A Novel Machine Learning Method for Estimating Biomass of Grass Swards Using a Photogrammetric Canopy Height Model, Images and Vegetation Indices Captured by a Drone
Full text
ag icul u e
A icle
A No el Machine Lea ning Me hod o Es ima ing
Biomass o G ass Swa ds Using a Pho og amme ic
Canopy Heigh Model, Images and Vege a ion Indices
Cap u ed by a D one
Niko Viljanen 1,*ID , Eija Honka aa a 1ID , Roope Näsi 1ID , Teemu Hakala 1, Oi a Niemeläinen 2
and Je e Kai osoja 2ID
1
Depa men o Remo e Sensing and Pho og amme y, Finnish Geospa ial Resea ch Ins i u e, Geodee in inne
2, 02430 Masala, Finland; [email p o ec ed] (E.H.); [email p o ec ed] (R.N.); [email p o ec ed] (T.H.)
2G een Technology Uni , Na u al Resou ces Ins i u e Finland (LUKE), Vakolan ie 55, 03400 Vih i, Finland;
[email p o ec ed] (O.N.); [email p o ec ed] (J.K.)
*Co espondence: [email p o ec ed]; Tel.: +35-850-400-9527
Recei ed: 29 Ma ch 2018; Accep ed: 14 May 2018; Published: 17 May 2018
Abs ac :
Silage is he main eed in milk and uminan mea p oduc ion in No he n Eu ope. No el
d one-based emo e sensing echnology could be u ilized in many phases o silage p oduc ion,
bu ad anced me hods o u ilizing hese da a a e s ill de eloping. G ass swa ds a e ha es ed h ee
imes in season, and e ilize is applied simila ly h ee imes—once o each ha es when aiming
a maximum yields. Timely in o ma ion o he yield is hus necessa y se e al imes in a season o
making decisions on ha es ing ime and a e o e ilize applica ion. Ou objec i e was o de elop
and assess a no el machine lea ning echnique o he es ima ion o canopy heigh and biomass o
g ass swa ds u ilizing mul ispec al pho og amme ic came a da a. Va ia ion in he s udied c op
s and was gene a ed using six di e en ni ogen e ilize le els and ou ha es ing da es. The swa d
was a imo hy-meadow escue mix u e domina ed by imo hy. We ex ac ed a ious ea u es om
he emo e sensing da a by combining an ul a-high esolu ion pho og amme ic canopy heigh
model (CHM) wi h a pixel size o 1.0 cm and ed, g een, blue (RGB) and nea -in a ed ange in ensi y
alues and di e en ege a ion indices (VI) ex ac ed om o hopho o mosaics. We compa ed he
pe o mance o mul iple linea eg ession (MLR) and a Random Fo es es ima o (RF) wi h di e en
combina ions o he CHM, RGB and VI ea u es. The bes es ima ion esul s wi h bo h me hods
we e ob ained by combining CHM and VI ea u es and all h ee ea u e classes (CHM, RGB and VI
ea u es). Bo h es ima o s p o ided equally accu a e esul s. The Pea son co ela ion coe icien s
(PCC) and Roo Mean Squa e E o s (RMSEs) o he es ima ions we e a bes 0.98 and 0.34 /ha
(12.70%), espec i ely, o he d y ma e yield (DMY) and 0.98 and 1.22 /ha (11.05%), espec i ely,
o he esh yield (FY) es ima ions. Ou assessmen o he sensi i i y o he me hod wi h espec o
di e en de elopmen s ages and di e en amoun s o biomass showed ha he use o he machine
lea ning echnique ha in eg a ed mul iple ea u es imp o ed he esul s in compa ison o he simple
linea eg essions. These esul s we e ex emely p omising, showing ha he p oposed mul ispec al
pho og amme ic app oach can p o ide accu a e biomass es ima es o g ass swa ds, and could be
de eloped as a low-cos ool o p ac ical a ming applica ions.
Keywo ds:
pho og amme y; d one; unmanned ae ial ehicle; digi al su ace model; canopy heigh
model; g ass swa d; biomass; machine lea ning; Random Fo es ; mul iple linea eg ession
Ag icul u e 2018,8, 70; doi:10.3390/ag icul u e8050070 www.mdpi.com/jou nal/ag icul u e
Ag icul u e 2018,8, 70 2 o 28
1. In oduc ion
Silage is he main eed in uminan mea and milk p oduc ion in No he n Eu ope. In Finland,
23% o he cul i a ed a ea is used o silage p oduc ion, which is app oxima ely 513,500 ha. In he
silage p oduc ion, g ass swa ds a e ha es ed h ee imes each season, and e ilize is applied o
each ha es . Timely in o ma ion o he yield quali y and quan i y would be highly aluable o
decision-making on ha es ing ime and a e o e ilize applica ion o he nex ha es . In g ass
yield, he quan i y inc eases apidly in he sp ing g ow h. Simul aneously he quali y dec eases,
especially in g ass diges ibili y. Ha es ing ime op imiza ion en ails a balance be ween he highes
possible yield quan i y and an adequa ely high diges ibili y o eeding.
The g ass biomass has a s ong co ela ion wi h canopy heigh s [
1
–
3
]. The e o e, he g ass heigh
is a undamen al pa ame e o in e es when conce ning p ecision managemen o g azing and silage
ha es ing. Accu a e es ima es o g ass biophysical a iables a e impo an o moni o ing ege a ion
g ow h and o analysing impo an physiological pa ame e s du ing he g ass g ow h cycle [
4
].
In p ac ical a ming—pa icula ly in o a ional g azing managemen —physical measu emen s o
g ass heigh and biomass es ima ion a e usually done by using de ices such as he ising pla e
me e , capaci ance me e and me e s ick [
5
–
7
]. Howe e , hese in si u physical measu emen s a e
labo ious. Fu he mo e, i can be di icul o cha ac e ize he spa ial a iabili y due o ege a ion
g ow h cha ac e is ics by he physical sample collec ion, which limi s hei abili y o p o ide obus
es ima es. Se e al ac o s cause a ia ion in he g ass swa ds. In pa icula , swa ds a e composed o
mul iple species, each ha ing di e en g owing cha ac e is ics and a iabili y in soil condi ions and
opog aphy wi hin he ield [8].
Remo e sensing me hods, including digi al imaging, pho og amme y, hype spec al imaging,
lase scanning and a ious senso combina ions can wo k as high-pe o mance al e na i es o physical
measu emen me hods. Remo e sensing o e s a po en ial o apid and au oma ic measu emen o
la ge a eas wi h high spa ial esolu ion. Se e al s udies ha e in es iga ed he use o emo e sensing
echniques o calcula e plan heigh s o es ima ing c op pa ame e s. Mos o hose ha e been using
e es ial lase scanning [
9
,
10
] and mobile lase scanning [
8
] due o he equi emen s o high spa ial
esolu ion. Howe e , pho og amme ic imaging using unmanned ai c a ehicles (UAV o a d one),
s uc u e om mo ion (SFM) echniques and dense image ma ching a e becoming a e y in e es ing
ool o collec 3D in o ma ion o objec s due hei low cos , e iciency and lexibili y. These me hods
ha e al eady been in es iga ed in se e al s udies especially ela ed o g ass- o sh ub ansi ion zone [
3
],
c ops [
11
], win e ba ley [
12
,
13
], maize [
14
] and moss beds [
15
,
16
]. Se e al s udies ha e also u ilized
ege a ion indices (VI) based on mul ispec al da a [
17
–
20
] o hype spec al da a [
21
–
23
] o es ima e
he biomass and canopy heigh o c ops.
Addi ionally, some s udies ha e in eg a ed d one-based 3D and spec al da a o es ima e c op
pa ame e s. Yue e al. [
24
] combined c op heigh in o ma ion and spec al da a om he Cube
UHD185 “Fi e ly” hype spec al snapsho senso (Cube GmbH, Ulm, Ge many) o es ima e he
biomass o win e whea , and Bendig e al. [
25
] combined d one-based 3D da a wi h g ound-measu ed
spec ome e da a o biomass moni o ing o ba ley. Many g ound-based combina ions o spec al and
3D da a ha e also been u ilized [
26
–
28
]. Howe e , only a ew s udies ha e in eg a ed d one-based
3D and spec al da a o he es ima ion o g ass quali y and quan i y. Ba e h e al. [
29
] s udied he
possibili y o u ilize d one-based canopy heigh models (CHMs) in he es ima ion o g ass heigh
and biomass using linea eg ession. Thei esul s showed ha he pho og amme ic CHM ga e
accu a e heigh es ima es in he la e phases o g ow h bu we e no accu a e a he beginning o he
g ow h when he canopy was s ill spa se. The VIs had he opposi e pe o mance; hey p o ided good
es ima es in he ea ly phases o g ow h bu sa u a ed a he la e s ages o g ow h wi h inc easing
plan heigh , which is also a known beha iou om ea lie s udies [
30
,
31
]. Based on hese indings,
Ba e h e al. [29]
de eloped a G assland Index (G assI) which combines he ad an ages o he CHM
and VIs based on RGB o p o ide accu a e es ima es o he en i e g ow h season. Possoch e al. [
32
]
u ilized a low-cos RGB d one sys em o calcula e he CHM and RGB VIs o es ima ing biomass o
Ag icul u e 2018,8, 70 3 o 28
g assland. Thei esul s indica ed ha he pho og amme ic CHM ga e he bes esul s o he d y
ma e yield (DMY) when using linea eg ession models.
CHMs o he biomass es ima ion can be c ea ed in di e en ways [
1
,
2
]. Pi man e al. [
8
]
ecommended measu ing he digi al e ain model (DTM) and digi al su ace model (DSM) using
emo e sensing echnologies, such as ul asonic o lase scanne . They s udied he es ima ion o he
biomass and canopy heigh o be mudag ass, al al a and he mix (which con ained a mix u e o bo h
be mudag ass and al al a) using a g ound-based mobile pla o m—a gol ca wi h ul asonic, lase
and spec al senso s. Thei compa ison o he pe o mance o single-senso and a combina ion o h ee
senso s indica ed ha he use o mul isenso y sys ems imp o ed he biomass es ima ion accu acy
o g asslands.
Se e al s udies ha e e alua ed di e en eg ession echniques o biomass es ima ion.
Ma abel e al. [33]
in es iga ed biomass es ima ion o g asslands using ield spec ome e da a.
They e alua ed pe o mance o he suppo ec o machine (SVM) and Pa ial Leas Squa es Reg ession
(PLSR). The mos accu a e model o p edic he o al biomass was ob ained using he PLSR and spec al
bands be ween 916–1120 nm and 1079–1297 nm. Yue e al. [
34
] compa ed eigh di e en eg ession
echniques o win e whea biomass using nea -su ace spec oscopy. The esul s o he s udy showed
ha PLSR and mul i a iable linea eg ession we e mos sui able when high-accu acy and s able
es ima es a e equi ed om ela i ely ew samples. In addi ion, Random Fo es (RF) in oduced by
B eiman e al. [
35
] is highly obus agains noise and is bes sui ed o deal wi h epea ed obse a ions
in ol ing emo e-sensing da a ha a e usually a ec ed by a mosphe e, clouds, obse a ion imes
and senso noise [
34
]. Addi ionally, RF’s ad an ages o e o he me hods, such as mul iple linea
eg ession (MLR) and A i icial Neu al Ne wo k (ANN), a e high p edic ion accu acy, ea u e selec ion
is unnecessa y and i is less sensi i e o o e i ing [
36
–
38
]. RF has shown compe i i e accu acy
compa ed o o he me hods es ima ing he biomass o o es s [
39
–
41
] and ag icul u e [
14
,
42
,
43
].
The majo i y o biomass es ima ion s udies ha e u ilized o he es ima o s such as linea models
and nea es neighbou app oaches [
3
,
41
]. Mos o he d one-based biomass es ima ion s udies ha e
been ca ied ou using linea models based on only a ew ea u es [
18
,
23
,
25
,
28
]. An RF was used
by
Liu e al. [43]
o es ima e he le el o ice seeds, by Li e al. [
14
] o es ima e he biomass o maize,
and by Tu ne e al. [
16
] o p edic An a c ic moss heal h. Howe e , no s udies ha e pe o med an RF
o de e mine he biomass o g asslands.
Ou objec i e was o de elop and assess a machine lea ning echnique o he es ima ion o
canopy heigh and biomass o g ass swa ds based on a d one-based mul ispec al pho og amme ic
app oach. In pa icula , ou objec i e was o s udy he po en ial o ul a-high esolu ion canopy heigh
models (CHMs) and ege a ion indices (VIs) ex ac ed om ed, g een and blue (RGB) and colou
in a ed (CIR) images. To gene a e high a ia ion in o ma ion o s udy swa ds, he Na u al Resou ces
Ins i u e o Finland (LUKE) es ablished in he Jokioinen es si e an expe imen using six di e en
ni ogen e ilize applica ion a es and ou ha es ing da es. We i s e alua ed he easibili y o
CHMs and VIs sepa a ely in he heigh and biomass es ima ion by using a simple linea eg ession
echnique. We hen e alua ed he pe o mance o he combina ion o a ious heigh ea u es and VIs
in g ass quan i y es ima ion using machine lea ning echniques based on MLR and RF.
2. Ma e ials and Me hods
2.1. S udy A ea and Re e ence Measu emen s
The expe imen was conduc ed a he LUKE esea ch a m, which is loca ed in he municipali y
o Jokioinen in sou hwes Finland (app oxima ely 60
◦
48
0
N, 23
◦
30
0
E) (Figu e 1). The expe imen
was es ablished on a second-yea silage p oduc ion ield. The g ass swa d was es ablished in 2015
in sp ing ba ley as a companion c op wi h a imo hy/meadow escue (Phleum p a ense and Fes uca
p a ensis) seed mix u e a a 25 kg ha
−1
sowing a e (65% imo hy and 35% meadow escue on a weigh
basis). The ow wid h was 12.5 cm. In 2016, he ield was managed as a silage p oduc ion swa d.
Ag icul u e 2018,8, 70 4 o 28
A uni o m and e en si e o he ield was selec ed o he expe imen . The soil ype a he es si e
was clay, and he soil e ili y alues we e as ollows: pH 6.2, 377 mg K L
−1
soil, 6.6 mg P L
−1
soil,
1101 mg Mg L
−1
soil, and 2580 mg Ca L
−1
soil. The size o he expe imen al a ea was app oxima ely
50 m by 20 m. The expe imen al se up was a spli plo design wi h ou eplica es. The e ilize
ea men was in he 24 main plo s (plo size 12 m
×
3 m), and he ha es ing ime was in he sub-plo .
The expe imen had a o al o 96 plo s. Fou eplica es, 6 ni ogen e ilize le els (0 kg/ha, 50 kg/ha,
75 kg/ha, 100 kg/ha, 125 kg/ha and 150 kg/ha), and ou ha es ing/measu ing da es (6 June, 15 June,
19 June and 28 June) we e used in he p ima y g ow h. The e ilize applica ion was ca ied ou on
10 May 2017 by an expe imen al su ace e ilize b oadcas e ( ailo -made model) wi h a wo king
wid h o 1.5 m. To main ain he swa d ee o weeds, a con ol sp aying by S a ane XL he bicide
(ac i e ing edien s: 100 g L
−1
lu oxypy + 2.5 g L
−1
lo asulam) a he a e o 1.5 L ha
−1
was ca ied
ou on 24 May by he a m scale sp aye Ha di win s eam 363 MA 1200 EEEC/5 15 HAL wi h a
12 m sp aying boom (HARDI INTERNATIONAL A/S, No e Alsle , Denma k). The bo de s be ween
he main plo s ( e ilize ea men s) we e egula ly cu by a lawn mowe o make he ha es ing
easy. The ne size o he ha es ed and d one-measu ed plo was 1.5 m by app oxima ely 2.6 m.
A e he ha es , he ac ual leng h o each ha es ed plo was measu ed, and he hec a e yield was
adjus ed acco dingly.
Ag icul u e 2018, 8, x FOR PEER REVIEW 5 o 27
we e aken pe plo , and he mean alue o he h ee measu emen s was used. Heigh s ick
measu emen s we e ca ied ou acco ding o Finnish guidelines o p oducing an es ima e o
biomass o a g ass swa d pa cel [44]. In ha me hod, o each measu emen a clus e o g ass ille s
and lea es is s aigh ened up and he a e age heigh o ha clus e is measu ed wi h a heigh s ick.
In his me hod, indi idual ille s ha a e highe han a e age a e igno ed in he measu emen .
We compa ed he pla e me e and he heigh s ick me hods in he i s da ase . We made i e
measu emen s in each e iliza ion le el plo s (12 m × 1.5 m) wi h he pla e me e and calcula ed
a e age heigh o each plo . The pla e me e ’s bo om s ick is placed on g ound le el bu wi hou
pene a ing he g ound, and he pla e pa ee alls down o he swa d. While he pla e pa is going
down, i comp esses he swa d down a ew cen ime es (Figu e 2b). The pla e me e wo ks well
when he swa d is dense and has a heigh o a ound 20–30 cm. On he o he hand, he pla e me e
s uggles o wo k when he swa d is high and has no la op s uc u e o when he g ass swa d is
spa se and i s g ow h is poo and non-uni o m. In his s udy, swa d heigh s we e 60–70 cm du ing
he las ha es ing da es, and he plo s wi hou ni ogen inpu we e spa se and had low heigh ,
which was no ideal o he pla e me e . Addi ionally, p e ious s udies ha e ecei ed wo se
co ela ions wi h biomass es ima ion wi h a pla e me e han he heigh s ick [8]. The compa isons
showed ha he heigh s ick p o ided sligh ly highe heigh alues han he pla e me e ; he a e age
di e ence was 1.46 cm and he RMSE was 1.82 cm.
Figu e 1. O hopho o mosaic om he Jokioinen es si e om 15 June. In he i s eplica e ( he
le mos column), he e ilize a es we e om 0 o 150 kg/ha (indica ed wi h N0 o N150) in o de
om op o bo om in his pho o. Ni ogen e ilize applica ion a es we e andomized in Replica es
1–4 (in Columns 2–4) as well as he loca ion o he ha es ing da e o e e ence ha es s.
(a)
(b)
Figu e 2. In si u measu emen s o g ass heigh by: (a) heigh s ick; and (b) pla e me e .
2.3. Remo e Sensing Da a Acquisi ion
Figu e 1.
O hopho o mosaic om he Jokioinen es si e om 15 June. In he i s eplica e ( he le mos
column), he e ilize a es we e om 0 o 150 kg/ha (indica ed wi h N0 o N150) in o de om op
o bo om in his pho o. Ni ogen e ilize applica ion a es we e andomized in Replica es 1–4 (in
Columns 2–4) as well as he loca ion o he ha es ing da e o e e ence ha es s.
Ha es ing was ca ied ou by a Hald up o age plo ha es e (Model GR, HALDRUP GmbH,
Ilsho en, Ge many). The wid h o he cu ing ba was 1.5 m. The s ubble heigh was om 6 cm o 7 cm.
The p ima y g ow h was ha es ed on ou da es in June: 6 June was a he e y ea ly de elopmen al
s age o silage ha es ing; 15 June was jus p io and 19 June was e y close o he a ge ed silage
p oduc ion de elopmen al s age, and heading was jus s a ing in he swa ds; and 28 June was clea ly
a e he desi ed silage ha es ing s age. The esh yield (FY) was measu ed by he Hald up o age
plo ha es e . Howe e , on he i s ha es da e, due o ope a ion ailu e in Hald up scale, he plo
yield was collec ed and measu ed by weigh indoo s. A sample was aken om each plo o d y
ma e yield (DMY) and quali y analyses. On he i s ha es da e, he whole ha es was aken,
and, on he la e ha es da es, a 1 kg FY sample was aken om he ha es . The samples we e
chopped in o 3–4 cm long pieces by a Win e s eige (Model Hege 44, Win e s eige AG, Ried, Aus ia)
sample choppe , and he DMY was de e mined a e 17 h d ying a 100
◦
C in o ced ai d ying o ens.
Ag icul u e 2018,8, 70 5 o 28
On he day be o e each ha es , he e e ence canopy heigh s (H
e
) we e measu ed wi h a heigh s ick
and wi h a heigh pla e on he i s da e.
Idea o he expe imen al se up was o gene a e g ea a ia ion o DMY, FY and ni ogen amoun
in he s udy swa d. Gene al guidelines o ni ogen e ilize applica ion a e is 100 kg/ha o he
p ima y g ow h in clay soil o comme cial g ass silage p oduc ion, and ha es ing is a ge ed a a
D- alue (o ganic ma e diges ibili y in d y ma e ) o a ound 690 g kg
−1
which occu ed on his ield
in Jokioinen a ound 19 June in 2017.
The s a o he g owing season was la e in 2017 (5 May), and he ea ly summe was cool. The mean
mon hly empe a u e was 8.9
◦
C in May and 12.9
◦
C in June 2017, while long- e m a e ages (1980–2010)
a e 9.8
◦
C and 14.0
◦
C, espec i ely. The ain all igu es we e 13 mm and 101 mm in May and June
2017, and 40 mm and 63 mm as long- e m a e ages (1980–2010), espec i ely. The de elopmen
o g ass swa d in he p ima y g ow h in Finland depends on he accumula ed empe a u e sum.
The accumula ed e ec i e empe a u e (abo e + 5
◦
C) on he ha es da es we e he ollowing. 6 June:
160 (long- e m a e age o he da e (LTA): 224); 15 June: 241 (LTA: 300); 19 June: 289 (LTA: 336);
and 28 June: 347 (LTA: 425).
The swa d wi h ze o ni ogen applica ion was spa se and weak pa icula ly on he i s obse a ion
da es. The highes ni ogen applica ion a es—125 and 150 kg ha
−1
—p oduced a dense swa d in
mid-June which was suscep ible o lodging, and his a ec ed he g ow h o he s and. In addi ion,
he highes ni ogen applica ion a es seemed o inc ease he sha e o meadow escue in he swa d,
pa icula ly on he la es ha es ing da es. O he wise, he swa ds we e p edominan ly o imo hy.
2.2. Re e ence Field Da a and Biomass Sampling
De ices such as he ising pla e me e , capaci ance me e , and me e s ick a e examples o de ices
used o physical measu emen s o ege a ion heigh and biomass es ima ion [
5
,
6
]. We used he heigh
s ick o measu e he H
e
o he g assplo s on all he da es (Figu e 2a). Th ee measu emen s we e aken
pe plo , and he mean alue o he h ee measu emen s was used. Heigh s ick measu emen s we e
ca ied ou acco ding o Finnish guidelines o p oducing an es ima e o biomass o a g ass swa d
pa cel [
44
]. In ha me hod, o each measu emen a clus e o g ass ille s and lea es is s aigh ened
up and he a e age heigh o ha clus e is measu ed wi h a heigh s ick. In his me hod, indi idual
ille s ha a e highe han a e age a e igno ed in he measu emen .
Ag icul u e 2018, 8, x FOR PEER REVIEW 5 o 27
we e aken pe plo , and he mean alue o he h ee measu emen s was used. Heigh s ick
measu emen s we e ca ied ou acco ding o Finnish guidelines o p oducing an es ima e o
biomass o a g ass swa d pa cel [44]. In ha me hod, o each measu emen a clus e o g ass ille s
and lea es is s aigh ened up and he a e age heigh o ha clus e is measu ed wi h a heigh s ick.
In his me hod, indi idual ille s ha a e highe han a e age a e igno ed in he measu emen .
We compa ed he pla e me e and he heigh s ick me hods in he i s da ase . We made i e
measu emen s in each e iliza ion le el plo s (12 m × 1.5 m) wi h he pla e me e and calcula ed
a e age heigh o each plo . The pla e me e ’s bo om s ick is placed on g ound le el bu wi hou
pene a ing he g ound, and he pla e pa ee alls down o he swa d. While he pla e pa is going
down, i comp esses he swa d down a ew cen ime es (Figu e 2b). The pla e me e wo ks well
when he swa d is dense and has a heigh o a ound 20–30 cm. On he o he hand, he pla e me e
s uggles o wo k when he swa d is high and has no la op s uc u e o when he g ass swa d is
spa se and i s g ow h is poo and non-uni o m. In his s udy, swa d heigh s we e 60–70 cm du ing
he las ha es ing da es, and he plo s wi hou ni ogen inpu we e spa se and had low heigh ,
which was no ideal o he pla e me e . Addi ionally, p e ious s udies ha e ecei ed wo se
co ela ions wi h biomass es ima ion wi h a pla e me e han he heigh s ick [8]. The compa isons
showed ha he heigh s ick p o ided sligh ly highe heigh alues han he pla e me e ; he a e age
di e ence was 1.46 cm and he RMSE was 1.82 cm.
Figu e 1. O hopho o mosaic om he Jokioinen es si e om 15 June. In he i s eplica e ( he
le mos column), he e ilize a es we e om 0 o 150 kg/ha (indica ed wi h N0 o N150) in o de
om op o bo om in his pho o. Ni ogen e ilize applica ion a es we e andomized in Replica es
1–4 (in Columns 2–4) as well as he loca ion o he ha es ing da e o e e ence ha es s.
(a)
(b)
Figu e 2. In si u measu emen s o g ass heigh by: (a) heigh s ick; and (b) pla e me e .
2.3. Remo e Sensing Da a Acquisi ion
Figu e 2. In si u measu emen s o g ass heigh by: (a) heigh s ick; and (b) pla e me e .
We compa ed he pla e me e and he heigh s ick me hods in he i s da ase . We made i e
measu emen s in each e iliza ion le el plo s (12 m
×
1.5 m) wi h he pla e me e and calcula ed
a e age heigh o each plo . The pla e me e ’s bo om s ick is placed on g ound le el bu wi hou
pene a ing he g ound, and he pla e pa ee alls down o he swa d. While he pla e pa is going
down, i comp esses he swa d down a ew cen ime es (Figu e 2b). The pla e me e wo ks well when
he swa d is dense and has a heigh o a ound 20–30 cm. On he o he hand, he pla e me e s uggles
o wo k when he swa d is high and has no la op s uc u e o when he g ass swa d is spa se and i s
Ag icul u e 2018,8, 70 6 o 28
g ow h is poo and non-uni o m. In his s udy, swa d heigh s we e 60–70 cm du ing he las ha es ing
da es, and he plo s wi hou ni ogen inpu we e spa se and had low heigh , which was no ideal
o he pla e me e . Addi ionally, p e ious s udies ha e ecei ed wo se co ela ions wi h biomass
es ima ion wi h a pla e me e han he heigh s ick [8]. The compa isons showed ha he heigh s ick
p o ided sligh ly highe heigh alues han he pla e me e ; he a e age di e ence was 1.46 cm and
he RMSE was 1.82 cm.
2.3. Remo e Sensing Da a Acquisi ion
The Finnish Geospa ial Resea ch Ins i u e’s (FGI’s) d one, a emo ely pilo ed ai c a sys em
(RPAS), was u ilized o collec ing he emo e sensing da ase s. The ame o he FGI d one was he
G yphon Dynamics quadcop e wi h de achable a ms, and i was equipped wi h Pixhawk au opilo
(Compu e Vision and Geome y Lab, Zu ich, Swi ze land) wi h A duPilo APM Cop e (Ve sion 3.4,
Open-sou ce, Raleigh, NC, USA) i mwa e [
45
]. Endu ance o he d one is app oxima ely 25 min wi h
a maximum payload o 2.5 kg. The d one was equipped wi h a posi ioning sys em consis ing o an
NV08C-CSM L1 Global Na iga ion Sa elli e Sys em (GNSS) ecei e (NVS Na iga ion Technologies
L d., Mon lingen, Swi ze land), a Vec o na VN-200 IMU (Vec o Na Technologies, Dallas, TX, USA)
and a Raspbe y Pi single-boa d compu e (Raspbe y Pi Founda ion, Camb idge, Uni ed Kingdom).
The d one was ca ying an RGB digi al came a, a Sony A7R (Sony Co po a ion, Mina o, Tokyo, Japan)
equipped wi h a Sony FE 35 mm /2.8 ZA Ca l Zeiss Sonna T* lens (Sony Co po a ion, Mina o,
Tokyo, Japan). Sony A7R has a 35.90 mm by 24.00 mm complemen a y me al-oxide semiconduc o
(CMOS) senso wi h 36.4 megapixels. The size o aw images is 7360 pixels
×
4910 pixels. The came a
is igge ed o cap u e images in wo-second in e als, and a GNSS ecei e is used o eco d he
exac ime o each igge ing pulse. Fu he mo e, we calcula ed Pos P ocessed Kinema ic (PPK)
GNSS posi ions o each came a using Na ional Land Su ey o Finland (NLS) RINEX se ice, which
o e s obse a ion da a om FinnRe s a ions [
46
], in RTKlib (RTKlib e sion 2.4.2, Open-sou ce,
Raleigh, NC, USA) so wa e kpos ool [
47
]. A hype spec al came a based on a uneable Fab y
Pé o in e e ome e (FPI) ope a ing in he isible o nea -in a ed spec al ange (500–900 nm) (VTT
Technical Resea ch Cen e o Finland L d, Espoo, Finland) [
22
] was used o collec he spec al da a
cubes o each ligh . The FPI came a is a ligh weigh , ame o ma hype spec al image ope a ing in
he ime-sequen ial p inciple collec ing spec al bands wi h 648 by 1024 pixels. In his s udy, we used
i in a mul ispec al mode o p o ide mul ispec al bands in ed (cen al wa eleng h L0 = 669.0 nm;
ull wid h a hal maximum (FWHM) o 27.0 nm) and he nea in a ed (NIR; L0 = 804.1 nm, FWHM:
28.3 nm) spec al ange.
The ligh pa ame e s and condi ions a e in oduced in Table 1. We used lying heigh s o 30 m
and 50 m and a lying speed o 2 m/s. The g ound sampling dis ances (GSD) we e 3.9 mm and 6.4 mm
o he RGB images and 30 mm and 50 mm o he FPI images wi h he lying heigh s o 30 m and 50 m,
espec i ely. These se ings esul ed in 84–87% and 65–81% o wa d and side o e laps, espec i ely,
o he RGB and FPI images, which a e sui able o he pho og amme ic p ocessing o hese scenes.
In he esul ing pho og amme ic blocks, he a ea o in e es was cap u ed in he block J2_F1 in mo e
han six images and in o he blocks in mo e han nine images. The eason o he lowe numbe s o
o e lapping images in he block J2_F1 was igge ing p oblems o he RGB-came a, howe e , as he
o e laps we e app op ia e, we decided o use his da a.
Fi e pe manen g ound e e ence poin s we e a ge ed in he co ne s and cen e o he block o
be used as he g ound con ol poin s (GCPs) and checkpoin s (CP). The a ge s we e black-pain ed
plywood boa ds o size 0.5 m by 0.5 m wi h a whi e pain ed ci cle wi h a diame e o 0.3 m and hey
we e moun ed on wooden pilla s. The e e ence poin s we e measu ed wi h he T imble R10 RTK
DGNSS (T imble Inc., Sunny ale, CA, USA) wi h an accu acy wi hin 0.03 m ho izon ally and 0.04 m
e ically [
48
,
49
]. Addi ionally, h ee e lec ance panels wi h nominal e lec ance o 0.03, 0.09 and 0.50
we e ins alled in he a ea o enable ans o ma ion o he image digi al numbe alues o e lec ance.
Ag icul u e 2018,8, 70 7 o 28
Table 1.
Da ase s wi h hei collec ion da e, ime, cloud condi ions, sun azimu h, and sola ele a ion.
GNSS: global na iga ion sa elli e sys em; FH: ligh heigh .
Da ase Da e Time (GNSS) Cloud Condi ions Sun Azimu h (◦) Sola Ele a ion (◦) FH (m)
J1_F1 6 June 11:49 o 11:59 Va ying 212.51 48.80 50
J1_F2 6 June 11:59 o 12:08 Va ying 215.83 48.12 30
J2_F1 15 June 08:59 o 09:14 Sunny 150.21 49.96 30
J3_F1 19 June 9:09 o 09:26 Va ying 153.34 50.58 50
J4_F1 28 June 07:13 o 07:29 Sunny 117.47 40.45 50
2.4. Remo e Sensing Da a P ocessing
Agiso Pho oscan P o essional ( e sion 1.3.4, Agiso , S . Pe e sbu g, Russia) so wa e was used
o he pho og amme ic p ocessing [
50
]. We ollowed simila p ocessing wo k low in Pho oscan
as in oduced in p e ious s udies by se e al au ho s [
3
,
51
–
53
]. In he i s s age, Pho oscan uses
SFM o de e mine he in e io (IOP) and ex e io o ien a ion pa ame e s (EOP) o each image and
o calcula e a spa se poin cloud. We used he sel -calib a ing op ion and included he ocal leng h,
p incipal poin coo dina es, and adial and angen ial lens dis o ions. In he o ien a ion p ocessing,
he high quali y se ing was selec ed wi h 40,000 key poin s and 4000 ie poin s pe image. In addi ion,
we used PPK p ocessed GNSS coo dina es, o each image o p eselec image pai s o he o ien a ion
p ocess. The e e ence poin s (GCPs) we e measu ed on he images manually; each o he GCPs
we e measu ed in en o mo e images. We used he accu acy se ings o
±
0.005 m o he GCPs and
±
5 m o he came a posi ions coo dina es o he images. All IOPs, EOPs and poin coo dina es
we e op imized using “op imize came a alignmen ” ool in Pho oscan [
54
]. A e he op imiza ion,
an au oma ic ou lie emo al was pe o med using he g adual selec ion ools o he so wa e based
on he e-p ojec ion e o and econs uc ion unce ain y. Addi ionally, some poin s we e manually
emo ed om he spa se poin cloud, pa icula ly poin s unde g ound and up in he ai . O e all,
app oxima ely 10% o he wo s poin s we e emo ed du ing he g adual selec ion and manual poin
emo ing o each da ase . Then he inal op imiza ion o he spa se poin cloud, IOPs and EOPs was
ca ied ou . Nex , dense poin cloud gene a ion was ca ied ou using he high quali y pa ame e and
mild dep h il e ing. Acco ding o ou es ings and p e ious s udies [
3
,
51
–
53
], hese pa ame e s a e
sui able o la a eas such as g ass ields o p o ide accu a e esul s. The coo dina e sys em in he
p ocessing was he ETRS89-TM35FIN.
E en hough all da ase s we e p ocessed using he same pa ame e s in Pho oscan, small
di e ences in he lying heigh and o e laps be ween images esul ed in sligh ly di e en p ocessing
esul s (Table 2). In pa icula , he lowe lying heigh esul ed in a smalle GSD and a highe poin
densi y. Addi ionally, he e-p ojec ion e o s we e smalle o he 30 m ligh s (0.534–0.589) han o
he 50 m ligh s (0.783–1.25). Roo Mean Squa e E o s (RMSEs) o block adjus men s we e calcula ed
using a lea e-one-ou c oss- alida ion (LOOCV) me hod. The LOOCV was ca ied ou by pe o ming
he block adjus men i e imes by using ou e e ence poin s as GCPs and one e e ence poin as
an independen CP. The e o be ween he adjus ed coo dina e and he e e ence coo dina e o he
CP was calcula ed in each adjus men and inally he LOOCV RMSE was calcula ed using e o s
o each CP [
55
]. The RMSEs we e 0.5–2.4 cm in he X- and Y-coo dina es, and 1.0–4.8 cm in heigh .
These esul s indica ed ha he block adjus men s we e o good accu acy and he blocks we e no
de o med. The ligh s om a 30 m lying heigh p o ided sligh ly be e 3D RMSE (2.7–2.9 cm) han
he 50 m lying heigh (2.8–5.0 cm).
Ag icul u e 2018,8, 70 8 o 28
Table 2.
Da ase pa ame e s: Da e, FH (Fligh Heigh ), N Images (Numbe o Images), e-p ojec ion
e o , poin densi y, and RMSE (Roo Mean Squa e E o ) o X, Y and Z coo dina es and 3D.
Da ase Da e FH N Re-P ojec ion Poin Densi y RMSE (cm)
(m) Images E o (pix) Poin s/m2X Y Z 3D
J1_F1 6 June 50 156 0.783 5920 1.1 0.6 4.8 5.0
J1_F2 6 June 30 171 0.534 14,600 1.0 1.1 2.5 2.9
J2_F1 15 June 30 174 0.589 17,100 0.5 0.7 2.6 2.7
J3_F1 19 June 50 320 1.12 5860 1.0 2.4 1.0 2.8
J4_F1 28 June 50 350 1.25 5230 0.6 0.9 3.7 3.9
The RGB o homosaics we e calcula ed wi h a GSD o 1.0 cm in he Pho oscan using he
o homosaic blending mode. The o homosaics o he FPI ed and NIR bands we e calcula ed wi h a
GSD o 5 cm using FGI in-house C++ so wa e [
22
]. In his case, he o homosaics we e c ea ed using
he mos nadi image pa s, and no blending was pe o med. The o homosaic DNs we e no malized
o e lec ance alues using he empi ical line me hod [
56
] using he h ee e lec ance panels. We used
exponen ial unc ion o he RGB da ase and linea unc ion o he FPI da ase .
We used bo h au oma ic and manual app oaches o c ea e he DTM in he Pho oscan. The DTM
au o
was gene a ed using Pho oscan’s au oma ic g ound poin classi ica ion p ocedu e. In his p ocedu e,
he dense cloud was i s di ided in o cells o a ce ain size, and he lowes poin o each cell was
de ec ed. The i s app oxima ion o he DTM was calcula ed using hese poin s. A e ha , all poin s
o he dense cloud we e checked, and a new poin was added o he g ound class i he poin was
wi hin a selec ed dis ance om he e ain model and i he angle be ween app oxima ion o he
DTM and he line o connec he new poin on i we e less han he selec ed angle [
3
,
52
,
54
]. Based
on ou p elimina y es ing we selec ed s a ing pa ame e s o au oma ic classi ica ion o g ound
poin s and i e a i ely selec ed he mos sui able pa ame e s o he ecosys em o his s udy by isually
compa ing he classi ica ion esul s o he RGB o homosaics. Fo all he da ase s, a cell size o 3 m,
a maximum angle o 0.5 deg ee and a maximum dis ance o 2.0 cm we e selec ed. These pa ame e s a e
sligh ly di e en om he pa ame e s selec ed by o he au ho s in di e en ecosys ems, o example,
Cunli e e al. [3]
used o g ass-domina ed-sh ubs ecosys ems he cell-size o 3 m, maximum angle o
3
◦
and maximum dis ance o 5.0 cm, and Méndez-Ba oso e al. [
57
] used o classi ying g ound poin s
in medium densi y o es si es he cell-size o 10 m, maximum angle o 3
◦
and maximum dis ance o
10 cm. The DTM
manual
was gene a ed using Pho oscan’s ee- o m-selec ion ool o manually selec
and classi y g ound poin s. The classi ied g ound poin s we e used o in e pola e DTM o he whole
a ea. The DSM, DTM
manual
and DTM
au o
we e expo ed as TIFF images wi h a 1.0 cm esolu ion.
Finally, we calcula ed he CHM o bo h DTM
manual
and DTM
au o
by sub ac ing DTM om DSM o
each da ase , using QGIS ( e sion 2.18.14, Open-sou ce, Raleigh, NC, USA) so wa e.
2.5. Fea u e Ex ac ion om he Remo e Sensing Da ase s
2.5.1. Heigh Fea u es
We c ea ed shape iles o each plo using a ma gin o 0.25 m o he plo bo de o exclude possible
bo de e ec s. Then, using he CHM and he bo de shape ile, we calcula ed he heigh ea u es,
including a e age heigh (H
mean
), median heigh (H
median
), minimum heigh (H
min
), maximum heigh
(H
max
), heigh s anda d de ia ion (H
s d
), heigh 50% pe cen ile (H
p50
), heigh 70% pe cen ile (H
p70
),
heigh 80% pe cen ile (H
p80
) and heigh 90% pe cen ile (H
p90
) (Table 3), o each plo using Ma lab
( e sion 2016b, Ma hWo ks, Na ick, MA, USA) so wa e.
Ag icul u e 2018,8, 70 9 o 28
Table 3.
De ini ions and o mulas o CHM me ics in his s udy. h
i
is he heigh o he i h heigh alue,
N is he o al numbe o heigh alues in he plo , Z is he alue om he s anda d no mal dis ibu ion
o he desi ed pe cen ile (1.282, he 90 h pe cen ile) and σis he s anda d de ia ion o he a iable.
Index Name Equa ion
Mean heigh Hmean 1
N(N
∑
i=1
hi)
Median heigh Hmedian median(hi), 1 ≤i≤N
Minimum heigh Hmin min(hi), 1 ≤i≤N
Maximum heigh Hmax max(hi), 1 ≤i≤N
S anda d de ia ion heigh Hs d sN
∑
i=1
(hi−1
N
N
∑
i=1
hi)2
N−1
90 h pe cen ile Hp90 1
N
N
∑
i=1
hi+Zσ
2.5.2. Vege a ion Indices
We calcula ed he VIs om he o homosaics (and, in some cases, also u ilizing he CHM) using
QGIS ( e sion 2.18.14, Open-sou ce, Raleigh, NC, USA) so wa e. The polygonal shape ile o each plo
was used o ex ac digi al numbe s (DN) om he ed, g een, blue and NIR bands and he CHM. Then
mean alues o each plo we e calcula ed by “zonal s a is ics” implemen a ion in QGIS. The mean
alues we e used as inpu alues in he VI equa ions shown in Table 4.
We also in oduced a new VI o g ass ields, he ExG + CHM. The idea o he index is simila o he
G assI-Index, which aims o compensa e o he weaknesses o CHM and VIs a di e en g ow h s ages
o he g ass [
29
]. The di e ence is ha we used he ExG, while he G assI is based on he RGBVI [
25
].
The ExG was in oduced by Woebbecke e al. [
58
], and i is commonly used o ege a ion g eenness
iden i ica ion. I has been widely used in di e en s udies, such as maize biomass es ima ion [
14
] and
ege a ion ac ion mapping o whea [59].
Table 4.
Vege a ion index (VI) abb e ia ion, name, o mula and e e ence. * g=G/(R+G+B),
=R/(R+G+B).
VI Name Equa ion Re e ence
GRVI G een Red Vege a ion Index RG−RR
RG+RRTucke [60]
MGRVI Modi ied G een Red Vege a ion Index (RG)2−(RR)2
(RG)2+(RR)2Bendig e al. [13]
RGBVI Red G een Blue Vege a ion Index (RG)2−(RB×RR)
(RG)2+(RB×RR)Bendig e al. [25]
ExG Excess G een Index 2 ×g× −bWoebbecke e al. [58]
ExR Excess Red Index 1.4 × −bMeye e al. [61]
ExGR Excess G eenRed Index ExG −ExR Ne o [62]
G assI G assland Index RGBVI +CHM Ba e h e al. [29]
ExG + CHM Excess G een combined wi h CHM ExG +CHM In oduced he e
NDVI No malized Di e ence Vege a ion Index R800−R670
R800+R670 Rouse e al. [63]
RVI Ra io Vege a ion Index R800
R670
Pea son & Mille [
64
]
MSAVI Modi ied Soil Adjus ed Vege a ion Index (2×R800+1−√2×R800+1)2−8×(R800−R670)
2Qi e al. [65]
OSAVI
Op imiza ion o Soil Adjus ed Vege a ion Index
1.16×(R800−R670)
R800+R670+0.16 Rondeaux e al. [66]
Ag icul u e 2018,8, 70 16 o 28
Ag icul u e 2018, 8, x FOR PEER REVIEW 15 o 27
(a)
(b)
Figu e 6. Simple linea eg ession o (a) MSAVI and d y ma e yield (DMY) and (b) ExG + Hp90 and
DMY o di e en ime se ies.
Table 9. Pea son co ela ion coe icien s o VIs and DMY, FY and H e on di e en da es and
di e en Ni ogen e ilize le els (0–150 kg/ha). DMY: d y ma e yield; FY: esh yield; and H e :
e e ence heigh measu emen .
Da e
N-Le el (kg/ha)
6 June
15 June
19 June
28 June
0
50
75
100
125
150
DMY
MSAVI
0.95
0.94
0.96
0.95
0.62
0.91
0.90
0.91
0.87
0.94
NDVI
0.92
0.94
0.94
0.89
0.75
0.95
0.94
0.95
0.88
0.91
ExG
0.77
0.75
0.87
0.89
0.75
0.88
0.68
0.84
0.85
0.90
ExG + Hp90
0.91
0.94
0.96
0.90
0.80
0.92
0.96
0.96
0.93
0.88
G assIp90
0.88
0.91
0.96
0.90
0.87
0.89
0.95
0.94
0.92
0.89
FY
MSAVI
0.96
0.95
0.97
0.99
0.59
0.95
0.87
0.90
0.94
0.92
NDVI
0.94
0.92
0.92
0.81
0.73
0.95
0.92
0.94
0.96
0.91
ExG
0.82
0.71
0.84
0.89
0.77
0.94
0.79
0.89
0.92
0.91
ExG + Hp90
0.95
0.92
0.94
0.83
0.75
0.98
0.99
0.98
0.98
0.91
G assIp90
0.93
0.88
0.92
0.85
0.87
0.98
0.99
0.98
0.98
0.94
H e
MSAVI
0.85
0.94
0.93
0.81
0.71
0.89
0.89
0.85
0.88
0.87
NDVI
0.86
0.97
0.94
0.88
0.77
0.93
0.95
0.92
0.89
0.86
ExG
0.72
0.85
0.89
0.76
0.75
0.88
0.71
0.87
0.79
0.85
ExG + Hp90
0.84
0.97
0.96
0.89
0.76
0.93
0.99
0.97
0.88
0.94
G assIp90
0.82
0.96
0.96
0.89
0.80
0.93
0.98
0.97
0.89
0.95
3.4. Biomass Es ima ion Using MLR and RF
We used he MLR o es ima e he DMY and he FY using he RGB, he VI and he 3D ea u es
sepa a ely and in di e en combina ions (Table 10). The bes esul s when using he RGB, he VI o
he 3D ea u es sepa a ely we e ob ained wi h he VI ea u es: he PCC and RMSE we e 0.96 and 0.44
/ha (16.7%) o he DMY and 0.91 and 2.94 /ha (26.6%) o he FY, espec i ely. Using he 3D
ea u es p o ided sligh ly wo se esul s: he PCC and RMSE we e 0.93 and 0.59 /ha (22.4%) o he
DMY, and 0.92 and 2.79 /ha (25.27%) o he FY, espec i ely. The RGB ea u es p o ided clea ly he
wo s esul s. Combining he 3D and VI ea u es and all he ea u es (3D, VI, and RGB) p o ided he
Figu e 6.
Simple linea eg ession o (
a
) MSAVI and d y ma e yield (DMY) and (
b
) ExG + H
p90
and
DMY o di e en ime se ies.
3.4. Biomass Es ima ion Using MLR and RF
We used he MLR o es ima e he DMY and he FY using he RGB, he VI and he 3D ea u es
sepa a ely and in di e en combina ions (Table 10). The bes esul s when using he RGB, he VI o
he 3D ea u es sepa a ely we e ob ained wi h he VI ea u es: he PCC and RMSE we e 0.96 and
0.44 /ha (16.7%) o he DMY and 0.91 and 2.94 /ha (26.6%) o he FY, espec i ely. Using he 3D
ea u es p o ided sligh ly wo se esul s: he PCC and RMSE we e 0.93 and 0.59 /ha (22.4%) o he
DMY, and 0.92 and 2.79 /ha (25.27%) o he FY, espec i ely. The RGB ea u es p o ided clea ly he
wo s esul s. Combining he 3D and VI ea u es and all he ea u es (3D, VI, and RGB) p o ided he
bes esul s; o example, in he case wi h all he ea u es, he PCC and RMSE we e 0.98 and 0.34 /ha
(12.7%) o he DMY and 0.98 and 1.25 /ha (11.4%) o he FY, espec i ely.
Table 10.
Mul ilinea eg ession (MLR) and Random Fo es (RF) classi ica ion esul s. PCC: Co ela ion
coe icien s; RMSE: Roo Mean Squa e E o ; NRMSE: No malized Roo Mean Squa e E o ; DMY: d y
ma e yield; FY: esh yield; RGB: Red, G een and Blue spec al ea u es; VI: Vege a ion Index ea u es;
3D: CHM 3D ea u es.
DMY FY
PCC RMSE ( /ha) NRMSE (%) PCC RMSE ( /ha) NRMSE (%)
MLR
RGB 0.65 1.19 44.82 0.66 5.17 46.83
VI 0.96 0.44 16.71 0.91 2.94 26.62
3D 0.93 0.59 22.37 0.92 2.79 25.27
RGB + VI 0.96 0.45 17.03 0.94 2.47 22.34
RGB + 3D 0.97 0.40 15.22 0.95 2.14 19.41
VI + 3D 0.98 0.34 12.94 0.98 1.22 11.05
RGB + VI + 3D
0.98 0.34 12.70 0.98 1.25 11.35
RF
RGB 0.77 1.00 37.65 0.79 4.22 38.19
VI 0.96 0.46 17.37 0.97 1.67 15.13
3D 0.93 0.56 21.16 0.93 2.58 23.34
RGB + VI 0.96 0.42 15.94 0.97 1.63 14.78
RGB + 3D 0.96 0.43 16.34 0.97 1.80 16.32
VI + 3D 0.97 0.37 14.06 0.98 1.51 13.66
RGB + VI + 3D
0.97 0.40 15.12 0.98 1.49 13.49
Ag icul u e 2018,8, 70 17 o 28
We pe o med a simila analysis using he RF es ima o (Table 10). The bes esul s when using
he RGB, he VI o he 3D ea u es indi idually we e ob ained using he VIs: he PCC and RMSE
we e 0.96 and 0.46 /ha (17.4%) o he DMY and 0.97 and 1.67 /ha (15.1%) o he FY, espec i ely.
In addi ion, in he case o he RF, he 3D ea u es p o ided sligh ly wo se esul s han he VI ea u es
and he RGB ea u es p o ided he wo s esul s. Simila o MLR, using he combina ions o he 3D
and VI ea u es and all he ea u es, p o ided he bes esul s; o example, he 3D and VI ea u es
p o ided he PCC and RMSE 0.97 and 0.37 /ha (14.1%) o he DMY, and 0.98 and 1.51 /ha (13.7%)
o he FY, espec i ely. When compa ing he wo es ima o s, he RF es ima o p o ided be e esul s
han he MLR when using indi idual RGB and VI ea u es, and hei combina ions; he esul s wi h he
VI and 3D ea u es we e on he same le el. The RF es ima o p o ided sligh ly wo se esul s han he
MLR when combining all he ea u es. Rega ding he impo ance o di e en ea u es, he 3D ea u es
we e he mos impo an i hey we e included in he ea u e combina ions: he pe cen ile heigh s and
ExG + CHM we e he mos impo an o he DMY and ExG + CHM, G assI and he pe cen ile heigh s
o he FY (Table 11 and Appendix B, Tables A5 and A6). Thus, he new ExG + CHM index appea ed
o be signi ican in g ass biomass es ima ions.
Table 11.
The mos impo an ea u es o he Random Fo es (RF) (in he o de o impo ance). DMY:
d y ma e yield; FY: esh yield; RGB: Red, G een and Blue spec al ea u es; VI: Vege a ion Index
ea u es; 3D: CHM 3D ea u es.
Case DMY FY
Fea u es
RGB B, R, G B, R, G
VI RVI, OSAVI, NDVI, MGRVI, ExG,
MSAVI, ExGR, ExR, RGBVI, GRVI
RVI, NDVI, MGRVI, OSAVI, ExG, MSAVI,
RGBVI, ExGR, GRVI, ExR
3D Hp90, Hp80, Hp70, Hmin, Hmax, Hp50,
Hmean, Hmedian, Hs d
Hp70, Hp80, Hp90, Hmax, Hp50, Hmin, Hmean,
Hmedian, Hs d
RGB + VI NDVI, RVI, OSAVI, MGRVI, ExG,
MSAVI, B, RGBVI, GRVI, ExR
RVI, OSAVI, NDVI, ExG, MSAVI, B, ExGR,
RGBVI, G
RGB + 3D Hp90, Hp80, Hmin, Hp70, Hp50, Hmean,
Hmedian, Hmax, G, R
Hp90, Hp80, Hp70, Hp50, Hmax, Hmean, Hmedian,
Hmin, G, B
VI + 3D Hp90, Hmin, G assImax, Hp80, G assIp90,
Hmean, Hp50, Hp70, Hmax, ExG + Hp90
ExG + Hmax, ExG + Hp90, G assImax, G assIp90,
Hp90, Hp80, Hp70, Hmax, Hp50, Hmeadian
RGB + VI + 3D Hp90, Hmin, Hp70, Hp80, Hmean, Hmax,
H
p50
, H
median
, ExG + H
max
, ExG + H
p90
ExG + Hmax, ExG + Hp90, G assImax, Hp90,
Hmin, Hmean, Hp70, G assIp90, Hp80, Hp50
We analysed he sensi i i y o he biomass es ima ion wi h espec o he da e and ni ogen
applica ion a e using he RF es ima o wi h he combina ion o 3D, VI and RGB ea u es. Simila
o he analysis wi h he linea eg ession in Sec ion 3.2, bo h he da e and he ni ogen e iliza ion
le el had an impac on he esul s (Table 12). When conce ning he impac o he measu emen da e,
he esul s we e he wo s o he i s da e: he PCC and RMSE we e 0.91 and 0.15 /ha (13.1%) o
he DMY, and 0.95 and 0.48 /ha (12.1%) o he FY, espec i ely. The bes pe o mance was ob ained
du ing he hi d measu emen da e: he PCC and RMSE we e 0.97 and 0.27 /ha (9.1%) o he DMY,
and 0.98 and 1.30 (10.2%) o he FY, espec i ely. The esul s o he second and las measu emen da es
we e almos as good. Rega ding he impac o he ni ogen applica ion a e, he bes pe o mance o
he DMY was achie ed wi h he applica ion a e o 100 kg/ha, which p o ided he PCC and RMSE
o 0.98 and 0.35 /ha (11.0%), espec i ely. In he case o he FY, he ni ogen applica ion a es o
50–125 kg/ha pe o med e enly well gi ing he PCC and he RMSE o 0.97–0.98 and 0.95–1.71 /ha
(11.3–12.0%), espec i ely. A ni ogen applica ion a e o 150 kg/ha p o ided sligh ly wo se es ima ion
accu acy, and he ni ogen le el 0 kg/ha p o ide clea ly he poo es es ima ion accu acy. In he i s
measu emen da e, he mos impo an ea u es we e ExG + H
p90
, ExG + H
max
, G assI
p90
, G assI
max
Ag icul u e 2018,8, 70 18 o 28
and NIR-based VIs; less impo an ea u es we e he 3D ea u es (Appendix B, Table A6). Du ing
he second and hi d measu emen da es, he impo ance o 3D ea u es inc eased, e en hough he
mos impo an ea u es we e s ill ExG + H
p90
, ExG + H
max
and G assI
p90
. On he las measu emen
da e, he VI ea u es s a ed o domina e: he mos impo an ea u es we e NDVI, OSAVI and MSAVI.
Be ween di e en ni ogen e iliza ion le els, he selec ed ea u es had less a ia ion han di e en
da es (Appendix B, Table A6).
Table 12.
Random Fo es (RF) classi ica ion esul s o di e en da es and ni ogen e ilize le els. PCC:
Co ela ion coe icien s; RMSE: Roo Mean Squa e E o ; NRMSE: No malized Roo Mean Squa e E o ;
DMY: d y ma e yield; FY: esh yield; RGB: Red, G een and Blue spec al ea u es; VI: Vege a ion
Index ea u es; 3D: CHM 3D ea u es; 0–150: Ni ogen e ilize le els 0–150 kg/ha.
DMY FY
PCC RMSE ( /ha) NRMSE (%) PCC RMSE ( /ha) NRMSE (%)
Da e
6 June 0.91 0.15 13.09 0.95 0.48 12.13
15 June 0.95 0.24 10.44 0.95 1.27 11.98
19 June 0.97 0.27 9.13 0.98 1.30 10.22
28 June 0.94 0.52 12.11 0.97 1.92 11.30
Ni ogen
0 0.75 0.19 24.41 0.73 0.71 27.78
50 0.93 0.41 17.88 0.97 0.95 11.37
75 0.95 0.41 14.61 0.97 1.35 11.99
100 0.98 0.35 10.99 0.97 1.48 11.25
125 0.98 0.42 12.45 0.98 1.71 11.55
150 0.93 0.64 18.36 0.96 2.39 14.86
4. Discussion
We de eloped and assessed a no el d one-based machine lea ning echnique o es ima ing he
heigh , esh yield (FY) and d y ma e yield (DMY) o g ass swa ds. Ou app oach was o de i e
a ious ea u es om he mul ispec al pho og amme ic da a se s, including he heigh ea u es om
he CHM and he colou s and di e en VIs om o hopho os in he ed (R), g een (G), blue (B) and
nea -in a ed (NIR) spec al bands. The MLR and RF es ima o s we e ained using high a ia ion
imo hy/meadow escues mix u e swa ds domina ed by imo hy.
Ou s udy was he i s o in eg a e a ious s uc u al and spec al ea u es om a d one
mul ispec al pho og amme ic sys em using machine lea ning echniques o he g ass swa d biomass
es ima ion in he con ex o silage p oduc ion. The bes esul s we e ob ained when combining di e en
heigh , RGB and VI ea u es. The co ela ions and RMSEs we e a bes 0.98 and 0.34 /ha (12.7%)
o he DMY and 0.98 and 1.22 /ha (11.05%) o he FY, espec i ely. The MLR and RF p o ided
qui e simila esul s (Table 10). O e all, he mos impo an ea u es o he RF we e he new indices
ExG + H
p90
and ExG + H
max
(in oduced in his s udy), he heigh ea u es and he G assI (Table 11
and Appendix B, Table A5). Addi ionally, he CHM- ea u es ga e be e co ela ions o he DMY
and he FY han he eg essions wi h he physical canopy heigh measu emen s by he heigh s ick
(H
e
) a he h ee g ow h s ages o he silage swa d s udied esul ing om he h ee i s measu emen
da es. We ob ained he bes esul s on he a ge ed silage ha es ing da e (19 June) and jus be o e
ha (15 June) when he canopy was well-g own and homogeneous. Poo e es ima ion esul s we e
ob ained ea ly in he g owing season when he swa d olume and densi y was low, as well as a e he
a ge ed silage ha es ing da e when he s and was al eady heading, and lodging occu ed in he mos
hea ily e ilized plo s.
In he ial a ea, we gene a ed he DTM u ilizing he pa hs ha we e cu down be ween he
sample plo s (Figu e 1). We e alua ed he pe o mance o manual and semi-au oma ic DTM gene a ion
Ag icul u e 2018,8, 70 19 o 28
app oaches. In he case o he manual DTM, we classi ied all poin s o he pa hs o g ound poin s.
The semi-au oma ic me hod also classi ied mos o hese same poin s as g ound poin s. Bo h me hods
p o ided simila DTMs and co ela ions o he biomasses. The esul s indica ed ha i was possible o
gene a e an accu a e CHM om a single ligh ’s da a wi hou he need o collec DTM da a sepa a ely
be o e o a e ha es ing. The second da e’s au oma ic DTM’s wo se RMSE o 6.80 cm was caused by
he lowe numbe o o e lapping images in some pa s o he model ha was due o some came a
igge ing p oblems du ing he ligh . Ou plo s had he cu pa hs (Figu e 3), which p obably imp o ed
g ound poin de ec ion in all cases. Howe e , he g ass heigh in cu pa hs was 6–7 cm on each da e,
in o which he pho og amme ic DSM could no pene a e well. Because o his, he in e pola ed DTMs
we e abo e he eal g ound le el. The e o e, he canopy heigh alues based on he CHM we e lowe
han he physical heigh measu emen s. The limi a ion o pho og amme ic DSM has been disco e ed
in ea lie s udies [
32
,
52
]. One solu ion o he unde es ima ion could be o use oblique images o
combine oblique and nadi images in he DSM gene a ion as sugges ed in p e ious s udies [
3
,
68
].
Accu a e geo e e encing and a non-de o med pho og amme ic block a e impo an equi emen s
in he p oposed me hod. We pe o med he geo e e encing using GCPs, which can be conside ed
a labo ious app oach when aiming o ully au oma ic p ocedu es and also expensi e equipmen
o pe o ming ield measu emen s a e equi ed. In u u e s udies, ou objec i e will be o e alua e
app oaches ha do no equi e in si u GCPs, pa icula ly ou objec i e is o implemen be e di ec
geo e e encing p ocess [
69
] u ilizing a mo e accu a e L1/L2 GNSS/IMU ecei e and o in es iga e
ela i e geo e e encing o he mul i empo al da ase s. This imp o emen would also educe he o e all
cos o he comple e measu emen sys em. Ou esul s also suppo ed he gene al expec a ions ha
he la ge image o wa d and side o e laps o app oxima ely 80% and he sel -calib a ion du ing he
pho og amme ic p ocessing p o ided a non-de o med pho og amme ic block.
In he simple linea eg essions, he VIs pe o med well. The bes pe o ming indi idual VIs
we e he MSAVI o he DMY and FY (PCC: 0.94–0.99 o di e en da es) and ExG + H
p90
o he
H
e
(PCC: 0.84–0.97 o di e en da es) (Table 9). O e all, he NIR-based VIs ga e be e esul s han
he RGB-based VIs. Se e al s udies ha e shown ha he NIR spec al ange pe o ms be e in he
es ima ion o c op biomass han he RGB spec al ange [
25
,
28
]. Howe e , he new ExG + CHM indices,
sugges ed by us, p o ided good esul s and ou pe o med all he RGB-based VIs in eg essions wi h
all o he physical measu emen s, and in addi ion ou pe o med he bes pe o ming VI (MSAVI) in
he eg essions wi h he DMY and he H
e
. Simila ly, in o he s udies, combined CHM and RGB
based VIs ha e p o ided be e eg essions han he CHM and he RGB based VIs sepa a ely [
29
,
32
].
Wi h inc easing g ass heigh , he RGB-based VIs we e sa u a ing, which is consis en wi h p e ious
s udies [
18
,
32
]. Howe e , he sa u a ion had less impac on he new ExG + CHM indices and he
G assI han o he RGB-based VIs; and e en less impac on he NIR-based VIs (Table 9and Figu e 6a).
All o he CHM ea u es pe o med o e all be e han he VIs; H
p90
being he bes pe o ming
CHM ea u e (Appendix A, Table A4). The simple linea eg ession esul s o Possoch e al. [
32
]
indica ed also ha he g ass heigh was a sui able ea u e o g ass yield es ima ion; eg ession o
CHM o DMY p o ided a co ela ion o 0.80, and he physically measu ed heigh o he DMY p o ided
he sligh ly wo se co ela ion o 0.79. P e ious s udies ha e p o en in di e en g asses and c ops
ha he isible spec al ange VIs (VI
VIS
) pe o m well a he boo ing s age and do no pe o m as
well in he o he g owing s ages [
25
,
61
,
70
–
72
]. Wang e al. [
42
] ound ha lase scanning de i ed
me ics combined wi h hype spec al da a can p o ide be e biomass es ima es o maize using pa ial
leas squa es (PLS) eg ession. Pi man e al. [
8
] s udied es ima ion o biomass and canopy heigh in
be mudag ass, al al a, and he mix (con aining a mix u e o be mudag ass and al al a) using a gol
ca wi h mobile ul asonic, lase and spec al senso s. Fo single senso es ima ions, lase -es ima ed
heigh measu emen s had he bes co ela ions o he physically measu ed canopy heigh s and o he
DMY: he PCC o heigh eg essions was 0.88 o be mudag ass and 0.78 o he mix; co ela ions o
he DMY we e 0.88 o be mudag ass and 0.80 o al al a. Howe e , combining wo o h ee me hods
imp o ed co ela ions o heigh es ima ion o be mudag ass gi ing he PCC o 0.92; and o DMY
Ag icul u e 2018,8, 70 20 o 28
es ima ion in all cases, gi ing he PCC o 0.92 o be mudag ass, 0.83 o al al a and 0.89 o he mix.
Addi ionally, combined lase - and ul asonic-es ima ed heigh measu emen s p o ided equi alen
and/o be e es ima es when compa ed o he physical canopy heigh measu emen s and pla e me e
DMY es ima ion me hods. Di e ences be ween me hod-p edic ed alues e sus measu ed DMY
we e minimal. The a e age pe cen e o was 11.2% o he di e ences be ween p edic ed alues
e sus o age ha es e and quad a measu emen s o o age biomass alues (1.64 and 4.91 /ha),
excep a he lowes measu ed DMY, whe e he e o s we e la ge , he a e age pe cen e o was 89%
and he absolu e e o <0.79 /ha. Wi h he g ea es measu ed DMY, he a e age e o was 18% and
>6.4 /ha [8].
Moeckel e al. [
73
] es ima ed he DMY and FY wi h a g ound-based ul asonic and spec ome e
in g asslands wi h a he e ogeneous swa d s uc u e on ou da es ep esen ing di e en g ow h s ages.
The MPLSR app oach esul ed in a PCC o 0.69 (0.39–0.89 o da e-speci ic models) o he DMY and a
PCC o 0.82 (0.57–0.93 o da e-speci ic models) o he FY. F icke and Wachendo [
74
] had simila
esul s es ima ing biomass o legume-g ass swa ds wi h combined ul asonic and hype spec al
VIs: PCC we e 0.91 in common swa ds and 0.94–0.95 o species-speci ic calib a ions o DMY.
Ma abel e al. [33]
s udied Suppo Vec o Machine (SVM) and Pa ial Leas Squa es Reg ession
(PLSR) o es ima ing he biomass o g asslands om ield spec ome e da a. The bes esul s o
DMY es ima ion we e ob ained using PLSR, and he maximum band dep h index de i ed om he
con inuum emo ed e lec ance in he abso p ion ea u es be ween 916–1120 nm and 1079–1297 nm;
he PCC was 0.97 and he RMSE was 71.2 /ha o he DMY. We can hus conclude ha ou esul s
wi h he pho og amme ic d one da a we e in mos cases be e han he p e ious esul s ob ained
wi h ypically mo e expensi e measu emen sys ems, and e en wi h e es ial measu emen s.
To ob ain accu a e es ima ions o he g ass swa d biomass using a low-cos d one-based sys em,
we sugges u ilizing a high- esolu ion RGB came a equipped wi h a good quali y lens wi h a global
shu e and po en ially combined wi h an NIR band, e.g., a colou -in a ed modi ied came a. Senso s
such as he Pa o Sequoia [
75
] a e no ideal o 3D econs uc ion due o he olling shu e , bu he
mul ispec al da a a e ele an . Da ase s should be cap u ed wi h high image o e laps a a low lying
heigh o p o ide ul ahigh spa ial esolu ion and dense poin clouds. Wi h he suppo o an accu a e
onboa d di ec geo e e encing sys em he o e all sys em and measu emen cos could be educed o a
low le el. The image p ocessing and machine lea ning me hods in oduced in his s udy we e p o en
o p o ide accu a e esul s. We ecommend in eg a ing 3D, VI and RGB ea u es using ad anced
machine lea ning me hods.
5. Conclusions
Ou s udy de eloped and assessed a machine lea ning echnique based on mul ispec al
o hopho os and pho og amme ic 3D geome ic emo e sensing da a o he es ima ion o esh
and d y ma e yield o g ass swa ds o silage p oduc ion. Ou app oach was o ex ac a ious
ea u es om a emo e sensing da ase by combining an ul a-high esolu ion pho og amme ic canopy
heigh model (CHM) wi h a pixel size o 1.0 cm and ed, g een, blue and nea -in a ed ange in ensi y
alues and di e en ege a ion indices (VI) ex ac ed o m o hopho o mosaics. We compa ed he
pe o mance o he Mul iple Linea Reg ession (MLR) and he Random Fo es es ima o (RF). The bes
es ima ion esul s we e ob ained by combining he 3D, RGB and VI ea u es. The RF p o ided simila
o be e esul s han he MLR. The Pe son’s co ela ion coe icien (PCC) and RMSEs we e a bes
0.98 and 0.34 /ha
(12.70%), espec i ely, o he d y ma e yield combining he 3D, RGB and VI
ea u es, and 0.98 and 1.22 /ha (11.05%), espec i ely, o he esh yield combining he 3D and VI
ea u es. We also e alua ed he sensi i i y o he me hod by c ea ing a iabili y in he swa ds by
using di e en ni ogen e iliza ion a es and by epea ing he da a cap u e on ou da es o he
p ima y g ow h o he g ass swa d. Se e al poin s may ha e educed he accu acy o he es ima es:
(1) nonhomogeneous spa se s and as a esul o ea ly season; (2) nonhomogeneous spa se s and as a
Ag icul u e 2018,8, 70 21 o 28
esul s o low e ilize applica ion le els; (3) he heading o he swa d due o a la e ha es ing da e;
and (4) lodging caused by high e iliza ion le els.
Ou esul s a e consis en wi h p e ious scien i ic esul s ega ding he impac o da e and
combina ion o 3D and RGB ea u es on he heigh and biomass es ima ion esul s. Fu he mo e,
ou no el app oach in eg a ing machine lea ning algo i hms and he a ious 3D, spec al and VI
ea u es om he ul a-high esolu ion CHM and o homosaics p o ided be e esul s han mos o
he p e ious s udies. The esul s we e also highly p ecise in absolu e e ms. The yield es ima ion
esul s had an excellen accu acy and ou pe o med he es ima ion esul s based on he measu emen
on he ield using he heigh s ick. Ou esul s showed ha he p oposed me hod o e s an accu a e
ool o es ima ing bo h he esh yield and he d y ma e yield o g ass swa ds pa icula ly close o
he a ge ed silage ha es ing s age.
In he nex phases o ou esea ch, we will in eg a e he g ass swa d quali y pa ame e s o he
es ima ion p ocess, including ni ogen con en and diges ibili y. We will also compa e he p oposed
app oach o mul ispec al da ase s as well as o hype spec al da a. I is necessa y also o in eg a e
e icien di ec geo e e encing app oach o he me hod o u he imp o e he le el o au oma ion
and cos o he o e all sys em. An impo an u u e esea ch opic will be o assess he gene aliza ion
po en ial o he de eloped me hods—in o he wo ds, o use he ained es ima o s in o he ields
wi hou in si u aining da a. This is a highly ele an opic o de elop e icien emo e sensing ools
wi h minimum in si u e o s.
Au ho Con ibu ions:
O.N., J.K. and E.H. designed he expe imen . O.N. and J.K. planned, acqui ed and
analysed he ield e e ence. N.V., T.H. and R.N. acqui ed he emo e sensing da a and p ocessed he da a. N.V.
co esponded on he s a is ical analyses. R.N. co esponded on he machine lea ning me hods. N.V., E.H, R.N., J.K.
and O.N. analysed he da a and w o e he i s d a o he pape . All au ho s assis ed in w i ing and imp o ing
he pape .
Acknowledgmen s:
The au ho s would like o acknowledge unding by he Business Finland D oneKnowledge
p ojec (Dn o 1617/31/2016) and he ICT Ag i ERA-NET 2015 Enabling p ecision a ming p ojec
G assQ-De elopmen o g ound based and Remo e Sensing, au oma ed “ eal- ime” g ass quali y measu emen
o enhance g assland managemen in o ma ion pla o ms (P ojec id 35779). We a e g a e ul o Raquel Oli e a,
Somayeh Nezami and Lau i Ma kelin o hei help in p ocessing he mul ispec al da ase s.
Con lic s o In e es :
The au ho s decla e no con lic s o in e es . The unding sponso s had no ole in he design
o he s udy, in he collec ion, analyses, o in e p e a ion o da a, in he w i ing o he manusc ip , o in he decision
o publish he esul s.
Appendix A
Table A1.
Pea son co ela ion o coe icien s (PCC) o all VI’s wi h he DMY, he FY and he H
e
om
di e en imes. DMY: d y ma e yield; FY: esh yield; H e : e e ence heigh measu emen s.
DMY FY H e
Fea u e 6 June 15 June 19 June 28 June 6 June 15 June 19 June 28 June 6 June 15 June 19 June 28 June
RGBVI 0.51 0.49 0.77 0.76 0.58 0.43 0.73 0.79 0.51 0.63 0.79 0.59
GRVI 0.80 0.93 0.90 0.95 0.79 0.91 0.88 0.93 0.70 0.96 0.92 0.87
MGRVI 0.23 0.93 0.90 0.95 0.17 0.91 0.88 0.92 0.16 0.96 0.92 0.87
ExG 0.77 0.75 0.87 0.89 0.82 0.71 0.84 0.89 0.72 0.85 0.89 0.76
ExR 0.60 0.96 0.86 0.91 0.55 0.96 0.85 0.85 0.50 0.94 0.87 0.91
ExGR 0.80 0.92 0.90 0.95 0.79 0.90 0.88 0.92 0.70 0.96 0.91 0.87
NDVI 0.92 0.94 0.94 0.89 0.94 0.92 0.92 0.81 0.86 0.97 0.94 0.88
MSAVI 0.95 0.94 0.96 0.95 0.96 0.95 0.97 0.99 0.85 0.94 0.93 0.81
OSAVI 0.94 0.94 0.96 0.97 0.96 0.94 0.96 0.97 0.86 0.96 0.95 0.87
RVI 0.95 0.94 0.96 0.64 0.98 0.96 0.97 0.56 0.86 0.92 0.93 0.74
ExG + Hp90 0.91 0.94 0.96 0.90 0.95 0.92 0.94 0.83 0.84 0.97 0.96 0.89
ExG + Hmax 0.91 0.92 0.96 0.91 0.94 0.90 0.93 0.85 0.83 0.97 0.96 0.90
G assIp90 0.88 0.91 0.96 0.90 0.93 0.88 0.92 0.85 0.82 0.96 0.96 0.89
G assImax 0.88 0.92 0.96 0.90 0.94 0.90 0.93 0.83 0.83 0.96 0.96 0.89
Ag icul u e 2018,8, 70 22 o 28
Table A2.
Pea son co ela ion o coe icien s (PCC) o all VI’s wi h he DMY and he FY om di e en
ni ogen e ilize le els. DMY: d y ma e yield; FY: esh yield; 0–150: Ni ogen e ilize le els
0–150 kg/ha.
DMY FY
Fea u e 0 50 75 100 125 150 0 50 75 100 125 150
RGBVI 0.77 0.85 0.67 0.82 0.89 0.87 0.80 0.90 0.76 0.86 0.89 0.87
GRVI 0.22 0.55 0.48 0.56 0.62 0.55 0.17 0.64 0.59 0.65 0.75 0.60
MGRVI 0.62 0.72 0.72 0.74 0.71 0.78 0.69 0.87 0.85 0.85 0.87 0.86
ExG 0.75 0.88 0.68 0.84 0.85 0.90 0.77 0.94 0.79 0.89 0.92 0.91
ExGR 0.17 0.55 0.49 0.58 0.62 0.59 0.13 0.64 0.60 0.67 0.77 0.64
ExR 0.30 0.25 0.23 0.29 -0.02 0.48 0.35 0.20 0.22 0.26 0.37 0.45
NDVI 0.75 0.95 0.94 0.95 0.88 0.91 0.73 0.95 0.92 0.94 0.96 0.91
MSAVI 0.62 0.91 0.90 0.91 0.87 0.94 0.59 0.95 0.87 0.90 0.94 0.92
OSAVI 0.72 0.94 0.92 0.93 0.88 0.94 0.70 0.96 0.90 0.92 0.96 0.93
RVI 0.74 0.83 0.84 0.86 0.78 0.88 0.69 0.79 0.74 0.79 0.82 0.84
ExG + Hp90 0.80 0.92 0.96 0.96 0.93 0.88 0.75 0.98 0.99 0.98 0.98 0.91
ExG + Hmax 0.85 0.88 0.95 0.94 0.91 0.87 0.83 0.97 0.99 0.97 0.98 0.92
G assIp90 0.87 0.89 0.95 0.94 0.92 0.89 0.87 0.98 0.99 0.98 0.98 0.94
G assImax 0.83 0.93 0.96 0.96 0.93 0.91 0.81 0.99 0.99 0.99 0.99 0.94
Table A3.
Pea son co ela ion coe icien s (PCC) o all VI’s wi h he H
e
om di e en ni ogen
e ilize le els. H e : e e ence heigh measu emen s; 0–150: Ni ogen e ilize le els 0–150 kg/ha.
H e
Fea u e 0 50 75 100 125 150
RGBVI 0.76 0.83 0.69 0.84 0.89 0.81
GRVI 0.19 0.62 0.54 0.61 0.62 0.63
MGRVI 0.59 0.79 0.80 0.84 0.68 0.88
ExG 0.75 0.88 0.71 0.87 0.79 0.85
ExGR 0.15 0.62 0.55 0.63 0.62 0.66
ExR 0.32 0.16 0.19 0.27 0.06 0.37
NDVI 0.77 0.93 0.95 0.92 0.89 0.86
MSAVI 0.71 0.89 0.89 0.85 0.88 0.87
OSAVI 0.79 0.92 0.92 0.89 0.92 0.88
RVI 0.71 0.80 0.81 0.76 0.73 0.75
ExG + Hp90 0.76 0.93 0.99 0.97 0.88 0.94
ExG + Hmax 0.78 0.93 0.98 0.96 0.87 0.93
G assImax 0.80 0.93 0.97 0.97 0.88 0.94
G assIp90 0.80 0.93 0.98 0.97 0.89 0.95
Table A4.
Pea son co ela ion coe icien s (PCC) o all VI’s and CHM ea u es wi h he DMY, he FY
and he H
e
in all da ase s. DMY: d y ma e yield; FY: esh yield; H
e
: e e ence heigh measu emen s.
Fea u e DMY FY H e
RGBVI 0.81 0.80 0.83
GRVI 0.70 0.74 0.68
MGRVI 0.71 0.75 0.80
ExG 0.86 0.87 0.86
ExGR 0.70 0.74 0.68
ExR 0.27 0.33 0.24
NDVI 0.82 0.81 0.81
MSAVI 0.89 0.92 0.82
OSAVI 0.86 0.88 0.81
RVI 0.81 0.73 0.79
Ag icul u e 2018,8, 70 23 o 28
Table A4. Con .
Fea u e DMY FY H e
Hmean 0.91 0.9 0.93
Hmedian 0.9 0.89 0.93
Hmin 0.82 0.78 0.81
Hmax 0.91 0.92 0.94
Hs d 0.54 0.55 0.56
Hp50 0.9 0.89 0.93
Hp70 0.91 0.9 0.94
Hp80 0.92 0.91 0.94
Hp90 0.92 0.92 0.94
ExG + Hp90 0.93 0.93 0.95
ExG + Hmax 0.92 0.93 0.95
G assImax 0.94 0.93 0.96
G assIp90 0.93 0.93 0.96
H e 0.95 0.93 -
Appendix B
Table A5.
Selec ed ea u es o he Mul ilinea Reg ession (MLR) (no in he o de o impo ance).
DMY: d y ma e yield; FY: esh yield; RGB: Red, G een and Blue spec al ea u es; VI: Vege a ion
Index ea u es; 3D: CHM 3D ea u es.
Case DMY FY
MLR
RGB B B
VI RGBVI, GRVI, MGRVI, ExR,
MSAVI, RVI RGBVI, MGRVI, ExR, RVI
3D Hmin, Hmax, Hs d Hmedian, Hmin, Hmax, Hs d, Hp50
RGB + VI R, G, B, GRVI, ExGR, ExR, NDVI,
MSAVI, RVI
R, B, MGRVI, ExG, NDVI, MSAVI, RVI
RGB + 3D R, G, B, Hmin, Hp80 R, G, B, Hmedian, Hmax, Hs d, Hp50
VI + 3D
RGBVI, MGRVI, ExG, ExR,
MSAVI, RVI, Hmedian, Hp50, Hp80,
G assImax, ExG + Hp90
RGBVI, MGRVI, ExR, NDVI, MSAVI,
OSAVI, Hmean, Hmax, Hs d, Hp90,
G assImax, G assIp90, ExG + Hmax
RGB + VI + 3D
R, B, RGBVI, MGRVI, ExG, ExR,
NDVI, MSAVI, OSAVI, Hmedian,
Hmax, Hp50, Hp90, ExG + Hmax,
ExG + Hp90
R, B, MGRVI, ExR, NDVI, OSAVI,
Hmean, Hmax, Hs d, Hp90, G assImax,
G assIp90, ExG + Hmax, ExG + Hp90
Ag icul u e 2018,8, 70 24 o 28
Table A6.
The mos impo an ea u es o he Random Fo es (RF) (in he o de o impo ance) o
di e en da es and di e en ni ogen e ilize le els. DMY: d y ma e yield; FY: esh ma e ; RGB:
Red, G een and Blue spec al ea u es; VI: Vege a ion Index ea u es; 3D: CHM 3D ea u es; 0–150:
Ni ogen e ilize le els 0–150 kg/ha.
Case DMY FY
Da e
6 June G assImax, ExG + Hmax, G assIp90, ExG +
Hp90, MSAVI, OSAVI, NDVI, Hmax, B, RVI
G assImax, ExG + Hmax, MSAVI, NDVI,
G assIp90, OSAVI, ExG + Hp90, RVI,
Hp80, Hmax
15 June ExR, MSAVI, B, ExGR, GRVI, Hmean,
Hmedian, NDVI, G, ExG + Hmax
G assImax, OSAVI, RVI, NDVI, MSAVI,
Hp70, Hp90, Hp80, MGRVI, ExG + Hp90
19 June
ExG + H
max
, ExG, OSAVI, ExR, NDVI, ExG
+ Hp90, MGRVI, G assIp90, Hmean, Hp50
ExG + Hmax, RVI, NDVI, ExG + Hp90,
Hmean, OSAVI, Hp50, Hp90,
Hmedian, G assIp90
28 June ExG + Hmax, ExR, ExG, GRVI, MGRVI,
MSAVI, OSAVI, ExGR, G assImax, NDVI
ExGR, OSAVI, MSAVI, ExR, GRVI, ExG,
MGRVI, ExG + Hp90, NDVI, Hp90
Ni ogen
0Hmax, Hp90, Hp80, G assIp90, Exg + Hmax,
G assImax, ExG + Hp90, Hs d, Hp70, RVI
ExG + Hp90, Hmax, G assIp90, G assImax,
Exg + Hmax, Hs d, Hp90, Hp80, RVI, NDVI
50
G assI
p90
, H
min
, MSAVI, OSAVI, G assI
max
,
RVI, B, Hp70, Hp90, Hp80
Hp70, ExG + Hp90, Hmax, G assIp90, RVI,
MGRVI, B, ExG + Hmax, MSAVI, G assImax
75
Hmax, G assImax, ExG + Hmax, Hp80,
G assIp90, Hmin, Hp90, Hp50,
ExG + Hp90, Hmedian
OSAVI, G assIp90, RGBVI, MSAVI, MGRVI,
G, ExG + Hp90, RVI, ExG
100 ExG + Hmax, Hp80, Hp90, G assImax, Hp70,
Hmax, Hmean, Hp50, G assIp90
H
p90
, G assI
max
, H
p80
, H
median
, RVI, H
mean
,
G assIp90, Hmin, NDVI
125 Hp90, Hp80, G assImax, ExG + Hmax, Hmax,
G assIp90, OSAVI, MSAVI, Hp70
ExG + Hp90, MSAVI, G assImax, RVI, ExG,
B, Hp80, RGBVI, Hp90
150 Hp90, RVI, Hmin, OSAVI, Hp70, Hmedian,
Hmean, MSAVI, Hmax
ExG + Hp90, Hmin, G assImax, MGRVI,
G assIp90, RVI, ExG + Hmax, Hp80,
MSAVI, NDVI
Re e ences
1.
Vi kajä i, P. Compa ison o Th ee Indi ec Me hods o P edic ion o He bage Mass on Timo hy-Meadow
Fescue Pas u es. Ac a Ag ic. Scand. Sec . B Soil Plan Sci. 1999,49, 75–81. [C ossRe ]
2.
Paka inen, K.; Hy käs, M.; Juu inen, E. De elopmen and alida ion o p ac ical me hods o de e mina ion
o d y ma e yield in g ass silage swa ds. In P oceedings o he 12 h Cong ess o he Eu opean Socie y o
Ag onomy, Helsinki, Finland, 20–24 Augus 2012; Volume 14, pp. 542–543.
3.
Cunli e, A.M.; B azie , R.E.; Ande son, K. Ul a- ine g ain landscape-scale quan i ica ion o d yland
ege a ion s uc u e wi h d one-acqui ed s uc u e- om-mo ion pho og amme y. Remo e Sens. En i on.
2016,183, 129–143. [C ossRe ]
4.
Vi kaja i, P.; Ja en an a, K. Lea dynamics o imo hy and meadow escue unde No dic condi ions.
G ass Fo age Sci. 2001,56, 294–304. [C ossRe ]
5.
Tucke , C.J. A c i ical e iew o emo e sensing and o he me hods o non-des uc i e es ima ion o s anding
c op biomass. G ass Fo age Sci. 1980,35, 177–182. [C ossRe ]
6.
Sande son, M.A.; Ro z, C.A.; Ful z, S.W.; Raybu n, E.B. Es ima ing Fo age Mass wi h a Comme cial
Capaci ance Me e , Rising Pla e Me e , and Pas u e Rule . Ag on. J. 2001,93, 1281–1286. [C ossRe ]
7.
Vi kajä i, P. G ow h and U iliza ion o Timo hy: Meadow Fescue Pas u es; Uni e si y o Helsinki: Helsinki,
Finland, 2004.
Ag icul u e 2018,8, 70 25 o 28
8.
Pi man, J.; A nall, D.; In e an e, S.; Mo e , C.; Bu le , T. Es ima ion o Biomass and Canopy Heigh in
Be mudag ass, Al al a, and Whea Using Ul asonic, Lase , and Spec al Senso s. Senso s
2015
,15, 2920–2943.
[C ossRe ] [PubMed]
9.
Ho meis e , D.; Waldho , G.; Cu d , C.; Tilly, N.; Bendig, J.; Ba e h, G. Spa ial a iabili y de ec ion o c op
heigh in a single ield by e es ial lase scanning. In P ecision Ag icul u e ’13: Pape s P esen ed a he 9 h
Eu opean Con e ence on P ecision Ag icul u e; S a o d, J.V., Ed.; Wageningen Academic: Lleida, Spain, 2013;
pp. 267–274.
10.
Tilly, N.; Ho meis e , D.; Cao, Q.; Huang, S.; Lenz-Wiedemann, V.; Miao, Y.; Ba e h, G. Mul i empo al c op
su ace models: Accu a e plan heigh measu emen and biomass es ima ion wi h e es ial lase scanning
in paddy ice. J. Appl. Remo e Sens. 2014,8, 083671. [C ossRe ]
11.
Zhang, C.; Ko acs, J.M. The applica ion o small unmanned ae ial sys ems o p ecision ag icul u e: A e iew.
P ecis. Ag ic. 2012,13, 693–712. [C ossRe ]
12.
Bendig, J.; Bol en, A.; Ba e h, G. UAV-based Imaging o Mul i-Tempo al, e y high Resolu ion C op
Su ace Models o moni o C op G ow h Va iabili yMoni o ing des P lanzenwachs ums mi Hil e
mul i empo ale und hoch au lösende Obe lächenmodelle on Ge eidebes änden au Basis on Bilde n
aus UAV-Be liegungen. Pho og amm. Fe ne kund. Geoin . 2013, 551–562. [C ossRe ]
13. Bendig, J.; Bol en, A.; Benne z, S.; B oschei , J.; Eich uss, S.; Ba e h, G. Es ima ing Biomass o Ba ley Using
C op Su ace Models (CSMs) De i ed om UAV-Based RGB Imaging. Remo e Sens.
2014
,6, 10395–10412.
[C ossRe ]
14.
Li, W.; Niu, Z.; Chen, H.; Li, D.; Wu, M.; Zhao, W. Remo e es ima ion o canopy heigh and abo eg ound
biomass o maize using high- esolu ion s e eo images om a low-cos unmanned ae ial ehicle sys em.
Ecol. Indic. 2016,67, 637–648. [C ossRe ]
15.
Luciee , A.; Tu ne , D.; King, D.H.; Robinson, S.A. Using an Unmanned Ae ial Vehicle (UAV) o cap u e
mic o- opog aphy o An a c ic moss beds. In . J. Appl. Ea h Obs. Geoin . 2014,27, 53–62. [C ossRe ]
16.
Tu ne , D.; Luciee , A.; Maleno ský, Z.; King, D.; Robinson, S.A. Assessmen o An a c ic moss heal h om
mul i-senso UAS image y wi h Random Fo es Modelling. In . J. Appl. Ea h Obs. Geoin .
2018
,68, 168–179.
[C ossRe ]
17.
Be ni, J.; Za co-Tejada, P.J.; Sua ez, L.; Fe e es, E. The mal and Na owband Mul ispec al Remo e Sensing o
Vege a ion Moni o ing F om an Unmanned Ae ial Vehicle. IEEE T ans. Geosci. Remo e Sens.
2009
,47, 722–738.
[C ossRe ]
18.
Hun , E.R.; Hi ely, W.D.; Fujikawa, S.; Linden, D.; Daugh y, C.S.; McCa y, G. Acquisi ion o NIR-G een-Blue
Digi al Pho og aphs om Unmanned Ai c a o C op Moni o ing. Remo e Sens.
2010
,2, 290–305. [C ossRe ]
19.
Candiago, S.; Remondino, F.; De Giglio, M.; Dubbini, M.; Ga elli, M. E alua ing Mul ispec al Images and
Vege a ion Indices o P ecision Fa ming Applica ions om UAV Images. Remo e Sens.
2015
,7, 4026–4047.
[C ossRe ]
20.
Geipel, J.; Link, J.; Wi wahn, J.; Claupein, W. A P og ammable Ae ial Mul ispec al Came a Sys em o
In-Season C op Biomass and Ni ogen Con en Es ima ion. Ag icul u e 2016,6, 4. [C ossRe ]
21.
Za co-Tejada, P.J.; González-Dugo, V.; Be ni, J.A.J. Fluo escence, empe a u e and na ow-band indices
acqui ed om a UAV pla o m o wa e s ess de ec ion using a mic o-hype spec al image and a he mal
came a. Remo e Sens. En i on. 2012,117, 322–337. [C ossRe ]
22.
Honka aa a, E.; Saa i, H.; Kai osoja, J.; Pölönen, I.; Hakala, T.; Li key, P.; Mäkynen, J.; Pesonen, L. P ocessing
and Assessmen o Spec ome ic, S e eoscopic Image y Collec ed Using a Ligh weigh UAV Spec al Came a
o P ecision Ag icul u e. Remo e Sens. 2013,5, 5006–5039. [C ossRe ]
23.
Aasen, H.; Bu ka , A.; Bol en, A.; Ba e h, G. Gene a ing 3D hype spec al in o ma ion wi h ligh weigh
UAV snapsho came as o ege a ion moni o ing: F om came a calib a ion o quali y assu ance. ISPRS J.
Pho og amm. Remo e Sens. 2015,108, 245–259. [C ossRe ]
24.
Yue, J.; Yang, G.; Li, C.; Li, Z.; Wang, Y.; Feng, H.; Xu, B. Es ima ion o Win e Whea Abo e-G ound Biomass
Using Unmanned Ae ial Vehicle-Based Snapsho Hype spec al Senso and C op Heigh Imp o ed Models.
Remo e Sens. 2017,9, 708. [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
UAV-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. Geoin . 2015,39, 79–87. [C ossRe ]