ISPRS Open Jou nal o Pho og amme y and Remo e Sensing 15 (2025) 100081
A ailable online 24 Decembe 2024
2667-3932/© 2024 The Au ho s. Published by Else ie B.V. on behal o In e na ional Socie y o Pho og amme y and Remo e Sensing. This is an open access a icle
unde he CC BY license (h p://c ea i ecommons.o g/licenses/by/4.0/).
Con en s lis s a ailable a ScienceDi ec
ISPRS Open Jou nal o Pho og amme y and Remo e Sensing
jou nal homepage: www.jou nals.else ie .com/isp s-open-jou nal-o -pho og amme y-and- emo e-sensing
T ans e lea ning and single-pola ized SAR image p ep ocessing o oil spill
de ec ion
Na aliia Kussul a,b,c, Ye henii Salii a,b,∗, Volodymy Kuzin a,b, Bohdan Yailymo b,
And ii Sheles o a,b
aNa ional Technical Uni e si y o Uk aine ‘‘Igo Siko sky Kyi Poly echnic Ins i u e’’, Depa men o Ma hema ical Modelling and Da a Analysis, Kyi , Uk aine
bSpace Resea ch Ins i u e NAS Uk aine and SSA Uk aine, Depa men o Space In o ma ion Technologies and Sys ems, Kyi , Uk aine
cUni e si y o Ma yland, Depa men o Geog aphical Sciences, College Pa k, Uni ed S a es
ARTICLE INFO
Keywo ds:
Oil spill de ec ion
Syn he ic ape u e ada (SAR)
Deep lea ning
Image p ep ocessing
T ans e lea ning
ABSTRACT
This s udy add esses he challenge o oil spill de ec ion using Syn he ic Ape u e Rada (SAR) sa elli e image y,
employing deep lea ning echniques o imp o e accu acy and e iciency. We in es iga ed he e ec i eness o
a ious neu al ne wo k a chi ec u es and encode s o his ask, ocusing on scena ios wi h limi ed aining
da a. The esea ch p oblem cen e ed on enhancing ea u e ex ac ion om single-channel SAR da a o imp o e
oil spill de ec ion pe o mance.
Ou me hodology in ol ed de eloping a no el p ep ocessing pipeline ha con e s single-channel SAR
da a in o a h ee-channel RGB ep esen a ion. The p ep ocessing echnique no malizes SAR in ensi y alues
and encodes ex ac ed ea u es in o RGB channels.
Th ough an expe imen , we ha e shown ha a combina ion o he LinkNe wi h an E icien Ne -B4 is
supe io o pai s o o he well-known a chi ec u es and encode s.
Quan i a i e e alua ion e ealed a signi ican imp o emen in F1-sco e o 0.064 compa ed o adi ional
dB-scale p ep ocessing me hods. Quali a i e assessmen on independen SAR scenes om he Medi e anean
Sea demons a ed be e de ec ion capabili ies, albei wi h inc eased sensi i i y o look-alike.
We conclude ha ou p oposed p ep ocessing echnique shows p omise o enhancing au oma ic oil spill
segmen a ion om SAR image y. The s udy con ibu es o ad ancing oil spill de ec ion me hods, wi h po en ial
implica ions o en i onmen al moni o ing and ma ine ecosys em p o ec ion.
1. In oduc ion
The Medi e anean Sea, a biodi e si y ho spo hos ing 11% o he
wo ld’s ma ine species in less han 1% o he global ocean a ea, aces
inc easing h ea s om pollu ion. Sa elli e image y has e ealed 757
oil slicks spanning 1.9 million hec a es in he Medi e anean be ween
2020 and 2024 (SkyT u h,2024), p ima ily om illegal discha ges
by essels. These inciden s a e pa icula ly ala ming as hey impac
key ma ine p o ec ed a eas, including he Medi e anean Ce acean
Mig a ion Co ido and he Pelagos Sanc ua y. Fu he mo e, se e al
a eas, such as hose nea Cyp us, emain unde -assessed despi e hei
ulne abili y o pollu ion.
Gi en he scale and numbe o hese inciden s, adi ional oil spill
de ec ion me hods, such as isual inspec ions and on-si e sampling,
a e limi ed in scalabili y and wea he dependency, Syn he ic Ape u e
Rada (SAR) o e s a p omising al e na i e wi h i s la ge a ea co e age,
∗Co esponding au ho a : Space Resea ch Ins i u e NAS Uk aine and SSA Uk aine, Depa men o Space In o ma ion Technologies and Sys ems, Kyi , Uk aine.
E-mail add ess: [email p o ec ed] (Y. Salii).
all-wea he , day-and-nigh imaging capabili ies (Dong e al.,2023;
He nández-Hamón e al.,2023;V înceanu e al.,2023). Howe e ,
accu a e and imely de ec ion o oil spills om SAR image y emains
challenging, pa icula ly due o look-alike phenomena ha can be
mis aken o oil slicks and he gene al di icul y in dis inguishing oil
om simila wa e da k spo s (K es eni is e al.,2019;Chen e al.,
2024;Wu e al.,2024b). Fu he mo e, a signi ican challenge in he
deploymen o da a-d i en me hods o SAR analysis is he lack o
su icien labeled aining da a o many geog aphic egions, including
ou pilo e i o y nea Cyp us.
Ad ances in deep lea ning, pa icula ly con olu ional neu al ne -
wo ks (CNNs), ha e become in eg al o imp o ing oil spill de ec ion
om SAR image y. These echniques ha e e olu ionized he ield by
add essing he complexi y o SAR da a and he sub le ies o oil spill sig-
na u es ha adi ional image p ocessing me hods o en s uggle wi h.
h ps://doi.o g/10.1016/j.opho o.2024.100081
Recei ed 1 July 2024; Recei ed in e ised o m 15 Decembe 2024; Accep ed 16 Decembe 2024
ISPRS Open Jou nal o Pho og amme y and Remo e Sensing 15 (2025) 100081
2
N. Kussul e al.
A a ie y o CNN a chi ec u es, including U-Ne (de Mou a e al.,2022;
Mahmoud e al.,2022), LinkNe (Ding e al.,2020;Yan e al.,2023),
DeepLabV3+(Kong e al.,2021;Li and Kong,2023), U-Ne ++ (Chen
e al.,2023;Yu e al.,2023), MAne (Wu e al.,2024a;Huang e al.,
2023), FPN (Zhang e al.,2021;Sun e al.,2022), and PSPNe (Li
e al.,2020;E en e al.,2023), o e di e se app oaches o image
segmen a ion.
The selec ion o an encode backbone signi ican ly impac s model
pe o mance, complemen ing he choice o a chi ec u e. Op ions like
E icien Ne , MobileNe V3, and MobileOne o e a ious ade-o s be-
ween accu acy and compu a ional e iciency (Li e al.,2021;Koloso
e al.,2022;Vasu e al.,2023). This balance is pa icula ly c ucial in
oil spill de ec ion, whe e apid p ocessing o la ge olumes o sa elli e
da a is essen ial o imely esponses o en i onmen al eme gencies.
In ou s udy, we conduc an expe imen h ough a comp ehensi e
e alua ion o s a e-o - he-a a chi ec u es and encode s o de e mine
hei e ec i eness in he con ex o he applica ion ask.
While deep lea ning models ha e shown imp essi e esul s, hei
pe o mance hea ily depends on he quali y and ep esen a ion o
inpu da a. This dependency highligh s he c i ical ole o p ep ocess-
ing echniques o sa elli e images, which can signi ican ly enhance
model pe o mance by accen ua ing ele an ea u es and supp essing
noise (C is ea e al.,2020;Hong e al.,2020;Lin e al.,2021). Howe e ,
he op imal pipeline emains an open ques ion because o he di e si y
in SAR image cha ac e is ics ac oss di e en en i onmen al condi ions
and geog aphic egions.
Ou s udy add esses hese challenges wi hin he amewo k o he
HORIZON Eu ope iMERMAID p ojec , which aims o de elop inno-
a i e s a egies o p e en ing, moni o ing, and mi iga ing pollu ion
in he Medi e anean Sea. Speci ically, we p opose an app oach ha
combines ad anced p ep ocessing echniques wi h he u iliza ion o
he bes a chi ec u e-encode pai , de e mined h ough a comp ehen-
si e expe imen . By le e aging ans e lea ning, we mi iga e he da a
sca ci y issue by ine- uning models p e- ained on la ge da ase s o
adap hem o he cha ac e is ics o ou s udy egion.
Ou con ibu ions and hei impac s a e as ollows:
•A compa a i e e alua ion o a ious CNN a chi ec u es and en-
code backbones o oil spill de ec ion, iden i ying e ec i e com-
bina ions o his ask.
•A no el p ep ocessing echnique ha no malizes SAR in ensi y
alues and encodes ex ac ed ea u es in o RGB channels, leading
o enhancemen o ele an ea u es o oil spills
•An expe imen al e alua ion using a da ase spanning mul iple ge-
og aphic egions, demons a ing he model’s abili y o gene alize
o unseen da a om he Medi e anean Sea
These con ibu ions no only ad ance he echnical ield o oil spill
de ec ion bu also di ec ly suppo he b oade goals o he iMERMAID
p ojec in p o ec ing he Medi e anean Sea om chemical pollu ion.
2. Da a and me hods
In his sec ion, we de ail he comp ehensi e p ocess in ol ed in de-
ec ing oil spills om sa elli e image y using deep lea ning models. The
en i e wo k low, om da a acquisi ion o he aining o segmen a ion
models, is illus a ed in Fig. 1.
2.1. CleanSeaNe
CleanSeaNe (Ca pen e ,2015), a Eu opean sa elli e-based oil spill
de ec ion se ice, ep esen s a co ne s one in ma i ime en i onmen-
al moni o ing. Ope a ed by he Eu opean Ma i ime Sa e y Agency
(EMSA), his sys em le e ages ad anced SAR echnology o p o ide
nea - eal- ime de ec ion o po en ial oil spills ac oss Eu opean wa e s.
The se ice’s s eng h lies in i s abili y o cap u e high- esolu ion
image y ega dless o cloud co e o dayligh condi ions, making i an
in aluable ool o con inuous ma i ime su eillance. CleanSeaNe ’s de-
ec ion capabili ies ex end beyond me e iden i ica ion, o e ing c ucial
de ails such as spill loca ion, a eal ex en , and con idence assessmen s
o de ec ed anomalies.
Despi e i s sophis ica ed capabili ies, he publicly accessible da a
om CleanSeaNe o he Medi e anean egion p esen s limi a ions o
esea ch pu poses. The a ailable in o ma ion is es ic ed o annual
poin da a, which lacks he empo al linkage needed o i machine
o deep lea ning models. This cons ain poses a signi ican challenge
o esea che s aiming o de elop and ain high-p ecision oil spill
de ec ion models speci ic o he Medi e anean Sea.
2.2. Ma ine Pollu ion Su eillance P og am
The Ma ine Pollu ion Su eillance P og am (Shazi ,2022), ope a ed
by he NOAA/NESDIS Sa elli e Analysis B anch (SAB), s ands as a
pi o al esou ce in he moni o ing and de ec ion o oil spills in U.S.
wa e s. Ope a ional since la e 2010, his p og am has e ol ed in o a
obus , ound- he-clock su eillance sys em, u ilizing a combina ion o
SAR and mul i-spec al sa elli e image y.
The p og am’s p ima y mission ex ends beyond me e de ec ion,
aiming o p o ide comp ehensi e analyses o oil spill inciden s. These
analyses include c ucial in o ma ion such as he loca ion o oil slicks,
hei spa ial ex en , and, when possible, assessmen s o ela i e oil
hickness. This de ailed app oach suppo s apid esponse e o s coo -
dina ed by he Na ional Ocean Se ice Eme gency Response Di ision.
A key s eng h o he Ma ine Pollu ion Su eillance P og am lies in
i s ex ensi e his o ical da ase . The p og am o e s openly accessible
ec o da a (NOAA,2022) o oil spills om 2011 o 2024, co e ing
a as a ay o wa e bodies including he Gul o Mexico, he Paci ic
Ocean, he A lan ic Ocean, and he G ea Lakes. This ich empo al and
spa ial co e age p o ides an in aluable esou ce o de eloping and
e ining oil spill de ec ion algo i hms.
The da ase ’s comp ehensi e na u e, encompassing a ious ma ine
en i onmen s and condi ions, makes i pa icula ly sui able o aining
machine lea ning models. I o e s esea che s he oppo uni y o de-
elop obus algo i hms capable o iden i ying oil spills ac oss di e se
scena ios and en i onmen al condi ions.
Howe e , i is impo an o no e ha while his da ase is ex ensi e
o U.S. wa e s, i does no include da a om he Medi e anean Sea.
This geog aphical limi a ion p esen s bo h a challenge and an oppo -
uni y o esea che s. The challenge lies in he di ec applicabili y o
models ained on his da a o Medi e anean con ex s. Con e sely, i
opens a enues o explo ing ans e lea ning echniques, whe e models
ained on he ich U.S. da ase can be adap ed and ine- uned o
applica ion in Medi e anean wa e s, po en ially b idging he da a gap
in his egion.
2.3. Sen inel-1 SAR da a
The backbone o ou s udy is Sen inel-1 SAR da a, chosen o i s
consis en quali y and global co e age.
Fo ou analysis, we speci ically u ilized he G ound Range De ec ed
(GRD) p oduc s om Sen inel-1, which o e a spa ial esolu ion o
10 m. These p oduc s a e p e-p ocessed o emo e he mal noise and
p o ide e ain-co ec ed da a in g ound ange geome y. We ocused
on he VV (Ve ical–Ve ical) pola iza ion channel, as i has been
shown o be pa icula ly e ec i e in dis inguishing oil spills om he
su ounding wa e su ace.
The Sen inel-1 da a acquisi ion p ocess was wo old, ailo ed o
ou speci ic s udy a eas. Fo he images co esponding o he Ma ine
Pollu ion Su eillance Repo s (MPSR), we u ilized he NASA Ea h
Da a pla o m wi h he Alaska Sa elli e Facili y (ASF). This app oach
allowed us o p ecisely a ge and download he ele an SAR images
o he epo ed oil spill e en s in U.S. wa e s.
ISPRS Open Jou nal o Pho og amme y and Remo e Sensing 15 (2025) 100081
3
N. Kussul e al.
Fig. 1. Pipeline illus a ing he s eps om da a acquisi ion o segmen a ion model aining and e alua ion.
In con as , we used he Google Ea h Engine pla o m o he
Medi e anean s udy a ea. This cloud-based solu ion allowed us o
e icien ly access and p ocess Sen inel-1 image y co e ing ou a ea
o in e es be ween Cyp us and Ana olia. Using he cloud pla o m
was pa icula ly ad an ageous o handling he la ge amoun o da a
equi ed o comp ehensi e co e age o he Medi e anean egion. We
applied addi ional p e-p ocessing s eps o he Sen inel-1 image y o
ensu e op imal da a quali y. These included speckle il e ing o educe
inhe en SAR noise and adiome ic calib a ion o ensu e consis ency
be ween acquisi ion da es and senso modes.
2.4. Da ase
Ou da ase was me iculously compiled om Ma ine Pollu ion Su eil-
lance P og am epo s, ocusing speci ically on Sen inel-1 da a om
2018 o 2023. A e ca e ul cu a ion, which in ol ed excluding epo s
wi hou bina y segmen a ion, we assembled a collec ion o 300 dis inc
a eas o in e es .
Thus, he esul ing da ase comp ises 300 VV-pola ized SAR images,
each wi h dimensions o 640x640 pixels and a spa ial esolu ion o
10 m. To ensu e obus model e alua ion, we ese ed all 58 a eas
om he yea 2021 o es ing pu poses. The emaining 242 a eas we e
andomly di ided in o aining and alida ion subse s, ollowing an
80:20 a io espec i ely.
To complemen his p ima y da ase and acili a e quali a i e as-
sessmen wi hin he iMe maid p ojec amewo k, we cons uc ed an
addi ional one. This supplemen a y se consis s o 114 unlabeled SAR
images, each measu ing 6275 ×3489 pixels. These images co e an ex-
pansi e a ea o 176,000 hec a es in he Medi e anean Sea, speci ically
he egion be ween Cyp us and Ana olia. Acco ding o CleanSeaNe
eco ds, his a ea expe ienced 32 oil spill e en s du ing 2022, making i
a aluable esou ce o e alua ing model pe o mance in a eal-wo ld,
Medi e anean con ex 2.
This comp ehensi e da ase s uc u e allows o ho ough model
aining, alida ion, and es ing, while also p o iding a means o assess
he model’s ans e abili y o he Medi e anean egion, which is c ucial
o he b oade objec i es o he iMe maid p ojec .
2.5. P e ained models
Gi en he limi ed size o ou oil spill da ase , we adop ed a ans-
e lea ning app oach o le e age knowledge om models p e ained
on la ge-scale da ase s. This s a egy allows us o bene i om ea-
u es lea ned on di e se image da a, po en ially imp o ing ou model’s
abili y o de ec oil spills in SAR image y.
We u ilized models p e ained on he ImageNe da ase , a da abase
o o e 14 million images ac oss 20,000 ca ego ies. The ImageNe
p e aining p o ides ou models wi h a obus se o low-le el ea u es
and high-le el seman ic unde s anding, which can be ad an ageous
e en when ans e ing o a domain as specialized as SAR image y.
Fo ou expe imen s, we explo ed a ange o mode n a chi ec-
u es known o hei e ec i eness in image segmen a ion asks. These
Fig. 2. A pilo a ea in he Medi e anean o es ing an oil spills de ec ion model.
include U-Ne , LinkNe , DeepLabV3+, U-Ne ++, DeepLabV3, MAne ,
FPN, and PSPNe . Each o hese a chi ec u es o e s unique s eng hs in
handling mul i-scale ea u es and p ese ing spa ial in o ma ion, which
a e c ucial o accu a e oil spill delinea ion.
To complemen hese a chi ec u es, we in es iga ed se e al encode
backbones, speci ically e icien ne -b4, imm-mobilene 3_small_100
and mobileone_s4. These encode s we e chosen o hei balance o
accu acy, compu a ional e iciency and size, which is pa icula ly im-
po an o po en ial eal- ime applica ions o oil spill de ec ion.
Th ough his ans e lea ning app oach, we aim o le e age he
powe o la ge-scale p e aining while ine- uning he models o he
speci ic challenges o oil spill de ec ion in SAR image y. This s a -
egy allows us o po en ially achie e highe accu acy and as e con-
e gence han aining om sc a ch, e en wi h ou ela i ely small
domain-speci ic da ase .
2.6. Compu ing in as uc u e
The aining o ou deep lea ning models was pe o med on a dedi-
ca ed ins ance wi hin he C eoDIAS pla o m (Kuzin e al.,2022). C eo-
DIAS is a cloud-based in as uc u e ha p o ides access o Cope nicus
sa elli e da a and compu ing esou ces, enabling e icien p ocessing
and analysis o la ge-scale Ea h obse a ion da ase s.
Fo ou expe imen s, we u ilized a cus om i ual machine ins ance
speci ically con igu ed o deep lea ning asks. The ins ance, designa ed
as ‘‘ m.a6000.2’’, was equipped wi h he ollowing speci ica ions:
ISPRS Open Jou nal o Pho og amme y and Remo e Sensing 15 (2025) 100081
4
N. Kussul e al.
•Vi ual GPU: NVIDIA RTX A6000 wi h 12 GB VRAM
•RAM: 28 GB
• CPUs: 4
•S o age: 80 GB SSD
This high-pe o mance con igu a ion, pa icula ly he NVIDIA RTX
A6000 GPU, allowed us o e icien ly ain and e alua e ou deep lea n-
ing models on he men ioned Sen inel-1 da ase . The ample GPU mem-
o y was c ucial o handling he complex neu al ne wo k a chi ec u es
and ela i ely la ge ba ch sizes used in ou expe imen s.
2.7. Me ics
In e alua ing he pe o mance o ou oil spill de ec ion model, we
ocused on se e al key me ics ha a e pa icula ly well-sui ed o
assessing segmen a ion asks in imbalanced da ase s, which is cha ac-
e is ic o oil spill de ec ion in as ocean en i onmen s.
The p ima y me ics we employed we e he In e sec ion o e Union
(IoU), also known as he Jacca d index, and he F1 sco e, which is
equi alen o he Dice coe icien . These me ics we e chosen o hei
abili y o p o ide a comp ehensi e assessmen o ou model’s accu acy
in iden i ying oil spills agains he p edominan backg ound o clean
wa e .
The IoU me ic measu es he o e lap be ween he p edic ed oil spill
segmen a ion and he g ound u h, calcula ed as he a io o he in e -
sec ion o he union o he p edic ed and ac ual oil spill a eas. I anges
om 0 o 1, wi h 1 indica ing pe ec o e lap. The IoU is pa icula ly
aluable in ou con ex as i penalizes bo h alse posi i es and alse
nega i es, p o iding a s ic measu e o segmen a ion accu acy.
The F1 sco e, calcula ed as he ha monic mean o p ecision and
ecall, o e s a balanced measu e o he model’s pe o mance. In he
con ex o oil spill de ec ion, whe e he posi i e class (oil spills) is
signi ican ly smalle han he nega i e class (clean wa e ), he F1 sco e
helps in assessing how well he model iden i ies oil spills wi hou being
o e ly in luenced by he la ge numbe o ue nega i es.
Addi ionally, we u ilized he Dice loss unc ion du ing model ain-
ing. The Dice loss, being he complemen o he Dice coe icien ,
se ed as an e ec i e op imiza ion a ge . I helped in add essing he
class imbalance issue inhe en in ou da ase by ocusing he model’s
a en ion on co ec ly iden i ying he ela i ely small oil spill a eas.
2.8. Da a p epa a ion
The use o p e ained models, howe e , comes wi h a no able
challenge: ImageNe - ained models a e designed o p ocess colo im-
ages composed o h ee channels ( ed, g een, and blue) whe eas ou
applica ion in ol es using only single-channel VV-pola ized images.
To add ess his disc epancy, we adap ed ou app oach by u ilizing
con e sa ion om VV o RGB, which can be done in a ious ways, bu
in he scope o ou esea ch, we ocus on 2 o hem.
2.8.1. O iginal app oach
The es ablished echnique in ol es ans o ming calib a ed sa elli e
image y (𝑉 𝑉) o a loga i hmic decibel (𝑑 𝐵) scale:
𝑑 𝐵= 20 ⋅log10 𝑉 𝑉
This con e sion enhances he isibili y o ela i e di e ences in
backsca e in ensi y (Blondeau-Pa issie e al.,2023). I is pa icu-
la ly e ec i e o highligh ing oil spill pixels in sa elli e images while
ensu ing hey emain da ke han su ounding ocean and land a eas.
Oil spills in SAR image y ypically ha e a limi ed ange o in ensi y
alues, making hem challenging o dis inguish om o he su aces. The
dB scale add esses his by expanding he lowe in ensi y ange asso-
cia ed wi h oil slicks while comp essing highe in ensi y alues om
sea and e ain. This ans o ma ion ampli ies sub le con as s be ween
oil slicks and hei su oundings, main aining he cha ac e is ic da ke
appea ance o oil spills in SAR image y due o hei smoo he su ace
ex u e.
To ep esen he 𝑑 𝐵 alues as colo componen s, we apply min–
max no maliza ion o scale hem o he ange [0,255], sui able o RGB
channels.
2.8.2. P oposed app oach
Ou no el p ep ocessing app oach aims o enhance he di e en ia-
ion o oil spills by assigning dis inc in o ma ion o each colo channel,
he eby emphasizing a ious aspec s o oil slicks while minimizing he
in luence o con ounding ac o s.
Ini ially, we con e he VV-pola ized SAR da a o he decibel scale,
simila o he o iginal app oach and apply min–max scaling:
𝑟= log10 𝑉 𝑉
Howe e , in cases whe e no oil slicks a e p esen in he image,
min–max scaling can be expec ed o c ea e a signi ican numbe o
look-alikes, as he sp ead o wa e b igh ness alues will be ela i ely
small.
To minimize look-alike ecogni ion caused by ed channel no -
maliza ion, i was decided o con e he VV channel o a no mal
dis ibu ion. This con e sion wo ks on he hypo hesis ha he as
majo i y o pixels in he image belong o he wa e class, which is ue
o bo h aining da a and eal-wo ld images. Since a e con e sion he
alues a e no limi ed, bu mos a e close o 0, we used no maliza ion
based on he a c angen unc ion.
𝑔=1
𝜋a c an (𝑉 𝑉−𝑚𝑒𝑎𝑛(𝑉 𝑉)
𝑠𝑡𝑑(𝑉 𝑉))+ 0.5
One o he ea u es o he aining da a is ha some images ha e
signi ican in a-image a iabili y in wa e su ace b igh ness. In hese
cases, he spa ial b igh ness o he wa e o ms a nonlinea su ace wi h
a low equency o a ia ion om he mean. In many cases, such images
ha e la ge da k egions ha could be mis aken o oil spills. To mi iga e
his issue, we p opose a local b igh ness equaliza ion echnique ha
ies o app oxima e he spa ial nonlinea i y o wa e b igh ness:
𝐸𝑃=𝑝𝑜𝑜𝑙 𝑖𝑛𝑔𝑎𝑣𝑔
32 (𝑉 𝑉)𝐷𝑃=𝑝𝑜𝑜𝑙 𝑖𝑛𝑔𝑠𝑡𝑑
32 (𝑉 𝑉)
𝐸𝑃and 𝐷𝑃a e compu ed using a e age pooling and pooled s an-
da d de ia ion, espec i ely, o VV alues wi h a ke nel size and a s ide
o 32. This esul s in a i s app oxima ion o he desi ed su ace. A e
his ope a ion, he dimensions o he inpu image a e educed by a
ac o o 32 in bo h axes.
𝐸𝐵=𝑏𝑙 𝑢𝑟𝑔 𝑎𝑢𝑠𝑠𝑖𝑎𝑛
7(𝐸𝑃)𝐷𝐵=𝑏𝑙 𝑢𝑟𝑔 𝑎𝑢𝑠𝑠𝑖𝑎𝑛
7(𝐷𝑃)
The pooled alues a e hen smoo hed using Gaussian blu wi h a
ke nel size o 7. Smoo hing educes he impac o ou lie s caused by
he possible p esence o oil spills in pooled pixels.
𝐸𝑅=𝑟𝑒𝑠𝑖𝑧𝑒𝑙 𝑖𝑛𝑒𝑎𝑟
𝑉 𝑉(𝐸𝐵)𝐷𝑅=𝑟𝑒𝑠𝑖𝑧𝑒𝑙 𝑖𝑛𝑒𝑎𝑟
𝑉 𝑉(𝐷𝐵)
Then we achie e he inal app oxima ion by esizing he smoo hed
alues o he o iginal VV image dimensions using linea in e pola ion.
𝑏∗=(𝑉 𝑉−𝐸𝑅)∕𝐷𝑅
Finally, we ob ain he blue channel by no malizing VV alues
based on app oxima ed local wa e b igh ness (𝐸𝑅) and local s anda d
de ia ion (𝐷𝑅). Fo scaling we use a c angen unc ion:
𝑏=1
𝜋a c an (𝑏∗)+ 0.5
ISPRS Open Jou nal o Pho og amme y and Remo e Sensing 15 (2025) 100081
5
N. Kussul e al.
Table 1
Quan i a i e esul s o ained models so ed by weigh ed 1 sco e.
A chi ec u e Encode Ba ch T aining ime T ain Valida ion Tes Weigh ed
size (minu es) F1 F1 F1 F1
LinkNe e icien ne -b4 4 25 0.774 0.703 0.765 0.746
DeepLabV3Plus imm-mobilene 3_small_100 16 12 0.765 0.676 0.777 0.741
MAne imm-mobilene 3_small_100 16 5 0.780 0.642 0.787 0.737
FPN e icien ne -b4 4 33 0.777 0.658 0.768 0.733
LinkNe imm-mobilene 3_small_100 16 6 0.693 0.675 0.785 0.728
Une PlusPlus e icien ne -b4 2 35 0.758 0.670 0.753 0.726
PSPNe mobileone_s4 4 18 0.719 0.672 0.750 0.717
Une PlusPlus imm-mobilene 3_small_100 12 8 0.731 0.649 0.741 0.708
Une e icien ne -b4 4 26 0.705 0.679 0.726 0.706
DeepLabV3 imm-mobilene 3_small_100 12 19 0.768 0.557 0.778 0.702
FPN imm-mobilene 3_small_100 16 8 0.731 0.614 0.740 0.696
LinkNe mobileone_s4 4 33 0.667 0.710 0.690 0.691
Une mobileone_s4 4 32 0.703 0.696 0.669 0.686
PSPNe e icien ne -b4 4 13 0.723 0.630 0.692 0.678
Une imm-mobilene 3_small_100 16 6 0.746 0.637 0.655 0.669
MAne e icien ne -b4 4 24 0.690 0.633 0.681 0.667
FPN mobileone_s4 4 14 0.650 0.653 0.629 0.641
PSPNe imm-mobilene 3_small_100 16 3 0.588 0.584 0.642 0.611
Une PlusPlus mobileone_s4 2 45 0.375 0.693 0.557 0.562
DeepLabV3Plus e icien ne -b4 2 61 0.572 0.620 0.498 0.555
MAne mobileone_s4 4 23 0.496 0.534 0.541 0.529
P io o assigning he ob ained , g, b alues o he R, G, B colo
channels, we apply min–max scaling o o i he ange [0, 255], while
g and b a e simply mul iplied by 255.
This p ep ocessing pipeline aims o enhance he isibili y o oil
spills while no malizing o bo h global and local b igh ness a ia ions,
imp o ing he model’s abili y o dis inguish be ween oil spills and o he
da k su aces in sa elli e images.
2.9. Expe imen I: A chi ec u e and encode selec ion
Fi s o all, we wan o know which combina ion o a chi ec u e
and encode is bes sui ed o building a model o sol e he oil slick
de ec ion p oblem.
We ained and e alua ed 24 dis inc models, comp ising combi-
na ions o 8 a chi ec u es (DeepLabV3, DeepLabV3+, FPN, LinkNe ,
MAne , PSPNe , Une , Une ++) and 3 encode s (E icien Ne -B4,
MobileOne-S4, Timm-MobileNe V3-Small-100). This allows o a ho -
ough assessmen o a ious s a e-o - he-a segmen a ion a chi ec u es
and e icien encode backbones.
All models will be ained wi h a lea ning a e o 10−3. Ba ch size
will a y due o di e en equi emen s o RAM o each model, bu will
no exceed 16. T aining ime is limi ed o 300 epochs, howe e , i he
model is no imp o ing on alida ion o 10 epochs hen aining will
be s opped ea lie .
All models in his expe imen u ilized he o iginal VV o RGB
con e sion app oach as desc ibed in Sec ion 2.8.1
2.10. Expe imen II: Model e inemen and Medi e anean Sea e alua ion
Building upon he esul s o Expe imen I, ou second expe imen
aims o maximize model pe o mance and assess gene aliza ion ca-
pabili ies o he Medi e anean Sea con ex . We ocused on he mos
e ec i e a chi ec u e-encode combina ion iden i ied in he i s expe -
imen . Two a ian s o his model we e ained: one using he o iginal
VV o RGB con e sion app oach, and ano he using he p oposed
p ep ocessing app oach desc ibed in Sec ion 2.8.2
Fo his ine- uning phase, we educed he lea ning a e o 10−4
while keeping he o he aining pa ame e s he same as in Expe imen
I. This allows a mo e p ecise adjus men o he model weigh s.
3. Resul s
3.1. Expe imen I
Quan i a i e pe o mance o ained models alongside wi h aining
ime and ba ch size can be seen in Table 1.
Due o he high sp ead o sco e be ween alida ion and es , we
decided o add weigh ed a e age sco e. This sco e was cons uc ed
by aking ain, alida ion and es sco es in a a io o 2:3:4. Such
a io was chosen o ewa d high es sco es and punish low alida ion
while aking in o accoun he in luence o co esponding subse s o he
aining p ocess.
F om Table 1we can see ha only 21 models was ained while in
Sec ion 2.9 was s a ed ha 24 models would be ained. This misma ch
was caused due o he ac ha DeepLabV3 wi h e icien ne -b4 and
mobileone_s4 encode was unable o i in o a ailable VRAM wi h a
ba ch size o 2. The aining o he las missing model (DeepLabV3Plus
wi h mobileone_s4 encode ) was canceled due o poo pe o mance and
he long aining ime o he model wi h e icien ne -b4 encode .
As shown in Table 1, he bes esul s we e achie ed by he LinkNe
and FPN models wi h he E icien Ne -B4 encode , as well as by
DeepLabV3Plus and MAne wi h he imm-mobilene 3_small_100 en-
code . Howe e , while hese models pe o med simila ly on he aining
and es se s, only he LinkNe model demons a ed good esul s on he
alida ion se .
DeepLabV3Plus and MAne exhibi ed in e es ing beha io , pe o m-
ing be e wi h smalle encode s. Fo ins ance, bo h models achie ed a
weigh ed sco e o app oxima ely 0.74 when using he imm-
mobilene 3_small_100 encode , which has 0.93 million pa ame e s.
In con as , hei pe o mance d opped o 0.67 when u ilizing he
E icien Ne -B4 encode , which has 17 million pa ame e s.
The o e all conclusion om his expe imen is ha he combina-
ion o LinkNe and E icien Ne -B4 shows he highes po en ial o
e ec i ely de ec ing oil spills.
3.2. Expe imen II
Since, acco ding o he esul s o he p e ious expe imen , he
bes pe o mance was shown by LinkNe in combina ion wi h he
e icien -b4 encode , we will use i o aining.
Fi s o all, i is wo h no ing ha he esul s ob ained o he
o iginal app oach (Table 2) show ha educing he lea ning a e o
10−4helped o signi ican ly imp o e he model’s pe o mance.
ISPRS Open Jou nal o Pho og amme y and Remo e Sensing 15 (2025) 100081
6
N. Kussul e al.
Fig. 3. Segmen a ion o Medi e anean es si e using he o iginal app oach o con e sion om VV o RGB. (a) - RGB image; (b) - segmen a ion esul .
Fig. 4. Segmen a ion o Medi e anean es si e using he p oposed app oach o con e sion om VV o RGB. (a) - RGB image; (b) - segmen a ion esul .
Table 2
Quan i a i e esul s o ained LinkNe models wi h e icien ne -b4 encode using
di e en VV o RGB app oaches.
App oach Ba ch T aining ime T ain Valida ion Tes Weigh ed
size (minu es) F1 F1 F1 F1
P oposed 4 55 0.875 0.748 0.831 0.813
O iginal 4 52 0.850 0.728 0.767 0.773
A compa ison o he esul s using he o iginal app oach om his
expe imen wi h hose om he p e ious one (Table 1) clea ly shows
an inc ease in bo h aining (by 0.076) and alida ion (by 0.025) sco es.
Howe e , he inc ease in es sco es is negligible. This sugges s ha he
model may ha e become mo e p one o o e i ing.
Rega ding he p oposed app oach, he da a in Table 2indica e
a signi ican imp o emen in model accu acy ac oss all cases. Fo
ins ance, he new app oach enhanced he pe o mance on es by 0.064,
and on aining and alida ion by 0.025 and 0.020, espec i ely. This
indica es a be e gene aliza ion abili y o he model wi h he new
app oach.
A e aining he models, we applied hem o da a om he Medi e -
anean Sea.
As can be seen om Fig. 3, he model using he o iginal app oach
con iden ly iden i ies he main pa o he oil slick, bu misses he
hinne ail o i . Addi ionally, i in e p e s a da k pa ch o he igh
o he slick as oil.
The model using he p oposed app oach, as seen in Fig. 4, mo e
con iden ly ecognized he main pa o he slick and also highligh ed
pa o i s ail. This model also highligh ed a da k pa ch o he igh o
he slick as oil, inc easing he likelihood ha i was, al hough he lack
o g ound u h p e en s us om e i ying his.
Howe e , he model is much mo e sensi i e, which imp o es i s
abili y o de ec oil spills, bu leads o an inc eased numbe o alse
posi i es, especially nea he no-da a egions.
In es iga ion in o he easons o his beha io shows ha he mos
likely cause is he blue channel o he p ep ocessed image. Due o he
use o con olu ions, i is sensi i e o how no-da a is p ocessed. In ou
case, he no-da a alue is eplaced by he maximum alid alue o VV,
leading o an o e es ima ion o backg ound wa e b igh ness nea no-
da a egions. A solu ion o his p oblem could be an adap i e selec ion
o he alue o ill no-da a ins ead o using maximum.
The appea ance o o he alse posi i es can be explained by a small
alue o local s anda d de ia ion leading o da k spo s appea ing in
egions wi h mos ly cons an wa e b igh ness. A possible solu ion o
his p oblem would be a limi on he minimum alue o he local
s anda d de ia ion.
Ou expe imen s wi h a ying hese ac o s ha e con i med ha his
can educe he numbe o alse posi i es. Howe e , inding he balance
be ween de ec ion capabili ies and alse posi i e a e by hese ac o s
equi es a sepa a e in es iga ion.
4. Discussion
This s udy has con i med he e ec i eness o using p e ained mod-
els on ImageNe o segmen a ion asks in o he domains. The con-
duc ed expe imen s demons a ed ha among he a chi ec u es and
encode s examined, he LinkNe model wi h an E icien Ne -B4 encode
p oduced he bes esul s o de ec ing oil spills based on VV-pola ized
SAR image y.
An unexpec ed inding was ha he DeepLabV3+and MAne models
pe o med be e wi h smalle encode s. Conside ing hei aining
speed and in e ence e iciency, hese models could po en ially be used
ISPRS Open Jou nal o Pho og amme y and Remo e Sensing 15 (2025) 100081
7
N. Kussul e al.
o p elimina y and/o ensemble segmen a ion me hods. Howe e , his
possibili y was no explo ed wi hin he scope o his wo k and wa an s
u he in es iga ion.
The expe imen in ol ing di e en app oaches o con e ing VV im-
ages o RGB showed ha enginee ing wi hin his s ep can signi ican ly
enhance model pe o mance. Ou p oposed me hod o his con e sion
signi ican ly imp o ed he model’s quan i a i e me ics.
When he model was ans e ed o a di e en geog aphical a ea,
he model show signi ican ly be e capabili y o de ec hinne pa s
o oil spills, howe e i also exhibi ed a highe le el o alse posi i es.
Ou expe imen s ou side he a icle’s scope show ha eplacing
no-da a alues wi h he mean alue o alid pixels ins ead o he
maximum leads o sol ing he p oblem o alse posi i es a he bo de
wi h no-da a. Limi ing he minimal alue o local s anda d de ia ion in
blue channel calcula ion helps dec ease alse posi i e a es. Howe e ,
inding op imal alues equi es sepa a e s udies.
Apa om he no-da a cases, he only known limi a ion o he p o-
posed app oach lies in coas al a eas wi h a signi ican amoun o land,
whe e, due o he peculia i ies o he calcula ions, he 𝐵and 𝐺channels
o he p ep ocessed image end o be da kened o b igh ened, hus
becoming less in o ma i e. This limi a ion can be pa ially o e come
by masking land wi h no-da a alues.
5. Conclusion
This s udy p esen s a signi ican s ep o wa d in he applica ion
o deep lea ning o oil spill de ec ion using Sen inel-1 VV-pola ized
SAR da a. Ou comp ehensi e e alua ion o a ious a chi ec u es and
encode s iden i ied he LinkNe a chi ec u e wi h an E icien Ne -B4
encode as a pa icula ly e ec i e combina ion o his ask, especially
in scena ios wi h limi ed aining da a.
A key con ibu ion o his wo k is ou no el p ep ocessing app oach
o con e ing SAR da a o RGB ep esen a ions. This me hod demon-
s a ed no able imp o emen s, inc easing he F1 sco e by 0.064 on ou
es da ase compa ed o he adi ional dB-scale ans o ma ion.
The ans e abili y o ou model om U.S. coas al wa e s o he
Medi e anean Sea showed p omising esul s. The model demons a ed
an imp o ed abili y o de ec oil spills in a new geog aphic con ex , bu
he model’s highe sensi i i y inc eases he numbe o alse posi i es.
The esul s ob ained wi hin he amewo k o VV o RGB con e sion
enginee ing can be used o sol e o he SAR-based segmen a ion asks.
And wi h some modi ica ions, i can be adap ed o o he ypes o
sa elli e da a.
This wo k aligns well wi h he objec i es o he HORIZON Eu ope
iMERMAID p ojec , o e ing po en ial enhancemen s o moni o ing ca-
pabili ies o he Medi e anean ecosys em. The indings open up new
a enues o esea ch in SAR image p ep ocessing and segmen a ion,
wi h po en ial applica ions ex ending beyond oil spill de ec ion o o he
a eas o emo e sensing.
CRediT au ho ship con ibu ion s a emen
Na aliia Kussul: W i ing – e iew & edi ing, Supe ision, P ojec
adminis a ion, Funding acquisi ion. Ye henii Salii: W i ing – o iginal
d a , Valida ion, So wa e, Me hodology, In es iga ion, Concep ual-
iza ion. Volodymy Kuzin: W i ing – e iew & edi ing, W i ing –
o iginal d a , Me hodology, Da a cu a ion. Bohdan Yailymo : W i ing
– o iginal d a , Da a cu a ion. And ii Sheles o : Supe ision.
Decla a ion o Gene a i e AI and AI-assis ed echnologies in he
w i ing p ocess
S a emen : Du ing he p epa a ion o his wo k, he au ho s used
Cha GPT se ice, in o de o assis wi h he d a ing and e inemen
o ce ain sec ions o he manusc ip . A e using his ool, he au ho s
ho oughly e iewed and edi ed he con en as needed and ake ull
esponsibili y o he con en o he publica ion. The use o his AI ool
was solely o language e inemen and o ganiza ion pu poses, and did
no con ibu e o he scien i ic con en , analysis, o conclusions o he
s udy.
Decla a ion o compe ing in e es
The au ho s decla e ha hey ha e no known compe ing inan-
cial in e es s o pe sonal ela ionships ha could ha e appea ed o
in luence he wo k epo ed in his pape .
Acknowledgmen s
The s udy is suppo ed by he Eu opean Commission h ough he
HORIZON Eu ope p ojec iMERMAID ‘‘Inno a i e solu ions o Medi e -
anean Ecosys em Remedia ion ia Moni o ing and Decon amina ion
om Chemical Pollu ion’’ (101112824). (h ps://ime maid.eu/).
Da a a ailabili y
Da a a e a ailable on eques , please ge in ouch wi h e-mail:
[email p o ec ed].
Re e ences
Blondeau-Pa issie , D., Sch oede , T., Su esh, G., Li, Z., Diakogiannis, F.I., I ing, P.,
Wi e, C., S e en, A.D., 2023. De ec ion o ma ine oil-like ea u es in sen inel-1 SAR
images by supplemen a y use o deep lea ning and empi ical me hods: Pe o mance
assessmen o he G ea Ba ie Ree ma ine pa k. Ma ine Poll. Bull. 188, 114598.
h p://dx.doi.o g/10.1016/j.ma polbul.2023.114598.
Ca pen e , A., 2015. Eu opean ma i ime sa e y agency CleanSeaNe ac i i ies in he
no h sea. In: Oil Pollu ion in he No h Sea. Sp inge In e na ional Publishing, pp.
33–47. h p://dx.doi.o g/10.1007/698_2015_429.
Chen, Y.T., Chang, L., Wang, J.H., 2024. Full-scale agg ega ed MobileUNe : An
imp o ed U-ne a chi ec u e o SAR oil spill de ec ion. Senso s 24 (12), 3724.
h p://dx.doi.o g/10.3390/s24123724.
Chen, S., Lei, F., Zang, Z., Zhang, M., 2023. Fo es mapping using a VGG16-UNe ++&;
s acking model based on google ea h engine in he u ban a ea. IEEE Geosci.
Remo e Sens. Le . 20, 1–5. h p://dx.doi.o g/10.1109/lg s.2023.3274128.
C is ea, A., an Hou e, J., Doulge is, A.P., 2020. In eg a ing incidence angle depen-
dencies in o he clus e ing-based segmen a ion o SAR images. IEEE J. Sel. Top.
Appl. Ea h Obs. Remo e Sens. 13, 2925–2939. h p://dx.doi.o g/10.1109/js a s.
2020.2993067.
de Mou a, N.V.A., de Ca alho, O.L.F., Gomes, R.A.T., Guima ães, R.F., de Ca -
alho Júnio , O.A., 2022. Deep-wa e oil-spill moni o ing and ecu ence analysis
in he B azilian e i o y using sen inel-1 ime se ies and deep lea ning. In . J. Appl.
Ea h Obs. Geoin . 107, 102695. h p://dx.doi.o g/10.1016/j.jag.2022.102695.
Ding, Y., Wu, M., Xu, Y., Duan, S., 2020. P-linkne : Linkne wi h spa ial py amid pooling
o high- esolu ion sa elli e image y. In . A ch. Pho og amm. Remo e Sens. Spa ial
In . Sci. XLIII-B3-2020, 35–40. h p://dx.doi.o g/10.5194/isp s-a chi es-xliii-b3-
2020-35-2020.
Dong, X., Li, J., Li, B., Jin, Y., Miao, S., 2023. Ma ine oil spill de ec ion om
low-quali y SAR emo e sensing images. J. Ma . Sci. Eng. 11 (8), 1552. h p:
//dx.doi.o g/10.3390/jmse11081552.
E en, H., Bos anci, E., Acici, K., Guzel, M.S., Asu oglu, T., Aydin, A., 2023. Seman ic
segmen a ion wi h high- esolu ion sen inel-1 SAR da a. Appl. Sci. 13 (10), 6025.
h p://dx.doi.o g/10.3390/app13106025.
He nández-Hamón, H., Ramí ez, P.Z., Za aza, M., Micalle , A., 2023. Google ea h
engine app using sen inel 1 SAR and deep lea ning o ocean seep me hane
de ec ion and moni o ing. Remo e Sens. Appl. Soc. En i on. 32, 101036. h p:
//dx.doi.o g/10.1016/j. sase.2023.101036.
Hong, S.J., Baek, W.K., Jung, H.S., 2020. Ship de ec ion om X-Band SAR images using
M2de deep lea ning model. Appl. Sci. 10 (21), 7751. h p://dx.doi.o g/10.3390/
app10217751.
Huang, L., Tian, Q., Tang, B.-H., Le, W., Wang, M., Ma, X., 2023. Siam-EMNe :
A siamese E icien Ne –mane ne wo k o building change de ec ion in e y
high esolu ion images. Remo e Sens. 15 (16), 3972. h p://dx.doi.o g/10.3390/
s15163972.
Koloso , D., Kele ou as, V., Kou essis, P., Mpo as, I., 2022. Ana omy o deep lea ning
image classi ica ion and objec de ec ion on comme cial edge de ices: A case s udy
on ace mask de ec ion. IEEE Access 10, 109167–109186. h p://dx.doi.o g/10.
1109/access.2022.3214214.
Kong, Y., Liu, Y., Yan, B., Leung, H., Peng, X., 2021. A no el Deeplab 3+ ne wo k
o SAR image y seman ic segmen a ion based on he po en ial ene gy loss
unc ion o gibbs dis ibu ion. Remo e Sens. 13 (3), 454. h p://dx.doi.o g/10.3390/
s13030454.
K es eni is, M., O anidis, G., Ioannidis, K., A ge inakis, K., V ochidis, S., Kompa -
sia is, I., 2019. Oil spill iden i ica ion om sa elli e images using deep neu al
ne wo ks. Remo e Sens. 11 (15), 1762. h p://dx.doi.o g/10.3390/ s11151762.
ISPRS Open Jou nal o Pho og amme y and Remo e Sensing 15 (2025) 100081
8
N. Kussul e al.
Kuzin, V., Musial, J., Sheles o , A., 2022. EO4UA ini ia i e: Scien i ic Eu opean suppo
o uk ainian scien i ic communi y. In: 2022 12 h In e na ional Con e ence on
Dependable Sys ems, Se ices and Technologies. DESSERT, IEEE, h p://dx.doi.o g/
10.1109/desse 58054.2022.10018706.
Li, Q., Kong, Y., 2023. An imp o ed SAR image seman ic segmen a ion Deeplab 3+
ne wo k based on he ea u e pos -p ocessing module. Remo e Sens. 15 (8), 2153.
h p://dx.doi.o g/10.3390/ s15082153.
Li, S., Tan, M., Pang, R., Li, A., Cheng, L., Le, Q.V., Jouppi, N.P., 2021. Sea ching o
as model amilies on da acen e accele a o s. In: P oceedings o he IEEE/CVF
Con e ence on Compu e Vision and Pa e n Recogni ion. CVPR, pp. 8085–8095.
Li, J., Xiu, J., Yang, Z., Liu, C., 2020. Dual pa h a en ion ne o emo e sensing
seman ic image segmen a ion. ISPRS In . J. Geo-In . 9 (10), 571. h p://dx.doi.
o g/10.3390/ijgi9100571.
Lin, S.Y., Lin, C.W., an Gassel , S., 2021. P ocessing amewo k o landslide de ec ion
based on syn he ic ape u e ada (SAR) in ensi y-image analysis. Remo e Sens. 13
(4), 644. h p://dx.doi.o g/10.3390/ s13040644.
Mahmoud, A.S., Mohamed, S.A., El-Kho iby, R.A., AbdelSalam, H.M., El-Khoda y, I.A.,
2022. Oil spill iden i ica ion based on dual a en ion UNe model using syn he ic
ape u e ada images. J. Indian Soc. Remo e Sens. 51 (1), 121–133. h p://dx.doi.
o g/10.1007/s12524-022-01624-6.
NOAA, 2022. Ma ine pollu ion su eillance epo s. URL: h ps://ospo.noaa.go /
P oduc s/ocean/ma inepollu ion/2022_a chi e.h ml.
Shazi , M., 2022. Oil spill clean-up me hods; s a e o he a .
SkyT u h, 2024. Medi e anean Sea Ch onic Oil Pollu ion Analysis. Technical Repo ,
SkyT u h, URL: . Analysis o oil slicks om July 2020 o Janua y 2024.
Sun, Z., Meng, C., Cheng, J., Zhang, Z., Chang, S., 2022. A mul i-scale ea u e py amid
ne wo k o de ec ion and ins ance segmen a ion o ma ine ships in SAR images.
Remo e Sens. 14 (24), 6312. h p://dx.doi.o g/10.3390/ s14246312.
Vasu, P.K.A., Gab iel, J., Zhu, J., Tuzel, O., Ranjan, A., 2023. MobileOne: An imp o ed
one millisecond mobile backbone. In: P oceedings o he IEEE/CVF Con e ence on
Compu e Vision and Pa e n Recogni ion. CVPR, pp. 7907–7917.
V înceanu, C.A., G ebby, S., Ma sh, S., 2023. The pe o mance o speckle il e s on
cope nicus sen inel-1 SAR images con aining na u al oil slicks. Q. J. Eng. Geol.
Hyd ogeol. 56 (3), h p://dx.doi.o g/10.1144/qjegh2022-046.
Wu, K., Cai, F., Wang, H., 2024a. CTMANe : A CNN- ans o me hyb id seman ic
segmen a ion ne wo k o ine-g ained ai po ex ac ion in complex SAR scenes.
IEEE J. Sel. Top. Appl. Ea h Obs. Remo e Sens. 17, 4689–4704. h p://dx.doi.o g/
10.1109/js a s.2024.3361927.
Wu, W., Wong, M.S., Yu, X., Shi, G., Kwok, C.Y.T., Zou, K., 2024b. Composi ional oil
spill de ec ion based on objec de ec o and adap ed segmen any hing model om
SAR images. IEEE Geosci. Remo e Sens. Le . 21, 1–5. h p://dx.doi.o g/10.1109/
lg s.2024.3382970.
Yan, Q., Chen, Y., Jin, S., Liu, S., Jia, Y., Zhen, Y., Chen, T., Huang, W., 2023. Inland
wa e mapping based on GA-LinkNe om CyGNSS da a. IEEE Geosci. Remo e Sens.
Le . 20, 1–5. h p://dx.doi.o g/10.1109/lg s.2022.3227596.
Yu, J., He, X., Yang, P., Mo agh, M., Xu, J., Xiong, J., 2023. Coas al aquacul u e
ex ac ion using GF-3 ully pola ime ic SAR image y: A amewo k in eg a ing
UNe ++ wi h ma ke -con olled wa e shed segmen a ion. Remo e Sens. 15 (9),
2246. h p://dx.doi.o g/10.3390/ s15092246.
Zhang, T., Zhang, X., Ke, X., 2021. Quad-FPN: A no el quad ea u e py amid ne wo k
o SAR ship de ec ion. Remo e Sens. 13 (14), 2771. h p://dx.doi.o g/10.3390/
s13142771.
