Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
AI-D i en Quali y Con ol and
Unce ain y Quan ifica ion in
Essen ial Ocean Va iables
ObsSea4Clim T aining Wo kshop/21-11-2025
Thania Papapos olou (HCMR)
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
In oduc ion
Pa 1: How can A ificial In elligence (AI) be applied o he Quali y
Con ol (QC) o Essen ial Ocean Va iables (EOVs) and Essen ial
Clima e Va iables (ECVs)?
Pa 2: How/Can we use AI-d i en QC o add ess EOV unce ain ies?
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Pa 1
How can A ificial In elligence (AI) be applied o he Quali y
Con ol (QC) o Essen ial Ocean Va iables (EOVs) and
Essen ial Clima e Va iables (ECVs)?
→ Why QC ma e s o all ocean obse a ions and wha is
he added challenge wi hin he EOV amewo k?
→ Wha a e he po en ial ad an ages o applying AI-d i en
s “ adi ional” QC me hods?
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
EOV/ECV amewo k
1 GOOS EOV defini ion
2 GCOS ECV defini ion
“[...] EOVs a e defined as he minimum se
o ocean a iables ha a e needed o
assess ocean s a e and a iabili y o
impo an global ocean phenomena [...],
and o p o ide essen ial da a o
applica ions ha suppo socie al
benefi .” 1
“An [...] ECV is a physical, chemical o
biological a iable o a g oup o linked
a iables ha c i ically con ibu es o he
cha ac e iza ion o Ea h’s clima e.” 2
Based on he F amewo k o Ocean Obse ing (FOO, Linds om e al.,
2012), EOV cha ac e is ics a e:
●sus ainable ocean measu emen s
●collec ed ollowing he Ocean Bes P ac ice guidelines (OBPS)
●e alua ed based on an o e all scien ific consensus depending on
he a iable
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Quali y con ol wi hin
he EOV/ECV amewo k
Quali y con ol is a p ocess ha is ollowed o e alua e
and label he measu ed da a o ensu e hei quali y
E alua e: h ough se ies o
es s (depending on he
measu ed a iable and
measu ing pla o m)
Label: assign a quali y flag
based on a flagging sys em (i.e.
IOC 54:V3)
Bu , we need o conside :
1) QC in Nea -Real Time (NRT) o Delayed Time (DT)?
2) QC is pla o m specific bu how does his wo k wi hin he EOV
amewo k?
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Example: ope a ional
mul i-pla o m moni o ing
POSEIDON/HCMR DATA FLOW SCHEME
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Example: ope a ional
mul i-pla o m moni o ing
Map showing he loca ion o he POSEIDON sys em pla o ms a CRETAN PSS (AB:
A hos Buoy, MB: Mykonos Buoy, PB: Pylos Buoy, SB: Sa onikos Buoy, E1-M3A Buoy),
glide endu ance line ( ed line) and Fe ybox ou es (yellow and g een line)
(h ps://www.je ico- i.eu/p ojec s/je ico-s3/pilo -supe si es/c e an-sea-pilo -supe si e/ )
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Obse ing pla o ms No 2015 - No 2025
●Con inuous need o deli e
he “bes ” quali y da a
●Add ess s akeholde needs
⇒ Wha is he Aegean Sea
Sea Su ace Tempe a u e
du ing he las decade?
EOV amewo k
Sou ce: Poseidon/HCMR Yellow: moo ed buoys
Cyan: A go floa s
G een: Bo le and CTDs
Magen a: glide
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Why use AI in QC?
⇒ Ocean da a QC is labo -in ensi e, especially
when i in ol es isual inspec ion
Example: A go equi es manual e iew o each
p ofile, e en hough da a a e needed in nea – eal
ime o modeling (Bi ig e al., 2019)
⇒ G owing demand o high-quali y ope a ional
da a o be used in o ecas ing, ea ly-wa ning
sys ems, and moni o ing
●Au oma ion could educe human
wo kload and speed up da a
deli e y
●ML could imp o e he
iden ifica ion o anomalous da a
by “lea ning” he beha iou o
ocean a iable
●Po en ially p o ide mo e ealis ic
alue anges, suppo ing
unce ain y quan ifica ion
… be e , as e , s onge …
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Sou ces o unce ain y
1) Δx1 = Ins umen al accu acy
2) Δx2 = Senso biases
3) Δx3 = Da a collec ion echniques
4) Δx4 = Calib a ion
5) Δx5 = Da a collec ion condi ions
….
n) Δxn = he lis goes on …
JCMC, 2008; Cowley e al.,
2021; Waldmann e al, 2022
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
EOV/ECV unce ain y challenges
Mul i-pla o m app oach
Wha happens when he e a e
simul aneous measu emen s o an
EOV om mul iple pla o ms,
emembe ou C e an Sea slide?
1) Which pla o m do we us
mo e? hie a chy?
2) A e age? need o know ALL he
Δxs ( om equa ions (1) AND (2)!!!
E o ????
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Ocean su ace s ess
EOV/ECV unce ain y challenges
Mul i-pa ame e EOV/ECVs
W óbel-Niedźwiecka e al.,
2019 (Oceanologia)
Vec o a iable: Added complexi y (need o
con e o pola coo dina es e c)
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
EOV/ECV unce ain y challenges
Implemen a ion
How can we use AI o achie e his? ⇒
EXPLORATORY wo k = we don’ ha e he answe s ye
Wha we DO know:
la ge olumes o da a can be
used o ain AI/ML model(s)
o ep oduce he beha io o
an EOV/ECV (success ul in
empe a u e)
Wha we DON’T know YET:
Can we also ain he AI/ML
models o ep oduce he
unce ain y?
Tes ing hypo hesis:
I a model accu a ely p edic s a a iable du ing he es ing pe iod,
he di e ence be ween p edic ed and obse ed alues
→ can be used o flag ques ionable da a and
→ con ibu es di ec ly o unce ain y es ima es.
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean
Resea ch Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
AI-d i en QC and unce ain y
es ima es o SST in he
Medi e anean Sea: Me hodology
APPLICATION
●Ex end o mo e
loca ions
●Consolida e esul s
PREPARATION
●Iden i y egions o
in e es
●Define a se o
models o be
conside ed o
u he analysis
TESTING
●Which model
cap u es/ ep oduce
s be e he EOV o
in e es
●Tes wi h da a
classifica ion, based
on QC
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean
Resea ch Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Candida e Models
LSTM: Long-Sho Te m Memo y
•RNN (Recu en Neu al Ne wo k)
ha u ilizes memo y cells
•Cap u es and models long e m
dependencies
•Compu a ionally expensi e
•1D, canno cap u e spa ial ea u es
•Widely used o SST ime se ies
CNN: Con olu ion Neu al Ne wo k
•U ilizes con olu ional laye s o
ex ac pa e ns
•Compa able o LSTM
•Compu a ionally less expensi e
han LSTM
•Used o spa ial SST pa e ns
(2D,3D)
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean
Resea ch Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Expe imen al se up o model es ing
DATASETS
EOV: SST
REGION: Eas e n Medi e anean (No h Aegean)
DATASETS: Remo e Sensing SST (Medi e anean Sea - High Resolu ion L4 Sea Su ace Tempe a u e
Rep ocessed, A hos Moo ed buoy empe a u e (Medi e anean Sea In-Si u Real Time Obse a ion)
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean
Resea ch Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Expe imen al se up o model es ing
DATASETS (mul i-pla o m app oach)
Cope nicus Ma ine Se ice, Medi e anean Sea In-Si u Nea Real Time Obse a ions
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean
Resea ch Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Expe imen al se up o model es ing
EOV: SST
DATASETS: Cope nicus Ma ine Se ice Sa elli e SST L4
70% 🡪 aining
15% 🡪 alida ion
15% 🡪 es ing
🡺 3-4 days o calcula e he op imal hype pa ame e s o LSTM (190 nodes), CNN is much as e ( o a p e-se
numbe o laye s)
🡺 12 days window o a ge 13 h day as p edic ion
TRAINING
PERIOD
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean
Resea ch Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
Model T aining Assessmen
T aining loss: e o he model makes on he aining da ase du ing he aining pe iod
Valida ion loss: e o on he alida ion da ase du ing he aining pe iod ( his da a is no included in he
aining da ase )
Epoch: he numbe o imes he model ains on he en i e aining da ase
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean
Resea ch Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
IQuOD 1.0, Cowley e al, 2021 (F on ie s in Ma ine Science)
Pla o m-specific
empe a u e unce ain ies a e
well documen ed o in-si u
measu emen s
Funded by he Eu opean Union. Views and opinions exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he Eu opean Union o he Eu opean Resea ch
Execu i e Agency (REA). Nei he he Eu opean Union no he g an ing au ho i y can be held esponsible o hem.