Integration of Statistical Methods and Artificial Neural Networks for

the Detection of Oil Stains in the Aquatic Environment

Monik Silva Sousa and João Viana da Fonseca Neto

Departamento de Engenharia Elétrica, Universidade Federal do Maranhão, São Luís, Brazil

Keywords: Linear Discriminant Analysis, Machine Learning, Oil Stain Detection, Classification, Predictive Model,

Artificial Neural Networks.

Abstract: The growth in oil exploration and transport increases the risk of accidents in the aquatic environment. Early

detection of oil slicks in the aquatic environment is essential to minimize the risk of accidents, as well as

effective decision-making. Thus, a method for detecting oil stains is needed to reduce the damage caused by

industrial activities to the environment. This article presents statistical methods of classification and machine

learning to detect oil slicks on the ocean surface. For this, images from a Synthetic Aperture Radar (SAR)

were used. The proposed model for detecting oil slicks uses Linear Discriminant Analysis (LDA) to generate

an estimate of the class to which the database images belong (image without oil slick, and image with oil

slick), and the Artificial Neural Network (ANN) to classify the data, in which these data come from the

grouping of the image with the result of the LDA. With the results obtained, it is concluded that the proposed

method of detecting oil slicks on the ocean surface can detect oil slicks with good accuracy.

1 INTRODUCTION

The world production of hydrocarbons, especially oil,

began in the mid-twentieth century and has since

grown exponentially. According to data from the

National Agency of Petroleum, Natural Gas and

Biofuels (ANP), total oil production in Brazil in 2019

was 1.017 billion barrels (ANP, 2020).

The growth in production, consumption, exports

and imports of oil and its derivatives increases the

probability of oil spill accidents. Consequently, it is

important to develop efficient techniques to prevent,

detect and monitor oil slicks.

Several studies are presented in the literature

related to oil spills for the detection of oil stains.

Some studies treat stain detection as a segmentation

problem (Krestenitis, M. O., 2019), in which they use

deep convolutional neural networks (DCNN) to

perform semantic segmentation of the image in

various areas of interest, including oil spill, which

makes it easier to detect the oil stain.

Remote sensing based on Synthetic Aperture

Radar (SAR) has been widely used for monitoring oil

slicks in the ocean, due to its wide area coverage and

its ability to operate in all climates. However, recently

Andreotti and Peixoto (2020) developed an

autonomous system (called ARIEL), combining an

autonomous vessel (unmanned surface vehicle) and a

drone. These two subsystems work together to

monitor a region of interest where oil spills are likely

to occur.

In this context of oil detection, the objective of

this work is the elaboration of an algorithm for the

detection of anomalies (oil stain on the surface of the

ocean), with time and memory restrictions, so that it

can be embedded in mobile sensors. To achieve this

goal, the integration of statistical methods and

artificial neural networks is proposed.

The focus of this work is the integration of

statistical methods (linear discriminant analysis) and

the simplest ANN (perceptron). Well, this detection

is part of a work in progress, which wants to adjust

the positioning of the sensor to monitor (track) the oil

stain, based on the geographic positioning of the

anomaly. Consequently, the effective contribution of

this work lies in the development of an algorithm for

detecting oil slicks on the ocean surface to be

embedded in the sensor, so that the detection of the

slick is done online, alerting if it detects an oil slick,

avoiding greater damage to the aquatic environment

and the coastal population.

Throughout this article, the steps of the proposed

methodology for the detection of oil slicks in the

550

Sousa, M. and Neto, J.

Integration of Statistical Methods and Artiﬁcial Neural Networks for the Detection of Oil Stains in the Aquatic Environment.

DOI: 10.5220/0011799300003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 550-557

ISBN: 978-989-758-626-2; ISSN: 2184-4313

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

aquatic environment are presented, based on the

integration of statistical methods and artificial neural

networks (ANNs). In Section 2, the proposed oil stain

detection system is presented, in which the block

diagram of the proposed predictive model is exposed

and from it it is reported on its blocks that are being

explored: SAR measurements and critical modeling.

Section 3 presents the results of computational

experiments resulting from the implementation of the

proposed integration method. In Section 4 the

conclusion is presented.

2 OIL STAIN DETECTION

SYSTEM

In this section, the proposed integration method is

presented, in which its scheme is illustrated by the

block diagram of Figure 1, then each block of the

diagram is reported as: SAR measurement block,

which exposes the base used in this work, block of the

critic, which is where the proposed method is located,

that is, where the Linear Discriminant Analysis

(LDA) and the Artificial Neural Network (ANN) are

located. In addition to also exposing the metrics to

evaluate the performance of this proposed method.

2.1 Proposed Predictive Method

The proposed predictive method for detecting oil

slicks on the oceanic surface is based on the

integration of statistical methods, multivariate data

analysis with the artificial neural network. The

multivariate data analysis technique is used to

estimate the class that the data belongs, that is, if

SAR image has an anomaly or not, then this data is

added to the concatenated vector of the image to add

more information, helping in the classification. This

method is contextualized as the critical module of the

system dedicated to decision making, considering the

process as the aquatic environment and oil slicks as

disturbances, and this process is monitored by

Synthetic Aperture Radar. Figure 1 illustrates this

system in a block diagram.

According to Figure 1, the block diagram of the

proposed predictive method, having as reference

signal the clean aquatic environment (without stains),

the critic is responsible for acquiring the classifier

model from the interaction of the LDA and of ANN,

which has as answer the classification of the image

with oil stain or without oil stain. This answer is used

for decision-making, because if an oil stain is

detected, it is necessary to apply certain measures to

contain the stains in order to minimize the

environmental impacts that they can cause.

Figure 1: Block diagram of the proposed predictive method.

The process that is aquatic that you want to

monitor, which can be close to the oil and gas

exploration and production industries and where there

is a large flow of ships, since most of the oil is

transported by ships. Oil slicks are considered as

process disturbances, and measurements of the

aquatic environment are performed by SAR. These

SAR measurements are inputs to the critic, which is

based on multivariate analysis of data and ANN, and

the key between the decision-making process and the

measures applied by an external individual.

2.2 SAR Measurements

Remote sensing systems have been widely used to

detect stains resulting from oil spills at sea. The radar

is a simple system that basically consists of the

transmission and reception of electromagnetic pulses,

the Synthetic Aperture Radar (SAR) is a form of radar

widely used to capture images, because as long as the

monitored systems are active, that is, they provide

with its own lighting, the SAR is capable of acquiring

images during the day and also at night, as its

radiation belongs to the microwave region.

In this study, the database provided by the Oil

Spill Detection Dataset – MKLab, which contains

1112 images, 1002 for training and 110 for testing,

used. In total there are 880 images with oil slicks on

the ocean surface and 232 clean images without oil

slicks. In Figure 1, this database is illustrated by the

block of images of SAR measurements that describe

instances of oil spills, similar (which look a lot like

an oil slick, but are not), land, sea and sea areas.

For the monitoring of the aquatic environment, the

SAR images contained in that base were used, which

were acquired through the missions of the European

satellite Sentinel-1, during the period from September

28, 2015 to October 31, 2017. Geographic

coordinates, and date and time of the pollution event

were provided by the European Maritime Safety

Integration of Statistical Methods and Artiﬁcial Neural Networks for the Detection of Oil Stains in the Aquatic Environment

551

Agency (EMSA) through the CleanSeaNet service

(Krestenitis, M. O., 2019).

The terrestrial range coverage of the SAR sensor

used by the Sentinel-1 mission is approximately 250

km, with pixel spacing equal to 10 × 10 m, therefore,

this radar can cover a large area of interest, in addition

to capturing relatively small-sized instances. This

employed system operates in the c-band, and the

polarization of its radar is double, that is, transmitted

vertical polarization – received vertical polarization

(VV) and transmitted vertical polarization – received

horizontal polarization (VH). To build the dataset,

only the raw data from the VV band was processed,

following a series of pre-processing steps to extract

common views (Krestenitis M. O., 2019). The

captured images have a dimension of 1250 × 650 ×

3 pixels, and in this work these images were resized

to 63 × 33 × 3 pixels.

Figure

2 illustrates an image

captured by this SAR system.

Figure 2: SAR measurements.

Based on Figure 2, an elongated black spot can be

seen, which stands out in the image. This format of

stains are usually oil spills and can be identified in the

database.

2.3 Critic's Modelling

In this section, the modelling of the critical block is

exposed, which is the integration of the linear

discriminant analysis with the artificial neural

network, to then perform the image classification. In

which, the LDA adds more information at the time of

classification.

2.3.1 Statistic Analysis

The statistical modelling proposed for detecting and

classifying oil slicks on the surface determined in

four steps: capture of non-segmented images,

dimensionality reduction, class estimation and finally

image classification.

Considering the oceanic surface, images of the

surface are first acquired, that is, the surface to be

monitored, then the dimensionality is reduced and the

classes are estimated via multivariate analysis

techniques. Finally, artificial neural networks are

used to classify the images, with the images being

classified into two groups: surface with oil stains and

clean surface (without oil stain).

 Linear Discriminant Analysis

The linear discriminant analysis (LDA) technique

was used to obtain class estimates from the 1.112

images (clean ocean surface or oil slick). The LDA is

used to analyse the relationship between a non-metric

dependent variable and a metric independent variable

(explanatory variables or also called predictors)

(Hair, J. F et al, 2019).

Algorithm 1: LDA.

The development of the LDA is carried out from

the following steps: the first step is to calculate the

mean of the classes, to then calculate the covariance

of the classes and then the common covariance matrix

of the classes. The second step is to calculate the

discriminant vector, which is used to construct the

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

552

differential equation that satisfactorily classifies the

classes, increasing the variance between classes and

minimizing the variance within classes (Duda, R. O.,

Hart, P. E., Stork, D. G., 2001). The last step is the

classification rule based on Fisher's discriminant

function, but for that it is also necessary to calculate

the midpoint between the classes. These steps of the

LDA algorithm are presented below, in Algorithm 1.

According to the LDA algorithm, we have 𝜇



the

average of class 𝑗, 𝑥



the samples of class 𝑗, 𝑛



the

number of images acquired from class j, c the number

of classes, 𝜙



the subtraction of data from class 𝑗 by

the average of class 𝜇



, 𝑆





the covariance matrix of

class 𝑗 , 𝑆



the common covariance matrix of the

classes, 𝑥 the input data for the LDA classifier, 𝑚 the

threshold, the midpoint between the classes, and

𝐷(𝑥) the output of the classifier.

2.3.2 Artificial Neural Networks

For classification purposes, a feedforward artificial

neural network (ANN) was also implemented, which

is widely used for the classification of separable

patterns. The ANN consists of several neurons with

their synaptic weights and bias, which are processed

by the linear combiner, given by

𝑢



∑

𝑤



∗𝑥









(1)

and

𝑦



=𝜑

(

𝑢



+𝑏



)

(2)

where 𝑥



,𝑥



,…,𝑥



are input signals, 𝑤



,𝑤



,…,𝑤



the synaptic weights of the neuron, 𝑢



the output of

the linear combiner due to the input signals,𝑏



or low,

𝜑(. ) the activation function, and 𝑦



the output signal

of the neuron. The adjustment of synaptic weights can

be done according to the error correction learning rule

(Haykin, 2009), which is given by

𝑤

(

𝑛+1

)

=𝑤

(

𝑛

)

+𝜂



𝑑

(

𝑛

)

𝑦

(

𝑛

)



𝑥

(

𝑛

)

(3)

where 𝑤(𝑛) is the current weight and 𝑤(𝑛 + 1) the

next iteration, 𝜂 the learning rate,𝑑(𝑛)𝑦(𝑛) the

error signal, which is the difference between the

desired signal and the output of the perceptron.

The learning algorithm adopted for training the

MLP network was error backpropagation, or also

called backpropagation, which is widely used for

supervised training. The algorithm adjusts the

network weights by calculating the difference

between the value estimated by the ANN and the

observed value (HAYKIN, 2011). Algorithm 2 shows

the steps to achieve the goal, which is to classify the

images via ANN.

In the present study, the variable of interest,

denoted by D, is the presence of an oil slick on the

ocean surface. Therefore, the training targets in the

ANN output layer were defined as follows: 1 –

presence of stain and 0 – otherwise.

Algorithm 2: MLP.

Different ANN architectures are considered in this

study. The choice of the number of neurons in the

hidden layer of the network, as well as the number of

hidden layers were done through experiments, always

looking for networks with a smaller number of hidden

neurons and with a good generalization power,

reducing the problem of overfitting (overfitting).

The formation of the training and test sets for the

implementation of the ANN used the division that

already came in the database.

2.3.3 Metrics for Performance Evaluation

The classifier may present an error in its

classification, consequently, it is necessary to use the

evaluation metrics to evaluate the performance of the

obtained predicted model. Consequently, the main

Integration of Statistical Methods and Artiﬁcial Neural Networks for the Detection of Oil Stains in the Aquatic Environment

553

objective of these metrics is to measure how far the

model is from the ideal classifier (classifier that does

not present errors). The following metrics used:

accuracy, precision, recall, F1 measure, and ROC

curve.

Accuracy (𝐴𝐶) measures how many images were

correctly classified, regardless of class, so the greater

the AC, the more the predicted model approaches an

ideal classifier, and is given by the ratio between the

number of correctly classified samples and the total

number of samples, given by

𝐴

𝐶=

𝑇𝑃 + 𝑇𝑁

𝑇𝑃

𝐹𝑁

𝑇𝑁

𝐹𝑃

(4)

where 𝑇𝑃 is the true positive are images showing oil

slicks on the ocean surface (positive) that are

classified correctly, 𝑇𝑁 being true negative are

images without oil slicks (negatives) that are

classified correctly, 𝐹𝑃 the false positive which are

the images without stains that are classified as

positive and 𝐹𝑁 the false negative which are the

positive images that are classified as negative.

Precision (𝑃) is defined by the ratio between the

number of images with oil stains classified correctly

and the total number of images classified as positive,

given by

𝑃=

𝑇𝑃

𝐹𝑃

(5)

Therefore, Equation (5) can be understood as the

number of images with oil stains that were classified

correctly.

Recall, also known as sensitivity, is defined by the

ratio between the number of images with oil stains

classified correctly and the number of images with oil

stains in the sample under study, given by

𝑟𝑒𝑐𝑎𝑙𝑙 =

𝑇𝑃

𝑇𝑁

(6)

The F1 metric takes into account both precision

and recall, it is defined as the harmonic mean between

precision and recall, given by

𝐹1 = 2 × (𝑃 × 𝑟𝑒𝑐𝑎𝑙𝑙)/(𝑃 + 𝑟𝑒𝑐𝑎𝑙𝑙).

(7)

According to Equation (7), it is noticed that if the

precision or the recall is equal to zero or very close to

it, the F1 will be low, so the classifier model is not a

model capable of getting its predictions right.

The ROC curve (Receiver Operating

Characteristic) and the Area Under the Curve (AUC)

also built to evaluate the performance of the ANN, as

they are important tools to compare the performance

of binary classification models. The ROC curve is a

graphical representation of the performance of a

quantitative data model according to its sensitivity

rate (fraction of true positives) and the fraction of

false positives.

It is used to evaluate the ability of a classifier to

distinguish between existing classes, in addition to

allowing visual analyses between precision and recall

relative to different cut-off points, where the false

positive rate is represented on the abscissa axis and

the true positive rate is plotted on the ordinate axis.

The AUC contributes to the interpretation of the ROC

curve, because, as its name implies, it is the value of

the area under the curve, and this value varies from 0

to 1 (or from 0% to 100%), in which the value of AUC

for 1 indicates that the classifier model is great and

that it did not make a prediction error, whereas if

AUC is 0 it indicates that the model is bad and that it

misclassified all inputs.

3 RESULTS

Considering that the objective of this work is to

develop an algorithm for detection of oil slicks on the

surface of the ocean. The proposed algorithm is based

on the integration of a statistical method and an

artificial neural network, as presented in the previous

sections.

To obtain the results, the following steps are

performed. First, the acquisition of the SAR

measurements is performed. Next, the discriminant

function is computed from the training data. Finally,

the tests are carried out and the results of these two

analyses (training and test) are stored, as they will be

used as additional information in the ANN, along

with the SAR image. These results of this procedure

are presented in the rest of the section.

First, an analysis of the distribution of images was

carried out, seeking to identify similarity between the

distribution of images without oil stain and images

with oil stain. This distribution analysis contributes to

the choice of the predictive model. For this, four

images from each group were randomly selected and

the histograms of the pixels of the images were

plotted, as shown in Figures 3 and 4.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

554

Figure 3: Pixel histogram of 4 images WITH oil stain.

According to Figure 3, the histograms of the oil

stain images also show similarity to each other, where

their peaks are close to 100, which shows that it has

more pixels in the darkest colour, close to the black

colour, and a normal distribution. When comparing

the histograms of each group, it is observed that there

is a difference between the distributions.

According to Figure 4, the histograms of the four

images without oil slicks on the ocean surface show a

similarity, where the peaks are closer to 150, that is,

there is a greater number of pixels that approach the

white colour. Another fact to be observed is that the

histograms present characteristics of a normal

distribution, which is the reference distribution for the

statistical methods used in this study.

Then, the linear discriminant function was

calculated using this training database, and the

discriminant function obtained from this database

was 100% correct, that is, it presented a true positive

rate of 100% and a false positive rate also 100%.

To verify the effectiveness of the discriminant

model, a test was performed with the test database,

which contains 110 images. The values of the

evaluation metrics are presented in Table 1.

Table 1: Test metrics with the LDA.

METRICS PERCENTAGE

Accurac

59.09

AUC 60.80

F1 69.39

Precision 86.44

Recall 57.95

According to Table 1, it is observed that the

discriminant function calculated from the LDA

presents a good classification, in which an accuracy

of 59.09% and a precision of 86.44%. Looking at the

Figure 4: Pixel histogram of 4 images WITHOUT oil stain.

ROC curve, and knowing that it is a graphical

technique used to assess the ability of the predictive

model to perform proper classification, Figure 5

illustrates the ROC curve for the predictive model.

Figure 5: ROC curve for the LDA.

In order to improve the classification, the

multilayer perceptron network was used to classify

the database, plus the LDA estimate added to the SAR

image database, adding more information and the

results are presented below.

The first test carried out using a multilayer

perceptron (MLP), consisting of an intermediate layer

with 6237 neurons, which are the number of

parameters of the input data, having as input only the

images, in which the training bank consists of 1002

images and the test bank consists of 110. The results

obtained for this test are shown in Table 2.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Taxa de Falso Positiv

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

pç

Integration of Statistical Methods and Artiﬁcial Neural Networks for the Detection of Oil Stains in the Aquatic Environment

555

Table 2: Training and testing metrics, input - images.

METRICS TRAINING TEST

Accurac

74.35 69.09

F1 62.00 70.00

Precision 75.00 71.00

Recall 74.00 69.00

According to Table 2, it is observed that the MLP

network for images as input presented better metric

values for training than for testing. In addition to

being more efficient in classifying images that report

oil spills on the ocean surface.

The second analysis performed on the MLP

network, with an intermediate layer and 6238

neurons, which is the number of parameters of the

input data with the classification obtained by the

LDA. The results obtained for this test are shown in

Table 3.

Table 3: Training and testing metrics, input - images plus

LDA rating.

METRICS TRAINING TEST

Accurac

99.90 65.45

F1 100.00 68.00

Precision 100.00 72.00

Recall 100.00 65.00

According to Table 3, the MLP was ideal in

training, as metrics equal to 100% were obtained, but

in the test, it presented values close to the test using

only the images as input. However, it increased

precision, and like the previous one, it was more

effective in classifying images with oil stain.

The third analysis performed on the MLP

network, with an intermediate layer and 6238

neurons, which is the number of parameters of the

input data with the value obtained by the linear

discriminant function. The results obtained for this

test are shown in Table 4.

Table 4: Training and testing metrics, input - images plus

response from the discriminant function.

METRICS TRAINING TEST

Accurac

78.94 63.63

F1 71.00 67.00

Precision 82.00 73.00

Recall 79.00 64.00

According to Table 4, the MLP was good in

classifying the images with oil stains, and observing

the three tests, it was concluded that the added LDA

information helped in the training and testing of the

network, presenting better values for the metrics of

evaluation.

The proposed method integrates approaches from

the multivariate analysis technique, Linear

Discriminant Analysis (LDA) and Artificial Neural

Network (ANN) to allow the detection of oil slicks on

the ocean surface, alerting if an oil spill is detected.

According to the results presented, the developed

algorithm presented a good average for the evaluation

metrics and proved to be more efficient than the

classification performed by the methods of the two

separate approaches.

4 CONCLUSIONS

In view of the growth in oil production and

transportation, techniques are needed to monitor the

aquatic environment, and to detect oil spills or leaks.

Consequently, the proposed method of integrating

statistical techniques and artificial neural networks

presents a means for detecting oil slicks on the water

surface.

This study shown the possibility of detecting oil

slicks on the surface of the aquatic environment

without the need for image segmentation, being able

to automate the detection of oil slicks on the surface

of the ocean, alerting if an oil spill is detected.

In possession of the results, the statistical

technique and artificial neural network integration

method showed greater accuracy than the separate

techniques, showing that the proposed method is

more efficient for oil stain detection than the linear

discriminant analysis methods. and multilayer

perceptron network, isolated. This algorithm,

therefore, proved to be able to be embedded in a

sensor.

Due to the integration of multivariate analysis

methods and artificial neural networks, the results

showed that the proposed method is more efficient for

oil stain detection than linear discriminant analysis

methods and multilayer perceptron network without

integration. It was observed that the integration of the

two approaches presented greater precision than the

same ones applied separately. Therefore, the

proposed methodology is satisfactory to be embedded

in a sensor node to perform local classification based

on oil spills or other types of intrusions in coastal

areas.

To continue the work and compare the proposed

method with a network that uses deep learning, a Unet

algorithm is being developed to detect the oil stain,

and the next step is the algorithm to adjust the

positioning of the sensor so that it can monitor the oil

slick.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

556

ACKNOWLEDGEMENTS

We thank the ANP Human Resources Program,

which financed the research, the Federal University

of Maranhão (UFMA) and its Graduate Program in

Electrical Engineering (PPGEE), CNPQ, and Higher

Education Personnel Improvement Coordination

which encouraged the study in this area of great

interest and importance technological.

REFERENCES

ANP. (2020). Agência Nacional do Petróleo, Gás Natural e

Biocombustíveis. Acesso em 25 de março de 2020,

disponível em ANP: http://www.anp.gov.br/

Brekke, C.; Solberg, A.H. Oil spill detection by satellite

remote sensing. Remote Sens. Environ. 2005, 95, 1–13.

Carlos Alberto Alves Varella. Análise Multivariada

Aplicada as Ciências Agrárias. Rio de Janeiro.

Carvalho, G. d. (2015). Multivariate Data Analysis of

Satellite-Derived Measurements to Distinguish Natural

From Man-Made Oil Slicks On The Sea Surface of

Campeche Bay (Gulf of Mexico). Rio de Janeiro,

Brasil.

CETESB. (2020). Planos de Contingência para Vazamento

de Óleo no Mar. Acesso em 09 de Abril de 2020,

disponível em CETESB: https://cetesb.sp.gov.br/

emergencias-quimicas/tipos-de-acidentes/vazamentos-

de-oleo/preparacao-para-resposta/planos-de-contingen

cia-para-vazamento-de-oleo-no-mar/.

DONGMEI SONG, Z. Z. (2020). A Novel Marine Oil

Spillage Identification Scheme Based on Convolution

Neural Network Feature Extraction From Fully

Polarimetric SAR Imagery. IEEE Access, 08, 59801 a

59820.

Hair, J. F., Black, B., Babin, B., Anderson, R. E., & Tatham,

R. L. (2009). Análise Multivariada de Dados (6 ed.).

Porto Alegre, Brasil: Bookman.

Haykin, S. (2009). Neural networks and learning machines.

Pearson.

HOTELLING, H. “Multivariate Quality Control” in

Techniques of Statistical Analysis. McGraw Hill, New

York, 1947.

Ivan Nunes da Silva; Danilo Hernane Spatti; Rogério

Andrade Flauzino (2010). Redes neurais artificiais:

para engenharia e ciências aplicadas. São Paulo:

Artliber.

Joseph F. Hair, J., Black, W. C., Babin, B. J., Anderson, R.

E., & Tatham, R. L. (2009). Análise Multivariada de

Dados (6 ed.). Porto Alegre, Brasil: Bookman.

Krestenitis, M. O. (Janeiro de 2019). Early Identification of

Oil Spills in Satellite Images Using Deep CNNs. (C.

Springer, Ed.) In International Conference on

Multimedia Modeling, 424 - 435.

Krestenitis, M. O. (2019). Oil Spill Identification from

Satellite Images Using Deep Neural Networks. Remote

Sensing, 11, 1762.

Marcelo Andreotti, R. S., Alessandro Jacoud Peixoto, J. C.

(2020). ARIEL: An Autonomous Robotic System for

Oil Spill Detection. Offshore Technology Conference.

Pereira, S. B. (2001). Investimentos em automação no

Projeto PEGASO da PETROBRAS chegam a US$ 225

milhões. Revista InTech nº 32.

Petrobras. (2020). Exploração e Produção de Petróleo e Gás.

Acesso em 08 de julho de 2020, disponível em Petrobras:

https://petrobras.com.br/pt/nossas-atividades/areas-de-

atuacao/exploracao-e-producao-de-petroleo-e-gas/.

Richard O. Duda, P. E. (2001). Pattern Classification

(Second ed.). Canada: Wiley Interscience Publication.

Selmic, R. R., Phoha, V. V., & Serwadda, A. (2016).

Wireless Sensor Networks Security, Coverage, and

Localization. USA: Springer International Publishing.

Woods, R. C. (2010). Processamento digital de imagens (3

ed.). São Paulo, Brasil: Pearson Prentice Hall.

Integration of Statistical Methods and Artiﬁcial Neural Networks for the Detection of Oil Stains in the Aquatic Environment

557