Deep lea ning models o es ima ing olume and Lo ey's
1
heigh ac oss No dic coun ies using op ical and SAR
2
sa elli e images
3
Zso ia Koma1 Oleg An opo 2 Oli e Ca us3 Jukka Mie inen2 and Johannes B eidenbach1
4
1 No wegian Ins i u e o Bioeconomy (NIBIO), Di ision o Fo es and Fo es Resou ces, Na ional
5
Fo es In en o y, Høgskole eien 7, 1433 Ås, No way
6
2 VTT Technical Resea ch Cen e o Finland, P.O. Box 1000, 02044 Espoo, Finland
7
3Gamma Remo e Sensing, Wo bs asse 225, 3073 Gümligen, Swi ze land
8
Ta ge jou nal: In e na ional Jou nal o Applied Ea h Obse a ion and Geoin o ma ion
9
Highligh s
10
- A la ge-scale, ine- esolu ion o es esou ce map was p oduced in No way om mul isou ce
11
op ical and SAR image y, ained on an ALS-based e e ence map using a UNe deep lea ning
12
amewo k.
13
- The bes UNe model achie ed coe icien s o de e mina ion (R²) o 0.63 and 0.55 o imbe
14
olume and Lo ey’s heigh , espec i ely, compa ed o 0.39 and 0.37 om he e e ence
15
app oach (k-nea es neighbou s wi h k = 7).
16
- We also analyse he geog aphical ans e abili y o he p e- ained model on Finnish da a o
17
ine- une o No wegian condi ions.
18
Abs ac
19
Spa ially explici in o ma ion on o es esou ces and s uc u e is essen ial o sus ainable o es
20
managemen and e idence-based policy-making. In he No dic egion, la ge-scale o es mapping o en
21
elies on in eg a ing Na ional Fo es In en o y (NFI) ield plo s wi h ai bo ne lase scanning (ALS) da a.
22
Howe e , he in equen co e age o na ionwide ALS campaigns limi s hei use o con inuous
23
moni o ing. Sa elli e image y, wi h i s high empo al and spa ial esolu ion, p o ides a p omising
24
al e na i e. We e alua e UNe -based deep lea ning models ained on wall- o-wall ALS-de i ed o es
25
esou ce maps o p edic ing olume and Lo ey’s heigh in No way using op ical (Sen inel-2) and SAR
26
(Sen inel-1, PALSAR-2) da a. The UNe models, ained on bo h Finnish and No wegian ALS maps, a e
27
benchma ked agains kNN models based on No wegian NFI da a. T ans e lea ning is u he explo ed
28
by ine- uning models using No wegian NFI plo s. Model unce ain ies a e assessed using 541 ese ed
29
NFI plo s and 44 independen alida ion s ands mainly including high- olume bo eal o es s (>200 m³
30
ha⁻¹). The bes -pe o ming UNe model, ained on No wegian ALS da a, achie ed an R² o 0.63 o
31
olume and 0.55 o Lo ey’s heigh , consis en ly ou pe o ming kNN. Fine- uning imp o ed model
32
ans e abili y, wi h gains in R² o up o 0.13 o olume and 0.46 o Lo ey’s heigh , when adap ing he
33
Finnish model o No wegian condi ions. U ilizing SAR da a alongside op ical da a enhanced model
34
accu acy, wi h R2 imp o emen anging om 0.03 o 0.19. O e all, ou indings demons a e he
35
po en ial o UNe models ained on wall- o-wall ALS maps, o e ing a ans e able app oach o o es
36
esou ce mapping ac oss No dic coun ies.
37
Keywo ds: UNe , deep lea ning, o es esou ce mapping, sa elli e image y, SAR, mul isou ce
38
39
1. In oduc ion
40
Quan i ying he s a us and ends o o es esou ces and s uc u e—such as imbe olume and Lo ey’s
41
heigh —is c ucial o in o ming e idence-based policy and managemen decisions ac oss local,
42
na ional, and in e na ional le els (Tomppo e al., 2010; Vidal e al., 2016a). Sampling-based Na ional
43
Fo es In en o ies (NFIs) p o ide his in o ma ion a na ional and egional scales, o ming he
44
ounda ion o s a egic decision and policy-making ega ding key ecosys em se ices, including
45
bioene gy po en ial, biodi e si y, and changes in ca bon s ock (Vidal e al., 2016b; B eidenbach e al.,
46
2020). Howe e , because NFIs a e sampling-based su eys, hei capaci y o p oduce eliable es ima es
47
a local scales is es ic ed (B eidenbach and As up, 2012; As up e al., 2019). To b idge his gap, NFI
48
ield plo s a e o en combined wi h emo ely sensed da a o model ela ionships be ween ield-
49
obse ed o es a ibu es and emo e sensing-based p edic o a iables, enabling mapping be ween
50
NFI plo loca ions (McRobe s and Tomppo, 2007; Kangas e al., 2018). This app oach aligns wi h he
51
equi emen s o se e al in e na ional policy amewo ks—including he EU Fo es S a egy o 2030,
52
he EU Biodi e si y S a egy, and he EU De o es a ion Regula ion—which emphasize he need o
53
spa ially explici o es moni o ing a bo h na ional and local scales. Consequen ly, in eg a ing emo ely
54
sensed da a has become essen ial o p oducing de ailed o es esou ce maps and in o ma ion.
55
Some o he ea lies na ional o es maps we e based on a combina ion o NFI and Landsa da a
56
(Gje sen, 2007; McRobe s and Tomppo, 2007). In many coun ies, na ional o es esou ce maps a e
57
oday based on in eg a ing ield plo s om NFIs wi h coun y-wide Ai bo ne Lase Scanning (ALS) da a
58
(Kangas e al., 2018). ALS o e s de ailed 3D in o ma ion on o es s uc u e, enabling accu a e
59
es ima ion o key a ibu es such as imbe olume and Lo ey’s heigh . Howe e , na ional ALS
60
campaigns lack he empo al equency needed o e ec i e o es moni o ing. Fo example, No way’s
61
na ionwide ALS campaign equi ed app oxima ely en yea s o comple e ull co e age (Hauglin e al.,
62
2021), compa ed o a se en-yea ime ame in Sweden (Nilsson e al., 2017). In Finland, he u u e ALS
63
campaigns a e planned o be conduc ed in nine-yea cycles (NLS, 2025). In con as , Ea h Obse a ion
64
(EO) da ase s—such as op ical image y om Sen inel-2 and Syn he ic Ape u e Rada (SAR) da a om
65
Sen inel-1 o ALOS-2 PALSAR-2—o e bo h high empo al and spa ial co e age. Se e al s udies ha e
66
u ilized op ical emo e sensing o o es esou ce mapping, such as Sen inel-2 o LANDSAT (As ola e
67
al., 2019; Lang e al., 2019; Mie inen e al., 2021) o ha e combined EO wi h ALS da a (Es eban e al.,
68
2019; As ola e al., 2021). SAR da a wi h clea sensi i i y o o es s uc u e ha e also been used o
69
o es mapping, p ima ily a s and-le el o coa se mapping uni s and in wide-a ea s udies (San o o e
70
al., 2011; An opo e al., 2017; Ge e al., 2023b). Recen s udies ha e shown ha he combina ion o
71
op ical and SAR da a imp o es he quan i ica ion o o es s uc u al cha ac e is ics (Wi ke e al., 2019;
72
Pe sson e al., 2021; Ge e al., 2022, 2023a), especially i SAR ope a es in lowe equencies (e.g., L-
73
band used on boa d ALOS-2), which opens new possibili ies o imp o ing he accu acy o maps o
74
es ima ing o es a ibu es.
75
EO-based o es esou ce maps ha e been c ea ed using s a is ical me hods such as linea eg ession,
76
k-nea es neighbou s (kNN), and ensemble-based machine lea ning echniques such as andom o es s.
77
S udies using only Sen inel-2 da a wi h hese app oaches ha e epo ed RMSEs anging om 27.2% o
78
59.3% o es ima es o olume and Lo ey’s heigh in bo eal o es con ex s (As ola e al., 2019;
79
Mie inen e al., 2021; Pe sson e al., 2021). When using SAR da a alone, RMSEs o olume ange om
80
32% o 54.3% (Wi ke e al., 2019; Ge e al., 2023b; Pe sson e al., 2021). One s udy sys ema ically
81
es ed he in eg a ion o op ical and SAR da ase s using kNN me hod and epo ed imp o ed accu acies
82
in Sweden, wi h olume RMSE educed om 37.9% o 30.2% in bo eal o es s (Pe sson e al., 2021).
83
Howe e , a ecu ing limi a ion is, o ins ance, a sa u a ion e ec in imbe olume p edic ions—
84
pa icula ly abo e 300 m³/ha (San o o e al., 2011) —which con ibu es o lowe o e all accu acy
85
compa ed o models based on ai bo ne lase scanning (ALS) da a (Wi ke e al., 2019). Recen ad ances
86
in Deep Lea ning (DL) show p omise o imp o ing EO-based o es esou ce mapping. Ge e al. (2023a)
87
demons a ed ha he p edic ion o Lo ey’s heigh using Sen inel-2 da a alone, as well as combined
88
Sen inel-2 and SAR da a, imp o ed by 0.3% o 1.6% RMSE when applying he UNe DL me hod
89
compa ed o adi ional app oaches such as kNN, andom o es , o linea eg ession. Simila ly, As ola
90
e al. (2021) epo ed RMSE imp o emen s o 7.9% o olume and 2.5% o heigh when using DL
91
models compa ed o e e ence me hods. Ne e heless, DL emains unde u ilized in eg ession asks o
92
EO-based o es a ibu e es ima ion, especially when compa ed o i s mo e widesp ead applica ion in
93
classi ica ion asks such as dis u bance mapping (Yuan e al., 2020).
94
One o he key challenges in applying DL o o es esou ce mapping is he limi ed a ailabili y and
95
quali y o sui able aining da a (Ball e al., 2017; Li e al., 2019). NFI da ase s a e o en conside ed weak
96
labels in he con ex o DL because hey lack spa ial con ex due o a ela i ely small plo size. To add ess
97
his challenge, ans e lea ning has eme ged as a p omising s a egy in a ious emo e sensing asks
98
e.g. (Wu m e al., 2019; Zhou e al., 2023; Kuzu e al., 2024), bu only a ew s udies ha e been used o
99
in he con ex o o es esou ce mapping (As ola e al., 2021; Ge e al., 2023a). Ins ead o aining a
100
model solely on da a om he a ge egion, ans e lea ning allows he use o a p e- ained model—
101
de eloped on a ela ed ask —which can hen be ine- uned using a smalle se o local aining da a.
102
This app oach le e ages he gene alizable ea u es lea ned du ing p e- aining, enabling e ec i e
103
model adap a ion e en wi h limi ed labelled da a. In addi ion, ALS-based o es esou ce maps can
104
o e high-quali y, wall- o-wall aining da a ha can complemen o subs i u e adi ional plo -based
105
labels (Ge e al., 2022, 2023a). Fo example, Ge e al. (2023a) demons a ed he success ul aining o
106
models in Finnish Lapland using ALS-de i ed maps, along wi h op ical and SAR sa elli e image y,
107
ollowed by ine- uning wi h plo -le el measu emen da a in sou he n Finland. To u he e alua e he
108
e ec i eness o his app oach, applying ans e lea ning o neighbou ing coun ies—such as No way—
109
p o ides an ideal es case.
110
In his s udy, we de elop and e alua e he po en ial o UNe -based deep lea ning models in mul iple
111
EO da a combina ions o es ima e imbe olume and Lo ey’s heigh using op ical (Sen inel-2) and SAR
112
(Sen inel-1 and PALSAR-2) sa elli e image y. ALS-based o es esou ce maps se e as wall- o-wall
113
aining da a o models ained in bo h Finland and No way. We in es iga e ans e lea ning by ine-
114
uning ALS-based p e- ained models wi h No wegian NFI plo s, bo h by ans e ing he Finnish model
115
o No wegian condi ions and explo ing u he op ions o imp o ing he No wegian model. The
116
pe o mance and unce ain y o he UNe models ained on exis ing o es esou ce maps a e
117
compa ed wi h kNN models ained on NFI ield plo s. Model unce ain y is assessed using bo h a es
118
da ase o NFI plo s, which was no used in modelling, and an independen alida ion da ase o o es
119
s ands. Ou esea ch ques ions a e: (1) Which modelling me hod -- UNe o kNN -- achie es he highes
120
accu acy? (2) Can ans e lea ning, whe e NFI ield plo s a e inco po a ed in he model, imp o e he
121
UNe accu acy, especially in he case o applying he Finnish model o No wegian condi ions? (3) Can
122
he inclusion o SAR image y enhance he model’s p edic ion accu acy?
123
2. Ma e ials and Me hods
124
The o e all wo k low de eloped o his s udy (Fig. 1) consis s o ou main p ocessing s eps: (a) kNN
125
modelling; (b) p e aining UNe models using ALS-based o es esou ce maps om Finland and
126
No way; (c) ine- uning UNe models om Finland and No way o he No wegian s udy si e; and (d)
127
assessing he unce ain y associa ed wi h each model ype (k-NN, p e ained UNe , and ine uned
128
UNe ).
129
130
Figu e 1. O e iew o he wo k low used o assessing he accu acy o di e en models (kNN, UNe , ine uned UNe ). The
131
inpu da ase s a e ma ked wi h ec angles (sa elli e image y da a, o es esou ce maps, No wegian NFI plo s, and
132
No wegian alida ion s ands), da k g ey indica ing inpu o aining he models and ligh g een o alida ion models. The
133
in e media e models (kNN, UNe , and ine uned UNe ) a e ma ked wi h a snipped ec angle. The main p ocessing s eps a e
134
shown wi h ounded ec angles (a-d co esponding o he main ex ). V is olume, and H is Lo ey’s heigh , S2 is Sen inel-2
135
op ical images, S1 is Sen inel-1 SAR images, and P2 is PALSAR-2 SAR images. FI = Finland and NO = No way.
136
2.1. Da ase s
137
2.1.1. Sa elli e image y da a and p e-p ocessing
138
ESA Cope nicus Sen inel-2 sa elli es ope a ing wi h he Mul i-Spec al Ins umen (MSI) p o ide op ical
139
image y da a on 13 spec al bands a a 2–3-day e isi equency in he No dic coun ies, in isible,
140
nea -in a ed, and he sho -wa e in a ed pa o he elec omagne ic spec um. In his s udy, we used
141
he Le el-2 su ace e lec ance p oduc s p o ided by ESA a he isible and nea in a ed bands (B2, B3,
142
B4, B5, B8) and sho -wa e in a ed bands (B11, B12). The 20 m esolu ion bands (B5, B11, B12) we e
143
esampled in o 10 m esolu ion using he bilinea in e pola ion me hod o ma ching he esolu ion
144
wi h he o he bands (10 m). The Sen inel-2 da a we e u he p ocessed wi h he Te amoni o
145
composi ing sc ip (Mie inen e al., 2021), p oducing g owing season composi es by weigh ing based
146
on cloudiness, haze, and shadows. The composi ing p ocess also deli e ed a quali y lag ha was used
147
o il e ing cloud-a ec ed o bad-quali y pixels.
148
In his s udy, we ha e used wo ypes o SAR da ase s. The i s consis ed o C-band backsca e image y
149
acqui ed by Sen inel-1A in In e e ome ic Wide (IW) swa h mode, using VV and VH pola isa ion. The
150
images in G ound-Range-De ec ed (GRD) o ma we e p e-p ocessed using he GAMMA Remo e
151
Sensing so wa e (We ne e al., 2000), which esul ed in geocoded and adiome ic e ain co ec ed
152
g0 backsca e images. The images we e subsequen ly esampled o 10 m o ma ch he Sen inel-2
153
image y. The Sen inel-1 da a acqui ed wi hin a yea (30 o 31 images) we e hen used o calcula e he
154
annual mean backsca e pe pola iza ion. Maps depic ing a eas o layo e o shadow we e applied o
155
mask such a eas.
156
The second ype o SAR da ase was based on he Japan Ae ospace Explo a ion Agency (JAXA), which
157
eleases ee and open annual mosaics o he L-band da a acqui ed by he Phased-A ay ype L-band
158
SAR (PALSAR-2) onboa d he Ad anced Land Obse ing Sa elli e-2 (ALOS-2).The backsca e mosaics,
159
wi h a esolu ion o ~ 25 × 25 m², a e p oduced om image y acqui ed in Fine Beam Dual Pola iza ion
160
mode (HH and HV), using s ipmap mode unde JAXA’s global basic obse a ion scena io. The images
161
used o gene a e he mosaics we e adiome ic e ain-co ec ed, and geocoded o a geog aphic spa ial
162
e e ence sys em, as wi h (Shimada and Oh aki, 2010). In his s udy, he HH and HV mosaics we e
163
esampled o a 10 m g id and masked o layo e and shadow using he layo e /shadow maps p oduced
164
by JAXA.
165
All sa elli e image y da ase s we e p ocessed o he s udy a eas in Finland and No way. In Finland, he
166
da ase s we e p ocessed o he yea 2020, and in No way o 2021, o align as closely as possible wi h
167
he iming o he ALS-based o es esou ce maps.
168
2.1.2. Finnish e e ence da ase s
169
The Finnish UNe model was ained based on he wall- o-wall Finnish ALS-based o es esou ce map
170
p oduced by he Finnish Fo es Cen e using ALS acquisi ions om 2020 and associa ed ield in en o y
171
plo s as aining o k-NN based app oaches (Mal amo and Packalen, 2014). The map co e s he co e
172
cen al and sou he n egions o Finland (Fig. 2), cap u ing bo eal o es s in 16 m spa ial esolu ion. The
173
s udy a ea has ela i ely small ele a ion di e ences, wi h he highes poin s gene ally less han 300 m
174
abo e sea le el. The o es in he e e ence a eas has a ound 40% pine, mo e han 30% sp uce, and he
175
es deciduous ees. The a e age o es olume ac oss he s udy si es was 120 m3 ha-1, wi h an a e age
176
heigh o 14 m.
177
2.1.3. No wegian e e ence da ase s
178
The No wegian s udy si e co e s app oxima ely 210,000 km² o land, encompassing he Sou h-Eas e n,
179
Cen al, Sou he n, and Wes e n egions, ex ending om he coun y o T øndelag and sou h (Fig. 2). I
180
is loca ed in he bo eal clima e zone and con ains mos o he p oduc i e o es o No way. The bo eal
181
o es s a e domina ed by h ee main species: No way sp uce, Sco s pine, and deciduous ees, p ima ily
182
bi ch, which a e o en mixed wi h coni e ous species. The s udy si e has a di e se opog aphical
183
a ia ion.
184
185
Figu e 2 The No wegian s udy a ea (highligh ed in g een on he No dic map) is whe e he di e en model ypes—kNN, UNe ,
186
and UNe ine- uned —we e e alua ed. G een do s ep esen he NFI plo s used o es ing, while pink colou ed do s indica e
187
he aining NFI plo s used o he kNN model and UNe ine- uning. B own a eas ma k he independen alida ion s ands.
188
Panel A shows a zoomed-in iew o he s udy a ea, wi h panels B and C highligh ing wo example alida ion s ands used in
189
he accu acy assessmen . The backg ound maps a e he ALS-based imbe olume map om Kilden (NIBIO, 2025) and
190
Google sa elli e image y in panel A.
191
Di e en ield obse a ion da ase s we e u ilized in his s udy (Fig.2, Table 1). Fi s , he No wegian NFI
192
is a con inuously ope a ing in en o y based on a pe manen sample g id, whe e each ield plo is
193
emeasu ed e e y i e yea s (B eidenbach e al., 2020). The NFI ield measu emen s a e ca ied ou
194
wi hin ci cula 250 m2 plo s whe e he diame e a b eas heigh (dbh) and species o all ees wi h a
195
dbh >= 5 cm a e eco ded, and he heigh o en ees is measu ed o calcula e imbe olume and
196
Lo ey’s heigh . Only NFI plo s measu ed in 2021 wi hin a alid emo e sensing-based o es mask we e
197
included in he s udy, which esul ed in 1566 NFI plo s. 541 NFI plo s we e se aside o es ing and no
198
used o kNN o DL modelling, and he emaining 1025 NFI plo s we e used in model aining (Table 1).
199
Addi ional s and-le el ield measu emen s (Fig. 2, Table 1), independen o he NFI, we e used o
200
alida ion (Koma and B eidenbach, 2025). The da ase includes wo subse s: (a) ma u e s ands in Aske
201
(eas e n egion) and Al e (wes e n egion), whe e 10–15 plo s pe s and we e sampled ollowing NFI
202
p o ocols. Mean imbe olume and Lo ey’s heigh we e calcula ed o cha ac e ise each s and. (b) Long-
203
e m ield ials assessing o es managemen me hods, consis ing o 1–30 adjacen ec angula s ands.
204
Wi hin each s and, dbh and species we e eco ded o all ees, and heigh was measu ed o e e y
205
ou h ee. S and-le el imbe olume and Lo ey’s heigh we e calcula ed and agg ega ed o ial-le el
206
a e ages. A e applying a emo e sensing-based o es mask and excluding cloud-co e ed s ands om
207
2021, 44 alida ion s ands co e ing 62 ha we e a ailable o unce ain y assessmen .
208
Table 1 Summa y s a is ics o aining NFI, es ing NFI, and alida ion s and da ase s. The able includes he numbe o
209
obse a ions (N), minimum (Min), maximum (Max), mean, and s anda d de ia ion (SD) o each a iable (Volume, Lo ey’s
210
heigh , a ea o he plo , and he s ands).
211
Da ase
Va iable
Uni
N
Min
Max
Mean
SD
T aining NFI plo s
Volume
m3ha-1
1025
0.18
1157.59
118.13
127.99
Lo ey's heigh
m
2.40
28.18
11.96
5.11
A ea
m2
250
Tes ing NFI plo s
Volume
m3ha-1
541
0.30
779.70
125.91
118.66
Lo ey's heigh
m
3.40
28.00
12.38
4.66
A ea
m2
250
Valida ion s ands
Volume
m3ha-1
44
100.62
906.70
369.14
206.17
Lo ey's heigh
m
12.24
26.55
18.58
4.11
A ea
ha
0.11
4.32
1.40
1.06
212
Thi d, he No wegian Fo es Resou ce Map (SR16) was used as a wall- o-wall aining da ase o he
213
UNe model. These spa ially con inuous maps p o ide es ima es o imbe olume, Lo ey’s heigh , and
214
o he o es a ibu es a a 16 m × 16 m esolu ion and a e ope a ionally upda ed each yea ollowing
215
he wo k low desc ibed by Hauglin e al. (2021). They a e p oduced by combining NFI plo da a wi h
216
p edic o a iables de i ed om na ionwide ALS acquisi ions conduc ed by he No wegian Mapping
217
Agency. Species-speci ic uni a ia e linea mixed-e ec s eg ession models a e used o accoun o
218
sys ema ic di e ences ac oss ALS p ojec s and egions. To ensu e consis ency, NFI plo alues a e
219
p ojec ed o a common e e ence da e, and a emo e sensing-based o es mask is applied.
220
2.2. Modelling
221
2.2.1. kNN modeling
222
The kNN me hod has been chosen as he baseline me hod o be compa ed wi h he DL me hod, since
223
i has been success ully applied in es ima ing o es a iables and i is ope a ionally used, o example,
224
in Finland (McRobe s and Tomppo, 2007). I is a non-pa ame ic algo i hm, whe e he p edic ions o
225
he a ge a iables a e ob ained om as he a e ages o he k neighbou s wi h he smalles dis ances
226
o he a ge uni in he auxilia y space. The kNN model can be desc ibed as ollows:
227
𝑦𝑝=∑𝑤𝑖,𝑝
𝑖 𝜖 𝐼 𝑦𝑖
228
whe e he p edic ed ec o o each pixel is 𝑦𝑝 , 𝑦𝑖 is he ec o o obse a ions o he 𝑖- h con ibu ing
229
uni in he e e ence se , 𝐼 is he subse o con ibu ing uni s ha a e nea es wi h espec o he
230
dis ance me ic, and 𝑤𝑖,𝑝 is he weigh o i- h con ibu ing uni calcula ed as
231
𝑤𝑖,𝑝= {𝑑𝑖,𝑝
−𝑡
∑𝑑𝑖,𝑝
−𝑡
𝑖 𝜖 𝐼
0 𝑖 𝜖 𝐼
232
whe e 𝑡>0 weigh uni s in e sely o hei ea u e space dis ance o dis ance squa ed om pixel 𝑝. -
233
In his s udy, 𝑡 = 1 was used, wi h Euclidean dis ance and se en nea es neighbou s (k=7).
234
The kNN aining da abase was cons uc ed by in e sec ing aining NFI plo loca ions (Fig.2, Table 1)
235
wi h he p e-p ocessed EO da ase s, including Sen inel-2, Sen inel-1, and PALSAR-2 bands. Fo each
236
plo , all pixels o e lapping he 250 m² plo a ea we e conside ed, and hei con ibu ions we e
237
weigh ed by he ac ion o pixel a ea alling wi hin he plo . A weigh ed mean o he in e sec ing pixel
238
alues was hen compu ed, esul ing in a aining able ha combined ield-measu ed a ibu es wi h
239
co esponding EO ea u es. P edic ions we e gene a ed o each es NFI plo and alida ion s and
240
loca ions.
241
2.2.2. UNe model
242
A a ian o a UNe model (Ronnebe ge e al., 2015) has been adap ed o his s udy, which is a
243
con olu ional neu al ne wo k (CNN) widely used in biomedical image and o he seman ic segmen a ion
244
asks. Compa ed o a basic CNN, UNe models allow ex ac ion o deepe ea u e ep esen a ions om
245
he inpu image da ase s and keep he ea u e map size unchanged, which makes i sui able o pixel-
246
le el eg ession asks.
247
The UNe a chi ec u e applied (Fig.3.) in his s udy is a symme ic encode –decode ne wo k designed
248
o pixel-le el p edic ions. The encode con ains a double-con olu ion s uc u e, which includes a 2D
249
con olu ion, ba ch no maliza ion, and ReLU ac i a ion. A each s ep, 2×2 max-pooling hal es he
250
ea u e map. This allows he model o g adually ansi ion om cap u ing ine, local de ails o lea ning
251
mo e global spa ial pa e ns. The decode mi o s his p ocess by es o ing he o iginal esolu ion
252
h ough upsampling ope a ions. To eco e in o ma ion los du ing downsampling, skip connec ions
253
link co esponding encode and decode laye s, as he pooling ope a ion disca ded some de ails.
254
Applying skip-connec ion, he shallow ea u e maps a e conca ena ed o deep ea u es eco e ed om
255
up-sampling. Finally, a 1×1 con olu ion p ojec s he es o ed ea u e map in o he a ge space,
256
p oducing he desi ed pixel-wise p edic ions wi hou al e ing he spa ial dimensions.
257
258
259
Figu e 3 The UNe model a chi ec u e used in he s udy. Each box ep esen s a ea u e map wi h i s size indica ed nea by.
260
2.2.3. UNe model aining
261
Fi s , all inpu da ase s we e p epa ed o aining he UNe model (Fig. 4). Du ing p e- aining, he
262
Finnish and No wegian wall- o-wall ALS-based o es esou ce maps, desc ibed in Sec ions 2.1.2–2.1.3,
263
we e used as e e ence da ase s. The p e- aining da ase comp ised 1,068 ec angula pa ches o
264
Finland and 1,025 pa ches o No way, each wi h a size o 2.56 × 2.56 km² (256 × 256 pixels). These
265
pa ches we e andomly dis ibu ed ac oss he o es ed a eas co e ed by he ALS-based o es esou ce
266
maps. The esolu ion o he o es esou ce maps was aligned wi h he sa elli e image y da ase s by
267
esampling he na i e spa ial esolu ion o 16 × 16 m o 10 × 10 m. Fo ine- uning, 1,025 No wegian
268
NFI plo s selec ed o aining we e used. When p epa ing image pa ches ela ed o he aining NFI
269
plo s, only he 10 × 10 m pixels in e sec ing he ield plo loca ions con ained alid alues; all o he
270
pixels we e masked ou (se o no-da a).
271
P edic o a iables we e ex ac ed o he selec ed image pa ches ( o bo h p e- aining and ine-
272
uning) based on he p e-p ocessed and il e ed EO bands om Sen inel-2 only (S2), and om he
273
mul isou ce Sen inel-2, Sen inel-1, and PALSAR-2 da ase s (S2S1P2), as desc ibed in Sec ion 2.1.1. Fo
274
independen alida ion, EO image pa ches ela ed o he NFI es se —which included 541 ield plo s
275
se aside a he beginning o he modelling p ocess—and alida ion s ands we e p epa ed using he
276
same pa ch size, con aining he ele an plo s and s ands. Wi hin all image pa ches, non- o es pixels
277
we e masked ou . All EO pa ches we e u he no malised on a pe -band basis using z-sco e
278
no maliza ion, whe e each band’s mean was sub ac ed and he esul di ided by i s s anda d
279
de ia ion.
280
281
Figu e 4 O e iew o he UNe model aining wo k low. Inpu pa ches a e p epa ed o he Finnish and No wegian da ase s.
282
An independen es NFI plo s we e se aside be o e model aining and used only o inal unce ain y assessmen . Du ing
283
UNe p e- aining and ine- uning, he o es esou ce maps and aining NFI plo s used we e andomly spli 10 imes in o
284
aining, alida ion and es subse s. The in e nal alida ion and es a e used solely o un he DL model and a e no used o
285
unce ain y assessmen in his s udy.
286
Du ing UNe model p e- aining, he o e all da ase s we e andomly spli in o aining (70%), alida ion
287
(15%), and es ing (15%) pa ches. The Adam op imise was used wi h an ini ial lea ning a e o 1×10⁻⁴.
288
A ReduceLROnPla eau schedule was applied o adap i ely educe he lea ning a e by hal when he
289
alida ion loss pla eaued, wi h a pa ience o en epochs. The model was ained o 200 epochs wi h a
290
mild weigh decay o 1×10⁻⁵. The bes -pe o ming model was selec ed based on he lowes alida ion
291
loss, de e mined using he in e nal alida ion da a, and he co esponding weigh s we e sa ed o
292
subsequen ine- uning o each a ge a iable. T aining was conduc ed wi h a ba ch size o 16. The
293
inal model was used o p edic EO pa ches co esponding o he es NFI plo loca ions and alida ion
294
s ands.
295
esul s also showed ha in mos cases bias was educed, whe eas models wi hou ine- unning o en
439
exhibi ed highe bias, which aligns wi h Ge e al., (2023a).
440
Ou esul s indica e ha combining op ical and SAR p edic o s gene ally imp o ed model pe o mance
441
compa ed o op ical-only inpu s (Tables 2–3). Simila ends ha e been epo ed in o he bo eal o es
442
s udies. Fo example, Pe sson e al. (2021) demons a ed ha including SAR da a educed he RMSE
443
o olume es ima es in Sweden om 37.2% o 30.2%. Likewise, Ge e al. (2022, 2023a) epo ed an
444
imp o emen in Lo ey’s heigh p edic ion in Finland, wi h RMSE dec easing om 30.3% o 18.1%. In
445
ou s udy, he magni ude o imp o emen was smalle : he maximum educ ion in RMSE was om
446
61.3% o 57.1% o olume and om 26.4% o 25.3% o Lo ey’s heigh . O e all, we also ound ha he
447
addi ion o SAR, can imp o e cap u ing o es s uc u e, which aligns wi h se e al s udies (Wi ke e al.,
448
2019; Pe sson e al., 2021; Ge e al., 2022, 2023a). Two main ac o s can explain he mino
449
imp o emen in RMSE. Fi s , he mo e complex opog aphy in No way compa ed o Sweden o Finland
450
can in luence he SAR backsca e signal quali y. Second, Pe sson e al. (2021) u ilized TanDEM-X da a,
451
which o e s a ine spa ial esolu ion (10 m) and a mo e equen acquisi ion cycle (11-day epea
452
cycle), and, mos impo an ly, in e e ome ic cohe ence ha is inhe en ly sensi i e o e ical
453
s uc u e o o es s (Olesk e al., 2016). In con as , ou s udy also elied on PALSAR-2 da a addi ionally
454
o S1 wi h coa se esolu ion (25 m) and used a single annual mosaic.
455
5. Conclusion
456
Ou s udy demons a es he po en ial o using spa ially con inuous ALS-based o es esou ce maps as
457
aining da a o UNe -based deep lea ning models. In con as o mos p e ious bo eal o es mapping
458
app oaches—which mos commonly ely on end- o-end aining using ield plo s o speci ied sizes and
459
sa elli e image y wi hin a single coun y—we show ha ALS-de i ed maps can se e as an e ec i e
460
aining da ase ha also le e ages spa ial con ex and can be applied ac oss di e en No dic coun ies.
461
We ound ha deep lea ning models ou pe o m kNN me hod by educing o e es ima ion in low-
462
olume (<200 m³ ha⁻¹) o es s and unde es ima ion in high- olume (>200 m³ ha⁻¹) o es s. A key
463
ad an age o deep lea ning is i s ans e abili y wi h a limi ed se o local ield plo s. Ou esul s indica e
464
ha a model p e- ained on Finnish da a pe o med compa ably o kNN ained wi h No wegian da a.
465
Wi h he de elopmen o mo e comp ehensi e ounda ion models ained on mul i-coun y da ase s,
466
his app oach may enable obus , la ge-scale, c oss-coun y o es esou ce mapping.
467
Acknowledgmen s
468
This esea ch was unded by he Eu opean Space Agency (ESA) p ojec on Fo es Ca bon Moni o ing
469
unde con ac numbe 4000135015/21/I-NB—Fo es Ca bon Moni o ing, and by he EU unde
470
GA101056907 (Pa hFinde ). The sa elli e da a p ocessing was suppo ed by he ESA Ne wo k o
471
Resou ces Ini ia i e. Mic oso Copilo assis ed in enhancing he cla i y and co ec ness o he English
472
ex .
473
Au ho con ibu ions
474
ZK: Concep ualiza ion, Da a cu a ion, Fo mal analysis, Valida ion, W i ing o iginal d a , OA:
475
Concep ualiza ion, Da a cu a ion, Fo mal analysis, Me hodology, W i ing e iew and edi ing, OC: Da a
476
cu a ion, W i ing e iew and edi ing, JM: Fo mal analysis, P ojec adminis a ion, Funding acquisi ion,
477
W i ing e iew and edi ing, JB: P ojec adminis a ion, Funding acquisi ion, W i ing e iew and edi ing
478
Da a and code s a emen
479
The p e-p ocessed EO da ase is a ailable upon eques om he co esponding au ho . The used
480
me hods a e a ailable wi hin he Fo es y-TEP (h ps:// - ep.com/) de eloped by VTT.
481
Re e ences
482
An opo O, Raus e Y, Häme T, P aks J. 2017. Pola ime ic ALOS PALSAR Time Se ies in Mapping
483
Biomass o Bo eal Fo es s. Remo e Sensing 9 : 999. DOI: 10.3390/ s9100999
484
As ola H, Häme T, Si o L, Molinie M, Kilpi J. 2019. Compa ison o Sen inel-2 and Landsa 8 image y
485
o o es a iable p edic ion in bo eal egion. Remo e Sensing o En i onmen 223 : 257–273. DOI:
486
10.1016/j. se.2019.01.019
487
As ola H, Sei sonen L, Halme E, Molinie M, Lönnq is A. 2021. Deep Neu al Ne wo ks wi h T ans e
488
Lea ning o Fo es Va iable Es ima ion Using Sen inel-2 Image y in Bo eal Fo es . Remo e Sensing 13
489
: 2392. DOI: 10.3390/ s13122392
490
As up R, Rahl J, Bjø kelo K, Debella-Gilo M, Gje sen A-K, B eidenbach J. 2019. Fo es in o ma ion a
491
mul iple scales: de elopmen , e alua ion and applica ion o he No wegian o es esou ces map
492
SR16. Scandina ian Jou nal o Fo es Resea ch 34 : 484–496. DOI: 10.1080/02827581.2019.1588989
493
Ball JE, Ande son DT, S CSC. 2017. Comp ehensi e su ey o deep lea ning in emo e sensing:
494
heo ies, ools, and challenges o he communi y. Jou nal o Applied Remo e Sensing 11 : 042609.
495
DOI: 10.1117/1.JRS.11.042609
496
Becke A, Russo S, Puli i S, Lang N, Schindle K, Wegne JD. 2023. Coun y-wide e ie al o o es
497
s uc u e om op ical and SAR sa elli e image y wi h deep ensembles. ISPRS Jou nal o
498
Pho og amme y and Remo e Sensing 195 : 269–286. DOI: 10.1016/j.isp sjp s.2022.11.011
499
B eidenbach J, As up R. 2012. Small a ea es ima ion o o es a ibu es in he No wegian Na ional
500
Fo es In en o y. Eu opean Jou nal o Fo es Resea ch 131 : 1255–1267. DOI: 10.1007/s10342-012-
501
0596-7
502
B eidenbach J, G anhus A, Hylen G, E iksen R, As up R. 2020. A cen u y o Na ional Fo es In en o y
503
in No way – in o ming pas , p esen , and u u e decisions. Fo es Ecosys ems 7 : 46. DOI:
504
10.1186/s40663-020-00261-0
505
Es eban J, McRobe s RE, Fe nández-Landa A, Tomé JL, Nӕsse E. 2019. Es ima ing Fo es Volume and
506
Biomass and Thei Changes Using Random Fo es s and Remo ely Sensed Da a. Remo e Sensing 11 :
507
1944. DOI: 10.3390/ s11161944
508
Ge S, An opo O, Häme T, McRobe s RE, Mie inen J. 2023a. Deep Lea ning Model T ans e in Fo es
509
Mapping Using Mul i-Sou ce Sa elli e SAR and Op ical Images. Remo e Sensing 15 : 5152. DOI:
510
10.3390/ s15215152
511
Ge S, Gu H, Su W, P aks J, An opo O. 2022. Imp o ed Semisupe ised UNe Deep Lea ning Model o
512
Fo es Heigh Mapping Wi h Sa elli e SAR and Op ical Da a. IEEE Jou nal o Selec ed Topics in Applied
513
Ea h Obse a ions and Remo e Sensing 15 : 5776–5787. DOI: 10.1109/JSTARS.2022.3188201
514
Ge S, Tomppo E, Raus e Y, McRobe s RE, P aks J, Gu H, Su W, An opo O. 2023b. Sen inel-1 Time
515
Se ies o P edic ing G owing S ock Volume o Bo eal Fo es : Mul i empo al Analysis and Fea u e
516
Selec ion. Remo e Sensing 15 : 3489. DOI: 10.3390/ s15143489
517
Gje sen AK. 2007. Accu acy o o es mapping based on Landsa TM da a and a kNN-based me hod.
518
Remo e Sensing o En i onmen 110 : 420–430. DOI: 10.1016/j. se.2006.08.018
519
Hauglin M, Rahl J, Schumache J, As up R, B eidenbach J. 2021. La ge scale mapping o o es
520
a ibu es using he e ogeneous se s o ai bo ne lase scanning and Na ional Fo es In en o y da a.
521
Fo es Ecosys ems 8 : 65. DOI: 10.1186/s40663-021-00338-4
522
Kangas A e al. 2018. Remo e sensing and o es in en o ies in No dic coun ies – oadmap o he
523
u u e. Scandina ian Jou nal o Fo es Resea ch 33 : 397–412. DOI: 10.1080/02827581.2017.1416666
524
Ka enbo n T, Lei lo J, Schie e F, Hinz S. 2021. Re iew on Con olu ional Neu al Ne wo ks (CNN) in
525
ege a ion emo e sensing. ISPRS Jou nal o Pho og amme y and Remo e Sensing 173 : 24–49. DOI:
526
10.1016/j.isp sjp s.2020.12.010
527
Koma Z, B eidenbach J. 2025. La ge-scale alida ion o o es a ibu e maps ac oss di e en spa ial
528
esolu ions. Sil a Fennica 59 [online] A ailable om: h ps://www.sil a ennica. i/a icle/24061
529
(Accessed 16 Sep embe 2025)
530
Kuzu RS, An opo O, Molinie M, Dumi u CO, Saha S, Zhu XX. 2024. Fo es Dis u bance De ec ion ia
531
Sel -Supe ised and T ans e Lea ning Wi h Sen inel-1&2 Images. IEEE Jou nal o Selec ed Topics in
532
Applied Ea h Obse a ions and Remo e Sensing 17 : 4751–4767. DOI: 10.1109/JSTARS.2024.3361183
533
Lang N, Schindle K, Wegne JD. 2019. Coun y-wide high- esolu ion ege a ion heigh mapping wi h
534
Sen inel-2. Remo e Sensing o En i onmen 233 : 111347. DOI: 10.1016/j. se.2019.111347
535
Li J, Huang X, Gong J. 2019. Deep neu al ne wo k o emo e-sensing image in e p e a ion: s a us and
536
pe spec i es. Na ional Science Re iew 6 : 1082–1086. DOI: 10.1093/ns /nwz058
537
Mal amo M, Packalen P. 2014. Species-Speci ic Managemen In en o y in Finland. In Fo es y
538
Applica ions o Ai bo ne Lase Scanning: Concep s and Case S udies , Mal amo M, Næsse E, and
539
Vauhkonen J (eds). Sp inge Ne he lands: Do d ech ; 241–252. [online] A ailable om:
540
h ps://doi.o g/10.1007/978-94-017-8663-8_12
541
McRobe s RE, Tomppo EO. 2007. Remo e sensing suppo o na ional o es in en o ies. Remo e
542
Sensing o En i onmen 110 : 412–419. DOI: 10.1016/j. se.2006.09.034
543
Mie inen J e al. 2021. Demons a ion o la ge a ea o es olume and p ima y p oduc ion es ima ion
544
app oach based on Sen inel-2 image y and p ocess based ecosys em modelling. In e na ional Jou nal
545
o Remo e Sensing 42 : 9467–9489. DOI: 10.1080/01431161.2021.1998715
546
Nilsson M e al. 2017. A na ionwide o es a ibu e map o Sweden p edic ed using ai bo ne lase
547
scanning da a and ield da a om he Na ional Fo es In en o y. Remo e Sensing o En i onmen 194
548
: 447–454. DOI: 10.1016/j. se.2016.10.022
549
NLS. 2025. Na ional Land Su ey (NLS) Finland. 2025. Kansallise ilmaku aus- ja lase keilausohjelma
550
saa a ja koa – lu assa en is ä a kempaa kolmiulo eis a ie oa Suomes a. A ailable a :
551
h ps://www.maanmi auslai os. i/ajankoh ais a/kansallise -ilmaku aus-ja-lase keilausohjelma -
552
saa a -ja koa-lu assa-en is a (accessed 19 Sep 2025). [In Finnish and Swedish]
553
Olesk A, P aks J, An opo O, Zali e K, A umäe T, Voo mansik K. 2016. In e e ome ic SAR Cohe ence
554
Models o Cha ac e iza ion o Hemibo eal Fo es s Using TanDEM-X Da a. Remo e Sensing 8 : 700.
555
DOI: 10.3390/ s8090700
556
Pe sson HJ, Jonzén J, Nilsson M. 2021. Combining TanDEM-X and Sen inel-2 o la ge-a ea species-
557
wise p edic ion o o es biomass and olume. In e na ional Jou nal o Applied Ea h Obse a ion and
558
Geoin o ma ion 96 : 102275. DOI: 10.1016/j.jag.2020.102275
559
Ronnebe ge O, Fische P, B ox T. 2015. U-Ne : Con olu ional Ne wo ks o Biomedical Image
560
Segmen a ion. Cham. 234–241 pp.
561
San o o M, Bee C, Ca us O, Schmullius C, Sh idenko A, McCallum I, Wegmülle U, Wiesmann A.
562
2011. Re ie al o g owing s ock olume in bo eal o es using hype - empo al se ies o En isa ASAR
563
ScanSAR backsca e measu emen s. Remo e Sensing o En i onmen 115 : 490–507. DOI:
564
10.1016/j. se.2010.09.018
565
Shimada M, Oh aki T. 2010. Gene a ing La ge-Scale High-Quali y SAR Mosaic Da ase s: Applica ion o
566
PALSAR Da a o Global Moni o ing. IEEE Jou nal o Selec ed Topics in Applied Ea h Obse a ions and
567
Remo e Sensing 3 : 637–656. DOI: 10.1109/JSTARS.2010.2077619
568
Tomppo E, Gschwan ne T, Law ence M, McRobe s RE (eds). 2010. Na ional Fo es In en o ies:
569
Pa hways o Common Repo ing . Sp inge Ne he lands: Do d ech [online] A ailable om:
570
h ps://link.sp inge .com/10.1007/978-90-481-3233-1 (Accessed 25 Janua y 2024)
571
Vidal C, Albe di I, Redmond J, Ves man M, Lanz A, Schadaue K. 2016a. The ole o Eu opean Na ional
572
Fo es In en o ies o in e na ional o es y epo ing. Annals o Fo es Science 73 : 793–806. DOI:
573
10.1007/s13595-016-0545-6
574
Vidal C, Albe di IA, He nández Ma eo L, Redmond JJ (eds). 2016b. Na ional Fo es In en o ies .
575
Sp inge In e na ional Publishing: Cham [online] A ailable om:
576
h p://link.sp inge .com/10.1007/978-3-319-44015-6 (Accessed 13 Augus 2025)
577
We ne C, Wegmülle U, S ozzi T, Wiesmann A. 2000. GAMMA SAR AND INTERFEROMETRIC
578
PROCESSING SOFTWARE
579
Wi ke S, Yu X, Ka jalainen M, Hyyppä J, Pu onen E. 2019. Compa ison o wo-dimensional
580
mul i empo al Sen inel-2 da a wi h h ee-dimensional emo e sensing da a sou ces o o es
581
in en o y pa ame e es ima ion o e a bo eal o es . In e na ional Jou nal o Applied Ea h
582
Obse a ion and Geoin o ma ion 76 : 167–178. DOI: 10.1016/j.jag.2018.11.009
583
Wu m M, S a k T, Zhu XX, Weigand M, Taubenböck H. 2019. Seman ic segmen a ion o slums in
584
sa elli e images using ans e lea ning on ully con olu ional neu al ne wo ks. ISPRS Jou nal o
585
Pho og amme y and Remo e Sensing 150 : 59–69. DOI: 10.1016/j.isp sjp s.2019.02.006
586
Yuan Q e al. 2020. Deep lea ning in en i onmen al emo e sensing: Achie emen s and challenges.
587
Remo e Sensing o En i onmen 241 : 111716. DOI: 10.1016/j. se.2020.111716
588
Zhou J e al. 2023. A deep ans e lea ning amewo k o mapping high spa io empo al esolu ion
589
LAI. ISPRS Jou nal o Pho og amme y and Remo e Sensing 206 : 30–48. DOI:
590
10.1016/j.isp sjp s.2023.10.017
591
592
593
594
595
Appendix
596
Table S1 O e all unce ain y assessmen o imbe olume compa ing kNN, UNe , and ine- uned UNe models ained on
597
Finnish (FI) and No wegian (NO) da a. Resul s a e shown o NFI plo s, alida ion s ands, and wo ypes o p edic o a iable
598
combina ions (P ed. ype): S2 (Sen inel-2) and S2S1P2 (Sen inel-2 wi h Sen inel-1 and PALSAR
‑
2). Fine- uned models we e
599
p e- ained on ALS-based o es esou ce maps om Finland (FI) o No way (NO) and ine- uned u he using No wegian
600
aining NFI plo s.
601
Tes
da a
ype
P ed.
ype
Model
ype
kNN
UNe
Fine-
uned
UNe
kNN
UNe
Fine-
Tuned
UNe
kNN
UNe
Fine-
Tuned
UNe
RMSE [m3ha-1]
BIAS [m3ha-1]
R2
NFI
S2
FI
93.8
91.4
-20.4
-9.9
0.37
0.41
NFI
S2
NO
94.0
77.2
77.2
9.9
3.1
2.7
0.37
0.58
0.58
NFI
S2S1P2
FI
90.4
90.3
-5.1
-3.1
0.42
0.42
NFI
S2S1P2
NO
92.8
72.6
72.0
7.0
10.2
3.4
0.39
0.62
0.63
S and
S2
FI
190.1
176.7
-116.6
-98.5
0.04
0.17
S and
S2
NO
148.6
128.1
128.5
-67.6
-45
-46.0
0.42
0.57
0.56
S and
S2S1P2
FI
170.6
168.8
-74.2
-70.8
0.23
0.25
S and
S2S1P2
NO
156.8
119.8
123.5
-67
-4.6
-23.3
0.35
0.62
0.60
602
Table S2 O e all unce ain y assessmen o Lo ey’s heigh compa ing kNN, UNe , and ine- uned UNe models ained on
603
Finnish (FI) and No wegian (NO) da a. Resul s a e shown o NFI plo s, alida ion s ands, and wo ypes o p edic o a iable
604
combina ions (P ed. ype): S2 (Sen inel-2) and S2S1P2 (Sen inel-2 wi h Sen inel-1 and PALSAR
‑
2). Fine- uned models we e
605
p e- ained on ALS-based o es esou ce maps om Finland (FI) o No way (NO) and ine- uned u he using No wegian
606
aining NFI plo s.
607
Tes
da a
ype
P ed.
ype
Model
ype
kNN
UNe
Fine-
Tuned
UNe
kNN
UNe
Fine-
Tuned
UNe
kNN
UNe
Fine-
Tuned
UNe
RMSE [m]
BIAS [m]
R2
NFI
S2
FI
37.3
36.6
3.0
-2.0
0.36
0.38
NFI
S2
NO
36.9
33.0
32.6
3.4
5.2
1.7
0.37
0.50
0.51
NFI
S2S1P2
FI
45.9
36.7
7.5
4.3
0.03
0.38
NFI
S2S1P2
NO
37.0
31.9
31.3
3.2
6.2
1.8
0.37
0.53
0.55
S and
S2
FI
28.9
29.7
1.4
-5.6
0.48
0.46
S and
S2
NO
36.8
31.2
32.3
-15.1
-4.8
-9.2
0.17
0.40
0.36
S and
S2S1P2
FI
43.6
33.8
27.4
5.7
-0.17
0.29
S and
S2S1P2
NO
36.5
30.0
30.5
-13.7
0.8
-4.6
0.18
0.44
0.43
608
609
610
