emo e sensing
A icle
Es ima ing Biomass and Ni ogen Amoun o Ba ley
and G ass Using UAV and Ai c a Based Spec al and
Pho og amme ic 3D Fea u es
Roope Näsi 1,*ID , Niko Viljanen 1ID , Je e Kai osoja 2, Ka ja Alhonoja 3, Teemu Hakala 1ID ,
Lau i Ma kelin 1ID and Eija Honka aa a 1ID
1Depa 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] (N.V.); [email p o ec ed] (T.H.);
[email p o ec ed] (L.M.); [email p o ec ed] (E.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]
3Ya a Ko kaniemi Resea ch S a ion, Ya a Suomi Oy, Ko kaniemen ie 100, 03250 Ojakkala, Finland;
[email p o ec ed]
*Co espondence: [email p o ec ed]; Tel.: +358-29-531-4860
Recei ed: 31 May 2018; Accep ed: 5 July 2018; Published: 7 July 2018


Abs ac :
The imely es ima ion o c op biomass and ni ogen con en is a c ucial s ep in a ious
asks in p ecision ag icul u e, o example in e iliza ion op imiza ion. Remo e sensing using d ones
and ai c a s o e s a easible ool o ca y ou his ask. Ou objec i e was o de elop and assess a
me hodology o c op biomass and ni ogen es ima ion, in eg a ing spec al and 3D ea u es ha can
be ex ac ed using ai bo ne minia u ized mul ispec al, hype spec al and colou (RGB) came as.
We used he Random Fo es (RF) as he es ima o , and in addi ion Simple Linea Reg ession (SLR)
was used o alida e he consis ency o he RF esul s. The me hod was assessed wi h empi ical
da ase s cap u ed o a ba ley ield and a g ass silage ial si e using a hype spec al came a based
on he Fab y-Pé o in e e ome e (FPI) and a egula RGB came a onboa d a d one and an ai c a .
Ag icul u al e e ence measu emen s included esh yield (FY), d y ma e yield (DMY) and amoun o
ni ogen. In DMY es ima ion o ba ley, he Pea son Co ela ion Coe icien (PCC) and he no malized
Roo Mean Squa e E o (RMSE%) we e a bes 0.95% and 33.2%, espec i ely; and in he g ass DMY
es ima ion, he bes esul s we e 0.79% and 1.9%, espec i ely. In he ni ogen amoun es ima ions o
ba ley, he PCC and RMSE% we e a bes 0.97% and 21.6%, espec i ely. In he biomass es ima ion,
he bes esul s we e ob ained when in eg a ing hype spec al and 3D ea u es, bu he in eg a ion
o RGB images and 3D ea u es also p o ided esul s ha we e almos as good. In ni ogen con en
es ima ion, he hype spec al came a ga e he bes esul s. We concluded ha he in eg a ion o
spec al and high spa ial esolu ion 3D ea u es and adiome ic calib a ion was necessa y o op imize
he accu acy.
Keywo ds:
hype spec al; pho og amme y; UAV; d one; machine lea ning; andom o es ;
eg ession; p ecision ag icul u e; biomass; ni ogen
1. In oduc ion
The moni o ing o plan s du ing he g owing season is he basis o p ecision ag icul u e. Wi h he
suppo o quan i y and quali y in o ma ion on plan s (i.e., c op pa ame e s), a me s can plan he c op
managemen and inpu use ( o example, nu ien applica ion and c op p o ec ion) in a con olled way.
Biomass is he mos common c op pa ame e indica ing he amoun o he yield [
1
]; and oge he wi h
ni ogen con en in o ma ion, i can be used o de e mine he need o addi ional ni ogen e iliza ion.
Remo e Sens. 2018,10, 1082; doi:10.3390/ s10071082 www.mdpi.com/jou nal/ emo esensing
Remo e Sens. 2018,10, 1082 2 o 32
When a m inpu s a e co ec ly aligned, bo h he en i onmen and he a me bene i by ollowing he
p inciple o sus ainable in ensi ica ion [2]
Remo e sensing has p o ided ools o p ecision ag icul u e since he 1980s [
3
]. Howe e , d ones
(o UAV (Unmanned Ae ial Vehicles) o RPAS (Remo ely Pilo ed Ai c a Sys em)) ha e de eloped
apidly, o e ing new al e na i es o adi ional emo e sensing echnologies [
1
,
4
]. Remo e sensing
ins umen s ha collec spec al e lec ance measu emen s ha e ypically been ope a ed om sa elli es
and ai c a o es ima e c op pa ame e s. Due o echnological inno a ions, ligh weigh mul i- and
hype -spec al senso s ha e become a ailable in ecen yea s. These senso s can be ca ied by
small UAVs ha o e no el emo e sensing ools o p ecision ag icul u e. One ype o ligh weigh
hype spec al senso is based on he Fab y-Pé o in e e ome e (FPI) echnique [
5
–
8
], and his was used
in his s udy. This echnology p o ides spec al da a cubes wi h a ame o ma . The FPI senso has
al eady shown po en ial in a ious en i onmen al mapping applica ions [
7
,
9
–
20
]. In addi ion o spec a,
da a abou he 3D s uc u e o plan s can be collec ed a he same ime because he ame-based senso s
and mode n pho og amme y enable he gene a ion o spec al Digi al Su ace Models (DSM) [
21
,
22
].
The use o d one-based pho og amme ic 3D da a has al eady p o ided p omising esul s in biomass
es ima ion, bu combining he 3D and spec al e lec ance da a has u he imp o ed he es ima ion
esul s [23–25].
A la ge numbe o s udies ega ding c op pa ame e es ima ion using emo e sensing echnologies
ha e been published du ing he las decades. The as majo i y o hem ha e been conduc ed using
spec al in o ma ion cap u ed om sa elli e o manned ai c a pla o ms. Since lase scanning
became widesp ead, 3D in o ma ion on plan heigh and s uc u e became a ailable o c op
pa ame e es ima ion. Te es ial app oaches ha e mos ly been used hus a [
26
–
28
] due o he
equi emen s o high spa ial esolu ion and he ela i ely la ge weigh o high-pe o mance sys ems.
The as de elopmen o d one echnology and pho og amme y, especially he s uc u e om
mo ion (SFM) echnologies, ha e made 3D da a collec ion mo e e icien , lexible and low in cos .
No su p isingly, pho og amme ic 3D da a om d ones we e aken unde sc u iny o p ecision
ag icul u e applica ions [
16
,
25
,
29
–
32
]. Ins ead o 3D da a, a ious s udies ha e exploi ed Vege a ion
Indices (VI) adop ed om mul ispec al [
33
–
37
] o hype spec al da a [
21
,
38
,
39
]. Howe e , only a
ew s udies ha e in eg a ed UAV-based spec al and 3D in o ma ion o c op pa ame e es ima ion.
Yue e al. [
24
] combined spec al and c op heigh in o ma ion om a Cube UHD 180 hype spec al
senso (Cube GmbH, Ulm, Ge many) o es ima e he biomass o win e whea . They concluded ha
combining he c op heigh in o ma ion wi h wo-band VIs imp o ed he es ima ion esul s. Bu hey
sugges ed ha he accu acy o hei es ima ions could be imp o ed by u ilizing ull spec a, mo e
ad anced es ima ion me hods, and g ound con ol poin s (in he geo e e encing p ocess o imp o e
geome ic accu acy). In he s udy by Bendig e al. [
23
], pho og amme ic 3D da a was combined
wi h spec ome e measu emen s om he g ound. G ound-based app oaches, which ha e combined
spec al and 3D da a, ha e also been pe o med [
28
,
40
,
41
]. Comple ely d one-based app oaches
we e in es iga ed by Geipel e al. [
37
], Schi mann e al., [
42
] and Li e al. [
32
] o c op pa ame e
es ima ion based on RGB poin clouds wi h uncalib a ed spec al da a. The s udy by Li e al. [
32
])
showed ha poin cloud me ics o he han he mean heigh o he c op a e also ele an in o ma ion
o biomass modelling.
In he as majo i y o biomass es ima ion s udies, es ima o s such as linea models and nea es
neighbou app oaches ha e been applied [
43
]. In pa icula , d one-based c op pa ame e es ima ion
s udies ha e been pe o med mos ly by eg ession echniques using a ew ea u es and linea
models
[4,21,23,28,37]
o using he nea es neighbou echnique [
7
,
14
]. Thus, he use o es ima o s
which a e able o exploi he ull spec a, such as he Random Fo es (RF), ha e been sugges ed in
UAV-based c op pa ame e es ima ion [
21
,
25
]. Since he publica ion o he RF echnique [
44
], i has
ecei ed inc easing a en ion in emo e sensing applica ions [
45
]. The main ad an ages o he RF o e
many o he me hods include high p edic ion accu acy, he possibili y o in eg a e a ious ea u es
in he es ima ion p ocess, no need o ea u e selec ion (because calcula ions include measu es o
Remo e Sens. 2018,10, 1082 3 o 32
ea u e impo ance o de ), and i is less sensi i e o o e i ing and in pa ame e selec ion [
45
–
47
].
In biomass es ima ion, RF has shown compe i i e accu acy among o he es ima ion me hods applied
in o es y [
43
,
48
] and in ag icul u al [
32
,
49
–
51
] applica ions. Only some s udies ha e used RF in
c op pa ame e es ima ions. Liu e al. [
50
] used RF o es ima e he ni ogen le el o whea using
mul ispec al da a. Li e al. [
32
] and Yue e al. [
51
] used success ully RF o es ima ing he biomass o
maize and win e whea . P e iously, Viljanen e al. [
5
] used RF o he esh and d y ma e biomass
es ima ion o g ass silage, using 3D and mul ispec al ea u es. Exis ing s udies ha e ocused mo e
on biomass es ima ion han on ni ogen con en es ima ion. Especially he s udies on he use o
hype spec al da a in ni ogen es ima ion ha e commonly used e es ial app oaches (e.g., [52–54]).
The objec i e o his in es iga ion was o de elop and assess a no el op imized wo k low
based on he RF algo i hm o es ima ing c op pa ame e s employing bo h spec al and 3D ea u es.
Hype spec al and pho og amme ic image y was collec ed using he FPI came a and a egula
consume RGB came a. This s udy employed he ull hype spec al and s uc u al in o ma ion o he
biomass and ni ogen con en es ima ion o mal ba ley and g ass silage u ilizing da ase s cap u ed
using a d one and ai c a . We also e alua ed he impac o he adiome ic p ocessing le el on
he esul s. This pape ex ends ou p e ious wo k [
55
], which pe o med a p elimina y s udy wi h
he ba ley da a using linea eg ession echniques. The majo con ibu ions o his s udy we e he
de elopmen and assessmen o he in eg a ed use o spec al and 3D ea u es in he c op pa ame e
es ima ion in di e en condi ions, he compa ison o RGB and hype spec al imaging based emo e
sensing echniques and he conside a ion o impac s o a ious pa ame e s, especially he lying heigh
and he adiome ic p ocessing le el on he esul s.
2. Ma e ials and Me hods
2.1. Tes A ea and G ound T u h
A es si e o ag icul u al emo e sensing was es ablished in 2016 by he Na u al Resou ces
Ins i u e Finland (LUKE) and he Finnish Geospa ial Resea ch Ins i u e in he Na ional Land Su ey o
Finland (FGI) in Vih i, Ho i (60
◦
25
0
21
00
N, 24
◦
22
0
28
00
E). The en i e es a ea included h ee pa cels wi h
ba ley (35 ha in o al) and wo pa cels wi h g ass (11 ha) (Figu e 1).
The mal ba ley T ekke pa cels we e seeded be ween 29 May and 6 June 2016. The combined
d illing se ings o seeding densi y was 200 kg/ha and o ni ogen inpu 67.5 kg/ha. Due o ela i ely
cold wea he condi ions and a sho g owing season, he ba ley yield was small (1900 kg/ha) and had
a a iance o 23.3% [
18
]. The ba ley ha es ing was made be ween 23 Sep embe and 11 Oc obe 2016.
The ela i ely la ge span in da es was due o he di icul wea he condi ions. In his s udy, we used he
ba ley pa cel o 20 ha in size. The ba ley e e ence measu emen s we e ca ied ou on 8 July 2016 on
36 sample a eas ha we e 50 cm
×
50 cm. The ield was e enly ea ed, al hough a 12-m wide s ipe
spli ing he ield was le un ea ed o p o ide a ba e soil e e ence. The measu emen s included
he a e age plan heigh , esh yield (FY), d y ma e yield (DMY) and amoun o ni ogen (Table 1).
The coo dina es o he sample a eas we e measu ed using di e en ially co ec ed T imble GeoXH GPS
wi h an accu acy o 10 cm in he X- and Y-coo dina es. The a e age plan heigh o each sample spo
was measu ed using a measu emen s ick. The sample plo s we e selec ed so ha he ege a ion was
as homogeneous as possible inside and a ound he sample a eas. Thi een o he sample plo s we e
loca ed on he sp aying acks ha did no include ba ley (0-squa es), howe e , some weeds we e
g owing in hese sample a eas, which was impo an o no e du ing he analysis.
The g ass silage ield was a i e-yea -old mix u e o imo hy and meadow escue. Sample a eas
we e based on eigh ea men ial plo s, wi h ou eplica es conduc ed by Ya a Ko kaniemi Resea ch
S a ion, Ya a Suomi Oy, Vih i, Finland (Ya a) (h ps://www.ya a. i/). The ni ogen ou pu o he
i s cu in e e y ea men was 120 kg/ha, and he yield le el a ied be ween 4497 and 4985 kg/ha
o d y ma e . The phospho us (P) le el o he g ass ield si e was e y low (2.9 mg/L), and di e en
ea men s wi h a iable P le els pa ly explains he yield di e ences. The e e ence measu emen s o a
Remo e Sens. 2018,10, 1082 4 o 32
g ass pa cel we e ca ied ou by Ya a in he i s cu on 13 June 2016 in 32 sample a eas (
1.5 m ×10 m
).
Sample a eas we e ha es ed wi h a Hald up 1500 o age plo ha es e . A e ha es ing, d ied
samples we e analysed in he labo a o y. The ea men s we e combined in he labo a o y analysis;
hus, he e e ence FY, DMY and ni ogen measu emen s we e a ailable o eigh samples (Table 1).
Al oge he 32 pe manen g ound con ol poin s (GCPs) we e buil and measu ed in he
a ea. They we e ma ked by wooden poles and a ge ed wi h ci cula a ge s 30 cm in diame e .
Thei coo dina es in he ETRS-TM35FIN coo dina e sys em we e measu ed using he T imble R10
(L1 + L2)
RTK-GPS. The es ima ed accu acy o he GCPs was 2 cm in ho izon al coo dina es and
3 cm in heigh [
56
]. Fu he mo e, h ee e lec ance panels wi h a nominal e lec ance o 0.03, 0.09 and
0.50 [57] we e ins alled in he a ea o suppo he adiome ic p ocessing.
Remo e Sens. 2018, 10, x FOR PEER REVIEW 4 o 33
ha es ing, d ied samples we e analysed in he labo a o y. The ea men s we e combined in he
labo a o y analysis; hus, he e e ence FY, DMY and ni ogen measu emen s we e a ailable o
eigh samples (Table 1).
Al oge he 32 pe manen g ound con ol poin s (GCPs) we e buil and measu ed in he a ea.
They we e ma ked by wooden poles and a ge ed wi h ci cula a ge s 30 cm in diame e . Thei
coo dina es in he ETRS-TM35FIN coo dina e sys em we e measu ed using he T imble R10 (L1 +
L2) RTK-GPS. The es ima ed accu acy o he GCPs was 2 cm in ho izon al coo dina es and 3 cm in
heigh [56]. Fu he mo e, h ee e lec ance panels wi h a nominal e lec ance o 0.03, 0.09 and 0.50
[57] we e ins alled in he a ea o suppo he adiome ic p ocessing.
Figu e 1. Tes si e whe e ba ley and g ass ields a e ma ked using hick black lines on he
o homosaic based on RGB images om d one. Loca ions o g ound con ol poin s and 36 sample
plo s in ba ley and 32 sample plo s in g ass ield (zoom) is also ma ked.
Table 1. Ag icul u al sample e e ence measu emen s o ba ley and g ass ields: Min: minimum;
Max: maximum; Mean: a e age and s anda d de ia ion o he a ibu e; N o plo s: numbe o
sample plo s.
Plan
A ibu e
Min
Max
Mean
S anda d De ia ion
N o Plo s
Ba ley
F esh biomass (kg/m2)
0
1.66
0.46
0.52
36
Ba ley
D y biomass (kg/m2)
0
0.24
0.07
0.08
36
Ba ley
Ni ogen (kg/m2)
0
0.01
0.00
0.00
36
Ba ley
Ni ogen %
0
4.23
1.71
0.49
36
Ba ley
Heigh (m)
0
0.31
0.13
0.11
36
G ass
F esh biomass (kg/m2)
1.5
1.88
1.73
0.10
8
G ass
D y biomass (kg/m2)
0.45
0.50
0.48
0.02
8
G ass
Ni ogen (kg/m2)
0.01
0.01
0.01
0.00
8
G ass
Ni ogen %
1.47
1.96
1.70
0.19
8
2.2. Remo e Sensing Da a
Figu e 1.
Tes si e whe e ba ley and g ass ields a e ma ked using hick black lines on he o homosaic
based on RGB images om d one. Loca ions o g ound con ol poin s and 36 sample plo s in ba ley
and 32 sample plo s in g ass ield (zoom) is also ma ked.
Table 1.
Ag icul u al sample e e ence measu emen s o ba ley and g ass ields: Min: minimum;
Max: maximum; Mean: a e age and s anda d de ia ion o he a ibu e; N o plo s: numbe o sample plo s.
Plan A ibu e Min Max Mean S anda d De ia ion N o Plo s
Ba ley F esh biomass (kg/m2)0 1.66 0.46 0.52 36
Ba ley D y biomass (kg/m2)0 0.24 0.07 0.08 36
Ba ley Ni ogen (kg/m2)0 0.01 0.00 0.00 36
Ba ley Ni ogen % 0 4.23 1.71 0.49 36
Ba ley Heigh (m) 0 0.31 0.13 0.11 36
G ass F esh biomass (kg/m2)1.5 1.88 1.73 0.10 8
G ass D y biomass (kg/m2)0.45 0.50 0.48 0.02 8
G ass Ni ogen (kg/m2)0.01 0.01 0.01 0.00 8
G ass Ni ogen % 1.47 1.96 1.70 0.19 8
Remo e Sens. 2018,10, 1082 5 o 32
2.2. Remo e Sensing Da a
Remo e sensing da a cap u es we e ca ied ou using a d one and a manned ai c a (Table 2).
A hexacop e d one wi h a Ta o 960 oldable ame belonging o he FGI was equipped wi h
a hype spec al came a based on a uneable FPI and a high-quali y Samsung NX500 RGB came a.
In his s udy, he FGI2012b FPI came a [
6
,
7
,
58
] was used; i was con igu ed wi h 36 spec al bands in
he 500 nm o 900 nm spec al ange (Table 3). The d one had a NV08C-CSM L1 GNSS ecei e
(NVS Na iga ion Technologies L d., Mon lingen, Swi ze land) and a Raspbe y Pi single-boa d
compu e (Raspbe y Pi Founda ion, Camb idge, UK). The RGB came a was igge ed o ake images
a wo-second in e als, and he GNSS ecei e was used o eco d he exac ime o each igge ing
pulse. The FPI came a had i s own GNSS ecei e , which collec ed he exac ime o each image.
We calcula ed pos -p ocessed kinema ic (PPK) GNSS posi ions o he RGB and FPI came as’ images,
using he NV08C-CSM and he Na ional Land Su ey o Finland (NLS) RINEX se ice [
59
], using
RTKlib so wa e (RTKlib, e sion 2.4.2, Open-sou ce, Raleigh, NC, USA). UAV da a in g ass ields
was collec ed using lying heigh s o 50 m and 140 m and lying speeds o 3.5 m/s and 5 m/s, which
p o ided g ound sampling dis ances (GSDs) o 0.01 m and 0.05 m o RGB images and 0.05 m and
0.14 m o FPI images, espec i ely. In he ba ley ield, only he lying heigh o 140 m was used, bu
ou di e en ligh s du ing 3.5 h we e necessa y o co e he en i e es ield.
In he ba ley ield, emo e sensing da ase s we e also cap u ed using a manned small ai c a
(Cessna, Wichi a, KS, USA) ope a ed by Len oku a Vallas. The came as we e a RGB came a
(Nikon D3X, Tokyo, Japan) and he FPI came a. The RGB da a om he ai c a was collec ed using
lying heigh s o 450 m and 900 m and lying speed o 55 m/s, p o iding GSDs o 0.05 m and 0.10 m,
espec i ely, o 450 m and 900 m al i udes. The ai c a -based FPI images we e cap u ed using a lying
heigh o 700 m and a lying speed o 65 m/s, which p o ided a GSD o 0.6 m (Table 2). GNSS ajec o y
da a was no a ailable o he ai c a da a.
The ligh pa ame e s p o ided image blocks wi h 73–93% o wa d and 65–82% side o e laps, which
a e sui able o accu a e pho og amme ic p ocessing. In he ollowing, we will e e o he UAV-based
senso s as UAV FPI and UAV RGB and he manned ai c a (AC)-based senso s as AC FPI and AC RGB.
Table 2.
Fligh pa ame e s o each da ase : da e, ime, wea he , sun azimu h, sola ele a ion, FH:
ligh heigh and FL: numbe o ligh lines. AC RGB: ai c a wi h RGB came a; AC FPI: ai c a
wi h FPI (Fab y–Pé o in e e ome e ) came a. (In he UAV da ase s FPI and RGB came as we e
used simul aneously).
Da ase Da e Time
Wea he
Exposu e ime (ms) Sun Azimu h Sola Ele a ion FH FL
(UTC +3) (◦) (◦) (m)
G ass UAV 140 m 13 June
13:31 o 13:58
a ying 8 188.47 52.63
140
6
G ass UAV 50 m 13 June
15:09 o 15:40
a ying 8 223.47 47.21
50 10
Ba ley UAV 140 m 4 July
12:42 o 16:15
cloudy 20–25 166–233 43–52
140 28
Ba ley AC RGB 450 m 6 July
11:49 o 12:04
sunny 146.59 48.92
450 10
Ba ley AC RGB 900 m 6 July
12:06 o 12:49
sunny 158.76 50.9
900
6
Ba ley AC FPI 700 m 6 July
10:18 o 11:23
a ying 8 126.38 43.41
700
7
Table 3.
Spec al se ings o he hype spec al came a. L0: cen al wa eleng h; FWHM: ull wid h a
hal maximum.
Band
1 2 3 4 5 6 7 8 9 10 11 12
L0
(nm): 512.3 514.8 520.4 527.5 542.9 550.6 559.7 569.9 579.3 587.9 595.9 604.6
FWHM (nm): 14.81 17.89 20.44 21.53
19.5
20.66 19.56 22.17 17.41 17.56 21.35 20.24
Band
13 14 15 16 17 18 19 20 21 22 23 24
L0
(nm): 613.3 625.1 637.5 649.6 663.8 676.9 683.5
698
705.5 711.4 717.5 723.8
FWHM (nm):
25.3
27.63 24.59 27.86 26.75
27
28.92 24.26 24.44 25.12 27.45 27.81
Band
25 26 27 28 29 30 31 32 33 34 35 36
L0
(nm): 738.1 744.9
758
771.5 800.5 813.4
827
840.7 852.9 865.3 879.6 886.5
FWHM (nm): 26.95 25.56 27.78 27.61 23.82 28.28 26.61 26.85 27.54 28.29 25.89 23.69

Remo e Sens. 2018,10, 1082 6 o 32
2.3. Da a P ocesing
2.3.1. Geome ic P ocessing
Geome ic p ocessing included he de e mina ion o he o ien a ions o he images using bundle
block adjus men (BBA) and he gene a ion o pho og amme ic 3D poin cloud. We used Agiso
Pho oscan comme cial so wa e ( e sion 1.2.5) (AgiSo LLC, S . Pe e sbu g, Russia). We p ocessed
he RGB da a sepa a ely o ob ain a good quali y dense poin cloud. To ob ain good o ien a ions o
he FPI images, we pe o med in eg a ed geome ic p ocessing wi h he RGB images and h ee bands
o he FPI images. The o ien a ions o he es o he bands o FPI images we e calcula ed using he
me hod de eloped by Honka aa a e al. [60].
The BBA using Pho oscan was suppo ed wi h i e GCPs, and he es o hem [
27
] we e used
as checkpoin s. The GNSS coo dina es o all UAV images, compu ed using he PPK p ocess, we e
also applied in he BBA. Fo he ai c a images, GNSS da a was no a ailable. The se ings o BBA
we e selec ed so ha ull esolu ion images we e used (quali y se ing: ‘High’). The se ings o he
numbe o key poin s pe image we e 40,000 and he numbe o ie poin s pe image was se a 4000.
Fu he mo e, an au oma ed came a calib a ion was pe o med simul aneously wi h image o ien a ion
(sel -calib a ion). The es ima ed pa ame e s we e ocal leng h, p incipal poin , and adial and angen ial
lens dis o ion. A e ini ial p ocessing, 10% o he poin s wi h he la ges unce ain y and ep ojec ion
e o s we e emo ed au oma ically, and mo e clea ou lie s we e emo ed manually. The ou pu s o
he geome ic p ocess we e he came a pa ame e s (In e io O ien a ion Pa ame e s—IOP), he image
ex e io o ien a ions in he objec coo dina e sys em (Ex e io O ien a ion Pa ame e s—EOP) and he
3D coo dina es o he ie poin s (spa se poin cloud). The spa se poin cloud and he es ima ed IOP
and EOP o h ee FPI bands (band 3: L0 = 520.4 nm; band 11: L0 = 595.9; band 14: L0 = 625.1 nm) we e
used as inpu s in he 3D band egis a ion p ocess [
58
]. The p ocessing achie ed band egis a ion
accu acy be e han 1 pixel o e he a ea.
The canopy heigh model (CHM) was gene a ed using he DSM and digi al e ain model (DTM)
c ea ed by Pho oscan using a simila p ocedu e desc ibed by Viljanen e al. [
25
] (Figu es 2and 3).
Fi s , he dense poin cloud was c ea ed using he quali y pa ame e se ing ‘Ul ahigh’ and dep h il e ing
se ing ‘Mild’; hus, he highes image esolu ion was used in he dense poin cloud gene a ion p ocess.
A e wa ds, all he poin s in he dense poin cloud we e u ilized o in e pola e he DSM. The DTM was
gene a ed om he dense poin cloud using Pho oscan’s au oma ic classi ica ion p ocedu e o g ound
poin s. A i s , he dense poin cloud was 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 e ain model was calcula ed using hese poin s. A e ha ,
all poin s o he dense poin cloud we e checked, and a new poin was added o he g ound class i he
poin was wi hin a ce ain dis ance om he e ain model and i he angle be ween he app oxima ion o
he DTM and he line o connec he new poin on i was less han a ce ain angle. Finally, he DTM was
in e pola ed using he poin s ha we e classi ied as g ound poin s.
The bes pa ame e s o he au oma ic classi ica ion p ocedu e o g ound poin s we e selec ed by
isually compa ing classi ica ion esul s o he o homosaics. Hence, he cell size o 5 m o he lowes poin
selec ion was chosen o all he da ase s. Fo he RGB and FPI da ase s, he maximum angle o 0
◦
and 3
◦
,
espec i ely, and he maximum dis ance o 0.03 m and 0.05 m, espec i ely, we e selec ed. The pa ame e s
a e en i onmen - and senso - esolu ion-speci ic, and hey di e sligh ly om he pa ame e s ha we used
in ou p e ious s udy on a g ass ial si e [
25
] and om he pa ame e s used by Cunli e e al. [
61
] in he
g ass-domina ed-sh ub ecosys ems and by Méndez-Ba oso e al. [62] in he o es en i onmen .
The geome ic p ocessing indica ed good esul s (Tables 4and 5; Figu es 2and 3). The ep ojec ion
e o s we e wi hin 0.46–1.59 pixels. We used 27 checkpoin s o e alua e he accu acy o he p ocessing
o he ba ley da ase s and 4 checkpoin s o he g ass da ase s. The RMSEs in X and Y coo dina es
we e 1.3–11.3 cm and 5.5–50.9 cm in heigh (Table 5). A lowe lying heigh esul ed in a smalle GSD
and also a highe poin densi y. Addi ionally, inc easing he lying heigh inc eased he RMSEs in a
consis en way. Fo example, in he case o he g ass ield, he RMSE in Z coo dina e was 6.9 cm and
Remo e Sens. 2018,10, 1082 7 o 32
13.8 cm o he lying heigh s o 50 m and 140 m, espec i ely. Fo he ai c a RGB da ase s, he RMSEs
in Z-coo dina e we e 9 cm and 14 cm o he lying heigh s o 450 m and 900 m, espec i ely (Table 5).
The accu acy o he ba ley CHMs we e e alua ed using he plan heigh measu emen s o
he sample plo s as e e ence and calcula ing linea eg essions be ween hem (Table 5). The 90 h
pe cen ile o he CHM was used as he heigh es ima e ( o mula in Sec ion 2.4.2). The bes RMSEs we e
7.3 cm o he da ase cap u ed using a 140 m lying heigh (‘Ba ley UAV 140 m (RGB)’) (Figu e 2a).
The ai c a -based CHMs o he RGB image y (’Ba ley AC 450 m (RGB)’, ‘Ba ley AC 900 m (RGB)’)
also appea ed o be non-de o med, bu showed lowe canopy heigh s (RMSE: 9.7–10.3 cm) han he
UAV-based RGB image y CHMs (Figu e 2a,b,c). In he UAV FPI, CHM s iping ha ollowed he
ligh lines appea ed. This indica ed ha he block was de o med (Figu e 2d), and he RMSE o CHM
(12.7 cm) was sligh ly wo se han he RGB image y CHMs. The ai c a FPI-based CHM was clea ly
de o med and noisie (Figu e 2e); i also had he wo s RMSE (50.9 cm). De o ma ion o he FPI-based
CHMs was caused by he poo e spa ial and adiome ic esolu ion o he images. Excep o he
poo -quali y da ase o CHM “Ba ley AC 700 m (FPI)”, he bias was nega i e o all da ase s, which
indica ed ha CHM was unde es ima ing he eal heigh o he c op, which is gene ally an expec ed
esul [25] (Table 5).
Table 4.
Da ase pa ame e s: GSD: G ound Sampling Dis ance, FH: Fligh Heigh , O e laps in : ligh
di ec ion and c : c oss- ligh di ec ions; N Images: Numbe o Images, Re-p ojec ion e o and Poin densi y.
Da ase GSD FH O e lap ; c N Re-P ojec ion Poin Densi y
(m) (m) (%) Images e o (pix) poin s/m2
G ass UAV 140 m (RGB) 0.037 140 93;82 375 1.59 325
G ass UAV 140 m (RGBFPI) 0.14 140 760 1.13
G ass UAV 50 m (RGB) 0.013 50 86;77 468 0.77 2230
G ass UAV 50 m (RGBFPI) 0.05 50 586 1.06
Ba ley UAV 140 m (RGB) 0.037 140 90;75 500 0.79 297
Ba ley UAV 140 m (RGBFPI) 0.14 140 2034 0.46
Ba ley UAV 140 m (FPI) 0.14 140 79;65 1196 0.5 58.5
Ba ley AC 450 m (RGB) 0.05 450 76;66 160 0.63 380
Ba ley AC 900 m (RGB) 0.1 900 73;72 56 0.6 98.3
Ba ley AC 700 m (FPI) 0.62 700 78;68 1604 0.72 2.6
Table 5.
RMSE: Roo Mean Squa e E o s o X, Y, Z and 3D coo dina es we e calcula ed using 27 check
poin s in Ba ley da ase s and 4 in G ass da ase s. CHM (Canopy heigh model) s a is ics (Mean: a e age
canopy heigh , S d: S anda d de ia ion o canopy heigh s; PCC: Pea son Co ela ion Coe icien o
linea eg ession o e e ence and CHM-heigh s, RMSE and Bias: a e age e o ) we e calcula ed
compa ing 90 h pe cen ile o CHM in sample plo s and g ound e e ence da a.
Da ase Check Poin s RMSE (cm) CHM S a is ics in Sample Plo s (cm)
X Y Z 3D Mean S d PCC RMSE Bias
G ass UAV 140 m (RGB) 3.7 2.7 13.8 4.49
G ass UAV 50 m (RGB) 1.3 1.7 6.9 3.15
Ba ley UAV 140 m (RGB) 4 2.9 5.5 3.52 9.04 6.44 0.87 7.33 −
3.63
Ba ley UAV 140 m (FPI) 8.3 11.3 10.8 5.51 5.15 7.46 0.43 12.71 −
7.52
Ba ley AC 450 m (RGB) 3.6 6.5 9 4.37 6.92 6.68 0.63 10.34 −
5.75
Ba ley AC 900 m (RGB) 6.2 7.5 13.9 5.25 9.29 8.00 0.58 9.67 −
3.38
Ba ley AC 700 m (FPI) 2.4 4.5 23.2 5.49
44.68 39.94
0.12 50.96
32.02
Remo e Sens. 2018,10, 1082 8 o 32
Remo e Sens. 2018, 10, x FOR PEER REVIEW 8 o 33
Figu e 2. CHMs (Canopy heigh model) wi h ba ley c op es ima es om di e en da ase s: (a) Ba ley
UAV 140 m (RGB) (b) Ba ley AC 450 m (RGB) (c) Ba ley AC 900 m (RGB) (d) Ba ley UAV 140 m (FPI)
(e) Ba ley AC 700 m (FPI).
(a)
(b)
Figu e 3. CHMs (Canopy heigh model) wi h g ass es ima es om di e en da ase s: (a) G ass UAV
50 m (RGB) (b) G ass UAV 140 m (RGB).
Figu e 2.
CHMs (Canopy heigh model) wi h ba ley c op es ima es om di e en da ase s: (
a
) Ba ley
UAV 140 m (RGB) (
b
) Ba ley AC 450 m (RGB) (
c
) Ba ley AC 900 m (RGB) (
d
) Ba ley UAV 140 m (FPI)
(e) Ba ley AC 700 m (FPI).
Remo e Sens. 2018, 10, x FOR PEER REVIEW 8 o 33
Figu e 2. CHMs (Canopy heigh model) wi h ba ley c op es ima es om di e en da ase s: (a) Ba ley
UAV 140 m (RGB) (b) Ba ley AC 450 m (RGB) (c) Ba ley AC 900 m (RGB) (d) Ba ley UAV 140 m (FPI)
(e) Ba ley AC 700 m (FPI).
(a)
(b)
Figu e 3. CHMs (Canopy heigh model) wi h g ass es ima es om di e en da ase s: (a) G ass UAV
50 m (RGB) (b) G ass UAV 140 m (RGB).
Figu e 3.
CHMs (Canopy heigh model) wi h g ass es ima es om di e en da ase s: (
a
) G ass UAV
50 m (RGB) (b) G ass UAV 140 m (RGB).
Remo e Sens. 2018,10, 1082 9 o 32
2.3.2. Radiome ic P ocessing
Radiome ic p ocessing o he hype spec al da ase s was ca ied ou using FGI’s RadBA
so wa e [
7
,
63
]. The objec i e o he adiome ic co ec ion was o p o ide accu a e e lec ance
o homosaics. The adiome ic modelling app oach de eloped a he FGI included senso co ec ions,
a mosphe ic co ec ion, co ec ion o adiome ic nonuni o mi ies due o he illumina ion changes,
and he no maliza ion o he objec e lec ance aniso opy due o illumina ion and iewing di ec ion
ela ed nonuni o mi ies using bidi ec ional e lec ance dis ibu ion unc ion (BRDF) co ec ion.
Fi s he senso esponse was co ec ed o he FPI images using he da k signal co ec ion
and he pho on esponse nonuni o mi y co ec ion (PRNU) [
6
,
7
]. The da k signal co ec ion was
calcula ed using a black image collec ed igh be o e he da a cap u e wi h a co e ed lens, and he
PRNU co ec ion was de e mined in he labo a o y.
The empi ical line me hod [
64
] was used o calcula e he ans o ma ion om g ey alues in
images (DN) o e lec ance (Re l) o each channel sol ing ollowing o mula:
DN =aabsRe l +babs (1)
whe e
aabs
and
babs
a e he pa ame e s o he ans o ma ion. Two e e ence e lec ance panels (nominal
e lec ance 0.03 and 0.10), which we e measu ed wi h ASD du ing ield wo k, in he es a ea we e
used o de e mine he pa ame e s.
Because o a iable wea he condi ions du ing he ime o he measu emen and o he adiome ic
phenomena, addi ional adiome ic co ec ions we e necessa y o ob ain uni o m o homosaics.
The basic p inciple o he me hod is o use he DNs o he adiome ic ie poin s in he o e lapping
images as obse a ions and o de e mine he model pa ame e s desc ibing he di e ences in DNs
in di e en images ( he adiome ic model) indi ec ly ia he leas squa es p inciple. The model o
e lec ance was
Rjk(θi,θ ,ϕ) = ( DNjk
a el j
−babs)/aabs (2)
whe e
Rjk(θi,θ ,ϕ)
is he bi-di ec ional e lec ance ac o (BRF) o he objec poin , k, in image j;
θi
and
θ
a e he illumina ion and e lec ed ligh (obse a ion) zeni h angles,
ϕi
and
ϕ
a e he azimu h angles,
espec i ely, and
ϕ=ϕ −ϕi
is he ela i e azimu h angle and
a el j
is he ela i e co ec ion pa ame e
wi h espec o he e e ence image. The pa ame e s used can be selec ed acco ding o he demands o
he da ase in conside a ion.
In he case o mul iple ligh s in he UAV based ba ley da ase , he ini ial alue
a el j
was based on
he i adiance measu emen s by he ASD and in o ma ion abou in eg a ion (exposu e) ime used in
image acquisi ion:
a elj=ASDj(nm)
ASD e (nm)×ITj
IT e
(3)
whe e ASD
j
and ASD
e
a e he i adiance measu emen s and IT
j
and IT
e
in eg a ion ime o senso
du ing he acquisi ion o image jand e e ence image e . This alue was u he enhanced in he
adiome ic block adjus men .
A p io i alues and s anda d de ia ions used in his s udy (Table 6) we e selec ed based on
sugges ions by Honka aa a e al. [
7
,
63
]. Du ing he d one-based g ass da a collec ion, wea he was
mainly sunny (see Table 2); he e o e, we used he BRDF co ec ion o compensa e o he e lec ance
aniso opy e ec s. Fo he ba ley da ase s cap u ed by he d one and ai c a , he aniso opy e ec s
did no appea due o he cloudy wea he du ing da a collec ion. In all da ase s, i was possible o
lea e some de ian images unused om he o homosaics because o he good o e laps be ween he
images. These included some pa ially shaded images due o clouds in he case o he g ass da ase
and some images collec ed unde sunshine in he case o he ba ley da ase .
The adiome ic block adjus men imp o ed he uni o mi y o he image o homosaics, bo h
s a is ically (Figu e 4) and isually (Figu e 5). Fo he unco ec ed da a, he coe icien o a ia ion
Remo e Sens. 2018,10, 1082 16 o 32
Remo e Sens. 2018, 10, x FOR PEER REVIEW 15 o 33
ea u es, since adding he 3D ea u es did no signi ican ly imp o e he es ima ion esul s. In he
cases wi h he RGB came a, he RGB spec al ea u es yielded be e es ima ion accu acy han only
using 3D ea u es, and combining bo h ga e sligh ly be e es ima ion accu acy. Fo example, o he
FY, he PCC and RMSE% we e 0.95% and 34.5%, espec i ely, o he combined RGB ea u es (‘RGB
all’).
In he cases wi h he ai c a da ase s, he bes esul s we e ob ained when using he RGB
spec al ea u es o a combina ion o he RGB spec al and 3D ea u es (cases: ‘RGB all’ and ‘RGB
spec’). The lying heigh o 900 m ga e sligh ly be e esul s. A bes , he PCC and RMSE% we e
0.96% and 31.5%, espec i ely, in he FY es ima ion. The es ima ion esul s we e poo e wi h he FPI
came a han wi h he RGB came a. This was possibly due o he a ying illumina ion condi ions
du ing he FPI-came a ligh , which did no p o ide su icien ly good da a quali y.
In all cases, he es ima ions wi h only he 3D ea u es yielded he wo s esul s. The es ima ion
o FY was mo e accu a e han he es ima ion o DMY. The RF pe o med well wi h a ious ea u es
and combina ions and p o ided in mos cases be e esul s han he SLR, bu when a limi ed
numbe o ea u es om one senso (‘RGB 3D’ and ‘RGB spe’) was used, he SLR yielded be e
es ima ion esul s han he RF (Appendix A; Table A1).
RF p o ided impo ance o de o he di e en ea u es used in he expe imen s. In mos cases
he indices (such as Cl- ed-edge) we e mo e signi ican spec al ea u es han single e lec ance
bands. F om he 3D ea u es pe cen iles, p90 was he mos impo an in many cases (Appendix B;
Tables A4–A5).
Figu e 7. RMSE% o esh (FY) and d y (DMY) biomass using a ied ea u e se s in he ba ley es
ield. pi/FPI: FPI came a; spec: spec al ea u es; RBA: adiome ic block adjus men ; all: all ea u es
(spec al and 3D); RGB: RGB came a; 3d: 3D ea u es; UAV: unmanned ae ial ehicle; AC: Cessna
manned ai c a .
Table 10. Pea son Co ela ion Coe icien s (PCC), Roo mean squa ed e o (RMSE) and RMSE% o
esh (FY) and d y (DMY) biomass using a ied ea u e se s in ba ley es ield. pi/FPI: FPI came a;
spec: spec al ea u es; RBA: adiome ic block adjus men ; all: all ea u es (spec al and 3D); RGB:
RGB came a; 3d: 3D ea u es; UAV: unmanned ae ial ehicle; AC: Cessna manned ai c a .
FY Ba ley
DMY Ba ley
CC
RMSE
RMSE%
CC
RMSE
RMSE%
Flying heigh 140 m UAV
pi spec
0.911
0.219
47.1
0.891
0.035
48.3
pi spec RBA
0.956
0.156
33.6
0.94
0.026
36
pi all
0.910
0.224
48.1
0.885
0.037
49.8
pi all RBA
0.955
0.159
34.2
0.941
0.026
35.9
RGB 3d
0.867
0.255
54.9
0.852
0.04
54.9
RGB spec
0.939
0.177
38.0
0.914
0.031
42.6
Figu e 7.
RMSE% o esh (FY) and d y (DMY) biomass using a ied ea u e se s in he ba ley es
ield. pi/FPI: FPI came a; spec: spec al ea u es; RBA: adiome ic block adjus men ; all: all ea u es
(spec al and 3D); RGB: RGB came a; 3d: 3D ea u es; UAV: unmanned ae ial ehicle; AC: Cessna
manned ai c a .
3.1.2. Ni ogen
In he case o he UAV da ase s, he bes es ima ion accu acy o he ba ley ni ogen amoun and
N% we e ob ained when ea u es om bo h senso s we e applied (‘all RBA’) and wi h he FPI-based
adiome ically co ec ed spec al ea u es (‘ pi spec RBA’) (Figu e 8, Table 11). The adiome ic
calib a ion o he FPI da a clea ly imp o ed he es ima ion accu acy. The bes PCC and RMSE% we e
0.97% and 21.6% o he ni ogen amoun , espec i ely, and 0.92% and 34.4% o he N%, espec i ely.
In he case o he UAV RGB senso , he bes accu acy was achie ed wi h he combined da a (‘RGB
all’). The bes PCC and RMSE% we e sligh ly wo se han wi h he FPI da a—0.94% and 25.2% o he
ni ogen amoun , espec i ely, and 0.92% and 34.5% o he N%, espec i ely.
The ai c a based FPI da ase s p esen ed he wo s es ima ion accu acy, whe eas he es ima ion
esul s wi h he RGB ea u es we e a he same le el as he esul s o he UAV es ima ion. Fo example,
he PCC and RMSE% we e 0.95% and 25.2%, espec i ely, o he ni ogen amoun and 0.94% and
28.7% o he N% wi h he RGB spec al ea u es (‘RGB spec’) wi h he 450 m ligh heigh .
The es ima ion o he ni ogen amoun was mo e accu a e han he es ima ion o he N%.
Rega ding he impo ance o he ea u es, he indi idual e lec ance bands a he ed-edge (670–710 nm)
we e he mos impo an especially o he es ima ion o N%. In many cases he indices we e also
conside ed as he mos impo an . The pe cen iles (3D ea u es) we e impo an in many cases
(Appendix B, Tables A6 and A7).
Remo e Sens. 2018, 10, x FOR PEER REVIEW 16 o 33
RGB all
0.951
0.162
34.7
0.935
0.027
37.4
pi spec; RGB 3d
0.939
0.187
40.2
0.924
0.03
41.2
pi spec RBA; RGB 3d
0.964
0.144
31.0
0.95
0.024
33.2
all
0.947
0.178
38.2
0.928
0.029
40.1
all RBA
0.966
0.141
30.4
0.95
0.024
33.3
Flying heigh 450–700 m AC
pi spec RBA
0.853
0.271
58.2
0.815
0.045
60.9
pi all RBA
0.841
0.280
60.1
0.813
0.045
61.3
RGB 3d
0.656
0.389
83.7
0.595
0.063
85.3
RGB spec
0.932
0.187
40.1
0.919
0.03
41.2
RGB all
0.921
0.201
43.2
0.903
0.033
45.1
pi spec RBA; RGB 3d
0.862
0.265
56.9
0.825
0.044
59.7
all; RBA
0.920
0.210
45.2
0.901
0.034
46.7
Flying heigh 900 m AC
RGB 3d
0.828
0.291
62.7
0.818
0.045
60.9
RGB spec
0.941
0.175
37.6
0.918
0.031
41.6
RGB all
0.962
0.146
31.5
0.94
0.027
36.2
3.1.2. Ni ogen
In he case o he UAV da ase s, he bes es ima ion accu acy o he ba ley ni ogen amoun
and N% we e ob ained when ea u es om bo h senso s we e applied (‘all RBA’) and wi h he
FPI-based adiome ically co ec ed spec al ea u es (‘ pi spec RBA’) (Figu e 8, Table 11). The
adiome ic calib a ion o he FPI da a clea ly imp o ed he es ima ion accu acy. The bes PCC and
RMSE% we e 0.97% and 21.6% o he ni ogen amoun , espec i ely, and 0.92% and 34.4% o he
N%, espec i ely. In he case o he UAV RGB senso , he bes accu acy was achie ed wi h he
combined da a (‘RGB all’). The bes PCC and RMSE% we e sligh ly wo se han wi h he FPI
da a—0.94% and 25.2% o he ni ogen amoun , espec i ely, and 0.92% and 34.5% o he N%,
espec i ely.
The ai c a based FPI da ase s p esen ed he wo s es ima ion accu acy, whe eas he es ima ion
esul s wi h he RGB ea u es we e a he same le el as he esul s o he UAV es ima ion. Fo
example, he PCC and RMSE% we e 0.95% and 25.2%, espec i ely, o he ni ogen amoun and
0.94% and 28.7% o he N% wi h he RGB spec al ea u es (‘RGB spec’) wi h he 450 m ligh heigh .
The es ima ion o he ni ogen amoun was mo e accu a e han he es ima ion o he N%.
Rega ding he impo ance o he ea u es, he indi idual e lec ance bands a he ed-edge (670–710
nm) we e he mos impo an especially o he es ima ion o N%. In many cases he indices we e
also conside ed as he mos impo an . The pe cen iles (3D ea u es) we e impo an in many cases
(Appendix B, Tables A6 and A7).
Figu e 8. RMSE% o ni ogen (N) and Ni ogen-% in ba ley es ield. pi/FPI: FPI came a; spec:
spec al ea u es; RBA: adiome ic block adjus men ; all: all ea u es (spec al and 3D); RGB: RGB
came a; 3d: 3D ea u es; UAV: unmanned ae ial ehicle; AC: Cessna manned ai c a .
Figu e 8.
RMSE% o ni ogen (N) and Ni ogen-% in ba ley es ield. pi/FPI: FPI came a; spec:
spec al ea u es; RBA: adiome ic block adjus men ; all: all ea u es (spec al and 3D); RGB: RGB
came a; 3d: 3D ea u es; UAV: unmanned ae ial ehicle; AC: Cessna manned ai c a .

Remo e Sens. 2018,10, 1082 17 o 32
Table 11.
Pea son Co ela ion Coe icien s (PCC), Roo mean squa ed e o (RMSE) and RMSE%
o ni ogen (N) and Ni ogen-% in ba ley es ield. pi/FPI: FPI came a; spec: spec al ea u es;
RBA: adiome ic block adjus men ; all: all ea u es (spec al and 3D); RGB: RGB came a; 3d: 3D
ea u es; UAV: unmanned ae ial ehicle; AC: Cessna manned ai c a .
N Ba ley N% Ba ley
CC RMSE RMSE% CC RMSE RMSE%
Flying heigh 140 m UAV
pi spec 0.9131
0.0009
32.4
0.8287
0.82 48.0
pi spec RBA 0.9643
0.0006
21.6 0.863 0.75 43.6
pi all 0.9072 0.001 36.0
0.8234
0.84 48.9
pi all RBA 0.962
0.0006
21.6
0.8701
0.73 42.6
RGB 3d 0.8732
0.0011
39.6 0.908 0.62 35.9
RGB spec 0.9373
0.0008
28.8
0.8501
0.77 45.1
RGB all 0.9441
0.0007
25.2
0.9157
0.59 34.5
pi spec; RGB 3d 0.9381
0.0008
28.8
0.9168
0.59 34.5
pi spec RBA; RGB 3d 0.9648
0.0006
21.6
0.9125
0.61 35.6
all 0.9451
0.0008
28.8
0.9185
0.59 34.4
all RBA 0.9663
0.0006
21.6
0.9091
0.62 36.3
Flying heigh 450–700 m AC
pi spec RBA 0.8522
0.0012
43.2
0.6267
1.15 67.3
pi all RBA 0.8413
0.0012
43.2
0.6146
1.17 68.1
RGB 3d 0.6388
0.0017
61.2
0.4862
1.32 77.2
RGB spec 0.9453
0.0007
25.2
0.9427
0.49 28.7
RGB all 0.9329
0.0008
28.8
0.9332
0.54 31.3
pi spec RBA; RGB 3d 0.8653
0.0011
39.6
0.6597
1.11 64.6
all; RBA 0.925
0.0009
32.4
0.8534
0.78 45.5
Flying heigh 900 m AC
RGB 3d 0.7694
0.0014
50.4
0.5832
1.23 71.6
RGB spec 0.953
0.0007
25.2
0.8793
0.70 41.0
RGB all 0.9682
0.0006
21.6 0.886 0.68 39.8
3.2. G ass Pa ame e Es ima ion
The a ia ion o he biomass and ni ogen amoun was low and we had a limi ed amoun o
g ound samples a ailable in he g ass es ield. We e alua ed he pe o mance using he a e age
o he samples as he es ima e. The RF p o ided be e esul s han using he a e age alue o he
biomass es ima ion, whe eas o he ni ogen amoun he a e age was as good as he RF. The e o e, we
only s udied he biomass es ima ion. The da ase s we e cap u ed om he lying heigh s o 50 m and
140 m using he UAV.
The gene al iew o he esul s is ha he es ima ion e o s we e low because o he small a ia ion
in he da ase s (Table 1). The bes PCC and RMSE% we e 0.640% and 4.29% o he FY, espec i ely,
and 0.79% and 1.91% o he DMY, espec i ely (Figu e 9, Table 12). These esul s we e ob ained wi h
he RGB came a spec al ea u es (‘RGB spec’) cap u ed om he lying heigh o 140 m. Wi h he
FPI came a, he bes esul s we e nea ly as good, and hey we e ob ained wi h he adiome ically
co ec ed da ase (‘FPI spec RBA’) om he lying heigh o 50 m. In his case, he PCC and he
RMSE% we e 0.538% and 4.63% o he FY, espec i ely and 0.72% and 2.09% o he DMY, espec i ely.
The adiome ic co ec ion wi h he adiome ic block adjus men sligh ly imp o ed he es ima ion
accu acy in he case o he 50 m da ase , bu i did no impac he esul s o he da ase om he 140 m
lying heigh . I was expec ed ha he adiome ic co ec ion would imp o e he es ima ion esul s
wi h he 50 m lying heigh da ase because i clea ly imp o ed he uni o mi y o he o homosaics
om he 50 m lying heigh , while o he o homosaics om he 140 m lying heigh he co ec ion
had a mino impac (Figu e 5). The 3D ea u es alone p o ided he poo es es ima ion esul s o
Remo e Sens. 2018,10, 1082 18 o 32
bo h lying heigh s (‘RGB 3D’), and hei use oge he wi h he spec al ea u es did no imp o e he
es ima ion accu acy. The impac o he lying heigh was mino . The FPI came a da ase wi h he 50 m
lying heigh p o ided be e esul s han he da ase wi h he 140 m ligh heigh . And o he RGB
came a, he 140 m da ase p o ided sligh ly be e esul s. The es ima ion accu acies we e be e o
he DMY han o he FY.
Simila o ba ley analysis, he indices, om spec al ea u es, and pe cen iles, om 3D ea u es,
we e he mos ep esen a i e ea u es o mos o he cases o he g ass es ima ion analysis (Appendix B;
Tables A8 and A9).
Remo e Sens. 2018, 10, x FOR PEER REVIEW 18 o 33
m ligh heigh . And o he RGB came a, he 140 m da ase p o ided sligh ly be e esul s. The
es ima ion accu acies we e be e o he DMY han o he FY.
Simila o ba ley analysis, he indices, om spec al ea u es, and pe cen iles, om 3D ea u es,
we e he mos ep esen a i e ea u es o mos o he cases o he g ass es ima ion analysis
(Appendix B; Tables A8 and A9).
Figu e 9. RMSE% o esh (FY) and d y (DMY) biomass using da a collec ed om 50 m and 140 m
lying heigh in g ass es ield. pi/FPI: FPI came a; spec: spec al ea u es; RBA: adiome ic block
adjus men ; all: all ea u es (spec al and 3D); RGB: RGB came a; 3d: 3D ea u es; UAV: unmanned
ae ial ehicle; AC: Cessna manned ai c a .
Table 12. Pea son Co ela ion Coe icien s (PCC), Roo mean squa ed e o (RMSE) and RMSE%, o
esh (FY) and d y (DMY) biomass using da a collec ed om 50 m and 140 m lying heigh in g ass
es ield. pi/FPI: FPI came a; spec: spec al ea u es; RBA: adiome ic block adjus men ; all: all
ea u es (spec al and 3D); RGB: RGB came a; 3d: 3D ea u es; UAV: unmanned ae ial ehicle; AC:
Cessna manned ai c a .
PCC
RMSE
RMSE%
PCC
RMSE
RMSE%
FY G ass
FY
DMY G ass
DMY
A e age
0.100
5.8
0.015
3.2
Flying heigh 50 m
FPI spec
0.443
0.085
4.9
0.722
0.010
2.1
FPI spec RBA
0.538
0.080
4.6
0.722
0.010
2.1
RGB 3D
0.410
0.089
5.1
0.103
0.016
3.4
RGB spe
0.395
0.089
5.1
0.706
0.010
2.2
RGB all
0.433
0.086
5.0
0.415
0.014
2.8
FPI spec; RGB 3D
0.451
0.085
4.9
0.644
0.011
2.3
FPI spec RBA; RGB 3D
0.493
0.083
4.8
0.668
0.011
2.3
Flying heigh 140 m
FPI spec
0.446
0.085
4.9
0.711
0.011
2.2
FPI spec RBA
0.433
0.086
5.0
0.645
0.011
2.4
RGB 3D
0.213
0.093
5.4
0.097
0.015
3.1
RGB spe
0.640
0.074
4.3
0.786
0.009
1.9
RGB all
0.542
0.080
4.6
0.632
0.011
2.4
FPI spec; RGB 3D
0.436
0.085
4.9
0.683
0.011
2.3
FPI spec RBA; RGB 3D
0.441
0.085
4.9
0.601
0.012
2.5
4. Discussion
We de eloped and assessed a machine lea ning echnique in eg a ing 3D and spec al ea u es
o he es ima ion o esh and d y ma e yield (FY, DMY), ni ogen amoun and ni ogen
pe cen age (N%) o mal ba ley c op and g ass silage ields. Ou app oach was o ex ac a a ie y o
emo e sensing ea u es om he da ase s ha we e collec ed using RGB and imaging hype spec al
Figu e 9.
RMSE% o esh (FY) and d y (DMY) biomass using da a collec ed om 50 m and 140 m
lying heigh in g ass es ield. pi/FPI: FPI came a; spec: spec al ea u es; RBA: adiome ic block
adjus men ; all: all ea u es (spec al and 3D); RGB: RGB came a; 3d: 3D ea u es; UAV: unmanned
ae ial ehicle; AC: Cessna manned ai c a .
Table 12.
Pea son Co ela ion Coe icien s (PCC), Roo mean squa ed e o (RMSE) and RMSE%, o
esh (FY) and d y (DMY) biomass using da a collec ed om 50 m and 140 m lying heigh in g ass es
ield. pi/FPI: FPI came a; spec: spec al ea u es; RBA: adiome ic block adjus men ; all: all ea u es
(spec al and 3D); RGB: RGB came a; 3d: 3D ea u es; UAV: unmanned ae ial ehicle; AC: Cessna
manned ai c a .
PCC RMSE RMSE% PCC RMSE RMSE%
FY G ass FY DMY G ass DMY
A e age 0.100 5.8 0.015 3.2
Flying heigh 50 m
FPI spec 0.443 0.085 4.9 0.722 0.010 2.1
FPI spec RBA 0.538 0.080 4.6 0.722 0.010 2.1
RGB 3D 0.410 0.089 5.1 0.103 0.016 3.4
RGB spe 0.395 0.089 5.1 0.706 0.010 2.2
RGB all 0.433 0.086 5.0 0.415 0.014 2.8
FPI spec; RGB 3D 0.451 0.085 4.9 0.644 0.011 2.3
FPI spec RBA; RGB 3D 0.493 0.083 4.8 0.668 0.011 2.3
Flying heigh 140 m
FPI spec 0.446 0.085 4.9 0.711 0.011 2.2
FPI spec RBA 0.433 0.086 5.0 0.645 0.011 2.4
RGB 3D 0.213 0.093 5.4 0.097 0.015 3.1
RGB spe 0.640 0.074 4.3 0.786 0.009 1.9
RGB all 0.542 0.080 4.6 0.632 0.011 2.4
FPI spec; RGB 3D 0.436 0.085 4.9 0.683 0.011 2.3
FPI spec RBA; RGB 3D 0.441 0.085 4.9 0.601 0.012 2.5
Remo e Sens. 2018,10, 1082 19 o 32
4. Discussion
We de eloped and assessed a machine lea ning echnique in eg a ing 3D and spec al ea u es
o he es ima ion o esh and d y ma e yield (FY, DMY), ni ogen amoun and ni ogen pe cen age
(N%) o mal ba ley c op and g ass silage ields. Ou app oach was o ex ac a a ie y o emo e
sensing ea u es om he da ase s ha we e collec ed using RGB and imaging hype spec al came as.
The ea u es included 3D ea u es om he canopy heigh model (CHM) and spec al ea u es as a
spec al, and a ious ege a ion indices (VI) om he o homosaics. Fu he mo e, we in es iga ed he
impac o he adiome ic co ec ion and lying heigh on he es ima ion esul s. Ou app oach was o
use he Random Fo es es ima o (RF), bu he esul s o he Simple Linea Reg ession (SLR) es ima o
was also calcula ed o alida e he pe o mance o he RF.
The bes es ima ion esul s o he ba ley biomass and ni ogen con en es ima ions we e ob ained
by combining ea u es om he FPI and RGB came as. In mos cases, he spec al ea u es om he
FPI came a p o ided he mos o nea ly he mos accu a e esul s. Adding he FPI came a 3D ea u es
did no imp o e he esul s, which was an expec ed esul since FPI based CHM did no ha e high
quali y (Table 5, Figu e 2d) due o ela i ely la ge GSD o 0.14 m. The da a om he RGB came a
p o ided good es ima ion esul s— ypically almos as good as he FPI came a and in some cases he
bes esul s. We could also obse e ha he combina ion o RGB spec al and 3D ea u es imp o ed
he es ima ion accu acy, especially in he case o biomass es ima ion. The RF pe o med well wi h
a ious ea u es and combina ions and p o ided in mos cases be e esul s han he SLR, bu some
excep ions also appea ed (Appendix A; Tables A1–A3). Especially when only a limi ed numbe o
ea u es om one senso (‘RGB 3D’ and ‘RGB spe’) was used, he SLR yielded compe i i e o e en
be e es ima ion esul s han he RF, bu when he amoun and a ia ion o ea u es was high, he RF
p o ided egula ly be e es ima ion esul s han he SLR. This is a logical pe o mance, because wi h
a small numbe o ea u es he e a e no g ea di e ence in he SLR and RF models, bu wi h la ge
numbe o ea u es, he SLR s ill uses only single ea u e in he es ima ion bu RF can ake ad an age
o a ious ea u es du ing model building. A simila obse a ion was also made by Li e al. [
32
], whe e
he d y biomass o maize was es ima ed; hey ob ained an R
2
o 0.52 and an RMSE% o 18.8% wi h
SLR and an R
2
o 0.78 and an RMSE% o 16.7% wi h he RF. The RF hus p o ided mo e accu a e
es ima ion esul s. They also concluded ha pho og amme ic 3D ea u es s ongly con ibu ed o
he es ima ion models, in addi ion o he spec al ea u es om he RGB came a. They sugges ed
ha hype spec al da a could imp o e he es ima ion esul s, and ou s udy showed ha his was a
alid assump ion in many si ua ions. Yue e al. [
51
] compa ed eigh di e en eg ession echniques
o he win e whea biomass es ima ion, using nea -su ace spec oscopy and achie ed R
2
alues o
0.79–0.89. They concluded ha machine lea ning echniques such as RF we e less sensi i e o noise
han con en ional eg ession echniques.
In he biomass es ima ions o ba ley, he PCC and RMSE% we e a bes 0.95% and 33.2%,
espec i ely, o he DMY, and 0.97% and 31.0%, espec i ely, o he FY. The co esponding s a is ics
o he g ass da ase wi h he 140 m lying heigh we e 0.79% and 1.9% o he DMY, and 0.64% and
4.3% o he FY, and o he da ase wi h he lying heigh o 50 m, he esul s we e on he same le el.
Conce ning he impac s o di e en ea u es used in he es ima ions o ba ley DMY, he inclusion o he
3D ea u es om he RGB came a in addi ion o he spec al ea u es om he FPI came a imp o ed
he RMSEs by 14.7% o uncalib a ed FPI, and 7.95% o calib a ed FPI. The esul s we e he simila o
he ba ley FY. The possible explana ion o his is ha he es ima ion accu acies eached almos he
bes possible quali y wi h he calib a ed spec al ea u es and so he 3D ea u es could no p o ide
u he imp o emen whe eas o he uncalib a ed spec al ea u es hey imp o ed s ill signi ican ly
accu acy. Inclusion o he 3D ea u es based on he FPI came a did no imp o e he accu acy wi h
ei he uncalib a ed o calib a ed da a. The eason o his was he insu icien quali y o he heigh
da a wi h he FPI came a, and he e o e i could no p o ide quan i a i e in o ma ion o di e ences o
a ious samples o he es ima ion p ocess. Conside ing he RGB senso , he 3D ea u es imp o ed he
RMSE% in he DMY and FY es ima ion by 12.18% and 8.6%, espec i ely, o ba ley. The co esponding
Remo e Sens. 2018,10, 1082 20 o 32
imp o emen s we e 24.2% o he g ass DMY and 8.1% o he FY. In he s udy by Bendig e al. [
23
],
adding he heigh ea u es wi h he spec al indices ei he did no imp o e o only sligh ly imp o ed
he es ima ion accu acy o ba ley biomass when using mul ilinea eg ession models. In he s udy
by Yue e al. [
24
], he co ela ion be ween he win e whea d y biomass and he pa ial leas squa es
eg ession (PLS) model based on spec al ea u es was imp o ed om 0.53 o 0.74 and he RMSE om
1.69 o 1.20 /ha when 3D ea u es we e included. These esul s a e compa able o ou esul s o he
ba ley DMY. In s udies wi h spec ally uncalib a ed RGB alues and 3D ea u es, R
2
alues o 0.74 ha e
been epo ed o he co n g ain yield es ima ion [
37
] and 0.88 o he maize biomass es ima ion [
32
],
which indica ed lowe co ela ions han ou esul s using he RGB da a o he ba ley FMY es ima ion
(PCC = 0.95, RMSE% = 34.74%).
In he ni ogen es ima ions o ba ley, he PCC and RMSE% we e a bes 0.966 and 21.6%, espec i ely,
o he ni ogen amoun and 0.919% and 34.4%, espec i ely, o he N%. Conce ning he impac s o
di e en ea u es used in he es ima ions, he inclusion o he RGB came a 3D ea u es wi h he spec al
ea u es o he FPI came a imp o ed he RMSEs (0–30%), which indica ed ha he 3D ea u es p o ided
addi ional in o ma ion o he es ima ion model. Also combining he 3D ea u es o he RGB spec al
ea u es imp o ed he es ima ion accu acy o he ni ogen amoun and he N% by 12.5% and 23.6%,
espec i ely, p o iding simila accu acy as he FPI based spec al ea u es. I is wo h no ing, ha e en
hough he a ia ion o N% on sample e e ences was no high (Table 1), good accu acies we e achie ed.
The a ia ion in he ni ogen amoun was mainly ela ed o a ia ion in he biomass amoun , which
explains he simila es ima ion accu acies o he wo quan i ies. Liu e al. [
50
] used se e al di e en
algo i hms o es ima e he ni ogen con en (N%) o win e whea based on mul ispec al da a and achie ed
he bes esul s wi h an R
2
o 0.79 and an RMSE% o 11.56% wi h he RF. Geipel e al. [
37
] used SLR
models based on a mul ispec al senso o es ima e he N con en and achie ed accu acies wi h an R
2
o
0.58–0.89 and an RMSE% o 7.6–11.7%. Schi mann e al. [
42
] achie ed a he bes R
2
alue o 0.65 be ween
he ni ogen con en and he p incipal componen s o RGB image. Ou esul s we e on he same le el wi h
Liu e al. [
50
], Geipel e al. [
37
] and Schi mann e al. [
42
]; bu wi h e es ial app oaches, e en highe
accu acies ha e been achie ed [52]. Fu he mo e, da a om ac o -moun ed Ya a N-senso has epo ed
good co ela ions o R
2
0.80 wi h N-up ake in g ass swa d [
54
]. Howe e , i is impo an o no ice ha he
es ima ion accu acies o di e en s udies a e no di ec ly compa able because hey a e also impac ed by
he p ope ies o he c op sample da a, such as he a ia ion in hei alues.
When compa ing he es ima ion accu acy wi h he spec al ea u es only om he FPI and RGB
came as, he FPI came a p o ided 15.4% and 18.5% be e RMSEs han he RGB came a o he ba ley DMY
and FY, espec i ely, bu up o 16.5% and 14.4% wo se RMSEs han he RGB came a o he g ass DMY and
FY, espec i ely. Be e pe o mance o he FPI came a was expec ed since he hype spec al images p o ide
mo e spec al in o ma ion han he RGB images. The challenges wi h he g ass s udy we e he small
numbe o samples and he small a ia ion in he biomass amoun , and he e o e he g ass esul s should be
conside ed as indica i e. In he es ima ion o he ni ogen, he FPI came a ou pe o med he RGB came a
by 25.0% o he ba ley ni ogen amoun and by 21.1% o he ba ley N%. The ni ogen con en o plan s is
ela i ely small (Table 1), hus i is expec ed ha hey only sligh ly a ec he spec a. Consequen ly, FPI
p o ides highe accu acy han RGB, because i is collec ing mo e in o ma ion om spec a.
In mos cases, he adiome ic calib a ion o he da ase s using he adiome ic block adjus men
imp o ed he es ima ion esul s. In he case o ba ley pa ame e es ima ions wi h all ea u es, he
adiome ic co ec ion imp o ed he RMSE by 17.0% o he DMY, 20.3% o he FY and 25.0% o
he ni ogen amoun . In he case o he g ass es ima ion, he impac was smalle — he co ec ion
ei he sligh ly dec eased o imp o ed he RMSE by
−
6.3–3.6% o he DMY and 0–2.4% o he FY.
The imp o emen was la ges in he da ase s ha ing many ligh lines (Table 2). The e ec was he mos
no iceable in he ‘Ba ley UAV 140 m’ da ase , which was collec ed du ing 4.5 h, when illumina ion
changed signi ican ly, and in he ‘G ass UAV 50 m’ da ase , which was collec ed du ing sunny
condi ions a a low lying heigh ha caused ema kable aniso opy e ec s (Figu e 5). Mul iple s udies
ha e shown ha adiome ic co ec ion using he RBA me hod imp o ed he uni o mi y o image
Remo e Sens. 2018,10, 1082 21 o 32
o homosaics [
7
,
11
,
12
,
63
,
77
]. Ou esul s showed ha he co ec ions also imp o ed he accu acy o
he c op pa ame e es ima ions.
The ba ley da ase s we e collec ed om he UAV and ai c a using a ious lying heigh s, which
p o ided di e en GSDs. In he case o he RGB came a, he GSDs we e 0.05 and 0.10 m, and he
es ima ion esul s we e simila when spec al ea u es we e applied. Howe e , he lying heigh and
GSD had a signi ican impac on he accu acy o he 3D ea u es, which we could al eady deduce
based on he CHM quali y (Table 5, Figu es 2and 3). The mos eliable CHM was ob ained using
da a wi h he smalles GSD, ‘Ba ley UAV 140 m RGB’, whe e he co ela ion be ween in si u e e ence
measu emen s and he CHM we e highes , e en hough he CHM egula ly unde es ima ed in si u
measu emen s (Figu e 2a). The quali y o he DSM dec eased when he GSD inc eased, and when he
GSD was oo la ge (like in he case o ‘Ba ley AC 700 m FPI’ wi h a GSD o 0.60 m), he 3D ea u es
we e useless. I is also impo an o no ice ha in all cases, he heigh accu acy o he blocks was good
and acco ding o expec a ions—on he le el o 0.5–2 imes he GSD. A he smalles GSDs, he UAV
and ai c a p o ided compa able accu acies. Thus he low-cos senso s used in his s udy can also be
ope a ed om small ai c a . The ad an age o he ai c a -based me hod is ha la ge a eas can be
co e ed mo e e icien ly. Howe e , in smalle a eas d ones a e mo e a o dable.
I is wo h no ing ha in he ba ley ield he g ow h was no ideal due o poo wea he condi ions
a he beginning o he g owing season. In he g ass canopy, he numbe o ield e e ence samples
we e ela i ely low (8 samples) and a ia ion in he biomass and ni ogen amoun s was low, which
gene ally dec eases he co ela ion and es ima ion esul s. Howe e , i we hink p ac ical solu ion
o c op pa ame e es ima ion, collec ion o e en a small numbe o samples is ime-consuming and
inc ease cos s. The esul wi h a ew samples wi h a small a ia ion was sligh ly be e han when using
he a e age alue as he es ima e; his indica ed ha wi h he comp ehensi e machine lea ning me hod
he es ima ion accu acy could be imp o ed om he case o using only a e age alues, as i e ealed
ela i ely small spa ial a ia ions. Al hough we ob ained p omising esul s using da ase s om he
140 m o highe ligh heigh s, he use o lowe heigh da a, and hus mo e p ecise CHMs, can imp o e
he es ima ions, as shown in p e ious s udies using ligh heigh s o 50 m o less [
4
,
21
,
25
]. We assume
ha he spa ial and adiome ic esolu ion o he images a e he undamen al ac o s impac ing he
quali y o CHM hus we expec ha al e na i ely a be e -quali y imaging sys em could also p o ide
good esul s om highe al i udes; his would be ad an ageous i aiming a mapping la ge a eas.
To ou knowledge, his s udy was he i s one ha comp ehensi ely in eg a ed and compa ed
UAV-based hype spec al, RGB and poin cloud ea u es in c op pa ame e es ima ion. We de eloped
an app oach o u ilizing a combina ion o spec al da a and 3D ea u es in he es ima ion p ocess ha
simul aneously and e icien ly u ilizes all a ailable in o ma ion. Fu he mo e, ou esul s showed ha he
in eg a ion o spec al and 3D ea u es imp o ed he accu acy o he biomass es ima ion; bu in he ni ogen
es ima ion, he spec al ea u es we e mo e impo an . The esul s also indica ed ha he hype spec al
da a p o ided only a sligh o no imp o emen o he es ima ion accu acy o he biomass compa ed
o he RGB da a. This esul hus sugges s ha he low-cos RGB senso s a e sui able o he biomass
es ima ion ask. Howe e , mo e s udies a e ecommended o alida e his esul in di e en condi ions.
In he ni ogen es ima ion, he hype spec al da a appea ed o be mo e ad an ageous han he RGB da a.
The ai c a -based da a cap u e also p o ided esul s compa able o he UAV-based esul s.
In he u u e, u he s udies using mo e accu a e hype spec al senso s and highe a iabili y
es si es will be o in e es . The da ase s also gi e possibili ies o new ypes o analysis, such as
u ilizing he spec al DSM mo e igo ously [
21
,
22
] and u ilizing he mul i iew spec al da ase s in
he analysis [
78
–
80
]. Ou u u e objec i e will be o de elop gene alized es ima o s ha can be used
wi hou in si u aining da a, o example, aining an es ima o wi h a da ase om one sample a ea
and hen using i in o he a eas. Va ious machine lea ning echniques exis ha can be used in his
p ocess. Ou esul s showed ha he SLR was no ideal o his ask. The RF beha ed well, and u he
s udies will be necessa y o e alua e i s sui abili y o he gene alized p ocedu es. Fo example, he
deep lea ning neu al ne wo k es ima o s will be e y in e es ing al e na i es [81].

Remo e Sens. 2018,10, 1082 22 o 32
5. Conclusions
We de eloped and assessed a machine lea ning echnique in eg a ing 3D and spec al ea u es o
he es ima ion o esh and d y ma e yield (FY, DMY), ni ogen con en and ni ogen pe cen age (N%)
o ba ley c ops and g ass silages. Ou app oach was o ex ac a la ge numbe o emo e sensing ea u es
om he da ase s, including 3D ea u es om he canopy heigh model (CHM) and hype spec al
ea u es and a ious ege a ion indices (VI) om o homosaics. Fu he mo e, we in es iga ed he
impac o adiome ic co ec ion on he es ima ion esul s. We compa ed he pe o mance o Simple
Linea Reg ession (SLR) and he Random Fo es es ima o (RF). To ou knowledge, his s udy was one
o he i s s udies o in eg a e and compa e UAV-based hype spec al, RGB and poin cloud ea u es
in he es ima ion p ocess. Gene ally, he bes esul s we e ob ained when in eg a ing hype spec al
and 3D ea u es. The in eg a ion o RGB and 3D ea u es also p o ided nea ly as good esul s as he
hype spec al ea u es. The in eg a ion o spec al and 3D ea u es especially imp o ed he biomass
es ima ion esul s. The adiome ic calib a ion imp o ed he es ima ion accu acy, and we expec ha i
will be one o he p e equisi es in de eloping gene alized analysis ools. Ou impo an u u e esea ch
objec i e will be o de elop gene alized es ima ion ools ha do no equi e in si u aining da a.
Au ho Con ibu ions:
The expe imen was designed by E.H. and J.K. J.K. and K.A. co esponded abou he
ag icul u al es ield design and E.H. co esponded abou he emo e sensing es ield design. T.H., R.N., and
N.V. ca ied ou he imaging ligh s. N.V. ca ied ou he geome ic and R.N. and L.M. adiome ic p ocessing
s eps. R.N. de eloped he es ima ion me hods and pe o med he es ima ions; analysis o he esul s was made by
R.N. and E.H. The a icle was w i en by R.N., N.V. and E.H., wi h he assis ance o o he au ho s.
Funding:
This wo k was unded by 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), and he Academy o Finland p ojec
“Quan i a i e emo e sensing by 3D hype spec al UAVs—F om heo y o p ac ice” (g an No. 305994).
Con lic s o In e es : The au ho s decla e no con lic o in e es .
Appendix A
Table A1.
Pea son Co ela ion Coe icien s (PCC), Roo mean squa ed e o (RMSE) and RMSE% o
esh (FY) and d y (DMY) biomass using a ied ea u e se s in ba ley es ield using SLR.
FY Ba ley DMY Ba ley
PCC RMSE RMSE% PCC RMSE RMSE%
Flying heigh 140 m UAV
pi spec 0.924 0.196 42.2 0.872 0.038 51.2
pi spec RBA 0.919 0.203 43.7 0.918 0.030 41.4
pi all 0.873 0.250 53.7 0.872 0.038 51.2
pi all RBA 0.935 0.182 39.1 0.918 0.030 41.4
RGB 3d 0.924 0.196 42.2 0.906 0.032 44.2
RGB spec 0.956 0.151 32.5 0.940 0.026 35.6
RGB all 0.956 0.151 32.5 0.940 0.026 35.6
pi spec; RGB 3d 0.873 0.250 53.7 0.873 0.038 51.2
pi spec RBA; RGB 3d 0.935 0.182 39.1 0.918 0.030 41.4
all 0.956 0.151 32.5 0.940 0.026 35.6
all RBA 0.956 0.151 32.5 0.940 0.026 35.6
Flying heigh 450–700 m AC
pi spec RBA 0.788 0.315 67.8 0.771 0.049 66.6
pi all RBA 0.788 0.315 67.8 0.771 0.049 66.6
RGB 3d 0.731 0.351 75.5 0.689 0.056 76.3
RGB spec 0.919 0.201 43.3 0.911 0.032 43.1
RGB all 0.919 0.201 43.3 0.911 0.032 43.1
pi spec RBA; RGB 3d 0.714 0.361 77.6 0.771 0.049 66.6
all; RBA 0.919 0.201 43.3 0.911 0.032 43.1
Flying heigh 900 m AC
RGB 3d 0.738 0.349 75.0 0.693 0.056 76.3
RGB spec 0.935 0.181 39.0 0.925 0.029 39.6
RGB all 0.935 0.181 39.0 0.925 0.029 39.6
Remo e Sens. 2018,10, 1082 23 o 32
Table A2.
Pea son Co ela ion Coe icien s (PCC), Roo mean squa ed e o (RMSE) and RMSE% o
ni ogen (N) and Ni ogen-% in ba ley es ield using SLR.
N Ba ley N% Ba ley
PCC RMSE RMSE% PCC RMSE RMSE%
Flying heigh 140 m UAV
pi spec 0.927 0.001 28.8 0.770 0.937 54.7
pi spec RBA 0.917 0.001 32.4 0.795 0.893 52.2
pi all 0.834 0.001 43.2 0.770 0.937 54.7
pi all RBA 0.938 0.001 28.8 0.795 0.893 52.2
RGB 3d 0.927 0.001 28.8 0.704 1.046 61.1
RGB spec 0.964 0.001 21.6 0.764 0.948 55.4
RGB all 0.964 0.001 21.6 0.764 0.948 55.4
pi spec; RGB 3d 0.834 0.001 43.2 0.770 0.937 54.7
pi spec RBA; RGB 3d 0.938 0.001 28.8 0.795 0.893 52.2
all 0.964 0.001 21.6 0.722 1.022 59.7
all RBA 0.964 0.001 21.6 0.795 0.893 52.2
Flying heigh 450–700 m AC
pi spec RBA 0.741 0.002 54.0 0.507 1.320 77.1
pi all RBA 0.741 0.002 54.0 0.507 1.320 77.1
RGB 3d 0.725 0.002 54.0 0.228 1.555 90.8
RGB spec 0.949 0.001 25.2 0.847 0.781 45.6
RGB all 0.949 0.001 25.2 0.847 0.781 45.6
pi spec RBA; RGB 3d 0.741 0.002 54.0 0.507 1.320 77.1
all; RBA 0.949 0.001 25.2 0.847 0.781 45.6
Flying heigh 900 m AC
RGB 3d 0.715 0.002 57.6 0.193 1.514 88.5
RGB spec 0.959 0.001 21.6 0.854 0.765 44.7
RGB all 0.959 0.001 21.6 0.854 0.765 44.7
Table A3.
Pea son Co ela ion Coe icien s (PCC), Roo mean squa ed e o (RMSE) and RMSE%,
o esh (FY) and d y (DMY) biomass using da a collec ed om 50 and 140 m lying heigh in g ass
es ield.
FY Ba ley DMY Ba ley
PCC RMSE RMSE% PCC RMSE RMSE%
Flying heigh 50 m
FPI spec 0.332 0.107 6.2 0.436 0.015 3.1
FPI spec RBA 0.712 0.077 4.5 0.847 0.008 1.6
RGB 3D 0.332 0.105 6.1 0.378 0.015 3.1
RGB spe 0.367 0.100 5.8 0.446 0.015 3.1
RGB all 0.203 0.114 6.6 0.446 0.015 3.1
FPI spec; RGB 3D 0.332 0.107 6.2 0.436 0.015 3.1
FPI spec RBA; RGB 3D 0.712 0.077 4.5 0.847 0.008 1.6
Flying heigh 140 m
FPI spec 0.515 0.087 5.0 0.732 0.011 2.3
FPI spec RBA 0.527 0.081 4.7 0.577 0.013 2.6
RGB 3D 0.822 0.055 3.2 0.390 0.017 3.5
RGB spe 0.711 0.069 4.0 0.742 0.012 2.5
RGB all 0.754 0.064 3.7 0.742 0.012 2.5
FPI spec; RGB 3D 0.596 0.088 5.1 0.732 0.011 2.3
FPI spec RBA; RGB 3D 0.683 0.076 4.4 0.577 0.013 2.6
Remo e Sens. 2018,10, 1082 24 o 32
Appendix B
Table A4. The mos impo an ea u es o he Random Fo es (RF) (in he o de o impo ance) o FY es ima ion in ba ley.
Flying heigh 140 m UAV
pi spec Cl-RE Cl-G RDVI MTVI b34 NDVI b19
pi spec RBA b21 b20 Cl-G Cl-RE b17 b15 b19
pi all Cl-RE MTVI Cl-G OSAVI RDVI NDVI b33
pi all RBA Cl-RE Cl-G b20 GNDVI b15 b19 b17
RGB 3d RGB_CHMp90 RGB_CHMp80 RGB_CHMp50 RGB_CHMp70 RGB_CHMmax RGB_CHMmin
RGB_CHMmean
RGB spec RGB-GRVI RGB-ExG RGB-R RGB-B RGB-G NaN NaN
RGB all RGB-R RGB-GRVI RGB-ExG RGB_CHMp90 RGB_CHMp80 RGB-B RGB_CHMp70
pi spec; RGB 3d MTVI Cl-RE OSAVI Cl-G NDVI RGB_CHMmax b18
pi spec RBA; RGB 3d Cl-G b16 b20 b17 Cl-RE MTVI b18
all Cl-RE RGB-GRVI RGB-R RGB-ExG OSAVI MTVI RGB-B
all RBA RGB-R b21 RGB-ExG b15 b19 b17 OSAVI
Flying heigh 450–700 m AC
pi spec RBA MTVI OSAVI NDVI RDVI Cl-RE b20 GNDVI
pi all RBA MTVI Cl-RE OSAVI NDVI b20 b19 b18
RGB 3d CHMp90 CHMp50 CHMp80 CHMmin CHMp70 CHMmax CHMmean
RGB spec GRVI B R ExG G NaN NaN
RGB all GRVI B R CHMp50 ExG CHMp90 G
pi spec RBA; RGB 3d MTVI b20 OSAVI Cl-RE b19 NDVI RDVI
all; RBA MTVI Cl-RE OSAVI GRVI R RDVI NDVI
Flying heigh 900 m AC
RGB 3d CHMp80 CHMp90 CHMp70 CHMmax CHMp50 CHMmin CHMmean
RGB spec B GRVI G ExG R NaN NaN
RGB all ExG GRVI B G CHMp90 CHMp80 R
Remo e Sens. 2018,10, 1082 25 o 32
Table A5. The mos impo an ea u es o he Random Fo es (RF) (in he o de o impo ance) o DMY es ima ion in ba ley.
Flying heigh 140 m UAV
pi spec Cl-G Cl-RE RDVI MTVI NDVI OSAVI GNDVI
pi spec RBA Cl-RE MTVI RDVI GNDVI b32 NDVI b30
pi all Cl-G Cl-RE OSAVI MTVI RDVI NDVI b36
pi all RBA Cl-RE MTVI NDVI GNDVI b32 RDVI Cl-G
RGB 3d RGB_CHMp90 RGB_CHMp80 RGB_CHMp70 RGB_CHMmax RGB_CHMp50
RGB_CHMmean
RGB_CHMmin
RGB spec RGB-ExG RGB-GRVI RGB-R RGB-B RGB-G NaN NaN
RGB all RGB_CHMp90 RGB-ExG RGB-GRVI RGB_CHMp80 RGB-R RGB_CHMp70 RGB_CHMmax
pi spec; RGB 3d MTVI Cl-G Cl-RE RGB_CHMp80 RGB_CHMp90 RDVI RGB_CHMmax
pi spec RBA; RGB 3d MTVI Cl-RE RDVI GNDVI b32 Cl-G OSAVI
all Cl-RE RGB-ExG RGB-GRVI RGB-R OSAVI RGB_CHMp90 Cl-G
all RBA Cl-RE OSAVI RDVI Cl-G RGB_CHMp80 RGB-GRVI GNDVI
Flying heigh 450–700 m AC
pi spec RBA MTVI OSAVI NDVI GNDVI b17 Cl-RE b36
pi all RBA OSAVI MTVI NDVI Cl-RE GNDVI RDVI b20
RGB 3d CHMp50 CHMp90 CHMp70 CHMp80 CHMmin CHMmax CHMs d
RGB spec ExG B GRVI R G NaN NaN
RGB all ExG GRVI B R G CHMp80 CHMmin
pi spec RBA; RGB 3d MTVI b20 NDVI OSAVI Cl-RE RDVI b19
all; RBA OSAVI GRVI ExG MTVI Cl-RE R RDVI
Flying heigh 900 m AC
RGB 3d CHMp80 CHMp90 CHMp70 CHMmax CHMp50 CHMmin CHMmean
RGB spec GRVI ExG B R G NaN NaN
RGB all GRVI CHMp80 ExG R B CHMp90 G
Remo e Sens. 2018,10, 1082 32 o 32
61.
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 ]
62.
Méndez-Ba oso, L.A.; Zá a e-Valdez, J.L.; Robles-Mo úa, A. Es ima ion o hyd omo phological a ibu es
o a small o es ed ca chmen by applying he S uc u e om Mo ion (S M) app oach. In . J. Appl. Ea h
Obs. Geoin . 2018,69, 186–197. [C ossRe ]
63.
Honka aa a, E.; Kho amshahi, E. Radiome ic Co ec ion o Close-Range Spec al Image Blocks Cap u ed Using
an Unmanned Ae ial Vehicle wi h a Radiome ic Block Adjus men . Remo e Sens. 2018,10, 256. [C ossRe ]
64.
Smi h, G.M.; Mil on, E.J. The use o he empi ical line me hod o calib a e emo ely sensed da a o e lec ance.
In . J. Remo e Sens. 1999,20, 2653–2662. [C ossRe ]
65.
Tucke , C.J. Red and pho og aphic in a ed linea combina ions o moni o ing ege a ion.
Remo e Sens. En i on. 1979,8, 127–150. [C ossRe ]
66.
Woebbecke, D.M.; Meye , G.E.; Von Ba gen, K.; Mo ensen, D.A. Colo Indices o Weed Iden i ica ion Unde
Va ious Soil, Residue, and Ligh ing Condi ions. T ans. ASAE 1995,38, 259–269. [C ossRe ]
67.
Roujean, J.-L.; B eon, F.-M. Es ima ing PAR abso bed by ege a ion om bidi ec ional e lec ance
measu emen s. Remo e Sens. En i on. 1995,51, 375–384. [C ossRe ]
68.
Rouse, J.W. Moni o ing Vege a ion Sys ems in he G ea Plains wi h ERTS; NASA: Washing on, DC, USA, 1974.
69.
Rondeaux, G.; S e en, M.; Ba e , F. Op imiza ion o soil-adjus ed ege a ion indices. Remo e Sens. En i on.
1996,55, 95–107. [C ossRe ]
70.
Guyo , G.; Ba e , F. U ilisa ion de la hau e esolu ion spec ale pou sui e l’e a des cou e s ege aux.
Spec . Signa . Objec s Remo e Sens. 1988,287, 279.
71.
Gi elson, A.A.; Kau man, Y.J.; Me zlyak, M.N. Use o a g een channel in emo e sensing o global ege a ion
om EOS-MODIS. Remo e Sens. En i on. 1996,58, 289–298. [C ossRe ]
72.
Daugh y, C.S.T.; Wal hall, C.L.; Kim, M.S.; de Cols oun, E.B.; McMu ey, J.E. Es ima ing Co n Lea
Chlo ophyll Concen a ion om Lea and Canopy Re lec ance. Remo e Sens. En i on.
2000
,74, 229–239.
[C ossRe ]
73.
Haboudane, D.; Mille , J.R.; Pa ey, E.; Za co-Tejada, P.J.; S achan, I.B. Hype spec al ege a ion indices
and no el algo i hms o p edic ing g een LAI o c op canopies: Modeling and alida ion in he con ex o
p ecision ag icul u e. Remo e Sens. En i on. 2004,90, 337–352. [C ossRe ]
74.
Dash, J.; Cu an, P.J. The MERIS e es ial chlo ophyll index. In . J. Remo e Sens.
2004
,25, 5403–5413.
[C ossRe ]
75.
Gi elson, A.A.; G i z, Y.; Me zlyak, M.N. Rela ionships be ween lea chlo ophyll con en and spec al
e lec ance and algo i hms o non-des uc i e chlo ophyll assessmen in highe plan lea es. J. Plan Physiol.
2003,160, 271–282. [C ossRe ] [PubMed]
76.
He nández-Clemen e, R.; Na a o-Ce illo, R.M.; Suá ez, L.; Mo ales, F.; Za co-Tejada, P.J. Assessing s uc u al
e ec s on PRI o s ess de ec ion in coni e o es s. Remo e Sens. En i on. 2011,115, 2360–2375. [C ossRe ]
77.
Miyoshi, G.T.; Imai, N.N.; Tommaselli, A.M.G.; Honka aa a, E.; Näsi, R.; Mo iya, É.A.S. Radiome ic block
adjus men o hype spec al image blocks in he B azilian en i onmen . In . J. Remo e Sens.
2018
, 1–21. [C ossRe ]
78.
Danne , M.; Be ge , K.; Woche , M.; Mause , W.; Hank, T. Re ie al o Biophysical C op Va iables om
Mul i-Angula Canopy Spec oscopy. Remo e Sens. 2017,9. [C ossRe ]
79.
Roosjen, P.P.J.; B ede, B.; Suomalainen, J.M.; Ba holomeus, H.M.; Koois a, L.; Cle e s, J.G.P.W.
Imp o ed es ima ion o lea a ea index and lea chlo ophyll con en o a po a o c op using mul i-angle spec al
da a—Po en ial o unmanned ae ial ehicle image y. In . J. Appl. Ea h Obs. Geoin . 2018,66, 14–26. [C ossRe ]
80.
Näsi, R.; Viljanen, N.; Oli ei a, R.; Kai osoja, J.; Niemeläinen, O.; Hakala, T.; Ma kelin, L.; Nezami, S.;
Suomalainen, J.; Honka aa a, E. Op imizing adiome ic p ocessing and ea u e ex ac ion o d one based
hype spec al ame o ma image y o es ima ion o yield quan i y and quali y o a g ass swa d. ISPRS In .
A ch. Pho og amm. Remo e Sens. Spa . In o m. Sci. 2018,XLII–3, 1305–1310.
81.
Zhang, L.; Zhang, L.; Du, B. Deep Lea ning o Remo e Sensing Da a: A Technical Tu o ial on he S a e o he
A . IEEE Geosci. Remo e Sens. Mag. 2016,4, 22–40. [C ossRe ]
©
2018 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/).