Utilizing a multi-source forest inventory technique, MODIS data and Landsat TM images in the production of forest cover and volume maps for the Terai physiographic zone in Nepal
Full text
Remo e Sens. 2012, 4, 3920-3947; doi:10.3390/ s4123920
Remo e Sensing
ISSN 2072-4292
www.mdpi.com/jou nal/ emo esensing
A icle
U ilizing a Mul i-Sou ce Fo es In en o y Technique, MODIS
Da a and Landsa TM Images in he P oduc ion o Fo es Co e
and Volume Maps o he Te ai Physiog aphic Zone in Nepal
Ee o Muinonen 1,*, Heikki Pa ikka 1, Yam P. Pokha el 2, Sahas M. Sh es ha 2
and Kalle Ee ikäinen 1
1 Finnish Fo es Resea ch Ins i u e, Joensuu Uni , P.O. Box 68, FI-80101 Joensuu, Finland;
E-Mails: [email p o ec ed] (H.P.); kalle.ee ikainen@me la. i (K.E.)
2 Depa men o Fo es Resea ch and Su ey, G.P.O. 3339, Baba mahal, Ka hmandu, Nepal;
E-Mails: [email p o ec ed] (Y.P.), [email p o ec ed] (S.S.)
* Au ho o whom co espondence should be add essed; E-Mail: [email p o ec ed];
Tel.: +358-40-801-5282; Fax: +358-29-532-3113.
Recei ed: 15 Oc obe 2012; in e ised o m: 4 Decembe 2012 / Accep ed: 5 Decembe 2012 /
Published: 10 Decembe 2012
Abs ac : An app oach based on he nea es neighbo s echniques is p esen ed o
p oducing hema ic maps o o es co e ( o es /non- o es ) and o al s and olume o he
Te ai egion in sou he n Nepal. To c ea e he o es co e map, we used a combina ion o
Landsa TM sa elli e da a and isual in e p e a ion da a, i.e., a sample g id o isual
in e p e a ion plo s o which we ob ained he land use classi ica ion acco ding o he FAO
s anda d. These isual in e p e a ion plo s oge he wi h he ield plo s o olume mapping
o igina e om an ope a i e o es in en o y p ojec , i.e., he Fo es Resou ce Assessmen
o Nepal (FRA Nepal) p ojec . The ield plo s we e also used in checking he classi ica ion
accu acy. MODIS sa elli e da a we e used as a e e ence in a local co ec ion app oach
conduc ed o he ela i e calib a ion o Landsa TM images. This s udy applied a
non-pa ame ic k-nea es neighbo echnique (k-NN) o he o es co e and olume
mapping. A ee heigh p edic ion app oach based on a nonlinea , mixed-e ec s (NLME)
modeling p ocedu e is p esen ed in he Appendix. The MODIS image da a pe o med well
as e e ence da a o he calib a ion app oach applied o make he Landsa image mosaic.
The ag eemen be ween he o es co e map and he ield obse ed alues o o es co e
was subs an ial in Wes e n Te ai (KHAT 0.745) and s ong in Eas e n Te ai (KHAT
0.825). The o es co e and olume maps ha we e es ima ed using he k-NN me hod and
he in en o y da a om he FRA Nepal p ojec a e al eady app op ia e and aluable da a
OPEN ACCESS
Remo e Sens. 2012, 4 3921
o esea ch pu poses and o he planning o o hcoming o es in en o ies. Adap a ion o
he me hods and echniques was ca ied ou using Open Sou ce so wa e ools.
Keywo ds: da a impu a ion; GIS; k-nea es neighbo s; mapping o es a iables; nonlinea
mixed-e ec s model; Open Sou ce so wa e; Landsa ; MODIS
1. In oduc ion
A gene al aim o o es in en o y p ojec s is o o m a basis o decision-making and sus ainable use
o o es esou ces in he o m o objec i e s a is ics a he na ional and egional le els and o
adminis a i e uni s. Con en ionally, he s a is ics a e ob ained o p e-de ined a eas o he
species-speci ic measu es o he o al olume o li ing and dead ees and me chan able olume
cha ac e is ics. Ques ions and issues aised by he Uni ed Na ions F amewo k Con en ion on Clima e
Change (UNFCCC) h ough i s “Reducing Emissions om De o es a ion and Fo es Deg ada ion”
(REDD) p og am ha e inc eased he need o o es esou ce in en o ies and widened he applica ion
o hei esul s. Coun ies al eady ha ing sui able na ional o es in en o y (NFI) amewo ks can
ha ness hei NFIs o suppo also he assessmen o g een house gas (GHG) emissions. When
de eloping coun ies, such as Nepal, o ins ance, a e add essing his kind o ask, echnical and
me hodological suppo p o ided o assessing he o es esou ces may become e en mo e impo an
han be o e [1]. Due o he epe i ion equi emen s and empo al and geog aphical consis ency
condi ions, he possibili ies o u ilizing emo e sensing echnology oge he wi h ield in en o y da a
a e inc easingly gaining he in e es o he o ganiza ions esponsible o he o es in en o y ac i i ies
ela ed o he REDD mechanism (e.g., [2,3]). The ope a i e in en o y p ojec “Fo es Resou ce
Assessmen o Nepal” (FRA Nepal) belongs o he Finnish Go e nmen ’s de elopmen aid oolki . The
FRA Nepal p ojec is an example o an ac ion ha aims o de elop cos -e icien in en o y echniques
based on in o ma ion ob ained om bo h con en ional ield in en o ies and emo e sensing ma e ials.
Biome ical modeling is also needed o in en o y calcula ions ha make use o emo e sensing: o
example, he calcula ion sys em o he FRA Nepal p ojec consis s o models o heigh o olume by
ee species, which a e necessa y o impu ing he alues missing om he ield da a. Ob aining da a
om he ee le el o he s and o plo le el, o o he egional, le el is a p ocess by which in en o y
calcula ions and epo ing modules a e de eloped. To pu all o his oge he , a o es in en o y
da abase is needed.
To combine ield sample plo obse a ions, digi al map da a and sa elli e image da a, o
classi ica ion o p edic ion pu poses, mul i-sou ce o es in en o y (MSFI) echniques ha e been
de eloped [4,5]. E en i he pu e in en o y esul s a e di ec ly de i ed om he cha ac e is ics
measu ed by he ield sample plo s, a mul i-sou ce app oach aiming a he wall- o-wall mapping o
o es cha ac e is ics equi es emo e sensing da a. The necessa y inpu da a, i.e., he g ound- u h da a,
comes om he in en o y calcula ion sys em and om he Fo es In en o y Da abase. Tools o
p ocessing GIS and image da a a e u he u ilized, o example, when classi ying sa elli e image
pixels and pe o ming o e laying analysis.
Remo e Sens. 2012, 4 3922
A ep esen a i e and la ge enough ield sample o MSFI is a p e equisi e, so ha he es ima ion
esul s a e o a sa is ac o y quali y and accu a e enough. Howe e , i only a small o e y spa se
sample o ield plo s is a ailable, i mus be possible o no only u ilize plo s on he same image, bu
also on neighbo ing images. Fo he FRA Nepal p ojec , a minimum, i would be necessa y o ha e a
ela i e calib a ion p ocedu e ha would make i possible o combine se e al sa elli e images wi h
espec o he a ge in en o y a eas. In ou s udy case in Te ai egion, he main ac o s a ec ing he
a ailabili y and he usabili y o Landsa TM images a e cloudiness and seasonal a ia ions. I cloud- ee
images a e a ailable, one al e na i e o ela i e calib a ion is hen o adjus he adiome ic p ope ies
o a subjec image o hose o a e e ence image [6–11]. Seasonal a ia ion e ec s be ween images can
pa ly be dec eased by aiming o use images om he same ime o yea [6]. Fo ackling he p oblems
caused by cloudiness, gap- illing app oaches ha e been p esen ed, whe e se e al images a e used o
co e he gaps in a selec ed image [12].
The aim o his s udy was o p oduce hema ic maps o o es co e ( o es /non- o es ) and o al
s and olume o he Te ai egion in sou he n Nepal using a non-pa ame ic, k-nea es neighbo
echnique. The p e-p ocessing o Landsa TM images was conduc ed by using MODIS sa elli e da a as
a e e ence in a local co ec ion app oach. Mixed modeling was u ilized o he gene aliza ion o
sample ee cha ac e is ics ha was conduc ed as a pa o calcula ing he s and olumes by sample
plo s (see Appendix). The app oach used o c ea e he o es co e map was based on a combina ion o
Landsa TM sa elli e da a and isual in e p e a ion da a, i.e., a sample g id o isual in e p e a ion plo s
by which we ob ained he land use classi ica ion acco ding o he FAO s anda d. Fo he olume
mapping p ocedu e, we u ilized Landsa TM da a, he o es co e map and he o al s and olumes
ob ained by he ield measu ed sample plo s.
2. Ma e ials
2.1. S udy A ea
The a ge a ea o he s udy, i.e., he Te ai egion, is loca ed app oxima ely be ween
26°15′–29°15′N and 80°00′–88°15′E. The Te ai egion is he sou he nmos o he i e physiog aphic
zones o Nepal and comp ises 14% o he coun y’s land (Figu e 1). In con as o he o he ou
physiog aphic zones, he e ain in his sub opical lowland pla eau is opog aphically less complex,
and i s ele a ion a ies om 60 o 330 m abo e sea le el [13]. In his s udy, he wes e n Te ai
comp ises he wo le mos sub- egions o Te ai, whe eas he eas e n Te ai consis s only o he
igh mos sub- egion o Te ai (Figu e 1).
2.2. Image Ma e ials
Th ee Landsa TM sa elli e images om he wes e n pa o Te ai and ou images om he eas e n
pa o Te ai we e included in wo image mosaics, one co e ing wes e n Te ai in he UTM zone o 44
and ano he co e ing eas e n Te ai in he UTM zone o 45 (Table 1).
MODIS su ace e lec ance p oduc s p o ide an es ima e o he su ace spec al e lec ance as i
would be measu ed a g ound le el in he absence o a mosphe ic sca e ing o abso p ion. The
Remo e Sens. 2012, 4 3923
MODIS Su ace Re lec ance p oduc “MYD09A1” is a combina ion o he bes possible obse a ions
in an eigh -day ime pe iod a a 500-m esolu ion a he sinusoidal p ojec ion [14].
MYD09A1 ma e ial om he h25 06 and h24 06 sinusoidal iles we e u ilized as image mosaic
e e ence ma e ials. The eigh -day ime in e al o bo h MODIS p oduc s was om 10 Feb ua y 2010
o 17 Feb ua y 2010. The MODIS’s da a co e age could be ound om a ime in e al close o he
da es o he Landsa images, which p o ided a qui e easonable s a ing poin o image mosaic
building and image in e p e a ion pu poses. One Landsa TM image om he no h-wes e n side o
Te ai was aken in 2011, hus ep esen ing a di e ence o jus one yea om he ime-poin o he
MODIS da a.
Figu e 1. Map o Nepal wi h bo de s o he physiog aphic zones. The s udy a ea, i.e.,
Te ai, is deno ed by he g ey- ill colo .
Table 1. In o ma ion abou he Landsa TM sa elli e image y used in his s udy. Pa h and
ow in o ma ion is gi en in he WRS-2 sys em.
A ea
Da e
Pa h
Row
Wes e n Te ai
2011-03-07
144
40
2010-02-25
143
41
2010-02-18
142
41
Eas e n Te ai
2010-02-11
141
41
2010-02-04
140
41
2010-02-04
140
42
2010-01-28
139
42
Remo e Sens. 2012, 4 3924
The MODIS Rep ojec ion Tool [15] was used o ans o m he MODIS da a in o a UTM zone44
and zone45 coo dina e sys em wi h wgs84 da um and 500-m esolu ion. All o he image p ocessing
ope a ions we e ca ied ou using Open Sou ce ools and he u ili ies o GDAL, Quan um GIS and
GRASS GIS [16–18].
2.3. Two-Phase Sampling and Visual In e p e a ion Da a
The se up o he ongoing na ional o es in en o y was designed o he FRA Nepal p ojec and
o med he basis o his s udy. In he FRA Nepal p ojec , a s a i ied sys ema ic clus e sampling
design wi h a wo-phase clus e sampling me hod was applied [19,20].
In o de o loca e he sampling uni s, i.e., he sample plo s wi hin squa e clus e s, h oughou he
whole coun y, a 4 km by 4 km g id was c ea ed. Each clus e consis ed o six sample plo s in wo
pa allel lines in a No h-Sou h di ec ion; he sample plo s we e 150 m apa and he dis ance be ween
he No h-Sou h lines was 300 m (see Figu e 2(a)). The sampling o clus e s wi h he sample plo s o
he ield measu emen s was conduc ed in wo phases [20]. Du ing he i s phase, each clus e o six
plo s along he 4 km by 4 km g id co e ing he en i e coun y was isually classi ied acco ding o land
use class (1 = Fo es a ea (n = 1,603), 2 = O he wooded a ea (n = 131), 3 = Ag icul u al a ea wi h ee
co e (n = 796), 4 = Ag icul u al a ea wi hou ee co e (n = 4,168), 5 = Buil -up a ea wi h ee co e
(n = 202), 6 = Buil -up a ea wi hou ee co e (n = 119), 7 = Roads (n = 3), 8 = O he a ea (n = 229),
9 = Wa e a ea (n = 282), whe e n is he numbe o obse a ions in Te ai) and accessibili y.
Technically, he in e p e a ion was implemen ed using Google Ea h, a i ual globe and sa elli e
image y iewe [21] consis ing o eely a ailable sa elli e images and addi ional da a laye s, including
RapidEye image y, opog aphical maps and he sampling g id o Nepal. In his s udy, he isual
in e p e a ion da a o he Te ai physiog aphic zone comp ised al oge he 7,533 plo wise obse a ions
wi hin 1,281 clus e s. Du ing he second phase, he selec ion o clus e s o ield measu emen s was
done by s a a. The i s s a um con ained he clus e s ha ing a leas one plo in he o es a ea and he
second s a um con ained he clus e s wi h no o es a ea plo s in he 1s phase sample. Du ing his
sampling p ocess, mo e weigh was gi en o he clus e s comp ising plo s on o es a ea. Only he
ou -co ne plo s o he clus e in a squa e wi h a side leng h o 300 m we e used in hese analyses (see
Figu e 2(a)).
2.4. Field Da a
The ield da a o his s udy we e collec ed om he plo s sampled o ield measu emen s be ween
2010 and 2011 in he FRA Nepal p ojec . The ield-measu ed in en o y da a we e inally ob ained
om 217 sample plo s wi hin 56 clus e s on o es land (171 plo s) and non- o es land (46 plo s).
The ype o sample plo employed in he FRA Nepal p ojec is a Concen ic Ci cula Sample Plo
(CCSP), which is ecommended o in en o ies o na u al o es s ha a e la ge in size and
cha ac e ized by many old ees and a wide a ie y o ee species (see [20,22]). In he FRA Nepal
p ojec , he CCSPs consis o ou ci cula plo s ha ing adii o 20, 15, 8 and 4 m and diame e -a -
b eas -heigh h eshold anges o ≥30.0, 20.0–29.9, 10.0–19.9 and 5.0–9.9 cm, espec i ely. Since all
Remo e Sens. 2012, 4 3925
he FRA Nepal sample plo s ha e been es ablished on a pe manen basis, he posi ions o he ees
we e ob ained ia bea ing and dis ance measu emen s (Figu e 2(b)).
Figu e 2. Layou o he in en o y clus e (a) and Concen ic Ci cula Sample Plo (CCSP)
(b) used in he FRA Nepal p ojec . In he case o Te ai, one clus e comp ises ou CCSPs
(1, 3, 4 and 6) in a squa e wi h 300 m sides, i.e., plo s numbe 2 and 5 o he basic clus e ,
which a e used in he isual in e p e a ion ( i s s age sample) and a e excluded om he
ield in en o y (second s age sample). In (b), he symbols 1,..., 4 a e o he adii o he
ou ci cula plo s (4, 8, 15 and 20 m, espec i ely) wi hin he CCSP.
(a)
(b)
Remo e Sens. 2012, 4 3926
Bo h he diame e a b eas heigh and he ee species we e measu ed o e e y allied ee, whe eas
e e y i h ee was ea ed as a sample ee, o which he o al ee heigh was also measu ed. Gene al
s and cha ac e is ics and plo -speci ic a iables, such as he land use class, o es ype, aspec , x and y
coo dina es ob ained o he cen e poin using a GPS de ice, soil dep h, main si e ype, o es ype,
de elopmen s a us and o igin o he o es we e also collec ed. Desc ip i e s a is ics o he sample
plo da a collec ed om Te ai a e gi en in Table 2.
Table 2. Desc ip i e s a is ics o he CCSP da a, (n = 217): Dg is he basal a ea weigh ed
mean diame e a b eas heigh (cm), Hg is he basal a ea weigh ed mean heigh (m), N is
he numbe o li ing ees (ha−1), G is he s and basal a ea o li ing ees (m2∙ha−1) and V is
he s and olume o li ing ees (m3∙ha−1).
Va iable
Minimum
Median
Mean
S anda d De ia ion
Maximum
Dg
0.00
35.64
36.95
25.70
145.11
Hg
0.00
17.44
16.13
9.23
40.00
N
0
288.5
473.6
546.3
2735.5
G
0.00
16.09
15.00
10.45
43.31
V
0.00
130.94
137.53
112.66
499.59
The exis ing ee-le el olume models de eloped by Sha ma and Pukkala [23] o Nepalese ee
species a e applicable when he species, diame e a b eas heigh and o al heigh , measu ed o
p edic ed, a e known. The loga i hmically linea ized allome ic models de eloped by Sha ma and
Pukkala [23] we e a ailable o a o al o 21 indi idual ee species and wo addi ional ee species
g oups. Heigh s we e needed o all ally ees in o de o apply hese exis ing s em olume models
based on he heigh o he ee and i s diame e a b eas heigh (see Sha ma and Pukkala [23]).
The e o e, i was necessa y o de elop a heigh gene aliza ion model wi h species-speci ic pa ame e s
designed o he in en o y da a om Te ai.
The heigh p edic ion app oach based on a nonlinea , mixed-e ec s (NLME) modeling p ocedu e
(see, e.g., Pinhei o and Ba es [24]) is p esen ed in he Appendix. The heigh s o he ally ees (see
Appendix) we e p edic ed i s . S em olumes we e he ea e p edic ed o all he allied ees using
species-speci ic olume unc ions o Sha ma and Pukkala [23]. Finally, o al s and olumes by sample
plo s we e calcula ed (Table 2).
3. Me hods
3.1. Building a Sa elli e Image Mosaic
The FAO NAFORMA p ojec has p oduced a echnique o c ea ing image mosaics o he pu poses
o o es in en o y planning [8]. A MODIS (MYD09A1) image, i.e., an a mosphe ically co ec ed image
wi h a 500-m pixel size, was used as a e e ence o a mosphe ically co ec ing he Landsa images. The
basic p inciple was o ma ch he mean and he a iance o he da a in bo h images by aking in o accoun
he di e ence in he pixel sizes [8]. In he echnique de eloped by Tomppo e al. [8], simple linea
mapping was calcula ed o each band and he co ec ion was calcula ed sepa a ely o each Landsa
Remo e Sens. 2012, 4 3927
image when making he mosaics. The a e aging was used o he Landsa da a o accoun o he la ge
esolu ion in he MODIS da a.
The app oach de eloped by Tomppo e al. [8], which hey ini ially used in Tanzania, is close o ha
o he one de eloped by Tuominen and Pekka inen [7] o local adiome ic co ec ion o assis wi h
he mul i-sou ce o es in en o y. In hei s udy, he idea was o adjus he BRDF-a ec ed in ensi ies in
digi ized ae ial pho og aphy o ma ch he local in ensi ies o he e e ence image y, i.e., he Landsa 5
TM sa elli e da a. Thus, he e e ence image y needs o be independen o he BRDF. The local
calib a ion was based on adjus ing he mean alues in a co ec ion uni , o example, a mo ing ci cle
wi h a adius o 40 m. Using a mo ing window, an adap i e co ec ion based on he e e ence image
digi al numbe s became possible; likewise, he pa ame e s o he me hod we e de ined empi ically.
In he case o Te ai, app oaches de eloped by bo h Tomppo e al. [8] and Tuominen and
Pekka inen [7] we e combined o c ea e a local co ec ion model. The sepa a e Landsa TM images
we e co ec ed by applying he co ec ion unc ion below (see Equa ion (1)) and by using MODIS
image da a as he unde lying e e ence image. The objec i es o he p ocedu e p esen ed he e a e as
ollows: (1) o ma ch he mean and he a iance o he da a in bo h images by aking in o accoun he
di e ences in pixel sizes [8] and (2) o apply he co ec ion locally [7] by using as e map algeb a and
a mo ing window app oach o calcula e he pa ame e alues o he model.
The co ec ion unc ion (see Tomppo e al. [8]) applied o a pixel (x, y) was as ollows:
),(),(),(),(
ˆyxbyx yxayx iiii
(1)
whe e
),(
ˆyx i
is he co ec ed da a o he pixel (x, y) in band i, i (x, y) is he unco ec ed da a in band
i, and ai (x, y) and bi (x, y) a e he pa ame e s compu ed o he gi en pixel (x, y) in band i.
To calcula e he pa ame e s o he co ec ion unc ion (Equa ion (1)), a e aging was i s used o
ans o m he Landsa image da a o ma ch he esolu ion o he MODIS e e ence image. The common
pixel size o he images a e a e aging is la e deno ed by c. This s udy used he c- alue o 500 m.
The as e maps “
Li
a g
”, “
Li
sd
”, “
Mj
a g
” and “
Mj
sd
”—which ep esen he mean and s anda d
de ia ion alues o he Landsa TM da a and he MODIS da a in he compa ible spec al wa eleng h
bands i and j, espec i ely—we e he ea e calcula ed using a mo ing window app oach wi h a
window size o w × w pixels. He e, a 21 × 21 window neighbo hood was u ilized (w = 21),
ep esen ing app oxima ely a 10 km neighbo hood o each pixel wi h he size o c. The as e maps
wi h a pixel size o c we e compu ed o bo h image da ase s: a ge images (Landsa TM) and
e e ence images (MODIS).
Pa ame e s ai and bi in he co ec ion unc ion (Equa ion (1)) we e calcula ed o each pixel (x, y) as
ollows:
),(),(),( yx/sdyxsdyxa LiMji
and
),(),(),(),( yxayxa gyxa gyxb iLiMji
(2)
whe e i deno es a Landsa TM band and j deno es a MODIS band ha is compa ible wi h i. This s udy
used he MODIS bands 3, 4, 1, 2, 6 and 7 as compa ible bands o he Landsa TM bands 1, 2, 3, 4, 5
and 7.
Remo e Sens. 2012, 4 3928
Finally, he co ec ion model (Equa ion (1)) was applied in he o iginal Landsa TM image da a o
p oduce locally co ec ed Landsa da a in a MODIS e lec ance alue scale and a he o iginal 30-m
esolu ion. The coe icien s ai and bi (see Equa ion (2)) o he co ec ion unc ion (Equa ion (1)) we e
local, being based on s a is ics om he neighbo hood o each pixel (x, y). As a esul , he Landsa TM
pixel alues we e escaled acco ding o he local dis ibu ion pa ame e s (mean and s anda d de ia ion)
o he pixel alues in he e e ence image bands.
In he co ec ion app oach o his s udy, each Landsa TM image had o i s be p ocessed
sepa a ely, a e which a mosaic o he co ec ed Landsa TM images was c ea ed. This mosaic was
used as inpu o he o es /non- o es classi ica ion and he wall- o-wall map gene a ion o a o es
a iable (s and olume, m3∙ha−1).
The esolu ion (c) o he as e maps ep esen ing he mean and s anda d de ia ion and he size o
he mo ing window (w) a e use -de ined pa ame e s. He e, we used a esolu ion co esponding o ha
o he e e ence MODIS image, i.e., a e aging was used only o Landsa images, as men ioned abo e.
The size o he mo ing window (w) de e mines he lexibili y o local co ec ion. Howe e , he
window size (w) should be la ge enough o enable alid compu a ions o he mean and s anda d
de ia ion. In his case, i was no possible o conduc an exhaus i e sea ch o he esolu ion and
neighbo hood size (w), and he e o e, we e alua ed he esul by checking he isual appea ance o he
esul ing image mosaic. A good-quali y e e ence image wi hou any ex eme pixel alues is needed
o a success ul esul .
3.2. Nea es Neighbo s Techniques
A non-pa ame ic, mul i-sou ce o es in en o y me hod based on he k-nea es neighbo (k-NN)
es ima ion (see Tomppo [4], Tomppo e al. [5]), an app oach ecen ly e iewed by McRobe s [25],
was applied in he p oduc ion o o es co e and olume maps in he Te ai egion. In his s udy, he
popula ion uni s ( he a ge se o pixels) we e he pixels in he sa elli e image. The sa elli e image
bands we e used as he ancilla y a iables ( ea u e a iables) wi h obse a ions o all uni s o he
popula ion. The o es a iables wi h obse a ions only a ailable o a sample ( he e e ence se o
pixels) we e deno ed as esponse a iables. In p ac ice, he e e ence se is buil by que ying he
sa elli e image in he loca ions o sample plo s, whe eas he aim o he nea es neighbo s app oach is o
impu e he esponse a iable alues o he a ge se elemen s.
The classi ica ion was based on a pixel-by-pixel analysis, whe e he nea es neighbo s o a a ge
pixel among all he e e ence pixels ( he pixels co e ing he cen e poin o a sample plo ) we e
de e mined using weigh ed Euclidean dis ance in spec al ea u e space (see Tokola [26]; o his in
ma ix o m, see, e.g., McRobe s [25]):
b
1
2
)]([
n
hjhihhij bbpd
,
(3)
whe e nb is he numbe o bands, i is he a ge se elemen o which a p edic ion is sough and j is a
e e ence se elemen , bih and bjh a e spec al band alues o he pixels i and j on band h, espec i ely,
and ph is he empi ical pa ame e o band h.
Remo e Sens. 2012, 4 3935
we e 85.4 m3∙ha−1 (62.0%) and −0.541 m3∙ha−1 (sligh o e es ima ion e ec ), espec i ely. The
boxplo [31] o esiduals o he olume p edic ion ca ego ies o 50 m3∙ha−1 show he p esence o some
ex eme obse a ions, whe eas he median o esiduals is close o ze o in all ca ego ies (Figu e 6).
Figu e 5. A o es co e map ( o es : b igh g een colo ) o he Te ai physiog ahic zone
p oduced by he k-NN echnique and o e laid on a composi e o MODIS image bands 1, 4
and 3 ( gb).
Figu e 6. Volume esiduals in olume p edic ion ca ego ies o 50 m3∙ha−1.
Remo e Sens. 2012, 4 3936
Plo ing he means o he o de ed g oups o obse a ions agains he means o he p edic ions o
he same o de ed g oups (see McRobe s [33]) e eals ha olume p edic ions su e om a e aging:
he la ge olumes we e unde es ima ed and he small olumes we e o e es ima ed (Figu e 7).
Figu e 7. Mean obse a ion e sus mean p edic ion o olume (m3∙ha−1). The
obse a ions ha e been so ed and g ouped in o g oups o 20 obse a ions a he minimum
based on he obse ed olume (m3∙ha−1).
5. Discussion
In he case o he Te ai egion, i was no possible o apply he nea es neighbo ’s echniques o
each Landsa image sepa a ely, because he numbe o plo s was inadequa e. This mean ha we
needed o apply ela i e calib a ion o he Landsa images, despi e he ac ha he e could be qui e a
long amoun o ime be ween he da es o he a ailable good-quali y images. Fo his pu pose, we
applied a p e-p ocessing app oach based on using a e e ence image co e ing he whole egion. Fo
aiming a a seamless esul , he po en ial o e lapping a ea o he neighbo ing a ge images (Landsa
TM images in ou case) needs o be u ilized when he as e maps ep esen ing means and s anda d
de ia ions a e compu ed. By de ec ing pixels ha ing ex eme alues (ou lie pixels) in he images and
excluding hem om his s age, one can imp o e he esul o his local app oach. Depending on image
ma e ials, his me hod also allows one o apply a pixel size c di e en han he o iginal pixel size in he
e e ence image.
A obus eg ession app oach would o e ano he empi ically o ien ed app oach o a ela i e
calib a ion o image ma e ials. Reg essing he images o he selec ed e e ence image would make i
possible o concu en ly use neighbo ing images. In he case o he geog aphically wide Te ai egion,
Remo e Sens. 2012, 4 3937
he obus eg ession app oach was less applicable. Howe e , in a mul i- empo al case, a eg ession
app oach o ela i e adiome ic calib a ion could possibly be applied when de ec ing changes in
landscape, such as cu ings (e.g., [11,34,35]).
The ag eemen be ween he k-NN o es co e classi ica ion and he isual in e p e a ion in Te ai
was s ong, since he esul ing alue o KHAT was g ea e han 0.80 (see, e.g., Congal on [27] and
G een [32]). The OA alues we e also high, which can a leas be pa ly explained by he high
p opo ion o he non- o es ca ego y. I is wo h no ing ha we did no examine he e ec o he
pos -p ocessing phase (s eps 1 o 3) when calcula ing he accu acy measu es in he es samples o he
selec ed alue o k. Also, he classi ica ion accu acy o he o es co e map e alua ed agains ield
obse ed alues showed ha he ag eemen be ween hese o es co e delinea ions is subs an ial
(wes e n Te ai) o s ong (eas e n Te ai).
The isual land use class in e p e a ion p ocess was conduc ed using Google Ea h sa elli e image y
iewe [21], whe e he backg ound image y o igina ed om he pe iod be ween 2003 and 2010,
oge he wi h addi ional RapidEye image y om he yea 2010. Due o shaded a eas, haze and
cloudiness, he isual in e p e a ion wo k p o ed di icul in he case o some RapidEye images, and
hen, he in e p e a ion was suppo ed using sa elli e image y a ailable in Google Ea h. The e o e, he
ime di e ence be ween hese image ma e ials could make a sou ce o unce ain y o he in e p e a ion
p ocess i sel . In his s udy, an unwan ed mixing o he o es class and he land use classes wi h some
ee co e was ackled by d opping ou he plo s in classes “Ag icul u al a ea wi h ee co e ” and
“Buil -up a ea wi h ee co e ” om he e e ence se . Fo classi ica ion accu acy checking, a co ec
ma ching o he loca ions be ween ield plo s and he isual in e p e a ion plo s on he image is also
c ucial. Un o una ely, in o ma ion o check his loca ion accu acy was no a ailable.
Besides RMSE and bias, obus measu es o he quali y o he k-NN-based es ima ion in e ms o
ea u e selec ion and c oss- alida ion could be use ul, especially in cases whe e he numbe o plo s is
small, which co esponds o his s udy (n = 217). Fo ins ance, in he k-nea es neighbo analysis o
s and olume, he mos ex eme obse a ions may also ha e a ec ed he model pa ame e sea ch,
because he numbe o ield plo s was qui e small. The e o e, he diagnos ic app oach sugges ed by
McRobe s [33] o de ec ing ou lie s and in luen ial obse a ions could be e y impo an o ou
wo k in he u u e.
The k-nea es neighbo echnique, applied o he mapping o o es co e and s and olume in he
Te ai egion o Nepal, is a s aigh o wa d p ocedu e ha has been e icien ly u ilized in Finland (see,
e.g., Tomppo e al. [5] and Tomppo [4]). Mo eo e , his echnique was also applied ea lie by
Tokola e al. [36] o classi ying land use and o es ima ing imbe olume and biomass in Nepal.
This was, howe e , one o he i s o es mapping s udies comple ely conduc ed using Open Sou ce
so wa e ools and ee packages. In his espec , i was na u al ha subs an ial e o s we e needed o
inco po a e app op ia e da a p ocessing echniques and sui able p ocedu es and, especially, ha he
echniques and p ocedu es we e compa ible o p ocessing he emo e sensing and in en o y da a.
The o es co e map now a ailable o Te ai (Figu e 5) is one sou ce o in o ma ion ha can be
u ilized when es ima ing abo e-g ound o es olumes and biomasses based on he LiDAR da a as pa
o he ongoing su ey conduc ed by he FRA Nepal p ojec in he cen al pa s o Te ai. In he u u e,
simila o es co e maps will also be needed o o he physiog aphic zones beyond Te ai, whe e he
LiDAR-o ien ed es ima ions o o es a ibu es will be conduc ed. La e , expe ience will show i he
Remo e Sens. 2012, 4 3938
o es co e maps p o e use ul in o he ields o o es y. Po en ial applica ions in Nepalese condi ions
include in en o y planning and moni o ing p ac ices. The o es co e and olume maps (Figu e 8)
es ima ed using he k-NN me hod and in en o y da a om he FRA Nepal p ojec a e al eady
app op ia e and aluable da a o esea ch pu poses and o planning o hcoming o es in en o ies
when es ing op imal in en o y designs (see [37]).
Figu e 8. A le : o es co e map ( ec o bounda ies) and a ield sample plo clus e ,
(plo n = sample plo numbe ; zone = UTM zone numbe ; ba_ha = basal a ea, m2∙ha−1;
ol_ha = olume, m3∙ha−1; mcl1– mcl6 = pixel alues in he Landsa TM mosaic
bands 1–6). A igh : A hema ic map o he olume. An example om eas e n Te ai
(backg ound: a composi e o Landsa TM mosaic bands 5, 3 and 2 ( gb)). (Coo dina e
e e ence sys em: WGS 84/UTM zone 45N).
Spa ially explici biomass maps could also o m a basis o he o hcoming g eenhouse gas
in en o y, i.e., he REDD- ela ed es ima ion o g oss p ima y p oduc ion and soil ca bon change. The
k-NN-based o es biomass mapping echnique, which co esponds o ha o s and olume mapping,
has al eady been demons a ed by Tuominen e al. [38] in Finnish condi ions. One sha ed ea u e
be ween he app oach by Tuominen e al. [38] and he one in his s udy has o do wi h he da a-
e icien , mixed-e ec s modeling-based gene aliza ion o sample ee cha ac e is ics. This s udy,
howe e , used an NLME model o gene alizing he sample ee heigh s (see Appendix), whe eas he
model u ilized by Tuominen e al. [38] made use o an LME model p o ided by Ee ikäinen [39].
Adap ing he me hods and echniques and applying he Open Sou ce so wa e ools p esen ed he e
equi es capaci y building in Nepal: de elopmen wo k and educa ion measu es ha e al eady been
launched by he ICI p ojec , which is an in e -ins i u ional de elopmen coope a ion p ojec be ween
Nepalese, Vie namese and Finnish go e nmen al ins i u es and ha was ini ia ed wi h he suppo o
he Minis y o Fo eign A ai s (MFA) o Finland. The de elopmen ac i i ies o he p ojec ha e
Remo e Sens. 2012, 4 3939
been designed and implemen ed wi h a special ocus on human capaci y de elopmen wi hin he
go e nmen al o es esea ch o ganiza ions pa icipa ing in he p ojec . Special emphasis has he e o e
been gi en o hands-on aining pe iods and wo kshops and o dissemina ing in o ma ion on o es
in en o y- ela ed echniques and p ocedu es. O hese objec i es, he la e co e s he goal o he
p esen s udy: o inc ease and sha e knowledge abou he exis ing echniques and p ocedu es and o
make i easie o adap hem o he condi ions in, o ins ance, Nepal.
Local expe ise is always equi ed o he mos pi o al pa o he o es in en o y, ha is o say, he
wo k conduc ed in he ield and, a he momen , by he FRA Nepal p ojec in Nepal. Th ough scien i ic
wo k and collabo a ion, howe e , i is possible o achie e mo e di e se and de ailed esul s in e ms o
con en ional s a is ics and ad anced elec onic maps o he o es a ibu es o in e es . This s udy is
also one example o how o imp o e he use o ield in en o y da a and exploi hem oge he wi h
addi ional emo e sensing da a in o de o sa is y he new epo ing equi emen s se o la ge-scale
o es in en o ies.
6. Conclusions
In his s udy, we in oduced an app oach o a MODIS-based ela i e calib a ion o Landsa TM
images o enable he use o a mosaic o se e al Landsa TM images in a k-nea es neighbo (k-NN)
es ima ion. The p esen ed app oach o ela i e calib a ion combined aspec s p esen ed ea lie by
Tuominen and Pekka inen [7] and Tomppo e al. [8]. The me hod elies only on image da a (see [7])
and is easy o implemen . Op ionally, he e e ence image could ha e been con e ed o a e lec ance
scale, bu o he k-NN es ima ion me hod ha was no necessa y. A p e equisi e o he success o he
local co ec ion app oach o he ela i e calib a ion o images is ha he a ge images and he
good-quali y e e ence image ma e ial need o be empo ally and seasonally close o each o he .
Howe e , mo e s udies a e needed on selec ing he pa ame e s o he app oach when using image
esolu ion scales o he han he ones used in he Te ai case, which u ilized Landsa TM and MODIS.
The k-nea es neighbo echnique was e y applicable o he o es co e mapping in Te ai using
isual in e p e a ion plo s as a e e ence ma e ial. The e was a s ong ag eemen (KHAT > 0.80)
be ween he o es co e delinea ions based on isual in e p e a ion and ield obse a ions. The
ag eemen be ween he o es co e delinea ion in he o es co e map and he one de i ed om he
ield obse ed alues was subs an ial in Wes e n Te ai (KHAT 0.745) and s ong in Eas e n Te ai
(KHAT 0.825).
One ea u e ela ed o he de elopmen o he mul i-sou ce echnique and, especially, o i s
applica ion o di e en geog aphical condi ions in Nepal in he u u e has o do wi h using digi al
ele a ion models (DEMs). In he case o Te ai, i.e., wi hin he lowlands o Nepal, no adiome ic
co ec ions we e conduc ed using DEMs (e.g., Tomppo e al. [5]; Tokola e al. [36]), nei he was he
DEM-based mo ing geog aphical e ical e e ence a ea used in he k-NN me hod (see Ka ila and
Tomppo [29]). I is he e o e ecommended ha he signi icance o DEM be es ed in o he
physiog aphic ege a ion zones no h o Te ai, whe e he opog aphy is e y moun ainous compa ed o
he sou he nmos egion o Nepal.
Remo e Sens. 2012, 4 3940
Acknowledgmen s
This s udy was conduc ed a he Finnish Fo es Resea ch Ins i u e’s (Me la) Regional Uni in Joensuu,
Finland and he Depa men o Fo es Resea ch and Su ey (DFRS) in Baba mahal, Ka hmandu, Nepal.
I was pa ly unded by he Minis y o Fo eign A ai s o Finland h ough he in e -ins i u ional
de elopmen coope a ion p ojec en i led “Imp o ing Resea ch Capaci y o Fo es Resou ce In o ma ion
Technology in Vie nam and Nepal” (con ac numbe : HELM431-14, in e en ion code: 79811501). We
a e g a e ul o hese ins i u ions o hei esou ces and unding. Suppo gi en by he Fo es Resou ce
Assessmen o Nepal p ojec is also g a e ully acknowledged. Many hanks a e also due o Juho Pi känen
o his help in de eloping he calcula ion p ocedu es o his s udy.
Re e ences
1. Acha ya, K.P., Dangi, R.B., T ipa hi, D.M., Bushley, B.R., Bhanda y, R.R., Bha a ai, B., Eds.;
Ready o REDD? Taking S ock o Expe ience, Oppo uni ies and Challenges in Nepal; Nepal
Fo es e s’ Associa ion: Ka hmandu, Nepal, 2009.
2. Pea son, T.; Walke , S.; B own, S. Sou cebook o Land Use, Land-Use Change and
Fo es y P ojec s; Win ock In e na ional and he BioCa bon Fund o he Wo ld Bank:
Washing on, DC, USA, 2005. A ailable online: h p://www.win ock.o g/ecosys ems/ iles/
win ock-bioca bon_ und_sou cebook-comp essed.pd (accessed on 30 May 2012).
3. GOFC-GOLD. A Sou cebook o Me hods and P ocedu es o Moni o ing and Repo ing
An h opogenic G eenhouse Gas Emissions and Remo als Caused by De o es a ion, Gains and
Losses o Ca bon S ocks in Fo es s Remaining Fo es s, and Fo es a ion; GOFC-GOLD Repo
e sion COP17-1; GOFC-GOLD P ojec O ice, Na u al Resou ces Canada: Edmon on, AB,
Canada, 2011. A ailable online: h p://www.go cgold.wu .nl/ edd/ (accessed on 20 Sep embe
2012).
4. Tomppo, E. The Finnish Mul i-Sou ce Na ional Fo es In en o y—Small A ea Es ima ion and
Map P oduc ion. In Fo es In en o y. Me hodology and Applica ions; Managing Fo es
Ecosys ems; Kangas, A., Mal amo, M., Eds.; Sp inge : Do d ech , The Ne he lands, 2006;
Volume 10; pp. 195–224.
5. Tomppo, E.; Haakana, M.; Ka ila, M.; Pe äsaa i, J. Mul i-Sou ce Na ional Fo es In en o y
Me hods and Applica ions; Managing Fo es Ecosys ems, Sp inge Science+Business Media:
New Yo k, NY, USA, 2008; Volume 18.
6. Coppin, P.; Jonckhee e, I.; Nackae s, K.; Muys, B.; Lambin, E. Digi al change de ec ion me hods
in ecosys em moni o ing: A e iew. In . J. Remo e Sens. 2004, 25, 1565–1596.
7. Tuominen, S.; Pekka inen, A. Local adiome ic co ec ion o digi al ae ial pho og aphs o mul i
sou ce o es in en o y. Remo e Sens. En i on. 2004, 89, 72–82.
8. Tomppo, E.; Ka ila, M.; Mäkisa a, K.; Pe äsaa i J.; Malimbwi, R.; Chamuya, N.; O ieno, J.;
Dalsgaa d, S.; Leppänen, M. A Repo o he Food and Ag icul u e O ganiza ion o he Uni ed
Na ions (FAO) in Suppo o Sampling S udy o Na ional Fo es y Resou ces Moni o ing and
Assessmen (NAFORMA) in Tanzania; FAO: Rome, I aly, 10 Ma ch 2010. A ailable online:
h p://www.mp-discussion.o g/NAFORMA.pd (accessed on 24 Augus 2012).
Remo e Sens. 2012, 4 3941
9. Boda , C.; E a, H.; Beuchle, R.; Raši, R.; Simone i, D.; S ibig, H.-J.; B ink, A.; Lindquis , E.;
Acha d, F. P e-p ocessing o a sample o mul i-scene and mul i-da e Landsa image y used o
moni o o es co e changes o e he opics. ISPRS J. Pho og amm. 2011, 66, 555–563.
10. Tomppo, E.; Czaplewski, R.L.; Mäkisa a, K. FRA 2000 The Role o Remo e Sensing in Global
Fo es Assessmen . A Remo e Sensing Backg ound Pape o Ko ka IV Expe Consul a ion
01.07–05.07.2002, Ko ka, Finland; Wo king Pape 61; Fo es Resou ces Assessmen P og amme
(FRA) o FAO: Rome, I aly, 2002. A ailable online: p:// p. ao.o g/doc ep/ ao/006/ad650e/
ad650e00.pd (accessed on 20 No embe 2012).
11. Tokola, T.; Lö man, S.; E kkilä, A. Calib a ion o mul i empo al landsa da a o o es co e
change de ec ion. Remo e Sens. En i on. 1999, 68, 1–11.
12. Open Fo is Wiki. A ailable online: h p://km. ao.o g/OFwiki/index.php/Main_Page (accessed on
26 June 2012).
13. Lillesø, J.-P.B.; Sh es ha, T.B.; Dhakal, L.P.; Nayaju, R.P.; Sh es ha, R. The Map o Po en ial
Vege a ion o Nepal: A Fo es y/Ag oecological/Biodi e si y Classi ica ion Sys em; De elopmen
and En i onmen Se ies 2–2005 and CFC-TIS Documen Se ies No. 110; Fo es & Landscape:
Aalbo g, Denma k, 2005.
14. MODIS P oduc s Table. A ailable online: h ps://lpdaac.usgs.go /p oduc s/modis_p oduc s_ able
(accessed on 18 June 2012).
15. MODIS Rep ojec ion Tool. A ailable online: h ps://lpdaac.usgs.go / ools/
modis_ ep ojec ion_ ool (accessed on 26 June 2012).
16. GDAL: Geospa ial Da a Abs ac ion Lib a y. A ailable online: h p://www.gdal.o g/index.h ml
(accessed on 18 June 2012).
17. Quan um GIS. A ailable online: h p://www.qgis.o g/ (accessed on 18 June 2012).
18. Ne ele , M.; Mi aso a, H. Open Sou ce GIS: A GRASS GIS App oach, 3 d ed.; Sp inge
Science+Business Media: New Yo k, NY, USA, 2008.
19. Kleinn, C. Fo es Resou ces In en o ies in Nepal: S a us Quo, Needs, Recommenda ions; FRIS
P ojec Pape No. 1; Fo es Resou ce In o ma ion Sys em P ojec (FRISP), HMGN/FINNIDA,
Finnish Fo es and Pa k Se ice: Ka hmandu, Nepal, 1994.
20. Fo es Resou ce Assessmen o Nepal. D a Field Manual 2010; Ve sion o Oc obe 16, 2010;
Fo es Resou ce Assessmen o Nepal P ojec : Ka hmandu, Nepal, 2010.
21. Google Ea h. A ailable online: h p://www.google.com/ea h/index.h ml (accessed on 26 June
2012).
22. an Laa , A.; Akça, A. Fo es Mensu a ion; Cu illie Ve lag: Gö ingen, Ge many, 1997.
23. Sha ma, E.R.; Pukkala, T. Volume Equa ions and Biomass P edic ion o Fo es T ees o Nepal;
Publica ion 47; Fo es Su ey and S a is ics Di ision, Minis y o Fo es s and Soil Conse a ion:
Ka hmandu, Nepal, 1990.
24. Pinhei o, J.C.; Ba es, D.M. Mixed-E ec s Models in S and S-PLUS; Co ec ed hi d p in ing;
Sp inge -Ve lag: New Yo k, NY, USA, 2002.
25. McRobe s, R.E. Es ima ing o es a ibu e pa ame e s o small a eas using nea es neighbo s
echniques. Fo es Ecol. Manag. 2012, 272, 3–12.
26. Tokola, T. The In luence o Field Sample Da a Loca ion on G owing S ock Volume Es ima ion in
Landsa TM-based Fo es In en o y in Eas e n Finland. Remo e Sens. En i on. 2000, 74, 422–431.
Remo e Sens. 2012, 4 3942
27. Congal on, R.G.; G een, K. Assessing he Accu acy o Remo ely Sensed Da a: P inciples and
P ac ices, 2nd ed.; CRC P ess: Boca Ra on, FL, USA, 2009.
28. Hy önen, P.; An ila, P. Change de ec ion in bo eal o es s using bi- empo al ae ial pho og aphs.
Sil a Fenn. 2006, 40, 303–314.
29. Ka ila, M.; Tomppo, E. Selec ing es ima ion pa ame e s o he Finnish mul isou ce Na ional
Fo es In en o y. Remo e Sens. En i on. 2001, 76, 16–32.
30. Haapanen, R.; Tuominen, S. Da a combina ion and ea u e selec ion o mul i-sou ce o es
in en o y. Pho og amm. Eng. Remo e Sensing 2008, 74, 869–880.
31. R De elopmen Co e Team. R: A Language and En i onmen o S a is ical Compu ing;
R Founda ion o S a is ical Compu ing: Vienna, Aus ia, 2011. A ailable online:
h p://www.R-p ojec .o g/ (accessed on 2 Decembe 2011).
32. Landis, J.R.; Koch, G.G. The measu emen o obse e ag eemen o ca ego ical da a. Biome ics
1977, 33, 159–174.
33. McRobe s, R.E. Diagnos ic ools o nea es neighbo s echniques when used wi h sa elli e
image y. Remo e Sens. En i on. 2009, 113, 489–499.
34. Va jo, J. Change de ec ion and con olling o es in o ma ion using mul i- empo al Landsa TM
image y. Ac a Fo . Fenn. 1997, 258, 1–64.
35. Olsson, H. Reg ession unc ions o mul i empo al ela i e calib a ion o Thema ic Mappe da a
o e bo eal o es . Remo e Sens. En i on. 1993, 46, 89–102.
36. Tokola, T.; Sa keala, J.; an de Linden, M. Use o opog aphic co ec ion in Landsa TM-based
o es in e p e a ion in Nepal. In . J. Remo e Sens. 2001, 22, 551–563.
37. Ka ila, M.; Tomppo, E. Sampling Simula ion on Mul i-Sou ce Ou pu Fo es Maps—An
Applica ion o Small A eas. In P oceedings o 7 h In e na ional Symposium on Spa ial Accu acy
Assessmen in Na u al Resou ces and En i onmen al Sciences, Lisbon, Po ugal, 5–7 July 2006;
pp. 614–623.
38. Tuominen, S.; Ee ikäinen, K.; Schibalski, A.; Haakana, M.; Leh onen, A. Mapping biomass
a iables wi h a mul i-sou ce o es in en o y echnique. Sil a Fenn. 2010, 44, 109–119.
39. Ee ikäinen, K. A mul i a ia e linea mixed-e ec s model o he gene aliza ion o sample ee
heigh s and c own a ios in he Finnish Na ional Fo es In en o y. Fo es Sci. 2009, 55, 480–493.
Appendix: Heigh P edic ion Model
A1.1. Modeling Da a
The sample ee da a a ailable o modeling he ela ionship be ween heigh and diame e
comp ised 1,048 ees measu ed on 182 ield sample plo s wi hin 56 clus e s. The o al numbe o
species was 93, including an “unknown” g oup. Desc ip i e s a is ics o he ee-le el and sample
plo -le el cha ac e is ics used in he modeling a e gi en in Table A1.
Remo e Sens. 2012, 4 3943
Table A1. Desc ip i e s a is ics o he ee-le el and plo -le el cha ac e is ics o he
heigh modeling da a: h is he o al ee heigh (m); zdomina ed and zdominan a e dummy
a iables o he domina ed and dominan ee, espec i ely; G is he s and basal a ea a he
plo -le el (m2∙ha−1); zKS/SK, zSB and zTMH a e plo -le el dummy a iables o he o es ypes
o Khai Sissoo Fo es , Sh ub, and Te ai Mixed Ha dwood Fo es , espec i ely; zSG1–zSG4
a e ee-le el dummy a iables o species g oups 1–4, espec i ely; and d is he ee
diame e a b eas heigh (cm).
Va iable
Minimum
Median
Mean
S anda d De ia ion
Maximum
h
2.00
15.95
16.88
7.671
41.10
zdomina ed
0
0
0.05
0.225
1
zdominan
0
1
0.82
0.385
1
G
0.51
19.93
20.58
7.908
43.31
zKS/SK
0
0
0.02
0.126
1
zSB
0
0
0.00
0.053
1
zTMH
0
1
0.60
0.491
1
zSG1
0
0
0.43
0.495
1
zSG2
0
0
0.13
0.333
1
zSG3
0
0
0.25
0.432
1
zSG4
0
0
0.20
0.400
1
d
5.0
27.4
32.30
21.516
187.9
A1.2. Model Speci ica ion
The sample ee heigh s we e gene alised o co e he ally ees using a mixed modeling echnique
p oposed, o ins ance, by Lappi [A1] and Ee ikäinen [39]. Ins ead o linea model o ms, which we e
used by Lappi [A1] and Ee ikäinen [39], his s udy u ilized a nonlinea , mixed-e ec s modeling
(NLME) p ocedu e (see, e.g., Pinhei o & Ba es [24]). Ee ikäinen [39] es ima ed species-speci ic
e ec s using dummy a iables, which we e ea ed no only as ixed bu also as andom a iables, and
de e mined he species e ec s o Sco s pine, No way sp uce and a g oup o b oadlea ed ees. In
Te ai case, howe e , an au oma ed pooling p ocedu e o g ouping species in o c own laye classes
was cons uc ed. In he pooling p ocedu e, a candida e model localised by clus e s and sample plo s,
espec i ely, was used o de e mine whe he he majo i y o he species-wise obse a ions belonged o
he g oup o heigh s ha we e equal o o g ea e han he model p edic ion o o he g oup o
obse a ions ha we e less han he model p edic ion. A e sepa a ing he heigh da a by species in o
wo c own laye s, he same modeling and so ing p ocedu e was conduc ed o bo h se s o da a,
esul ing in he ac ha he obse a ions we e inally agg ega ed in o ou g oups o consecu i e
c own laye s by species.
The andom componen s o he models we e ela ed o he in e cep e m and he coe icien o he
ee size a iable, i.e., he diame e a b eas heigh , espec i ely, o which he la e e m was also
associa ed wi h he componen s o andom species e ec s (see also Ee ikäinen [39]). An NLME
model o a ee heigh (hijk, m) ha ing a andom in e cep and slope e ms wi h espec o he clus e
Remo e Sens. 2012, 4 3944
(u(1)) and sample plo (u(2)) e ec s as well as he ee species e ec s o he ou c own laye g oups
can now be desc ibed unde condi ions o no mali y (N) as ollows:
))())(
)(
)(
)((exp(
)
)ln(((3.1
,
9
SG4
(2)
8.4,
(1)
8.4,4.8
SG3
(2)
8.3,
(1)
8.3,3.8
SG1
(2)
8.1,
(1)
8.1,1.8
(2)
8,
(1)
8,8
TMH7SB6KS/SK5
4dominan 3domina ed2
(2)
1,
(1)
1,
1
1)
ijkdijk
ijk
iji
ijk
iji
ijk
iji
iji
ij
ijij
ij
ijkijk
iji
ijk
edzuuβ
zuuβ
zuuβ
uuβ
zβzβzβ
Gβzβzββh
β
uu
o which
2
2
)1(
8.4,
)1(
1,
(1)
8.4
(1)
1
(1)
8.4
(1)
8.4
(1)
1
(1)
1
(1)(1) :),(~
u
uu
uu
u
uu
N
u
u
i
i
ΩΩ0
2
2
)2(
8.4,
)2(
1,
(2)
8.4
(2)
1
(2)
8.4
(2)
8.4
(2)
1
(2)
1
(2)(2) :),(~
u
uu
uu
u
uu
N
u
u
ij
ij
ΩΩ0
,
),(~ 2
e
N
ijk
e
0
,
(A1)
whe e
1,
2,...,
8.4,
9 and
d a e ixed model pa ame e s; i, j and k e e o he clus e , plo and ee,
espec i ely;
ijk
zdomina ed
and
ijk
zdominan
a e dummy a iables o he domina ed and dominan ee,
espec i ely;
ij
G
is he s and basal a ea a he plo j o clus e i (m2∙ha−1);
ij
zKS/SK
,
ij
zSB
and
ij
zTMH
a e
dummy a iables o h ee o es ypes, a Khai Sissoo Fo es , a Sh ub Fo es and a Te ai Mixed
Ha dwood Fo es , espec i ely; dijk is he diame e a b eas heigh (cm);
ijk
zSG1
,
ijk
zSG3
and
ijk
zSG4
a e
dummy a iables o he species g oups 1, 3 and 4, espec i ely (no e: when
ijk
zSG1
,
ijk
zSG3
and
ijk
zSG4
a e se o ze o, he p edic ions a e ob ained o species g oup 2 (SG2)); and eijk is he andom e o
e m o he model.
The pa ame e
d in Equa ion (A1) was added o he independen a iable “dijk” in o de o dec ease
he esidual a ia ion among small-sized ees. The selec ion o ixed alues o
d and o he o m
pa ame e
9 was based on he analysis o esiduals as well as he Akaike In o ma ion C i e ion (AIC)
alues (see [24]) and Roo Mean Squa e E o s (RMSEs, see Equa ion (6)) ob ained using di e en
alues o he wo pa ame e s (see also Ee ikäinen e al. [A2]). A combina ion o alues o 10 and 0.65
o he espec i e pa ame e s
d and
9 was ound o p o ide he bes model i in he PSP da a
om Te ai.
The a iance unc ions a e use ul o accoun ing o he wi hin-g oup he e ogeneous a iances
when, o ins ance, he e o a iance inc eases as he ee size inc eases (see [24,A3]). Thus, hey a e
also use ul o imp o ing he con e gence p ope ies o he algo i hms used o he es ima ion o