Indonesian
J
our
nal
of
Electrical
Engineering
and
Computer
Science
V
ol.
21,
No.
3,
March
2021,
pp.
1634
1647
ISSN:
2502-4752,
DOI:
10.11591/ijeecs.v21i3.pp1634-1647
r
1634
Remote
sensing
data
dri
v
en
bathing
water
quality
assessment
using
sentinel-3
Antonia
Senta,
Ljiljana
ˇ
Seri
´
c
F
aculty
of
Electrical
Engineering,
Mechanical
Engineering
and
Na
v
al
Architecture,
Uni
v
ersity
of
Split,
R.
Bo
ˇ
sk
o
vi
´
ca
32,
Split
21000,
Croatia
Article
Inf
o
Article
history:
Recei
v
ed
Oct
2,
2020
Re
vised
Dec
2,
2020
Accepted
Dec
23,
2020
K
eyw
ords:
Bathing
w
ater
quality
DT
KNN
Machine
learning
Remote
sensing
Sentinel-3
ABSTRA
CT
In
this
paper
we
are
in
v
estig
ating
the
possibility
of
usage
of
remote
sensing
satellite
data,
more
precisely
sentinel-3
OLCI
and
SLSTR
data,
for
assessment
of
ba
thing
w
ater
quality
.
In
this
resea
rch
we
used
data
dri
v
en
approach
and
analysis
of
data
in
order
to
pinpoint
aspects
of
remote
sensing
data
that
can
be
useful
for
bathing
w
ater
quality
assessment.
F
or
this
purpose
we
collected
satellite
images
for
period
from
start
of
June
till
end
of
September
of
2019
and
results
of
in-situ
measurement
for
the
same
period
.
Results
of
in-situ
measurement
were
correlated
with
sat
ellite
images
bands
and
analyzed.
W
e
propose
a
simple
method
for
rapid
assessment
of
possible
deterioration
of
bathing
w
ater
quality
to
be
used
by
public
health
authorities
for
better
planning
of
in
situ
measurements.
Results
of
im
plementation
of
predicti
v
e
models
based
on
k-nearest
neighbour
(KNN)
and
decision
tree
(DT)
are
described.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
Antonia
Senta
F
aculty
of
Electrical
Engineering
Mechanical
Engineering
and
Na
v
al
Architecture
Uni
v
ersity
of
Split,
R.
Bosk
o
vica
32,
Split
21000,
Croatia
Email:
Antonia.Senta.00@fesb
.hr
1.
INTR
ODUCTION
Bathing
w
ater
quality
has
man
y
health
implications
[1,
2].
Assessments
on
sea
bathing
w
ater
quality
are
determined
on
the
basis
of
criteria
defined
by
re
gulation
on
sea
bathing
w
ater
quality
(OG
73/08)
and
the
EU
directi
v
e
on
management
of
bathing
w
ater
quality
No
2006/7/EC.
T
raditionally
,
bathing
w
ater
quality
is
measured
and
published
by
public
health
authorities.
Ho
we
v
er
,
although
ef
fort
is
m
ade
to
pro
vide
timely
information,
information
about
bathing
w
ater
quality
is
often
not
published
on
time
and
does
not
reach
the
public
on
time
to
pre
v
ent
bathing
in
polluted
w
ater
.
As
an
e
xample,
if
a
measurement
for
one
location
sho
ws
good
w
ater
quality
the
ne
xt
measurement
is
scheduled
in
15
days.
State
of
w
ater
quality
in
a
gi
v
en
moment
is
result
of
pre
vious
nearby
acti
vities
and
e
v
ents
and
an
y
deterioration
of
w
ater
quality
that
happens
in
that
period
will
not
be
detected.
In
this
paper
we
are
in
v
estig
ating
if
it
is
possible
to
mak
e
assessment
of
the
bathing
w
ater
quality
based
on
v
alues
of
indi
vidual
bands
or
combination
of
bands
from
collected
images.
Collected
images
we
used
in
this
study
are
pro
vided
by
satellite
sentinel-3
using
ocean
and
land
colour
instrument
(OLCI)
and
sea
and
land
surf
ace
temperature
instrument
(SLSTR).
These
images
are
a
v
ailable
for
scientific
and
commercial
e
xploitation.
In
order
to
e
v
aluate
and
compare
this
collected
data,
we
implemented
k-near
est
neighbors
(KNN)
and
decision
tree
algorithms
(DT)
and
applied
them
on
in
situ
measurements
data
on
sea
bathing
w
ater
quality
J
ournal
homepage:
http://ijeecs.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1635
at
the
beaches
in
the
Republic
of
Croatia.
This
measurements
data
is
public
and
published
by
Institute
of
Oceanograph
y
and
Fisheries
from
Split,
Croatia
[3].
Responding
quickly
to
e
v
ents
that
pose
a
risk
to
human
health
is
a
remote
sensing
mission,
so
we
propose
a
method
and
a
softw
are
system
to
be
used
by
authorities
to
mak
e
rapid
assessm
ent
of
bat
hing
w
ater
quality
and
detect
possible
deterioration
of
bathing
w
ater
quality
to
schedule
additional
in
situ
measurements
when
necessary
.
This
tool
w
ould
be
be
n
e
ficial
for
public
health,
since
deterioration
of
bathing
w
ater
quality
w
ould
be
detected
more
quickly
than
ha
ving
only
15
days
scheduled
measurements.
All
of
these
research
and
implemented
algorithms
lead
us
to
the
first
prototype
of
softw
are
system
which
wi
ll
be
used
for
classifying
bathing
w
ater
quality
in
the
area
in
the
Bay
of
Ka
ˇ
stela
and
the
Bra
ˇ
c
channel.
2.
RELA
TED
W
ORK
Based
on
directi
v
e
2006/7/EC
of
the
European
parliament
and
of
the
Council,
each
member
state
of
the
European
Union
need
to
control
coastal
bathing
w
ater
quality
by
measuring
the
concentration
of
Escherichia
Coli,
Intestinal
Enterococci,
T
otal
Coliforms
and
F
aecal
Coliforms
.
The
results
of
the
monitoring
of
bathing
w
ater
quality
are
disseminated
to
the
public
via
web
page
[3,
4].
In
recent
years
data
driven
pr
ediction
has
found
its
application
in
man
y
domains.
Model
driven
appr
oac
h
requires
h
ypothesis
of
ph
ysical
bac
kg
r
ou
nd
of
phenomenon
and
modeled
e
v
ent,
while
data
driven
appr
oac
h
uses
data
and
mathematical
operations
on
lar
ge
set
of
data
to
find
la
ws
and
correlat
ions
between
v
ariables
using
statistics
and
filtering.
Aside
from
in
situ
measurements,
there
is
increasing
number
of
the
studies
using
remote
sensing
data
and
techniques
for
assessment
of
bathing
w
ater
quality
[5].
Their
focus
is
on
optical
properties
as
well
as
chloroph
yll
[6],
turbidity
,
total
suspended
matters
[7]
and
dissolv
ed
or
g
anic
matters
[8].
Dif
ferent
sensors
onboard
satellites
or
aircraft
platforms
pro
vide
an
opportunit
y
to
choose
the
most
adequate
one
for
the
analyzing
bathing
w
ater
quality
.
In
[9]
authors
used
Landsat-8
thermal
images
for
determination
of
bathing
w
ater
quality
.
Beside
remote
sensing
techniques,
there
is
a
wide
use
of
h
yperspectral
images
from
drone
based
coastal
w
ater
quality
monitoring.
Specifically
,
h
yperspectral
monitoring
can
be
used
for
measuring
turbidity
,
colored
dissolv
ed
or
g
anic
matter
(CDOM)
and
ph
ytoplankton
us
ed
for
characterisation
of
the
spectra
of
bacteria
responsible
for
w
ater
quality
[10].
In
this
paper
we
decided
to
use
Sentinel-3
data
as
input
data
for
assessment
of
bathing
w
ater
quality
on
a
case
study
area
in
the
Bay
of
Ka
ˇ
stela
and
the
Bra
ˇ
c
channel.
3.
MA
TERIALS
AND
METHODS
Thanks
to
the
go
v
ernments
programs,
such
as
Copernicus
and
N
ASA,
there
is
a
v
ast
amount
of
freely
a
v
ailable
data
which
originates
from
a
remote
sensing.
Remote
sensing
data
comes
from
v
arious
sensors
mounted
on
satellites.
The
data
pro
vided
by
sensors
is
characterized
by
w
a
v
elengths
(bands
or
channels),
time
resolution
and
spatial
resolution.
Usually
,
a
sensors
measures
se
v
eral
bands
of
w
a
v
elength
in
a
single
capturing.
Specific
bands
or
a
combination
of
bands
v
alues
can
be
used
for
detecting
dif
ferent
phenomenon
on
the
Earth
surf
ace.
As
sho
wn
in
pre
vious
analysis,
ef
forts
ha
v
e
been
made
to
find
correct
equations
to
measure
features
such
as
surf
ace
temperature,
salinity
and
chloroph
yll.
In
this
paper
we
in
v
estig
ate
if
we
can
identify
features
of
v
arious
satellite
bands
that
can
i
ndicate
changes
in
bathing
w
ater
quality
.
The
whole
w
ork
is
strictly
data
dri
v
en
-
we
e
xamined
changes
in
features
of
remote
sensing
signals
in
case
of
bad
b
a
thing
w
ater
quality
.
W
e
use
data
obtained
from
satellite
sensors
monitoring
both
land
and
sea
and
associa
te
their
v
alues
to
measured
w
ater
quality
in
search
for
indicators
that
can
detect
or
predict
incidents
of
lo
wer
bathing
w
ater
quality
.
The
research
question
of
this
study
can
be
formulated
as.
Is
it
possible
to
pinpoint
rele
v
ant
remote
sensing
bands
for
assessment
of
bathing
w
ater
quality
using
data
dri
v
en
approach?
T
o
answer
this
question
we
selected
a
study
area
and
appropriate
data
and
performed
a
case
study
to
demonstrate
the
data
dri
v
en
assessment
of
bathing
w
ater
quality
using
data
obtained
from
Sentinel-3
satellite
and
discussed
results.
3.1.
Study
ar
ea
The
study
area
is
located
in
the
Bay
of
Ka
ˇ
stela
and
the
Bra
ˇ
c
Channel,
as
sho
wn
in
Figure
1.
Both
locations
are
located
near
the
to
wn
of
Split,
which
is
the
second
lar
gest
city
in
the
Republic
of
Croatia.
The
main
moti
v
ation
for
using
this
study
area
w
as
the
e
xceedance
of
the
limit
v
alues
in
summer
of
2019
defined
by
Re
gulation
on
Sea
bathing
w
ater
quality
(OG73/08)
for
the
microbiological
indicator
Escerichia
Coli
and
Enterococci.
Ha
ving
in
mind
these
in
situ
results,
we
were
encouraged
to
use
remote
sensing
techniques,
Remote
sensing
data
driven
bathing
water
quality
assessment
using
sentinel-3
(Antonia
Senta)
Evaluation Warning : The document was created with Spire.PDF for Python.
1636
r
ISSN:
2502-4752
machine
learning
algorithms
and
data
mining,
in
ef
fort
to
relate
in
situ
data
and
data
obtained
from
satellite
image
processing.
Figure
1.
The
satellite
image
(Sentinel-2
L2A)
of
study
area
3.2.
Data
sour
ces
In
this
study
we
used
images
related
to
Sentinel-3
SLSTR
and
OLCI
instrument.
The
ar
gumentation
for
using
this
specific
satellite
is
that
it
pro
vides
the
most
suitable
spatial
and
temporal
resolution
for
the
selected
area
and
the
time
period
of
the
study
.
Apart
from
spatial
and
temporal
resolution,
the
a
v
ailability
of
images
for
all
bands
via
the
API
is
another
reason
wh
y
we
used
the
Sentinel-3
satellite
for
this
study
.
F
or
comparison,
data
originating
from
moderate
resolution
imaging
spectroradiometer
(MODIS)
consists
of
tw
o
satellites,
terra
and
aqua,
has
good
time
resolution,
same
as
Sentinel-3
data,
b
ut
w
orse
spatial
resolution
for
majority
of
bands
so
it
w
ouldn’
t
be
a
good
choice
for
collecting
and
anal
y
z
ing
data
for
a
short
time
period.
T
able
1
sho
ws
a
comparison
of
the
spatial
and
temporal
resolution
for
the
Sentinel-2,
Sentinel-3,
Landsat-8
and
MODIS
satellites
data
and
sho
ws
that
Sentinel-3
is
the
most
suitable
for
our
research.
T
able
1.
Comparison
of
satellite
spatial
and
temporal
resolution
Satellite
T
emporal
Resolution
Spatial
Resolution
Reference
Sentinel-2
<
=
5
days
10m,
20m,
and
60m,
depending
on
the
w
a
v
elength
[11]
Sentinel-3
<
=
2
days
SLSTR:
500m
for
bands
S1
to
S6,
1km
for
bands
S7
to
S9,
F1,
and
F2
[12]
OLCI:
300m
[13]
Landsat-8
<
=
8
days
130m,
thermal
(100m
resampled
to
30m)
and
panchromatic
(15m)
bands
[14]
MODIS
<
=
2
days
250m
(bands
1-2),
500m
(bands
3-7),
1000m
(bands
8-36)
[15]
Satellite
Sentinel-3
is
de
v
eloped
by
european
space
agenc
y
(ESA)
as
a
part
of
copernicus
programme.
Sentinel-3
is
a
satellite
mission
consisting
of
tw
o
satellites,
Sentinel-3A
and
Sentinel-3B,
de
v
eloped
to
pro-
vide
accurate
and
timely
information
for
better
measurements
of
sea
surf
ace
topograph
y
,
sea
and
land
surf
ace
temperature
and
ocean
and
land
surf
ace
colour
.
F
our
main
instruments
on
Sentinel-3
spacecraft
are
Ocean
and
land
colour
instrument
(OLCI),
sea
and
land
Ssurf
ace
temperature
instrument
(SLSTR),
SAR
radar
altimeter
(SRAL)
and
micro
w
a
v
e
radiometer
(MWR)
[16].
The
moti
v
ation
to
use
sensors
rel
ated
both
to
the
sea
and
land
remote
sensing
is
found
in
the
f
act
that
the
locations
of
measurements
where
we
inspect
bathing
w
ater
quality
are
located
on
w
aters
in
near
vicinity
of
land.
Bathing
w
ater
can
be
considered
as
an
interf
ace
between
land
and
sea
and
its
current
state
is
influenced
by
both
sea
and
land
features.
Images
related
to
SLSTR
instrument
ha
v
e
11
bands
from
which
nine
are
spectral
bands
S1-S9
and
tw
o
additional
bands
are
optimised
for
fire
monitoring
F1
and
F2.
Spatial
resolution
of
images
obtained
by
this
instrument
is
500
m
for
bands
S1
to
S6,
1km
for
bands
S7
to
S9,
F1
and
F2.
Re
visit
time
for
SLSTR
is
less
than
or
equal
to
one
day
using
both
satellites.
OLCI
instrument
has
21
spectral
bands,
ranging
from
the
visible
to
the
near
-infra-red
(400
nm
to
1020
nm).
Spatial
resolution
for
this
instrument
is
300
m,
while
re
visit
time
is
tw
o
days
using
both
satellites
[16].
All
images
collected
from
these
satellites
are
tak
en
in
period
of
June
1
st
2019
to
October
1
st
2019,
with
cloud
co
v
er
percentage
less
than
20%
for
SLSTR
images.
Data
acquired
by
OLCI
and
SLSTR
sensors
are
a
v
ailable
on
Sentinel
Hub
EO
Bro
wser
[17].
Measurements
for
Sentinel-3
SLSTR
are
processed
at
Le
v
el
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
21,
No.
3,
March
2021
:
1634
–
1647
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1637
1B
and
represent
top
of
atmosphere
reflectance
or
brightness
temperature,
while
data
acquired
with
the
OLCI
sensor
are
at
full
resolution
and
processed
at
Le
v
el
1
reflectance.
Locations
of
interest
are
related
to
geograph-
ical
coordinates
of
in
situ
measurements
locations
for
which
we
ha
v
e
a
v
ailable
data
published
by
Institute
of
Oceanograph
y
and
Fisheries
from
Split,
Republic
of
Croatia
[18].
The
band
v
alues
for
each
location
of
the
study
area
were
e
xtracted
from
each
image.
W
eb
scrapping,
data
filtering,
consolidation
and
all
data
manipula-
tion
w
as
automated
in
Python
[19]
scripts.
3.3.
Methodology
The
methodology
consists
of
se
v
eral
phases
required
to
obtain
results,
and
in
this
section
we
will
describe:
data
collection
and
consolidation,
implementation
of
machine
learning
techniques
for
assessment
and
prediction
of
bathing
w
ater
quality
.
3.3.1.
Data
collection
and
consolidation
The
Sentinel-3
satellite,
due
to
its
short
re
visit
time,
is
ideal
for
e
xtracting
data
that
is
time-conti
nuous.
This
satellite
sensors
also
ha
v
e
good
spatial
resolution
suitable
for
a
small
study
area.
Acquisition
of
bathing
w
ater
quality
data
is
done
by
web
scrapping
from
the
web
site
of
Institute
of
Oceanograph
y
and
Fisheries
from
Split,
Croatia.
The
quality
of
bathing
w
ater
at
a
particular
location
can
ha
v
e
one
of
four
ratings:
e
xcellent
,
good
,
satisfactory
,
and
unsatisfactory
.
In
situ
measurement
is
performed
e
v
ery
15
days
unless
the
measurement
sho
ws
pollution.
In
case
of
lo
wer
quality
,
measurement
is
done
e
v
ery
day
until
the
results
are
impro
v
ed.
Due
to
the
v
ery
small
number
of
dif
ferent
ratings
of
samples
for
indi
vidual
locations,
we
replaced
these
four
classes
of
ratings
with
tw
o
ratings,
in
a
manner
that
e
xcellent
rating
is
denoted
as
good
,
and
all
other
ratings
are
considered
suboptimal
and
denoted
as
notGood
.
Remote
sensing
data
w
as
also
retrie
v
ed
using
a
web
API
from
the
Sentinel
Hub
EO
Bro
ws
er
[17].
Images
were
collected
in
.ti
f
f
format
georeferenced
in
the
WGS
84
coordinate
s
ystem
(EPSG:
4326).
Collected
images
co
v
erage
corresponds
to
study
area
and
selected
time
period
-
from
June
1
st
2019
till
October
1
st
2019.
The
v
alues
of
each
band
for
the
observ
ed
tim
e
period
were
e
xtracted
from
image
related
to
the
location
(latitude
and
longitude)
of
the
in
situ
measurement.
W
e
constructed
data
set
where
each
measurement
represent
a
sample
of
data.
Sample
is
described
by
features
of
in
situ
measurement
and
v
alues
of
each
band
for
that
day
for
tw
o
instruments
of
Sentinel-3
satellite.
Each
sample
is
described
with
a
set
of
features:
the
measurement
date,
the
bathing
w
ater
quality
score,
and
the
band
v
alue
for
the
observ
ed
measurement.
The
T
able
2
sho
ws
the
number
of
samples
with
respect
to
bathing
w
ater
quality
e
v
aluation
and
used
instruments
of
Sentinel-3
satellite.
In
this
w
ay
we
associated
each
measurement
with
a
series
of
v
alues
of
satellite
sensor
bands
[20].
T
able
2.
The
number
of
consolidated
data
samples
in
period
of
1st
June
2019
to
1st
October
2019
Ev
aluation
SLSTR
OLCI
good
865
475
notGood
119
32
T
otal
984
507
3.3.2.
Data
analysis
and
pr
epr
ocessing
In
this
section
we
will
present
beha
viour
of
all
bands,
separately
for
SLSTR
and
OLCI
instrum
ents.
W
e
used
all
a
v
ailable
data
collected
from
satellite
images
consolidated
with
in
situ
measurements,
so
we
could
get
tw
o
classes
of
bathing
w
ater
quality
,
where
one
is
classified
as
good
and
other
is
classified
as
notGood
bathing
w
ater
quality
.
Figures
2
and
3
present
graphs
of
data
distrib
ution
for
instruments
SLSTR
and
OLCI,
corresponding
bands
v
alues
and
bathing
w
ater
quality
scores.
SLSTR
data
distrib
ution
sho
ws
that
distrib
ution
of
v
alues
for
good
and
notGood
bathing
w
ater
quali
ty
are
most
distincti
v
e
for
bands
S5
and
S6.
Similarly
OLCI
instruments
bands
v
alue
distrib
ution
graph
sho
ws
the
most
distincti
v
e
v
alues
for
bands
B7,
B8,
B9,
B10
and
B11.
After
analysis
of
the
distrib
ution
of
all
bands
v
alues
in
order
to
select
bands
appropriate
for
clas
sifi-
cation
problem,
we
decided
to
analyse
bands
beha
viour
in
time
domain
for
tw
o
locations,
for
which
we
kne
w
that
o
v
erall
rating
w
as
notGood
opposed
to
those
whose
o
v
erall
rating
w
as
good
.
T
o
perform
further
analysis
we
singled
out
tw
o
locations:
one
located
at
Kamp
in
Ka
ˇ
stel
Gomilica
that
w
as
rated
with
an
o
v
erall
rating
as
notGood
and
the
other
located
at
Zv
on
ˇ
cac
in
Split
with
o
v
erall
rating
as
good
.
W
e
collected
v
alues
from
Remote
sensing
data
driven
bathing
water
quality
assessment
using
sentinel-3
(Antonia
Senta)
Evaluation Warning : The document was created with Spire.PDF for Python.
1638
r
ISSN:
2502-4752
SLSTR
and
OLCI
satellite
instruments
bands
that
c
orrespond
to
selected
tw
o
location
for
the
continuous
time
period
from
July
1
st
2019
to
August
31
st
2019.
Based
on
the
e
xtracted
data,
we
created
time
graphs
of
band
v
alues
for
each
location
and
tried
to
isolate
bands
whose
v
alues
de
viated
the
most
for
locations
with
good
and
notGood
bathing
w
ater
quality
.
The
time
series
graphs
for
all
bands
for
both
locations
are
sho
wn
belo
w
,
one
for
SLSTR
instrument,
as
sho
wn
in
Figure
4
and
the
other
for
OLCI
instrument,
as
sho
wn
in
Figure
5.
The
SLSTR
graph
in
Figure
4
sho
ws
the
inconsistenc
y
in
the
bands
v
alues
during
13
July
2019
to
16
July
2019
and
23
August
2019
to
27
August
2019,
especially
for
bands
S5
and
S6.
Based
on
the
band
v
alue
graphs
and
in
situ
data
measurements
of
bathing
w
ater
quality
,
we
can
notice
that
inconsistenc
y
in
the
data
of
band
v
alues
usually
occurred
during
the
time,
before,
or
after
the
determined
bathing
w
ater
pollution.
In
Figure
5
we
can
noti
ce
the
change
of
OLCI
bands
v
alues
for
the
period
from
August
20,
2019
till
September
7,
2019.
A
sudden
change
in
bands
v
alues
can
be
read
for
August
30,
2019,
where
the
peak
of
the
line
representing
notGood
bathing
w
ater
quality
is
higher
than
the
line
representing
good
bathing
w
ater
quality
.
If
we
pay
attention
to
the
pre
vious
v
alues
of
these
tw
o
lines
before
the
isolated
peak,
we
can
conclude
that
up
to
that
point
the
band
v
alues
were
follo
wing
the
same
trend
in
periods
when
both
locations
had
satisf
actory
measurements.
Based
on
these
observ
ations,
we
can
conclude
that
our
analysis
for
all
locations
and
for
tw
o
singled
out
locations
picks
out
the
same
bands
that
ha
v
e
de
viation
in
v
alues
between
good
and
notGood
measurements.
W
e
choose
bands
S5
and
S6
for
SLSTR
and
B7,
B8,
B9,
B10,
B11
for
OLCI
instrument
for
further
modeling
of
more
accurate
algorithms
for
prediction
of
bathing
w
ater
quality
.
This
approach
aims
to
w
ard
de
v
elopment
of
a
softw
are
system
for
remote
sensing
screening
of
bathing
w
ater
quality
prior
to
in
situ
measurements.
Such
a
system
can
be
used
as
a
tool
for
scheduling
in
situ
measurements,
focused
on
the
locations
where
deterioration
of
bathing
w
ater
quality
is
suspected.
Figure
2.
SLSTR-data
distrib
ution
of
all
bands
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
21,
No.
3,
March
2021
:
1634
–
1647
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1639
Figure
3.
OLCI-data
distrib
ution
of
all
band
Remote
sensing
data
driven
bathing
water
quality
assessment
using
sentinel-3
(Antonia
Senta)
Evaluation Warning : The document was created with Spire.PDF for Python.
1640
r
ISSN:
2502-4752
Figure
4.
T
ime
series
graph
of
all
bands
of
SLSTR
instrument
for
tw
o
selected
locations
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
21,
No.
3,
March
2021
:
1634
–
1647
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1641
Figure
5.
T
ime
series
graph
of
all
bands
of
OLCI
instrument
for
tw
o
selected
locations
3.3.3.
Implementation
of
assessment
and
pr
ediction
In
this
paper
we
implemented
tw
o
machine
learning
algorithms.
K-nearest
neighbors
and
deci
sion
tree
classification
as
a
tools
for
predicti
v
e
modeling.
Both
of
these
al
gorithms
are
good
recognition
techniques
and
are
widely
used
for
analysis
and
predictions
in
remote
sensing
data
sets
[21].
K-nearest
neighbors
(
KNN
)
algorithm
[22]
is
an
ef
ficient
lazy
learning
algorithm
widely
applied
in
pattern
recognition
and
data
mining
in
machine
learning
classification.
It
is
analytically
tract
able
and
highly
adapti
v
e
to
local
information.
KNN
algorithm
uses
the
closest
data
points
for
estimation.
F
or
each
data
point
the
algorithm
checks
ag
ainst
the
training
table
for
the
KNN.
Since
each
data
point
is
independent
of
the
others,
the
e
x
ecution
of
search
and
score
can
be
conducte
d
in
parallel.
Samples
for
training
are
stored
in
an
n-dimensional
space.
When
an
unkno
wn
test
label
is
gi
v
en,
the
KNN
classifier
searches
these
samples
which
are
close
st
to
Remote
sensing
data
driven
bathing
water
quality
assessment
using
sentinel-3
(Antonia
Senta)
Evaluation Warning : The document was created with Spire.PDF for Python.
1642
r
ISSN:
2502-4752
the
unkno
wn
sample.
Closeness
is
in
this
paper
defined
in
terms
of
Euclidean
distance.
Euclidean
distance
closeness
w
as
calculated
on
the
basis
of
the
band
v
alues
and
not
geographical
closeness.
Decision
tree
(DT)
classification
algorithm
[23]
creates
a
tree
model
from
the
data.
Model
is
created
by
using
v
alues
of
only
one
attrib
ute
at
a
time.
T
ree
structure
is
made
of
r
oo
t
(formed
from
all
of
the
data),
internal
(splits)
and
leaf
nodes.
Each
node
in
a
decision
tree
has
only
one
parent
node
and
tw
o
or
more
descendant
nodes.
The
algorithm
sorts
the
data
set
on
the
attrib
ute’
s
v
al
ue
and
then
looks
for
re
gions
in
the
data
set
that
clearly
contain
only
one
class
and
mark
those
re
gions
as
lea
v
es.
F
or
the
rem
aining
re
gions
that
ha
v
e
more
than
one
classes,
the
algorithm
chooses
another
attrib
ute
and
continues
the
branching
process
with
only
the
number
of
instances
in
those
re
gions,
until
it
produces
all
lea
v
es
or
there
is
no
attrib
ute
t
hat
can
be
used
to
produce
one
or
more
lea
v
es
in
the
conflicted
re
gions.
These
tw
o
algorithms
were
first
applied
to
data
set
consisting
of
all
bands
of
OLCI
and
SLSTR
in-
strument
and
results
are
discussed.
After
the
discussion,
we
selected
distinguished
bands
v
alues
based
on
data
analysis
of
all
bands
for
all
locations
and
tw
o
singled
out
locations
described
in
section
3.3.2.
The
results
of
the
applied
algorithms
will
be
depicted
belo
w
.
4.
RESUL
TS
AND
DISCUSSION
In
this
section
the
analysis
performed
on
prepared
data
set
will
be
presented
and
the
results
discussed.
The
discussion
is
aimed
to
w
ards
ar
gumentation
whether
selection
of
smal
ler
subset
of
bands
can
be
more
useful
for
assessment
and
prediction
of
bathing
w
ater
quality
.
4.1.
Pr
edicti
v
e
models
used
on
all
SLSTR
bands
The
data
set
constructed
from
the
SLSTR
instrum
ent,
ra
w
bands
images
are
used
as
an
input
set
for
K-nearest
neighbor
and
Decision
tree
algorithms.
F
or
each
algorithm,
obtained
results
sho
w
influence
of
the
indi
vidual
bands
of
the
SLSTR
instrument
in
decision
if
sample
will
be
classified
as
good
or
notGood
.
Since
the
data
set
we
used
for
training
the
models
is
imbalanced-selected
case
study
area
has
much
lar
ger
number
of
good
measurements
then
notGoot
we
trained
models
with
50%
of
our
data
set,
and
used
remaining
50%
for
e
v
aluation
of
results.
W
e
created
a
confusion
matrix
for
each
of
the
algorithms
(KNN
and
DT)
which
sho
ws
the
performance
of
the
classification
model
on
a
set
of
test
data,
for
which
the
true
v
alues
are
kno
wn.
The
results
of
the
confusion
matrix
of
both
algorithms
are
sho
wn
on
Figure
6
and
7.
Both
of
these
algorithms
are
highly
accurate,
b
ut
since
data
set
is
imbalanced,
model
can
f
ail
to
identify
ne
g
ati
v
e
labels
[24].
In
our
study
,
Decision
tree
algorithm
produced
better
and
more
precise
results
than
K-nearest
neighbor
algorithm.
The
reason
for
this
is
that
decision
tree
algorithm
identifies
man
y
more
true
positi
v
e
v
alues
than
the
K-nearest
neighbor
algorithm
does,
so
e
v
en
if
accurac
y
of
decision
tree
algorithm
is
smaller
than
accurac
y
of
K-nearest
neighbor
algorithm,
true
positi
v
e
and
f
alse
positi
v
e
v
alues
are
in
better
ratio
and
mak
e
accurac
y
more
reliable.
(a)
(b)
Figure
6.
KNN
algorithm
for
SLSTR
data,
(a)
Classification
report,
(b)
Normalized
confusion
matrix
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
21,
No.
3,
March
2021
:
1634
–
1647
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1643
(a)
(b)
Figure
7.
DT
algorithm
for
SLSTR
data,
(a)
Classification
report,
(b)
Normalized
confusion
matrix
4.2.
Pr
edicti
v
e
models
used
on
all
OLCI
bands
W
e
performed
the
same
methodology
on
data
from
OLCI
instrument
and
implemented
K-neares
t
neighbor
and
Decision
tree
algorithms.
Amount
of
SLSTR
data
is
almost
double
in
v
olume
compared
to
OLCI
data
because
of
the
OLCI
instrument’
s
lar
ger
re
visit
time.
Also,
OLCI
data
is
e
v
en
more
imbalanced,
where
data
classified
as
ha
ving
good
bathing
w
ater
quality
is
in
the
majority
compared
to
data
classified
as
ha
ving
notGood
bathing
w
ater
quality
.
The
results
of
the
applied
K-nearest
neighbor
and
Decision
tree
algorithms
are
sho
wn
on
Figure
8
and
9.
Confusion
matrix
for
both
algorithms
applied
to
OLCI
data
sho
ws
high
accurac
y
.
But
if
we
look
closely
,
we
will
notice
that
because
of
imbalanced
data
set
,
neither
of
these
tw
o
models
are
suitable
for
pre-
dicting
deterioration
of
bathing
w
ater
quality
based
only
on
OLCI
data.
T
rue
positi
v
e
v
alues
in
re
g
ards
to
true
ne
g
ati
v
e
v
alues
are
in
the
minority
,
so
our
accurac
y
is
not
reliable
parameter
,
although
it
has
a
high
v
alue.
Our
intuition
w
as
that
both
instruments
-
OLCI
typically
used
for
ocean
monitoring
and
SLSTR
typically
used
for
land
monitoring
should
be
used
for
bathing
w
ate
r
quality
assessment
since
specific
area
of
interest
is
located
near
interf
ace
between
land
and
sea.
(a)
(b)
Figure
8.
KNN
algorithm
for
OLCI
data,
(a)
Classification
report,
(b)
Normalized
confusion
matrix
Remote
sensing
data
driven
bathing
water
quality
assessment
using
sentinel-3
(Antonia
Senta)
Evaluation Warning : The document was created with Spire.PDF for Python.