Individual tree detection and classification with UAV-based photogrammetric point clouds and hyperspectral imaging

emo e sensing A icle Indi idual T ee De ec ion and Classi ica ion wi h UAV-Based Pho og amme ic Poin Clouds and Hype spec al Imaging Olli Ne alainen 1,*, Eija Honka aa a 1, Saka i Tuominen 2, Niko Viljanen 1, Teemu Hakala 1, Xiaowei Yu 1, Juha Hyyppä 1, Heikki Saa i 3, Ilkka Pölönen 4, Nil on N. Imai 5and An onio M. G. Tommaselli 5 1Finnish Geospa ial Resea ch Insi i u e, Na ional Land Su ey o Finland, Geodee in inne 2, 02430 Masala, Finland; [email p o ec ed] (E.H.); [email p o ec ed] (N.V.); [email p o ec ed] (T.H.); [email p o ec ed] (X.Y.); [email p o ec ed] (J.H.) 2Na u al Resou ces Ins i u e Finland, PL 2 00791 Helsinki, Finland; [email p o ec ed] 3VTT Mic oelec onics, P.O. Box 1000, FI-02044 VTT, Finland; [email p o ec ed] 4Depa men o Ma hema ical In o ma ion Tech., Uni e si y o Jy äskylä, P.O. Box 35, FI-40014 Jy äskylä, Finland; [email p o ec ed] 5Depa men o Ca og aphy, Uni . Es adual Paulis a (UNESP), P esiden e P uden e, SP 19060-900, B azil; [email p o ec ed] (N.N.I.); [email p o ec ed] (A.M.G.T.) *Co espondence: [email p o ec ed]; Tel.: +358-50-911-9062 Academic Edi o s: Fa id Melgani, F ancesco Nex, No man Ke le and P asad S. Thenkabail Recei ed: 8 Decembe 2016; Accep ed: 18 Feb ua y 2017; Published: 23 Feb ua y 2017 Abs ac : Small unmanned ae ial ehicle (UAV) based emo e sensing is a apidly e ol ing echnology. No el senso s and me hods a e en e ing he ma ke , o e ing comple ely new possibili ies o ca y ou emo e sensing asks. Th ee-dimensional (3D) hype spec al emo e sensing is a no el and powe ul echnology ha has ecen ly become a ailable o small UAVs. This s udy in es iga ed he pe o mance o UAV-based pho og amme y and hype spec al imaging in indi idual ee de ec ion and ee species classi ica ion in bo eal o es s. Ele en es si es wi h 4151 e e ence ees ep esen ing a ious ee species and de elopmen al s ages we e collec ed in June 2014 using a UAV emo e sensing sys em equipped wi h a ame o ma hype spec al came a and an RGB came a in highly a iable wea he condi ions. Dense poin clouds we e measu ed pho og amme ically by au oma ic image ma ching using high esolu ion RGB images wi h a 5 cm poin in e al. Spec al ea u es we e ob ained om he hype spec al image blocks, he la ge adiome ic a ia ion o which was compensa ed o by using a no el app oach based on adiome ic block adjus men wi h he suppo o in- ligh i adiance obse a ions. Spec al and 3D poin cloud ea u es we e used in he classi ica ion expe imen wi h a ious classi ie s. The bes esul s we e ob ained wi h Random Fo es and Mul ilaye Pe cep on (MLP) which bo h ga e 95% o e all accu acies and an F-sco e o 0.93. Accu acy o indi idual ee iden i ica ion om he pho og amme ic poin clouds a ied be ween 40% and 95%, depending on he cha ac e is ics o he a ea. Challenges in e e ence measu emen s migh also ha e educed hese numbe s. Resul s we e p omising, indica ing ha hype spec al 3D emo e sensing was ope a ional om a UAV pla o m e en in e y di icul condi ions. These no el me hods a e expec ed o p o ide a powe ul ool o au oma ing a ious en i onmen al close- ange emo e sensing asks in he e y nea u u e. Keywo ds: UAV; hype spec al; pho og amme y; adiome y; poin cloud; o es ; classi ica ion Remo e Sens. 2017,9, 185; doi:10.3390/ s9030185 www.mdpi.com/jou nal/ emo esensing Remo e Sens. 2017,9, 185 2 o 34 1. In oduc ion Knowing he ee species composi ion o a o es enables he es ima ion o he o es ’s economic alue and p oduces aluable in o ma ion o s udying o es ecosys ems. Today, o es in en o y in Scandina ia is based on an a ea-based app oach [ 1 , 2 ] whe e lase scanning (poin densi y abou 1 p s/m 2 ) and ae ial images ( esolu ion ypically 0.5 m) a e used as in en o y da a. Howe e , using hese app oaches, species-speci ic diame e dis ibu ions ha e poo p edic ion accu acy and imp o emen in ee species de ec ion is needed. Fo es da a using emo e sensing me hods has mos ly been concen a ing on o es s and (i.e., he ecologically homogeneous and spa ially con inuous pa o a o es ) and plo le el da a. Howe e , s and-le el o es a iables a e ypically an a e age o sum om he se o ees composing he s and. In calcula ing o es in en o y a iables such as he olume and biomass o he g owing s ock, ee-le el models a e ypically used nowadays [ 3 – 5 ]. Ve y high esolu ion emo e sensing da a allows mo ing om he s and le el o he le el o indi idual ees, which has ce ain bene i s, o example in p ecision o es y, o es managemen planning, biomass es ima ion and modeling o es g ow h [6]. Inc easing he le el o de ail o he o es can also imp o e de ailed modeling o o es s, which can be used o p edic o es g ow h and imp o e sa elli e-based emo e sensing o o es s by mo e accu a ely modeling he adia i e ans e wi hin he o es canopy. Fo es and ee species classi ica ion using mul i- o hype spec al imaging o lase scanning has been widely s udied [ 7 – 9 ]. Howe e , he da a has mainly been cap u ed om manned ai c a s o sa elli es, whe e o e s udies ha e been ocusing mo e on he o es o plo le el. The challenge in passi e imaging has been he dependence on sunligh and he high impac o he changing and di e en illumina ion condi ions on he adiome y o he da a [ 9 – 13 ]. Shadowing and b igh ening o indi idual ee c owns cause he pixels o a single ee c own o scale om eally da k pixels o eally b igh pixels. A ew me hods ha e been de eloped and sugges ed o educe he e ec o he changing illumina ion in he o es canopy [ 14 ], bu none o hem ha e been ex ensi ely es ed wi h high esolu ion spec al imaging da a om indi idual ees. In addi ion, illumina ion changes can be bene icial in classi ica ion asks, since hey po en ially p o ide species-speci ic in o ma ion o he ee s uc u e [15,16]. Ai bo ne Lase Scanning (ALS), bo h disc e e and ull-wa e o m, has been used o classi y ees in bo eal o es s [ 17 – 21 ]. The mos demanding ask using passi e imaging sys ems has always been disc imina ing pines om sp uces due o hei spec al simila i y. Howe e , s uc u al da a om lase scanning has shown i can cap u e he s uc u al di e ences o hese species, such as he e ical ex en o he lea y canopy [18,22,23]. Indi idual ees ha e been de ec ed using passi e da a [ 24 ], mainly using image segmen a ion using ex u al ea u es [ 25 , 26 ], bu he de elopmen o dense image ma ching me hods and compu ing powe has enabled he p oduc ion o high esolu ion pho og amme ic poin clouds. Thei abili y o p oduce s uc u al da a om a o es has been p esen ed in he li e a u e [ 27 – 30 ]. Pho og amme ic poin clouds p oduce g ea op-o - he-canopy in o ma ion ha enables he compu a ion o Canopy Heigh Models (CHM) [ 29 ] and hus de ec ion o indi idual ees [ 31 ]. Howe e , he no able limi a ion o pho og amme ic poin clouds compa ed o lase scanning is ha passi e imaging does no ha e good pene a ion abili y [ 30 ], especially wi h ae ial da a om manned ai c a s. Thus, lase scanning ha can mo e deeply pene a e o es canopies has been he main me hod o p o iding in o ma ion on o es s uc u e [ 1 , 2 ], especially a indi idual ee le el [ 32 ]. Pho og amme ic poin clouds de i ed om da a cap u ed using manned ai c a s ha e been used, bu he spa ial esolu ion is no usually accu a e enough o p oduce good de ec ion esul s. Small unmanned ae ial ehicles (UAVs) ha e been apidly inco po a ed in a ious emo e sensing applica ions, including o es y [ 33 – 35 ]. The use o UAVs in ae ial imaging has enabled measu emen s wi h highe spa ial esolu ion, imp o ing he esolu ion o pho og amme ic poin clouds and he acquisi ion o h ee-dimensional (3D) s uc u al da a om he o es [ 36 , 37 ]. Some s udies ha e been using da a om UAVs in indi idual ee heigh de e mina ion wi h good esul s [ 38 ]. Mul ispec al da a Remo e Sens. 2017,9, 185 3 o 34 acqui ed om UAVs has been used in ee species classi ica ion, bu he da a has been limi ed o RGB images and one nea -in a ed (NIR) channel using NIR-modi ied came as [ 39 – 41 ]. The de elopmen o low-weigh hype spec al imaging senso s is likely o inc ease he use o spec al da a collec ed om UAVs in o es y applica ions [34]. Recen ly, he de elopmen o small hype spec al imaging senso s has enabled high spec al and spa ial esolu ion measu emen s om UAVs. Se e al pushb oom hype spec al senso s ha e been implemen ed in UAVs [ 42 – 46 ]. The no el hype spec al came as ope a ing in he ame o ma p inciple o e in e es ing possibili ies o UAV emo e sensing by s able imaging geome y and by gi ing an oppo uni y o make 3D hype spec al measu emen s [ 47 – 49 ]. Näsi e al. [ 50 ] p esen ed he i s s udy wi h 3D hype spec al UAV imaging in indi idual ee-le el analysis o ba k bee le damage in sp uce o es s. To he bes o he au ho s’ knowledge, he classi ica ion o indi idual ees using hype spec al image y om UAVs has no ye been s udied. No el hype spec al imaging echnology based on a a iable ai gap Fab y–Pé o in e e ome e (FPI) ope a ing in he isible o nea -in a ed spec al ange (500–900 nm; VNIR) [ 47 – 49 ] was used in his s udy. The FPI echnology makes i possible o manu ac u e a ligh weigh , ame o ma hype spec al image ope a ing on he ime-sequen ial p inciple. The FPI came a can be easily moun ed on a small UAV oge he wi h an RGB came a, which enables he simul aneous hype spec al imaging wi h high spa ial esolu ion pho og amme ic poin cloud c ea ion [50]. The objec i e o his s udy is o es he use o high esolu ion pho og amme ic poin clouds oge he wi h hype spec al UAV imaging in indi idual ee de ec ion and classi ica ion in bo eal o es s. In pa icula , he da a p ocessing challenges in a eal o es en i onmen will be s udied and he impo ance o di e en spec al and s uc u al ea u es in ee species classi ica ion will be assessed. P e ious s udies using UAV da a ha e no u ilized bo h spec al and s uc u al da a in ee species classi ica ion, no ha e hey p o ided eliable comple e wo k lows o de ec and classi y indi idual ees om la ge o es ed scenes. The ma e ials and me hods a e p esen ed in Sec ion 2. The esul s a e gi en in Sec ion 3and discussed Sec ion 4. Conclusions a e p o ided in Sec ion 5. 2. Ma e ials and Me hods 2.1. S udy A ea and Re e ence T ee Da a Collec ion The s udy a ea was he Vesijako esea ch o es a ea in he municipali y o Padasjoki in sou he n Finland (app oxima ely 61 ◦ 24 0 N and 25 ◦ 02 0 E). The a ea has been used as a esea ch o es by he Na u al Resou ces Ins i u e o Finland (and i s p edecesso , he Finnish Fo es Resea ch Ins i u e). Ele en expe imen al es si es om s ands domina ed by pine (Pinus syl es is), sp uce (Picea abies), bi ch (Be ula pendula) and la ch (La ix sibi ica) we e used in his s udy. The es si es ep esen ed de elopmen s ages om young o middle-aged and ma u e s ands (no seedling s ands o clea cu a eas). Wi hin each es si e, he e we e 1–16 expe imen al plo s ea ed wi h di e ing sil icul u al schemes and cu ing sys ems (al oge he 56 expe imen al plo s). The size o he expe imen al plo s was 1000–2000 m 2 . Wi hin he expe imen al plo s, all ees wi h b eas -heigh diame e o a leas 50 mm we e measu ed as ally ees in 2012–2013. Fo each ally ee, he ollowing a iables we e eco ded: ela i e loca ion wi hin he plo , ee species, and diame e a b eas heigh . Heigh was measu ed o a numbe o sample ees in each plo and es ima ed o all ally ees using heigh models calib a ed by sample ee measu emen s. The geog aphic loca ion o he expe imen al plo co ne poin s was measu ed wi h a Global Posi ioning Sys em (GPS) de ice, and he loca ions we e p ocessed wi h local base s a ion da a, wi h an a e age e o o app oxima ely 1 m. Fo his s udy, a o al o 300 ixed- adius (9 m) ci cula sample plo s we e placed on he expe imen al plo s in he ele en es si es (4–8 ci cula plo s pe each expe imen al plo , depending on he size and shape o he expe imen al plo s). The plo a iables o he ci cula sample plo s we e calcula ed on he basis o he ee maps o he expe imen al plo s. The s a is ics o he main o es a iables in he ield measu emen da a a e p esen ed in Tables 1and 2. Some o he sample plo s Remo e Sens. 2017,9, 185 4 o 34 ha e an excep ionally high amoun o g owing s ock compa ed o alues ypical o his geog aphic a ea, which can be seen in he maximum alues and s anda d de ia ion o he olumes in he ield obse a ions o his s udy a ea. The a eas wi h he highes amoun o g owing s ock a e domina ed by pine and la ch. Figu e 1shows he loca ions o he s udy a eas. Table 1. A e age, maximum (Max), minimum (Min) and s anda d de ia ion (S d.) o ield a iables in he ield da a in he s udy a ea. Fo es Va iable A e age Max Min S d. To al olume, m3/ha 328.3 1160.8 33.2 220.4 Volume o Sco s pine *, m3/ha 240.0 1110.8 0 212.2 Volume o No way sp uce, m 3 /ha 46.6 420.0 0 88.7 Volume o b oadlea ed, m3/ha 41.7 352.6 0 84.5 Mean diame e , cm 22.9 55.4 13.9 7.6 Mean heigh , m 21.2 39.4 14.3 5.1 Basal a ea, m2/ha 31.3 78.5 3.7 14.6 * Including La ix sp. Table 2. Fo es plo densi ies in he s udy a ea. Densi ies a e p esen ed as ees pe hec a e. Tes Si e Min Max Mean Median Numbe Plo s 01 676 2964 1699 1756 16 02 310 1938 898 625 11 0304 333 1247 657 594 9 05 173 1666 800 715 6 06 1381 1381 1381 1381 1 07 850 850 850 850 1 08 766 983 855 818 3 09 435 2701 1909 2250 4 10 691 2016 1354 1354 2 11 468 468 468 468 1 Remo e Sens. 2017, 9, 185 4 o 33 obse a ions o his s udy a ea. The a eas wi h he highes amoun o g owing s ock a e domina ed by pine and la ch. Figu e 1 shows he loca ions o he s udy a eas. Table 1. A e age, maximum (Max), minimum (Min) and s anda d de ia ion (S d.) o ield a iables in he ield da a in he s udy a ea. Fo es Va iable A e age Max Min S d. To al olume, m 3 /ha 328.3 1160.8 33.2 220.4 Volume o Sco s pine *, m 3 /ha 240.0 1110.8 0 212.2 Volume o No way sp uce, m 3 /ha 46.6 420.0 0 88.7 Volume o b oadlea ed, m 3 /ha 41.7 352.6 0 84.5 Mean diame e , cm 22.9 55.4 13.9 7.6 Mean heigh , m 21.2 39.4 14.3 5.1 Basal a ea, m 2 /ha 31.3 78.5 3.7 14.6 * Including La ix sp. Table 2. Fo es plo densi ies in he s udy a ea. Densi ies a e p esen ed as ees pe hec a e. Tes Si e Min Max Mean Median Numbe Plo s 01 676 2964 1699 1756 16 02 310 1938 898 625 11 0304 333 1247 657 594 9 05 173 1666 800 715 6 06 1381 1381 1381 1381 1 07 850 850 850 850 1 08 766 983 855 818 3 09 435 2701 1909 2250 4 10 691 2016 1354 1354 2 11 468 468 468 468 1 Figu e 1. The geog aphical loca ions o he s udy a eas. The a ea is loca ed in he Vesijako esea ch o es a ea in he municipali y o Padasjoki in sou he n Finland (app oxima ely 61°24′N and 25°02′E). Figu e 1. The geog aphical loca ions o he s udy a eas. The a ea is loca ed in he Vesijako esea ch o es a ea in he municipali y o Padasjoki in sou he n Finland (app oxima ely 61 ◦ 24 0 N and 25 ◦ 02 0 E). Remo e Sens. 2017,9, 185 5 o 34 The e e ence ees collec ed in he ield measu emen s we e isually compa ed o he collec ed UAV o homosaics in o de o check hei geome ic co espondence a he indi idual ee le el (o homosaic calcula ion is desc ibed in Sec ion 2.3). Some misalignmen could be obse ed, which was due o challenges in ee posi ioning in he g ound condi ions, due o he geo e e encing quali y o he UAV o homosaics, as well as due o di e en cha ac e is ics a he objec iew and he ae ial iew. The misaligned e e ence ees we e manually aligned wi h he ees in he UAV o homosaics o emo ed i he co esponding ee could no be iden i ied eliably om he UAV o homosaics. This was pe o med using Quan um GIS (QGIS) by o e laying he ield da a o e he RGB and FPI mosaics. 2.2. Remo e Sensing Da a Cap u e Al oge he , 11 es si es we e cap u ed using a small UAV in eigh sepa a e ligh s on 25–26 June 2014, in he Vesijako es si e (Tables 3and 4). Fo geo e e encing pu poses, h ee o nine G ound Con ol Poin s (GCPs) wi h c oss-shaped signals wi h an a m leng h o 3 m and wid h o 30 cm we e ins alled in each es a ea. Re lec ance panels o size 1 m × 1 m and wi h a nominal e lec i i y o 0.03, 0.1 and 0.5 we e ins alled in he a ea o e lec ance ans o ma ion pu poses [51]. The e e ence e lec ance alues we e measu ed in a labo a o y wi h an es ima ed accu acy o 2%–5% using he FIGIFIGO goniospec ome e [52]. The UAV emo e sys em belonging o he Finnish Geospa ial Resea ch Ins i u e (FGI) was used in he da a cap u e. The UAV pla o m ame was a Ta o 960 hexacop e wi h Ta o 5008 (340 KV) b ushless elec ic mo o s. The au opilo was a Pixhawk equipped wi h A ducop e 3.15 i mwa e. The sys em’s payload capaci y is abou 3 kg and he ligh ime is 10–30 min, depending on payload, ba e y, condi ions, and ligh s yle. A no el hype spec al came a based on a uneable Fab y–Pé o in e e emo e (FPI) [ 47 – 49 ] was used o cap u e he spec al da a. The FPI came a cap u es ame- o ma hype spec al images in a ime-sequen ial mode. Due o he sequen ial exposu e o he indi idual bands (0.075 s be ween adjacen exposu es, 1.8 s du ing he en i e da a cube wi h 24 exposu es), each band o he da a cube has a sligh ly di e en posi ion and o ien a ion, which has o be aken in o accoun in he pos -p ocessing phase. The image size was 1024 × 648 pixels, and he pixel size was 11 µ m. The FPI came a has a ocal leng h o 10.9 mm; he 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 . The came a sys em has an i adiance senso (based on he In e sil ISL29004 pho ode ec o ) o measu e he wideband i adiance du ing each exposu e [ 53 ]. The UAV was also equipped wi h he Ocean Op ics spec ome e and i adiance senso o moni o illumina ion condi ions; due o some echnical p oblems, he da a quali y was no sui able o he adiome ic co ec ion. A GPS ecei e is used o eco d he exac ime o he i s exposu e o each da a cube. Spec al se ings can be selec ed acco ding o he equi emen s. In his s udy, al oge he , 33 bands we e used wi h he ull wid h o he hal maximum (FWHM) o 11–31 nm. The se ings used in his s udy a e gi en in Table 5. In o de o cap u e high spa ial esolu ion da a, he UAV was also equipped wi h an o dina y RGB compac digi al came a, he Samsung NX1000. The came a has a 23.5 × 15.7 mm complemen a y me al-oxide semiconduc o (CMOS) senso wi h 20.3 megapixels and a 16 mm lens. The UAV sys em is p esen ed in Figu e 2. The lying heigh was 83–94 m om he g ound le el, p o iding an a e age G ound Sampling Dis ance (GSD) o 8.6 cm o he FPI images and 2.3 cm o he RGB images on he g ound le el. The ligh heigh was 62–73 m om he ee ops (calcula ed o an a e age ee heigh o 21 m). Thus, he a e age GSDs we e 6.5 cm and 1.8 cm a ee ops o he FPI and RGB da a se s, espec i ely. The ligh speed was abou 4 m/s. The FPI image blocks had a e age o wa d and side o e laps o 67% and 61%, espec i ely, a he nominal g ound le el and 58% and 50%, espec i ely, a he ee op le el. Fo he RGB blocks, he a e age o wa d and side o e laps we e 78% and 73%, espec i ely, a he g ound le el. A he le el o ee ops, he a e age o wa d and side o e laps we e 72% and 65%, espec i ely. The o e laps o RGB blocks we e sui able o he o ien a ion p ocessing and poin cloud gene a ion. The o e lapping o FPI images we e qui e low in he o es ed scene wi h la ge Remo e Sens. 2017,9, 185 6 o 34 heigh di e ences, hus he combined p ocessing wi h RGB images was necessa y o enable he highes quali y geome ic econs uc ion. Table 3. Fligh condi ions and came a se ings du ing he ligh s. Median i adiance was aken om In e sil ISL29004 i adiance measu emen s. Tes Si e Da e Time (GPS*) Wea he Sola Ele a ion Sun Azimu h Median I ad Exposu e (ms) 01 26.6 11:07 o 11:23 Cloudy 50.91 199.22 2602 10 02 26.6 12:09 o 12:22 Cloudy 47.38 219.63 4427 12 0304 25.6 10:38 o 10:51 Va ying 51.79 188.36 a iable 6 05 25.6 09:26 o 09:40 Va ying 50.93 160.69 a iable 6 0607 25.6 12:14 o 12:24 Cloudy 47 221.30 3773 10 08 26.6 09:58 o 10:09 Sunny 51.84 173.41 13894 10 0910 25.6 13:51 o 14:12 Cloudy 37.20 249.45 2546 8 11 26.6 08:49 o 08:58 Va ying 49.1 148.44 13982 10 *GPS, Global Posi ioning Sys em. Table 4. P ope ies o image blocks calcula ed a g ound le el and a ee op le el (N gcp: numbe o g ound con ol poin s (GCPs), FH: lying heigh ; w;sl: o wa d and side o e laps; FPI: Fab y–Pé o in e e ome e , GSD: g ound sampling dis ance). Block N GCP G ound T ee ops FH (m) FPI RGB FH (m) FPI RGB GSD (m) w;sl (%) GSD (m) w;sl (%) GSD (m) w;sl (%) GSD (m) w;sl (%) 01 7 94 0.094 64;54 0.025 76;68 73 0.073 54;41 0.020 69;59 02 4 88 0.088 65;57 0.024 77;70 67 0.067 53;43 0.018 69;61 0304 9 85 0.085 69;62 0.023 79;73 64 0.064 59;49 0.017 73;65 05 5 86 0.086 59;34 0.023 73;54 65 0.065 59;34 0.017 73;54 06 3 94 0.094 78;79 0.025 85;86 73 0.073 71;73 0.020 80;81 07 4 94 0.094 73;71 0.025 82;80 73 0.073 65;63 0.020 77;74 08 5 86 0.086 64;64 0.023 76;75 65 0.065 52;52 0.017 68;67 09 4 84 0.084 69;57 0.023 79;70 63 0.063 58;43 0.017 72;60 10 3 83 0.083 69;67 0.022 79;77 62 0.062 58;56 0.017 72;70 11 4 83 0.083 60;61 0.022 74;73 62 0.062 47;47 0.017 65;63 A e age 86 0.086 67;61 0.023 78;73 65 0.065 58;50 0.018 72;65 Remo e Sens. 2017, 9, 185 6 o 33 heigh di e ences, hus he combined p ocessing wi h RGB images was necessa y o enable he highes quali y geome ic econs uc ion. Table 3. Fligh condi ions and came a se ings du ing he ligh s. Median i adiance was aken om In e sil ISL29004 i adiance measu emen s. Tes Si e Da e Time (GPS*) Wea he Sola Ele a ion Sun Azimu h Median I ad Exposu e (ms) 01 26.6 11:07 o 11:23 Cloudy 50.91 199.22 2602 10 02 26.6 12:09 o 12:22 Cloudy 47.38 219.63 4427 12 0304 25.6 10:38 o 10:51 Va ying 51.79 188.36 a iable 6 05 25.6 09:26 o 09:40 Va ying 50.93 160.69 a iable 6 0607 25.6 12:14 o 12:24 Cloudy 47 221.30 3773 10 08 26.6 09:58 o 10:09 Sunny 51.84 173.41 13894 10 0910 25.6 13:51 o 14:12 Cloudy 37.20 249.45 2546 8 11 26.6 08:49 o 08:58 Va ying 49.1 148.44 13982 10 *GPS, Global Posi ioning Sys em. Table 4. P ope ies o image blocks calcula ed a g ound le el and a ee op le el (N gcp: numbe o g ound con ol poin s (GCPs), FH: lying heigh ; w;sl: o wa d and side o e laps; FPI: Fab y–Pé o in e e ome e , GSD: g ound sampling dis ance). Block N GCP G ound T ee ops FH (m) FPI RGB FH (m) FPI RGB GSD (m) w;sl (%) GSD (m) w;sl (%) GSD (m) w;sl (%) GSD (m) w;sl (%) 01 7 94 0.094 64;54 0.025 76;68 73 0.073 54;41 0.020 69;59 02 4 88 0.088 65;57 0.024 77;70 67 0.067 53;43 0.018 69;61 0304 9 85 0.085 69;62 0.023 79;73 64 0.064 59;49 0.017 73;65 05 5 86 0.086 59;34 0.023 73;54 65 0.065 59;34 0.017 73;54 06 3 94 0.094 78;79 0.025 85;86 73 0.073 71;73 0.020 80;81 07 4 94 0.094 73;71 0.025 82;80 73 0.073 65;63 0.020 77;74 08 5 86 0.086 64;64 0.023 76;75 65 0.065 52;52 0.017 68;67 09 4 84 0.084 69;57 0.023 79;70 63 0.063 58;43 0.017 72;60 10 3 83 0.083 69;67 0.022 79;77 62 0.062 58;56 0.017 72;70 11 4 83 0.083 60;61 0.022 74;73 62 0.062 47;47 0.017 65;63 A e age 86 0.086 67;61 0.023 78;73 65 0.065 58;50 0.018 72;65 Figu e 2. Unmanned ae ial ehicle (UAV) emo e sensing sys em o he Finnish Geospa ial Resea ch Ins i u e (FGI) is based on he Ta o 960 hexacop e . Figu e 2. Unmanned ae ial ehicle (UAV) emo e sensing sys em o he Finnish Geospa ial Resea ch Ins i u e (FGI) is based on he Ta o 960 hexacop e . Remo e Sens. 2017,9, 185 7 o 34 Table 5. Spec al se ings o he Fab y–Pe o in e e ome e (FPI) came a a isible (VIS) and nea -in a ed (NIR) channels. L0: cen al wa eleng h; FWHM: ull wid h a hal maximum. L0 (nm): 507.60, 509.50, 514.50, 520.80, 529.00, 537.40, 545.80, 554.40, 562.70, 574.20, 583.60, 590.40, 598.80, 605.70, 617.50, 630.70, 644.20, 657.20, 670.10, 677.80, 691.10, 698.40, 705.30, 711.10, 717.90, 731.30, 738.50, 751.50, 763.70, 778.50, 794.00, 806.30, 819.70 FWHM (nm): 11.2, 13.6, 19.4, 21.8, 22.6, 20.7, 22.0, 22.2, 22.1, 21.6, 18.0, 19.8, 22.7, 27.8, 29.3, 29.9, 26.9, 30.3, 28.5, 27.8, 30.7, 28.3, 25.4, 26.6, 27.5, 28.2, 27.4, 27.5, 30.5, 29.5, 25.9, 27.3, 29.9 Imaging condi ions we e qui e windless, bu illumina ion a ied a lo in di e en ligh s. Illumina ion condi ions we e cloudy and qui e uni o m du ing ligh s 01, 02, 0607 and 0910. Tes si e 08 was cap u ed unde sunny condi ions, and du ing ligh s 0304, 05 and 11 he illumina ion condi ions a ied be ween sunny o cloudy. I adiance eco dings by he In e sil ISL29004 i adiance senso du ing he ligh s a e p esen ed in Figu e 3. The eco dings we e qui e uni o m du ing he ligh s cap u ed in cloudy condi ions. The a iabili y in i adiance measu emen s du ing ligh s 0304, 05, 08 and 11 we e pa ially due o changing wea he and pa ially due o il ing o he i adiance senso in di e en ligh di ec ions, hus ob aining di e en le els o i adia ion. Remo e Sens. 2017, 9, 185 7 o 33 Table 5. Spec al se ings o he Fab y–Pe o in e e ome e (FPI) came a a isible (VIS) and nea -in a ed (NIR) channels. L0: cen al wa eleng h; FWHM: ull wid h a hal maximum. L0 (nm): 507.60, 509.50, 514.50, 520.80, 529.00, 537.40, 545.80, 554.40, 562.70, 574.20, 583.60, 590.40, 598.80, 605.70, 617.50, 630.70, 644.20, 657.20, 670.10, 677.80, 691.10, 698.40, 705.30, 711.10, 717.90, 731.30, 738.50, 751.50, 763.70, 778.50, 794.00, 806.30, 819.70 FWHM (nm): 11.2, 13.6, 19.4, 21.8, 22.6, 20.7, 22.0, 22.2, 22.1, 21.6, 18.0, 19.8, 22.7, 27.8, 29.3, 29.9, 26.9, 30.3, 28.5, 27.8, 30.7, 28.3, 25.4, 26.6, 27.5, 28.2, 27.4, 27.5, 30.5, 29.5, 25.9, 27.3, 29.9 Imaging condi ions we e qui e windless, bu illumina ion a ied a lo in di e en ligh s. Illumina ion condi ions we e cloudy and qui e uni o m du ing ligh s 01, 02, 0607 and 0910. Tes si e 08 was cap u ed unde sunny condi ions, and du ing ligh s 0304, 05 and 11 he illumina ion condi ions a ied be ween sunny o cloudy. I adiance eco dings by he In e sil ISL29004 i adiance senso du ing he ligh s a e p esen ed in Figu e 3. The eco dings we e qui e uni o m du ing he ligh s cap u ed in cloudy condi ions. The a iabili y in i adiance measu emen s du ing ligh s 0304, 05, 08 and 11 we e pa ially due o changing wea he and pa ially due o il ing o he i adiance senso in di e en ligh di ec ions, hus ob aining di e en le els o i adia ion. The na ional ALS da a by he Na ional Land Su ey o Finland (NLS) was used o p o ide he g ound le el, and i was also used as he geome ic e e ence o e alua ing he geome ic quali y o pho og amme ic p ocessing [54]. The minimum poin densi y o he NLS’s ALS da a is hal a poin pe squa e me e, and he ele a ion accu acy o he poin s in well-de ined su aces is 15 cm. The ho izon al accu acy o he da a is 60 cm. The ALS da a used in his s udy was collec ed on 12 May, 2012. Figu e 3. I adiance eco dings du ing di e en ligh s measu ed using he onboa d i adiance senso . 2.3. UAV Da a P ocessing Rigo ous p ocessing was equi ed in o de o de i e quan i a i e in o ma ion om he image y. The p ocessing o FPI came a images is simila o any ame o ma came a images ha co e he a ea o in e es wi h a la ge numbe o images. The majo di e ence is he p ocessing o he nonaligned spec al bands due o he ime-sequen ial imaging p inciple (Sec ion 2.2) o which a egis a ion p ocedu e has been de eloped. The image da a p ocessing chain o ee pa ame e es ima ion included he ollowing s eps: 1. Applying adiome ic labo a o y calib a ion co ec ions o he FPI images. 2. De e mina ion o he geome ic imaging model, including in e io and ex e io o ien a ions o he images. 3. Using dense image ma ching o c ea e a Digi al Su ace Model (DSM). 4. Regis a ion o he spec al bands o FPI images. Figu e 3. I adiance eco dings du ing di e en ligh s measu ed using he onboa d i adiance senso . The na ional ALS da a by he Na ional Land Su ey o Finland (NLS) was used o p o ide he g ound le el, and i was also used as he geome ic e e ence o e alua ing he geome ic quali y o pho og amme ic p ocessing [ 54 ]. The minimum poin densi y o he NLS’s ALS da a is hal a poin pe squa e me e, and he ele a ion accu acy o he poin s in well-de ined su aces is 15 cm. The ho izon al accu acy o he da a is 60 cm. The ALS da a used in his s udy was collec ed on 12 May 2012. 2.3. UAV Da a P ocessing Rigo ous p ocessing was equi ed in o de o de i e quan i a i e in o ma ion om he image y. The p ocessing o FPI came a images is simila o any ame o ma came a images ha co e he a ea o in e es wi h a la ge numbe o images. The majo di e ence is he p ocessing o he nonaligned spec al bands due o he ime-sequen ial imaging p inciple (Sec ion 2.2) o which a egis a ion p ocedu e has been de eloped. The image da a p ocessing chain o ee pa ame e es ima ion included he ollowing s eps: 1. Applying adiome ic labo a o y calib a ion co ec ions o he FPI images. 2. De e mina ion o he geome ic imaging model, including in e io and ex e io o ien a ions o he images. 3. Using dense image ma ching o c ea e a Digi al Su ace Model (DSM). Remo e Sens. 2017,9, 185 8 o 34 4. Regis a ion o he spec al bands o FPI images. 5. De e mina ion o a adiome ic imaging model o ans o m he digi al numbe s (DNs) da a o e lec ance. 6. Calcula ing he hype spec al and RGB image mosaics. 7. Subsequen emo e sensing analysis The image p ep ocessing and pho og amme ic p ocessing (s eps 1–3) we e ca ied ou using 3.5 GHz quad-co e PC wi h 88 GB RAM and a GeFo ce GTX 980 g aphics p ocessing uni (GPU). Fo he band egis a ion, adiome ic p ocessing and mosaic calcula ion (s eps 4–6), se e al egula o ice PCs we e used in pa allel. Remo e sensing analysis was pe o med using 3.5 GHz quad-co e PC wi h 24 GB RAM. In he ollowing sec ions, he geome ic (2, 3, 4) and adiome ic (1, 5, 6) p ocessing s eps and es ima ion p ocess (7) used in his in es iga ion a e desc ibed. 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 and de e mina ion o he 3D objec model. The Agiso Pho oScan P o essional comme cial so wa e (AgiSo LLC, S . Pe e sbu g, Russia) and he FGI’s in-house C++ so wa e (spec al band egis a ion) we e used o he geome ic p ocessing. Pho oScan pe o ms pho o-based 3D econs uc ion based on ea u e de ec ion and dense ma ching, and i is widely used in p ocessing UAV images [ 50 , 55 ]. In he o ien a ion p ocessing, he quali y was se o “high”, which means ha he ull esolu ion images we e used. The se ings o he numbe o key poin s pe image we e 40,000 and o he inal numbe o ie poin s pe image 1000; 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 ini ial p ocessing p o ided image o ien a ions and spa se poin clouds in he in e nal coo dina e sys em o he so wa e. Fo he da a, an au oma ic ou lie emo al was pe o med on he basis o he esiduals ( e-p ojec ion e o ) (10% o he poin s wi h he la ges e o s we e emo ed), as well as s anda d de ia ions o he ie poin 3D coo dina es ( econs uc ion unce ain y) (10% o he poin s wi h he la ges unce ain y we e emo ed). Some ou lie poin s we e also emo ed manually om he spa se poin cloud (poin s on he sky and below he g ound). The image o ien a ions we e ans o med o he ETRS-TM35FIN coo dina e sys em using he GCPs in he a ea. The ou pu s o he p ocess we e he came a calib a ions (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. O ien a ions o he FPI hype cubes we e de e mined in a simul aneous p ocessing wi h he RGB images in he Pho oScan. Th ee FPI image bands ( e e ence bands) we e included simul aneously in he p ocessing: band 4: L 0 = 520.8 nm, d = 1.125 s; band 12: L 0 = 590.4 nm, d = 1.725 s; band 16: L 0 = 630.7 nm, d = 0.15 s. (L0 is he cen al wa eleng h and d is he ime di e ence o he i s exposu e o he da a cube.) The o ien a ions o hose FPI image bands ha we e no included in he Pho oScan p ocessing we e de e mined using he space esec ion me hod de eloped a he FGI. This me hod used he ie poin s calcula ed du ing he Pho oScan p ocessing as GCPs and de e mined he co esponding image coo dina es by co ela ion ma ching o each uno ien ed band o he e e ence band 4 and inally calcula ed he space esec ion. This p ocessing p oduced EOPs o all he bands in all hype cubes. Fo accu a e dense poin cloud gene a ion, he o ien a ions o he RGB da ase s we e also de e mined sepa a ely wi hou he FPI images. Dense poin clouds wi h 5 cm poin in e al we e hen gene a ed using wo- imes downsampled RGB images. A mild il e ing was used o elimina e ou lie s, which allowed high heigh di e ences o he da a se . Remo e Sens. 2017,9, 185 9 o 34 2.3.2. Radiome ic P ocessing Ou p e e ed app oach was o analyze di e en es a eas simul aneously. Thus, i was necessa y o scale he adiome ic alues in each da a se o a simila e e ence scale. Calib a ion o he DNs o e lec ance ac o s was he p e e ed app oach. We ins alled he e lec ance e e ence panels close o he akeo place in each es si e in o de o ca y ou he e lec ance ans o ma ion using he empi ical line me hod. Un o una ely, he a ge s we e inside he o es and su ounded by all ees, hus he illumina ion condi ions in e e ence a ge s did no co espond o he illumina ion condi ions on op o he canopy. Because o his, hey did no p o ide accu a e e lec ance calib a ion by he empi ical line me hod. A u he challenge was ha he illumina ion condi ions we e a iable du ing many o he ligh s, p o iding la ge ela i e di e ences in adiome ic alues wi hin he blocks and be ween di e en blocks (Figu e 3). A modi ied app oach was used o elimina e adiome ic di e ences o images. We applied he adiome ic block adjus men and in- ligh i adiance da a o no malize he di e ences wi hin and be ween he blocks [ 49 , 53 , 56 ]. We selec ed es si e 06 as he e e ence es si e and calcula ed he empi ical line calib a ion using his a ea (a abs_ e , b abs_ e ). The a ea o he e lec ance panels was ela i ely open in block 06, and he su ounding ege a ion did no shade he panels. We selec ed a e e ence image in each block i( i∈ 1,...,11) and no malized he es o he images jin he block o his image using he ela i e scaling ac o a el_ i_j . The e e ence image was selec ed so ha i s illumina ion condi ions ep esen ed well he condi ions o he majo i y o he da a se . The a p io i alues o he ela i e co ec ion pa ame e s a el_ i_j we e de i ed om he In e sil ISL29004 i adiance measu emen s: a el_ i_j=i adiance i_ e i adiance i_j whe e i adiance i_j and i adiance i_ e a e he i adiance measu emen s du ing he acquisi ion o image jand e e ence image e . The a el_ i_j alues we e adjus ed using he adiome ic block adjus men me hod [ 49 , 53 , 56 ]. The ela i e scaling ac o (s i_ e ) was calcula ed be ween each block and he e e ence block 06 using he In e sil i adiance alues o e e ence images and scaled using he exposu e imes o he e e ence block and he block i( exp_ e , exp_ i ). Median il e ing was used in he i adiance alues o elimina e he ins abili y o he i adiance measu emen s: s i_ e =i adiance e _ e × exp_ e i adiance i_ e × exp _ i Thus, he inal equa ion o he ans o ma ion om DNs o e lec ance was Re lec ance =aabs_ e s i_ e a el_ i_jDN i_j+babs_ e We did no apply co ec ion o he bidi ec ional e lec ance (BRDF) e ec s in he da a se . BRDF e ec s we e nonexis en in he image blocks cap u ed unde cloudy condi ions. Fo he da a se s cap u ed in sunny condi ions, he maximum iew angles o he objec o di e en blocks we e 2.6 ◦ –4.8 ◦ in he ligh di ec ion and 3.9 ◦ –9.4 ◦ in he c oss- ligh di ec ion a he op o he canopy; hus, he BRDF e ec s could be assumed o be insigni ican . The adiome ic model pa ame e s we e calcula ed sepa a ely o each band. 2.3.3. Mosaic Calcula ion Hype spec al o hopho o mosaics we e calcula ed wi h 10 cm GSD om he FPI images using he FGI’s in-house mosaicking so wa e. Mosaics we e calcula ed o each band sepa a ely and hen combined o o m a uni o m hype spec al da a cube o e each block a ea. The adiome ic p ocessing desc ibed abo e (absolu e calib a ion and ela i e no maliza ions wi hin and be ween image blocks) Remo e Sens. 2017,9, 185 16 o 34 Remo e Sens. 2017, 9, 185 15 o 33 na ow, which educed he accu acy o he de i ed poin s. Howe e , he quali y o geo e e encing could be conside ed good and app op ia e o his in es iga ion. Figu e 5. Compa ison o pho og amme ic and ai bo ne lase scanning (ALS) poin cloud p o iles in es si es 01 and 05. Figu e 6. Nadi (le ) and oblique ( igh ) iews o es si e 02 RGB poin cloud colo ed by heigh . G een indica es lowe heigh and ed indica es highe heigh . Nadi and oblique iews. The esul s o he block adjus men s we e u ilized in he egis a ion o he all he bands o he e e ence band. The egis a ion p ocess was success ul, and only mino misalignmen s could be obse ed be ween he bands in some cases. A isual assessmen o mosaics a e adiome ic p ocessing indica ed ha he adiome y uni o mi y was su icien in mos o he cases (Figu e 7) and sui able e lec ance alues we e ob ained om indi idual image blocks. The mos challenging es si es we e 0304, 05, and 11, which all had a a iable illumina ion. The adiome ic no maliza ion me hod did no p o ide su icien uni o mi y o he es si e 0304, so i was le ou o he analysis. In he o he wo es si es ( 05 and 11), he p ocessing elimina ed he majo adiome ic di e ences. The esul ing spec a a e discussed in Sec ion 3.3. Figu e 6. Nadi ( le ) and oblique ( igh ) iews o es si e 02 RGB poin cloud colo ed by heigh . G een indica es lowe heigh and ed indica es highe heigh . Nadi and oblique iews. Remo e Sens. 2017, 9, 185 16 o 33 Figu e 7. Hype spec al image mosaics o he es si es. The spec al alues ha e been op imized o isual pu poses, and he scaling is no he same o all mosaics. 3.2. Re e ence Da a P ocessing Due o inaccu acies in he ield da a collec ion and he RGB and FPI mosaics, he e we e no able misalignmen s be ween he ield da a ee loca ions and he ees isually obse ed om he mosaics as discussed in Sec ion 2.1. The misalignmen was wo s in he a eas wi h high ee densi y. In some a eas, he misalignmen was sys ema ic and he da a could be aligned jus by mo ing all he e e ence ees in ha speci ic a ea o some di ec ion. Howe e , in some a eas, he misalignmen was no sys ema ic, he e o e e e y ee had o be mo ed one-by-one o he co ec posi ion i he co ec posi ion and species o he ee could be iden i ied om he o homosaic. Some es si es included se e al ee species a di e en canopy heigh s. In such a eas, only he highes ees could be iden i ied. I e e ence da a had o be mo ed, pa icula a en ion was gi en o mo ing he e e ence da a o he ma ching species. In mos cases, i was possible o iden i y he ee species om he RGB o homosaic. Howe e , a some a eas wi h high ee densi y and mul iple ee species, i was no possible o iden i y he co ec ees, and hus he e e ence da a had o be omi ed. Re e ence da a o he es si e 0304 could no be aligned p ope ly, and hus i was omi ed comple ely om he analysis. Al hough i was possible o ix he loca ions o he misaligned e e ence ees o ma ch he co ec ee species in he mosaics, i was no always possible o be ce ain ha he e e ence ee was Figu e 7. Hype spec al image mosaics o he es si es. The spec al alues ha e been op imized o isual pu poses, and he scaling is no he same o all mosaics. Remo e Sens. 2017,9, 185 17 o 34 3.2. Re e ence Da a P ocessing Due o inaccu acies in he ield da a collec ion and he RGB and FPI mosaics, he e we e no able misalignmen s be ween he ield da a ee loca ions and he ees isually obse ed om he mosaics as discussed in Sec ion 2.1. The misalignmen was wo s in he a eas wi h high ee densi y. In some a eas, he misalignmen was sys ema ic and he da a could be aligned jus by mo ing all he e e ence ees in ha speci ic a ea o some di ec ion. Howe e , in some a eas, he misalignmen was no sys ema ic, he e o e e e y ee had o be mo ed one-by-one o he co ec posi ion i he co ec posi ion and species o he ee could be iden i ied om he o homosaic. Some es si es included se e al ee species a di e en canopy heigh s. In such a eas, only he highes ees could be iden i ied. I e e ence da a had o be mo ed, pa icula a en ion was gi en o mo ing he e e ence da a o he ma ching species. In mos cases, i was possible o iden i y he ee species om he RGB o homosaic. Howe e , a some a eas wi h high ee densi y and mul iple ee species, i was no possible o iden i y he co ec ees, and hus he e e ence da a had o be omi ed. Re e ence da a o he es si e 0304 could no be aligned p ope ly, and hus i was omi ed comple ely om he analysis. Al hough i was possible o ix he loca ions o he misaligned e e ence ees o ma ch he co ec ee species in he mosaics, i was no always possible o be ce ain ha he e e ence ee was exac ly he co ec ee in he o homosaics. This means ha o he e e ence da a (such as heigh and DBH) migh be inco ec , and hus could no be u ilized in his s udy. O e hal o he o iginal da a (app oxima ely 8800 ees) had o be omi ed om he analysis, since i could no be aligned o co ec ee species eliably. The inal e e ence da a included 4151 ees om nine di e en es si es. The p opo ion o di e en ee species and hei dis ibu ion o di e en es si es a e summa ized in Table 9. Table 9. The inal e e ence ee da a. Re e ence da a o he es si e 0304 could no be aligned p ope ly, and hus i was omi ed om he analysis. Tes Si e Pine Sp uce Bi ch La ch To al Numbe o Re e ence T ees 01 1769 116 0 0 1885 02 0 24 525 0 549 05 540 80 25 0 645 06 1 179 0 0 180 07 0 102 2 0 104 08 62 40 2 0 104 09 114 259 25 0 398 10 141 0 1 0 142 11 0 22 0 122 144 To al numbe o e e ence ees 2627 822 580 122 4151 3.3. Fea u e Selec ion The mos signi ican ea u es acco ding o he ea u e selec ion me hods a e summa ized in Table 10. The i s se en ea u es o he ull ea u e se wi h no malizing spec al ea u es (i.e., MeanSpec aNo malized and Con inuumRemo edSpec a) we e always included in he inal ea u e se ega dless o he me hod used. The las i een ea u es we e among he bes p edic ing ea u es using wo o he h ee me hods. I he no malizing spec al ea u es we e omi ed om he analysis, he ea u es included 14 ea u es, om which he i s h ee we e selec ed by each ea u e selec ion me hod. Mos o he signi ican ea u es we e spec al ea u es. Examples o he mean and median spec a and he e ec o no maliza ion can be seen in Figu es 8–10. When conside ing he a e age spec a o species calcula ed om all he es si es, he di e ences be ween he sp uce and pine we e he smalles . The di e ences in he bi ch and la ch we e la ge han he o he wo (Figu e 8). The mean and median spec a a e almos equal which indica es ha he species-speci ic spec al alues a e qui e uni o mly dis ibu ed and he e a e no any signi ican ou lie s. When compa ing he spec al signa u es o he species in di e en es si es, he di e ences we e less han 25% in mos cases (Figu e 9). These Remo e Sens. 2017,9, 185 18 o 34 di e ences we e caused by he unce ain ies in he ela i e calib a ion p ocedu e, bu hey we e pa ially also due o he na u al a iabili y wi hin each species. The sp uce spec um o he es si e 11 is clea ly b igh e han wi h he o he es si es. The no maliza ion sligh ly educed he di e ences bu no able di e ences we e s ill p esen a e no maliza ion. The es si e 11 had changes in he illumina ion du ing he da a cap u e and he adiome ic co ec ion was no able o emo e all he illumina ion changes which a ec ed he adiome ic quali y o he inal da a. Table 10. The bes pe o ming ea u es selec ed by he ea u e selec ion me hods, wi h and wi hou no malized spec al ea u es. Fea u e Selec ion wi h No malized Spec al Fea u es Fea u e Selec ion wi hou No malized Spec al Fea u es MeanSpec aNo malized_515nm MeanSpec a_820 MeanSpec aNo malized_529nm MedianSpec a_711 MeanSpec aNo malized_606nm MedianSpec a6Max_820 MeanSpec aNo malized_657nm MeanSpec a_806 Con inummRemo edSpec a_764nm MeanSpec aDa k_657 b90 MedianSpec aDa k_657 b95 MeanSpec a6Max_806 MeanSpec a_820nm MeanSpec a6Max_820 MedianSpec a_711nm MedianSpec a6Max_806 MedianSpec aDa k_657nm Min MeanSpec aNo malized_508nm p90 MeanSpec aNo malized_510nm b90 MeanSpec aNo malized_521nm b95 MeanSpec aNo malized_590nm co MeanSpec aNo malized_599nm MeanSpec aNo malized_644nm MeanSpec aNo malized_670nm MeanSpec aNo malized_718nm MeanSpec aNo malized_806nm Con inummRemo edSpec a_644nm Con inummRemo edSpec a_657nm Min Remo e Sens. 2017, 9, 185 18 o 33 Mos o he signi ican ea u es we e spec al ea u es. Examples o he mean and median spec a and he e ec o no maliza ion can be seen in Figu es 8–10. When conside ing he a e age spec a o species calcula ed om all he es si es, he di e ences be ween he sp uce and pine we e he smalles . The di e ences in he bi ch and la ch we e la ge han he o he wo (Figu e 8). The mean and median spec a a e almos equal which indica es ha he species-speci ic spec al alues a e qui e uni o mly dis ibu ed and he e a e no any signi ican ou lie s. When compa ing he spec al signa u es o he species in di e en es si es, he di e ences we e less han 25% in mos cases (Figu e 9). These di e ences we e caused by he unce ain ies in he ela i e calib a ion p ocedu e, bu hey we e pa ially also due o he na u al a iabili y wi hin each species. The sp uce spec um o he es si e 11 is clea ly b igh e han wi h he o he es si es. The no maliza ion sligh ly educed he di e ences bu no able di e ences we e s ill p esen a e no maliza ion. The es si e 11 had changes in he illumina ion du ing he da a cap u e and he adiome ic co ec ion was no able o emo e all he illumina ion changes which a ec ed he adiome ic quali y o he inal da a. Figu e 8. Mean, Median and MeanNo malized spec a o di e en species a e aged o e all es si es ( op ow) and he same spec a no malized o Pine spec a (bo om ow). Tes si es wi h less han 10 ee samples ha e been omi ed om he plo . The e o ba s p esen he s anda d e o . When he ull ea u e se was used, mos o he signi ican spec al ea u es we e no malized spec al ea u es. No malizing he spec a educes he e ec s o shadowing and illumina ion di e ences, which is p obably he eason why hey pe o med bes wi h his da a, since some o he da a had been collec ed in di e en ligh s and ime o day wi h a ying illumina ion condi ions. Illumina ion and shadowing di e ences a e educed, especially when he illumina ion condi ions a e specula , bu i also educes he possible scale di e ences wi h da a collec ed on di e en occasions in di use illumina ion condi ions. The ela i e spec al di e ences we e educed, especially a he nea in a- ed wa eleng hs whe e changes in iew angle and ee s uc u e ha e he s onges impac . A he same ime, he di e ences a isible wa eleng hs a e s ill p esen , and hus mos o he bes ea u es a e a isible wa eleng hs, bu ew NIR ea u es a e also p esen . Figu e 8. Mean, Median and MeanNo malized spec a o di e en species a e aged o e all es si es ( op ow) and he same spec a no malized o Pine spec a (bo om ow). Tes si es wi h less han 10 ee samples ha e been omi ed om he plo . The e o ba s p esen he s anda d e o . Remo e Sens. 2017,9, 185 19 o 34 Remo e Sens. 2017, 9, 185 19 o 33 Figu e 9. Mean spec a o di e en species a di e en es si es ( op ow) and hei ela i e di e ences (bo om ow). In he ela i e di e ences, Pine spec a ha e been no malized o es si e 01, Sp uce o 06, and Bi ch o 02. The Pine and Bi ch spec a no malized o a eas which had mos ee samples. Sp uces we e no malized o a ea 06 which was he spec al e e ence a ea in his s udy. The e o ba s p esen he s anda d e o . Figu e 10. No malized mean spec a o di e en species a di e en es si es ( op ow) and hei ela i e di e ences (bo om ow). In he ela i e di e ences, Pine spec a ha e been no malized o es si e 01, Sp uce o 06 and Bi ch o 02. The e o ba s p esen he s anda d e o . Figu e 9. Mean spec a o di e en species a di e en es si es ( op ow) and hei ela i e di e ences (bo om ow). In he ela i e di e ences, Pine spec a ha e been no malized o es si e 01, Sp uce o 06, and Bi ch o 02. The Pine and Bi ch spec a no malized o a eas which had mos ee samples. Sp uces we e no malized o a ea 06 which was he spec al e e ence a ea in his s udy. The e o ba s p esen he s anda d e o . Remo e Sens. 2017, 9, 185 19 o 33 Figu e 9. Mean spec a o di e en species a di e en es si es ( op ow) and hei ela i e di e ences (bo om ow). In he ela i e di e ences, Pine spec a ha e been no malized o es si e 01, Sp uce o 06, and Bi ch o 02. The Pine and Bi ch spec a no malized o a eas which had mos ee samples. Sp uces we e no malized o a ea 06 which was he spec al e e ence a ea in his s udy. The e o ba s p esen he s anda d e o . Figu e 10. No malized mean spec a o di e en species a di e en es si es ( op ow) and hei ela i e di e ences (bo om ow). In he ela i e di e ences, Pine spec a ha e been no malized o es si e 01, Sp uce o 06 and Bi ch o 02. The e o ba s p esen he s anda d e o . Figu e 10. No malized mean spec a o di e en species a di e en es si es ( op ow) and hei ela i e di e ences (bo om ow). In he ela i e di e ences, Pine spec a ha e been no malized o es si e 01, Sp uce o 06 and Bi ch o 02. The e o ba s p esen he s anda d e o . Remo e Sens. 2017,9, 185 20 o 34 When he ull ea u e se was used, mos o he signi ican spec al ea u es we e no malized spec al ea u es. No malizing he spec a educes he e ec s o shadowing and illumina ion di e ences, which is p obably he eason why hey pe o med bes wi h his da a, since some o he da a had been collec ed in di e en ligh s and ime o day wi h a ying illumina ion condi ions. Illumina ion and shadowing di e ences a e educed, especially when he illumina ion condi ions a e specula , bu i also educes he possible scale di e ences wi h da a collec ed on di e en occasions in di use illumina ion condi ions. The ela i e spec al di e ences we e educed, especially a he nea in a- ed wa eleng hs whe e changes in iew angle and ee s uc u e ha e he s onges impac . A he same ime, he di e ences a isible wa eleng hs a e s ill p esen , and hus mos o he bes ea u es a e a isible wa eleng hs, bu ew NIR ea u es a e also p esen . I he no malizing spec al ea u es we e omi ed om he analysis, he inal ea u e se included mo e NIR ea u es and only ea u es in he ed band (i.e., 657 nm) om he isible spec al egion. Wi hou he no maliza ion, he ela i e di e ences in he di e en species a e highe in he NIR egion and smalle in he isible wa eleng h egion, and hus he NIR ea u es wi hou no maliza ion we e he bes p edic o s. In addi ion, a ew mo e 3D s uc u al ea u es we e included in he inal ea u e se wi h he ea u e se wi hou no malizing spec al ea u es. The mos signi ican 3D s uc u al ea u es we e mos ly hose ea u es a he canopy op laye s, which was an expec ed esul since passi e op ical imaging does no ha e good pene a ion abili y. The minimum heigh no malized by he maximum heigh pe o med well, as i p o ides a good es ima e o he canopy ex en downwa ds, which p o ides species-speci ic in o ma ion ( o example, be ween pines and sp uces), since he canopy o sp uces ex ends lowe han pines and bi ches. Howe e , i mus be no ed ha wi h pho og amme ically-p oduced poin clouds, he pene a ion abili y is also highly dependen on he densi y o he o es (gaps be ween ees), which is no always a species-speci ic p ope y. 3.4. Classi ica ion The classi ica ion was pe o med using a ious ea u e con igu a ions, including all ea u es, all ea u es wi hou no malizing spec al ea u es and mean spec a only. The e alua ion o di e en classi ica ion me hods o ea u e-selec ed da ase s and mean spec a only a e summa ized in Table 11. Table 11. Classi ica ion accu acies and Kappa alues o es ed classi ica ion me hods using ea u e-selec ed da ase s and mean spec a only. Algo i hm Fea u e Selec ed-All Fea u es Fea u e Selec ed-wi hou No malizing Spec al Fea u es Mean Spec a Only Acc. Kappa Acc. Kappa Acc. Kappa C4.5 91.4 0.84 90.7 0.83 89.5 0.80 RandomFo es 94.9 0.90 92.6 0.86 93.3 0.87 k-NN k = 1 93.1 0.87 89.5 0.80 92.3 0.86 k = 3 93.8 0.88 90.8 0.82 93.5 0.88 k = 7 93.3 0.87 91.2 0.83 93.0 0.87 MLP 95.2 0.91 92.4 0.86 95.4 0.91 Nai e Bayes 87.1 0.77 83.3 0.70 70.7 0.49 Abb e ia ions: k-NN, k-Nea es Neighbo s; MLP, Mul ilaye Pe cep on. The bes pe o ming classi ica ion algo i hms we e MLP and RandomFo es , which p o ided 94.9% and 95.2% o e all accu acies o ea u e-selec ed da ase wi h all ea u es, espec i ely. None o he classi ica ion algo i hms pe o med excep ionally poo ly in he classi ica ion, bu he Nai eBayes me hod did no p o ide good esul s (87.1%). The Nai eBayes me hod was p obably mos a ec ed by he imbalance da a, which esul ed in poo e esul s. The classi ica ion esul s wi h ea u e selec ion and using a ea u e se wi hou he no malizing spec al ea u es p oduced sligh ly wo se classi ica ion Remo e Sens. 2017,9, 185 21 o 34 esul s compa ed o hose achie ed wi h he no malizing ea u es. Wi hou he no malizing spec al ea u es, Random Fo es p o ided he bes esul wi h 93% o e all accu acy. The classi ica ion was also es ed using only he mean spec a. The esul s indica e ha using only he mean spec a, good classi ica ion esul s can be achie ed. The accu acy o classi ica ion using only he mean spec a and MLP me hod was 95.2%. The accu acy o he o he classi ica ion me hods di e ed by app oxima ely 1% om hose achie ed wi h he ull ea u e se wi h ea u e selec ion, excep he Nai eBayes me hod which p oduced only 70.7% accu acy. The e alua ion o classi ica ion pe o mance wi h di e en ea u e se s using k-NN and RandomFo es classi ie s a e p esen ed in Tables 12 and 13. The bes pe o mance was achie ed when using all 347 ea u es. The accu acies we e 95.5% o k-NN and 95.1% o RandomFo es . When all ea u es we e used, he ea u e selec ion did no imp o e he classi ica ion accu acy, bu nei he did i d as ically wo sen he esul s. Fea u e educ ion based on he ea u e selec ion signi ican ly imp o ed he lea ning ime o he classi ica ion, especially wi h mo e complex me hods such as MLP. Good pe o mance was also achie ed by using only he spec al ea u es (94.7% and 94.9% o Random Fo es and k-NN, espec i ely). Random Fo es and k-NN (k = 3) achie ed 94.8% and 94.9% accu acies, espec i ely, when all he ea u es we e used wi h he ea u e se wi hou no malizing spec al ea u es, which was less han 1% wo se han wi h he ull ea u e se . The wo s esul s we e ob ained when only he s uc u al ea u es we e used in classi ica ion. The Kappa alues a e sligh ly smalle han he accu acy, which indica es a small e ec o he imbalanced da a. Table 12. E ec o di e en ea u e se s o classi ica ion using he k-NN me hod wi h k = 3. Me ic Name All Fea u es Spec al Fea u es Fea u e Selec ion All Fea u es (No No m. Spec a) Spec al Fea u es (No No m. Spec a) Fea u e Selec ion (No No m. Spec a) S uc u al Fea u es Accu acy (%) 95.5 94.9 93.8 94.9 94.7 90.8 68.4 Kappa 0.92 0.90 0.88 0.90 0.90 0.82 0.37 Table 13. E ec o di e en ea u e se s o classi ica ion using RandomFo es . Me ic Name All Fea u es Spec al Fea u es Fea u e Selec ion All Fea u es (No No m. Spec a) Spec al Fea u es (No No m. Spec a) Fea u e Selec ion (No No m. Spec a) S uc u al Fea u es Accu acy (%) 95.1 94.7 94.9 94.8 94.3 92.6 72.0 Kappa 0.91 0.90 0.90 0.90 0.89 0.86 0.39 The species-speci ic esul s o he k-NN, RandomFo es , and MLP me hods wi h a ea u e-selec ed ull ea u e se a e p esen ed in con usion ma ices in Tables 14–17. As expec ed, he mos e o s in classi ica ion occu ed among he classi ica ion o pines and sp uces. The mos equen e o was sp uces classi ied as pines, which migh be a esul o he imbalance in he da a (mos o he aining da a a e pines). Bi ches could be classi ied wi h a high ecall and p ecision (bo h o e 95%) since hey ha e a s ong spec al di e ence compa ed o coni e ous species, especially in he NIR channels (Figu e 8). When using no malized ea u es, he bes balance be ween ecall and p ecision is ob ained wi h he RandomFo es and MLP which bo h had an F-sco e o 0.93. The con usion ma ix o RandomFo es using a ea u e se wi hou no malizing spec al ea u es (Table 17) and ea u e selec ion indica es ha mos e o s in his case occu due o he poo ue posi i e classi ica ion a e o La ch ees. Remo e Sens. 2017,9, 185 22 o 34 Table 14. Con usion ma ix o classi ica ion using he k-NN me hod wi h he k = 3 wi h a ea u e-selec ed da a se . F-sco e p esen s he mean F-sco e o all species. F-Sco e: 0.91 Classi ied as Recall Pine Sp uce Bi ch La ch T ue Class Pine 2583 43 0 1 0.983 Sp uce 152 660 3 7 0.803 Bi ch 13 9 553 5 0.953 La ch 17 4 3 98 0.803 P ecision 0.934 0.922 0.989 0.883 Table 15. Con usion ma ix o classi ica ion using RandomFo es wi h a ea u e-selec ed da a se selec ed om all ea u es. F-sco e p esen s he mean F-sco e o all species. F-Sco e: 0.93 Classi ied as Recall Pine Sp uce Bi ch La ch T ue Class Pine 2584 41 0 2 0.984 Sp uce 122 692 2 6 0.842 Bi ch 13 8 558 1 0.962 La ch 11 1 5 105 0.861 P ecision 0.947 0.933 0.988 0.921 Table 16. Con usion ma ix o classi ica ion using MLP wi h a ea u e-selec ed da a se selec ed om all ea u es. F-sco e p esen s he mean F-sco e o all species. F-Sco e: 0.93 Classi ied as Recall Pine Sp uce Bi ch La ch T ue Class Pine 2564 57 1 5 0.976 Sp uce 89 718 5 10 0.873 Bi ch 11 5 562 2 0.969 La ch 6 5 5 106 0.869 P ecision 0.960 0.915 0.981 0.862 Table 17. Con usion ma ix o classi ica ion using RandomFo es wi h a ea u e-selec ed da a se wi hou no malized spec al ea u es. F-sco e p esen s he mean F-sco e o all species. F-Sco e: 0.85 Classi ied as Recall Pine Sp uce Bi ch La ch T ue Class Pine 2555 67 0 5 0.973 Sp uce 130 680 9 3 0.827 Bi ch 8 21 535 16 0.922 La ch 17 12 20 73 0.598 P ecision 0.943 0.872 0.949 0.753 3.5. Indi idual T ee De ec ion and Species P edic ion The indi idual ee de ec ion was e alua ed isually and by using he e e ence da a. A isual examina ion o he loca ed ees and o homosaics indica ed ha mos o he ees in he es si es could be ound. The mos no able a eas whe e he posi ion o he de ec ed ees and he ees isually obse ed om he o homosaic di e ed we e a eas whe e he poin clouds we e incomple e due o ma ching ailu es o occlusions ( o example, due o s ong shadows o insu icien o e laps), and hus had local inaccu acies in he 3D poin cloud and CHM. The mos no able unde ec ed ees we e sho e ees a lowe le els o he canopy, which could be seen om he o homosaics bu could no be Remo e Sens. 2017,9, 185 23 o 34 de ec ed om he poin clouds due o he limi ed abili y o passi e senso s o pene a e he canopy. Some o he e o s migh ha e been induced by he ac ha he e e ence ees we e manually placed wi h he help o o homosaics. The ee op peak in he poin clouds migh ha e been a a di e en loca ion han he ee cen e isually de ec ed om he o homosaics. The ee de ec ion in he es si es was es ed using he e e ence da a wi h h ee di e en sea ch adiuses (Table 18). Inc easing he sea ch adius had a signi ican impac on he de ec ion a e in some o he es si es, which indica es ha he e was a no able di e ence be ween he ee c own cen e s de e mined om image mosaics and au oma ic de ec ion om CHM. The bes esul s wi h 64%–96% ee iden i ica ion a es we e ob ained wi h he 2 m sea ch adius. Howe e , due o he lack o e e ence da a o es ing, he accu acy es ima ion esul s o he ee de ec ion a e only indica i e. Table 18. Pe cen age o e e ence ees ound using he indi idual ee de ec ion me hod, and he o al numbe o ees de ec ed in each es si e. Tes Si e Sea ch Radius To al No. o T ees De ec ed 1 m 1.5 m 2 m 01 0.82 0.88 0.89 9723 02 0.88 0.95 0.96 5223 05 0.63 0.78 0.83 5942 06 0.80 0.82 0.82 1585 07 0.64 0.82 0.88 2023 08 0.26 0.46 0.64 1884 09 0.63 0.72 0.77 3191 10 0.61 0.79 0.87 1242 11 0.53 0.71 0.76 1611 Visual inspec ion o he classi ica ion o he indi idual ees indica ed ha mos o he ees we e classi ied co ec ly. Mos e o s occu ed be ween pines and sp uces and in a eas ha had a lo o shadowing ( o example in a ea 08, which was measu ed unde comple ely sunny skies). The species-speci ic and a ea-speci ic p edic ion p obabili ies p esen ed in Table 19 indica e ha he ained MLP classi ie could classi y mos o he de ec ed ees wi h a high p obabili y (90%–97%). The bes p edic ion could be p o ided o pines (97%), and hus also in es si es wi h many pines. The p obabili y o he p edic ion o la ch (90%) was he lowes , which also causes he 11 es si e o ha e lowes mean p obabili ies since i has many la ch ees. Figu e 11 p esen s an example o es si e 05 wi h de ec ed and classi ied ees. Table 19. Species- and a ea-speci ic p edic ion p obabili ies using he MLP model ained wi h a ea u e-selec ed da a se selec ed om all ea u es. N indica es he numbe o ees classi ied o a speci ic species a a speci ic a ea; Mean is he mean p edic ion p obabili y; S d. is he s anda d de ia ion. T ee Species 01 02 05 06 07 08 09 10 11 All Tes Si es Pine N 6102 593 3580 246 137 382 1029 579 415 13063 Mean 0.98 0.95 0.97 0.95 0.94 0.94 0.96 0.96 0.95 0.97 S d. 0.07 0.10 0.08 0.11 0.12 0.13 0.11 0.11 0.12 0.09 Sp uce N 2634 2376 1886 876 1623 1105 1789 308 780 13377 Mean 0.96 0.97 0.93 0.97 0.98 0.96 0.97 0.94 0.95 0.96 S d. 0.11 0.08 0.12 0.10 0.06 0.09 0.09 0.11 0.11 0.10 Bi ch N951 2237 421 451 255 246 365 325 154 5405 Mean 0.94 0.98 0.93 0.95 0.91 0.94 0.92 0.93 0.87 0.95 S d. 0.14 0.08 0.14 0.13 0.15 0.12 0.15 0.15 0.14 0.12 La ch N36 17 55 12 8 151 8 30 262 579 Mean 0.82 0.84 0.87 0.84 0.81 0.91 0.86 0.89 0.92 0.90 S d. 0.17 0.17 0.16 0.20 0.21 0.16 0.22 0.17 0.15 0.16 All species N 9723 5223 5942 1585 2023 1884 3191 1242 1611 32424 Mean 0.97 0.97 0.95 0.96 0.97 0.95 0.96 0.95 0.93 0.96 S d. 0.09 0.08 0.11 0.11 0.09 0.11 0.10 0.13 0.13 0.10 Remo e Sens. 2017,9, 185 24 o 34 Remo e Sens. 2017, 9, 185 23 o 33 Visual inspec ion o he classi ica ion o he indi idual ees indica ed ha mos o he ees we e classi ied co ec ly. Mos e o s occu ed be ween pines and sp uces and in a eas ha had a lo o shadowing ( o example in a ea 08, which was measu ed unde comple ely sunny skies). The species-speci ic and a ea-speci ic p edic ion p obabili ies p esen ed in Table 19 indica e ha he ained MLP classi ie could classi y mos o he de ec ed ees wi h a high p obabili y (90%–97%). The bes p edic ion could be p o ided o pines (97%), and hus also in es si es wi h many pines. The p obabili y o he p edic ion o la ch (90%) was he lowes , which also causes he 11 es si e o ha e lowes mean p obabili ies since i has many la ch ees. Figu e 11 p esen s an example o es si e 05 wi h de ec ed and classi ied ees. Figu e 11. Example o de ec ed and classi ied ees in he 05 es si e. Blue s a s a e bi ches, yellow iangles a e sp uces, g een ci cles a e pines, and ed squa es a e la ches. Figu e 11. Example o de ec ed and classi ied ees in he 05 es si e. Blue s a s a e bi ches, yellow iangles a e sp uces, g een ci cles a e pines, and ed squa es a e la ches. 3.6. P ocessing Times In he ollowing, we p esen es ima es o he compu ing imes needed o p ocess he da ase s. The gi en ime-es ima es a e compu ing imes o he esea ch sys em and non-op imized compu e s. In addi ion, he p ocess included some in e ac i e s eps, such as measu ing GCPs and e lec ance panels in images, as well as in e ac i e quali y con ol; hese s eps a e expec ed o impac minimally on he compu a ion imes when au oma ed in he u u e. Mos o he p ocessing ime was spen on compu ing he image o ien a ion and calcula ing he pho og amme ic poin clouds which ook a he minimum 3.5 h and a he maximum 11.5 h o es si es 02 and 0304, espec i ely. The band-wise ex e io o ien a ion calcula ion ook app oxima ely 5 s o he indi idual band and image, esul ing in p ocessing imes o 1.5 and 6 h o he smalles and la ges blocks, espec i ely. The adiome ic block adjus men and mosaic calcula ion ook app oxima ely 3–6 min pe hype cube depending on he pa ame e s and size o he a ea, hus esul ing in 2 and 13 h o he smalles and la ges blocks, espec i ely. These numbe s a e a ec ed by many ac o s, such as he quali y o app oxima e alues, selec ed quali y le els and complexi y o he a ea. Remo e Sens. 2017,9, 185 25 o 34 The spec al and s uc u al ea u e ex ac ion ook app oxima ely 2.3 h and 10 h o all es si es, espec i ely. The used ea u e selec ion me hods ook app oxima ely 5 min. The classi ica ion model lea ning and e alua ion had much a iabili y. The as es classi ica ion me hod using he ea u e selec ed da ase was he k-NN me hod which could be ained and e alua ed using he LOOCV me hod in less han one minu e. Fo RandomFo es and MLP me hods, i ook 3.6 h and 12.3 h, espec i ely. The comple e indi idual ee de ec ion wo k low (i.e., om poin cloud o ee loca ions) ook app oxima ely 40 min o all es si es and he ee species p edic ion o he de ec ed ees using he ained MLP model ook app oxima ely 5 s o all ees. 4. Discussion This in es iga ion de eloped a small UAV hype spec al imaging and a pho og amme y-based emo e sensing me hod o indi idual ee de ec ion and classi ica ion in a bo eal o es . The me hod was assessed using, al oge he , ele en o es es si es wi h 4151 e e ence ees. 4.1. Aspec s o Da a P ocessing Da ase s we e cap u ed in deep o es scene, and condi ions du ing he da a cap u e we e ypical o he no he n clima e zone du ing summe wi h a iable cloudiness and ain showe s. Va iable condi ions a e challenging o analysis me hods based on passi e spec al in o ma ion. The con en ional adiome ic co ec ion me hods based on e lec ance panels o adia i e ans e modeling [ 42 – 46 ] a e no sui able in such condi ions. The no el adiome ic p ocessing app oach based on adiome ic block adjus men and onboa d i adiance measu emen s p o ided p omising esul s. The e lec ance panels could no be u ilized in se e al es si es in his s udy because hey we e oo deep in he o es and shadowed by he ees. In many ope a ional applica ions, he ins alla ion o e lec ance panels is likely o be challenging; hus, i is impo an o de elop me hods ha a e independen o hose. Impo an u u e imp o emen s will include he enhancemen o he i adiance spec ome e o be in ole an o pla o m il ing as well as imp o ing he sys em calib a ion. The UAV sys em used in his s udy was equipped wi h he i adiance spec ome e bu he da a quali y was no su icien o pe o ming accu a e enough e lec ance co ec ion. Fu u e enhancemen s a e expec ed o imp o e he accu acy and le el o au oma ion and ul ima ely emo ing he need o g ound e lec ance panels. App oaches using objec cha ac e is ics could also be sui able. Fo example, one possible me hod would be o use e lec ance in o ma ion o known na u al a ge s in he scenes (such as ees) o pe o m adiome ic no maliza ion be ween di e en a eas whose da a has been collec ed in di e en ligh s. An impo an opic o u u e s udies will be how di e en illumina ion condi ions (di ec sun illumina ion and di e en shadow dep h) should be accoun ed o in he da a analysis in a o es ed en i onmen and which pa s o i should be compensa ed o . Fo example, Ko pela e al. [ 15 ] classi ied spec al ee ea u es based on he illumina ion condi ion de i ed om he 3D objec model. The p ocessing and analysis we e highly au oma ed. The majo in e ac i e p ocesses included he deploymen o geome ic and adiome ic in si u a ge s and he ield e e ence da a cap u e. We expec ha wi h u u e imp o emen s in di ec geo e e encing pe o mance and sys em adiome ic calib a ion, he da a p ocessing chain could be ully au oma ed. Fu he conside a ions a e needed o au oma e he ee species e e ence da a. Fo example, spec al lib a ies [ 68 ] could p o ide a good e e ence i he da a we e p ope ly calib a ed. Wi h his da a se , we also used in e ac i e analysis in some phases o he classi ica ion, bu he e a e possibili ies o au oma e hese s eps. A lo o di e en so wa e including comme cial, open and in-house was chained and much o he so wa e was no op imized o he ma e ial used in his s udy o wi h espec o speed. The compu ing imes in pho og amme ic and adiome ic p ocessing we e app oxima ely 8 h o 32 h depending on he size o he block; he combined ea u e ex ac ion, selec ion and classi ica ion ook app oxima ely 25 h wi h he MLP me hod, which was he slowes me hod. All he ope a ions can be highly au oma ed, op imized and pa allelized a e su icien in o ma ion abou he measu emen Remo e Sens. 2017,9, 185 32 o 34 23. Yu, X.; Li key, P.; Hyyppä, J.; Holopainen, M.; Vas a an a, M. Assessmen o Low Densi y Full-Wa e o m Ai bo ne Lase Scanning o Indi idual T ee De ec ion and T ee Species Classi ica ion. Fo es s 2014 ,5, 1011–1031. [C ossRe ] 24. Gougeon, F.A.; Moo e, T. Classi ica ion Indi iduelle des A b es àPa i d’images àHau e Résolu ion Spa iale. In P oceedings o he 6 h Cong ess o Associa ion québécoise de élédé ec ion, She b ooke, QC, Canada, 4–6 May 1989. 25. B and be g, T.; Wal e , F. Au oma ed delinea ion o indi idual ee c owns in high spa ial esolu ion ae ial images by mul iple-scale analysis. Mach. Vis. Appl. 1998,11, 64–73. [C ossRe ] 26. Wang, L.; Gong, P.; Biging, G.S. Indi idual ee-c own delinea ion and ee op de ec ion in high-spa ial- esolu ion ae ial image y. Pho og amm. Eng. Remo e Sens. 2004,70, 351–357. [C ossRe ] 27. S -Onge, B.; Vega, C.; Fou nie , R.A.; Hu, Y. Mapping canopy heigh using a combina ion o digi al s e eo-pho og amme y and lida . In . J. Remo e Sens. 2008,29, 3343–3364. [C ossRe ] 28. 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 ,2010, 99–115. [C ossRe ] [PubMed] 29. Bal sa ias, E.; G uen, A.; Eisenbeiss, H.; Zhang, L.; Wase , L.T. High-quali y image ma ching and au oma ed gene a ion o 3D ee models. In . J. Remo e Sens. 2008,29, 1243–1259. [C ossRe ] 30. Yu, X.; Hyyppä, J.; Ka jalainen, M.; Nu minen, K.; Ka ila, K.; Vas a an a, M.; Kanka e, V.; Kaa inen, H.; Holopainen, M.; Honka aa a, E. Compa ison o lase and s e eo op ical, SAR and InSAR poin clouds om ai -and space-bo ne sou ces in he e ie al o o es in en o y a ibu es. Remo e Sens. 2015 ,7, 15933–15954. [C ossRe ] 31. S -Onge, B.A. Es ima ing indi idual ee heigh s o he bo eal o es using ai bo ne lase al ime y and digi al ideog aphy. In . A ch. Pho og amm. Remo e Sens. Spa . In . Sci. 1999,XXXII-3/W14, 179–185. 32. Hyyppä, J.; Inkinen, M. De ec ing and es ima ing a ibu es o single ees using lase scanne . Pho og amm. J. Finl. 1999,16, 27–42. 33. Colomina, I.; Molina, P. Unmanned ae ial sys ems o pho og amme y and emo e sensing: A e iew. ISPRS J. Pho og amm. Remo e Sens. 2014,92, 79–97. [C ossRe ] 34. To esan, C.; Be on, A.; Ca o enu o, F.; Di Genna o, S.F.; Gioli, B.; Ma ese, A.; Miglie a, F.; Vagnoli, C.; Zaldei, A.; Wallace, L. Fo es y applica ions o UAVs in Eu ope: A e iew. In . J. Remo e Sens. 2016, 1–21. 35. Jaakkola, A.; Hyyppä, J.; Kukko, A.; Yu, X.; Kaa inen, H.; Leh omäki, M.; Lin, Y. A low-cos mul i-senso al mobile mapping sys em and i s easibili y o ee measu emen s. ISPRS J. Pho og amm. Remo e Sens. 2010 , 65, 514–522. [C ossRe ] 36. Lisein, J.; Pie o -Deseilligny, M.; Bonne , S.; Lejeune, P. A pho og amme ic wo k low o he c ea ion o a o es canopy heigh model om small unmanned ae ial sys em image y. Fo es s 2013 ,4, 922–944. [C ossRe ] 37. Puli i, S.; Ø ka, H.O.; Gobakken, T.; Næsse , E. In en o y o small o es a eas using an unmanned ae ial sys em. Remo e Sens. 2015,7, 9632–9654. [C ossRe ] 38. Za co-Tejada, P.J.; Diaz-Va ela, R.; Angile i, V.; Loudjani, P. T ee heigh quan i ica ion using e y high esolu ion image y acqui ed om an unmanned ae ial ehicle (UAV) and au oma ic 3D pho o- econs uc ion me hods. Eu . J. Ag on. 2014,55, 89–99. [C ossRe ] 39. Michez, A.; Piégay, H.; Lisein, J.; Claessens, H.; Lejeune, P. Classi ica ion o ipa ian o es species and heal h condi ion using mul i- empo al and hype spa ial image y om unmanned ae ial sys em. En i on. Moni . Assess. 2016,188, 1–19. [C ossRe ] [PubMed] 40. Gini, R.; Passoni, D.; Pin o, L.; Sona, G. Use o Unmanned Ae ial Sys ems o mul ispec al su ey and ee classi ica ion: A es in a pa k a ea o no he n I aly. Eu . J. Remo e Sens. 2014,47, 251–269. [C ossRe ] 41. Lisein, J.; Michez, A.; Claessens, H.; Lejeune, P. Disc imina ion o deciduous ee species om ime se ies o unmanned ae ial sys em image y. PLoS ONE 2015,10, e0141006. [C ossRe ] [PubMed] 42. Za co-Tejada, P.J.; González-Dugo, V.; Be ni, J.A.J. Fluo escence, empe a u e and na ow-band indices acqui ed om a UAV pla o m o wa e s ess de ec ion using a mic o-hype spec al image and a he mal came a. Remo e Sens. En i on. 2012,117, 322–337. [C ossRe ] 43. H uska, R.; Mi chell, J.; Ande son, M.; Glenn, N.F. Radiome ic and geome ic analysis o hype spec al image y acqui ed om an unmanned ae ial ehicle. Remo e Sens. 2012,4, 2736–2752. [C ossRe ] 44. Bü ne , A.; Röse , H.-P. Hype spec al Remo e Sensing wi h he UAS “S u ga e Adle ”—Sys em Se up, Calib a ion and Fi s Resul s. Pho og amm. Fe ne kund. Geoin . 2014,2014, 265–274. [C ossRe ] [PubMed] Remo e Sens. 2017,9, 185 33 o 34 45. Suomalainen, J.; Ande s, N.; Iqbal, S.; Roe ink, G.; F anke, J.; Wen ing, P.; Hünnige , D.; Ba holomeus, H.; Becke , R.; Koois a, L. A ligh weigh hype spec al mapping sys em and pho og amme ic p ocessing chain o unmanned ae ial ehicles. Remo e Sens. 2014,6, 11013–11030. [C ossRe ] 46. Luciee , A.; Maleno ský, Z.; Veness, T.; Wallace, L. Hype UAS—Imaging spec oscopy om a mul i o o unmanned ai c a sys em. J. Field Robo . 2014,31, 571–590. [C ossRe ] 47. Mäkynen, J.; Holmlund, C.; Saa i, H.; Ojala, K.; An ila, T. Unmanned ae ial ehicle (UAV) ope a ed megapixel spec al came a. In P oceedings o he SPIE 8186, Elec o-Op ical Remo e Sensing, Pho onic Technologies, and Applica ions, P ague, Czech, 19 Sep embe 2011; Kame man, G.W., S ein all, O., Bishop, G.J., Gonglewski, J.D., Lewis, K.L., Hollins, R.C., Me le , T.J., Eds.; 48. Saa i, H.; Pellikka, I.; Pesonen, L.; Tuominen, S.; Heikkilä, J.; Holmlund, C.; Mäkynen, J.; Ojala, K.; An ila, T. Unmanned Ae ial Vehicle (UAV) ope a ed spec al came a sys em o o es and ag icul u e applica ions. In P oceedings o he SPIE 8174, Remo e Sensing o Ag icul u e, Ecosys ems, and Hyd ology XIII, P ague, Czech, 19 Sep embe 2011; Neale, C.M.U., Mal ese, A., Eds. 49. Honka aa a, E.; Saa i, H.; Kai osoja, J.; Pölönen, I.; Hakala, T.; Li key, P.; Mäkynen, J.; Pesonen, L. P ocessing and Assessmen o Spec ome ic, S e eoscopic Image y Collec ed Using a Ligh weigh UAV Spec al Came a o P ecision Ag icul u e. Remo e Sens. 2013,5, 5006–5039. [C ossRe ] 50. Näsi, R.; Honka aa a, E.; Lyy ikäinen-Saa enmaa, P.; Blomq is , M.; Li key, P.; Hakala, T.; Viljanen, N.; Kan ola, T.; Tanhuanpää, T.; Holopainen, M. Using UAV-based pho og amme y and hype spec al imaging o mapping ba k bee le damage a ee-le el. Remo e Sens. 2015,7, 15467–15493. [C ossRe ] 51. Honka aa a, E.; Ma kelin, L.; Hakala, T.; Pel oniemi, J. The me ology o di ec ional, spec al e lec ance ac o measu emen s based on a ea o ma imaging by UAVs. Pho og amm. Fe ne kund. Geoin . 2014 ,2014, 175–188. 52. Pel oniemi, J.I.; Hakala, T.; Suomalainen, J.; Honka aa a, E.; Ma kelin, L.; G i se ich, M.; Eskelinen, J.; Jaanson, P.; Ikonen, E. Technical no es: A de ailed s udy o he p o ision o measu emen unce ain y and aceabili y o goniospec ome e s. J. Quan . Spec osc. Radia . T ans . 2014,146, 376–390. [C ossRe ] 53. Hakala, T.; Honka aa a, E.; Saa i, H.; Mäkynen, J.; Kai osoja, J.; Pesonen, L.; Pölönen, I. Spec al Imaging F om Ua s Unde Va ying Illumina ion Condi ions. In . A ch. Pho og amm. Remo e Sens. Spa . In . Sci. 2013 , XL-1/W2, 189–194. [C ossRe ] 54. NLS Na ional Land Su ey o Finland Open Da a License. A ailable online: h p://www.maanmi auslai os. i/en/openda a/acquisi ion (accessed on 8 Decembe 2016). 55. El ne , A.; Schneide , D. Analysis o Di e en Me hods o 3D Recons uc ion o Na u al Su aces om Pa allel-Axes UAV Images. Pho og amm. Rec. 2015,30, 279–299. [C ossRe ] 56. Honka aa a, E.; Hakala, T.; Ma kelin, L.; Rosnell, T.; Saa i, H.; Mäkynen, J. A p ocess o adiome ic co ec ion o UAV image blocks. Pho og amm. Fe ne kund. Geoin . 2012,2012, 115–127. [C ossRe ] 57. Cla k, R.N.; Roush, T.L. Re lec ance spec oscopy: Quan i a i e analysis echniques o emo e sensing applica ions. J. Geophys. Res. 1984,89, 6329–6340. [C ossRe ] 58. Io dache, M.-D. Ma lab Code and Demo o Con inuum Remo al. A ailable online: h ps://www. esea chga e.ne /publica ion/301289820_Ma lab_Code_and_Demo_ o _Con inuum_Remo al (accessed on 8 Decembe 2016). 59. Pea son, K. Con ibu ions o he Ma hema ical Theo y o E olu ion. II. Skew Va ia ion in Homogeneous Ma e ial. Philos. T ans. R. Soc. Lond. A 1895,186, 343–414. [C ossRe ] 60. Hall, M.A. Co ela ion-based Fea u e Subse Selec ion o Machine Lea ning. Ph.D. Thesis, The Uni e si y o Waika o, Hillc es , New Zealand, 1998. 61. Goldbe g, D.E. Gene ic Algo i hms in Sea ch, Op imiza ion and Machine Lea ning, 1s ed.; Addison-Wesley Longman Publishing Co., Inc.: Bos on, MA, USA, 1989. 62. Aha, D.W.; Kible , D.; Albe , M.K. Ins ance-based lea ning algo i hms. Mach. Lea n. 1991 ,6, 37–66. [C ossRe ] 63. John, G.H.; Langley, P. Es ima ing con inuous dis ibu ions in Bayesian classi ie s. In P oceedings o he Ele en h Con e ence on Unce ain y in A i icial In elligence, Mon eal, QC, Canada, 18–20 Augus 1995; Mo gan Kau mann Publishe s Inc.: San F ancisco, CA, USA, 1995; pp. 338–345. 64. Quinlan, J.R. C4.5: P og ams o Machine Lea ning; Mo gan Kau mann Publishe s Inc.: San F ancisco, CA, USA, 1993. 65. B eiman, L. Random Fo es . Mach. Lea n. 2001,45, 5–32. [C ossRe ] Remo e Sens. 2017,9, 185 34 o 34 66. Popescu, S.C.; Wynne, R.H.; Nelson, R.F. Es ima ing plo -le el ee heigh s wi h lida : Local il e ing wi h a canopy-heigh based a iable window size. Compu . Elec on. Ag ic. 2002,37, 71–95. [C ossRe ] 67. Popescu, S.C.; Wynne, R.H. Seeing he ees in he o es . Pho og amm. Eng. Remo e Sens. 2004 ,70, 589–604. [C ossRe ] 68. Zome , R.J.; T abucco, A.; Us in, S.L. Building spec al lib a ies o we lands land co e classi ica ion and hype spec al emo e sensing. J. En i on. Manag. 2009,90, 2170–2177. [C ossRe ] [PubMed] 69. Dalpon e, M.; Ø ka, H.O.; Ene, L.T.; Gobakken, T.; Næsse , E. T ee c own delinea ion and ee species classi ica ion in bo eal o es s using hype spec al and ALS da a. Remo e Sens. En i on. 2014 ,140, 306–317. [C ossRe ] 70. Lee, J.; Cai, X.; Lellmann, J.; Dalpon e, M.; Malhi, Y.; Bu , N.; Mo ec o , M.; Schönlieb, C.-B.; Coomes, D.A. Indi idual ee species classi ica ion om ai bo ne mul i-senso image y using obus PCA. IEEE J. Sel. Top. Appl. Ea h Obs. Remo e Sens. 2016,9, 2554–2567. [C ossRe ] 71. Vaughn, N.R.; Moskal, L.M.; Tu nblom, E.C. T ee species de ec ion accu acies using disc e e poin LiDAR and ai bo ne wa e o m LiDAR. Remo e Sens. 2012,4, 377–403. [C ossRe ] 72. Ko pela, I.; Ø ka, H.O.; Mal amo, M.; Tokola, T.; Hyyppä, J. T ee species classi ica ion using ai bo ne LiDAR–e ec s o s and and ee pa ame e s, downsizing o aining se , in ensi y no maliza ion, and senso ype. Sil a Fenn. 2010,44, 319–339. [C ossRe ] 73. Kaa inen, H.; Hyyppä, J.; Yu, X.; Vas a an a, M.; Hyyppä, H.; Kukko, A.; Holopainen, M.; Heipke, C.; Hi schmugl, M.; Mo sdo , F.; e al. An In e na ional Compa ison o Indi idual T ee De ec ion and Ex ac ion Using Ai bo ne Lase Scanning. Remo e Sens. 2012,4, 950–974. [C ossRe ] 74. Wang, Y.; Hyyppa, J.; Liang, X.; Kaa inen, H.; Yu, X.; Lindbe g, E.; Holmg en, J.; Qin, Y.; Malle , C.; Fe az, A.; e al. In e na ional Benchma king o he Indi idual T ee De ec ion Me hods o Modeling 3-D Canopy S uc u e o Sil icul u e and Fo es Ecology Using Ai bo ne Lase Scanning. IEEE T ans. Geosci. Remo e Sens. 2016,54, 5011–5027. [C ossRe ] 75. Spe lich, M.; Ka enbo n, T.; Koch, B.; Ka enbo n, G. Po en ial o Unmanned Ae ial Vehicle Based Pho og amme ic Poin Clouds o Au oma ic Single T ee De ec ion. A ailable online: h p://www.dgp . de/neu/P oc2014/p oceedings/pape s/Bei ag270.pd (accessed on 15 Janua y 2015). 76. Pan , P. Op imizing Spec al Bands o Ai bo ne Image o T ee Species Classi ica ion. Ph.D. Thesis, Uni e si y o Eas e n Finland, Joensuu, Finland, 11 June 2015. 77. Fe e , J.B.; Asne , G.P. T ee Species Disc imina ion in T opical Fo es s Using Ai bo ne Imaging Spec oscopy. IEEE T ans. Geosci. Remo e Sens. 2013,51, 73–84. [C ossRe ] © 2017 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/).