A po able and low-cos op ical de ice o pigmen -based axonomic
classi ica ion o mic oalgae using machine lea ning
Vi o Magalh˜
aes
a
, Vˆ
ania Pin o
a,c
, Paulo Sousa
a,c
, Jos´
e A. A onso
a,c
, Luís Gonçal es
a,c
,
Emilio Fe n´
andez
b
, G aça Minas
a,c,*
a
CMEMS-UMinho, Uni e si y o Minho, Campus de Azu ´
em, Guima ˜
aes 4800-058, Po ugal
b
Cen o de In es igaci´
on Ma ina. Uni e sidade de Vigo, Vigo 36310, Spain
c
LABBELS –Associa e Labo a o y, B aga, Guima ˜
aes, Po ugal
ARTICLE INFO
Keywo ds:
Mic oalgae iden i ica ion
HABs
Fluo ome y
Machine Lea ning
Po able de ice
ABSTRACT
The p oli e a ion o ce ain phy oplank on species may lead o ha m ul algal blooms (HABs) ha can a ec li ing
esou ces and human heal h. The e o e, an accu a e iden i ica ion o phy oplank on popula ions is essen ial o
he sus ainable managemen o some ac i i ies ele an o he blue economy, such as aquacul u e, being also
ele an o en i onmen al moni o ing and ma ine esea ch pu poses. Mic oalgae axonomic disc imina ion,
based on hei pigmen composi ion, is a e sa ile and p omising echnique o de ec and iden i y po en ial HABs.
In his wo k, a po able and low-cos de ice o axonomic iden i ica ion o mic oalgae, based on he pigmen
composi ion o 16 species belonging o 6 di e en phyla, was de eloped. I uses he luo escence in ensi y signal
emi ed by each species a h ee wa eleng hs (575 nm, 680 nm and 730 nm) when exci ed a i e wa eleng hs
(405 nm, 450 nm, 500 nm, 520 nm and 623 nm) o c ea e a luo escence signa u e o each species. Fu he mo e,
se e al machine lea ning classi ie s we e s udied using his luo escence signa u e as ea u es o ain and classi y
each species acco ding o hei espec i e axonomic g oup. The Ex eme G adien Boos ing (XGBoos ) classi ie
was able o co ec ly p edic mic oalgae monocul u es wi h 97 % accu acy a he phylum le el and 92 % ac-
cu acy a he o de le el. The ob ained esul s con i m he po en ial o his echnique o as , accu a e and low-
cos iden i ica ion o mic oalgae.
1. In oduc ion
Mic oalgae play a c ucial ole in assessing wa e quali y, as well as
ma ine and public heal h. They se e as impo an indica o s o he
o e all heal h and ecological balance o aqua ic ecosys ems [1]. How-
e e , hey can also cause en i onmen al p oblems when some species
p oli e a e exponen ially, a na u al phenomenon known as ha m ul
algal blooms (HABs). Some HABs a e cons i u ed by species capable o
p oducing oxins, o he s a e non- oxic high biomass p oli e a ions ha
can lead o wa e discolo a ion and oxygen deple ion [2]. These e en s
can ake place in la ge o small a eas depending on he species and
ex e nal condi ions. The e icien and imely moni o ing, wi h la ge
spa ial esolu ion, o hese isk si ua ions is essen ial o sa egua d public
heal h and blue economy, pa icula ly shell ish and in ish aquacul u e.
Howe e , he implemen a ion o hese moni o ing p og ammes is
cu en ly complex, expensi e and demands skilled pe sonnel.
Phy oplank on is composed o mic oscopic pho osyn he ic
unicellula o ganisms whe e sizes ange om less han 1
μ
m o colonies
la ge han 500
μ
m, and comp ises di e se e olu iona y lineages,
including euka yo es and cyanobac e ia, ha show di e ences in cell
mo phology, o namen a ion, pho osyn he ic pigmen s (chlo ophylls,
ca o enoids and phycobilins) and o he biochemical ma ke s. All hese
ea u es allow disc imina ion be ween species [3–5].
S anda d me hods o phy oplank on bloom iden i ica ion and
quan i ica ion ely mos ly on mo phological analysis and/o pigmen
composi ion [6,7]. The mos commonly used me hod is image-based
analysis, which is usually done h ough mic oscopy and equi es
well- ained pe sonnel. Despi e being e y obus in de ec ing and
quan i ying a ge species, i is ime-demanding and es ic ed o
lab-based ope a ions. Pigmen composi ion-based me hods a e able o
dis inguish mic oalgae due o hei spec al p ope ies using luo o-
me ic echniques. Se e al esea ch wo ks ha e subs an ia ed he
iabili y o employing luo escence echniques o he classi ica ion o
mic oalgae [4,7–10]. Howe e , exis ing de ices, such as FlowCAM [11],
* Co esponding au ho a : CMEMS-UMinho, Uni e si y o Minho, Campus de Azu ´
em, Guima ˜
aes 4800-058, Po ugal.
E-mail add esses: [email p o ec ed] (V. Pin o), [email p o ec ed] (G. Minas).
Con en s lis s a ailable a ScienceDi ec
Senso s and Ac ua o s: B. Chemical
jou nal homepage: www.else ie .com/loca e/snb
h ps://doi.o g/10.1016/j.snb.2024.136819
Recei ed 3 Decembe 2023; Recei ed in e ised o m 7 Oc obe 2024; Accep ed 19 Oc obe 2024
Senso s & Ac ua o s: B. Chemical 423 (2025) 136819
A ailable online 21 Oc obe 2024
0925-4005/© 2024 The Au ho s. Published by Else ie B.V. This is an open access a icle unde he CC BY-NC-ND license (
h p://c ea i ecommons.o g/licenses/by-
nc-nd/4.0/ ).
Imaging Flow Cy obo [12], Cy oBuoy [13] and Lase Op ical Plank on
Coun e [14], while accu a e, a e o en bulky and expensi e, which
migh hinde hei applica ion o eal- ime, on-si e de ec ion o ha m ul
algae o e la ge spa ial scales, especially in emo e o
esou ce-cons ained en i onmen s.
In addi ion o hese de ices, he e is an ongoing e o o analyze
phy oplank on and hus p edic HABS, using sa elli e ocean colo im-
aging [15] and me eo ological o ecas s [16], and while use ul o b oad
moni o ing, lack he esolu ion and speci ici y needed o iden i y di e se
phy oplank on species in coas al a eas whe e HABs a e mos p e alen .
Unmanned ae ial ehicles (UAVs) ha e been also used o moni o
possible HAB ou b eaks wi h high spa ial esolu ion and cos -
e ec i eness, bu ba e y powe consump ion and bad wea he a e
limi ing ac o s o ex ended sample collec ion and da a quali y when
using his app oach [17,18].
The in eg a ion o senso s and sa elli e da a wi h machine lea ning
(ML) and deep lea ning (DL) is a ho opic in he esea ch ield, which
allowed u he expanding he capabili y o accu a e, as e , cheape
and widesp ead moni o ing o phy oplank on, by combining da a om
sa elli es and/o au onomous senso s capable o in si u moni o ing
ea u ing also he capabili y o ML and DL o eal ime analysis o he
collec ed da a se [19–23]. The e is al eady ex ensi e esea ch showing
he easibili y o combining luo escence-based op ical senso s wi h ML
and DL wi h p omising esul s [24–27].
As he need o e icien and eliable in si u phy oplank on iden i i-
ca ion has g own, he de elopmen o po able and au onomous de ices
has become inc easingly impo an . Despi e some p og ess in his a ea,
cu en po able de ices s ill ace challenges ega ding cos , po abili y,
and eal- ime iden i ica ion accu acy. The e is an u gen need o a
compac , a o dable, and highly sensi i e de ice ha can be massi ely
deployed o sys ema ic in-si u moni o ing o bo h oxic and non- oxic
phy oplank on. Such a ool would ha e a signi ican impac on unde -
s anding HABs dynamics, he eby p o iding c i ical da a o p edic , o
ins ance, shell ish and in ish oxi ica ion episodes, enabling sus ainable
ha es ing decisions and sa egua ding public heal h.
Thus, in his wo k, a po able and low-cos (es ima ed less han 250
EUR) de ice o axonomic iden i ica ion o mic oalgae species based on
luo ome y and ML was de eloped. Ou de ice add esses he limi a-
ions o exis ing me hods by o e ing a compac , a o dable, and po able
solu ion, enabling on-si e classi ica ion o mic oalgae.
By making his echnology accessible, e en in de eloping coun ies,
he de ice p o ides ea ly de ec ion o impo an mic oalgal popula ions,
such as oxic species, o e ing a ans o ma i e app oach o managing
HABs and hei en i onmen al and economic impac s, b idging he gap
be ween cu en echnology and he need o eal- ime, accessible
moni o ing. The compac design and low powe equi emen s ensu e he
de ice can be widely used, e en in emo e o esou ce-limi ed egions,
whe e adi ional bulky lab equipmen canno be deployed.
The de ice uses he luo escence signal o se e al species emi ed a
h ee wa eleng hs egions (575 nm, 680 nm, 730 nm) when exci ed a
i e wa eleng hs (405 nm, 450 nm, 500 nm, 520 nm and 623 nm). These
exci a ion and emission wa eleng hs we e selec ed based on luo om-
e y analysis, using a comme cial equipmen , scanning he exci a ion
spec a and acqui ing he emission spec a o each species. The
di e en pa e ns o mic oalgae luo escence ha a ise om he com-
bina ion o a ious exci a ion sou ces a di e en emission egions allow
o classi ica ion o mic oalgae h ough ML. Se e al ML classi ie s we e
es ed and op imized in o de o ind he bes model ha p oduced he
mos accu a e esul s.
2. Ma e ials and me hods
2.1. Phy oplank on cul u es
A o al o 16 species belonging o 6 phyla we e selec ed o conduc
he s udy. The phy oplank on monocul u es we e collec ed om he
To alla Ma ine Science S a ion (ECIMAT), Vigo, Spain. A desc ip ion o
hei axonomic classi ica ion (Table S1) as well as hei main cha ac-
e is ics, including mo phological a iables, pigmen s and oxins
(Table S2) a e desc ibed in he supplemen a y ma e ial, sec ion S1.
2.2. Fluo escence cha ac e iza ion
The emission and exci a ion spec a o monocul u es species we e
ob ained using he Shimadzu RF-5301PC spec opho ome e a he
To alla Ma ine Science S a ion (ECIMAT). Knowledge on he exci a ion
spec a o each species was key o selec he ligh -emi ing diodes (LEDs)
sou ces ha allowed he bes po en ial disc imina ion be ween di e en
axonomic g oups.
2.3. Po able de ice o luo escence measu emen s
Based on a luo escence spec al disc imina ion app oach, a po able
de ice o de ec ing mul iple luo escence signals was implemen ed.
This me hod in ol es exci ing mic oalgae pigmen s wi h ligh sou ces a
di e en exci a ion egions and quan i ying he subsequen emission o
ligh in egions o in e es . The de ice includes ou main subsys ems:
(1) an exci a ion sou ce; (2) a pho ode ec ion sys em; (3) an elec onics
sys em; and (4) a powe sou ce consis ing o ba e ies. A 3D ep esen-
a ion o he op ical appa a us o he p o o ype is shown in Fig. S1 o
supplemen a y ma e ial.
The exci a ion sou ce comp ises a se o high-in ensi y LEDs (Poly-
ch oma ic LuxiGen Mul i-Colo LED 897-LZ7A4M2PD0000 plus a Bi a
UV3TZ-405–30) a wa eleng hs ha co e he ange o in e es (405 nm,
450 nm, 500 nm, 520 nm and 623 nm, ou pu ed om Sec ion 3.1)
(Fig. S2 o supplemen a y ma e ial); hey a e a anged in a compac a ea
(3.4 ×3.4 mm
2
) acili a ing i s in eg a ion in minia u ized de ices. Each
LED is ac i a ed o 1 second o limi he dec ease in luo escence in-
ensi y due o he pho ochemical quenching e ec [28].
The pho ode ec ion sys em includes a silicon pho odiode (Hama-
ma su, S1336–8BK) placed below he cu e e holde o cap u ing he
luo escence in ensi y, wi h a ange o sensi i i y om ul a iole o
in a ed. I s esponsi i y, high-quan um e iciency, non-uni o mi y,
non-linea i y and low-noise makes i widely used in op ical measu e-
men equipmen . Abo e he pho odiode he e is a mobile d awe wi h
h ee op ical bandpass il e s (cen ed wa eleng h ou pu ed om Sec-
ion 3.1). One cen ed a 575 nm wi h a FWHM o 10 nm (575BP10 om
Lase Componen s); o he cen ed a 680 nm wi h a FWHM o 10 nm
(FB680–10 om Tho labs); and he hi d cen ed a 730 nm wi h a
FWHM o 13 nm (730BP15 om Lase Componen s) (Fig. S3 o sup-
plemen a y ma e ial). They a e used o supp ess he exci a ion signals
and enhance sensi i i y by signi ican ly educing backg ound noise.
Addi ionally, he exci a ion and de ec ion subsys ems a e assembled in a
90◦con igu a ion using a 3D-p in ed suppo o ensu e he co ec
alignmen and posi ioning o he op ical componen s.
The signal acquisi ion is suscep ible o ambien noise in e e ence,
elec onic componen s and ci cui in e ac ions sou ces, hampe ing he
signal- o-noise a io. So, a lock-in ampli ie [29] was used o de ec ion
and measu emen o low ampli ude signals, e en when he signal o
in e es is comple ely embedded in noise. I uses a phase-sensi i e
de ec ion echnique, which isola es he measu ed signal h ough a
e e ence signal wi h a ce ain equency and phase. In his way, he
noise p esen a equencies di e en om he ones o he e e ence
signal will be ejec ed. He e, his is achie ed by modula ing he exci-
a ion signal a 1 kHz (pulsing he exci a ion LEDs a 1 kHz). The
pho ocu en ou pu ed by he pho odiode, a e con e ed o ol age, is
il e ed by a 1 kHz band-pass il e isola ing he desi ed signal om
unwan ed noise equencies. To es o e he o iginal signal, a synch o-
nous demodula o , ope a ing a 1 kHz, ansla es he modula ed signal
in o di ec cu en (DC), e ec i ely elimina ing un-synch onized signals
[29,30]. Finally, he use ul signal will be p ocessed by a compu e
h ough a mic ocon olle .
V. Magalh˜
aes e al.
Senso s and Ac ua o s: B. Chemical 423 (2025) 136819
2
The de eloped de ice o mic oalgae luo escence de ec ion is shown
in Fig. 1. Thei compac design and small size p o ides po abili y
allowing hei ope a ion ou side he labo a o y.
2.4. Machine lea ning algo i hm
Fo he ML implemen a ion, i s , he mic oalgae species we e
associa ed wi h hei espec i e axonomic phylum. Fo ins ance, he
Alexand ium ama ense and Alexand ium minu um species we e associ-
a ed o he Miozoa phylum hey belong o (see Table S2 o supplemen-
a y ma e ial). The wo species o cyanobac e ia despi e belonging o he
same phylum, we e conside ed as wo dis inc g oups (Cyanobac e ia 1
and Cyanobac e ia 2) due o hei e y dis inc pigmen composi ion.
Fluo escence in ensi y a ios we e calcula ed be ween he exci a ion
and emission wa eleng hs ha p o ided he g ea es dis inc ion,
c ea ing a unique spec al signa u e o each species (mo e de ails
desc ibed in Sec ion 3). These signa u es we e hen analyzed by ML
models o axonomic classi ica ion.
To a oid any possibili y o da a leakage and ensu e eliable model
e alua ion, he da ase was andomly spli in o aining (80 %) and
es ing (20 %) se s. This esul ed in 148 samples o aining and 38
samples o es ing. The andom spli ing was done a he species le el,
ensu ing ha he same species did no appea in bo h he aining and
es se s o a oid o e i ing o da a leakage. To u he ensu e he
obus ness o he model, a 3- old c oss- alida ion was employed on he
aining se . This means he aining se was spli in o h ee equal pa s,
wi h wo pa s used o aining and one pa o alida ion, cycling
h ough all combina ions. Fu he de ails o he es ed ML classi ie s a e
in Sec ion S6 o he Supplemen a y In o ma ion.
Due o he logis ical challenge o cul u ing 16 di e en mic oalgae
species a a quasi-synch onous g ow h phase, he aining se had a
ela i ely small numbe o samples. To o e come his limi a ion, a
di e se a ay o ML classi ie s wi h a ying hype pa ame e s was
employed o op imize model pe o mance.
3. Resul s and discussion
3.1. Fluo escence spec a
Fig. 2 shows he luo escence emission measu ed a 575 nm, 680 nm
and 730 nm, espec i ely Fig. 2a, b and c, o e a ange o exci a ion
wa eleng hs. The emission a 575 nm co esponds o he phycoe y h in
pigmen emission, an impo an ma ke pigmen o c yp ophy es (RI)
and cyanobac e ia (Syn034). The 680 nm and 730 nm co espond
mos ly o he chlo ophyll an emission. Only one species ep esen a i e
o each phylum (wi h excep ion o cyanobac e ia) is ep esen ed in
Fig. 2 o a oid con using isualiza ion due o excessi e o e lapping o
cu es.
The o e all luo escence o mic oalgae is a complex in e play o
a ious pigmen s and hei in e ac ions wi hin he pho osyn he ic
appa a us. Due o a ia ions in hei pigmen composi ion, di e en
algal g oups exhibi dis inc pa e ns o pho on abso p ion depending on
he exci a ion egion in he spec um. As a esul , hey emi luo escence
wi h a ying pa e ns allowing explo a ion o he ela ionships be ween
emi ed luo escence om di e en exci a ion sou ces, he eby iden i-
ying disc iminan ea u es be ween algal g oups.
Species like Rhodomonas lens (Rl) and Synechoccocus sp. (Syn033 and
Syn034) ha e e y dis inc luo escence p ope ies because o hei
unique pigmen composi ion, allowing easy dis inc ion be ween hem.
Fo ins ance, Rhodomonas lens (Rl) and Synechoccocus (Syn034) bo h
ha e he pigmen phycoe y h in, and, o ha eason, hey emi luo-
escence a 575 nm when exci ed a 500–540 nm. Ano he Synechocco-
cus species (Syn033) has he pigmen allophycocyanin wi h an exci a ion
peak a 620 nm, which also se s i apa om o he species. The
emaining species ha e exci a ion peaks much mo e simila and luo-
escence emission only a ound 680–730 nm egion, making disc imi-
na ion challenging.
We selec ed i e speci ic egions wi hin he spec um whe e he
a ia ions and magni ude in luo escence in ensi y we e mos p o-
nounced among he species (see Fig. 2). The chosen exci a ion wa e-
leng hs we e cen ed a 405 nm, 450 nm, 500 nm, 520 nm, and 623 nm.
Fig. 3 ep esen s he emission spec a o 8 species belonging o 7
di e en phyla, o he 5 selec ed wa eleng hs. Disc imina ion be ween
Rhodomonas lens (Rl) and he wo Synechoccocus species is qui e
s aigh o wa d. Thei unique pigmen composi ion gi es ise o e y
dis inc luo escence pa e ns when exci ed a he 5 selec ed wa e-
leng hs. The mo e challenging pa is he disc imina ion be ween species
ha only emi in he 680–730 nm egion, which is he case o mos
species in his s udy. Despi e ha ing luo escence emission a simila
egions, hei di e en pigmen composi ion a ec s how each species
will abso b each di e en wa eleng h, and, as a esul , he in ensi y o
emi ed luo escence a his speci ic wa eleng h will be sligh ly
di e en . These di e ences can p o ide dis inguishing ea u es o ML
classi ie s ha can be used o iden i y di e en axonomic g oups o
mic oalgae whe e luo escence emission o e laps.
Fig. 1. Po able de ice o de ec mul iple luo escence signals.
V. Magalh˜
aes e al.
Senso s and Ac ua o s: B. Chemical 423 (2025) 136819
3
3.2. De ice alida ion
A e an ini ial calib a ion (p ocedu es de ailed in sec ion S3 o
supplemen a y ma e ial), he de eloped de ice was used o measu ing
he emission signals o di e en monocul u es a he wa eleng hs p e-
iously s a ed. The samples we e collec ed a he s a iona y phase o
g ow h, al hough due o he di e en g ow h a es o he es ed species,
dilu ions we e pe o med o main ain simila concen a ions. The
de ec ed luo escence signal was alida ed by compa ing he ob ained
measu emen s wi h he exci a ion spec a o he mic oalgae. Co ec ion
ac o s, using he measu ed mean alue o he il e ed seawa e samples
(n =8), we e applied o he mic oalgae luo escence signal, yielding a
inal alue o 1 ( o co ec ing possible inhe en biases). These co ec-
ion ac o s, speci ic o each combined LED and emission il e , a e
de ailed in Table S3 o he supplemen a y ma e ial. When conduc ing
measu emen s, hese co ec ion ac o s a e applied o he pho odiode
ou pu ol age, esul ing in an adjus ed luo escence emission signal.
Fig. 4 shows he luo escence pa e n eco ded by he de ice
compa ed wi h he espec i e exci a ion spec a o a selec ion o
mic oalgae ep esen a i e o each phylum. In gene al, he luo escence
pa e ns a e in good ag eemen wi h he exci a ion cu es. Rhodomonas
lens (Fig. 4a) and Synechococcus species (Fig. 4b c) exhibi dis inc
luo escence pa e ns. Howe e , sub le di e ences we e ound in he
luo escence pa e ns o he o he species. Focusing only on he luo-
escence emission in he egions o 680 nm and 730 nm, bo h Alexan-
d ium minu um (Fig. 4d), Diac onema lu he i (Fig. 4e) and Skele onema
cos a um (Fig. 4 ) exhibi s signi ican ly lowe luo escence emission
when exci ed a 623 nm ( ed ba ) compa ed wi h exci a ion a 520 nm
(g een ba ). By con as , Nannochlo opsis gadi ana (Fig. 4g) and Te a-
selmis suecica (Fig. 4h) showed luo escence emission alues a 623 nm
ei he simila o e en highe compa ed o hose a 520 nm. Alexand ium
minu um (Fig. 4d) shows sligh ly lowe luo escence emission when
exci ed a 405 nm (pu ple ba ) when compa ed o 500 nm exci a ion
(cyan ba ). In con as , Nannochlo opsis gadi ana (Fig. 4g), Te aselmis
suecica (Fig. 4h) and Diac onema lu he i (Fig. 4e) display highe luo-
escence emission a an exci a ion o 405 nm compa ed o a 500 nm,
while Skele onema cos a um (Fig. 4 ) p esen s simila in ensi y be ween
bo h exci a ions. These esul s align well wi h he emission spec a
depic ed in Fig. 3 and allowed assigning a luo escence signa u e o each
species, based on he luo escence pa e ns measu ed by he de ice.
3.3. Fluo escence signa u e
A luo escence signa u e was ixed employing a a iome ic app oach
based on a ios be ween luo escence signals ob ained using he i e
di e en exci a ion LEDs o each emission egion (basically, de i ed
om Fig. 4). Speci ically, o emissions a 680 nm and 730 nm, we
de i ed a o al o en a ios o each emission. In he case o he 575 nm
emission, a single a io was de e mined, in ol ing he 520 nm and
405 nm LEDs. Each abb e ia ion le e co esponds o a speci ic a io
be ween wo exci a ion LEDs (see able S4 o supplemen a y ma e ial)
and he numbe , 575 nm, 680 nm o 730 nm, co esponds o he emis-
sion wa eleng h used (Fig. 5, x-axis). This means ha each species is
cha ac e ized by a o al o 21 a ios (10 a ios o he emission egion a
680 nm, 10 a ios o he emission egion a 730 nm, and one a io o
he 575 nm emission egion), each a io co esponding o a poin o he
luo escence signa u e o each species. Fig. 5 shows he luo escence
signa u e o all he species es ed, calcula ed om he a e age o
12–20 measu emen s o each species. As an example, he a io C575
( he i s in he x-axis) co esponds o he a io o emi ed luo escence
in ensi ies a 575 nm when exci ed wi h he 520 nm and he 405 nm
LEDs. Since only Rhodomonas lens (Rl) and Synechoccocus (Syn034) ha e
phycoe y h in, hey can be easily dis inguished om o he species by
using ha a io. Synechoccocus (Syn033) shows a high signal a 623 nm
Fig. 2. Exci a ion spec a o luo escence emission a a) 575 nm (dashed lines co espond o a zoomed sec ion o he plo ), b) 680 nm, and c) 730 nm. Each line
ep esen s he exci a ion spec um o a di e en species, wi h abb e ia ions co esponding o hose lis ed in Table S2 o he supplemen a y ma e ial.
V. Magalh˜
aes e al.
Senso s and Ac ua o s: B. Chemical 423 (2025) 136819
4
because o he p esence o phycocyanin, and o ha eason, hei I and J
a ios a 680 nm and 730 nm emissions a e clea ly dis inguished om
hose o o he species. Howe e , o mos a ios, iden i ying dis-
inguishing pa e ns be ween di e en species is mo e complex, and his
is he case whe e ML assumes an impo an ole, as hese a ios ( luo-
escence signa u es) can be used as ea u es in he ML models.
Le e aging hese ea u es ( a ios), he ML models will s i e o di e -
en ia e and ca ego ize mic oalgae, wi h he ul ima e goal o maximizing
accu acy in assigning da a poin s o hei app op ia e classes. The
adop ed me hodology is hus designed o iden i y he p esence o a
speci ic axonomic g oup in a gi en sample.
3.4. Machine lea ning classi ica ion esul s
A o al o 12 ML classi ie s we e es ed o ind he one ha ou pu ed
he bes pe o ming in he da ase . In he aining se , he model con-
s uc s he classi ie based on he labeled obse a ions and ies o
sepa a e he p ede ined classes as bes as i can. In he es ing se , he
model p edic s he class o unlabeled obse a ions.
The p ocess is di ided in o wo pa s. Fi s , each model was ained
and alida ed using only he aining se wi h a c oss- alida ion o 3.
This means ha he aining se was spli in a way ha 1/3 ep esen ed a
alida ion se and 2/3 he aining se . This was epea ed 3 imes,
allowing all aining samples o be included in he alida ion se a some
poin . E e y classi ie was also op imized by i e a ion o se e al
hype pa ame e s and he ones yielding he bes sco e we e sa ed. This
c oss- alida ion echnique along wi h he i e a ion o se e al hype -
pa ame e s maximized he obus ness and accu acy o he esul s,
mi iga ing he impac o he limi ed aining da a. In he second pa , he
op imized classi ie s wi h he bes accu acy we e ained wi h he whole
aining se , and hei classi ica ion pe o mance was e alua ed using
he es ing se . Accu acy was he main me ic used o e alua e he
pe o mance o he ML models. Accu acy is de ined as he a io be ween
he numbe o co ec ly p edic ed samples and he o al numbe o
samples. I measu es he p opo ion o co ec p edic ions made by he
model o e all p edic ions, bo h ue posi i es and ue nega i es.
Howe e , o he me ics such as p ecision and ecall we e also used o
p o ide a mo e comp ehensi e pe o mance e alua ion, especially when
handling po en ial imbalances in he da ase . These addi ional me ics
help assess he model’s pe o mance in classi ying less- ep esen ed
Fig. 3. Emission spec a o he exci a ion a 405 nm, 450 nm, 500 nm, 520 nm and 623 nm, o 8 phy oplank on species ep esen a i e o 7 di e en phyla.
V. Magalh˜
aes e al.
Senso s and Ac ua o s: B. Chemical 423 (2025) 136819
5
Fig. 4. Compa ison o he luo escence pa e n eco ded by he de ice (ba plo s) wi h he luo escence exci a ion spec um o a) Rhodomonas lens (n=5), b and c)
Synechoccocus species (n=23), d) Alexand ium minu um (n=14), ) Skele onema cos a um (n=13), g) Nannochlo opsis gadi ana (n=11), and h) Te aselmis suecica
(n=12). E o ba s ep esen he s anda d de ia ion in n samples measu ed.
V. Magalh˜
aes e al.
Senso s and Ac ua o s: B. Chemical 423 (2025) 136819
6
Fig. 4. (con inued).
V. Magalh˜
aes e al.
Senso s and Ac ua o s: B. Chemical 423 (2025) 136819
7
classes, mi iga ing he isk o bias owa d majo i y classes.
Table 1 shows he selec ed classi ie s, hei espec i e aining ime
and p edic ion ime, and he accu acy, which ep esen he accu acy o
he models on he aining da a o he bes hype pa ame e s. Mo e
de ailed in o ma ion abou he classi ie s and espec i e hype -
pa ame e s is p esen ed in supplemen a y ma e ial S6. The classi ie s
wi h accu acy equal o o abo e 0.90 we e selec ed o classi y he es
da a. The accu acy o he selec ed classi ie s o he es ing da a a e
p esen ed in Table 2.
XGBoos is an ensemble ML algo i hm widely popula o i s excep-
ional pe o mance on s uc u ed da a, consis en ly achie ing s a e-o -
he-a esul s ac oss a a ie y o ML asks [31]. I builds mul iple de-
cision ees sequen ially, each one imp o ing upon he e o s made by
he p e ious ees. XGBoos ’s key s eng h lies in i s abili y o co ec
esidual e o s i e a i ely, making i pa icula ly e ec i e o s uc u ed
da ase s and obus agains small da ase size and imbalances, which
would o he wise hinde models such as Random Fo es o Neu al Ne -
wo ks. While Neu al Ne wo ks equi e la ge da ase s o a oid o e -
i ing, XGBoos is mo e esis an o o e i ing due o i s use o g adien
boos ing and egula iza ion echniques, an impo an ad an age when
wo king wi h smalle da ase s like he one in his s udy. Addi ionally,
XGBoos employs pa alleliza ion and e icien memo y usage, making i
ela i ely as e and mo e e icien han many o he classi ie s [32,33].
Analyzing he selec ed hype pa ame e s o XGBoos , a lea ning a e
o 0.1 indica es ha he model lea ns g adually, wi h each ee
con ibu ing 10 % o he ensemble’s p edic ion. The ees ha e a max
dep h o 3, which helps p e en o e i ing and keeps he model simple .
Wi h 50 es ima o s, XGBoos was able o cap u e pa e ns in he da a
wi hou becoming o e ly complex. This combina ion o pa ame e s al-
lows XGBoos o s ike a balance be ween model complexi y and
accu acy.
The classi ica ion ma ix o he XGBoos , ained wi h he whole
aining se and espec i e es ing se , is ep esen ed in Fig. 6, showing
he co ec ly and inco ec ly classi ied phyla. This ma ix p o ides a
de ailed b eakdown o he p edic ions made by he model (p edic ed
label) and how hey compa e agains he eal axonomic classi ica ion o
Fig. 5. Fluo escence signa u e composed o 21 a ios o each species.
Table 1
Resul s o aining and alida ion o se e al ML classi ie s a he phylum le el.
Classi ie T aining
ime (s)
P edic ion
ime (s)
Bes pa ame e s Accu acy
o
aining
da ase
Nea es
Neighbo s
18.0059 0.0171 ’n_neighbo s’: 2 0.94
Decision T ee 0.1747 0.0003 ’max_dep h’: 6 0.87
Random
Fo es
70.5316 0.0136 ’max_dep h’: 2,
’max_ ea u es’: 4,
’min_samples_lea ’: 25,
’n_es ima o s’: 100
0.29
Logis ic
Reg ession
1.2498 0.0003 ’C’: 1e−05, ’penal y’:
’none’
0.93
Suppo
Vec o
Machine
0.2005 0.0015 ’C’: 10 0.93
Neu al
Ne wo k
190.5183 0.0033 ’ac i a ion’: ’ anh’,
’alpha’: 1,
’hidden_laye _sizes’:
(100, 100, 100),
’lea ning_ a e’:
’cons an ’
0.96
AdaBoos 5.3509 0.0120 ’lea ning_ a e’: 1,
’n_es ima o s’: 50
0.93
Nai e Bayes 0.0693 0.0021 ’ a _smoo hing’:
1e−09
0.87
Linea
Disc iminan
Analysis
0.7927 0.0002 ’sh inkage’: 0, ’sol e ’:
’lsq ’, ’ ol’: 0.0001
0.92
Quad a ic
Disc iminan
Analysis
0.3432 0.0007 ’ eg_pa am’: 0.01,
’ ol’: 0.0001
0.81
G adien
Boos ing
277.7135 0.0130 ’lea ning_ a e’: 0.01,
’max_dep h’: 3,
’n_es ima o s’: 500
0.90
Ex eme
G adien
Boos ing
34.6022 0.0020 ’lea ning_ a e’: 0.1,
’max_dep h’: 3,
’n_es ima o s’: 50
0.92
Table 2
Accu acy o he classi ie s o he es da a.
Classi ie Accu acy o es da ase
Nea es Neighbo s 0.95
Logis ic Reg ession 0.87
Suppo Vec o Machine 0.89
Neu al Ne wo k 0.87
AdaBoos 0.87
Linea Disc iminan Analysis 0.82
G adien Boos ing 0.87
Ex eme G adien Boos ing 0.97
The Ex eme G adien Boos ing (XGBoos ) s ood ou wi h an accu acy in he es
da a o 0.97. This means ha i co ec ly classi ied 37 o he 38 es ing samples
(97 %).
Fig. 6. Tes da a classi ica ion ma ix o XGBoos a he phylum le el.
V. Magalh˜
aes e al.
Senso s and Ac ua o s: B. Chemical 423 (2025) 136819
8
mic oalgae ( ue label). The model was able o co ec ly classi y almos
all es ing da a o he co ec phylum, wi h he excep ion o one sample
belonging o he Bacilla iophy a phylum which was inco ec ly classi-
ied as Hap ophy a.
The nex s ep was o in es iga e he capabili y o XGBoos o
co ec ly classi y he mic oalgae a he o de le el. Fo ha , each species
was associa ed wi h one o he 14 axonomic o de s. The XGBoos was
ained o he o de le el and i s classi ica ion pe o mance was e al-
ua ed using he es ing se . Fo he o de le el, he XGBoos classi ie
showed an accu acy o 0.85 o he aining da a and an accu acy sco e
o 0.92 o he es ing da a. The con usion ma ix o he XGBoos
ained wi h he whole aining se and espec i e es ing se is ep e-
sen ed in Fig. 7, showing he co ec ly and inco ec ly classi ied axo-
nomic o de . He e, he model misclassi ied h ee es ing samples, a
Chlo ellales sample w ongly p edic ed as a Chlamydomonadales and
ice e sa, and a Bacilla iales inco ec ly classi ied as a Thalassiosi ales,
making i a o al o 35 (ou o 38) co ec ly classi ied samples.
A summa y o he main classi ica ion me ics o he XGBoos clas-
si ie a he phylum and o de le el is p esen ed in Table 3. In addi ion o
accu acy, o he me is such as p ecision, ecall and F1-sco e, as well as
he mac o/mic o-a e age sco es, we e used o e alua e he pe o mance
o he model h ough a se o di e en me ics and o ci cum en he
some imes-misleading cha ac e o accu acy, as his me ic may be high
due o he model’s abili y o accu a ely p edic he majo i y class, while
pe o ming poo ly on he mino i y class. A he phylum le el, all me ics
show high alues o he es da a, showcasing he po en ial o his
model o classi y each phylum co ec ly, wi h a weigh ed a e age o
0.97 and 0.98 o ecall and p ecision, espec i ely, showcasing he
model’s abili y o accu a ely classi y mic oalgae ac oss di e se axo-
nomic g oups. High alues o all me ics we e ob ained excep o he
Chlamydomonadales, Chlo ellales, Thalassiosi ales and Bacilla iales
o de s, which showed lowe p ecision and/o ecall. Despi e ha , he
o e all pe o mance o he model is sa is ac o y, wi h a weigh ed
a e age o 0.92 and 0.93 o ecall and p ecision, espec i ely. XGBoos
demons a ed excellen pe o mance no only in accu acy bu also in
obus ness agains o e i ing and he abili y o gene alize om a ela-
i ely small da ase , making i pa icula ly sui ed o applica ions whe e
da a collec ion may be limi ed, bu high classi ica ion accu acy is s ill
equi ed.
Mic oalgae classi ica ion, employing a ange o me hods based on
luo escence analysis has unde gone a subs an ial e olu ion and ine-
uning wi h he in eg a ion o ML echniques. The combina ion o con-
en ional luo escence me hods wi h he capabili ies o AI has opened
up new possibili ies and signi ican ly enhanced p ecision and e iciency
in dis inguishing and cha ac e izing a ious algae species. In his ega d,
Young-Ho e al. [34] demons a ed he dis inc ion and he quan i ica-
ion o g een algae and cyanobac e ia and Pe siche i e al. [9] employed
a a iome ic app oach dis inguishing g een algae, dia oms, and
nume ous cyanobac e ia species. In ou esea ch, we ocus on seawa e
species aiming o di e en ia e be ween ep esen a i es o se en
di e en phyla (16 species). We used a smalle numbe o emission e-
gions o in e es ( h ee egions), when compa ed wi h hose in li e a u e,
bu we employed mul iple wa eleng h exci a ions ( i e LEDs) along wi h
a pho odiode equipped wi h op ical il e s o de ec he luo escence
signals, enabling a compac and ully po able de ice o classi ying
mic oalgae. Addi ionally, ou model achie es accu acy le els (97 %
accu acy o 7 phyla and 92 % accu acy o 14 axonomic o de s) simila
o hose epo ed in p e ious wo ks using simila pigmen -based ech-
niques wi h ML [7,35]. Ne e heless, he he ein epo ed model
ad an age he added ea u e o a compac and po able de ice, sui able
o use beyond he labo a o y en i onmen . As a as he au ho s’
knowledge is conce ned, he e a e no simila sys ems in he li e a u e
ha use luo ome y and ML in a compac and po able de ice o
phy oplank on axonomic classi ica ion.
The u u e in eg a ion o his echnology wi h speci ic so ing and
concen a ion o mic oalgae o in e es could enable a highe accu acy
o he axonomic disc imina ion o phy oplank on, especially in
seawa e samples wi h high-di e si y phy oplank on communi ies. This
could be add essed h ough s a egic in eg a ion wi h a spi al mic o-
channel de ice ecen ly de eloped by ou g oup [36] as a p e-s ep
p ocess, which showcased he e ec i eness o ine ial mic o luidics in
so ing, isola ing, and concen a ing mic oalgae o in e es based on
size. The isola ion o mic oalgae be o e examina ion na ows down he
analysis o a mo e con ined ange o species. Addi ionally, he abili y o
enhance cell concen a ion p io o measu emen s is aluable o inc ease
he luo escence signal. Also no ewo hy is he capabili y o isola e
ha m ul species and signi ican ly concen a e hem inc easing he
chance o ea ly de ec ion o oxic species du ing he ea ly phase o
bloom de elopmen .
Au onomous mic oalgae iden i ica ion and in si u moni o ing will
allow a massi ica ion o da a collec ion which is a key ac o in moni-
o ing p og ams and can complemen in o ma ion om sa elli e images
and modelling echniques [20,37–40]. While sa elli e images o e
la ge-scale co e age and can cap u e spa ial pa e ns, in si u de ec ion
ocuses on local-scale de ails and p o ides sea- u h da a o alida ion
and calib a ion o sa elli e obse a ions and models. By au onomously
iden i ying mic oalgae in hei na u al en i onmen , mo e p ecise in-
sigh s in o he composi ion and abundance o mic oalgae species in
speci ic loca ions a e gained. Addi ionally, eal- ime moni o ing capa-
bili ies enable o swi ly de ec and espond o sudden changes in
mic oalgae popula ions. This in eg a ed app oach would likely enhance
he e ec i eness and e iciency o wa e moni o ing and managemen ,
ul ima ely leading o imp o ed en i onmen al p o ec ion and op imized
esou ce u iliza ion.
4. Conclusions
The use o spec o luo ome ic echniques coupled wi h ML capa-
bili ies demons a ed o be a p omising ool o as and accu a e
disc imina ion o phy oplank on species. This wo k con i ms he po-
en ial o his app oach as a obus , low-cos , and po able de ice
implemen ed me hod o apid axonomic classi ica ion o mic oalgae
g owing in monocul u es, wi h up o 97 % accu acy a he phylum le el
and 92 % accu acy a he o de le el. Fu u e esea ch will p io i ize
minia u iza ion and au oma ion o he de ice o in si u es ing while
Fig. 7. Tes da a classi ica ion ma ices o XGBoos a he o de le el.
V. Magalh˜
aes e al.
Senso s and Ac ua o s: B. Chemical 423 (2025) 136819
9
