Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
C
om
put
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4197
~
4203
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
8
i
6
.
pp
4197
-
42
03
4197
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Classific
atio
n
of
M
acron
utrient
D
eficien
cies in
M
aize
P
l
an
t
U
sing
M
ac
hin
e
L
earning
Le
ena
N
1
,
K.
K.
S
aju
2
1
Depa
rt
m
ent
of Elec
trica
l
and
E
l
ec
tron
ic
s E
ng
ineeri
ng,
NS
S c
olle
ge
of
Engi
n
ee
rin
g
,
Indi
a
2
Depa
rt
m
ent
of
Mec
hanica
l
Eng
i
nee
ring
,
Co
chi
n
Univer
sit
y
of
Sc
ie
nc
e and
T
ec
hn
olog
y
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
23
, 201
8
Re
vised
Ju
l
15
,
201
8
Accepte
d
Aug
8
, 2
01
8
Dete
c
ti
on
of
nu
tri
ti
on
al
def
i
ci
e
nci
es
in
pl
ant
s
is
vit
a
l
for
improving
cro
p
produc
ti
vi
t
y
.
T
i
m
ely
ide
n
ti
f
ic
a
t
ion
of
nutri
ent
def
ic
i
ency
thr
ough
visual
s
y
m
ptoms
in
th
e
pl
ant
s
ca
n
h
e
lp
far
m
ers
ta
ke
quic
k
cor
r
ec
t
iv
e
a
ct
ion
b
y
appr
opriate
nut
r
ie
nt
m
ana
g
eme
nt
st
rategi
es.
T
he
application
of
computer
vision
and
m
a
chi
ne
le
a
rning
te
chn
ique
s
offe
rs
new
prospec
ts
in
non
-
destruc
t
ive
fi
el
d
-
base
d
anal
y
s
is
for
nutri
ent
d
eficie
n
c
y
.
Color
an
d
shape
are
important
par
a
m
et
ers
in
fea
tu
re
ext
ra
ct
ion
.
In
thi
s
work,
t
wo
diffe
ren
t
t
ec
hn
ique
s
are
used
for
image
segm
ent
at
ion
and
feature
ex
t
rac
t
ion
to
gene
ra
te
two
dif
fer
ent
feature
s
e
ts
from
the
sam
e
image
sets.
Th
e
se
are
the
n
used
for
c
la
ss
if
ic
a
ti
on
using
di
ffe
ren
t
m
ac
hin
e
learni
ng
techn
ique
s.
The
expe
riment
al
re
sults
are
a
naly
z
ed
and
compare
d
in
te
rm
s
of
cl
assificat
ion
ac
cur
acy
to
find the
b
est al
gori
th
m
for
the t
wo f
e
at
ure
sets.
Ke
yw
or
d:
Deep l
ear
ning
with a
uto
encode
rs
Feat
ur
e
ex
tr
act
ion
K
n
earest
n
ei
ghbo
rs
Nu
t
riti
on
al
d
ef
ic
ie
ncy
Suppor
t
v
ect
or m
achines
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Leena
N,
Dep
a
rtm
ent o
f El
ect
rical
an
d
Ele
ct
ro
nics
E
nginee
rin
g,
NS
S
Colle
ge
of E
ng
i
neer
i
ng,
Pala
kk
a
d, Ke
ra
la
, Ind
ia
.
Em
a
il
:
le
ena.am
ul
ya
@g
m
ail.co
m
1.
INTROD
U
CTION
The
increa
sin
g
cost
of
c
rop
pro
du
ct
io
n
and
prob
le
m
s
of
en
vir
on
m
ental
po
ll
utio
n
caused
by
agroc
hem
ic
a
ls
cal
ls
fo
r
the
ne
ed
to
ap
ply
m
i
ner
al
fe
rtil
iz
ers
m
or
e
eff
ic
ie
nt
ly
.
Plants
require
nutrie
nts
f
or
their
healt
hy
gro
wth
an
d
de
velo
pme
nt.
T
hese
incl
ud
e
m
acro
nu
t
r
ie
nts
li
ke
nitr
ogen
,
phospho
r
us
a
nd
pota
ssi
um
as
wel
l
as
m
ic
ro
nutrie
nts
li
ke
ca
lc
iu
m
,
boron
e
tc
.
Nu
t
rient
de
fici
ency
sym
pt
om
s
if
le
ft
un
i
den
ti
fie
d
res
ul
ts
in
low
yi
el
d
le
ve
ls
and
pr
of
it
.
Nu
t
riti
on
al
de
ficie
ncies
ha
ve
char
act
erist
ic
sy
m
pto
m
s
that
are
exh
i
bited
in
the
form
of
col
or
an
d
s
hap
e
,
e
sp
eci
al
ly
on
l
ea
ves.
Thes
e
inclu
de
necrosi
s,
Chl
orosis,
die
-
back
an
d
oth
e
rs
.
Kno
wled
ge of
these sym
pto
m
s can
h
el
p us
ta
ke
c
orrecti
ve
a
ct
ion
to
r
est
or
e
the
plant t
o norm
al
sta
te
.
The
co
nventio
nal tec
hn
iq
ues for nutrient de
f
ic
ie
ncy d
ia
gnosi
s includ
e ext
ensive
la
borat
ory
te
sti
ng
o
f
so
il
or
pla
nt
ti
s
su
e
or
m
anu
al
insp
ect
io
n
by
f
arm
ers.
These
are
te
dious
an
d
tim
e
con
su
m
ing
.
T
he
ap
plica
ti
on
of
m
achine
le
arn
i
ng
te
ch
niques
us
in
g
im
a
ge
processi
ng
offer
s
a
prom
i
sing
so
l
utio
n
f
or
ide
ntific
at
ion
of
nu
t
rient d
efici
encies.
Se
ve
ral work
s h
ave b
e
en
re
porte
d
on
the
us
e
s
of
m
achine
le
a
rn
i
ng
algorit
hm
s
com
bin
ed
with
im
age
processin
g
te
ch
ni
qu
es
t
o
cl
assify
an
d
predict
abiotic
an
d
bi
otic
stresses
on
plants
.
Ta
ble
1
pro
vid
es
a
n
over
view
of
m
a
chine
le
ar
ning
te
chn
i
qu
es
c
om
bin
ed
with
i
m
age
process
i
ng
m
et
hods
to
detect
nu
t
rient
def
ic
ie
ncy sym
pto
m
s in
pla
nts.
Ther
e
a
re
al
so
sever
al
resea
r
ches
on
the
use
of
dig
it
al
i
m
age
proces
sin
g
com
bin
ed
with
m
achine
le
arn
in
g
f
or
de
te
ct
ion
of
dise
ases
in
pla
nts.
Howe
ver
fe
w
l
it
eratur
e
ai
m
s
to
cl
assify
a
nd
identify
nutrit
ion
a
l
def
ic
ie
ncies
in
plants.
The
re
hav
e
al
so
no
t
been
any
at
tem
pts
to
m
ak
e
a
co
m
par
at
ive
stud
y
of
f
eat
ur
e
extracti
on and m
achine learn
i
ng
tech
niques
. Th
is pa
per
aim
s to
m
ake a
com
par
at
ive stud
y of
the p
er
for
m
ance
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4197
-
4203
4198
of
featu
re
e
xtr
act
ion
te
c
hn
i
ques
a
nd
m
achine
le
ar
ning
al
gorithm
s
for
cl
a
ssific
at
ion
of
nutrie
nt
def
ic
ie
ncy
in
m
ai
ze plants.
Table
1.
O
verv
ie
w
of
Im
age P
ro
ces
sin
g
a
nd
Ma
chine
Le
ar
ni
ng
Tech
niques
for
De
fici
ency D
et
ect
ion
Ap
p
licatio
n
Featu
re
Extractio
n
Techn
iq
u
e
Machin
e lea
rnin
g
Algo
rith
m
Plan
t Species
Ref
erence
Iden
tif
icatio
n
Sp
ectra
l and
sh
ap
e
f
eatu
res
Su
p
p
o
rt
Vector M
achi
n
e
Rice
[
1
]
Detectio
n
Co
lo
r
Clu
sterin
g
Bean
,
p
ea a
n
d
y
ello
w lup
in
e
[
2
]
Detectio
n
Geo
m
etri
c
m
o
m
en
ts
Grou
n
d
n
u
t
[
3
]
Detectio
n
Textu
re
an
d
colo
r
f
eatu
res
Rice
[
4
]
Detectio
n
Co
lo
r
an
d
size
Artif
icial Neu
ral
N
etwo
rks
Lettuce
[
5
]
Qu
an
tif
icatio
n
Co
lo
r
Barley
[
6
]
Detectio
n
Sh
ap
e,
Textu
re
Fu
zzy
Oil pal
m
[
7
]
Clas
sif
icatio
n
Textu
re
f
eatu
res
Fu
zzy
k
-
n
ea
rest Cl
ass
if
ier
To
m
ato
[
8
]
Clas
sif
icatio
n
SIFT
,
Co
lo
r
an
d
sh
ap
e
f
eatu
res
Ran
d
o
m
For
est
Co
ff
ee
[
9
]
2.
RESEA
R
CH MET
HO
D
This
work
f
oc
us
es
on
the
re
cogniti
on
an
d
cl
assifi
cat
ion
of
nutrit
ion
al
def
ic
ie
ncies
i
n
m
a
iz
e
plant.
The
m
ajo
r
ste
ps
of
cl
assif
ic
at
ion
inclu
de
sel
ect
ion
of
trai
nin
g
sam
ples,
im
age
pr
ep
ro
ce
ssin
g,
featur
e
extracti
on, se
le
ct
ion
of su
it
a
ble cla
ssific
at
ion ap
proac
h,
a
nd
accuracy as
ses
sm
ent.
2.1.
Ima
ge Dat
a
S
et
Def
ic
ie
ncies
c
ause
d
by
m
acro
nutrie
nts
li
ke
Nitro
ge
n,
Pho
sp
ho
ru
s
an
d
P
otassium
as
w
el
l
as
norm
al
(not
af
fected
)
plants
a
re
in
ve
sti
gated.
100
s
a
m
ple
le
af
im
ages
for
the
4
cl
asses
we
re
use
d
in
the
st
udy.
75
i
m
ages
fo
r
use
d
for
cl
assifi
ca
ti
on
an
d
25
i
m
ages
we
r
e
us
e
d
for
detect
io
n.
Im
ages
giv
en
in
Figu
re
1
ar
e
the
sam
ples o
f
m
aize leaves s
uffe
rin
g
f
r
om
n
utriti
on
al
d
e
fici
en
ci
es.
Figure
1
.
I
m
ages of
m
ai
ze le
a
f
a)
H
eal
thy l
e
af
b)
Nitr
og
e
n defici
ent lea
f
c
)
P
hosph
orus d
efici
ency
d) Po
ta
ssi
um
d
efici
ency
2.2.
Ima
ge
Pre
proc
essing
a
n
d F
eature E
xtr
ac
tion
Sele
ct
ing
s
uitable
feat
ur
es
is
a
crit
ic
al
ste
p
f
or
su
cce
ssf
ully
i
m
ple
m
enting
i
m
age
cl
assifi
cat
ion
.
The
sy
m
pto
m
of
nutrie
nt
def
ic
ie
ncy
in
plant
l
eaves
e
xhibit
diff
e
re
nt
pro
pe
rtie
s
in
te
rm
s
of
c
olou
r,
s
ha
pe,
a
nd
te
xtu
re
.
T
he
first
ste
p
in
vo
l
ve
s
pre
-
processi
ng
the
im
age
to
im
pr
ove
the
qu
al
it
y
of
the
im
age
an
d
t
o
re
m
ov
e
distor
ti
on.
T
hi
s
is
fo
ll
ow
e
d
by
segm
entat
i
on
te
ch
niques
to
segm
ent
the
reg
io
n
of
interest
from
the
whole
i
m
age.
F
ur
the
r
featu
re
e
xtract
ion
is
do
ne
fro
m
the
segm
ented
reg
i
on
ha
vi
ng
the
de
fici
en
t
par
t o
f
the
le
a
f.
T
his
work
c
om
par
es
the
accuracy
of
cl
assifi
cat
ion
us
in
g
us
e
s
two
m
ai
n
feat
ur
e
e
xtracti
on
te
chn
iq
ues
.
Th
e
firs
t
te
chn
iq
ue
us
es
colo
ur
h
ist
og
ram
for
im
ag
e
segm
entat
ion
an
d
fea
tu
re
extracti
on
w
her
eas
t
he
al
te
rn
at
iv
e
m
et
ho
d uses
K
-
m
eans clusteri
ng for se
gm
ent
at
ion
a
nd text
ural
f
eat
ures
are
ex
tract
e
d.
MATLAB
2016
was
use
d
f
or
design
i
ng
the
al
gorithm
fo
r
im
age
processi
ng
a
nd
cl
assifi
cat
ion
.
F
or
histo
gr
am
based
i
m
age
segm
e
ntati
on
an
d
cl
assifi
cat
ion
,
Ga
bor
filt
er
was
us
e
d
for
resizi
ng
a
nd
no
ise
fi
lt
ering
of
im
ages
.
Th
e
i
m
ages
wer
e
then
tra
ns
f
orm
ed
from
the
RGB
to
HSV
(Hue,
Sat
ur
at
i
on
a
nd
Value
)
colo
ur
sp
ace
as
HSV
separ
at
es
t
he
c
olour
c
om
po
ne
nts
(
HS)
f
r
om
the
lum
inance
com
po
ne
nt
(
V)
a
nd
is
le
ss
se
nsi
ti
ve
to
il
lu
m
inati
on
changes
.
The
def
ic
ie
nt
par
t
of
t
he
le
aves
s
a
m
ples
was
se
gm
ented
ou
t
from
the
colour
i
m
age
.
The
c
olour
feat
ur
es
are
e
xtract
ed
f
r
om
HS
V
c
olor
s
pace
of
s
el
ect
ed
re
gion
base
d
on
histo
gr
am
analy
sis.
Each
i
m
age
add
e
d
t
o
the
c
ollec
ti
on
is
a
naly
sed
t
o
c
om
pu
te
a
c
olour
histo
gra
m
,
wh
ic
h
s
hows
the
pr
opor
t
ion
of
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Cl
as
sif
ic
ation
of M
acron
utrient Defi
ci
encies
in
M
aize Pl
an
t
U
sin
g
…
(
Lee
na N)
4199
pix
el
s
of
eac
h
col
our
within
the
im
age.
T
hu
s
as
a
resu
l
t
256
feat
ur
e
s
we
re
obta
ine
d
for
eac
h
im
age.
The
c
olour
h
ist
ogram
f
or each
i
m
age is the
n st
or
e
d
in
the
da
ta
base.
In
t
he
sec
ond
te
chn
iq
ue
K
-
m
eans
cl
us
te
ri
ng
us
in
g
Eucli
de
an
distance
as
the
m
ini
m
iz
ation
crit
er
i
a
was
us
e
d
f
or
s
egm
entat
ion
.
T
he
i
m
age
is
segm
ented
in
to
t
hr
ee
sub
-
featu
r
e
i
m
ages
with
t
hr
ee
dif
fer
e
nt
t
ype
of
Re
gion
of
I
nterest
(R
OI).
Th
e
Re
gion
of
I
nt
erest
co
rr
es
po
nd
i
ng
to
nutrie
nt
de
fici
ency
i
s
m
anu
al
ly
sel
ect
ed
from
the
segm
ented
i
m
ages.
The
Gr
ey
Level
Occ
urre
nce
Ma
trix
is
then
c
om
pu
te
d
f
ro
m
the
ROI
a
fter
conve
rsion f
r
om
RGB to g
ray scal
e. 1
3
feat
ur
es a
re calculat
ed
for
each i
m
age b
ased
on textur
e pro
pe
r
ti
es l
ike
Sk
e
wn
es
s,
Sta
nd
a
r
d
De
viati
on,
Ho
m
og
e
neity
,
Con
t
rast,
S
m
oo
thn
ess,
Co
rr
el
at
ion,
Kurt
os
is,
E
ne
rg
y,
E
ntr
op
y,
Me
an,
Var
ia
nc
e,
RM
S,
an
d
I
DM.
Fig
ure
3
sh
ows
t
he
res
ul
ts
of
K
-
m
eans
cl
us
te
rin
g
f
or
a
le
af
with
nitr
og
e
n
def
ic
ie
ncy.
2.3.
Ma
c
hine
Le
ar
ning
Te
c
hniq
ues
M
achine
le
a
rn
i
ng
com
pr
ise
s
of
a
set
of
co
m
pu
ta
ti
on
al
m
od
el
li
ng
te
c
hn
i
qu
e
s
that
ca
n
l
earn
patte
r
ns
from
data
and
pe
rfor
m
auto
m
at
ed
ta
sk
s
li
ke
ide
ntific
at
ion,
pr
e
dicti
on
or
cl
assifi
cat
ion.
Ma
c
hin
e
l
earn
i
ng
te
chn
iq
ues
a
re
widely
us
e
d
in
ap
plica
ti
on
l
ike
ha
ndwr
it
in
g
rec
ogniti
on,
natu
ral
la
ngua
ge
proce
ssin
g,
sp
eec
h
processi
ng,
c
onsu
m
er
data
pr
edict
ive
a
naly
sis,
dru
g
desi
gn,
disease
ti
ssu
e
cl
assifica
ti
on
in
m
edici
ne
[
10]
,[
11]
netw
ork
flo
w
cl
assifi
cat
ion
[
12
]
.
Seve
ral
m
achine
le
a
rn
i
ng
te
ch
niques
ha
ve
been
us
e
d
in
nutrie
nt
de
fici
ency
detect
ion i
n pl
ants
as
sho
wn in Ta
ble
1.
2.3.1.
ANN Clas
sifie
r
Ar
ti
fici
al
neur
al
network
s
,
usual
ly
cal
le
d
neu
ral
net
works
hav
e
em
erg
ed
as
an
i
m
po
rtant
too
l
for
cl
assifi
cat
ion
.
Neural
netw
orks
a
re
sim
plifi
ed
m
od
el
s
of
the
bio
lo
gical
nerv
ou
s
syst
e
m
wh
ic
h
c
ons
ist
s
of
highly
interco
nn
ect
e
d
net
work
of
a
la
rg
e
num
ber
of
pr
oc
essing
el
em
ents
cal
le
d
neurons
in
an
ar
chit
ect
ur
e
insp
ire
d
by
th
e
br
ai
n
.
They
hav
e
the
pote
ntial
to
cl
assif
y
diff
e
ren
t
f
orm
s
(p
at
te
rn
s)
of
a
r
bitrary
c
om
plex
input/
ou
t
pu
t
m
app
i
ngs
.
2.3.2.
SVM
Clas
sifie
r
Suppor
t
Vecto
r
Ma
chine
(SV
M)
is
con
si
dered
as
one
of
th
e
eff
ic
ie
nt
m
ac
hin
e
le
ar
ning
m
et
ho
d
that
is
dev
el
oped
on
the
basis
of
t
he
sta
ti
sti
cal
le
arn
in
g
t
he
or
y.
They
a
r
e
sp
eci
fical
ly
i
m
ple
m
ented
f
or
t
he
cl
assifi
cat
ion
a
nd r
e
gr
es
sio
n wit
h hig
h dim
e
ns
io
nal s
pace. The aim
o
f
the
SV
M cl
assifi
e
r
is to
fin
d
a
n o
pti
m
a
l
hype
rp
la
ne.
S
uppo
rt
vecto
r
m
achines
a
re
a
ve
ry
popula
r
m
e
thod
in
cl
assifi
cat
ion
of
im
ag
es
du
e
to
t
heir
good
gen
e
rali
zat
ion
c
apab
il
it
y eve
n wit
h
a
lim
it
ed
nu
m
ber
of trai
ning
dataset
s
.
2.3.3.
k
N
N Classifie
r
Anothe
r
sim
pl
e
and
v
e
ry
po
wer
f
ull
s
up
e
rvi
sed
m
achine
le
arn
i
ng
te
c
hn
i
qu
e w
idely
u
se
d
in
sta
ti
sti
cal
est
i
m
ation
an
d
patte
rn
recog
niti
on
is
th
e
k
Nea
rest
N
ei
ghbour
(
k
NN)
.
The
kNN
cl
a
ssifie
r
use
s
a
non
par
am
et
ric
and
instance
-
ba
sed
le
arn
i
ng
al
go
rithm
.
In
cl
a
ssific
at
ion
pr
ob
le
m
s,
the
k
-
near
e
st
nei
ghbo
ur
al
gorithm
f
inds
m
ajo
rity
vote
betwee
n
the
k
m
os
t
si
m
ilar
instance
s
to
a
gi
ven
“
unsee
n”
obse
rvat
ion.
Si
m
il
arity is de
fine
d
acc
ordin
g
to
a
distance
m
et
ric b
et
ween t
wo d
at
a
point
s.
2.3.4.
Deep
Ne
twork
s U
sin
g
Autoe
ncod
er
s
Re
centl
y,
the
area
of
deep
le
arn
in
g
is
at
tract
ing
wides
pread
inter
est
by
pro
du
ci
ng
rem
a
rk
able
researc
h
i
n
al
m
os
t
ever
y
as
pect
of
a
rtific
ia
l
intel
li
gen
ce
[
13
]
.
A
uto
e
ncode
rs
a
re
unsupe
rv
ise
d
m
achine
le
arn
in
g
te
ch
ni
qu
e
s
that
ap
ply
back
pr
op
a
ga
ti
on
al
gorith
m
.
It
trie
s
to
l
earn
the
i
den
ti
fy
functi
on
by
placi
ng
const
raints
on
the
netw
ork
.
Au
t
o
en
c
oders
can
be
sta
cke
d
one
ov
e
r
the
oth
e
r
to
f
orm
deep
netw
ork
s
cal
le
d
sta
cked
aut
o
e
ncode
rs.
Her
e
le
arn
in
g
is
done
by
trai
ni
ng
one
la
ye
r
at
a
ti
m
e.
A
sta
cke
d
auto
e
nc
od
e
r
c
an
be
us
e
d
as
a
cl
assi
fier
by
re
placi
ng
t
he
decode
r
la
ye
r
with
a
s
of
tm
ax
la
ye
r
to
cl
assify
the
f
eat
ur
es
e
xtract
ed
f
ro
m
the en
c
oder
lay
ers.
3.
RESU
LT
S
A
ND D
I
SCUS
S
ION
Feat
ur
e
e
xtract
ion
is
do
ne
usi
ng
the
te
c
hn
i
qu
e
s
in
sect
io
n
2.2
.
Fig
ur
e
2
sho
ws
the
diff
e
ren
t
ste
ps
involve
d
in
the
segm
entat
i
on
process
f
or
a
le
af
with
nitrogen
def
i
ci
ency
us
in
g
the
first
te
ch
nique
.
The
se
gm
ented
par
t
is
s
how
n
in
Fig
ur
e
2(f
).
T
he
c
olour
hi
stog
ram
is
plot
te
d
f
or
t
he
se
gm
ented
i
m
age
an
d
a
total
o
f
25
6
f
ea
tures
a
re
obta
ined
.
T
he first
f
e
at
ur
e set
has
a
d
im
ension
of
75x
256 f
or
t
he 75 i
m
ages.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4197
-
4203
4200
Figure
2
.
Proce
ssing o
f
a
leaf i
m
age w
it
h nit
r
og
e
n de
fici
encies a)
ori
gin
al
I
m
age b) Hu
e
im
age
c)
Value im
age d)
Sat
ur
at
io
n I
m
age e)
H
ue
m
ask f)
seg
m
ented regi
on
s
Fig
ure
3
sho
w
s
the
res
ults
of
cl
us
te
rin
g
to
ob
ta
in
t
he
de
fi
ci
ent
par
ts
of
t
he
im
age
us
in
g
the
sec
on
d
m
et
ho
d.
The
cl
us
te
re
d
im
age
with
de
fici
ent
par
t
is
s
how
n
in
Fig
ure
3(
b)
.
This
is
co
nvert
ed
to
gr
ey
scal
e
an
d
13
featu
res
a
r
e
ob
ta
i
ned
us
i
ng
G
rey
le
vel
Occurre
nce
m
at
rix.
The
s
econd
featu
re
set
has
a
dim
ensio
n
of
75
x
13.
Cl
assifi
cat
ion
was
done
us
i
ng
the
cl
assifi
ers
i
n
s
ect
ion
2.3
a
nd
the
pe
r
form
ances
of
t
he
cl
assi
fier
s
in
te
rm
s
of
acc
ur
acy
of
cl
assifi
cat
ion
wer
e
c
om
par
ed
.
The
trai
ning
a
nd
te
sti
ng
a
re
car
rie
d
ou
t
with
bo
t
h
the
featur
e
ex
t
racti
on m
et
ho
ds
gi
ven in
sect
ion
2.2.
Figure
3
.
Proce
ssing o
f
a
le
af i
m
age w
it
h nit
r
og
e
n de
fici
encies a)
ori
gin
al
I
m
age b)
Cl
us
te
red im
age w
it
h
def
ic
ie
nt
portio
n
a
fter c
ontras
t enh
a
ncem
ent
A
m
ulti
la
ye
r
back
pr
op
a
gatio
n
neural
netw
ork
was
ch
os
e
n
for
t
he
ANN
cl
assifi
e
r
.
T
he
nu
m
ber
of
nodes
in
the
i
nput
eq
uals
the
nu
m
ber
of
feat
ur
es
wh
il
e
the
nu
m
ber
of
no
de
s
in
the
ou
t
pu
t
is
4
cor
res
po
nd
i
n
g
to
the
three
de
fici
ency
cl
asses
and
one
nor
m
al
c
la
ss.
Num
ber
of
nodes
in
the
hidden
la
ye
r
is
ta
ken
as
10
an
d
sig
m
oid
act
ivati
on
f
unc
ti
on is u
se
d.
Scaled c
onjug
at
e
gr
a
di
ent b
ac
kpr
op
a
ga
ti
on
te
ch
nique
is u
sed for trainin
g.
An
overall
cl
assifi
cat
ion
acc
ur
acy
of
90.9
%
was
ob
ta
ine
d
with
histo
gr
a
m
based
feat
ure
extracti
on
te
chn
i
qu
e
(f
eat
ure
set
1)
wh
e
reas
an
ov
erall
accu
racy
of
81.2%
wa
s
obta
i
ne
d
for
featu
re
e
xtrac
ti
on
us
in
g
k
-
m
eans
segm
entat
ion
foll
ow
e
d by s
ha
pe
a
nd textu
re
featur
e
s (feat
ur
e set 2
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Cl
as
sif
ic
ation
of M
acron
utrient Defi
ci
encies
in
M
aize Pl
an
t
U
sin
g
…
(
Lee
na N)
4201
Fig
ur
e
4
s
ho
ws
the
grap
h
of
cl
assi
ficat
ion
acc
ur
acy
of
m
acro
nutrie
nt
def
ic
ie
ncies
a
f
fecti
ng
m
ai
ze
plant
us
i
ng
t
he
two
feat
ur
e
se
ts
and
an
SVM
cl
a
ssifie
r.
The
ef
fect
of
ke
rn
el
s
on
cl
assif
ic
at
ion
accu
rac
y
was
al
so
inv
est
i
gated.
F
ro
m
the
gr
a
ph,
it
is
observ
e
d
that
th
e
m
axi
m
u
m
c
l
assifi
cat
ion
ac
cur
acy
of
95.
6
%
has
occurre
d
with
a
li
near
ke
rn
el
for
featu
re
se
t
1
and
90%
f
or
featur
e
s
et
2.
Hen
ce
S
V
M
with
li
near
kernel
functi
on
giv
es
bette
r
cl
assifi
c
at
ion
acc
ur
acy
com
par
ed
t
o other ke
rn
el
s
.
Figure
4
.
Cl
assifi
cat
ion
e
ff
ic
ie
ncy f
or S
VM c
la
ssifie
r
with
di
ff
ere
nt
kernel
functi
ons
The
res
ults
obta
ined
us
i
ng
a
KNN
cl
assifi
e
r
is
show
n
i
n
F
igure
5
(a
)
a
nd
(b)
for
t
he
tw
o
diff
e
re
nt
featur
e
e
xtracti
on
te
ch
niques.
The
cl
assifi
ca
ti
on
resu
lt
i
n
t
erm
s
of
ef
fici
ency
wa
s
ob
ta
i
ned
with
num
ber
of
neig
hbours
vari
ed
from
2
to
5.
As
k
N
N
cl
assifi
cat
ion
is
ba
sed
on
m
easur
in
g
the
distan
ce
between
th
e
te
st
data
an
d
eac
h
of
t
he
trai
ni
ng
data,
the
c
hose
n
dista
nce
f
un
ct
ion
ca
n
af
fec
t
the
cl
assifi
cat
ion
acc
ur
acy
.
Hen
ce
eff
ect
of d
if
fe
r
ent d
ist
a
nce me
tric
s li
ke
E
uclidean,
Cosi
ne,
Mi
nkowksi
a
nd C
heb
ys
hev
wer
e
also calc
ulate
d.
Figure
5
.
Cl
assifi
cat
ion
e
ff
ic
ie
ncy f
or
kN
N
cl
assifi
er
with
diff
e
ren
t
kernel
f
un
ct
io
ns (a
)
Fe
at
ur
e set
1
(b)
Feat
ur
e
set
2
The
res
ults
show
t
hat
best
cl
assifi
cat
ion
ac
cur
acy
wa
s
obta
ined
with
nu
m
ber
of
nei
ghbours
eq
ual
t
o
3
with
cosi
ne
distance
m
et
ri
c
fo
r
bo
t
h
the
featur
e
set
s.
T
he
ne
xt
cl
as
sifie
r
cho
se
n
was
a
deep
netw
ork
wi
t
h
two
e
ncoder
ne
tworks
for
fea
ture
ext
racti
on
fo
ll
owe
d
by
a
so
ftm
ax
la
ye
r
fo
r
cl
assifi
cat
io
n.
T
he
tw
o
enc
od
e
r
s
hav
e
a
hi
dd
e
n
la
ye
r
of
siz
e
10
an
d
a
li
ne
ar
trans
fer
functi
on.
T
he
L
2
we
igh
t
re
gula
rize
r,
s
par
sit
y
re
gula
rize
r
and
sp
a
rsity
pr
oport
ion
we
re
set
to
0.0
01,
4
and
0.1
res
pect
ively
.
A
n
ov
e
r
al
l
cl
assifi
cat
ion
acc
ur
acy
of
100
%
was
obta
ine
d
with
histo
gram
based
featu
re
extracti
on
te
ch
nique
(
featu
re
set
1)
wh
e
reas
an
overall
accu
racy
of
88%
was
obta
ined
f
or
featu
re
ext
racti
on
us
ing
k
-
m
eans
segm
entat
ion
fo
l
lowe
d
by
sh
a
pe
and
te
xture
f
eat
ur
es
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4197
-
4203
4202
(f
eat
ure
set
2)
.
A
c
om
par
iso
n
of
the
cl
assifi
cat
ion
ac
cu
ra
cy
of
di
ff
e
ren
t
m
achine
le
ar
ning
cl
assifi
e
r
s
f
or
cl
assifi
cat
ion
of m
acro
nutrie
nt
d
efici
encies
in
m
ai
ze plant usi
ng the t
wo f
e
at
ur
e sets
are s
how
n
in
Fig
ure
6
.
Figure
6
.
Com
par
is
on of e
ff
ic
ie
ncy of classi
f
ic
at
ion
of
dif
fe
ren
t cl
assifi
e
rs
The
res
ults
in
dicat
e
that
De
ep
net
work
usi
ng
s
parse
auto
e
ncoders
giv
es
highest
accuracy
of
cl
assifi
cat
ion
f
or
histo
gr
am
based
feat
ur
es
.
KNN
cl
assifi
er
work
e
d
well
with
both
feat
ur
e
set
s
for
a
cosine
distance
m
et
ric
and
k
value
ch
os
e
n
as
3.
Sim
i
la
rly
SV
M
with
li
near
ke
rn
el
giv
es
good
cl
assifi
cat
ion
acc
ur
acy
bo
t
h
the
f
eat
ure set
s.
T
he
l
owest
accur
a
cy
was o
btaine
d wit
h ANN classi
fi
er.
4.
CONCL
US
I
O
N
Cl
assifi
ca
ti
on
of
m
acro
nu
t
rient
de
fici
enci
es
in
m
ai
ze
plant
was
do
ne
us
in
g
dif
f
eren
t
featu
re
extracti
on
m
e
t
hods
a
nd
diff
e
ren
t
cl
assifi
ers
.
The
al
gorith
m
s
wer
e
te
ste
d
on
the
three
plant
m
acro
nu
trie
nt
def
ic
ie
ncies
i
n
m
ai
ze
plants
,
nam
ely
nitro
ge
n
,
po
ta
ssi
um
a
nd
phosp
hor
us.
Tw
o
featu
re
extracti
on
te
ch
niques
wer
e
us
e
d
to
de
velo
p
t
wo
dif
f
eren
t
featu
re
se
ts
for
t
he
sam
e
le
af
im
ages.
T
he
resu
lt
s
re
ve
al
that
dee
p
net
work
with
a
uto
e
nc
od
e
rs
giv
es
hi
gh
e
st
accu
rac
y
and
has
s
uperi
or
perf
orm
ance
fo
r
his
togram
based
featu
re
extracti
on
co
m
par
ed
to
the
oth
e
r
cl
assifi
cat
ion
m
et
ho
ds.
The
best
cl
assifi
cat
ion
ac
cur
acy
with
s
hap
e
a
nd
te
xtu
re
feat
ur
e
s
was
obta
ine
d
with
KNN
cl
assifi
er.
T
he
eff
ect
of
dif
f
eren
t
ke
rn
el
s
for
S
VM
was
al
so
inv
est
igate
d
for
bo
t
h
the
f
eat
ur
e
set
s.
It
wa
s
seen
that
SVM
with
li
near
kernel
ga
ve
th
e
hig
he
st
a
ccu
racy
of
cl
assifi
cat
ion
in
bo
th
ca
ses.
S
i
m
i
la
rly
the
effe
ct
of
distance
m
et
ric
and
nu
m
ber
of
nei
ghbours
on
acc
uracy
of
kNN
was
al
s
o
check
e
d
f
or
bo
th
the
fe
at
ure
set
s.
It
was
se
en
that
kNN
w
it
h
cosi
ne
m
et
r
ic
and
k
val
ue
of
3
giv
es
h
i
gh
est
a
ccur
acy
.
Th
e st
ud
y ca
n be e
xt
end
e
d
t
o
m
ic
ro
nu
t
rient
def
ic
ie
ncy an
d diseas
e inf
est
at
io
ns
a
lso.
REFERE
NCE
S
[1]
Chen
L.
et
a
l.
,
“
Ide
nti
f
ic
a
ti
on
of
nit
roge
n
,
phosp
horus,
and
pot
assium
def
ic
ie
n
cie
s
in
rice
base
d
o
n
stat
i
c
sca
nn
ing
te
chno
log
y
and
hie
rar
chi
c
al i
den
ti
ficat
ion
m
et
ho
d
,
”
PLoS
ONE
,
vol.
9
,
2014
.
[2]
W
iwar
t
M.,
et
al
.
,
“
Ea
r
l
y
di
ag
nostic
s
of
m
ac
ronutri
ent
de
fici
enc
i
es
in
thre
e
le
gum
e
spec
ie
s
b
y
col
or
imag
e
ana
l
y
sis
,
”
Comp
ute
rs
and
E
lectr
onic
s in
Agric
u
ltur
e
,
vol
.
65
,
200
9
.
[3]
Chanda
n
J.
B.
,
et
al
.
,
“
Dete
c
ti
o
n
and
Anal
y
sis
of
Defi
c
ie
nc
ie
s
in
Groundnu
t
Pl
ant
using
Geom
et
ri
c
Mom
ent
s
,
”
World
Ac
ademy
of
Sci
ence.
E
ngine
ering
and
Technol
ogy
In
te
rnational
Jou
rnal
of
Bi
ological,
Agri
cul
tura
l,
Bi
omole
cul
ar,
F
o
od
and
B
iot
e
ch
nologi
cal E
ng
in
ee
ring
,
vol
/
issue:
5
(
10
),
2011.
[4]
San
y
a
l
P
.
,
et
al.
,
“
Color
te
xture
ana
l
y
s
is
of
ric
e
le
av
es
dia
gnosing
def
ic
ie
n
c
y
in
the
bal
an
ce
of
m
ine
ral
le
vels
towar
ds
improvem
ent
of
cro
p
pr
oduct
ivit
y
,
”
10t
h
Inte
rnationa
l
Confe
renc
e
on
I
nform
ati
on
Tech
nology
,
I
EEE
,
p
p.
85
-
90
,
2007
.
[5]
Hetz
roni
A
.
,
e
t
al
.
,
“
Mac
h
ine
vi
sion
m
onit
oring
of
pla
nt
he
alth
,
”
Adv
Spac
e
Re
s
,
vol/
issue:
14(1
1),
pp.
203
–
21
2
,
2009
.
[6]
Pagola
M
.
,
et
al
.
,
“
New
m
et
hod
to
assess
bar
ley
n
it
roge
n
nutr
it
ion
stat
us
base
d
on
image
col
our
an
aly
s
is
,
”
Comput
El
e
ct
ron A
gri
c
,
vol/
issue:
65(2)
,
pp.
213
–
218
,
20
09.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Cl
as
sif
ic
ation
of M
acron
utrient Defi
ci
encies
in
M
aize Pl
an
t
U
sin
g
…
(
Lee
na N)
4203
[7]
H
ai
ruddin
M
.
A
.
,
e
t
al
.
,
“
Overv
ie
w
of
image
p
roc
essing
appr
o
ac
h
for
nutr
ie
nt
def
iciencie
s
de
te
c
ti
on
in
E
laeis
Guinee
nsis
,
”
IE
EE
In
te
rnationa
l
Confe
ren
ce on Syste
m E
ng
ineering
and
T
e
chnology
,
pp
.
116
-
120
,
2011
.
[8]
Xu
G
.
,
et
al.
,
“
Us
e
of
le
af
col
o
r
images
to
ide
nti
f
y
n
it
rog
en
an
d
pota
ss
i
um
def
ic
ie
n
t
tomatoe
s
,
”
Pattern
Recogn
it
Lett
,
vol/is
sue:
3
2(11)
,
pp
.
1584
-
1590
,
2011
.
[9]
Mons
al
ve
D.
,
et
al
.
,
“
Autom
at
ic
Cla
ss
ifi
c
at
ion
o
f
Nutrit
ion
al
De
f
ic
i
enc
i
es
in
Co
ffe
e
Plan
ts
,
”
6th
Latin
-
Ame
rican
Confe
renc
e
on
N
et
worked
and
Elec
troni
c
M
edi
a
,
2015.
[10]
E.
I.
Z
ac
h
ara
ki
,
et
al
.,
“
Cla
ss
i
fi
c
at
ion
of
bra
in
tu
m
or
ty
p
e
and
gr
ade
using
MRI
te
xtur
e
and
sha
pe
in
a
m
ac
hine
le
arn
ing
sch
eme
,
”
Magn.
R
eson
,
vol.
62
,
pp
.
1609
–
1618
,
2009
.
[11]
M
.
Ham
ia
ne
,
e
t
al
.
,
“
SV
M
Cla
ss
ifi
c
at
ion
of
MRI
Brai
n
Im
age
s
f
or
Com
pute
r
–
As
sisted
Diagnosis
,
”
Inte
rnat
ion
al
Journal
of
Elec
t
rical
and
Computer
Eng
ine
ering
,
vol
/i
ss
ue:
7
(
5
)
,
2
017.
[12]
A
.
Munther,
et
al.
,
“
A
Preli
m
ina
r
y
Perform
ance
Eva
lu
at
i
on
of
K
-
m
ea
ns,
KN
N
and
EM
Uns
up
erv
ised
Ma
chi
ne
Le
arn
ing
Meth
ods
for
Network
Flow
Cla
ss
ifi
cation
,
”
Int
ernati
onal
Journal
of
El
e
ct
rica
l
and
Comput
e
r
Engi
ne
ering
,
vol
/i
ss
ue:
6
(
2
)
,
pp
.
778
-
784
,
2016
.
[13]
S
.
Nare
jo,
et
al
.
,
“
EE
G
Based
E
y
e
Stat
e
Cl
assific
a
ti
on
using
Dee
p
Bel
i
ef
Netw
ork
and
Stac
ked
AutoEnc
oder
,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
,
vo
l
/i
ss
ue:
6
(
6
)
,
pp
.
3131
-
3141
,
201
6
.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Le
en
a.
N
is
pursuing
her
doctoral
degr
e
e
un
der
the
guida
n
ce
of
Dr.K.
K.
Saju
at
Cochi
n
Univer
sit
y
of
Sc
ie
nc
e
and
Techn
olog
y
,
Indi
a.
She
is
an
As
sistant
Profess
or
in
the
Depa
rtment
o
f
El
e
ct
ri
ca
l
and
El
e
ct
roni
cs
Eng
ine
er
ing
at
NS
S
Coll
ege
of
E
ngine
er
ing,
Ker
al
a
,
Indi
a.
Her
rese
arc
h
in
te
rest
s inc
lud
e
soft
co
m
puti
ng;
Mac
h
i
ne
l
ea
rning
and
El
e
ct
ri
c
dr
ive
s
a
nd
cont
ro
l.
K
K
Saju
is
Pr
ofe
ss
or
at
the
D
epa
rtment
of
M
ec
han
ic
a
l
Engi
n
ee
ring
at
Cochin
Univer
sit
y
of
Te
chno
log
y
,
In
dia
.
He
is
a
lso
the
Dir
ec
tor
of
the
In
te
rna
tion
al
Re
la
t
ions
and
Aca
d
emic
Adm
issions
at
C
ochi
n
Univer
si
t
y.
His
int
er
ests
lie
in
m
at
er
ia
l
s
cienc
e
,
robot
ic
s
a
nd
aut
om
at
ion
.
He
is a
u
thor
of
m
ore
tha
n
20
in
t
ern
ational publ
i
c
at
ions
and
h
as
h
andl
ed
seve
r
al
f
unded
proj
ec
ts
.
Evaluation Warning : The document was created with Spire.PDF for Python.