34 backbone model (34 encoding
other hand, the U-net decoder-block was re-trained (fined tuned) using blocks) from the Segmentation Model Repository described in Yaku­
our 3-channel augmented dataset (IM), to teach the DNN how to properly bovskiy (2019), and Keras framework (Chollet et al., 2015) for the
solve our oil pixel classification problem, this is called TL. optimization process during DNN fine-tuning. The network training was
performed on an Intel Xeon Silver 4214 CPU, 128 GB RAM with an
3. Experiments and results NVIDIA Titan RTX 24 GB GPU.

3.1. DNN settings


3.2. Evaluation
Our experimental setup is aimed at investigating the effectiveness of
To evaluate the effectiveness of our proposed method we make use of
different input-channel combinations for the problem of oil-spill clas­
sification. Each channel combination requires separate training of the both the Ground Truth B ∈ GT and the predicted B ̂ ∈ GT
̂ subimages. The
DNN (U-net + ResNet) and their own statistical performance metrics in predicted subimage corresponds to oil spill segmentation output of the
order to obtain the most effective combination for oil classification. trained DNN model. The evaluation of our DNN model is computed at
Channel combinations were set as follows: 1) IO = {IO}, 2) Iw = {IW}, 3) patch level and considers the correct segmentation of oil versus sea-
Iv = {IV}, 4) Iov = {IO, IV}, 5) Iow = {IO, IW}, Iwv = {IO, Iv} and 6) Iovw = {IO, water and continent. We choose the classification ACCuracy (ACC)
IV, IW}. Each full resolution image channel is split into sub-images of (Tsoumakas and Katakis, 2007; Elkan, 2012), the Intersection over
dimension w × h = 224 × 224, following U-net constraints. To get the Union (IoU) or Jaccard index (Rahman and Wang, 2016; Niwattanakul
et al., 2013), and the F1_score (Goutte and Gaussier, 2005) as the

7
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

performance metrics defined as: Table 2


Channel combination comparison in terms of ACC, IoU and F1-score metrics (%).
Combination ACC (oil IoU (oil and ACC-Oil IoU -Oil F1-oil
TP + TN and not oil) not oil)
ACC = ;
TP + TN + FP + FN Io 97.35 ± 94.85 ± 91.98 ± 88.25 ± 95.25
0.98 1.85 1.85 4.11
Iw 96.11 ± 92.53 ± 87.05 ± 82.88 ± 92.40
1.85 3.39 2.28 7.79
Iv 78.90 ± 65.22 ± 4.72 ± 4.67 ± 8.97
TP 5.10 6.96 5.14 4.40
IoU = ; Iov 98.41 ± 96.88 ± 94.82 ± 92.47 ± 96.68
FP + TP + FN
0.88 1.70 1.70 4.67
Iow 98.56 ± 97.17 ± 97.30 ± 93.55 ± 98.23
0.68 1.31 0.89 2.86
Iwv 97.91 ± 95.90 ± 94.61 ± 90.84 ± 96.66
1.00 1.91 1.30 4.18
precision • recall
F1 score = 2 • ; Iowv 98.54 ± 97.12 ± 96.26 ± 93.44 ± 97.68
precision + recall 0.70 1.36 1.36 2.85
TP
precision = ;
TP + FP
using the two-channel combination Iow (ACC = 0.9856, IoU = 0.9717,
TP
recall = ; and F1-score = 98.24 %), and very close behind are Iowv and Iov in terms
TP + FN
of ACC metric. The use of the variance channel by itself does not make
sense (very small values in all accuracy metrics.
are obtained for oil pixel detection), the DNN may find the most
where TP is number of pixels that were correctly classified as oil (true prominent edges but still needs the intensity values of each region from
positive), TN is number of pixels that were correctly classified as not oil, the original SAR image. Iw gets better results but still below the multi­
FP is the false positive value, that is the number of pixels that were channel inputs Iov, Iow, and Iovw by as much as 7 %–10 % and 4 %–5 % for
classified incorrectly as oil, and FN is the number of pixels that were ACC-Oil and IoU-Oil respectively. It appears that Iv and Iw provide
classified at not oil incorrectly. All these values are computed at a pixel- similar information (when Io is included) regarding to oil spill classifi­
level using the corresponding pair of binary sub-images B and B. ̂ ACC cation. In the analysis of the true pixel-label (oil, not-oil) against pre­
and IoU represent the average over all binary patches in the testing dicted pixel-label (DNN classified pixel), Io gets the lowest oil pixel
dataset, which are computed with their respective confidence interval as classification performance of 91.98 % (8.02 % oil-pixel misclassifica­
a statistical proof of our method. The F1-score is used to evaluate binary tion), compared to 97.30 % for the Iow channel (2.7 % oil-pixel
classification systems, and combines the precision and recall of the misclassification), for a total improvement of 5.32 %. The rest of
model. precision measures the quality of the model in the classification channel combinations also provide good results as in the case of Iow.
task (how many of the positive predictions made by the model were Based on these results, we can see that 2 or 3-channel combination
correct), while the recall is how many of the positive samples (oil pixels provide better segmentation performance than a sole channel Io, which
in this case) present in the dataset were correctly identified by the confirms that wind speed and variance information are useful variables
model. Therefore, F1-score is a harmonic mean of the model's precision to highlight those pixels associated to oil slicks when working with
and recall. To further assess the effectiveness of our pixel-level classifi­ radiometric SAR images.
cation method, we also present the confusion matrix (also called con­ A visual analysis of the segmentation results for all multi-channel
tingency table) as evaluation criteria of our classifier. Given the number combinations is shown in Fig. 8. The first and second rows correspond
of classes C in a particular classifier, the confusion matrix is represented to the original SAR images and corresponding ground truth, in which the
by C × C matrix describing the percentage of correct and incorrect pixels on white color represent the oil spill, and black color pixels are
predictions made by the classifier compared to the actual labels on the associated to not-oil, that is sea water, continent, drilling platforms,
test set. Since we provide a binary classification of oil spills, our ships, etc.). In most cases, Iow gets the best visual combination, which is
confusion matrix is represented by a 2 × 2 matrix of correct oil and not confirmed by the corresponding ACC and IoU values.
oil predictions. The confusion matrix represents the average accuracy We now compared our oil classification results against similar DNN
over all patches in the testing set: based state-of-the-art methods in the literature as shown in Tables 3 and
4. A comparative analysis of the dataset size in Table 3, shows that our
dataset is one of the largest in the literature for the oil-detection problem
using DNN (Deep Neural Networks) with 80,772 multichannel images
1∑ C
with dimension 224 × 224. In Table 4 we show the performance com­
ACC = ACCi
L i=1 parison of the different models based on ACC and IoU metrics. Our
proposed multichannel Iow model gets the highest scores in both global
(ACC and IoU) and per-class (ACC-Oil and IoU-Oil) classification accu­
racies. In particular, for ACC we are on average 1 % higher than (Wang
where ACCi is the accuracy of each patch and L is the number of patches
et al., 2021; Yu et al., 2018; Li et al., 2021), and 23 % higher on average
in B.
̂ A correct classification means that the learned model predicts the
than (Krestenitis et al., 2019; Zhu et al., 2022). None of the authors in
same class as the original class of the test case.
the comparative analysis report ACC-Oil (represented by NR—not re­
A comparison of the ACC, IoU, and F1-score (obtained from the
ported), and only one reports IoU-Oil (Krestenitis et al., 2019), which is
confusion matrices) metrics for the proposed method are summarized
40 % far below our reported value. It would have been interesting if
Table 2 and Fig. 7 (confusion matrices). Looking at Io classification
those authors we compare with, had reported the confusion matrices as
performance standalone, it detects quite well the absence or low-levels
well. Our performance results are statistically more trustable and stable
of ocean roughness associated to oil spill with an ACC = 97.35%, IoU
than other authors.
= 94.85 %, and F1-score = 95.41 %. The best performance was obtained
We have shown that our methodology is well suited to detect oil

8
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

(a) (b)

(c) (d)

(e) (f)

(g)
Fig. 7. Confusion matrices for each channel combination of our dataset.

patches; however, it is also important to determine its capacity to for which, some information has already been removed; b) We need
discriminate between oil and look-alike patches, which is a key chal­ original SAR radiometric values, otherwise we are not able to compute
lenge in oil spill detection. In the next experiment, we make use of a the wind field maps from JPEG images (not clear what additional pro­
publicly available dataset used in Krestenitis et al. (2019), for additional cessing the Krestenitis' dataset went through to be in the range [0,1]); c)
comparative analysis with the same labeled dataset. This dataset in­ Images have been rescaled (expanded or shorten) to meet the resolution
cludes original SAR images and corresponding ground truth classified as of 1250 × 650, which severely affects variance computation. Isolated
continent, oil, look-alike, and seawater, but there are important flaws: a) white or black pixels are widened during image expansion or interpo­
Original images are in JPEG format, which is a compression algorithm lation, introducing artifacts that affect the performance of our DNN; and

9
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

Table 4
Comparative analysis of our proposed multi-channel classification model against
most relevant state-of-the-art methods in the literature (%). Highest accuracy
values are represented in bold.
Authors Method ACC IoU ACC-oil IoU-oil Epochs
(oil, (oil, (iterations)
not- not-
oil) oil)

Yu et al. SAR/GAN 97.62 NR NR NR 4000


(2018)
Krestenitis SAR/U- NR 64.97 NR 53.79 600
et al. net +
(2019) ResNet
Wang et al. RGB/ 97.49 NR NR NR 2500
(2021) AlexNet
Li et al. SAR/ 97.60 NR NR NR 100
(2021) GAN/
Multiscale
Zhu et al. SAR/CBD- NR 83.42 NR NR 80
(2022) Net/
Multiscale
Ours (Iow) SAR/U- 98.56 97.17 97.30 93.55 100
net +
ResNet/
Multi-
channel
Average improvement 1 23 – 39.8 14×

using one and two channels (original SAR image and its corresponding
variance), obtaining the following performance results shown in
Table 5. It can be observed that our DNN model is on the average 3.43 %
and 55.0 % better in classifying oil and look-alikes respectively. Look-
alikes are detected with high accuracy (Io = 97.50 % and Iov = 97.93
%) by our model, because their coverage area is larger than the one
related to oil spills in the dataset (the variance channel in our model is
less affected by larger look-alikes), as shown in Table 5. On the contrary,
oil spills in the dataset are represented in their majority by small surfaces
that, as mentioned before, their ground truths are out of phase from the
original location, affecting significantly the overall accuracy of the
classification, as shown in Fig. 9. Red represents a look-alike, blue is the
ground truth, and yellow is the oil spill classification result of our DNN
model. It can be seen, the ground truth provided by Krestenitis et al.
(2019) is misplaced with respect to the real oil spill location. Despite
Fig. 8. Qualitative results obtained from a sample of Envisat SAR Images and these errors in the dataset, our results show much better performance
the multi-channel combinations of our dataset.
accuracy in the classification of both oil spills and look-alikes using the
original SAR image and corresponding variance.
As a final experiment, we have included a pair of Sentinel1-A SAR
Table 3
Number of images of our proposed multi-channel classification model against image acquired in 2022-03-07 at 00:15:51 h UTC and 2022-04-12 at
most relevant state-of-the-art methods in the literature. 00:15:52 h UTC from the Gulf of México, with potential (unconfirmed)
oil spills. These images delineate oil streaks at different resolution (20
Authors Total number of Image Approximate total number
images (training and size of images of size (224 ×
m), scales, and patterns to those present in our database, and conse­
testing phases) 224) quently were not considered in the training process. As shown in Fig. 10,
our scheme detects the corresponding SAR anomalies at a pixel level
Yu et al. 43 256 × 98
(2018) 256 quite well using Iow component. There is a clear correlation between
Krestenitis 1112 1250 × 17,792 these two images (Io and Iw), wherein Iw represents an image enhance­
et al. (2019) 650 ment of Io, that produces an increase in both classification performance
Wang et al. 18,810 300 × 48,906
and visual appearance.
(2021) 300
Li et al. (2021) 30 256 × 68
256
Zhu et al. 8070 256 × 9223 Table 5
(2022) 256 Comparative analysis of our DNN model against Krestenitis et al. (2019), using
Ours (Iow) 80,772 224 × 80,772 their own developed dataset (%).
224
Author IoU-Oil IoU-look- Percentage of improvement wrt [11] (oil,
alike look-alike)
d) In the majority of the dataset, the ground truth is wrong and Krestenitis 53.79 55.40 –
considerably misplaced (out of phase) from the real or original image et al.
scenario, affecting the final performance values. Despite all these (2019)
Ours IO 57.13 97.50 (3.34, 42.10)
problems, we did use Krestenitis' dataset to retrained our DNN model
IOV 54.27 97.93 (0.48, 42.53)

10
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

(a) SAR image. (b) Overlaping of GT and DNN segmentaon


into SAR image.
Fig. 9. Dataset characteristics.

Fig. 10. Sentinel-1A SAR sub-images of Campeche


Sound, southwestern Gulf of Mexico, rst row contains
a sub-image acquired in 2022-03-07 to 00:15:51 h
UTC has coordinates 19◦ 36′ 06.0′′ N, 92◦ 19′ 22.3′′ W
(upper left corner of quadrant), 19◦ 14′ 32.4′′ N,
91◦ 57′ 48.7′′ W (lower right corner of quadrant).
Second row shows a sub-image has acquired in 2022-
04-12 to 00:15:52 h UTC with coordinates
20◦ 07′ 18.1′′ N 94◦ 07′ 11.1′′ W (upper left corner of
quadrant), 19◦ 45′ 44.5′′ N 93◦ 45′ 37.6′′ W (lower right
corner of quadrant).

(a) 2022-03-07 (b) segmentaon

(c) 2022-04-12 (d) segmentaon

4. Conclusions during the training process. The Intersection over Union (IoU) and Ac­
curacy (ACC) performance values reached averages of 97.17 % and
Space-borne SAR is a very useful remote sensing technology to detect 98.56 % respectively for our proposed multi-channel component Iow,
and monitor oil spills in the ocean surface, providing large area coverage representing the highest scores reported in the literature. We demon­
with high spatial resolution. Several automatic methodologies have strated that our proposed multi-channel DNN methodology along with
been proposed for the segmentation of oil spills from SAR images; our dataset is better than using only one channel (original SAR images),
however, most of them do not work with full-resolution imagery nor providing a better data generalization for oil spill segmentation with
automatic semantic segmentation of oil and look-alikes by means of excellent accuracy. Our future work is aimed at different paths: a)
DNNs. Based on these aspects, the main contribution of this work is both improving the performance of our DNN model by dynamically adapting
the use of a multi-channel DNN model (that includes original SAR image the optimization step size during the training process; b) automatic
and derived variance and wind field maps) and the proposal of a new channel weighting (currently all channels in our model have the same
and carefully design dataset that takes into account different oil spill weight); that is the DNN should be able to detect which channel is more
geometries, texture and scales, providing high-accuracy image classifi­ important in the classification process; c) Detecting the type of oil and
cation values at a pixel level. Specifically, texture information of the oil thickness using additional multispectral information.
spill, seawater and look-alikes is introduced through a variance filter
which delineates or outlines the oil spill transition regions along the CRediT authorship contribution statement
wind speed as an additional channel. The dataset is expanded using data
augmentation, allowing the construction of a large and balanced dataset Rogelio Hasimoto-Beltran: Contributed to the main ideas of the
(oil and not oil regions) which in turn, reduces over-fitting of the DNN research proposal and responsible of research activity planning.

11
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

Mario Canul-Ku: Research activity planning and Implementation of Glorot, X., Bordes, A., Bengio, Y., 2011. Deep sparse rectifier neural networks. In:
Proceedings of the fourteenth international conference on artificial intelligence and
the computational techniques for data analysis and image database.
statistics. JMLR Workshop and Conference Proceedings, pp. 315–323.
Guillermo M. Díaz Méndez: Responsible for the computation of the Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S.,
wind field maps and acquisition of the financial support for the project Courville, A., Bengio, Y., 2020. Generative adversarial networks. Commun. ACM 63
leading to this publication. (11), 139–144.
Goutte, C., Gaussier, E., 2005. A Probabilistic Interpretation of Precision, Recall and F-
Francisco J. Ocampo-Torres: Co-responsible of wind field map Score, With Implication for Evaluation. In: Losada, D.E., Fernández-Luna, J.M.
interpretation. (Eds.), Advances in Information Retrieval. ECIR 2005, Lecture Notes in Computer
Bernardo Esquivel-Trava: Co-responsible of wind field map Science, vol 3408. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-540-
31865-1_25.
interpretation. He, K., Zhang, X., Ren, S., Sun, J., 2015. Deep residual learning for image recognition.
https://doi.org/10.48550/arXiv.1512.03385.
Hersbach, H., Stoffelen, A.D., Haan, S., 2003. An improved C-band scatterometer ocean
Declaration of competing interest geophysical model function: CMOD5. Technical memorandum. ECMWF.
Hou, Y., Canul-Ku, M., Cui, X., Hasimoto-Beltran, R., Zhu, M., 2021. Semantic
The authors declare that they have no known competing financial segmentation of vertebrate microfossils from computed tomography data using a
deep learning approach. J. Micropalaeontol. 40 (2), 163–173.
interests or personal relationships that could have appeared to influence Huang, X., Zhang, B., Perrie, W., Lu, Y., Wang, C., 2022. A novel deep learning method
the work reported in this paper. for marine oil spill detection from satellite synthetic aperture radar imagery. Mar.
Pollut. Bull. 179, 113666.
Jerneloev, A., Linden, O., 1981. Ixtoc 1: A Case Study of the World’s Largest Oil Spill.
Data availability
Jan.
Johnson, J.W., Croswell, W.F., 1982. Characteristics of 13.9 ghz radar scattering from oil
Data will be made available on request. films on the sea surface. Radio Sci. 17 (03), 611–617.
Kim, D., Jung, H.S., 2018. Mapping oil spills from dual-polarized Sar images using an
artificial neural network: application to oil spill in the Kerch strait in November
Acknowledgment 2007. Sensors 18 (7).
Kingma, D.P., Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
This research has been funded by the Mexican National Council for
Krestenitis, M., Orfanidis, G., Ioannidis, K., Avgerinakis, K., Vrochidis, S.,
Science and Technology (CONACyT) - Mexican Ministry of Energy - Kompatsiaris, I., 2019. Oil spill identification from satellite images using deep neural
Hydrocarbon Fund, project 201441: Implementation of networks of networks. Remote Sens. 11 (15), 1762.
ocean observations (physical, geochemical, ecological) to generate Kuilenburg, J.Van, 1975. Radar observations of controlled oil spills. In: International
Symposium on Remote Sensing of Environment, 10 th. Ann Arbor, Mich,
scenarios upon possible contingencies related to the exploration and pp. 243–250.
production of hydrocarbon in deep sea waters in the Gulf of Mexico. This Kvočka, D., Banovec, P., Žagar, D., 2021. A review of river oil spill modeling. Water 13
is a contribution of the Gulf of Mexico Research Consortium (CIGoM). (12).
Lang, H., Zhang, X., Xi, Y., Zhang, X., Li, W., 2017. Dark-spot segmentation for oil spill
Authors wish to acknowledge PEMEX's specific request to the Hydro­ detection based on multifeature fusion classification in single-pol synthetic aperture
carbon Fund to address the environmental effects of oil spills in the Gulf radar imagery. J. Appl. Remote. Sens. 11 (1), 015006.
of Mexico. Access to Envisat and Sentinel-1 imagery has been kindly LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436.
Leifer, I., Lehr, W.J., Simecek-Beatty, E.Bradley D., Clark, R., Dennison, P.,
granted by the European Space Centre. Authors also would like to Hu, Yongxiang Y., Matheson, C.Jones S., Holt, B., et al., 2012. State of the art
acknowledge CONACYT's support through the “Consorcio de Inteli­ satellite and airborne marine oil spill remote sensing: Application to the bp
gencia Artificial” under Project 296737 and to the “Laboratorio Nacio­ deepwater horizon oil spill. Remote Sensing of Environment 124, 185–209.
Li, Y., Zhang, Y., Yuan, Z., Guo, H., Pan, H., Guo, J., 2018. Marine oil spill detection
nal de Supercómputo del Bajío” under Project 300832. based on the comprehensive use of polarimetric Sar data. Sustainability 10 (12),
4408.
References Li, Y., Lyu, X., Frery, A.C., Ren, P., 2021. Oil spill detection with multiscale conditional
adversarial networks with small-data training. Remote Sens. 13 (12).
Meyers, R., 2012. Encyclopedia of Sustainability Science and Technology. Springer,
Akar, S., Süzen, M.L., Kaymakci, N., Dec 2011. Detection and object-based classification
Heidelberg.
of offshore oil slicks using envisat-asar images. Environ. Monit. Assess. 183 (1),
Ng, H.W., Nguyen, V.D., Vonikakis, V., Winkler, S., 2015. Deep learning for emotion
409–423.
recognition on small datasets using transfer learning. In: Proceedings of the 2015
Alpers, W., Holt, B., Zeng, K., 2017. Oil spill detection by imaging radars: challenges and
ACM on international conference on multimodal interaction. ACM, pp. 443–449.
pitfalls. Remote Sens. Environ. 201, 133–147.
Niwattanakul, S., Singthongchai, J., Naenudorn, E., Wanapu, S., 2013. Using of jaccard
Al-Ruzouq, R., Gibril, M.B.A., Shanableh, A., Kais, A., Hamed, O., Al-Mansoori, S.,
coefficient for keywords similarity. In: Proceedings of the International
Khalil, M.A., 2020. Sensors, features, and machine learning for oil spill detection and
MultiConference of Engineers and Computer Scientists, 1.
monitoring: a review. Remote Sens. 12 (20).
Pascanu, R., Mikolov, T., Bengio, Y., 2012. On the difficulty of training recurrent neural
Brekke, C., Solberg, A., 2005. Oil spill detection by satellite remote sensing. Remote Sens.
networks. CoRR abs/1211.5063.
Environ. 95 (1), 1–13.
Rahman, M.A., Wang, Y., 2016. Optimizing intersection-over-union in deep neural
Canul-Ku, M., Hasimoto-Beltran, R., Jiménez-Badillo, D., Ruiz-Correa, S., Román-
networks for image segmentation. In: International Symposium on Visual
Rangel, E., Jan 2019. Classification of 3d archaeological objects using multi-view
Computing. Springer, pp. 234–244.
curvature structure signatures. IEEE Access 7, 3298–3313.
Rascle, N., Ardhuin, F., 2013. A global wave parameter database for geophysical
Chollet, F., 2021. Deep Learning With Python. Simon and Schuster.
applications. part 2: model validation with improved source term parameterization.
Chollet, F., et al., 2015. Keras.
Ocean Modelling 70, 174–188.
Covington, P., Adams, J., Sargin, E., 2016. Deep neural networks for youtube
Rascle, N., Ardhuin, F., Queffeulou, P., Croizé-Fillon, Denis D., 2008. A global wave
recommendations. In: Proceedings of the 10th ACM Conference on Recommender
parameter database for geophysical applications. Part 1: wave-current–turbulence
Systems. ACM, pp. 191–198.
interaction parameters for the open ocean based on traditional parameterizations.
Deloor, G.P., 1983. Project noorwidj. part 4: the radar backscatter of oil spills as
Ocean Model. 25 (3–4), 154–171.
measured at platform noordwijk.
Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: convolutional networks for
Deng, L., Yu, D., 2014. Deep learning: methods and applications. Foundations and
biomedical image segmentation. In: International Conference on Medical Image
Trends®. Signal Process. 7 (3–4), 197–387.
Computing and Computer-assisted Intervention. Springer, pp. 234–241.
Díaz Méndez, G.M., Lehner, S., Ocampo-Torres, F.J., Ming Li, X., Brusch, S., 2010. Wind
Rosenblatt, F., 1958. The perceptron: a probabilistic model for information storage and
and wave observations off the south Pacific Coast of Mexico using TerraSAR-X
organization in the brain. Psychol. Rev. 65–386.
imagery. Int. J. Remote Sens. 31 (17–18), 4933–4955.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z.,
Elkan, C., 2012. Evaluating Classifiers. University of California, San Diego.
Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L., 2015. ImageNet large
eumetsat, 2020. Ocean surface characterisation of ocean surface roughness in nwp. https:
scale visual recognition challenge. Int. J. Comput. Vis. 115 (3), 211–252.
//www.eumetsat.int/ocean-surface-roughness-in-NWP.
SEMAR, 2019. Informe oficial 0001300181819, SEMAR, derrames de hidrocarburo en la
Fabijaśka, A., 2011. Variance filter for edge detection and edge-based image
terminal marítima de cayo arcas.
segmentation. In: Perspective Technologies and Methods in MEMS Design,
Severyn, A., Moschitti, A., 2015. Learning to rank short text pairs with convolutional
pp. 151–154.
deep neural networks. In: Proceedings of the 38th international ACM SIGIR
Fingas, M., Brown, C.E., 2018. A review of oil spill remote sensing. Sensors 18 (1), 91.
conference on research and development in information retrieval. ACM,
Glorot, X., Bengio, Y., 2010. Understanding the difficulty of training deep feedforward
pp. 373–382.
neural networks. In: Proceedings of the thirteenth international conference on
artificial intelligence and statistics, pp. 249–256.

12
R. Hasimoto-Beltran et al. Marine Pollution Bulletin 188 (2023) 114651

Shamsudeen, T.Y., 2020. Advances in remote sensing technology, machine learning and Velotto, D., Migliaccio, M., Nunziata, F., Lehner, S., 2011. Dual-polarized terrasar-x data
deep learning for marine oil spill detection, prediction and vulnerability assessment. for oil-spill observation. IEEE Trans. Geosci. Remote Sens. 49 (12), 4751–4762.
Remote Sens. 12 (20). Wan, J., Cheng, Y., 2013. Remote sensing monitoring of gulf of mexico oil spill using
Simonyan, K., Zisserman, A., 2014. VeryDeep Convolutional Networks for Large-scale envisat asar images. In: 2013 21st International Conference on Geoinformatics,
Image Recognition. arXiv preprint arXiv:1409.1556. pp. 1–5.
Soekhoe, D., Putten, P., Plaat, A., 2016. On the impact of data set size in transfer learning Wang, X., Liu, J., Zhang, S., Li, Y., Fan, J., Deng, Q., Wang, Z., 2021. Detection of oil spill
using deep neural networks. In: Bostróm, Henrik, Knobbe, Arno, Soares, Carlos, using Sar imagery based on alexnet model. Comput. Intell. Neurosci. 1–14.
Papapetrou, Panagiotis (Eds.), Advances in Intelligent Data Analysis XV. Springer Wang, S., Yang, D.M., Rong, R., Zhan, X., Xiao, G., 2019. Pathology image analysis using
International Publishing, Cham, pp. 50–60. segmentation deep learning algorithms. Am. J. Pathol. 189 (9), 1686–1698.
Solberg, A., 2012. Remote sensing of ocean oil-spill pollution. Proceedings of the IEEE Wang, S.C., 2003. Artificial neural network. In: Interdisciplinary Computing in Java
100 (10), 2931–2945. Oct. Programming. Springer, pp. 81–100.
Song, D., Ding, Y., Li, X., Zhang, B., Xu, M., 2017. Ocean oil spill classification with Yakubovskiy, P., 2019. Segmentation Models. https://github.com/qubvel/segmentati
radarsat-2 Sar based on an optimized wavelet neural network. Remote Sens. 9 (8). on_models.
Soto, L.A., Botello, Alfonso V., Licea-Durán, Sergio, Lizárraga-Partida, Marcial L., Yáñez- Yu, X., Zhang, H., Luo, C., Qi, H., Ren, P., 2018. Oil spill segmentation via adversarial f
Arancibia, Alejandro, 2014. The environmental legacy of the ixtoc-i oil spill in -divergence learning. IEEE Trans. Geosci. Remote Sens. 56 (9), 4973–4988.
Campeche sound, southwestern gulf of Mexico. Frontiers in marineScience 1. Zeng, K., Wang, Y., 2020. A deep convolutional neural network for oil spill detection
Tsoumakas, G., Katakis, I., 2007. Multi-label classification: an overview. Int. J. Data from spaceborne Sar images. Remote Sens. 12 (6), 1015.
Warehouse. Min. 3 (3), 1–13. Zhu, Q., Zhang, Y., Li, Z., Yan, X., Guan, Q., Zhong, Y., Zhang, L., Li, D., 2022. Oil spill
UNACAR, 2019. Posicionamiento ante el evento de derrame de hidrocarburos en contextual and boundary-supervised detection network based on marine Sar images.
terminal marÍtima de cayo arcas en octubre de 2019. IEEE Trans. Geosci. Remote Sens. 60, 1–10.
United States. Coast Guard and National Response Team (US), 2011. On Scene
Coordinator Report: Deepwater Horizon Oil Spill. US Department of Homeland
Security, US Coast Guard.

13

You might also like