Remo e Sens. 2013, 5, 5006-5039; doi:10.3390/ s5105006
Remo e Sensing
ISSN 2072-4292
www.mdpi.com/jou nal/ emo esensing
A icle
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
Eija Honka aa a 1,*, Heikki Saa i 2, Je e Kai osoja 3, Ilkka Pölönen 4, Teemu Hakala 1,
Paula Li key 1, Jussi Mäkynen 2 and Liisa Pesonen 3
1 Finnish Geode ic Ins i u e, Geodee in inne 2, P.O. Box 15, FI-02431 Masala, Finland;
E-Mails: [email p o ec ed] (T.H.), [email p o ec ed] (P.L.)
2 VTT Pho onic De ices and Measu emen Solu ions, P.O. Box 1000, FI-02044 VTT, Finland;
E-Mails: [email p o ec ed] (H.S.), [email p o ec ed] (J.M.)
3 MTT Ag i ood Resea ch Finland, Vakolan ie 55, FI-03400 Vih i, Finland;
E-Mails: je e.kai osoja@m . i (J.K.), liisa.pe[email p o ec ed] (L.P.)
4 Depa men o Ma hema ical In o ma ion Technology, Uni e si y o Jy äskylä, P.O. Box 35,
FI-40014 Jy äskylä, Finland; E-Mail: [email p o ec ed]
* Au ho o whom co espondence should be add essed: E-Mail: [email p o ec ed];
Tel: +358-40-1920-835; Fax: +358-9-2955-5211.
Recei ed: 16 Augus 2013; in e ised o m: 30 Sep embe 2013 / Accep ed: 6 Oc obe 2013 /
Published: 14 Oc obe 2013
Abs ac : Imaging using ligh weigh , unmanned ai bo ne ehicles (UAVs) is one o he
mos apidly de eloping ields in emo e sensing echnology. The new, unable,
Fab y-Pe o in e e ome e -based (FPI) spec al came a, which weighs less han 700 g,
makes i possible o collec spec ome ic image blocks wi h s e eoscopic o e laps using
ligh -weigh UAV pla o ms. This new echnology is highly ele an , because i opens up
new possibili ies o measu ing and moni o ing he en i onmen , which is becoming
inc easingly impo an o many en i onmen al challenges. Ou objec i es we e o
in es iga e he p ocessing and use o his new ype o image da a in p ecision ag icul u e.
We de eloped he en i e p ocessing chain om aw images up o geo e e enced e lec ance
images, digi al su ace models and biomass es ima es. The p ocessing in eg a es
pho og amme ic and quan i a i e emo e sensing app oaches. We ca ied ou an empi ical
assessmen using FPI spec al image y collec ed a an ag icul u al whea es si e in he
summe o 2012. Poo wea he condi ions du ing he campaign complica ed he da a
p ocessing, bu his is one o he challenges ha a e aced in ope a ional applica ions. The
OPEN ACCESS
Remo e Sens. 2013, 5 5007
esul s indica ed ha he came a pe o med consis en ly and ha he da a p ocessing was
consis en , as well. Du ing he ag icul u al expe imen s, p omising esul s we e ob ained
o biomass es ima ion when he spec al da a was used and when an app op ia e
adiome ic co ec ion was applied o he da a. Ou esul s showed ha he new FPI
echnology has a g ea po en ial in p ecision ag icul u e and indica ed many possible u u e
esea ch opics.
Keywo ds: pho og amme y; adiome y; spec ome y; hype spec al; UAV; DSM; poin
cloud; biomass; ag icul u e
1. In oduc ion
Mode n ai bo ne imaging echnology based on unmanned ai bo ne ehicles (UAVs) o e s
unp eceden ed possibili ies o measu ing ou en i onmen . Fo many applica ions, UAV-based
ai bo ne me hods o e he possibili y o cos -e icien da a collec ion wi h he desi ed spa ial and
empo al esolu ions. An impo an ad an age o UAV-based echnology is ha he emo e sensing
da a can be collec ed e en unde poo imaging condi ions, ha is, unde cloud co e , which makes i
uly ope a ional in a wide ange o en i onmen al measu ing applica ions.
We ocus he e on ligh weigh sys ems, which is one o he mos apidly g owing ields in UAV
echnology. The sys ems a e qui e compe i i e in local a ea applica ions and especially i epe i i e
da a collec ion o a apid esponse is needed.
An app op ia e senso is a undamen al componen o a UAV imaging sys em. The i s ope a ional,
ci il, ligh weigh UAV imaging sys ems ypically used comme cial ideo came as o cus ome s ill
came as ope a ing in selec ed h ee wide-bandwid h bands in ed, g een, blue and/o nea -in a ed
spec al egions [1–3]. The ecen senso de elopmen s ailo ed o ope a ion om UAVs o e
enhanced possibili ies o emo e sensing applica ions in e ms o be e image quali y, mul i-spec al,
hype -spec al and he mal imaging [4–10] and lase scanning [11–13].
One in e es ing new senso is a ligh weigh spec al came a de eloped by he VTT Technical
Resea ch Cen e o Finland (VTT). The came a is based on a piezo-ac ua ed, Fab y-Pe o
in e e ome e (FPI) wi h an adjus able ai gap [6,14]. This echnology makes i possible o
manu ac u e a ligh weigh spec al image ha can p o ide lexibly selec able spec al bands in a
wa eleng h ange o 400–1,000 nm. Fu he mo e, because he senso p oduces images in a ame
o ma , 3D in o ma ion can be ex ac ed i he images a e collec ed wi h s e eoscopic o e laps. In
compa ison o pushb oom imaging [7,9], he ad an ages o ame imaging include he possibili y o
collec image blocks wi h s e eoscopic o e laps and he geome ic and adiome ic cons ain s
p o ided by he igid ec angula image geome y and mul iple o e lapping images. We hink ha his
is impo an in pa icula o UAV applica ions, which ypically u ilize images collec ed unde
dynamic, ib a ing and u bulen condi ions.
Con en ional pho og amme ic and emo e sensing p ocessing me hods a e no di ec ly applicable
o ypical, small- o ma UAV image y, because hey ha e been de eloped o mo e s able da a and
images wi h a much la ge spa ial ex en han wha can be ob ained wi h ypical UAV imaging
Remo e Sens. 2013, 5 5008
sys ems. Wi h UAV-based, small- o ma ame imaging, a la ge numbe —hund eds o e en
housands—o o e lapping images a e needed o co e he desi ed objec a ea. Sys ems a e o en
ope a ed unde subop imal condi ions, such as below ull o pa ial cloud co e . Despi e he
challenging condi ions, he images mus be p ocessed accu a ely so ha objec cha ac e is ics can be
in e p e ed on a quan i a i e geome ic and adiome ic basis using he da a.
P ecision ag icul u e is one o he po en ial applica ions o hype spec al UAV
imaging [1,2,4,6,9,15–18]. In p ecision ag icul u e, he majo objec i es a e o enable e icien use o
esou ces, p o ec ion o he en i onmen and documen a ion o applied managemen ea men s by
applying machine guidance and si e-speci ic seeding, e iliza ion and plan p o ec ion. The
expec a ion is ha UAVs migh p o ide an e icien emo e sensing ool o hese asks [18]. A e iew
by Zhang and Ko acs [16] showed ha esea ch is needed on many opics in o de o de elop e icien
UAV-based me hods o p ecision ag icul u e. In his s udy, we will demons a e he use o he FPI
spec al came a in a biomass es ima ion p ocess o whea c ops; biomass is one o he cen al
biophysical pa ame e s o be es ima ed in p ecision ag icul u e [17].
The objec i es o his in es iga ion we e o in es iga e a comple e p ocessing me hodology o he
FPI spec al image y, as well as o demons a e i s po en ial in a biomass es ima ion p ocess o
p ecision ag icul u e. We depic a me hod o FPI image da a p ocessing in Sec ion 2. We desc ibe he
es se up used o he empi ical in es iga ion in Sec ion 3. We p esen he empi ical esul s in Sec ion
4 and discuss hem in mo e de ail in Sec ion 5.
2. A Me hod o P ocessing FPI Spec al Da a Cubes
2.1. An FPI-Based Spec al Came a
The FPI-based spec al came a de eloped by he VTT p o ides a new way o collec spec ome ic
image blocks. The image is based on he use o mul iple o de s o he Fab y-Pe o in e e ome e
oge he wi h he di e en spec al sensi i i ies o he ed, g een and blue pixels o he image senso .
Wi h his a angemen , i is possible o cap u e h ee wa eleng h bands wi h a single exposu e. When
he FPI is placed in on o he senso , he spec al sensi i i y o each pixel is a unc ion o he
in e e ome e ai gap. By changing he ai gap, i is possible o acqui e a new se o wa eleng hs. Wi h
smalle ai gaps, i is also possible o cap u e only one o wo wa eleng hs in each image. Sepa a e
sho -pass and long-pass il e s a e needed o cu ou unwan ed ansmissions a unused o de s o he
Fab y-Pe o in e e ome e . Du ing a ligh , a p ede ined sequence o ai gap alues is applied using
he FPI came a o econs uc he spec um o each pixel in he image (see mo e de ails in Sec ion 2.3.3).
O en, 24 di e en ai gap alues a e used, and by hese means, i is possible o collec 24 eely
selec able spec al bands in a single ligh , while he es o he bands (0–48) a e no independen . The
desi ed spec al bands can be selec ed wi h a spec al s ep o 1 nm. This echnology p o ides spec al
da a cubes wi h a ec angula image o ma , bu each band in he da a cube exposed o a di e en ai gap
alue has a sligh ly di e en posi ion and o ien a ion. The p inciples o he FPI spec al came a ha e
been desc ibed by Saa i e al. [6] and Mäkynen e al. [14]. Resul s p esen ed by Nackae s e al. [19]
demons a ed he easibili y o he senso concep . Honka aa a e al. [20] de eloped he i s
Remo e Sens. 2013, 5 5009
pho og amme ic p ocessing line o he FPI spec al came a, and Pölönen e al. [21] ca ied ou he
i s pe o mance assessmen s in p ecision ag icul u e using he 2011 p o o ype.
The 2012 p o o ype was used in his in es iga ion. I is equipped wi h cus om op ics wi h a ocal
leng h o 10.9 mm and an -numbe o less han 3.0. The came a has a CMOSIS CMV4000
Complemen a y Me al Oxide Semiconduc o (CMOS) ed, g een and blue (RGB) image senso wi h
an elec onic shu e ; he in a ed cu il e has been emo ed om he image senso . The senso has a
2,048 × 2,048 pixel esolu ion, a pixel size o 5.5 μm and a adiome ic esolu ion o 12 bi s. In
p ac ical applica ions, he senso is used in he wo- imes binned mode, while only pa o he senso
a ea is used. This p o ides an image size o 1,024 × 648 pixels wi h a pixel size o 11 μm. The ield o
iew (FOV) is ±18° in he ligh di ec ion, ±27° in he c oss- ligh di ec ion and ±31° a he o ma
co ne . Applica ion-based il e s can be used, o example 500–900, 450–700, 600–1,000 o 400–500 nm
il e s. The spec al esolu ion ange is 10–40 nm a he ull wid h a hal maximum (FWHM), and i is
dependen on he FPI ai gap alue, as well as he il e selec ion.
Table 1 shows he di e ences be ween he 2011 p o o ype and he 2012 p o o ype. Many o he
pa ame e s we e imp o ed o he 2012 p o o ype. The mos signi ican imp o emen was he
imp o emen o he -numbe in o de o imp o e he signal- o-noise a io (SNR). This was achie ed
by imp o ing he lens sys em and allowing o g ea e FPI ay angles (10° in compa ison o 4°). The
blu ing o images was educed by changing he olling shu e o an elec onic shu e .
Table 1. Spec al image speci ica ions o 2011 and 2012 p o o ypes. FOV, ield o iew;
FWHM, he ull wid h a hal maximum; FPI, Fab y-Pe o in e e ome e .
Pa ame e P o o ype 2011 P o o ype 2012
Ho izon al and e ical FOV (°) >36, >26 >50, >37
Nominal ocal leng h (mm) 9.3 10.9
Wa eleng h ange (nm) 400–900 400–900
Spec al esolu ion a FWHM (nm);
depending on he selec ion o he FPI ai gap alue 9–45 10–40
Spec al s ep (nm): adjus able by con olling he ai gap o he FPI <1 <1
-numbe <7 <3
Pixel size (μm); no binning/de aul binning 2.2/8.8 5.5/11
Maximum spec al image size (pixels) 2,592 × 1,944 2,048 × 2,048
Spec al image size wi h de aul binning (pixels) 640 × 480 1,024 × 648
Came a dimensions (mm) 65 × 65 × 130 <80 × 92 × 150
Weigh (g); including ba e y, GPS ecei e ,
downwelling i adiance senso s and cabling <420 <700
The senso se up is shown in Figu e 1. The en i e imaging sys em includes he FPI spec al came a,
a 32 GB compac lash memo y ca d, i adiance senso s o measu ing downwelling and upwelling
i adiance ( he i adiance senso o measu ing upwelling i adiance is no ope a ional in he cu en
se up), a GPS ecei e and a li hium polyme (LiPo) ba e y. The sys em weighs less han 700 g. Wi h
his se up, mo e han 1,000 da a cubes, each wi h up o 48 bands, can be collec ed in he wo- imes
binned mode wi hin a single ligh .
Remo e Sens. 2013, 5 5010
Figu e 1. Componen s o he FPI imaging sys em: FPI spec al came a, a compac lash
memo y ca d, wo i adiance senso s, a GPS ecei e and a LiPo ba e y.
2.2. FPI Spec al Came a Da a P ocessing
The FPI spec al came a da a can be p ocessed in a simila manne as small- o ma ame image y,
bu some senso -speci ic p ocessing is also equi ed. Figu e 2 shows a gene al da a p ocessing chain
o FPI spec al image da a. The p ocessing chain includes da a collec ion, FPI spec al da a cube
gene a ion, image o ien a ion, digi al su ace model (DSM) calcula ions, adiome ic model
calcula ions and ou pu p oduc gene a ion. When de eloping he da a p ocessing chain o he FPI
spec al came a, ou objec i e has been o in eg a e he senso -speci ic p ocessing s eps in o ou
exis ing p ocessing line based on comme cially a ailable pho og amme ic and emo e
sensing so wa e.
Figu e 2. FPI spec al came a da a p ocessing chain. DSM, digi al su ace model.
Da a collec ion cons i u es he i s phase in he imaging chain. The cen al pa ame e s ha need o
be se a e he spec al sensi i i ies o he bands and he in eg a ion ime. The spec al sensi i i ies a e
selec ed based on he equi emen s o he applica ion. The in eg a ion ime should be selec ed so ha
he images a e no o e exposed in ela ion o he b igh objec s and ha he e a e good dynamics o
he da k objec s. Fu he mo e, he ligh speed and lying al i ude will impac he da a quali y
(Sec ion 2.3.3).
The p e-p ocessing phase equi es ha he senso image y be accu a ely co ec ed adiome ically
based on labo a o y calib a ions [7,8,14]. Co ec ing he spec al smile and misma ch o di e en
spec al bands equi es senso -speci ic p ocessing, which a e desc ibed in mo e de ail in Sec ion 2.3.

Remo e Sens. 2013, 5 5011
The objec i e o geome ic p ocessing is o ob ain non-dis o ed 3D da a o a desi ed coo dina e
sys em. The geome ic p ocessing s eps in ol e de e mining he image o ien a ions and DSM
gene a ion. An impo an ecen ad ancemen in pho og amme ic p ocessing is he new gene a ion o
image ma ching me hods o DSM measu emen , which is an impo an ad an age o he image
analysis p ocess when using UAVs [20,22–27].
The objec i e o adiome ic p ocessing is o p o ide in o ma ion abou he objec ’s
e lec i i y [28]. Wi h his p ocess, all o he dis u bances ela ed o he imaging sys em, a mosphe ic
in luences and iew/illumina ion- ela ed ac o s need o be compensa ed o . We desc ibe ou
app oaches o adiome ic p ocessing in Sec ion 2.4.
The ou pu s esul ing om he abo e-men ioned p ocess include geo e e enced e lec ance image
mosaics and 3D p oduc s, such as poin clouds, poin clouds wi h e lec i i y in o ma ion, DSMs,
objec models and s e eomodels o isual e alua ions [20].
2.3. FPI Spec al Da a Cube Gene a ion
A c ucial s ep in he da a p ocessing chain is he cons uc ion o adiome ically and geome ically
consis en spec al da a cubes. The p ocess includes h ee majo phases: adiome ic image co ec ions
based on labo a o y calib a ions, spec al smile co ec ions and band ma ching.
2.3.1. Radiome ic Co ec ion Based on Labo a o y Calib a ion
The i s s eps in he da a p ocessing chain in ol e da k signal compensa ion and applying he
senso calib a ion in o ma ion o he images wi h he usual co ec ions o pho on esponse
nonuni o mi y (PRNU) (including he CMOS a ay nonuni o mi y and lens allo co ec ions). These
pa ame e s a e de e mined in he VTT’s calib a ion labo a o y. Da k signal compensa ion is ca ied ou
using a da k image collec ed be o e he image da a collec ion. A Baye -ma ix econs uc ion is ca ied
ou o ob ain h ee-band images. The p ocess has been desc ibed by Mäkynen e al. [14], and i will no
be emphasized in his in es iga ion.
2.3.2. Co ec ion o he Spec al Smile
Ideally, he op ics o he FPI spec al came a a e designed so ha ligh ays go h ough he FPI as a
collima ed beam o a speci ic pixel o he image. In o de o imp o e he -numbe , his equi emen
was comp omised, and a maximum FPI ay angle o 10° was allowed. This will cause a shi in he
cen al wa eleng h o he FPI’s spec al peak (a spec al smile e ec ). Fo he mos pa , he peak
wa eleng h (λ0) is linea ly dependen on he cosine o he ay angle (θ) a he image plane:
λ0 = 2·d·cos(θ)/m, m = 1, 2, ..., (1)
whe e d is he ai gap be ween he Fab y-Pe o in e e ome e mi o s and m is he in e e ence o de a
which he FPI is used.
The e a e di e en app oaches o handling he spec al smile:
1. Co ec ing images: Ou assump ion is ha we can calcula e smile-co ec ed images so
ha he co ec ed spec ums can be esampled om wo spec ally (wi h a di e ence in
Remo e Sens. 2013, 5 5012
peak wa eleng h p e e ably less han 10 nm) and empo ally (wi h a spa ial
displacemen less han 20 pixels) adjacen image bands.
2. Using he cen al a eas o he images: When images a e collec ed wi h a minimum o
60% o wa d and side o e laps and when he mos nadi pa s o images a e used, he
smile e ec is less han 5 nm and can be igno ed in mos applica ions (when he
FWHM is 10–40 nm).
3. Resampling a “supe spec um” o each objec poin o he o e lapping images
p o iding a iable cen al wa eleng hs. The en i e spec um can be u ilized in
he applica ions.
Fo ypical emo e sensing applica ions and so wa e, app oaches 1 and 2, a e he mos unc ional,
because hey can be ca ied ou in a sepa a e s ep du ing he p ep ocessing phase. In o de o u ilize he
en i e ex en o he images, app oaches 1 o 3 a e equi ed. The VTT has de eloped a me hod o
co ec ing he spec al smile by using wo spa ially and spec ally adjacen bands (app oach 1). A
co ela ion-based ma ching using shi s in he ow and column di ec ions is i s ca ied ou on he
images o accu a ely align wo o he bands. The co ec ed image is hen o med by in e pola ing he
desi ed wa eleng h om he adjacen bands. This simple app oach is conside ed su icien , because he
bands ha need o be ma ched a e spec ally and spa ially close o one ano he .
2.3.3. Band Ma ching
The bands o a single da a cube image do no o e lap pe ec ly, due o he imaging p inciple o he
FPI spec al came a (Sec ion 2.1). This makes i challenging o de e mine he o ien a ion o he bands,
because, in p inciple, all o he bands and images ha e a di e en o ien a ion.
Exposing and ans e ing a single band o a synch onous dynamic andom-access memo y
(SDRAM) akes app oxima ely 75 ms, which is he ime di e ence be ween wo empo ally adjacen
bands. Reco ding a single cube, o example, wi h 24 ai gaps, akes a o al o 1,800 ms. The esul ing
spa ial shi is dependen on he ligh speed o he UAV, and he shi in pixel coo dina es is
de e mined by he lying heigh o he sys em. Fo example, wi h a ligh speed o 5 m·s−1 and a
con inuous da a collec ion mode (no s opping du ing image exposu e), he ho izon al ansi ion o he
UAV in he ligh di ec ion be ween empo ally adjacen bands is 0.375 m, whe eas i is 9 m o he
en i e da a cube; wi h a lying al i ude o 150 m (g ound sample dis ance (GSD), 15 cm), he spa ial
displacemen s a e hus 2.5 and 60 pixels, espec i ely. The UAV is also swinging and ib a ing du ing
he ime o he da a cube exposu e, which gene a es nonde e minis ic posi ional and o a ional
misma ches o he bands. Due o hese small andom mo emen s, he a p io i in o ma ion has o be
imp o ed by using da a-based analysis.
Two di e en app oaches o p ocessing he FPI image da a a e as ollows:
1. De e mining he o ien a ions o and geo e e encing he indi idual bands sepa a ely.
Wi h his app oach, he e a e numbe -o -bands ( ypically 20–48) image blocks ha
need o be p ocessed. We used his app oach in ou in es iga ion wi h he 2011 came a
p o o ype, whe e we p ocessed i e bands [20,29].
Remo e Sens. 2013, 5 5013
2. Sampling he bands o indi idual da a cubes in ela ion o he geome y o a e e ence
band. O ien a ions a e de e mined o he image block wi h e e ence band images, and
his o ien a ion in o ma ion is hen applied o all o he bands. We s udied his app oach
o his in es iga ion.
The p inciple aspec o app oach 2 in ol es using one o he bands as he e e ence band and
ma ching all o he o he bands o i . In o de o ans o m a band o ma ch he geome y o he
e e ence band, a geome ic image ans o ma ion mus be ca ied ou . The pa ame e s o he
ans o ma ion a e de e mined by u ilizing ie poin s ha a e au oma ically measu ed be ween he
bands ia image ma ching. The e a e se e al challenges in his ma ching p ocess. A scene p o ides
di e en digi al numbe s (DN) in di e en spec al bands acco ding o he spec al esponses o he
di e en objec s, which complica es he image ma ching p ocess. Ano he challenge is ha , wi h
dynamic UAV imaging, he o e laps be ween he di e en bands can be qui e small. Wi h 3D objec s,
a u he complica ion is ha a simple 2D modeling o he objec ’s geome y is no likely o p o ide
accu a e esul s; wi h homogeneous objec s, he images migh no con ain su icien ea u es o
ma ching, and epe i i e ea u es, such as hose seen in c ops, can cause alse ma ches.
Ou cu en implemen a ion combines he abo e wo app oaches. We use a ew e e ence bands and
hen ma ch se e al adjacen bands o hese bands. By using se e al e e ence bands, we y o imp o e
he accu acy and obus ness o he image ans o ma ions. Fo example, we can selec e e ence bands
o each majo wa eleng h egion (blue, g een, ed, nea -in a ed (NIR)) o we could selec e e ence
bands o maximize he spa ial adjacency o he bands o be ma ched oge he . We also a oid
geo e e encing all laye s sepa a ely, which could slow down he p ocessing in gene al and be oo
labo ious in he case o challenging objec s ha equi e manual in e ac ions. Fo he i s
implemen a ion, we employed ea u e-based ma ching (FBM) o he band pai s using poin ea u es
ex ac ed ia he Fö s ne ope a o [30]. Finally, an image ans o ma ion is ca ied ou using an
app op ia e geome ic model; in ou sys em, an a ine o a p ojec i e model is used.
2.4. Radiome ic Co ec ion o F ame Image Block Da a
Ou expe ience is ha UAV emo e sensing da a o en need o be collec ed unde sub-op imal
illumina ion condi ions, such as a ying deg ees o cloudiness. The possibili y o collec da a below
cloud co e and unde di icul condi ions is also one o he majo ad an ages o he UAV echnology.
We ha e de eloped wo app oaches o he adiome ic p ocessing o UAV image blocks in o de o
p oduce homogeneous da a o non-homogeneous inpu da a.
The i s app oach is an image-in o ma ion-based adiome ic block adjus men me hod [20,29].
The basic p inciple o he app oach is o use he g ay alues (digi al numbe s (DNs)) o he adiome ic
ie poin s in he o e lapping images as obse a ions and o de e mine he pa ame e s o he adiome ic
model indi ec ly ia he leas squa es p inciple. Cu en ly, we use he ollowing model o a g ay
alue (DN):
DNjk = a el_j·(aabs·Rjk(θi,θ ,φ) + babs) +b el_j (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,
Remo e Sens. 2013, 5 5014
espec i ely, and φ = φ − φi is he ela i e azimu h angle; aabs and babs a e he pa ame e s o he
empi ical line model o ans o ming he e lec ance o DN and a el_j and b el_j a e ela i e co ec ion
pa ame e s wi h espec o he e e ence image.
The impac o iew/illumina ion geome y on he objec e lec ance is an impo an challenge in
passi e imaging. Ou app oach is o use a model-based, mul iplica i e, aniso opy ac o . A
e lec ance, Rjk(θi,θ ,φ), o he objec poin , k, can be gi en as ollows:
Rjk(θi,θ ,φ) = Rk(0,0,φ)·ρmodeled(θi,θ ,φ)/ρmodeled(0,0,φ) (3)
whe e R
k(0,0,φ) is he e lec ance ac o a he nadi iewing geome y, ρmodeled(θi,θ ,φ) is he
bi-di ec ional e lec ance dis ibu ion unc ion (BRDF) o he speci ic iew/illumina ion geome y and
ρmodeled(0,0,φ) is he modeled e lec ance in a desi ed nadi iewing geome y (see, also, [31]). We use
he simple BRDF model de eloped by Wal hall e al. [32]. The inal BRF wi h coe icien s a’ and b’
o he desi ed e lec ed geome y o θ = 0 is:
Rjk(θi,θ ,φ) = Rk(0,0,φ)(a’θ 2 + b’θ cos φ + 1) (4)
In he adiome ic model, he absolu e e lec ance ans o ma ion pa ame e s elimina e a mosphe ic
in luences. The ela i e addi i e and mul iplica i e pa ame e s elimina e di e ences be ween he
images ha a e mainly due o illumina ion changes and senso ins abili y. The BRDF model akes ca e
o he iew/illumina ion geome y- ela ed issues. The numbe s o he pa ame e s a e as ollows:
absolu e calib a ion: 2; ela i e calib a ion: (numbe o images − 1) × 2; BRDF model: 2 (i he same
model is used o he en i e objec ); nadi e lec ance o adiome ic ie poin s: numbe o adiome ic
ie poin s. The pa ame e s ha a e used depend on he condi ions du ing imaging. This model includes
many simpli ica ions, bu i can be ex ended using physical pa ame e s. Fo his app oach, a minimum
o wo e lec ance e e ence a ge s a e needed. Wi h he cu en implemen a ion, he pa ame e s a e
de e mined sepa a ely o each band. Du ing he image co ec ion phase, Rk(0,0,φ) is calcula ed based
on Equa ions (2) and (4).
A second al e na i e is o u ilize he i adiance measu emen s collec ed du ing he ligh campaign,
as desc ibed by Hakala e al. [33]. In his case, he ela i e adjus men o he images is ob ained by
selec ing one e e ence image and calcula ing he ela i e mul iplica i e co ec ion ac o s, Cj(λ), wi h
espec o i :
Cj(λ)=E e (λ)/Ej(λ) (5)
whe e Ej(λ) and E e (λ) a e he spec al i adiance measu emen s du ing he acquisi ion o image j and
e e ence image e . When mul iplying he DNs o image j using Cj(λ), he adiome ic di e ences
caused by changes in he i adiance a e elimina ed. Ou analysis, p esen ed in ano he a icle [33],
showed ha his model is unc ional unde ce ain condi ions. I he assump ions a e no alid, he
accu acy o he co ec ion will be educed.
3. Empi ical In es iga ion
3.1. Tes A ea
An empi ical campaign was ca ied ou a he MTT Ag i ood Resea ch Finland (MTT) ag icul u al
es si e in Vih i (60°25'21''N, 24°22'28''E) (Figu e 3). The es a ea, a 76 m by 385 m (2.9 ha) pa ch o
Remo e Sens. 2013, 5 5021
Figu e 7. Examples o image quali y. Bands om he le wi h cen al peak wa eleng hs
and FWHMs in pa en hesis: indi idual bands 7 (535.5 nm, 24.9 nm), 16 (606.2 nm, 44.0
nm) and 29 (787.5 nm, 32.1 nm) and a h ee-band, un-ma ched band composi e (7, 16, 29).
We es ima ed he signal- o-noise a io (SNR) as a a io o he a e age signal o he s anda d de ia ion
in a small image window when using a a paulin wi h a nominal e lec ance o 0.3 (Figu e 8). This is no
an accu a e es ima e o he SNR, because i was also in luenced by he nonuni o mi y o he a paulin;
s ill, i can be used as an indica i e alue, because he ela ionships be ween bands a e ealis ic. The SNR
was ypically a ound 80, bu a educ ion appea ed in some o he g een bands (wa eleng h 550–600 nm),
as well as in he NIR bands (wa eleng h >800 nm). The beha io was as expec ed. In he NIR band, he
quan um e iciency o he CMOS image senso dec eases as he wa eleng hs inc eases. The dec easing
SNR in some g een bands was a esul o he low ansmission a hose bands, due o he edge il e ,
which limi s he spec al bandwid h and dec eases he ansmission.
Figu e 8. Signal- o-noise a io (SNR) calcula ed using a a paulin wi h a nominal
e lec ance o 0.3 (x-axis: wa eleng h (nm); y-axis: SNR).
4.2. Band Ma ching
The assessmen o he band ma ching esul s indica ed ha he ma ching was success ul. The
numbe s o ie poin s we e 10–806, mos ly >100, and he median was 276. The s anda d de ia ions o
he shi pa ame e s we e 1.5, 0.9 and 0.8 pixels o he g een, ed and NIR e e ence bands, espec i ely.
The a e age shi s o he bands ma ched o he g een, ed and NIR e e ence bands a e shown in
Figu e 9. They we e be ween i e and −15 pixels in he ligh di ec ion and less han h ee pixels in he
di ec ion pe pendicula o he ligh di ec ion. The maximum shi s we e up o 25 pixels in he ligh
di ec ion and 43 pixels in he c oss- ligh di ec ion. The esul s p esen ed in Figu e 9 show ha in he
ligh di ec ion, he measu ed shi alues we e qui e close o he expec ed alues ha we e calcula ed
based on he ime di e ence be ween he e e ence band and he ma ched band. In he di ec ion
0
20
40
60
80
100
120
500 550 600 650 700 750 800 850 900
SNR
Wa eleng h (nm)

Remo e Sens. 2013, 5 5022
pe pendicula o he ligh , he shi s a e shown as a unc ion o he ime di e ence be ween he bands.
The a e age alues showed mino sys ema ic d i , which was la ge as he ime di e ence be ween
he bands became la ge ; his is likely due o he possible mino d i o he came a’s x-axis wi h
espec o he lying di ec ion.
These esul s show ha he band ma ching esul s a e consis en wi h ou expec a ions. Ou
conclusion is ha he me hod de eloped he e o band ma ching was app op ia e o he expe imen al
es ing done in his in es iga ion. I can be imp o ed upon in many ways i equi ed o u u e
applica ions. E icien , au oma ed quali y con ol p ocedu es can be de eloped by e alua ing he
ma ching and ans o ma ion s a is ics and he numbe o success ul ie poin s and by compa ing he
es ima ed ans o ma ion pa ame e s o he expec ed alues. The ma ching me hods can be u he
imp o ed, as well.
Figu e 9. A e age shi alues o he a ine ans o ma ion o he whole block in he ligh
(le ) and c oss- ligh di ec ion ( igh ). A e age shi s we e calcula ed o each e e ence
s. band combina ion.
4.3. Geome ic P ocessing
4.3.1. O ien a ions
We de e mined he o ien a ions o he e e ence bands in he g een, ed and nea -in a ed spec al
egions (bands 7, 16 and 29). Block s a is ics a e p o ided in Table 4. The s anda d de ia ions o he
uni weigh we e 0.5–0.7 pixels, and hey we e bes o he nea -in a ed band (band 29); he p ecision
es ima es o he o ien a ion pa ame e s also indica ed be e esul s o band 29. The be e esul s o
band 29 migh be due o he highe in ensi y alues in he NIR band ( ha p o ided a be e SNR),
which could p o ide be e quali y o he au oma ic ie poin measu emen s ( he ields a e e y da k
on he g een and ed bands). On he o he hand, he oo -mean-squa e e o s (RMSEs) o he GCPs
indica ed he poo es pe o mance o band 29, which was p obably due o he di icul ies in measu ing
he GCPs o his band ( he likely eason o his is ha he p ope ies o he pain used o he GCPs
we e no ideal in he NIR spec al egion). The RMSEs o he a ge ed GCPs we e on he le el o one
pixel, and hey we e expec ed o be a ep esen a i e es ima o o he geo e e encing accu acy.
-15
-10
-5
0
5
10
15
-15 -10 -5 0 5 10 15
Expec ed shi (pixels)
G
R
NIR
Measu ed
shi (pixels)
-15
-10
-5
0
5
10
15
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
Time di e ence (s)
G
R
NIR
Measu ed
shi (pixels)
Remo e Sens. 2013, 5 5023
Table 4. S a is ics on he block adjus men s o e e ence bands 7, 16 and 29: s anda d
e o o uni weigh (σ0), oo -mean-squa e e o (RMSE) alues o he s anda d de ia ions
o he pe spec i e cen e posi ions (X0, Y0, Z0) and image o a ions (ω, φ, κ) and he
RMSEs o he esiduals o he GCPs; he numbe o GCPs (N; a ge ed, na u al). Cen al
peak wa eleng hs and FWHMs o he e e ence bands: 7 (535.5 nm, 24.9 nm), 16
(606.2 nm, 44.0 nm) and 29 (787.5 nm, 32.1 nm).
Band σ0 RMSE Posi ions (m) RMSE Ro a ions (°) RMSE GCPs (m) N
(pixels) X0 Y0 Z0 ω φ κ X Y Z
7 0.68 0.28 0.27 0.06 0.105 0.107 0.019 0.08 0.10 0.05 9,10
16 0.73 0.26 0.25 0.06 0.094 0.100 0.018 0.06 0.13 0.05 11,13
29 0.49 0.24 0.24 0.06 0.091 0.093 0.016 0.14 0.13 0.03 10,13
Fo he sel -calib a ing bundle block adjus men , we es ima ed he p incipal poin o
au ocollima ion (x0, y0) and he adial dis o ion pa ame e , k1 ( he adial dis o ion co ec ion (d ) o
adial dis ance om he image cen e is d = k1 3) (Table 5). The k1 pa ame e alues we e simila in
he di e en bands, which was consis en wi h ou expec a ions. This p ope y is a o able o he band
ma ching p ocess. The es ima ed alues o he p incipal poin o au ocollima ion a ied in he
di e en bands; his was likely due o he non-op imali y o he block o his ask. The coo dina es o
he p incipal poin co ela ed wi h he coo dina es o he pe spec i e cen e . Mo e de ailed
in es iga ions o he came a calib a ion should be ca ied ou in a labo a o y using sui able
imaging geome y.
Table 5. Es ima ed coo dina es o he p incipal poin o au ocollima ion (x0, y0) and
adial dis o ion (k1) and hei s anda d de ia ions o e e ence bands 7, 16 and 29. Cen al
peak wa eleng hs and FWHMs o he e e ence bands: 7 (535.5 nm, 24.9 nm), 16
(606.2 nm, 44.0 nm), and 29 (787.5 nm, 32.1 nm).
Band Pa ame e s S anda d De ia ion
x0 (mm) y0 (mm) k1 (mm·mm−3) x0 (mm) y0 (mm) k1 (mm·mm−3)
7 0.270 −0.023 0.00251 0.003 0.006 1.26E-05
16 0.242 −0.091 0.00252 0.002 0.005 1.07E-05
29 0.208 −0.132 0.00251 0.002 0.005 1.06E-05
The quali y o he block adjus men was in acco dance wi h ou expec a ions. These esul s also
indica ed ha he o ien a ions o indi idual bands can be de e mined using s a e-o - he-a
pho og amme ic me hods a comme cial pho og amme ic wo ks a ions.
4.3.2. DSM
Figu e 10a (le ) shows a DSM ob ained ia au oma ic image ma ching using band 29 o he FPI
image block. The ma ching quali y was poo o ac o acks, which appea ed ei he as ma ching
ailu es (missing poin s) o as ou lie s (high poin s), bu o he wise, we ob ained a ela i ely good poin
cloud. We also es ed how o gene a e DSMs using he g een and ed e e ence bands, bu he ma ching
quali y was poo , likely due o he lowe SNR wi h he da k objec ( he ields a e da k in he g een and
Remo e Sens. 2013, 5 5024
ed bands). The e e ence DSM ex ac ed using he highe spa ial esolu ion image block collec ed on
he same day is shown in Figu e 10b (le ). This DSM has a highe quali y, bu some mino ma ching
ailu es in he ac o acks appea ed wi h his DSM, as well. The heigh RMSE o he FPI spec al
came a DSM was 35 cm a he GCPs; likewise, he compa ison o he FPI DSM o he e e ence DSM
a he ege a ion sample loca ions p o ided an RMSE o 35 cm.
Figu e 10. DSM (le ), ege a ion heigh s (middle) and a sca e plo o he ege a ion
heigh s ( igh ) as a unc ion o he d y biomass p o ided using (a) FPI spec al came a
images wi h a poin in e al o 20 cm and (b) high spa ial esolu ion UAV images wi h a
poin in e al o 10 cm. The da k blue a eas indica e ailu es in ma ching. Ai bo ne lase
scanning da a used as he ba e g ound in o ma ion is om he Na ional Land Su ey [37].
(a)
(b)
We e alua ed he po en ial o using ege a ion heigh s aken om he DSM du ing he biomass
es ima ion. We ob ained he es ima e o he c op heigh by calcula ing he di e ence be ween he FPI
spec al came a DSM and he DSM based on ai bo ne lase scanning (ALS DSM), which was
collec ed in sp ing ime on ba e g ound (Figu e 10a (middle)). The es ima ed heigh o he ege a ion
was be ween ze o and 1.4 m; he a e age heigh was 0.74 m, and he s anda d de ia ion was 0.36 m.
The linea eg ession be ween he d y biomass alues and he ege a ion heigh did no show any
FPI: y = 8E-05x + 0.6024
R² = 0.019
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
,0 1,000 2,000 3,000
D y biomass (kg/ha)
pi
Vege a ion heigh (m)
CIR: y = 0.0001x + 0.5575
R² = 0.3601
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
,0 1,000 2,000 3,000
D y biomass (kg/ha)
ci
Vege a ion heigh (m)
Remo e Sens. 2013, 5 5025
co ela ion ( he R2 alue o eg ession was ze o) (Figu e 10a ( igh )). In he mo e accu a e e e ence
DSM (Figu e 10b), he co esponding alues we e as ollows: a ege a ion heigh o 0.44 o 0.99 m, an
a e age heigh o 0.76 m and a s anda d de ia ion o 0.12 m; he R2 alue o he linea eg ession o
d y biomass and he ege a ion heigh was 0.36. Wi h bo h DSMs, he ege a ion heigh was sligh ly
o e es ima ed. I was possible o isually iden i y he a eas wi hou ege a ion using bo h DSMs.
Using he e e ence DSM, he a eas wi h 0% e iliza ion could also be isually de ec ed; his was no
possible wi h he FPI DSM. The FPI DSM was no o su icien quali y o e eal di e ences in he
ege a ion heigh s, which was an expec ed esul , due o he ela i ely high heigh de ia ions in he
DSM. Wi h bo h DSMs, i is expec ed ha he DSM is eliable only wi hin he a ea su ounded by
he GCPs.
The eason o he wo se DSM quali y when using he FPI spec al came a could be due o he
lowe SNR and na owe bands, as well as o he ac ha he ma ching so wa e and he pa ame e s
migh no ha e been ideal o his ype o da a; we es ed se e al pa ame e op ions using he NGATE
so wa e, bu hey did no imp o e he esul s. One possible imp o emen could be o use mo e ideal
band combina ions du ing he ma ching phase. Fu he mo e, he ma ching algo i hm could be u he
uned, and di e en ma ching me hods should be s udied. The quali y o he DSM can be conside ed
p omising and su icien o image mosaic gene a ion i he ma ching ailu es can be in e pola ed and
il e ed a a su icien le el o quali y.
4.3.3. Image Mosaics
We calcula ed he o hopho o mosaics using he o ien a ion in o ma ion and he DSM. The RMSE
when using he GCPs as checkpoin s was less han 0.2 m o he x and y coo dina es.
4.4. Radiome ic P ocessing
We conduc ed adiome ic p ocessing o he en i e block and o s ip 3. In addi ion o he new
esul s o s ip 3, we ep oduced ele an pa s o he esul s om a p e ious s udy [33] wi h he ull
block in o de o e alua e he impac o adiome ic co ec ion on he inal applica ion. Signi ican
adiome ic di e ences appea ed wi hin he image mosaics, due o he ex emely a iable illumina ion
condi ions. This was appa en especially o he ull block (Figu e 11b), while s ip 3 was o a mo e
uni o m quali y (Figu e 11a). The adiome ic co ec ion g ea ly imp o ed he homogenei y o he da a
(Figu e 11c– ). The de ail o a co ec ed mosaic (Figu e 11g) shows he good quali y o he da a.
We used he a e age coe icien o a ia ion o he adiome ic ie poin s as he indica o o he
adiome ic homogenei y o he block (Figu e 12) ( he co ec ion me hods a e desc ibed in Table 3).
The coe icien s o a ia ion alues o he ull block (Figu e 12a) we e 0.14–0.18 when a adiome ic
co ec ion was no used and 0.05–0.12 when a adiome ic co ec ion was applied. The bes esul s o
0.05–0.08 we e ob ained wi h he ela i e block adjus men (BA: elA). Fo s ip 3 (Figu e 12b), he
alues we e 0.06–0.08 when a adiome ic block adjus men was no pe o med. Wi h adiome ic
co ec ion, we ob ained he lowes a ia ion coe icien s, app oxima ely 0.02, o NIR bands; he
a ia ion coe icien s we e be e han 0.03 o he g een bands and be e han 0.04 o he ed bands
(BA: elB, BRDF). The be e alues o he single s ip a e mos likely due o he ac ha less
a iabili y needed o be adjus ed o he single s ip. The co ec ion based on he spec al i adiance
Remo e Sens. 2013, 5 5026
measu emen on he g ound (g ound) appea ed o p o ide be e homogenei y han he UAV-based
co ec ion (ua ). The image-based co ec ion p o ided he bes homogenei y, bu he e we e some d i
e ec s, which appea ed as a sligh endency o he block o b igh en in he no h-eas di ec ion. The
esul s om he p e ious s udy [33] also showed ha all o he co ec ion app oaches—UAV
i adiance (ua ), g ound i adiance (g ound) and adiome ic block adjus men (BA: elA)—p o ided
ela i ely simila es ima es o he illumina ion a ia ions.
Figu e 11. Image mosaics wi h bands 29, 7 and 16. (a) S ip 3 and (b) ull block wi hou
any adiome ic co ec ions. (c) S ip 3 wi h adiome ic block adjus men using addi i e
and BRDF co ec ion (BA: elB, BRDF) and mosaics (d) wi h co ec ions using he
i adiance measu emen in he UAV (ua ), (e) wi h co ec ions using he i adiance
measu emen on he g ound (g ound) and ( ) wi h he adiome ic block adjus men wi h a
mul iplica i e co ec ion (BA: elA). (g) De ail o he co ec ed mosaic (colo s a e
op imized o he image dynamic ange). The mosaics o he ull blocks a e based on ou
p e ious s udy [33].
(a) (b) (c)
(d) (e) ( )

Remo e Sens. 2013, 5 5027
Figu e 11. Con .
(g)
Figu e 12. A e age coe icien s o a ia ion da a in adiome ic ie poin s (a) o he ull
block [33] and (b) o s ip 3 o di e en co ec ion me hods: no co : no co ec ion; ua :
co ec ions using he i adiance measu emen in he UAV; g ound: co ec ions using he
i adiance measu emen on he g ound; BA: elA: adiome ic block adjus men wi h a
mul iplica i e co ec ion; BA: elB, BRDF: adiome ic block adjus men wi h addi i e and
BRDF co ec ion.
(a) (b)
Six sample spec al p o iles a e shown in Figu e 13 o di e en adiome ic p ocessing cases. The
e lec ances we e aken as he median alue in a 1 m by 1 m image window. The g ea es di e ences
be ween he samples appea ed in he e lec ance alues in he NIR spec al egion. In all cases, he NIR
e lec ance o samples wi h low biomass alues (<1,000 kg·ha−1) we e clea ly lowe han
he NIR e lec ance o samples wi h highe biomass. Wi h he adiome ic co ec ions, he high
(>2,500 kg·ha−1) and medium (1,500–2,000 kg·ha−1) biomass samples could be sepa a ed
(Figu e 13b–e), which was no possible in he da a wi hou adiome ic co ec ions (Figu e 13a). The e
we e also some conce ns wi h he spec al p o iles. The g een e lec ance peaks we e no dis inc ,
especially in cases wi hou adiome ic co ec ion (Figu e 13a), which indica ed ha he adiome ic
co ec ion was no pe ec ly accu a e in hese cases. Fu he mo e, he e lec ance alues we e low in
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
500 600 700 800 900
Wa eleng h (nm)
no co
ua
g ound
BA: elA
Coe icien o
a ia ion
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
500 600 700 800 900
Wa eleng h (nm)
s ip3, no co
s ip3, BA: elB, BRDF
Coe icien o
a ia ion
Remo e Sens. 2013, 5 5028
he g een and ed e lec ance egions, ypically lowe han he da k e lec ance a ge used in
adiome ic calib a ion (Figu e 13 ); his means ha he e lec ance a g een and ed spec al egions
was ex apola ed. Radiome ic calib a ion is e y challenging a low e lec ance, because small e o s
in calib a ion will cause ela i ely la ge e o s in e lec ance alues.
We did no ha e e e ence spec ums o he ege a ion samples, so i was no possible o e alua e he
absolu e e lec ance accu acy. The e lec ance spec a o he h ee e lec ance a paulins indica ed ha he
spec a measu ed in images i ed well wi h he e e ence spec a measu ed in he labo a o y (Figu e 13 ).
Figu e 13. Re lec ance spec ums o six whea samples wi h di e en d y biomass alues
(gi en as kg·ha−1 o each spec um). The e lec ances we e aken as he median alue in a
1 m by 1 m image window. Da ase s: (a) s ip 3 wi hou any adiome ic co ec ions,
(b) s ip 3 wi h adiome ic block adjus men using addi i e and BRDF co ec ion (BA: elB,
BRDF), (c) ull block wi h co ec ions using he i adiance measu emen in he UAV (ua ),
(d) ull block wi h co ec ions using he i adiance measu emen on he g ound (g ound) and
(e) ull block wi h he adiome ic block adjus men wi h a mul iplica i e co ec ion
(BA: elA); ( ) e lec ance spec ums o h ee e e ence a paulins, P05, P20 and P30, o he
case BA: elA (P05- e , P20- e and P30- e a e e e ence spec ums measu ed in
he labo a o y).
(a) (b) (c) (d) (e)
( )
0
0.1
0.2
0.3
0.4
0.5
500 600 700 800 900
Wa eleng h (nm)
no co
272
4
2512
2028
1497
99
4
632
Re lec ance
0
0.1
0.2
0.3
0.4
0.5
500 600 700 800 900
Wa eleng h (nm)
BA: elB,
BRDF
272
4
2512
2028
1497
99
4
632
Re lec ance
0
0.1
0.2
0.3
0.4
0.5
500 600 700 800 900
Wa eleng h (nm)
ua
272
4
2512
2028
1497
99
4
632
Re lec ance
0
0.1
0.2
0.3
0.4
0.5
500 600 700 800 900
Wa eleng h (nm)
g ound
272
4
2512
2028
1497
99
4
632
Re lec ance
0
0.1
0.2
0.3
0.4
0.5
500 600 700 800 900
Wa eleng h (nm)
BA: elA
272
4
2512
2028
1497
99
4
632
Re lec ance
0
0.05
0.1
0.15
0.2
0.25
0.3
500 600 700 800 900
Wa eleng h (nm)
P05
P20
P30
P05- e
P20- e
P30- e
Re lec ance
Remo e Sens. 2013, 5 5029
The esul s o he adiome ic co ec ion we e p omising. In he u u e, i would be o in e es o
de elop an app oach ha in eg a es he image-based in o ma ion and he ex e nal i adiance
measu emen s. Special a en ion mus be paid o si ua ions in which he illumina ion condi ions change
be ween he ligh lines om sunny o di use (sun behind a cloud). Fu he mo e, a means o
co ec ing he shadows and opog aphic e ec s should be in eg a ed in o he me hod.
4.5. Biomass Es ima ion Using Spec ome ic In o ma ion om he FPI Spec al Came a
We es ed he pe o mance o he e lec ance ou pu da a om he FPI spec al came a in a biomass
es ima ion p ocess using a KNN es ima o . Fu he mo e, we calcula ed No malized Di e ence
Vege a ion Index (NDVI) (NDVI = (NIR 815.7 − R648)/(NIR 815.7 + R648)). Figu e 14 shows he biomass
es ima ion and NDVI s a is ics o di e en adiome ic p ocessing op ions.
Figu e 14. Biomass es ima ion s a is ics o di e en adiome ic p ocessing op ions:
(a) s ip 3 wi hou any adiome ic co ec ions (no co ); (b) s ip 3 wi h adiome ic block
adjus men using addi i e and BRDF co ec ion (BA: elB, BRDF); (c) ull block wi h
co ec ions using he i adiance measu emen in he UAV (ua ); (d) ull block wi h
co ec ions using he i adiance measu emen on he g ound (g ound); and (e) ull block
wi h he adiome ic block adjus men wi h a mul iplica i e co ec ion (BA: elA). Figu es
om le : biomass es ima e map (kg·ha−1), a sca e plo o he measu ed and es ima ed
biomass alues, No malized Di e ence Vege a ion Index (NDVI) map and a sca e plo o
he NDVI wi h espec o measu ed biomass alues. In he sca e plo s, he measu emen s
we e ca ied ou in a eas o a size o 1 m by 1 m.
(a)
(b)
Remo e Sens. 2013, 5 5030
Figu e 14. Con .
(c)
(d)
(e)
In he biomass es ima e maps, he a eas ha ha e no plan s we e qui e isible in all cases (le mos
plo s in Figu e 14). The s ips wi h 0% e iliza ion could be iden i ied in mos cases. The con inuous
lowe biomass pa e n in he middle o he a ea in he co ec ed da ase co esponds o a sligh downhill
e ain slope, which la ens owa ds he no h and mo es he aluable nu ien s he e. I a adiome ic
co ec ion had no been ca ied ou , he adiome ic di e ences caused by he changes in he illumina ion
we e clea ly isible in he mosaic and dis o ed he biomass es ima es (Figu e 14a). In he case o he
co ec ion based on he i adiance measu emen wi h he UAV (Figu e 14c), some s ip- ela ed a i ac s
(high biomass alues in s ips 2 and 4) appea ed; we also iden i ied hese inaccu acies in he co ec ion
pa ame e s in ou ecen s udy [33] and concluded ha hese inaccu acies we e likely due o some
shadowing e ec s o he i adiance senso in image s ips 1, 3 and 5 (sol able in u u e campaigns). The
la ges di e ences be ween he es ima es wi h a co ec ion based on he g ound i adiance measu emen
(Figu e 14d) and he image-based co ec ion (Figu e 14e) appea ed in he sou h-wes pa o he mosaic,
which had lowe biomass es ima es, and in he no h-eas pa o he mosaic as highe biomass es ima es
o he image-based me hod. The esul s o he wo image-based co ec ions we e qui e simila
Remo e Sens. 2013, 5 5037
18. Zecha, C.W.; Link, J.; Claupein, W. Mobile senso pla o ms: Ca ego isa ion and esea ch
applica ions in p ecision a ming. J. Sens. Sens. Sys . 2013, 2, 51–72.
19. Nackae s, K.; Delau é, B.; E e ae s, J.; Michiels, B.; Holmlund, C.; Mäkynen, J.; Saa i, H.
E alua ion o a ligh weig h UAS-p o o ype o hype spec al imaging. In . A ch. Pho og amm.
Remo e Sens. Spa . In o . Sci. 2010, 38, 478–483.
20. Honka aa a, E.; Kai osoja, J.; Mäkynen, J.; Pellikka, I.; Pesonen, L.; Saa i, H.; Salo, H.;
Hakala, T.; Ma kelin, L.; Rosnell, T. Hype spec al e lec ance signa u es and poin clouds o
p ecision ag icul u e by ligh weigh UAV imaging sys em. ISPRS Annal. Pho og amm. Remo e
Sens. Spa . In . Sci. 2012, I-7, 353–358.
21. Pölönen, I.; Salo, H.; Saa i, H.; Kai osoja, J.; Pesonen, L.; Honka aa a, E. Biomass es ima o o
NIR image wi h a ew addi ional spec al band images aken om ligh UAS. P oc. SPIE 2012,
8369, doi:10.1117/12.918551.
22. Schol en, F.; Wewel, F. Digi al 3D-da a acquisi ion wi h he high esolu ion s e eo
came a-ai bo ne (HRSC-A). In . A ch. Pho og amm. Remo e Sens. Spa . In o . Sci. 2000, 33,
901–908.
23. Lebe l, F.; I scha a, A.; Pock, T.; Meixne , P.; G ube , M.; Scholz, S.; Wieche , A. Poin clouds:
Lida e sus 3D ision. Pho og amm. Eng. Remo e Sens. 2010, 76, 1123–1134.
24. Haala, N.; Has ed , H.; Wol , K.; Ressl, C.; Bal usch, S. Digi al pho og amme ic came a
e alua ion—Gene a ion o digi al ele a ion models. Pho og amm. Fe ne kund. Geoin . 2010, 2,
99–115.
25. Hi schmülle , H. Semi-Global Ma ching: Mo i a ion, De elopmen and Applica ions. In
Pho og amme ic Week 2011; F i sch, D., Ed.; Wichmann Ve lag: Heidelbe g, Ge many, 2011;
pp. 173–184.
26. Rosnell, T.; Honka aa a, E. Poin cloud gene a ion om ae ial image da a acqui ed by a
quad ocop e ype mic o unmanned ae ial ehicle and a digi al s ill came a. Senso s 2012, 12,
453–480.
27. Rosnell, T.; Honka aa a, E.; Nu minen, K. On geome ic p ocessing o mul i- emp oal image da a
collec ed by ligh UAV sys ems. In . A ch. Pho og amm. Remo e Sens. Spa . In o . Sci. 2011, 38,
63–68.
28. Schaepman-S ub, G.; Schaepman, M.E.; Pain e , T.H.; Dangel, S.; Ma onchik, J.V. Re lec ance
quan i ies in op ical emo e sensing—De ini ions and case s udies. Remo e Sen. En i on. 2006,
103, 27–42.
29. Honka a a, E.; Hakala, T.; Saa i, H.; Ma kelin, L.; Mäkynen, J.; Rosnell, T. A p ocess o
adiome ic co ec ion o UAV image blocks. Pho og amm. Fe ne kund. Geoin o . 2012,
doi: 10.1127/1432-8364/2012/0106.
30. Fö s ne , W.; Gülch, E. A Fas Ope a o o De ec ion and P ecise Loca ion o Dis inc Poin s,
Co ne s and Cen e s o Ci cula Fea u es. In P oceedings o In e commission Con e ence on Fas
P ocessing o Pho og amme ic Da a, In e laken, Swize land, 2–4 June 1987; pp. 281–305.
31. Beisl, U. New Me hod o Co ec ion o Bidi ec ional E ec s in Hype spec al Images. In
P oceedings o Remo e Sensing o En i onmen al Moni o ing, GIS Applica ions, and Geology,
Toulouse, F ance, 17 Sep embe 2001.

Remo e Sens. 2013, 5 5038
32. Wal hall, C.L.; No man, J.M.; Welles, J.M.; Campbell, G.; Blad, B.L. Simple equa ion o
app oxima e he bidi ec ional e lec ance om ege a i e canopies and ba e soil su aces. Appl.
Op . 1985, 24, 383–387.
33. Hakala, T.; Honka aa a, E.; Saa i, H.; Mäkynen, J.; Kai osoja, J.; Pesonen, L.; Pölönen, I.
Spec al imaging om UAVs unde a ying illumina ion condi ions. In . A ch. Pho og amm.
Remo e Sens. Spa . In o . Sci. 2013, XL-1/W2, 189–194.
34. Ku a ekniikka Oy Pa ik Raski. A ailable online: h p://ku a ekniikka.com/ (accessed on 5
Sep embe 2013).
35. In e sil ISL29004 Da ashee . A ailable online: h p://www.in e sil.com/con en /dam/
In e sil/documen s/ n62/ n6221.pd (accessed on 9 Oc obe 2013).
36. Ko sian is, S.B. Supe ised machine lea ning: A e iew o classi ica ion echniques. In o ma ica
2007, 31, 249–268.
37. Na ional Land Su ey o Finland Open Da a License. A ailable on line:
h p://www.maanmi auslai os. i/en/NLS_open_da a_licence_ e sion1_20120501 (accessed on 10
Oc obe 2013).
38. Chiang, K.-W.; Tsai, M.-L.; Chu, C.-H. The de elopmen o an UAV bo ne di ec geo e e enced
pho og amme ic pla o m o g ound con ol poin ee applica ions. Senso s 2012, 12,
9161–9180.
39. Tu ne , D.; Luciee , A.; Wa son, C. An au oma ed echnique o gene a ing geo ec i ied mosaics
om ul a-high esolu ion Unmanned Ae ial Vehicle (UAV) image y, based on S uc u e om
Mo ion (S M) poin clouds. Remo e Sens. 2012, 4, 1392–1410.
40. Sna ely, N. Bundle : S uc u e om Mo ion (SFM) o Uno de ed Image Collec ions. A ailable
online: pho o ou .cs.washing on.edu/bundle / (accessed on 12 July 2013).
41. Ma hews, A.J.; Jensen, J.L.R. Visualizing and quan i ying ineya d canopy LAI using an
Unmanned Ae ial Vehicle (UAV) collec ed high densi y s uc u e om mo ion poin cloud.
Remo e Sens. 2013, 5, 2164–2183.
42. Honka aa a, E.; A biol, R.; Ma kelin, L.; Ma inez, L.; C ame , M.; Bo e , S.; Chandelie , L.;
Il es, R.; Klonus, S.; Ma shal, P.; e al. Digi al ai bo ne pho og amme y—A new ool o
quan i a i e emo e sensing?—A s a e-o - he-A e iew on adiome ic aspec s o digi al
pho og amme ic images. Remo e Sens. 2009, 1, 577–605.
43. Rich e , R.; Schläp e , D. Geo-a mosphe ic p ocessing o ai bo ne imaging spec ome y da a.
Pa 2: A mosphe ic/ opog aphic co ec ion. In . J. Remo e Sens. 2002, 23, 2631–2649.
44. Rich e , R.; Kellenbe ge , T.; Kau mann, H. Compa ison o opog aphic co ec ion me hods.
Remo e Sens. 2009, 1, 184–196.
45. Beisl, U.; Telaa , J.; on Schönema k, M. A mosphe ic Co ec ion, Re lec ance Calib a ion and
BRDF Co ec ion o ADS40 Image Da a. In P oceedings o he XXI ISPRS Cong ess,
Commission VII, Beijing, China, 3–11 July 2008.
46. Chandelie , L.; Ma ino y, G. Radiome ic ae ial iangula ion o he equaliza ion o digi al ae ial
images and o hoimages. Pho og amm. Eng. Remo e Sens. 2009, 75, 193–200.
47. Collings, S.; Cace a, P.; Campbell, N.; Wu, X. Empi ical models o adiome ic calib a ion o
digi al ae ial ame mosaics. IEEE T ans. Geosci. and Remo e Sens. 2011, 49, 2573–2588.
Remo e Sens. 2013, 5 5039
48. López, D.H.; Ga cía, B.F.; Pique as, J.G.; Aöcáza , G.V. An app oach o he adiome ic
ae o iangula ion o pho og amme ic images. ISPRS J. Pho og amm. Remo e Sens. 2011, 66,
883–893.
49. A zbe ge , C. Ad ances in emo e sensing o ag icul u e: Con ex desc ip ion, exis ing ope a ional
moni o ing sys ems and majo in o ma ion needs. Remo e Sens. 2013, 5, 949–981.
50. Wa s, A.C.; Amb osia, V.G.; Hinkley, E.A. Unmanned Ai c a Sys ems in emo e sensing and
scien i ic esea ch: Classi ica ion and conside a ions o use. Remo e Sens. 2012, 4, 1671–1692.
© 2013 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 license
(h p://c ea i ecommons.o g/licenses/by/3.0/).