senso s
A icle
An Aga e Coun ing Me hodology Based on
Ma hema ical Mo phology and Images Acqui ed
h ough Unmanned Ae ial Vehicles
Gab iela Cal a io 1, Te esa E. Ala cón 2,*, Osca Dalmau 3, Basilio Sie a 4
and Ca men He nandez 4,5
1Depa men o Elec onics, Sys ems, and In o ma ics, ITESO—The Jesui Uni e si y o Guadalaja a,
Tlaquepaque, Jalisco 45604, Mexico; [email p o ec ed]
2Depa amen o de Ciencias Compu acionales e Ingenie ías, Cen o Uni e si a io de los Valles, Ameca,
Jalisco 46600, Mexico
3Cen o de In es igación en Ma emá icas, Guanajua o 36023, Mexico; [email p o ec ed]
4Depa amen o de Ciencias de la Compu ación e In eligencia A i icial, Uni e sidad del País Vasco
UPV/EHU, 20018 Donos ia-San Sebas ián, Spain; [email p o ec ed] (B.S.); [email p o ec ed] (C.H.)
5Cen e o he Resea ch and Technology o Ag o-En i onmen al and Biological Sciences, CITAB,
Uni e sidade de T ás-os-Mon es e Al o Dou o, UTAD, 5000-801 Vila Real, Po ugal
*Co espondence: e esa.ala con@p o eso es. alles.udg.mx
Recei ed: 29 Sep embe 2020; Accep ed: 28 Oc obe 2020; Published: 2 No embe 2020
Abs ac :
Blue aga e is an impo an comme cial c op in Mexico, and i is he main sou ce o
he adi ional mexican be e age known as equila. The a ie y o blue aga e c op known as
Tequilana Webe is a c ucial elemen o equila ag ibusiness and he ag icul u al economy in Mexico.
The numbe o aga e plan s in he ield is one o he main pa ame e s o es ima ing p oduc ion o
equila. In his manusc ip , we desc ibe a ma hema ical mo phology-based algo i hm ha add esses
he aga e au oma ic coun ing ask. The p oposed me hodology was applied o a se o eal images
collec ed using an Unmanned Ae ial Vehicle equipped wi h a digi al Red-G een-Blue (RGB) came a.
The numbe o plan s au oma ically iden i ied in he collec ed images was compa ed o he numbe
o plan s coun ed by hand. Accu acy o he p oposed algo i hm depended on he size he e ogenei y
o plan s in he ield and illumina ion. Accu acy anged om 0.8309 o 0.9806, and pe o mance o
he p oposed algo i hm was sa is ac o y.
Keywo ds: p ecision ag icul u e; UAV; da a mining; compu e ision; geoma ics; c op moni o ing
1. In oduc ion
Blue aga e is a succulen plan ha g ows in a id and wa m a eas [
1
]. Blue aga e (see Figu e 1) is
a na i e Mexican plan , and i is he p incipal sou ce o p oduc ion o he adi ional Mexican be e age
known as equila. Tequila p oduce s need o es ima e he yield o aga e plan a ion in o de o plan and
p edic p oduc ion o he d ink. The numbe o aga e plan s is one o he pa ame e s ha de e mines
he yield o plan a ion; hence, he elabo a ion o equila equi es p ecise con ol and moni o ing o he
numbe o aga e plan s. Despi e he exis ence o se e al mode n echniques in Mexican ag icul u e,
moni o ing o he aga e c op is mainly done manually wi h a m ools, which equi es g ea e o
o achie e good coun ing p ecision. Acco ding o ou esea ch, he e a e h ee main easons ha
explain he p edominan use o manual coun ing o aga e plan s in Mexico: he lack o in o ma ion
abou he use ulness o image p ocessing me hods in ag icul u e, he lack o a o dable and inno a i e
echniques o aga e moni o ing, and he loss o jobs ha his ype o echnology b ings.
Senso s 2020,20, 6247; doi:10.3390/s20216247 www.mdpi.com/jou nal/senso s
Senso s 2020,20, 6247 2 o 21
Figu e 1. Image o aga e plan a ion: he image is om [2].
Al hough he e a e algo i hms o coun ing plan s [
3
–
6
], o he bes o ou knowledge, he e is
no eliable algo i hm o aga e coun ing. One limi a ion when applying s a e-o - he-a algo i hms
o aga e plan coun ing is he o e lap among plan s ha can be seen in Figu es 1and 2. In his
manusc ip , we p esen an algo i hm o coun ing aga e plan s which is e y use ul in moni o ing
aga e plan a ion. The p oposal is e y impo an due o i allowing imp o emen s in he es ima ion
o Tequila p oduc ion. Ou algo i hm is based on he aga e segmen a ion me hodology p oposed
by Cal a io e al. [
2
], in which he scene was acqui ed h ough Unmanned Ae ial Vehicle (UAV) and
segmen ed by using k-means [
7
]. One d awback o he p e ious wo k is ha i does no sol e he
o e lap p oblem p esen ed in aga e plan a ion, which makes his algo i hm useless o di ec aga e
coun ing. The main con ibu ion o his wo k is he elabo a ion o a me hodology o aga e coun ing
in o de o be e con ol and moni o aga e yield. We desc ibe he implemen a ion o aga e coun ing
algo i hm based on he heo y o ma hema ical mo phology (MM) [
8
,
9
]. One ad an age o MM is
he compu a ional e iciency, addi ionally, i does no equi e la ge image da ase s o es ima e he
pa ame e s o algo i hms.
2. Rela ed Wo k
Remo e sensing and digi al image p ocessing ha e con ibu ed o p ecision ag icul u e.
In pa icula , Unmanned Ae ial Vehicles (UAVs) and image p ocessing ha e been used in ag icul u e
o be e moni o and con ol he yield pa ame e s o di e en c ops. UAV can p o ide a e y high
image esolu ion e en in e ms o millime e s depending on he al i ude o he ligh and esolu ion o
he came a [
10
]. This is e y ele an o managemen and moni o ing o he c op wi hou being in
di ec con ac and wi h a low cos [
2
]. On he o he hand, digi al image p ocessing allows au oma ic o
semiau oma ic analysis o images and ex ac s use ul in o ma ion o a me s. In he li e a u e, we can
ind me hods ha combine UAV echnique wi h compu e ision and a i icial in elligence ha allows
o he ex ac ion o use ul in o ma ion o ag icul u e such as segmen a ion o egions co esponding
o a speci ic c op. Meanwhile, he e a e o he me hods ha di ec ly add ess he coun ing p oblem.
In [
10
], he au ho s e alua ed he pe o mance o six di e en ege a ion indices: Colo index
o ege a ion (CIVE) [
11
], Excess g een (ExG) [
12
], Excess g een minus excess ed (ExGR) [
13
],
Woebbecke Index [
12
], No malized g een- ed di e ence index (NGRDI) [
14
], and Vege a i en
(VEG) [
15
] o s udying whea c op images acqui ed wi h a digi al Red-G een-Blue (RGB) came a
ins alled in UAV. In o de o disc imina e he whea c op, hey used O su’s me hod [
16
]. The au ho s
assessed he accu acy, and spa ial and empo al consis ency o he men ioned indices and conclude ha
ExG and VEG con ibu e wi h he highes classi ica ion accu acy. The in es iga o s in [
17
] p oposed
an OBIA [
18
] algo i hm based on O su’s h eshold me hod and ho oughly s udied how di e en
pa ame e s o a mul i esolu ion segmen a ion algo i hm du ing he segmen a ion s ep a ec he
classi ica ion o ege a ion co e age. In hei esea ch, hey conside ed images o ege a ion indices
ExG [
12
] and he No malized Di e ence Vege a ion Index (NDVI) [
19
] o enhance he in o ma ion
abou ege a ion. The au ho s in [
17
] es he p oposal o ege a ion de ec ion in UAV images
Senso s 2020,20, 6247 3 o 21
acqui ed o e h ee di e en c ops: whea , sun lowe , and maize. In [
20
], he au ho s ca ied ou
esea ch using images acqui ed om wo UAV pla o ms: e ical aking o landing mul i- o o
quadcop e and ixed wing UAV. In hei p oposal, hey combined he spec al and spa ial in o ma ion.
The pu pose o ha s udy was o de ec c op oma o egions and o classi y ee c own. In pa icula ,
hey used Bayesian in o ma ion c i e ion [
21
] o se he op imal numbe o clus e s o a gi en image.
Subsequen ly, hey applied k-means [
7
] and Expec a ion Maximiza ion (EM) [
22
] algo i hms o
clus e ing using spec al in o ma ion. To imp o e clus e ing, hey inco po a ed spa ial in o ma ion
h ough an agglome a i e app oach based on majo i y o ing me hod.
On he o he hand, Gnädinge and Schmidhal e [
3
] ca ied ou esea ch o demons a e he
e iciency o he UAV and image analysis as a possible echnique o coun ing plan s. In hei s udy,
hey analyzed ou di e en maize cul i a s and he images we e acqui ed by means o UAVs. In o de
o disc imina e he g ound co e om he cul i a s, hey used a h esholding echnique. On a e age,
hey achie ed an e o
≤
5% be ween isually and digi ally coun ed plan s. The au ho s in [
23
]
conside ed 10 di e en ege a ion indices and image classi ica ion om UAV-based RGB mul ispec al
images o es ima ing he lowe numbe in wo di e en oilseed ape ields. The ege a ion indices
wi h bigge con ibu ion o de ec lowe s a e NGRDI [
14
], Red G een Ra io Index (RGRI) [
24
],
and Modi ied G een Red Vege a ion Index (MGRVI) [
25
]. Fo classi ica ion, hey used k-means
applied on L*a*b colo space. Random Fo es [26] was also included o p edic he lowe numbe .
The au ho s in [
4
] p oposed a spec al-spa ial classi ie based on a single hidden laye eed- o wa d
neu al ne wo k. The classi ie uses RGB high-spa ial- esolu ion images acqui ed om UAVs. The goal
o he s udy was o de ec , delinea e, and coun ee c owns. The inpu o he classi ie was he RGB
alues o he pixels in he images, and he ou pu was a bina y esponse ( ee o non- ee pixels).
The au ho s ca ied ou he spa ial classi ica ion h ough h esholded geome ical p ope y il e ing
echniques. The coun ing and delinea ion we e done by means o he Wa e shed algo i hm [27].
In [
5
], a sys em o he au oma ic quan i ica ion o whea ea was p oposed based on images
acqui ed by an RGB con en ional came a. The algo i hm conside s 3 s eps: (1) de ec ion o ab up
changes by means o Laplacian equency il e [
28
], (2) median il e ing o smoo h he noise, and (3) a
segmen a ion s ep using Find Maxima. The esea che s de eloped an algo i hm h ough he image
analysis sys em ImageJ [
29
]. The algo i hm achie ed a success a e highe han 90% be ween he
algo i hm coun s and he manual ea coun s. The au ho s in [
6
] in es iga ed a segmen a ion app oach
based on U-ne [
30
] and a con olu ional ne wo k ne (CNN) [
31
], in combina ion wi h in e p e ed
aining da a acqui ed by means o UAV-based high- esolu ion RGB image y. The pu pose o he
esea ch was o ob ain a ine-g ained mapping o ege a ion species and communi ies. The au ho s
in [
6
] achie ed an accu acy o 84% and demons a ed ha usion be ween UAVs and a i icial
in elligence is applicable o a wide ange o asks in ag icul u e.
Ma hema ical mo phology (MM) is also a powe ul ool o disc imina ing a wide ange o shapes
and sizes in he image s uc u e. MM is based on a se ope a o s ha ans o m images aking in o
conside a ion hei opological and geome ical p ope ies. Among he p ope ies a e he size o he
egions and he shape. The mos impo an ope a o s a e dila ion, e osion, opening, and closing [
8
]. In [
32
],
he esea che s p oposed an app oach in which hey analyze images acqui ed h ough UAV by means
o MM in o de o de ec ailu es in co ee c ops. The objec i e o he p oposal was o e alua e p oduc
quali y and he op imal occupa ion o plan ed a eas. The esea che s in [
33
] p oposed a echnique based
on deep lea ning and compu a ional ision using UAV images wi h he pu pose o coun ing co n plan s.
Ki ano and collabo a o s [
33
] used he opening mo phological ope a o o sepa a e de ec ed objec s a e
he segmen a ion done by means o he U-ne a chi ec u e [
30
]. Fan e al. [
34
] implemen ed an algo i hm
ha combines a neu al ne wo k wi h mo phological ope a ions and he Wa e shed algo i hm [
27
].
The goal o hei esea ch was o de ec and coun obacco plan s. In [
32
–
34
], he MM allowed he
sepa a ion be ween objec s o in e es , emo ing o holes and p o usions, which is e y impo an o
p ese e he main ea u es o he de ec ed objec s and, la e , o ca y ou any coun ing algo i hm.
Senso s 2020,20, 6247 4 o 21
In [
2
], Cal a io e al. p oposed a me hodology o disc imina e aga e plan s using UAV, geoma ics,
and compu e ision. The au ho s conside ed an UAV ixed ligh o acqui e aga e plan images and
he k-means algo i hm o disc imina ing aga e c ops. Al hough his me hod has a good segmen a ion
pe o mance, i canno be di ec ly used o aga e coun ing. Now, we p opose a me hodology o
coun ing aga e plan s which is a con inua ion o he esea ch done in [
2
]. Ou p oposal is based
on he use o ma hema ical mo phology. This s a egy allows us o e ine he solu ion p o ided
in [
2
] and, a he same ime, o ob ain a good aga e coun ing me hod. MM is a adi ional image
p ocessing me hod [
8
,
9
] ha does no equi e a aining s age. Fo his kind o me hod, basically,
hype pa ame e s a e adjus ed conside ing he ea u es o he images o in e es .
A i icial neu al ne wo ks ha e ecen ly gained a lo o a en ion due o he applica ion o
deep lea ning echniques allowing o ob ainmen n o excellen esul s in di e en machine lea ning
asks [
35
]. The cos o ob aining such esul s is he inc easing complexi y o he unde lying model.
The e o e, his ype o model equi es la ge da abases du ing he aining s ep, and in gene al, hey a e
no p ac ical when he numbe o da a is e y low. In o de o educe his p oblem, some da a
augmen a ion has ecen ly been s udied. In he case o aga e plan s, he numbe o epo ed labeled
images is e y low, and o he bes o ou knowledge, he e a e no labeled aga e da abases ha allow
us o ain deep lea ning da a-based modeling. The selec ion o MM as a compu e ision ool is also
jus i ied by he p oblem we ace a e he segmen a ion in [
2
]: he o e lap be ween he de ec ed aga e
plan s, and he a ia ion in size and shape. MM has a good pe o mance o ex ac in o ma ion abou
he shape and size, which a e impo an ea u es ha we include in he p oposal desc ibed in Sec ion 3.
The s uc u e o he manusc ip in he ollowing is as ollows: Sec ion 3de ails he s udied
images and he p oposal; he discussion abou he ob ained esul s is p o ided in Sec ion 4; and inally,
he conclusions a e gi en in Sec ion 5.
3. Ma e ials and Me hods
3.1. S udy A eas
The s udy looked a h ee di e en ields, wi h aga e plan s o di e en sizes, concen a ions,
and ages 0. Table 1con ains he geog aphical posi ion, a ea, and age o aga e plan s o each s udied
ield. Figu es 2–4desc ibe he ea u es o each ield.
Table 1. Geog aphical posi ion and a ea o s udied sub egions
Field Geog aphical Posi ion Age A ea (ha)
1 20◦34032.2200 N, 102◦25010.0600 W 2–4 yea s 2.494
2 20◦44034.4900 N, 102◦27030.1400 W 4 yea s 0.4293
3 20◦40005.6900 N, 102◦39006.2300 W 4 yea s 4.018
Field 1 con ains aga e plan s o di e en ages and sizes. Weed is ano he componen o he ield.
Field 1 is o ed clay ype soil, and he e is a colo con as be ween he plan s and he soil. The o e lap
be ween plan s is ano he impo an ea u e. Obse e he a eas ep esen ed by polygons in ed, blue,
and pu ple in Figu e 2a–c. They demons a e he abo e explana ion.
Field 2 is o ocky soil, and he e is less con as be ween he plan s and he soil compa ed o
Field 1. Simila o Field 1, he e a e plan s o di e en sizes, bu he o e lap p oblem is less wi h espec
o Field 1.
Senso s 2020,20, 6247 5 o 21
(a)
(b)(c)(d)
Figu e 2.
Images o Field 1 ob ained wi h an Unmanned Ae ial Vehicle (UAV): (
a
) Field 1 wi h some
ma ked egions ha a e s udied in his wo k, (
b
) he egion in blue, (
c
) he egion in pu ple, and (
d
) he
egion in ed.
Senso s 2020,20, 6247 6 o 21
(a)
(b)(c)(d)
Figu e 3.
Images o Field 2 ob ained wi h an UAV: (
a
) Field 2 wi h some ma ked egions ha a e
s udied in his wo k, (b) he egion in blue, (c) he egion in pu ple, and (d) he egion in ed.
Field 3 is o he same ed clay as in Field 1. The o e lap p oblem and weeds a e also p esen .
No e ha , in his case, he a iabili y o plan sizes is lowe han he a iabili y in Fields 1 and 2.
The sowing sys em o he blue aga e is linea ; see Figu e 5. The lines can be made h ough a
ac o , s akes, o manually wi h a h ead. The u ows a e dis ibu ed a a dis ance allowing o
sepa a ion be ween he plan ows. A e wa ds, a h ead wi h ma ks is u ilised o ix he needed
dis ance be ween plan s.
Senso s 2020,20, 6247 7 o 21
(a)
(b)(c)(d)
Figu e 4.
Images o Field 3 ob ained wi h an UAV: (
a
) Field 3 wi h some ma ked egions ha a e
s udied in his wo k, (b) he egion in blue, (c) he egion in pu ple, and (d) he egion in ed.
Figu e 5.
Aga e plan a ion: he zone su ounded by a ci cle zooms in on he lines o he sowing sys em.
Senso s 2020,20, 6247 8 o 21
On a e age, he sepa a ion be ween each ow is 3 o 4 m and he dis ance be ween plan s is 1.00 o
1.20 m [
36
]. The diame e o he ci cle su ounding he plan wid h is abou 1–2 m o mo e [
37
], so ha
one plan can be bigge han he dis ance be ween plan s. This explains he o e lap be ween plan s.
3.2. Wo k low
This wo k imp o es he segmen a ion esul in [
2
] so ha we can use i o aga e coun ing
pu poses. The e o e, he inpu o ou me hod co esponds o he ou pu o he p oposal in [2].
Cal a io e al. in [
2
] p oposed a me hodology based on pho og amme y and k-means algo i hm.
They use unmanned ae ial ehicles o ype Phan om 4, DJI, and he images a e acqui ed wi h
an RGB senso wi h 6.25 mm
×
4.68 mm and lens wi h a Field o View (FOV) o 94
◦
20 mm.
The spa ial and bi esolu ions o he image a e 4000
×
3000 and 8 bi pe pixel, espec i ely [
2
,
38
].
The UAV ligh is planned aking in o accoun he opog aphy and Google Ea h in o ma ion.
Acco ding o he expe imen al wo k, he selec ed ligh al i ude was 60 me e s. In his wo k,
we ollow his ecommenda ion; see he de ails in [
2
]. Once he images a e acqui ed h ough he
UAV, he pho og amme ic p ocessing is ca ied ou . Following his, he au ho s in [
2
] gene a ed
he co esponding o homosaic; see Figu e 6a, wi h a spa ial esolu ion o abou 3 cm pe pixel.
The ob ained image is con e ed om RGB colo space o CIELab in o de o use he Euclidean
dis ance c i e ion in a pe cep ual uni o m space. Subsequen ly, k-means was applied o disc imina e
aga e plan s om he es o objec s in he image. The de ec ed classes, aga e plan s and backg ound,
a e alida ed using he g ound u h and geog aphic sys em in o ma ion. Fo mo e de ails, see [
2
].
An example o a segmen ed image, ob ained in [
2
], appea s in Figu e 6b. The inal ep esen a ion o
he esul s is gi en in RGB.
(a) (b)
Figu e 6.
Ex ac ion o he aga e laye : he image in (
a
) ep esen s he aga e c ops wi hou applying
he plan ex ac ion me hodology in [
2
]. The image in (
b
) shows he aga e de ec ion laye (segmen ed
image) a e applying he me hodology. Ci cles in ed highligh he o e lapping zones.
Figu e 6b depic s he inpu o ou coun ing algo i hm. No e he o e lap p oblem in he
segmen a ion esul s.
The g aphical ep esen a ion o he elabo a ed algo i hm is depic ed in Figu e 7. A de ailed
explana ion o he algo i hm is gi en in Sec ion 3.3.
Senso s 2020,20, 6247 9 o 21
Figu e 7.
Flow cha o he implemen ed aga e coun ing algo i hm based on ma hema ical mo phology:
he abb e ia ion CC s ands o connec ed componen . The g ouping s ep is ca ied ou by a h esholding
me hod. The segmen ed image is he ou pu o he p oposal in [2].
3.3. Implemen a ion o Ma hema ical Mo phology Ope a ions o Coun ing Aga e Plan s
The p oposal is composed o 3 s eps: p ep ocessing, objec sepa a ion, and coun ing.
1. P ep ocessing
: The main aim o his s ep is o ob ain a bina y image as clean as possible so ha
we can achie e a good objec sepa a ion in he nex s ep. Fo his pu pose, we ca y ou he
ollowing s eps:
•
Fi s ly, we ans o m he segmen ed RGB colo image in o a g ay-scale image by using he
ollowing equa ion:
I=0.299 ∗R+0.587 ∗G+0.114 ∗B(1)
whe e
R
,
G
, and
B
ep esen he ed, g een, and blue channels o he segmen ed image
(Figu e 8a) and Ico esponds o he g ay-scale ep esen a ion (Figu e 8b).
•Secondly, we bina ize he p e ious image Iacco ding o he ollowing equa ion:
I( ) = 1I( )>0
0o he wise (2)
•
A e he p e ious s ep, we ypically achie e a noisy bina y image wi h isola ed poin s and
holes. Fo his eason, we emo e hese a i ac s in o de o educe alse posi i es o alse
nega i es du ing he coun ing p ocess. Fo his s ep, we can use se e al image p ocessing
echniques. In pa icula , in his wo k, we use mo phological il e s, i.e., clean and ill
Senso s 2020,20, 6247 16 o 21
(a) (b)
(c) (d)
Figu e 14.
(
a
) A polygon in blue in Field 2 (Figu e 3), (
b
) an ex ac ed aga e laye by means o
he me hodology in [
2
], (
c
) aga e plan s coun ed by an expe (g ound u h), and (
d
) aga e plan s
de ec ed by he p oposal: hose no de ec ed appea in ed ( alse nega i e), while alse posi i es appea
in pu ple.
Unlike he example in Figu e 13a, he segmen ed image gi en in Figu e 14b h ough he
me hodology in [
2
] was a ec ed by illumina ion p oblems. This ac is e lec ed by he esul s o
ou coun ing algo i hm; see Figu e 14d. In his case, we ha e no only alse nega i es bu also alse
posi i es. False nega i es appea in ed, and alse posi i es appea in pu ple. The coun ing done by an
expe is illus a ed in Figu e 14c. Illumina ion p oblems in he gi en image a ec ed he ex ac ion o
he aga e plan s, and as a consequence, i was impossible o coun hem all. This explains he alse
nega i e esul s when he p oposal ies o de ec e y small plan s. Fu he , he illumina ion p oblems
b ing abou he alse posi i e esul s; see he ci cles in pu ple in Figu e 14d.
The esul s shown in Figu e 15 a e simila o hose in Figu e 13. Ve y small plan s we e no
de ec ed, and he o e lap p oblem a ec ed coun ing. False posi i es (a ci cle in pu ple) a e a i ac s in
he image due o he acquisi ion p ocess in [2].
Senso s 2020,20, 6247 17 o 21
(a) (b)
(c) (d)
Figu e 15.
(
a
) A polygon in blue in Field 3 (Figu e 4), (
b
) an ex ac ed aga e laye by means o he
me hodology in [
2
], (
c
) aga e plan s coun ed by an expe (g ound u h), and (
d
) aga e plan s de ec ed
by he p oposal; hose no de ec ed appea in ed ( alse nega i e).
The nume ical esul s in Table 2gi e in o ma ion abou he pe o mance o he coun ing algo i hm
applied on 3 egions in Field 1, Field 2, and Field 3; see Figu es 2–4. In his able, he symbol
S
e e s
o he samples in Figu es 2–4, while
GT
deno es he numbe o aga e plan s de ec ed by an expe ,
i.e., he g ound u h. The nume ical in o ma ion co esponds o applying he accu acy me ics
desc ibed in Sec ion 3.4.
Table 2. Compu ed accu acy me ics.
S TP FN FP GT Pacc Uacc Recall Acc
Field 1-Blue 198 25 1 223 0.8879 0.9950 0.8979 0.9414
Field 1-Pu ple 336 33 9 369 0.9106 0.9739 0.9106 0.9422
Field 1-Red 344 70 11 414 0.8309 0.9690 0.9309 0.9000
Field 2-Blue 326 60 25 386 0.8446 0.9288 0.8446 0.8867
Field 2-Pu ple 99 6 13 105 0.9429 0.8839 0.9429 0.9134
Field 2-Red 177 8 7 185 0.9568 0.9620 0.9568 0.9594
Field 3-Blue 101 2 10 103 0.9806 0.9099 0.9806 0.9452
Field 3-Pu ple 192 7 25 199 0.9648 0.8848 0.9648 0.9248
Field 3-Red 78 5 15 83 0.9398 0.8387 0.9398 0.8892
Senso s 2020,20, 6247 18 o 21
We obse e (see column
FN
) ha he numbe o alse nega i es in Field 1 is g ea e han in he
es o he ields. This is due mainly o he o e lap p oblem and he a iabili y in size and shape o
plan s in Field 1; see Figu e 2. On he o he hand, Field 3 has he leas numbe o alse nega i es.
In his ield, he aga e plan s a e mo e homogeneous in size and age. Al hough he o e lap p oblem is
also p esen , he a ea co esponding o each plan is mo e delimi ed in compa ison wi h o he wo
ields (Figu es 3and 4. Acco ding o column
FP
in Table 2, he samples in Fields 2 and 3 ha e he
highes alse posi i es. In his case, he weed segmen ed as aga e plan s explains he numbe o
FP
s.
F om Table 2, i could be obse ed ha he highes alues o
Pacc
we e ob ained o samples in Fields 2
and 3, which is in co espondence wi h he lowes
FN
alues o e hese samples. The highes alue o
Uacc
co esponds o samples in Field 1 because his Field has he lowes numbe o
FP
s. The highes
alues o
Acc
and
ecall
a e achie ed in samples o Fields 2 and 3. This is due o less a iabili y in size
and age and less o e lap among he aga e plan s. The measu e
Acc
achie ed a alue in he ange om
0.8892 o 0.9594. O e all, he Acc o he nine conside ed egions is 0.8955.
The p oposed coun ing algo i hm is able o de ec a wide ange o aga e plan s in size and age.
This p oposal has been applied on plan a ions loca ed in Jalisco S a e, Mexico, and was e i ied by he
Tequila Regula o y Council (CRT) in o de o p omo e p ecision ag icul u e in Mexico. Acco ding o
CRT, he ob ained esul s a e sa is ac o y and con ibu e owa ds he imp o emen o aga e c op
moni o ing and he e o e a be e p oduc ion con ol o he equila d ink [49].
5. Conclusions and Fu u e Wo k
The numbe o blue aga e plan s is an impo an pa ame e o es ima ing he yield o plan a ion
and o planning he p oduc ion o Tequila, a Mexican be e age ob ained om aga e. In his wo k,
an aga e coun ing algo i hm was p oposed in o de o imp o e he moni o ing o aga e c op and o
con ol he yield o he plan a ion in Jalisco, Mexico.
To he bes o ou knowledge, his is he i s algo i hm aimed a au oma ic coun ing o aga e
plan s in Mexico.
The p oposed coun ing algo i hm is based on ma hema ical mo phology. This is a s anda d
image-p ocessing echnique e y use ul o sepa a ing and coun ing s uc u es by hei size and
shape; his jus i ies ou choice. The p oposal combines he esul s o a segmen a ion algo i hm wi h
he applica ion o a well-s uc u ed sequence o mo phological ope a ions so as o achie e a good
plan sepa a ion and, acco dingly, eliable coun ing. The use o ma hema ical mo phology allows
us o educe he o e lap and ligh ing issues ha a e no ully esol ed in he segmen a ion s ep.
The p oposed algo i hm is able o sepa a e aga e plan s and o p ese e he main aga e plan pa e ns,
ob aining a eliable coun wi h a p oduce ’s accu acy in a ange o 0.8309 o 0.9806.
As ea u e wo k, we conside c ea ing a da abase inc easing he numbe o labeled aga e plan
images so ha we can use supe ised lea ning echniques like an a i icial neu al ne wo k and in
pa icula deep lea ning echniques.
Au ho Con ibu ions:
G.C. concei ed he idea o he pape and conduc ed he esea ch; B.S., T.E.A., and O.D.
helped in he expe imen al design, o mal analysis, and w i ing p ocess; and C.H. analyzed he da a and he
esul s o he expe imen s. All he au ho s ha e aken pa in he elabo a ion o his documen . All au ho s ha e
ead and ag eed o he published e sion o he manusc ip .
Funding:
This esea ch was suppo ed by he Spanish Minis e io de Economía y Compe i i idad,
con ac TIN2015-64395-R (MINECO/FEDER, UE), as well as by he Basque Go e nmen , con ac IT900-16.
This wo k was also suppo ed in pa by CONACYT (Mexico), g an 258033.
Acknowledgmen s:
We wish o acknowledge he Consejo Nacional de Ciencia y Tecnologia (CONACYT) o
i s inancial suppo o he PhD s udies o Gab iela Cal a io. We a e g a e ul o ITESO–The Jesui Uni e si y o
Guadalaja a o he s imula ing his wo k. In addi ion, we a e g a e ul o he suppo o he Tequila Regula o y
Council (CRT), which has allowed us o moni o se e al c ops.
Con lic s o In e es : The au ho s decla e no con lic o in e es .
Senso s 2020,20, 6247 19 o 21
Re e ences
1.
Flo es-Sahagun, T.H.; Dos San os, L.P.; Dos San os, J.; Mazza o, I.; Mikowski, A. Cha ac e iza ion o blue
aga e bagasse ibe s o Mexico. Compos. Pa A Appl. Sci. Manu . 2013,45, 153–161.
2.
Cal a io, G.; Sie a, B.; Ala con, T.; He nandez, C.; Dalmau, O. A mul i-disciplina y app oach o emo e
sensing h ough low-cos ua s. Senso s 2017,17, 1411. [C ossRe ] [PubMed]
3.
Gnädinge , F.; Schmidhal e , U. Digi al coun s o maize plan s by unmanned ae ial ehicles (UAVs).
Remo e Sens. 2017,9, 544. [C ossRe ]
4.
Kes u , R.; Angu al, A.; Bashi , B.; Omka , S.N.; Anand, G.; Meena a hi, M.B. T ee c own de ec ion,
delinea ion and coun ing in ua emo e sensed images: A neu al ne wo k based spec al–spa ial me hod.
J. Indian Soc. Remo e Sens. 2018,46, 991–1004. [C ossRe ]
5.
Fe nandez-Gallego, J.A.; Ke au e , S.C.; Gu ié ez, N.A.; Nie o-Talad iz, M.T.; A aus, J.L. Whea ea coun ing
in- ield condi ions: High h oughpu and low-cos app oach using RGB images. Plan Me hods
2018
,14, 22.
[C ossRe ]
6.
Ka enbo n, T.; Eichel, J.; Fassnach , F.E. Con olu ional Neu al Ne wo ks enable e icien , accu a e and
ine-g ained segmen a ion o plan species and communi ies om high- esolu ion UAV image y. Sci. Rep.
2019,9, 17656. [C ossRe ]
7.
MacQueen, J. Some me hods o classi ica ion and analysis o mul i a ia e obse a ions. In P oceedings
o he Fi h Be keley Symposium on Ma hema ical S a is ics and P obabili y; Volume 1: S a is ics; Uni e si y o
Cali o nia P ess: Be keley, CA, USA, 1967; pp. 281–297.
8. Doughe y, E. Ma hema ical Mo phology in Image P ocessing; CRC P ess: Boca Ra on, FL, USA, 1992.
9.
Jean, S.; Pie e, S. Ma hema ical Mo phology and I s Applica ions o Image P ocessing; Sp inge : Ams e dam,
The Ne he lands, 1994.
10.
To es-Sánchez, J.; Peña, J.M.; Cas o, A.I.D.; López-G anados, F. Mul i- empo al mapping o he ege a ion
ac ion in ea ly-season whea ields using images om UAV. Compu . Elec on. Ag ic.
2014
,103, 104–113.
[C ossRe ]
11.
Ka aoka, T.; Kaneko, T.; Okamo o, H.; Ha a, S. C op g ow h es ima ion sys em using machine ision.
In . Con . Ad . In ell. Mecha on. 2003,2, 1079–1083.
12.
Woebbecke, D.; Meye , G.; Von Ba gen, K.; Mo ensen, D. Colo indices o weed iden i ica ion unde a ious
soil, esidue, and ligh ing condi ions. T ans. ASAE 1995,38, 259–269. [C ossRe ]
13.
Cama go Ne o, J. A Combined S a is ical-So Compu ing App oach o Classi ica ion and Mapping Weed Species in
Minimum-Tillage Sys ems; ETD Collec ion o Uni e si y o Neb aska: Lincoln, NE, USA, 2004; pp. 1–170.
14.
Gi elson, A.A.; Kau man, Y.J.; S a k, R.; Rundquis , D. No el algo i hms o emo e es ima ion o ege a ion
ac ion. Remo e Sens. En i on. 2002,80, 76–87. [C ossRe ]
15.
Hague, T.; Tille , N.; Wheele , H. Au oma ed c op and weed moni o ing in widely spaced ce eals.
P ecis. Ag ic. 2006,7, 21–32. [C ossRe ]
16.
O su, N. A h eshold selec ion me hod om g ay-le el his og ams. IEEE T ans. Sys . Man Cybe .
1979
,
9, 62–69. [C ossRe ]
17.
To es-Sánchez, J.; López-G anados, F.; Peña, J.M. An au oma ic objec -based me hod o op imal h esholding in
UAV images: Applica ion o ege a ion de ec ion in he baceous c ops. Compu . Elec on. Ag ic.
2015
,114, 43–52.
[C ossRe ]
18.
Blaschke, T. Objec based image analysis o emo e sensing. ISPRS J. Pho og amm. Remo e Sens.
2010
,
65, 2–16. [C ossRe ]
19.
Guija o, M.; Paja es, G.; Riomo os, I.; He e a, P.; Bu gos-A izzu, X.; Ribei o, A. Au oma ic segmen a ion
o ele an ex u es in ag icul u al images. Compu . Elec on. Ag ic. 2011,75, 75–83. [C ossRe ]
20.
Sen hilna h, J.; Kanduku i, M.; Dokania, A.; Ramesh, K.N. Applica ion o UAV imaging pla o m o
ege a ion analysis based on spec al-spa ial me hods. Compu . Elec on. Ag ic. 2017,140, 8–24. [C ossRe ]
21. Schwa z, G. Es ima ing he Dimension o a Model. Ann. S a . 1978,6, 461–464. [C ossRe ]
22.
Demps e , A.; Lai d, N.; Rubin, D. Maximum likelihood om incomple e da a ia he EM algo i hm.
S a is . Soc. Se . B (Me hodol.) 1977,39, 1–22.
23.
Wan, L.; Li, Y.; Cen, H.; Zhu, J.; Yin, W.; Wu, W.; Zhu, H.; Sun, D.; Zhou, W.; He, Y. Combining UAV-based
ege a ion indices and image classi ica ion o es ima e lowe numbe in oilseed ape. Remo e Sens.
2018
,
10, 1484. [C ossRe ]
Senso s 2020,20, 6247 20 o 21
24.
Ve els , J.; Schaepman, M.; Koe z, B.; Kneubuehle , M. Angula sensi i i y analysis o ege a ion indices
de i ed om ch is/p oba da a. Remo e Sens. En i on. 2008,112, 2341–2353. [C ossRe ]
25.
Bendig, J.; Yu, K.; Aasen, H.; Bol en, A.; Benne z, S.; B oschei , J.; Gnyp, M.L.; Ba e h, G. Combining
ua -based plan heigh om c op su ace models, isible, and nea in a ed ege a ion indices o biomass
moni o ing in ba ley. In . J. Appl. Ea h Obs. 2015,39, 79–87. [C ossRe ]
26. B eiman, L. Random o es s. Mach. Lea n. 2001,45, 5–32. [C ossRe ]
27. Be and, G. On opological wa e sheds. J. Ma h. Imaging Vis. 2005,22, 217–230. [C ossRe ]
28.
Va ghese, J.; Subash, S.; Tai an, N.; Babu, B. Laplacian-Based F equency Domain Fil e o he Res o a ion o
Digi al Images Co up ed by Pe iodic Noise. Can. J. Elec . Compu . Eng. 2016,39, 82–91. [C ossRe ]
29.
Schneide , C.A.; Rasband, W.S.; Elicei i, K.W. NIH Image o ImageJ: 25 yea s o image analysis. Na . Me hods
2012,9, 671–675. [C ossRe ] [PubMed]
30.
Ronnebe ge , O.; Fische , P.; B ox, T. U-Ne : Con olu ional Ne wo ks o Biomedical Image Segmen a ion.
In Medical Image Compu ing and Compu e -Assis ed In e en ion—MICCAI 2015; Sp inge : Cham, Swi ze land,
2015; pp. 234–241.
31.
Gu, J.; Wang, Z.; Kuen, J.; Ma, L.; Shah oudy, A.; Shuai, B.; Liu, T.; Wang, X.; Wang, G.; Cai, J.; e al.
Recen ad ances in con olu ional neu al ne wo ks. Pa e n Recogni . 2018,77, 354–377. [C ossRe ]
32.
Oli ei a, H.C.; Guizilini, V.C.; Nunes, I.P.; Souza, J.R. Failu e De ec ion in Row C ops F om UAV Images
Using Mo phological Ope a o s. IEEE Geosci. Remo e Sens. Le . 2018,15, 991–995. [C ossRe ]
33.
Ki ano, B.T.; Mendes, C.C.; Geus, A.R.; Oli ei a, H.C.; Souza, J.R. Co n Plan Coun ing Using Deep Lea ning
and UAV Images. IEEE Geosci. Remo e. Sens. Le . 2019, 1–5. [C ossRe ]
34.
Fan, Z.; Lu, J.; Gong, M.; Xie, H.; Goodman, E.D. Au oma ic obacco plan de ec ion in UAV images ia deep
neu al ne wo ks. IEEE J. Sel. Top. Appl. Ea h Obs. Remo e Sens. 2018,11, 876–887. [C ossRe ]
35.
Ghosh, S.; Das, N.; Das, I.; Maulik, U. Unde s anding Deep Lea ning Techniques o Image Segmen a ion.
ACM Compu . Su . 2019,52, 73–107. [C ossRe ]
36.
Jose ina, L.U.P. E aluación de di e en es dosis de e ilización (npk) en el aga e equile o (aga e equilana
webe ). Ph.D. Thesis, Uni e sidad de Guadalaja a, Facul ad de Ag onomía, Zapopan, Jalisco, Mexico, 1987.
37.
Bueno, J.X.U.; Gu ié ez, C.V.; Figue oa, A.R. Mues eo y Análisis de Suelo en Plan aciones de Aga e.
In Conocimien o y P ác icas Ag onómicas pa a la P oducción de Aga e Tequilana Webe en la Zona de Denominación de
O igen del Tequila; Cen o de In es igación Regional Del PacíFico Cen o, Campo Expe imen al Cen o-Al os:
Tepa i lán de Mo elos, Jalisco, Mexico, 2007 p. 37.
38. DJI. A ailable online: h p://www.dji.com/mx/phan om-4 (accessed on 1 July 2020).
39. Se a, J. Mo phological il e ing: An o e iew. Signal P ocess. 1994,38, 3–11. [C ossRe ]
40.
La aˇc, N.; Flach, P.; Zupan, B. Rule E alua ion Measu es: A Uni ying View. In Induc i e Logic P og amming;
Dže oski, S.; Flach, P., Eds.; Sp inge : Be lin/Heidelbe g, Ge many, 1999; pp. 174–185.
41. Cohen, J. A coe icien o ag eemen o nominal scales. Educ. Psychol. Meas. 1960,20, 37–46. [C ossRe ]
42. Bazi, Y.; Malek, S.; Alajlan, N.; AlHich i, H. An au oma ic app oach o palm ee coun ing in UAV images.
In P oceedings o he 2014 IEEE Geoscience and Remo e Sensing Symposium, Québec Ci y, QC, Canada,
13–18 July 2014; IEEE: Pisca away, NJ, USA, 2014; pp. 537–540.
43.
Mo anduzzo, T.; Melgani, F. Au oma ic ca coun ing me hod o unmanned ae ial ehicle images. IEEE T ans.
Geosci. Remo e Sens. 2013,52, 1635–1647. [C ossRe ]
44.
Du ou , J.M.; Ne es, J. Chap e 1—Fini e-sample in e ence and nons anda d asymp o ics wi h Mon e
Ca lo es s and R. In Concep ual Econome ics Using R; Vinod, H.D., Rao, C., Eds.; Else ie : Ams e dam,
The Ne he lands, 2019; Volume 41, pp. 3–31.
45.
La ochelle, H.; E han, D.; Cou ille, A.; Be gs a, J.; Bengio, Y. An empi ical e alua ion o deep a chi ec u es
on p oblems wi h many ac o s o a ia ion. In ICML; Associa ion o Compu ing Machine y: New Yo k,
USA, 2007.
46.
Hin on, G.E. A P ac ical Guide o T aining Res ic ed Bol zmann Machines. In Neu al Ne wo ks: T icks o he
T ade, 2nd ed.; Sp inge : Be lin/Heidelbe g, Ge many, 2012; pp. 599–619.
47.
Chung Chang, C.; Jen Lin, C. LIBSVM: A Lib a y o Suppo Vec o Machines. 2001. A ailable online:
h p://www.csie.n u.edu. w/~cjlin/libs m (accessed on 30 June 2020).
Senso s 2020,20, 6247 21 o 21
48.
Ped egosa, F.; Va oquaux, G.; G am o , A.; Michel, V.; Thi ion, B.; G isel, O.; Blondel, M.; P e enho e , P.;
Weiss, R.; Dubou g, V.; e al. Sciki -lea n: Machine Lea ning in Py hon. J. Mach. Lea n. Res.
2011
,
12, 2825–2830.
49.
Cal a io, G.; Sie a, B.; Ma inez, J.; Mon e , E. Un En oque Mul idisciplina io de Sensado Remo o a T a és
de UAV de Bajo Cos o. In I Simposio de Aplicaciones Cien í icas y Técnicas de los Vehículos no T ipulados;
Uni e sidad Nacional Au onoma de Mexico, (UNAM): Mexico Ci y, Mexico, 2017; pp. 24–32.
Publishe ’s No e:
MDPI s ays neu al wi h ega d o ju isdic ional claims in published maps and ins i u ional
a ilia ions.
c
2020 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access
a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion
(CC BY) license (h p://c ea i ecommons.o g/licenses/by/4.0/).
