Oil Spill Detection and Visualization from UAV Images using

Convolutional Neural Networks

Val

erio N. Rodrigues Junior

, Roberto J. M. Cavalcante

, Jo

ao A. G. R. Almeida

, Tiago P. M. F

Ana C. M. Malhado

, Thales Vieira

1 a

and Krerley Oliveira

3 b

Institute of Computing, Federal University of Alagoas, Macei

o, AL, Brazil

Institute of Biological and Health Sciences, Federal University of Alagoas, Macei

o, AL, Brazil

Institute of Mathematics, Federal University of Alagoas, Macei

o, AL, Brazil

Keywords:

Oil Spill, Convolutional Neural Network, Deep Learning, Ummanned Aerial Vehicles, Geospatial Data

Analysis.

Abstract:

Marine oil spills may have devastating consequences for the environment, the economy, and society. The

2019 oil spill crisis along the northeast Brazilian coast required immediate actions to control and mitigate the

impacts of the pollution. In this paper, we propose an approach based on Deep Learning to efﬁciently inspect

beaches and assist response teams using UAV imagery through an inexpensive visual system. Images collected

by UAVs through an aerial survey are split and evaluated by a Convolutional Neural Network. The results are

then integrated into heatmaps, which are exploited to perform geospatial visual analysis. Experiments were

carried out to validate and evaluate the classiﬁers, achieving an accuracy of up to 93.6% and an F1 score of

78.6% for the top trained models. We also describe a case study to demonstrate that our approach can be used

in real-world situations.

1 INTRODUCTION

Marine oil spills are one of the highest proﬁles

and ecologically destructive polluting events with so-

cial and environmental consequences that can last

many years after the oil has been removed/dispersed

(Burger, 1997; Kingston, 2002).

For instance, the highly publicized 1989 spill of

oil from the Exxon Valdez into Prince William Sound,

Alaska was initially responsible for an enormous in-

crease in mortality followed by prolonged sub-lethal

effects that have led to the postponed recovery of

many species (Peterson et al., 2003). Likewise, a re-

cent review of the 2010 Deepwater Horizon oil spill

in the northern Gulf of Mexico (Beyer et al., 2016) re-

vealed a multiplicity of biological effects, with long-

term impacts on large ﬁsh species, deep-sea corals,

sea turtles, and cetaceans. Both the Exxon Valdez and

Deepwater Horizon spills had a clear point of origin in

time and space, with scientists able to closely monitor

the spread of oil and the consequences of the pollution

very soon after the spill had occurred.

https://orcid.org/0000-0001-7775-5258

https://orcid.org/0000-0002-7385-3114

This is very different from the recent (and still

mysterious) oil spill in Brazil whose origin and timing

are still unclear despite intensive investigation (Ma-

gris and Giarrizzo, 2020; Zacharias et al., 2021). The

ﬁrst indication that a major oil spill had occurred off

the northeast coast of Brazil was the presence of large

quantities of oil on beaches in August/September

2019 (Soares et al., 2020). Up to March 2020 re-

ports of oiled areas were recorded in nearly 550 sites

spanning 3000 km of coastline, affecting up to 55 en-

vironmental protection areas (Ladle et al., 2020). A

study in Alagoas State indicates that ﬁsh and seafood

sales decreased by more than 50% strongly impact-

ing some of the most economically vulnerable com-

munities of the northeast region. Likewise, tourism

in the area was also dramatically affected (Ribeiro

et al., 2021). Preliminary data suggests that diverse

ecosystems were affected, including seagrasses (Ma-

galh

aes et al., 2021), unique rhodolith beds (Sissini

et al., 2020) and reef-building corals (Miranda et al.,

2020). Soares et al. (Soares et al., 2020) identify four

characteristics that make this oil spill unique: 1) the

characteristics of the oil spill; 2) the characteristics

of the affected region in tropical Brazil; 3) the scale

of the disaster and; 4) the absence of measures and/or

Rodrigues Junior, V., Cavalcante, R., Almeida, J., Fé, T., Malhado, A., Vieira, T. and Oliveira, K.

Oil Spill Detection and Visualization from UAV Images using Convolutional Neural Networ ks.

DOI: 10.5220/0010802600003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 5: VISAPP, pages

331-338

ISBN: 978-989-758-555-5; ISSN: 2184-4321

331

ﬂaws in the measures taken by the federal government

to address this environmental and social emergency.

One of the major challenges in the early stage

of such a disaster is to quickly collect accurate spa-

tiotemporal oil pollution data (Soares et al., 2020).

In this sense, Deep Learning (DL) approaches have

achieved outstanding results in object detection and

localization tasks during the last years (Redmon et al.,

2016; Ren et al., 2016; Massa et al., 2021). Specif-

ically, geospatial data analysis has greatly beneﬁted

from Deep Learning algorithms to tasks such as crop

type detection (Kussul et al., 2017) and road extrac-

tion (Zhang et al., 2018), for instance. SAR remote

sensing has been proposed for microplastic pollution

in oceans (Davaasuren et al., 2018) and for marine

oil spill detection (Shaban et al., 2021). In particu-

lar, the latter study is focused on segmenting large oil

spills offshore using satellite images. In this work,

we tackle a different problem: detecting and localiz-

ing tiny oil spills that may be washing up on beaches

or inshore, using low cost unmanned aerial vehicles

(UAV’s).

A recent review (Fingas and Brown, 2018)

claimed that oil spill detection using the visible spec-

trum is challenging, limiting the use of UAV’s to per-

form such a task. More recently, Jiao et al. (Jiao et al.,

2019) proposed a DL method to inspect facilities us-

ing UAV imagery, based on the Faster R-CNN (Ren

et al., 2016). However, in that work oil spills’ appear-

ance, size and backgrounds are signiﬁcantly different

than in the problem we address herein. Consequently,

we consider it a substantially different Computer Vi-

sion problem. We refer the reader to (Cazzato et al.,

2020) for a comprehensive presentation on computer

vision methods for UAV object detection.

In this paper, we propose a DL approach to assist

response teams in the cleaning of oil spills washing

up on beaches. In this situation, oil is usually scat-

tered in many small connected compounds, which are

found in a diversity of different backgrounds (sand,

water, seaweed,...), as shown in Figures 1 and 5b.

We exploit images acquired by low-cost UAVs to per-

form oil detection through an automated image anal-

ysis technique based on DL, achieving up to 93.6% of

accuracy and an F

score of 78.6%. We also validate

and evaluate our approach through a visual system

that combines satellite imagery and heatmaps of oil

density, to promptly and accurately notify response

teams, thus avoiding signiﬁcant environmental im-

pacts.

The core of our approach is a Convolutional

Neural Network (CNN) classiﬁer, whose architecture

was optimized to accurately recognize oil in small

patches. Differently from more advanced deep neu-

ral network architectures, we opt for a small archi-

tecture specially trained to identify small oil spills in

coastal land covers (i.e., ﬂoating on the water surface

or washing up on beaches). To integrate numerous

CNN predictions in a visual manner, we propose a

practical method to promptly and accurately notify re-

sponse teams, thus avoiding signiﬁcant environmental

impacts.

Overall, our main contribution is to investigate

whether a low-cost DL classiﬁer, jointly with a visual

interface, may provide quick, accurate and intuitive

georeferenced information about small oil spills lo-

cated on the beach or inshore, for rapid response pur-

poses. In particular, the whole approach relies only

on an inexpensive UAV and on a small neural network

architecture, which is trainable in a low-cost desktop

computer.

2 METHODOLOGY

Our approach is based on georeferenced RGB images

collected through an aerial survey, by using UAVs. As

illustrated in Figure 1, each image is ﬁrst split into a

grid of small patches. In the training phase, patches

are manually annotated with a binary label indicating

the presence of oil. Next, the supervised dataset is

exploited to train a CNN binary classiﬁer, which is

then expected to recognize oil from small patches. In

the detection phase, UAV images acquired through an

aerial survey are ﬁrst split into patches that are eval-

uated by the CNN. The resulting predictions are or-

ganized in oil maps representing the occurrence of oil

in the region covered by each particular UAV image.

The oil maps are combined with the original UAV im-

ages geolocations to build a heatmap (Nogueira et al.,

2019) of the region covered in the aerial survey. Fi-

nally, the heatmap is superimposed over satellite im-

agery.

2.1 Imagery Acquisition, Preprocessing

and Annotation

Firstly, a UAV collects georeferenced RGB imagery

covering a region of interest, where oil may be po-

tentially found. Each collected image covers a region

that may contain superﬁcial oil only in small subar-

eas. Thus, to provide precise locations of oil spills,

we propose to ﬁrst split the captured images into small

squared patches before training a CNN classiﬁer ca-

pable of providing patch-level oil spills detections.

To train the CNN binary classiﬁer, a dataset com-

prised of many of the aforementioned patches must

be annotated by human trainers, in a binary fashion.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

332

annotation

training set

CNN

Oil maps

Heatmap

Georeferenced UAV images Patches split Satellite imagery

Geospatial Visualization

Figure 1: Overview of our approach, with two distinct phases: training a CNN classiﬁer from small RGB patches (green

arrows); and recognizing, localizing, and visualizing oil spots (blue arrows). As a common preprocessing step to both phases,

UAV’s georeferenced RGB imagery, acquired through an aerial survey, is split into small RGB patches.

In our experiments, we empirically concluded that

this may be accomplished visually, without relying on

prior oil spill geospatial information.

An appropriate patch size must satisfy both human

trainers, who must be capable of visually recognizing

oil in the patches; and the CNN classiﬁer. In our ex-

periments, we empirically found that a CNN could

still achieve top results when each input patch was set

to 128 × 128 pixels. This patch size was also satis-

factory for the human annotation task. More speciﬁ-

cally, we collected UAV images with spatial dimen-

sions of 4864 × 3648 pixels, which were split into

1064 patches with a size of 128 × 128 pixels. A thor-

ough investigation of optimal patch size parameters

may be conducted in future work.

2.2 Supervised Classiﬁcation

We adopt a classical CNN architecture but performed

many experiments to tune hyperparameters. The in-

put patches are ﬁrst ﬁltered by C blocks of convo-

lutional layers with F ﬁlters of size K, followed by

max-pooling layers. Then, the resulting feature maps

are ﬂattened and given as input to one or two dense

layers with D

and D

units each. The output layer

is made up of a single neuron with a sigmoid acti-

vation function, which outputs the probability of oil

occurrence in the input patch. The hyperparameter

search space is shown in Table 1 and the best classi-

ﬁer found in our experiments is revealed in Figure 2.

More details on our experiments on hyperparameter

optimization will be given in Section 3.

2.3 Postprocessing and Visual

Geospatial Analysis

We propose to visually analyze the oil spill distribu-

tion in a speciﬁc area by computing heatmaps that in-

tegrate the outputs of the CNN classiﬁer, and the geo-

coordinates of the images. Let I be an image, which

is split into a grid G

with m × n patches; and let

[i, j] ∈

{

0,1

}

be the binary CNN prediction of the

patch at the position [i, j] of G

. Note that f

is the oil

map of image I, as shown in Figure 1. We deﬁne the

density of oil in I, as the mean value of the oil map:

m · n

∑

i=1

∑

j=1

[i, j]. (1)

By combining information from many images col-

lected to cover a speciﬁc area of interest, we build a

polygonal heatmap by interpolating the densities d

using the geocoordinates g

of image I. To this aim,

we adopt the well-known Inverse Distance Weighting

(IDW). Finally, to perform visual geospatial analysis,

we employed the QGIS

system to plot the resulting

heatmap over satellite imagery of the area of interest.

3 EXPERIMENTS

We performed experiments to validate the whole ap-

proach, compare and optimize CNN hyperparameters,

and evaluate the best model performance. We also

present a case study to investigate whether the pro-

posed visual analysis system is appropriate to easily

reveal oil pollution patterns in the images.

With this aim, we collected a dataset from APA

Costa dos Corais – a coastal marine conservation unit

Table 1: Hyperparameter search space.

hyperparameter values

C 1, 2 or 3

F 16, 32 or 64

K 3 or 5

25, 50 or 100

0, 25, 50 or 100

https://qgis.org

Oil Spill Detection and Visualization from UAV Images using Convolutional Neural Networks

333

3@128x128

Input

Convolutional layer

32@128x128

Convolutional layer

32@64x64

Convolutional layer

32@32x32

Max-pool

Flatten

Dense

Output

Figure 2: Architecture of the top performing CNN.

located in the Brazilian states of Alagoas and Pernam-

buco. More speciﬁcally, data collection was carried

out in the municipality of Japaratinga, Alagoas, in the

peak phase of the crisis. A DJI Phantom 4 Advanced

drone and the Dronedeploy

tool were used. Four

ﬂights were made, all at 60 meters high, collecting

vertical images with a resolution of 1.8 centimeters

per pixel and no ﬁlter. On average, each ﬂight took

11 minutes and covered an area of roughly 70,000m

resulting in approximately 200 images. The paths of

the ﬂights were planned to cover different types of ar-

eas including sand, seaweed, trees, seawater, and river

water. Each ﬂight covered a disjoint region of inter-

est. Optimal ﬂight hyperparameters, such as ﬂight al-

titude, may be a topic of future research.

Some of the captured UAV images were manu-

ally chosen to compile a training set. Visual diver-

sity was considered the main selection criterion, to

assure that all types of covered areas were signiﬁ-

cantly represented. To avoid erroneous annotations,

annotation redundancy (Puttemans et al., 2018) was

adopted: each patch was annotated by at least two

humans, and inconsistent annotations were then re-

vised by a third human annotator. The ﬁnal training

set was comprised of 6,907 patches, which was split

into 5,180 negative samples (not containing oil) and

1,727 positive samples (containing oil).

3.1 CNN Validation and

Hyper-parameters Optimization

We performed a grid search over the hyperparame-

ter search space described in Section 2.2 and sum-

marized in Table 1. All possible combinations were

experimented with through a cross-validation proce-

dure by randomly splitting the dataset examples into

training and test sets, considering 75% of the patches

for training and 25% for testing, in a stratiﬁed man-

ner. This procedure was repeated 10 times, and the

average value was considered.

To train the networks, we adopted the Adam op-

timization algorithm (Kingma and Ba, 2014) of the

Keras library (Chollet et al., 2015), which was em-

ployed to minimize the well-known categorical cross-

https://www.dronedeploy.com

entropy loss with an initial learning rate of 0.001.

Class imbalance of the dataset was tackled by setting

each class weight to be inversely proportional to their

respective frequency. However, since false negatives

(patch containing oil, but not detected) are considered

more harmful than false positive results (patch does

not contain oil, but oil was detected) in this problem,

the positive class weight was doubled. To put it an-

other way, a false alarm would result in wasting re-

sponse personnel resources. Missing oil spills, how-

ever, is much more harmful, due to environmental im-

pacts.

Table 2 reveals the top 10 architectures, according

to their F

score. One can see promising results, with

many models achieving up to 93.59% of accuracy and

an F

score of 78.6%. By visually examining some

incorrectly predicted images shown in Figure 3, we

found challenging and unusual situations, including

very small oil spills hidden in seaweed; oil spills, re-

sembling shadows in the water; a person using a black

t-shirt, which looks similar to oil; and seaweed mixed

with small black dots, which is similar to previously

labeled small oil spills.

We also infer that, fortunately, the CNNs are not

much sensitive to hyperparameters values in terms of

accuracy: all top-performing networks achieved sim-

ilar accuracy, even though their number of trainable

weights varies greatly.

Nevertheless, this ﬁnding does not hold for the F

score, precision and recall, which are more suitable

measures since the dataset is imbalanced. Further-

more, a high recall is more relevant than a high pre-

(a) False negative patches.

(b) False positive patches.

Figure 3: Examples of incorrect predictions.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

334

Table 2: Conﬁguration, F

score, accuracy, precision, recall and number of trainable weights of the top 10 CNN conﬁgurations

and the compared methods, sorted by F

score. Best results in bold.

C F K D

acc. prec. rec. weights

(%) (%) (%) (%)

3 32 3 50 0 78.6 93.6 84.9 73.3 333,144

3 32 3 50 50 77.7 92.6 75.4 80.1 335,694

3 64 3 100 0 77.3 92.6 76.2 78.4 1,330,350

3 16 3 25 0 77.2 93.5 88.9 68.2 83,565

2 32 3 25 25 75.7 92.3 77.0 74.4 730,871

3 16 3 50 0 75.1 92.8 84.4 67.6 162,040

2 32 3 100 50 74.9 92.0 76.0 73.9 2,895,396

3 64 3 50 25 74.8 92.6 82.7 68.2 704,225

3 64 3 50 0 72.7 92.8 92.9 56.6 703,000

3 64 3 100 50 67.4 91.8 94.8 52.3 1,335,300

thresholding 56.9 71.0 48.2 69.6 -

HOG+SVM 22.0 72.0 24.0 20.0 -

cision, as aforementioned. The F

score of the top

CNNs ranges from 67.4% to 78.6%, revealing high

variance. This mostly occurred due to changes in

the precision-recall tradeoff. Overall, we consider the

second-best model to be an appropriate choice, since

it achieved a recall of 80.1%, while still keeping a

high accuracy of 92.6%. Besides, it is a small net-

work with only 335,694 trainable weights. It is also

worth considering that oil spills rarely occur in a sin-

gle patch of the UAV image. Thus, a single UAV im-

age comprised of several patches with oil spills will

have its density of oil (Equation 1) underestimated by

only approximately 20%, if the second-best model is

used.

In summary, the presented results validate our ma-

chine learning approach and indicate that it may be

effectively employed in real-world scenarios. As an

alternative to avoid the costs related to false positive

predictions, we suggest developing a more sophis-

ticated visual interface where a human would visu-

ally validate reddish spots, before sending a response

team.

3.2 Comparison with Baseline Image

Classiﬁcation Methods

We compare the results of our CNN classiﬁers to the

traditional HOG+SVM (Han et al., 2006) approach,

which combines Histogram of Gradients features with

a linear Support Vector Machines (SVM) classiﬁer;

and to a baseline thresholding method (Vyas et al.,

2015), since thresholding is a common oil spill detec-

tion approach (Al-Ruzouq et al., 2020). Class weight-

ing was similarly employed for the compared meth-

ods, as described in Section 3.1. The results shown

in Table 2 reveal inferior results for the HOG+SVM

method, with a poor F

score of 22%. Threshold-

ing achieved intermediate results, with an F

score of

56.9% but still inferior to the top 10 CNN conﬁgura-

tions. In this latter method, it is also worth noting a

reasonable recall of 69.6%, which may be exploited

in future work to improve the results of the CNN.

3.3 Case Study

In this case study, we followed the discussion of Sec-

tion 3.1 and opted for the second-best model of Ta-

ble 2, since it achieves the higher recall while keep-

ing high accuracy and a small number of trainable

weights, thus resulting in an accurate and compu-

tationally efﬁcient model. The selected model was

employed to predict all patches in a large area of

70,000m

composed of 200 images collected by a

drone ﬂying over Japaratinga. Prediction time was

approximately only 3 secs in a 2.5Ghz Intel i5 CPU

with 8GB of RAM, since patches prediction occurred

in batches.

After applying the postprocessing steps described

in Section 2.3, the heatmap shown in Figure 4 was ob-

tained. One can see the ﬂight area bounded by a black

polygon, with color-coded oil densities. The heatmap

reveals a red spot in the northern part of the polygon,

indicating a concentration of oil in that region that

should be cleaned. It is worth noting that the loca-

tion of this speciﬁc ﬂight was chosen because of high

concentrations of oil spills that were localized approx-

imately in the reddish spots of Figure 4. This is quali-

tative evidence that validates our method. Figure 5 ex-

hibit examples of a variety of correct predictions, in-

cluding challenging situations in the sand, water, and

seaweed. We thus conclude that the whole proposed

approach is practical to rapidly identify oil pollution

in beaches and allow a prompt response, minimizing

Oil Spill Detection and Visualization from UAV Images using Convolutional Neural Networks

335

Figure 4: Visualization of the resulting heatmap built from images collected in the municipality of Japaratinga, following the

pipeline shown in Figure 1. Our methodology allows to quickly identify a red spot indicating the presence of oil pollution in

the beach.

(a) True negative predictions. (b) True positive predictions.

Figure 5: Examples of correctly predicted patches of the heatmap shown in Figure 4.

environmental impacts. It is worth mentioning, how-

ever, that the trained model may not generalize to dif-

ferent coastal environments, requiring retraining the

model.

4 CONCLUSION

In this paper, we proposed a DL approach based on

low-cost UAVs images to efﬁciently identify and visu-

ally localize oil ﬂoating on the water surface or wash-

ing up on beaches. By combining a CNN, trained in

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

336

a supervised manner to evaluate small patches, with

a geospatial visual system, we showed that oil spills

can be rapidly detected and localized, allowing imme-

diate reactions and avoiding substantial environmen-

tal impact. Our classiﬁers achieved up to 93.6% of

accuracy and an F

score of 78.6% in a small dataset

comprised of 6,907 patches, revealing promising re-

sults that may be enhanced in future work by exploit-

ing larger datasets and more elaborate DL techniques,

such as image segmentation networks.

ACKNOWLEDGMENT

This work is part of the Long Term Eco-

logical Research – Brazil site PELD-CCAL

(Projeto Ecol

ogico de Longa Durac¸

ao -

Costa dos Corais, Alagoas) funded by CNPq

–(#441657/2016 − 8, #442237/2020 − 0), and

FAPEAL (#60030.1564/2016). This research was

partially ﬁnanced by the Justice Court of Alagoas

through the I Workshop on Mathematical Solutions

in Justice and Tourism.

REFERENCES

Al-Ruzouq, R., Gibril, M. B. A., Shanableh, A., Kais,

A., Hamed, O., Al-Mansoori, S., and Khalil, M. A.

(2020). Sensors, features, and machine learning for

oil spill detection and monitoring: A review. Remote

Sensing, 12(20):3338.

Beyer, J., Trannum, H. C., Bakke, T., Hodson, P. V., and

Collier, T. K. (2016). Environmental effects of the

deepwater horizon oil spill: a review. Marine pollu-

tion bulletin, 110(1):28–51.

Burger, J. (1997). Oil spills. Rutgers University Press.

Cazzato, D., Cimarelli, C., Sanchez-Lopez, J. L., Voos, H.,

and Leo, M. (2020). A survey of computer vision

methods for 2d object detection from unmanned aerial

vehicles. Journal of Imaging, 6(8):78.

Chollet, F. et al. (2015). Keras. https://keras.io.

Davaasuren, N., Marino, A., Boardman, C., Alparone, M.,

Nunziata, F., Ackermann, N., and Hajnsek, I. (2018).

Detecting microplastics pollution in world oceans us-

ing sar remote sensing. In IGARSS, pages 938–941.

IEEE.

Fingas, M. and Brown, C. E. (2018). A review of oil spill

remote sensing. Sensors, 18(1):91.

Han, F., Shan, Y., Cekander, R., Sawhney, H. S., and Kumar,

R. (2006). A two-stage approach to people and ve-

hicle detection with hog-based svm. In Performance

Metrics for Intelligent Systems 2006 Workshop, pages

133–140.

Jiao, Z., Jia, G., and Cai, Y. (2019). A new approach to oil

spill detection that combines deep learning with un-

manned aerial vehicles. Computers & Industrial En-

gineering, 135:1300–1311.

Kingma, D. P. and Ba, J. (2014). Adam: A

method for stochastic optimization. arXiv preprint

arXiv:1412.6980.

Kingston, P. F. (2002). Long-term environmental impact of

oil spills. Spill Science & Technology Bulletin, 7(1-

2):53–61.

Kussul, N., Lavreniuk, M., Skakun, S., and Shelestov, A.

(2017). Deep learning classiﬁcation of land cover and

crop types using remote sensing data. IEEE Geo-

science and Remote Sensing Letters, 14(5):778–782.

Ladle, R., Malhado, A. C., Campos-Silva, J., and Pinheiro,

B. (2020). Brazil’s mystery oil spill: an ongoing social

disaster. Nature, 578(7793):37–37.

Magalh

aes, K. M., de Souza Barros, K. V., de Lima, M.

C. S., de Almeida Rocha-Barreira, C., Rosa Filho,

J. S., and de Oliveira Soares, M. (2021). Oil spill+

covid-19: A disastrous year for brazilian seagrass

conservation. Science of The Total Environment,

764:142872.

Magris, R. A. and Giarrizzo, T. (2020). Mysterious oil

spill in the atlantic ocean threatens marine biodiversity

and local people in brazil. Marine pollution bulletin,

153:110961.

Massa, L., Barbosa, A., Oliveira, K., and Vieira, T. (2021).

Lrcn-retailnet: A recurrent neural network architec-

ture for accurate people counting. Multimedia Tools

and Applications, 80(4):5517–5537.

Miranda, R. J., Almeida, E. C., Pinto, T. K., Sampaio, C. L.,

Pereira, P. H., Nunes, J. A., and Ladle, R. J. (2020).

Re: Oil spill disaster in brazil: impact assessment ne-

glecting unique coral reefs. Science.

Nogueira, V., Oliveira, H., Silva, J. A., Vieira, T., and

Oliveira, K. (2019). Retailnet: A deep learning ap-

proach for people counting and hot spots detection in

retail stores. In SIBGRAPI, pages 155–162. IEEE.

Peterson, C. H., Rice, S. D., Short, J. W., Esler, D., Bodkin,

J. L., Ballachey, B. E., and Irons, D. B. (2003). Long-

term ecosystem response to the exxon valdez oil spill.

Science, 302(5653):2082–2086.

Puttemans, S., Van Beeck, K., and Goedem

e, T. (2018).

Comparing boosted cascades to deep learning archi-

tectures for fast and robust coconut tree detection in

aerial images. In Proceedings of the 13th interna-

tional joint conference on computer vision, imaging

and computer graphics theory and applications, vol-

ume 5, pages 230–241. SCITEPRESS.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time ob-

ject detection. In CVPR, pages 779–788.

Ren, S., He, K., Girshick, R., and Sun, J. (2016). Faster

r-cnn: towards real-time object detection with region

proposal networks. IEEE transactions on pattern

analysis and machine intelligence, 39(6):1137–1149.

Ribeiro, L., Souza, K., Domingues, E., and Magalh

aes,

A. S. (2021). Blue water turns black: economic im-

pact of oil spill on tourism and ﬁshing in brazilian

northeast. Current Issues in Tourism , 24(8):1042–

1047.

Oil Spill Detection and Visualization from UAV Images using Convolutional Neural Networks

337

Shaban, M., Salim, R., Abu, H., Kheliﬁ, A., Shalaby, A., El-

Mashad, S., Mahmoud, A., Ghazal, M., and El-Baz,

A. (2021). A deep-learning framework for the detec-

tion of oil spills from sar data. Sensors, 21(7):2351.

Sissini, M., Berchez, F., Hall-Spencer, J., Ghilardi-Lopes,

N., Carvalho, V., Schubert, N., Koerich, G., Diaz-

Pulido, G., Silva, J., et al. (2020). Brazil oil

spill response: Protect rhodolith beds. Science,

367(6474):156–156.

Soares, M., Teixeira, C., Bezerra, L., Rossi, S., Tavares, T.,

and Cavalcante, R. (2020). Brazil oil spill response:

Time for coordination. Science, 367(6474):155–155.

Vyas, G., Bhan, A., and Gupta, D. (2015). Detection of

oil spills using feature extraction and threshold based

segmentation techniques. In 2015 2nd International

Conference on Signal Processing and Integrated Net-

works (SPIN), pages 579–583. IEEE.

Zacharias, D. C., Gama, C. M., and Fornaro, A. (2021).

Mysterious oil spill on brazilian coast: Analysis and

estimates. Marine Pollution Bulletin, 165:112125.

Zhang, Z., Liu, Q., and Wang, Y. (2018). Road extraction

by deep residual u-net. IEEE Geoscience and Remote

Sensing Letters, 15(5):749–753.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

338