Remote Sensing of Environment 269 (2022) 112804

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Assessment of a photon recollision probability based forest reflectance

model in European boreal and temperate forests
Aarne Hovi a, *, Daniel Schraik a, Jan Hanuš b, Lucie Homolová b, Jussi Juola a, Mait Lang c, d,
Petr Lukeš b, Jan Pisek c, Miina Rautiainen a, e
a
Aalto University, School of Engineering, Department of Built Environment, P.O. Box 14100, 00760 Aalto, Finland
b
Global Change Research Institute of the Czech Academy of Sciences, Bělidla 986/4a, 603 00 Brno, Czech Republic
c
University of Tartu, Tartu Observatory, Observatooriumi 1, Tõravere, 61602 Tartumaa, Estonia
d
Estonian University of Life Sciences, Institute of Forestry and Rural Engineering, Kreutzwaldi 5, 51014 Tartu, Estonia
e
Aalto University, School of Electrical Engineering, Department of Electronics and Nanoengineering, Finland

A R T I C L E I N F O A B S T R A C T

Editor: Jing M. Chen We report a new version and an empirical evaluation of a forest reflectance model based on photon recollision
probability (p). For the first time, a p-based approach to modeling forest reflectance was tested in a wide range of
Keywords: differently structured forests from different biomes. To parameterize the model, we measured forest canopy
Spectral invariants structure and spectral characteristics for 50 forest plots in four study sites spanning from boreal to temperate
Radiative transfer
biomes in Europe (48◦ –62◦ N). We compared modeled forest reflectance spectra against airborne hyperspectral
Scattering
data at wavelengths of 450–2200 nm. Large overestimation occurred, especially in the near-infrared region,
Hyperspectral
Spectra when the model was parameterized considering only leaves or needles as plant elements and assuming a Lam­
Forest bertian canopy. The model root mean square error (RMSE) was on average 80%, 80%, 54% for coniferous,
Leaf area index broadleaved, and mixed forests, respectively. We suggest a new parameterization that takes into account the
Coniferous nadir to hemispherical reflectance ratio of the canopy and contribution of woody elements to the forest reflec­
Broadleaved tance. We evaluated the new parameterization based on inversion of the model, which resulted in average RMSE
of 20%, 15%, and 11% for coniferous, broadleaved, and mixed forests. The model requires only few structural
parameters and the spectra of foliage, woody elements, and forest floor as input. It can be used in interpretation
of multi- and hyperspectral remote sensing data, as well as in land surface and climate modeling. In general, our
results also indicate that even though the foliage spectra are not dramatically different between coniferous and
broadleaved forests, they can still explain a large part of reflectance differences between these forest types in the
near-infrared, where sensitivity of the reflectance of dense forests to changes in the scattering properties of the
foliage is high.

1. Introduction which are needed to interpret the remote sensing signals.

Physically-based forest reflectance models allow to estimate forest
Multi- and hyperspectral remote sensing provide a means of spatially variables from spectral observations while relying on a limited amount
and temporally continuous monitoring of forest extent, change, health, of empirical data (e.g., Rautiainen, 2005; Schraik et al., 2019). Appli­
phenology, energy exchange, photosynthesis, and biodiversity (Ryu cability of physically-based models can be limited by the large number
et al., 2019; Schneider et al., 2017; Wulder et al., 2019). Accurate and of model parameters, and consequently, ambiguity of the model inver­
cost-efficient monitoring is ever more important as, e.g., climate change sion (Baret and Buis, 2008). Ideally, the models should be simple, yet
and population growth induce various pressures to forest ecosystems. At realistic enough to be practically useful in interpreting remote sensing
the same time, the amount and diversity of available spectral remote signals and extracting the forest parameters of interest (Woodcock et al.,
sensing data is rapidly increasing (Rast and Painter, 2019). A persisting 1994, 1997). Modeling the complex multiple scattering behavior is
problem, however, is the lack of empirical data for constructing models needed for realistic characterization of forest reflectance, especially in

* Corresponding author.
E-mail address: [email protected] (A. Hovi).

https://doi.org/10.1016/j.rse.2021.112804
Received 3 June 2021; Received in revised form 10 November 2021; Accepted 11 November 2021
Available online 25 November 2021
0034-4257/© 2021 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

the near-infrared part of the spectrum where multiple scattering has a first time. The study is based on an exceptionally detailed set of field
substantial role. The spectral invariant and photon recollision proba­ measurements for 50 forest plots spanning from boreal to temperate
bility theories offer a useful framework for designing simple and realistic regions in Europe (48–62◦ N) coupled with airborne hyperspectral data
forest reflectance models which also account for multiple scattering. that cover the spectral region between 450 and 2200 nm. The aims are i)
The spectral invariants theory states that canopy reflectance, trans­ to evaluate the model performance when using its default parameteri­
mittance, and absorption can be approximated based on optical prop­ zation, and ii) to suggest ways of improving the parameterization to
erties of the foliage elements and spectrally invariant parameters yield more realistic forest reflectance simulations. To assess the model
(Knyazikhin et al., 1998). Smolander and Stenberg (2005) interpreted performance in detail, we remove the forest floor contribution from the
one of these spectrally invariant parameters as photon recollision airborne spectra, which allows to evaluate the model component that
probability (p), i.e., ‘the probability that a photon, being scattered by the simulates canopy scattering. We then report the model performance in
canopy, will interact with the canopy again’. This interpretation resulted simulating reflectance of both forest canopy and the entire forest
in a family of canopy reflectance models, referred to as PARAS (see re­ (including forest floor) when using the default parameterization.
view by Stenberg et al., 2016). These models simulate the forest Finally, we perform an inversion of the model, to demonstrate how to
reflectance (or transmittance, albedo, or absorption) based on the improve its parameterization.
spectral scattering or absorption properties of the forest canopy and the
background (forest floor), and the canopy gap fractions that determine 2. Theory
the relative contributions of the canopy and the forest floor (Rautiainen
and Stenberg, 2005). The modeling of scattering or absorption by the We evaluated a model belonging to the PARAS model family (see
canopy is based on plant element (usually leaf) albedo and p that sum­ review by Stenberg et al., 2016), the first model of which was published
marizes the effect of canopy structure on its radiation regime. The in Rautiainen and Stenberg (2005). In this study, we present a new,
benefits of PARAS models are simplicity, i.e., only a small number of extended version of the PARAS model that can model the dependence of
parameters are needed, and the fact that the input parameters can be forest reflectance on view direction while taking into account multiple
derived from field and laboratory measurements (e.g., Rautiainen and scattering between the canopy and the forest floor. Here, we define
Stenberg, 2005; Hadi and Rautiainen, 2018; Hovi et al., 2017; Majasalmi forest floor as all material (living, dead, and inorganic) up to height of
et al., 2014). 1.5 m above the ground. The multiple scattering between canopy and
Empirical evaluation of forest reflectance models is crucial to forest floor has previously been accounted for only by PARAS versions
quantify model errors and thus provide a measure of reliability of the that simulated canopy absorption (Majasalmi et al., 2014) or forest al­
model simulations. The evaluation is however difficult, because of the bedo (Manninen and Stenberg, 2009; Stenberg et al., 2013). In addition,
laborious work required for measuring the input and validation data. our approach differs from the previous studies in that we use empirical
Furthermore, the measurement uncertainties need to be accounted for in models to determine the fraction of incoming radiation that is scattered
the interpretation of the results. Inter-comparison of radiative transfer downwards by the canopy. This is done in the current study, in order to
models (e.g. Widlowski et al., 2015) is one solution proposed to identify reduce uncertainties in model evaluation, but it should be noted that the
model uncertainties. However, showing the model discrepancies does use of empirical models for downward scattering is not an inherent
not necessarily reveal systematic uncertainties due to, e.g., incorrect requirement of our model. This will be explained in more detail later in
parameterization. Only empirical evaluation, i.e., evaluation of Section 2. Similarly to Stenberg et al. (2013), our model assumes Lam­
modeling results against measurements, can reveal the model perfor­ bertian forest floor and vertical symmetry of the canopy. The latter
mance in real forests. means that the directional scattering properties of the canopy are in­
Previous studies have compared empirical reflectance, trans­ dependent of whether the canopy is illuminated from above or from
mittance, absorption, or albedo data from forest canopies to those below.
simulated with models based on photon recollision probability or the Reflectance of forest (or other vegetation) at any optical wavelength
related spectral invariants theory (Disney et al., 2005; Hadi et al., 2017; can be expressed as a combination of solutions to the ‘black soil’ and
Hadi and Rautiainen, 2018; Hovi et al., 2017; Huang et al., 2007; ‘soil’ problems (Knyazikhin et al., 1998; Wang et al., 2003; see also re­
Kuusinen et al., 2014; Majasalmi et al., 2014; Manninen and Stenberg, view by Wang et al., 2018). The black soil problem describes a situation
2009; Panferov et al., 2001; Rautiainen and Stenberg, 2005; Shabanov in which the canopy is illuminated from above, and the forest floor is
et al., 2003; Stenberg et al., 2008; Wang et al., 2003). Comparisons at optically black. The soil problem describes a situation in which the forest
the level of canopy components, such as shoots and small trees, have floor is assumed to act as an ideal Lambertian source that illuminates the
been reported by Hovi et al. (2020) and Rautiainen et al. (2012). Model canopy from below. The hemispherical-directional reflectance factor of
performance has also been assessed based on model inversion and the forest in view direction Ω, R(↓sky,Ω), can be written as
quantifying prediction accuracy of the inverted model parameters (e.g., ( )
( ) ( ) TBS ↓sky , ↓
Lukeš et al., 2011; Schraik et al., 2019). Based on the above literature, R ↓sky , Ω = RBS ↓sky , Ω + RG TS (↑, Ω), (1)
empirical evaluations have had only limited geographical extent. 1 − RG RS (↑, ↓)
Because of different quantities simulated (e.g., albedo, reflectance) and
where RBS(↓sky,Ω) is the canopy hemispherical-directional reflectance
versatile data sources used in model parameterization, it is difficult to
factor in direction Ω in the black soil case, TBS(↓sky,↓) is the canopy
directly compare results between studies. In addition, compromises and
bihemispherical transmittance, RG is forest floor reflectance, RS(↑,↓) is
generalizations in the model parameterization have often been made
the canopy downward bihemispherical reflectance in the soil case, and
because of limited measurements available, and the studies have often
TS(↑,Ω) is the hemispherical-directional transmittance factor of the
been restricted to certain spectral regions or few broad spectral bands.
canopy in direction Ω in the soil case (i.e., signal observed in direction Ω,
The above studies have contributed to an understanding of how p-based
relative to a signal that would be observed from an ideal Lambertian
approaches perform in forest reflectance simulations. However, rigorous
radiation source without the forest canopy obstructing the view). For
validations of the models themselves are still in their infancy. More
explanation of symbols and abbreviations, see also Table 1. The symbols
extensive efforts are needed to quantify predictive performance and
in brackets after each R or T term denote direction of illumination and
uncertainty of these models, and potentially also to point out how the
direction of view, respectively: arrow means hemisphere, Ω direction of
models could be improved, thus making them more readily useful for
the remote sensing sensor. The arrow with subscript ‘sky’ refers to
practical applications in remote sensing.
hemispherical incoming radiation that in clear-sky conditions is
In this paper, we report the performance of a photon recollision
composed of direct solar beam and a small diffuse component, as
probability based forest reflectance model in multiple biomes for the

2
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Table 1 opposed to hemispherical fluxes scattered by the canopy and the forest
Symbols and abbreviations used in this study. floor, which are assumed Lambertian in the model. Note that we omit
Symbol/ Explanation the wavelength sign in our equations here and elsewhere in the paper to
abbreviation simplify notation. The term
βS Shoot clumping coefficient
( )
TBS ↓sky , ↓ ( )
βC Clumping coefficient at higher than shoot level = T ↓sky , ↓ (2)
BRDF Bidirectional reflectance distribution function 1 − RG RS (↑, ↓)
c0 Current year needles
c1 One-year-old needles is the ratio of below canopy to top-of-canopy downward hemispherical
D Ratio of diffuse to total incoming radiation radiation flux. Eq. 1 indicates that forest reflectance is composed of
DBH Diameter at breast height (1.3 m) [cm]
canopy reflectance in the black soil case (first term on the right-hand
e Model residual
fsp Fraction of tree species sp from all tree species
side), and forest floor contribution (second term on the right-hand
Fraction of woody elements from effective (light-capturing) side). The calculation of forest floor contribution requires the ratio of
fW
plant area below canopy to top-of-canopy downward hemispherical radiation flux
H Tree height [m] (Eq. 2). It is modeled as a geometric series, with coefficient TBS(↓sky,↓),
iD Interception of diffuse radiation
and common ratio RGRS(↑,↓), that represents multiple scattering be­
iΩ Interception of radiation in the direction of view
iS Interception of radiation in the direction of the sun tween the forest floor and the canopy. The derivation of the geometric
i0 Interception of total incoming solar radiation (direct + diffuse) series is illustrated in Fig. 1.
Leff, L Effective and true plant area index [m2 m− 2] The terms in Eqs. 1–2 can be expressed as functions of canopy
MEE Mean estimation error
interception and scattering as
n Number of observations
NIR Near-infrared wavelength region, 700–1300 nm
( ) ( )
RBS ↓sky Ω = i0 ωC ↓sky , Ω , (3)
Ω Direction of view
Spherical scattering coefficient of forest canopy, i.e., canopy
ωC
albedo RS (↑↓) = iD ωC (↑↓), (4)
Directional scattering coefficient of forest canopy, for radiation
ωC(↓sky,Ω) ( ) ( )
coming from the sky TBS ↓sky , ↓ = (1 − i0 ) + i0 ωC ↓sky , ↓ , and (5)
Directional scattering coefficient of forest canopy, for radiation
ωC(↑,Ω)
coming from below
TS (↑Ω) = (1 − iΩ ) + iD ωC (↑Ω), (6)
Downward hemispherical scattering coefficient of forest
ωC(↓sky,↓)
canopy, for radiation coming from the sky
Downward hemispherical scattering coefficient of forest where i0 and iΩ are canopy interception of incoming radiation and
ωC(↑,↓)
canopy, for radiation coming from below canopy interception of radiation in the direction of view, respectively, iD
ωE Plant element albedo is canopy interception of diffuse radiation, and ωC(#,#) are illumination
Foliage (leaf or needle) albedo
ωL
and viewing geometry dependent canopy scattering coefficients. Canopy
ωS Shoot albedo
ωW Woody element albedo interception of incoming radiation is defined as i0 = D × iD+(1-D) × iS,
Photon recollision probability (at higher than shoot-level, where D is the ratio of diffuse to total incoming radiation and iS is canopy
p
unless otherwise specified) interception in the direction of the sun. The canopy scattering co­
pS Photon recollision probability within a shoot efficients quantify the fraction of intercepted radiation that is scattered
Pgap Gap fraction
into a hemisphere (in case of hemispherical scattering, ↓ or ↑) or ratio of
q Asymmetry parameter
Ratio of radiation scattered to the upper hemisphere to the total scattering observed in given direction to that observed from an ideal
Q
scattered radiation Lambertian surface that has the same interception and is observed in the
QΩ Directional to hemispherical scattering ratio of forest canopy same view direction (in the case of directional scattering, Ω). Canopy
r Pearson correlation coefficient
spherical scattering coefficient, often referred to as canopy albedo,
Hemispherical-directional reflectance factor of forest canopy in
RBS(↓sky,Ω)
the black-soil case
quantifies the ratio of total scattered radiation (in all directions) to that
RG Reflectance of forest floor intercepted by the canopy. Assuming that the photon recollision prob­
RL Reflectance of foliage (leaf or needle) ability (p) is independent of i) direction of illumination and ii) scattering
RMSE Root mean square error order, i.e., it does not depend on how many times the photon has already
Rref Reflectance of a white reference panel
collided with the canopy, canopy albedo (denoted as ωC without di­
RS(↑,↓) Downward bihemispherical reflectance of forest canopy
Hemispherical-directional reflectance factor of forest (i.e., rection signs in brackets) can be written as (e.g., Smolander and Sten­
R(↓sky,Ω)
including canopy and the forest floor) berg, 2005)
ρ(Ω), ρ(↑) Escape probability into direction Ω, or into a hemisphere
Limiting values of escape probability, reached when photons (1 − p)ωE
ρlim(Ω), ρlim(↑) ωC = , (7)
are evenly distributed in the canopy 1 − pωE
Reflected radiation signal measured from forest floor [digital
sG
numbers] where ωE is the plant element albedo. The value of p depends on canopy
Reflected or transmitted radiation signal measured from a leaf
sR,sT structure as p = 1-iD/L = 1-iDβC /Leff (Stenberg, 2007), where L is true
or needle [digital numbers]
Reflected radiation signal measured from a white reference and Leff is effective plant area index, and βC is canopy clumping coeffi­
sref
panel [digital numbers] cient. If there are several hierarchical levels in the canopy, the albedo of
SWIR Shortwave-infrared wavelength region, 1300–2210 nm an element at one hierarchical level can be calculated, using the p within
SZA Sun zenith angle
that element and the albedo of the lower-level element (Stenberg et al.,
TBS(↓sky,↓) Bihemispherical transmittance of forest canopy
TL Transmittance of foliage (leaf or needle)
2016). For example, the albedo of a coniferous shoot can be calculated
θi Angle of an ith zenith ring in hemispherical photographs with Eq. 7, by setting ωE equal to needle albedo and p equal to photon
θΩ View zenith angle recollision probability within a shoot (Smolander and Stenberg, 2003;
θS Sun zenith angle Rautiainen et al., 2012).
Ratio of below canopy to top-of-canopy downward
T(↓sky,↓) Modeling scattering towards the remote sensing sensor requires
hemispherical radiation flux
Hemispherical-directional transmittance factor of forest taking into account the directionality of scattering. This can be achieved
TS(↑,Ω)
canopy in direction Ω in the soil case by replacing the total escape probability (1-p) in Eq. 7 with directional
VIS Visible wavelength region, 400–700 nm escape probability, ρ(Ω). A complication arises from that the directional

3
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 1. Overview of radiation fluxes that compose the forest hemispherical-directional reflectance factor, R(↓sky,Ω), and ratio of below canopy to top-of-canopy
downward hemispherical radiation flux, T(↓sky,↓). The radiation fluxes form two geometric series, and the sums of these series (Eqs. 1 and 2) give R(↓sky,Ω) and
T(↓sky,↓), respectively. For explanation of symbols, see Section 2 and Table 1. Solid color of an arrow indicates a downward flux, and a striped pattern an upward flux.
The colors of the arrows represent coupled fluxes: the downward flux at the bottom of the forest canopy generates an upward flux when the radiation is reflected by
the forest floor. The relative decrease of widths of the arrows from left to right represents the diminishing amount of radiation from first to higher scattering orders,
but the absolute widths are compromised for sake of clarity. In reality, the upward reflected flux is always smaller than the downward flux, due to absorption by the
forest floor. The scattering between forest floor and canopy continues infinitely and attenuates at each interaction.

escape probability depends on scattering order (Mõttus and Stenberg, In the case of homogeneous canopy with Poisson-distributed leaves QΩ
2008; Yang et al., 2017). This is because photons hitting the canopy from equals iΩ/iD (Yang et al., 2017). This indicates that the ratio of direc­
the above are likely to hit the upper parts of the canopy and therefore tional to hemispherical scattering could potentially be approximated as
more likely to escape towards the upper than lower hemisphere. As the ratio of canopy interception in direction of view to canopy diffuse
photons scatter within the canopy several times, they become evenly interception.
distributed in the canopy and directional escape probability approaches In order to avoid uncertainties in modeling of Q, in this study, we
its limiting value, ρlim(Ω). Taking the dependence of directional escape obtained the downward hemispherical scattering coefficient directly
probability on scattering order into account, Mõttus and Stenberg from empirical measurements. Noting that QωC = ωC-(1-Q)ωC = ωC-
(2008) introduced the semi-empirical wavelength-dependent parameter ωC(↓sky,↓), our final model for the directional scattering coefficient
Q that quantifies the ratio of radiation scattered to the upper hemisphere becomes
to the total scattered radiation. Using the Q parameter, scattering into [ ]
( ) [ ( )] (1 − p)ωE ( )
the upper hemisphere can be modeled as ωC ↓sky Ω = QΩ ωC − ωC ↓sky ↓ = QΩ − ωC ↓sky , ↓ , (10)
1 − pω E
( )
ωC ↓sky ↑ = QωC , (8)
where ωC(↓sky,↓) is canopy downward hemispherical scattering coeffi­
where ωC is the canopy albedo (Eq. 7). In reality, the canopy scattering is cient for radiation coming from above (taken from empirical measure­
rarely Lambertian, and therefore modeling hemispherical scattering ments in our study). We note that an alternative approach would be to
without considering the exact direction of view is not necessarily suffi­ use the semiempirical Q parameter from Mõttus and Stenberg (2008), i.
cient. It can be shown (Supplementary material Section 1) that if the e., Eq. 9 instead of Eq. 10. Using Eq. 9 and setting QΩ = 1, and setting iΩ
ratio of limiting escape probabilities (directional vs. hemispherical) is = iD (i.e., assuming a hemispherical view), would produce the same
ρlim(Ω)/ρlim(↑) = QΩ, the formula for directional scattering coefficient results as the model presented in Stenberg et al. (2013). Further, setting
becomes T(↓sky,↓) = 1-i0 and TS(↑,Ω) = 1-iΩ, i.e., ignoring multiple scattering
( ) between canopy and forest floor, would result in the original PARAS
ωC ↓sky Ω = QΩ QωC . (9) model presented in Rautiainen and Stenberg (2005).
From Eq. 9 it is seen that QΩ can also be interpreted as the ratio of In a nutshell, our primary model of interest is Eq. 1, and particularly
directional to hemispherical scattering, i.e., QΩ = ωC(↓sky,Ω)/(QωC). its component RBS(↓,Ω) that models canopy scattering based on p-theory
Because in our experiment the forest reflectance data were acquired in a (Eqs. 3 and 10). The model requires as inputs spectral parameters forest
near-nadir observation geometry, we refer to QΩ as ‘nadir to hemi­ floor reflectance (RG), plant element albedo (ωE), and canopy downward
spherical scattering ratio’ when discussing our experiment and results. hemispherical scattering coefficient, ωC(↓sky,↓), and structural

4
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

parameters canopy interception (i0, iD, iΩ), photon recollision probabil­ 27◦ 19′ E), and Bílý Kříž (49◦ 30′ N, 18◦ 32′ E) and Lanžhot (48◦ 41′ N,
ity (p), and directional to hemispherical reflectance ratio (QΩ). The 16◦ 57′ E), in the Czech Republic (Fig. 2). The forests in the study sites are
structural parameters are spectrally invariant, except that i0 is boreal conifer-dominated with minor broadleaved mixture (Hyytiälä),
wavelength-dependent because the diffuse fraction of incoming solar hemiboreal mixed broadleaved and coniferous (Järvselja), temperate
radiation decreases as a function of wavelength. In implementing the coniferous mountain forests (Bílý Kříž), and temperate broadleaf-
calculation of forest floor contribution (second term on the right hand dominated floodplain forests (Lanžhot). The topography for Järvselja
side of Eq. 1), we assume that the forest canopy is vertically symmetric, and Lanžhot is flat (30–45 m and 150–155 m a.s.l.), Hyytiälä is
and therefore the canopy downward hemispherical scattering coeffi­ moderately flat (130–200 m a.s.l.), and Bílý Kříž is mountainous
cient for radiation coming from below is ωC(↑,↓) = 1/QΩ × ωC(↓sky,Ω). (700–950 m a.s.l.).
We also assume that the upward scattering of radiation that comes from Airborne hyperspectral data were acquired over all sites in summer
below follows a Lambertian distribution, and therefore ωC(↑,Ω) = 2019, and the field measurements were conducted close to the acqui­
ωC(↓sky,↓). sition, or in similar phenological conditions a year later (Table 2).
During the field measurements, we established a total of 50 forest plots,
3. Materials and methods located within 2 km from the site center coordinates. The plots were
selected to include both coniferous and broadleaved forests with
3.1. Study areas and forest plots different tree species and a wide range of plant area index and canopy
cover values, because these parameters were assumed to influence the
Field and airborne data were collected from four study sites con­ forest spectra. Coordinates of all plots were measured with a hand-held
taining forests of different structure and species composition. The sites GPS (Garmin GPSmap 62stc) and later refined to submeter accuracy
were Hyytiälä, Finland (61◦ 51′ N, 24◦ 18′ E), Järvselja, Estonia (58◦ 17′ N, through co-registering terrestrial laser scanning with airborne laser

Fig. 2. Location of the study sites with example photographs showing the studied forests. The study plots in Hyytiälä were in forests composed mainly of coniferous
pine (a) and spruce (b), and broadleaved birch (c). The plots in Järvselja were in forests composed mainly of coniferous pine (d) and spruce (not shown), and various
broadleaved species (e–f). The plots in Bílý Kříž were in spruce-dominated forests of varying age and density (g–h), and those in Lanžhot were in forests composed of
various broadleaved species (i–j).

5
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Table 2 Table 4
Acquisition dates, times of day (range), and sun zenith angles (range) during the Mean (and range) of tree species percentage (from basal area) in the study plots.
acquisitions of airborne and field data in the four study sites. Species that belong to the same genus, and were difficult to recognize reliably in
Hyytiälä Järvselja Bílý Kříž Lanžhot
the field, were grouped together. Bold font marks species that were sampled for
leaf or needle spectra.
Airborne hyperspectral measurements
Date Jul 13th, 2019 Jul 15th, 2019 Sep 4th, 2019 Sep 4th, 2019 Common name Scientific name Percent
Local Hyytiälä
08:57–10:21 1 12:57–14:07 1 11:01–11:07 2 12:14–12:22 2
time 38%
Sun Scots pine Pinus sylvestris L.
(0–100%)
zenith Norway spruce Picea abies (L.) Karst. 40% (0–99%)
51–60 37–38 47–48 42
angle Betula pendula Roth, B. pubescens
[◦ ] Silver birch, downy birch 22% (0–97%)
Ehrh.
Below canopy downward radiation flux measurements Other broadleaved – 0% (0–2%)
Jul 22–25th, Jun 26–28th, Sep 22nd, Järvselja
Date Sep 5th, 2019
2019 2020 2019 16%
Local Scots pine P. sylvestris L.
09:10–11:31 1 09:25–17:26 1 10:31–12:27 2 09:41–11:49 2 (0–100%)
time Norway spruce P. abies (L.) Karst. 18% (0–95%)
Sun Silver birch, downy birch B. pendula Roth, B. pubescens Ehrh. 30% (0–91%)
zenith European alder Alnus glutinosa (L.) Gaertn. 17% (0–82%)
47–60 40–55 50–56 45–57
angle European aspen Populus tremula L. 8% (0–50%)
[◦ ] Littleleaf linden Tilia cordata Mill. 6% (0–41%)
Forest inventory, hemispherical photography, spectral measurements of forest floor Goat willow Salix caprea L. 2% (0–13%)
and tree foliage Other willows Salix sp. 1% (0–13%)
Jun 17th–Jul Jun 24th–Jul Sep 17–30th, Sep 2–12th, European ash Fraxinus excelsior L. 1% (0–10%)
Date
26th, 2019 19th, 2020 2019 2019 Other broadleaved – 2% (0–8%)
1 2 Bílý Kříž
UTC + 3 h; UTC + 2 h.
91%
Norway spruce P. abies (L.) Karst.
(72–100%)
scanning measurements. Forest inventory measurements (Supplemen­ Broadleaved – 9% (0–28%)
tary material Section 2.1) were conducted in all plots, to obtain a general Lanžhot
Quercus robur L., Quercus petraea 50%
description of the forest in the plots, and to compute tree species frac­ English oak, sessile oak
(Matt.) Liebl. (0–100%)
tions needed in parameterization of the forest reflectance model. European ash, narrow-
F. excelsior L., F. angustifolia Vahl. 21% (0–74%)
Tables 3–4 summarize the forest variables and tree species composition leafed ash
in the plots. European hornbeam Carpinus betulus L. 11% (0–80%)
Hedge maple Acer campestre L. 11% (0–38%)
White poplar, European
3.2. Airborne hyperspectral data aspen
Populus alba L., P. tremula L. 5% (0–47%)

Littleleaf linden T. cordata Mill. 2% (0–12%)

Airborne hyperspectral measurements in all study sites were carried Other broadleaved – 1% (0–5%)
out using the same hyperspectral pushbroom sensors CASI-1500 and
SASI-600 by Itres Ltd., Canada, mounted on a Cessna C208B aircraft.
Sampling interval and spectral resolution were 15 nm. The flying alti­
Both sensors had been spectrally and radiometrically calibrated before
tude was approximately 1 km, resulting in pixel sizes of 0.5 m (CASI)
the flight season in March 2019. CASI-1500 sampled visible (VIS) to
and 1.25 m (SASI) on the ground. Both sensors had a field-of-view of 20◦
near-infrared (NIR) wavelengths from 382 to 1052 nm, and SASI-600
from nadir. The data were thus acquired in near-nadir observation ge­
sampled NIR and shortwave-infrared (SWIR) from 958 to 2443 nm.
ometry. The flying azimuth direction was approximately the same as sun
azimuth, which minimized spectral differences between plots that could
Table 3
occur due to large anisotropy of forest in the solar principal plane
Median (and range) of forest variables in the study plots, separately for conif­
(Deering et al., 1999; Kuusk et al., 2014). Sun zenith angle during the
erous, broadleaved, mixed, and all forest plots.
acquisitions varied from 37◦ to 60◦ (Table 2). The overlap of flight lines
Coniferous Broadleaved Mixed All was 60–80%. Simultaneous airborne laser scanning data were acquired
(more than (more than
85% conifers) 85%
with a Riegl LMS Q-780 laser scanner and used in the plot positioning as
broadleaved) well as in orthorectification of the hyperspectral data.
The raw digital number data collected by the hyperspectral sensors
Number of
plots
23 21 6 50 were first radiometrically corrected using the RadCor software (v11)
Basal area provided by Itres Ltd. The data were then geo-orthorectified using IMU/
[m2 28 (4–66) 21 (4–60) 26 (3–51) 25 (3–66) GNSS data and a canopy surface model (pixel size 0.5 m) computed from
ha− 1] the laser scanning data. The atmospheric correction was performed
Diameter at
breast
using ATCOR-4 software bundle v7.2.0 (Richter and Schläpfer, 2018),
27 (5–51) 20 (4–74) 18 (5–37) 22 (4–74) which utilizes a database of atmospheric look-up tables calculated by
height
[cm] means of the MODTRAN5 radiative transfer code. In the atmospheric
Tree height correction, signals measured by the sensors were corrected for path
25 (4–43) 21 (5–40) 21 (5–39) 23 (4–43)
[m]
radiance and adjacency radiance. For each site, inflight radiometric
Stand
density
670 660 1930 720 (vicarious) calibration was performed, using one bright ground target
(110− 20,000) (100–22,500) (460–8,400) (100–22,500)
[ha− 1] with known reflectance. The resulting data are at-surface (top-of-can­
Effective opy) hemispherical-directional reflectance factors, i.e., correspond to R
plant (↓sky,Ω) in our forest reflectance model. No topographic correction was
2.8
area 2.5 (0.4–4.7) 2.4 (1.2–6.3) 2.5 (0.4–6.3)
index
(0.4–4.9) applied.
[m2 m− 2] The hyperspectral data were manually checked in order to exclude
visible clouds or cloud shadows from the analysis. Clouds were present
particularly over the Hyytiälä site, where the conditions varied from

6
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

sunny to partly cloudy during the acquisition. Occasional clouds were same equations as used by the LAI-2200 instrument (LI-COR, 2012). See
also present in the Bílý Kříž site. The conditions during the Järvselja and Supplementary material Section 2.2 for equations used in the calcula­
Lanžhot acquisitions were cloud-free. If there were several cloud and tions, as well as detailed description of image acquisition settings.
shadow-free flight lines available for a plot, only the line acquired Linear interpolation between the five zenith rings was performed to
nearest to nadir was used. Finally, pixel values were extracted from a obtain the canopy gap fractions corresponding to the view and sun
rectangle of 30 × 30 m around each plot center and averaged to result in zenith angles during the airborne data acquisition, i.e., Pgap (θΩ ) and
a single forest reflectance spectrum per plot. Pgap (θS ). Interception in the direction of view and the sun were then
calculated as iΩ = 1-Pgap (θΩ ) and iS = 1-Pgap (θS ), and interception of
3.3. Model input data incoming radiation (i0) was obtained by weighting diffuse interception
(iD) and interception in the sun direction (iS) with the ratio of diffuse to
3.3.1. Hemispherical photography total incoming radiation (D) as described in Section 2. The wavelength-
We obtained effective plant area index and canopy interception dependent values of D were obtained from the diffuse and total irradi­
values for all plots using hemispherical photographs taken in diffuse ance spectra at ground-level produced by the ATCOR-4 software
illumination conditions. A total of 21 photographs per plot were taken (Richter and Schläpfer, 2018).
(Fig. 3). The height of the camera from the ground was 1.5 m in plots
with mean tree diameter at breast height (DBH) equal or larger than 10 3.3.2. Forest floor spectra
cm, and 1 m in plots with DBH less than 10 cm. We used a Nikon D5000 Reflectance spectra of the forest floor were measured at wavelengths
digital camera with a Sigma 4.5 mm 1:2.8 DC HSM Circular Fisheye lens of 350–2500 nm in each plot. The measurements (15 per plot) were
that was geometrically calibrated to enable correction for radial lens taken at approximately 0.8 m intervals along an 11 m long east-west
distortion in the data processing. The photographs were thresholded oriented transect, located 1 m south from the plot center (Fig. 3). We
into binary images by applying the algorithm of Nobis and Hunziker used an Analytical Spectral Devices (ASD) FieldSpec4 spectrometer (ser.
(2005) to the blue channel. Canopy gap fraction was computed in five nr. 18456) with a field-of-view of 25◦ . The measurements were taken at
concentric zenith rings with median zenith angles of 10.7◦ , 23.7◦ , 38.1◦ , nadir, from approximately 1.3 m height. White reference was measured
52.8◦ , and 66.6◦ , and averaged over all photographs in a plot to yield at the beginning and end of the transect, and at every third measurement
spot in between. The white reference was a 25 × 25 cm Spectralon panel
plot- and zenith angle-specific mean gap fractions, Pgap (θi ). All azimuths
with 99% nominal reflectance. The dark current was measured at the
were averaged, i.e., gap fraction was assumed independent of azimuth in
beginning and end of the transect. To minimize variation due to sun-
our study. Effective plant area index (Leff) and diffuse interception (iD) of
flecks and shadows, we performed the measurements always in diffuse
the forest canopy were calculated from the Pgap (θi ) values, using the illumination conditions, i.e., close to dusk or dawn, or on cloudy days.
The measured quantity is thus hemispherical-conical reflectance factor.
Integration time was adjusted to the prevailing light conditions, using
the automatic optimization by the instrument. The measured raw radi­
ation signals (digital numbers) were processed into forest floor reflec­
tance (RG) as RG = sG/sref × Rref, where sG is the signal from forest floor
and sref is the signal from the white reference (dark current subtracted
from both before calculations), and Rref is the reflectance of the white
reference panel. The value of sref for each measurement was obtained by
linearly interpolating in time between the white reference measure­
ments. The RG spectra were averaged to result in a single spectrum per
plot.

3.3.3. Measurements of below canopy downward radiation flux

Ratio of below canopy to top-of-canopy downward hemispherical
radiation flux, T(↓sky,↓), was measured at wavelengths of 350–2500 nm
for a subset of 22 plots (4–8 plots per site). The measurement protocol
and data processing have been described in detail in Hovi and Rautiai­
nen (2020). Total of 49 measurements per plot were performed in a 5 ×
5 m grid (Fig. 3). The measurements were performed in clear-sky con­
ditions, using cosine receptors attached to two field spectrometers. One
instrument (ASD FieldSpec4 ser. nr. 18641 or FieldSpec3 ser. nr. 16089)
was continuously measuring at 15 s intervals in an open area nearby the
study plots, and the other (ASD FieldSpec4 ser. nr. 18456) was used for
performing measurements in the plots. The integration time was opti­
mized before the measurements, and intercalibration of the instruments
was performed before and after the measurements on each measurement
day. The measurement height was 1.5 m from ground and the cosine
receptor was manually leveled during each measurement. Sun zenith
angle during the measurements varied from 40◦ to 60◦ (Table 2). The
measured spectra were averaged to result in a single spectrum per plot.

3.3.4. Foliage spectra

Foliage, i.e., leaf and needle, reflectance and transmittance spectra at
wavelengths of 350–2500 nm were measured for all major tree species in
the study sites (Table 4). Samples of both top-of-canopy (sun-exposed)
and bottom-of-canopy (shaded) foliage were taken. For coniferous
Fig. 3. Sampling layout of field measurements conducted in the study plots.

7
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

needles, we sampled two age cohorts: current year (c0) and one-year-old from the PARAS forest reflectance model (Eqs. 1–6). The calculation
(c1) needles. Only healthy foliage was measured. The sampling protocols started by estimating the canopy downward hemispherical scattering
differed slightly between Lanžhot and the other sites because the leaf coefficients, ωC(↓sky,↓), from the measured spectral ratio of below can­
spectral measurements in Lanžhot were conducted as part of another opy to top-of-canopy downward hemispherical radiation flux (Step 1 of
field campaign. For the other sites, we sampled three trees of each major the algorithm). The values of ωC(↓sky,↓) were then used together with
species, and took three samples of each foliage class per tree (‘sun- field-measured forest floor reflectance and canopy interception calcu­
exposed c0’ and ‘shaded c0’ for all species, and additionally ‘sun-exposed lated from hemispherical photographs to remove the contribution of
c1’ and ‘shaded c1’ for conifers). In Lanžhot, we sampled one to four trees forest floor from the airborne reflectance measurements (Step 2 of the
with one sample of each foliage class per tree (‘sun-exposed c0’ and algorithm). This resulted in estimates of the forest canopy directional
‘shaded c0’). For minority broadleaved species in Järvselja (littleleaf scattering coefficients, ωC(↓sky,Ω). Mathematical formulation and a
linden, goat willow, other willows, European ash, common hazel), one graphical illustration of the algorithm are given in the Supplementary
tree with three sun-exposed c0 leaves per species was sampled. Reflec­ material Section 2.5. The algorithm does not require a model for canopy
tance and transmittance spectra of the samples were measured with ASD scattering, except for initial estimates of ωC(↓sky,↓) at two wavelengths
RTS-3ZC integrating spheres, attached to an ASD spectrometer (Field­ (blue and red). Independence from the canopy scattering model was
Spec3 ser. nr. 16089, or FieldSpec4 ser. nr. 18456 or 18641). The important because our purpose was to evaluate the p-based model for
measurement protocol and data processing are described in detail in the canopy scattering. The two steps of the algorithm should be iterated
Supplementary material Section 2.3. Foliage albedo was calculated as until the solution stabilizes. We used five iterations. In the fifth iteration,
the sum of reflectance and transmittance. Finally, a representative fo­ the value of ωC(↓sky,Ω) changed less than 0.1%. As an intermediate
liage albedo spectrum for each tree species in each study plot was result, the algorithm also produced estimates of forest hemispherical-
calculated, as described in the Supplementary material Section 2.4. directional reflectance factors without forest floor contribution,
RBS(↓sky,Ω), which we used for illustrating the magnitude of forest floor
3.3.5. Woody element spectra contribution to airborne forest reflectance in different types of forests.
Our primary source for woody element spectra was Juola et al.
(2021), who collected stem bark reflectance spectra for boreal and 3.4.3. Modeling canopy scattering
temperate tree species in Finland and Estonia in summer 2020. The Canopy directional scattering coefficient values obtained through
spectra were measured at 1.3 m height in diffuse illumination condi­ the forest floor contribution removal were compared with those simu­
tions, using a Specim IQ hyperspectral camera covering wavelengths lated using our model (Eq. 10). We tested two model parameterizations,
400–1000 nm. For wavelengths above 1000 nm, we used data from which we call ‘default’ and ‘fitted’ models. Differences between them
other published sources: pine, spruce, and birch stem bark spectra were will be explained later in this section.
taken from Lang et al. (2002), and aspen stem bark spectrum from In both models, we considered the shoot as the basic green plant
Spencer and Rock (1999) was used for all other broadleaved species. We element. Shoot albedo (ωS) was calculated with Eq. 7, setting ωE equal to
used only spectra measured from the lower part of the stem, in order to foliage albedo and p equal to within-shoot photon recollision probability
obtain best correspondence with the measurements of Juola et al. (pS). The value of pS for each tree species was obtained as pS = 1-βS,
(2021). Because continuity of the spectra was required in the forest where βS is the shoot clumping coefficient (Smolander and Stenberg,
reflectance modeling, the spectra above 1000 nm were scaled with a 2003). For conifers, we used a βS value of 0.6 that represents a typical
linear scaling factor that was determined so that the reflectance at 1000 value for Scots pine and Norway spruce (Thérézien et al., 2007). For
nm matched with the measurements of Juola et al. (2021) at the same broadleaved species, value of 1 was used, i.e., we assumed no clumping
wavelength. The stem bark reflectance spectra were used as albedo at shoot level, and shoot albedo of broadleaved trees equaled leaf al­
spectra of woody elements in the forest reflectance model, i.e., trans­ bedo. The canopy-level photon recollision probability (p in Eq. 10) for
mittance of woody elements was assumed to equal zero. Finally, each plot was calculated from diffuse interception (iD), canopy clumping
representative woody element albedo spectrum for each tree species in coefficient (βC) and effective plant area index (Leff) as p = 1-iDβC/Leff
each study plot was calculated, as described in the Supplementary ma­ (Stenberg, 2007). The value of βC (mean of 0.95 and standard deviation
terial Section 2.4. of 0.04 for our study plots) was obtained from hemispherical photo­
graphs following Ryu et al. (2010) and its calculation is explained in the
Supplementary material Section 2.2. The canopy downward hemi­
3.4. Analyses
spherical scattering coefficient, ωC(↓sky,↓), was obtained from the
measured ratio of below canopy to top-of-canopy downward hemi­
3.4.1. Filtering of spectra
spherical radiation flux, using the algorithm for forest floor contribution
High spectral resolution data (forest floor reflectance, ratio of below
removal.
canopy to top-of-canopy downward hemispherical radiation flux, fo­
In the default model, we assumed a Lambertian canopy and did not
liage and woody element albedo spectra) were filtered with the
account for woody elements, i.e., nadir to hemispherical scattering ratio
Savitzky-Golay filter (Savitzky and Golay, 1964), and resampled by
(QΩ) was set to 1, and shoot albedo was used as plant element albedo
weighting with Gaussian functions that had mean and full-width-at-half-
(ωE). Our formulation, i.e., using shoot albedo as ωE and calculating p at
maximum corresponding to the band centers and spectral resolution of
higher than shoot level, is equal to using leaf or needle albedo as ωE and
the airborne forest reflectance data, respectively. From all spectral data,
taking into account shoot-level clumping in the calculation of canopy-
we excluded noisy regions caused by atmospheric absorption or low
level p, as done in previous studies (e.g., Rautiainen and Stenberg,
signal-to-noise ratio of the instruments. A total of 69 spectral bands
2005; Hovi et al., 2017; Hadi and Rautiainen, 2018; Schraik et al.,
remained for the analyses. These bands covered wavelength ranges of
2019). Plot-specific shoot albedo spectra were obtained by weighting
453–881 nm, 1018–1093 nm, 1183–1288 nm, 1543–1723 nm, and
species-specific shoot albedo values with the field-measured tree species
2053–2203 nm.
fractions in a plot.
In the fitted model, we allowed QΩ to deviate from 1, and took into
3.4.2. Removal of forest floor contribution from forest reflectance spectra
account woody elements in the estimation of ωE. The latter was achieved
We designed an algorithm for removal of forest floor contribution
by calculating the plot-specific plant element albedo (ωE) as weighted
from the airborne forest reflectance spectra. The algorithm was derived

8
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

average of species-specific woody element (ωW,sp) and shoot (ωS,sp) 4. Results

albedos as
∑{ [ ( ) ]} 4.1. Variation in forest reflectance spectra and structure
ωE = fsp fW,sp ωW,sp + 1 − fW,sp ωS,sp , (11)
sp
We present the results separately for coniferous (more than 85%
where fsp are tree species fractions in a plot, and fW,sp (constrained to 0 ≤ conifers, n = 23), broadleaved (more than 85% broadleaved trees, n =
fW,sp ≤ 1) are species-specific woody element fractions from effective 21), and mixed forest plots (n = 6), because coniferous and broadleaved
light-capturing plant area. The model had unknown parameters QΩ and forests showed distinct spectral characteristics. Overall, airborne forest
fW,sp, which were estimated by inverting the model. The inversion was reflectance spectra varied considerably between plots (Fig. 4). Broad­
performed by nonlinear least squares estimation, using function opti­ leaved forests showed systematically higher values of NIR reflectance
mize.least_squares() and its default method (Branch et al., 1999) in compared to the coniferous forests (Fig. 4a,b). Effective plant area index
SciPy Python package v1.1.0 (Virtanen et al., 2020). We minimized the (Leff), and consequently also canopy interception varied within each
sum of plot- and band-specific residuals (eij), calculated as species group. Overall, the Leff values ranged between 0.4 and 6.3 among
[ ] our study plots (Table 3). The dependence of canopy interception on
eij = QΩ,i
(1 − pi )ωE,ij ( ) ( )
− ωC,ij ↓sky ↓ − ωC,ij ↓sky Ω , (12) zenith angle was different for forests with different Leff. Sparse forests
1 − pi ωE,ij (low Leff) had low values of interception at nadir compared to the more
oblique angles, whereas in dense forests the interception in all zenith
where i refers to plot and j to spectral band, and ωC,ij(↓sky,Ω) is the angles tended to be similar (Fig. 5).
reference value of the canopy directional scattering coefficient, esti­ In coniferous forests, the average airborne forest reflectance, R
mated using the algorithm for forest floor contribution removal. For the (↓sky,Ω), decreased as a function of Leff in all wavelengths (Fig. 6 left
purpose of the inversion, we divided the tree species in three groups: column), with the mean (and standard deviation) of r between Leff and R
pine, spruce, and broadleaved. The value of fW,sp was constrained to be (↓sky,Ω) being − 0.71 (0.09). The densest coniferous plot in Bílý Kříž (Leff
equal for all species belonging to the same group, in order to avoid = 4.7) was an outlier, as its reflectance resembled that of broadleaved
overfitting. To take into account the uncertainty in ωC(↓sky,Ω), inverse forests. This could be because the few broadleaved beech trees in that
values of the 95% confidence intervals of ωC(↓sky,Ω) (see Section 3.4.5) plot, with their wide crowns, probably had a larger contribution to the
were used as weights for the model residuals. forest reflectance than indicated by their fraction from the basal area.
Topographic effects are also possible because the plot was on a southern
3.4.4. Modeling forest reflectance spectra slope (~10◦ ). In broadleaved forests, the correlation between Leff and R
The forest hemispherical-directional reflectance factors for each plot (↓sky,Ω) was weak, and in the mixed forests it resembled that observed in
were modeled using Eq. 1, together with the default and fitted param­ the coniferous plots (Fig. 6 middle and right columns). The mean (and
eterizations of Eq. 10 as described above. Canopy interception values (i0, standard deviation) of r between Leff and R(↓sky,Ω) was 0.00 (0.22) in the
iD, iΩ) were taken from hemispherical photographs, and forest floor broadleaved, and − 0.73 (0.17) in the mixed forests.
reflectance spectra (RG) from the field measurements. As explained in
Section 2, the upward scattering coefficient of the canopy for radiation 4.2. Forest floor contribution to forest reflectance spectra
coming from below, i.e., ωC(↑,Ω), was assumed equal to ωC(↓sky,↓). The
downward hemispherical scattering coefficient for radiation coming The forest floor contribution in each study plot and wavelength was
from below, i.e., ωC(↑,↓), was calculated as ωC(↑,↓) = 1/QΩ × ωC(↓sky,Ω). calculated as (R(↓sky,Ω)-RBS(↓sky,Ω))/R(↓sky,Ω) × 100%. The average
forest floor contribution (i.e., averaged over all study plots) depended on
3.4.5. Sensitivity analysis wavelength, and thus varied within the range of 18–28% in coniferous,
Sensitivity analysis was performed to quantify uncertainty in the 13–19% in broadleaved, and 25–37% in mixed forests in the VIS region.
results of forest floor contribution removal (Section 3.4.2) and forest In NIR and SWIR, the average forest floor contribution was larger:
reflectance modeling (Sections 3.4.3 and 3.4.4). We generated 1000 sets 26–35% in coniferous, 15–26% in broadleaved, and 32–43% in mixed
of results and computed 95% confidence intervals for the reference forests. The forest reflectance without forest floor contribution,
value of canopy directional scattering coefficient, ωC(↓sky,Ω), and for the RBS(↓sky,Ω), approached zero as Leff decreased (not shown). This is ex­
modeled ωC(↓sky,Ω) and forest reflectance, R(↓sky,Ω). In each of the 1000 pected because a small amount of intercepting canopy material means
sets, we randomly selected the values for the input parameters from that most of the incoming radiation would be absorbed by the (black)
within their estimated uncertainty ranges. See Supplementary material soil. Inaccuracies in the removal of forest floor contribution from the
Section 2.6 for the list of input parameters and their estimated reflectance spectra were observed, and they resulted in negative
uncertainties. RBS(↓sky,Ω) in some of the plots with sparse canopies (low Leff). Conse­
quently, also the canopy directional scattering coefficients, ωC(↓sky,Ω),
3.4.6. Model evaluation exhibited large variation, which was seen also in their uncertainty es­
The model simulations were evaluated, using wavelength-specific timates (Fig. 6). Overall, ωC(↓sky,Ω) values were very similar to R(↓sky,Ω)
root mean square error (RMSE) and mean estimation error (MEE) as in dense forests (high Leff) and tended to be somewhat lower than R
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (↓sky,Ω) in sparse forests (low Leff). Due to high uncertainty, we excluded
1∑ n
RMSE = ( ŷi − yi )2 , and (13) plots with Leff lower than 1, and used the remaining plots (n = 45) in the
n i=1
modeling of canopy scattering (Section 4.3). In this subset, the average
forest floor contribution was smaller than in the entire dataset, varying
1∑ n
within the range of 11–20% in coniferous, 13–19% in broadleaved, and
MEE = ( ŷi − yi ), (14)
n i=1 6–10% in mixed forests in the VIS region. In NIR and SWIR, the
respective values were 18–27% (coniferous), 15–26% (broadleaved),
where ŷi is the modeled and yi is the reference value for plot i, and n is and 8–14% (mixed forests).
the total number of plots. RMSE and MEE relative to the reference [%]
were obtained by multiplying RMSE and MEE by term 100%/y, where y 4.3. Modeled canopy scattering
is the mean of the reference values. Wavelength-specific Pearson cor­
relation coefficient (r) was also used to evaluate the agreement between The default model tended to overestimate canopy scattering
the modeled and reference. considerably, especially in the NIR region (Fig. 7). The maximum

9
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 4. Forest hemispherical-directional reflectance factors, R(↓sky,Ω), from the airborne measurements in the coniferous (more than 85% conifers), broadleaved
(more than 85% broadleaved trees), and mixed forest plots.

Fig. 5. Ratio of canopy interception (i) at four different zenith rings (median zenith angles of 10.7◦ , 23.7◦ , 38.1◦ , 52.8◦ ) to the interception at the largest zenith ring
of the hemispherical photographs (66.6◦ ) as function of effective plant area index (Leff).

overestimation was larger in coniferous (MEE ranging from − 24% to average weak, but varied between wavelengths and species groups, with
240%) than in broadleaved (MEE from 15% to 151%) and mixed forests mean r of 0.12, 0.09, and 0.94 for coniferous, broadleaved, and mixed
(MEE from − 21% to 137%) (Fig. 8). In coniferous forests, the over­ forests, respectively (Fig. 8).
estimation in the NIR region clearly exceeded the uncertainty estimates Contrary to the default model, the canopy directional scattering
(Fig. 7). In broadleaved forests, the uncertainty estimates in NIR over­ coefficients from the fitted model were close to the reference values
lapped in some (but not all) of the plots. The correlation between (Fig. 7). The absolute values of the MEE for the fitted model were
reference and modeled directional scattering coefficients was on notably lower than for the default model: the MEE of the fitted model

10
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 6. Measured forest hemispherical-directional reflectance factors, R(↓sky,Ω) (square symbols), and canopy directional scattering coefficients estimated from them,
ωC(↓sky,Ω) (x symbols), as a function of effective plant area index (Leff) at four wavelengths (rows) in coniferous, broadleaved, and mixed forest plots (columns). The
vertical lines illustrate 95% confidence intervals for ωC(↓sky,Ω).

varied from − 51% to 27% for coniferous, from − 31% to 34% for coniferous plots in Bílý Kříž (Fig. 7). The somewhat large errors in Bílý
broadleaved, and from − 33% to 22% for mixed forests (Fig. 8). Kříž also influenced the RMSE and MEE of the coniferous forests, which
Furthermore, the performance of the fitted model was not strongly went up to 54% and down to − 51% in the VIS region, respectively
dependent on wavelength (Fig. 8). Plots that had low (or high) values of (Fig. 8). The correlation between reference and modeled directional
directional scattering coefficients throughout the spectrum tended to scattering coefficients was consistently high for the fitted model, with
have so also in the modeled data (Fig. 7). The site- and wavelength- mean r of 0.88, 0.88, and 0.96 for coniferous, broadleaved, and mixed
specific differences in the data were also correctly reproduced by the forests, respectively (Fig. 8).
model. As an example, broadleaved canopies in Järvselja tended to have The plant element albedo obtained through model fitting (as a linear
strong scattering in NIR and weak scattering in red, and the same combination of shoot and woody element albedos) was lower than pure
behavior was observed also in the modeled data (middle column of shoot albedo in green and NIR wavelengths (Fig. 9). In the red and blue
Fig. 7). The largest relative differences between modeled and reference wavelengths, the plant element albedo obtained through model fitting
values of canopy scattering were observed in the VIS region in the was higher than shoot albedo, and in the SWIR the difference between

11
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 7. Modeled and reference (i.e., estimated from measurements) canopy directional scattering coefficients, ωC(↓sky,Ω), as function of effective plant area index
(Leff) at four wavelengths (rows) in coniferous, broadleaved, and mixed forest plots (columns). The diamond symbols represent the default model (i.e., assuming a
Lambertian canopy that is composed of foliage only), square symbols the fitted model (accounting for nadir to hemispherical scattering ratio and contribution of
woody elements), and the x symbols the reference. The vertical lines illustrate 95% confidence intervals for the default model, and for the reference ωC(↓sky,Ω).

the two depended on species but was relatively small. The fraction of showing higher QΩ values than spruce. Note that because the model
woody elements obtained in model fitting was 0.32 for pine, 0.30 for inversion was not performed for plots with Leff lower than 1, we set the
spruce, and 0.12 for broadleaved tree species. The mean (and range) of QΩ value for those plots equal to average QΩ of all remaining plots (0.58)
nadir to hemispherical scattering ratio of the canopy (QΩ) was 0.55 when modeling the forest reflectance (Section 4.4). For comparison,
(0.32–0.92) in coniferous, 0.58 (0.36–0.84) in broadleaved, and 0.71 theoretical estimates of QΩ (iΩ/iD), and canopy-level photon recollision
(0.65–0.75) in mixed forests. The value of QΩ depended on Leff (Fig. 10 probability (p), are shown in the middle and bottom rows of Fig. 10, and
top row), with sparse canopies (low Leff) having on average low QΩ will be interpreted in Section 5.2.
values. In other words, compared to sparse canopies, dense canopies
tended to have scattering properties closer to Lambertian. The depen­
dence of QΩ on Leff was particularly evident in broadleaved plots. In 4.4. Modeled forest reflectance spectra
conifers, pine and spruce plots deviated from each other, with pine
The results regarding forest reflectance, R(↓sky,Ω), were similar to the

12
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 8. Top and middle rows: Root mean square error (RMSE) and mean estimation error (MEE) of the modeled canopy directional scattering coefficient, ωC(↓sky,Ω),
relative to the mean of reference ωC(↓sky,Ω). Bottom row: Pearson correlation coefficient (r) between modeled and reference ωC(↓sky,Ω). Left column shows results for
the default model (i.e., assuming a Lambertian canopy that is composed of foliage only) and right column for the fitted model (accounting for nadir to hemispherical
scattering ratio and contribution of woody elements). Mean (and range) of the evaluation metrics are listed in each sub-figure.

results of canopy directional scattering coefficients shown above. Once 5. Discussion

again, the default model resulted in notable overestimation in the NIR
region (Fig. 11). The MEE values ranged from − 22% to 188% (conif­ 5.1. Evaluation of the model
erous), from 11% to 117% (broadleaved), and from − 17% to 124%
(mixed forests) (Fig. 12). The fitted model showed good agreement with For the first time, a photon recollision probability (p) based approach
the measured forest reflectance (Fig. 11). The MEE values ranged from to modeling forest reflectance was tested in a wide range of differently
− 45% to 20% (coniferous), from − 28% to 27% (broadleaved), and from structured forests from different biomes. The measurements were
− 30% to 19% (mixed forests) (Fig. 12). The mean r between measured detailed, comprising, e.g., measurement of forest floor reflectance
and modeled forest reflectance were 0.37 (coniferous), 0.10 (broad­ spectrum in each study plot individually. Our spectral data (450–2200
leaved), and 0.88 (mixed forests) for the default model, and 0.91 nm) covered almost all solar wavelengths commonly used in optical
(coniferous), 0.77 (broadleaved), and 0.97 (mixed forests) for the fitted remote sensing. The main findings are related to the parameterization of
model (Fig. 12). To support quantitative comparisons with other studies, the p-based part of the model that estimates canopy scattering. Large
the MEE, RMSE, and r values for selected spectral bands have been overestimation occurred, especially in the NIR region, when the model
tabulated in Supplementary material Tables S2–S5. was parameterized considering only leaves or needles as plant elements

13
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 9. Plant element albedo spectra of pine, spruce, and broadleaved species. Foliage albedo (ωL) was converted to shoot albedo (ωS), using photon recollision
probability within a shoot (for broadleaved species ωS equals ωL). The default model used the shoot albedos as such. In the fitted model, the plant element albedo (ωE)
was obtained by least squares fitting, as a linear combination of woody element (ωW) and shoot albedo. The fW (95% confidence intervals in brackets) is the fraction
of woody elements from effective plant area that produced the ωE values. Top row shows the entire spectrum, and bottom row an enlarged view of the visible region.
The colored regions show 95% confidence intervals for the ωL, ωW, and ωE spectra.

and assuming a Lambertian canopy. This is the way that the previous Stenberg (2008), we obtained NIR overestimation of similar magnitude
version of PARAS canopy reflectance model has often been parameter­ as in Hadi and Rautiainen (2018). We also note that the somewhat larger
ized (e.g., Hadi and Rautiainen, 2018; Hadi et al., 2017; Rautiainen and discrepancies between modeled and reference in the VIS region in the
Stenberg, 2005; Schraik et al., 2019), not because it is believed to be Bílý Kříž site compared to the other sites (Figs. 7, 11) could have been
perfectly correct, but because measurements of woody element spectra caused by slight inaccuracies in the airborne data, because a comparison
have not been readily available, and because the model itself has not with Landsat 8 OLI data indicated some deviation of the airborne data in
provided a formula for modeling the scattering directionality. Our re­ Bílý Kříž from the rest of the sites in the VIS region (Supplementary
sults now indicate that with careful attention to the parameterization, material Fig. S2). We discuss the other differences and similarities be­
the model can correctly simulate the species- and wavelength specific tween modeled and reference data, as well as the role of model pa­
relations of both canopy scattering and entire forest reflectance to plant rameters, in the next section.
area index.
Quantitatively, the results of our fitted model are in line with results 5.2. New parameterization of the model
obtained in a laboratory with small single trees, where relative RMSE up
to 57% was obtained for the modeled tree reflectance (Hovi et al., 2020). The inclusion of woody elements helped to reduce the model’s large
Similarly to our results, an evaluation of an earlier PARAS model version overestimation of canopy scattering in the NIR region. To explain this,
found that the largest overestimation occurred in the NIR region (Hadi we look into Eq. 7. In forest canopies, the values of photon recollision
and Rautiainen, 2018). The overestimation was however not as large as probability (p) are usually relatively high (0.53–0.85, Fig. 10) compared
we observed here. In the referred study, the model was parameterized to those observed for single trees (e.g., approximately 0.1–0.6, see
using a separate parameter (Q) for calculating the ratio of radiation Fig. 10 in Hovi et al., 2020). Thus, multiple scattering occurs and con­
scattered to the upper hemisphere to the total scattered radiation tributes strongly to the canopy scattering in forests. The effect of mul­
(Mõttus and Stenberg, 2008). This, together with the fact that multiple tiple scattering on canopy reflectance is the largest when the plant
scattering between forest floor and the canopy was ignored in the pre­ element albedo is high. Using Eq. 7, it can be calculated that when the
vious model version, meant that a certain fraction of radiation scattered element albedo is 0.9, a reduction of 0.1 units in the element albedo
by the canopy was always directed towards the forest floor and did not results in reductions of 21% to 40% in canopy scattering, assuming p is
contribute to the forest reflectance. We note that when running the 0.6 and 0.9, respectively. In other words, the canopy scattering is very
earlier model version and using the Q parameter from Mõttus and sensitive to changes in element albedo, when the canopy is dense and

14
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 10. Nadir to hemispherical scattering ratio of the forest canopy (QΩ) obtained through model fitting (top row), theoretical estimate of QΩ (middle row), and
photon recollision probability (p) calculated from hemispherical photographs (bottom row) as function of effective plant area index (Leff) in coniferous, broadleaved,
and mixed forest plots (columns). The vertical lines illustrate 95% confidence intervals.

the element albedo is high. Inclusion of woody elements in the model should, however, be interpreted as approximations. Measurement and
reduced the modeled canopy scattering in NIR, because the NIR albedo sampling uncertainties in foliage spectra (Fig. 9) can also partly explain
of woody elements is clearly lower than that of foliage or shoots (Fig. 9). the NIR overestimation, as seen in Fig. 7 and Fig. 11. Other parameters
At the same time, the woody elements increased the average plant (p and QΩ) are possible explanations as well, but, as will be discussed
element albedo in the red and blue wavelength regions, thus making the next, changing their values would not reduce the discrepancy in model
model performance more consistent across wavelength regions. performance between NIR and the other wavelengths.
Our study supports earlier findings that it is important to consider the Previous studies have emphasized the importance of p, and thus
woody elements in forest reflectance simulations (Kuusinen et al., 2021; canopy clumping, in controlling the reflectance properties of forest
Malenovský et al., 2008; Verrelst et al., 2010). A meta-analysis of studies canopies (e.g., Rautiainen and Stenberg, 2005; Hovi et al., 2017). We
conducted for different broadleaved and coniferous tree species reported estimated p from field measurements and accounted for clumping at
woody element percentage from total plant area to vary between 5 and shoot level (through using the shoots as a basic plant element) and
35% (Gower et al., 1999). Corresponding values for all our study species higher levels (through applying a canopy clumping coefficient in esti­
are not readily available, but in a limited case study, Stenberg et al. mating true plant area index, and thus canopy-level p). As discussed by
(2003) reported values of 13% and 26% for Scots pine forests on poor Ryu et al. (2010), our correction for higher than shoot level clumping is
and fertile soils, respectively. Our woody element percentage (32%, likely to represent a minimum, and the values of p can in reality be
30%, and 12% from effective plant area, i.e., 22%, 21%, and 12% from somewhat higher than what we estimated. Fine-tuning the p estimates
total plant area for pine, spruce, and broadleaved tree species, respec­ through model inversion was not possible, because in our model p is
tively) are therefore within a physically meaningful range. The values correlated with QΩ (i.e., both affect the directional scattering coefficient

15
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 11. Modeled and reference (measured) hemispherical-directional reflectance factors, R(↓sky,Ω), of the forest as function of effective plant area index (Leff) at four
wavelengths (rows) in coniferous, broadleaved, and mixed forest plots (columns). The diamond symbols represent the default model (i.e., assuming a Lambertian
canopy that is composed of foliage only), square symbols the fitted model (accounting for nadir to hemispherical scattering ratio and contribution of woody ele­
ments), and the x symbols the reference. The vertical lines illustrate 95% confidence intervals for the default model.

of the forest canopy in a similar manner) and thus both variables are photons that scatter from the canopy for the first time. The other
difficult to estimate reliably at the same time. Confidence intervals in extreme is a theoretical situation where the plant element albedo is one.
Fig. 10 indicate that an increase in p could increase the estimates of QΩ. In that case, the canopy scattering coefficient does not depend on p at all
However, we also argue that the effect of increasing the p value would be because, independent of the p value, eventually the photons will escape
the highest in wavelength regions where plant element albedo is low (e. the canopy. When plant element albedo increases (e.g., when moving
g., in the VIS region), and thus would result in undesired effect: reduc­ from VIS to the NIR region), we approach this theoretical situation, and
tion of simulated forest reflectance in the VIS region but not as much in the sensitivity of canopy scattering to p decreases.
the NIR region. This is because when plant element albedo is low, ra­ Together with woody elements, the nadir to hemispherical scattering
diation attenuates rapidly after first interaction with the canopy and, ratio of the canopy (QΩ) was important in reducing the overall level of
given unchanged plant element albedo, the canopy scattering coefficient simulated forest reflectance. We obtained QΩ values from 0.32 to 0.92 in
is almost directly determined by the escape probability (1 – p) for our inversion, and also the confidence intervals were heavily weighted

16
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Fig. 12. Top and middle rows: Root mean square error (RMSE) and mean estimation error (MEE) of the modeled forest hemispherical-directional reflectance factor,
R(↓sky,Ω), relative to the mean of reference R(↓sky,Ω). Bottom row: Pearson correlation coefficient (r) between modeled and reference R(↓sky,Ω). Left column shows
results for the default model (i.e., assuming a Lambertian canopy that is composed of foliage only) and right column for the fitted model (accounting for nadir to
hemispherical scattering ratio and contribution of woody elements). Mean (and range) of the evaluation metrics are listed in each sub-figure.

towards values below one. This indicates that in the majority of studied inversion are physically meaningful.
canopies, there is less scattering to nadir than to other hemispherical Based on our data and earlier publications (e.g., Hadi et al., 2017;
directions on average. The QΩ values are impossible to evaluate based on Hovi et al., 2017), the average gap fraction of forest canopies tends to be
our airborne data, because our data only cover close-to-nadir view an­ the largest in nadir and approaches zero when the canopy is viewed from
gles (±20◦ ). A previous study reported nadir and hemispherical reflec­ oblique angles. The contrast in gap fractions measured at nadir and
tance values based on above-canopy multi-angular measurements at red, oblique angles is the largest when the tree cover is low (Gerard and
NIR, and SWIR wavelengths in boreal coniferous and broadleaved for­ North, 1997). Thus, compared to dense canopies, sparse canopies have
ests (Deering et al., 1999). Based on those results, it can be calculated less plant area visible to the sensor in nadir (Fig. 5), which means that
that QΩ values ranged between 0.42 and 1.19. Those values cannot be less radiation is likely to be scattered to nadir than to oblique directions.
directly compared to our results because our QΩ is defined for the can­ An interesting question is whether a sufficiently accurate, yet simple
opy only (forest floor excluded), but they give some evidence that the analytical formula for QΩ could be included in the model. The good
obtained range of QΩ is correct. Interestingly, we observed QΩ to in­ qualitative agreement of QΩ with the theoretical estimate iΩ/iD (Fig. 10)
crease as a function of Leff. In preliminary ray tracing simulations at red suggests that such a formula could possibly utilize multi-angular canopy
wavelength (simulated forests composed of ellipsoid crowns filled with interception values measured from, e.g., hemispherical photos or
randomly distributed and oriented, circular leaves), we noticed a similar terrestrial laser scanning. Quantitatively, there remained some
increasing trend as a function of Leff, and the QΩ values ranged between discrepancy, i.e., inverted values of QΩ were on average 29% lower than
0.39 and 0.82. This suggests that the values obtained through our iΩ/iD, which can be due to both uncertainty of our inversion as well as

17
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

deviation of real canopies from the assumptions behind the theoretical match with the empirical measurements motivated us to find the
estimate. Other factors contributing to the angular distribution of re­ explanation and to provide a solution that improved the model perfor­
flected radiation, and thus probably also the nadir to hemispherical mance. We also made an interesting discovery about the spectral prop­
scattering ratio, are crown shape (Rautiainen et al., 2004), crown height erties of forests: we observed that even though the plant element albedos
distribution (Gerard and North, 1997), and sun zenith angle (Deering are not dramatically different between coniferous and broadleaved
et al., 1999; Gerard and North, 1997), which all influence shadow forests, they can still explain a large part of differences between these
fraction and its dependence on view-direction. forest types in the NIR reflectance. This is because, due to multiple
In this study, our focus was on near-nadir measurements, but the scattering, the relation between plant element albedo and forest canopy
model could be applied also to simulations in other view directions, as scattering is nonlinear. Thus, the forest reflectance, especially in dense
long as QΩ could be adequately estimated. We note, however, that forests, is very sensitive to variation in plant element albedo in the NIR
predicting QΩ near the hotspot region can be difficult, because close to wavelength region where the plant element albedo is high.
the hotspot the gap fraction in direction of view (and thus escape
probability) becomes highly correlated with gap fraction in direction of 6. Conclusions
the sun (Kuusk, 1991; Yang et al., 2017). This correlation depends, in
addition to other canopy structural factors, also on leaf size and orien­ We conducted an extensive (in terms of geographical coverage and
tation. Therefore, simple parameterizations of QΩ based on macroscale forest structural variation) empirical evaluation of a photon recollision
canopy structure such as discussed in the previous paragraph might not probability based forest reflectance model. Based on our results, we
be sufficient for hotspot simulations. conclude that using foliage albedo as the plant element albedo and
assuming a Lambertian canopy can lead to large errors when simulating
5.3. Implications the forest reflectance in a typical near-nadir observation geometry of
satellite sensors. On the other hand, the model simulates forest spectra
What are then the broader implications of our study, beyond the and their dependence on effective plant area index of the forest canopy
technical details and parameters of the p-based model? First, cost- correctly, if effective plant element albedo (which combines foliage and
efficient and computationally simple methods, such as the physically- woody elements) and directional scattering properties of the canopy are
based model tested here, are becoming more and more important, due used as inputs. Future studies should develop methods for approxi­
to the increase in the variety of satellite data that is available. New mating the directional scattering properties (i.e., forest BRDF), and
hyperspectral satellite missions (e.g., EnMAP), in particular, increase the determine effective plant element albedos for different tree species and
need for testing the models in a wide range of wavelengths. By using biomes based on measurements of spatial and temporal variation of both
hyperspectral data, we showed that, if parameterized accurately, the woody element and foliage spectra.
model based on photon recollision probability performs well throughout
the spectrum. In other words, the geometry (structure) of the canopy, Declaration of Competing Interest
together with visibility of the forest floor, determine the overall
‘brightness’ of the forest throughout the spectrum. Thus, few spectrally The authors declare that they have no known competing financial
invariant parameters and the plant element albedo can be enough to interests or personal relationships that could have appeared to influence
describe the reflectance characteristics of the forest. the work reported in this paper.
The concept of spectral invariants has been used in interpretation of
global multispectral satellite datasets: mapping of leaf area index and Acknowledgments
fraction of absorbed photosynthetically active radiation using MODIS
(Knyazikhin et al., 1998; Myneni et al., 2002) and VIIRS (Yan et al., We would like to thank Tomáš Fabiánek for significant contributions
2018) data, and mapping of leaf area index and its sunlit portion using in airborne data processing. Thanks also to Lukáš Fajmon, Petri For­
DSCOVR EPIC (Yang et al., 2017) data. Retrieval of leaf area index from sström, Karel Holouš, Mihkel Kaha, Bijay Karki, Lauri Korhonen, Nea
Landsat multispectral satellite data has also been demonstrated Kuusinen, Andres Kuusk, Titta Majasalmi, Matti Mõttus, Ville Ranta, and
(Ganguly et al., 2012). Models belonging to the PARAS family, in the staff of Hyytiälä forestry field station for assistance and comments in
particular, have been used in the inversion of forest leaf area index and various stages of field work, data processing, and interpretation. This
its seasonal variation from medium resolution (Landsat, Sentinel-2) study has received funding from the European Research Council (ERC)
multispectral satellite data (Schraik et al., 2019; Varvia et al., 2018). under the European Union’s Horizon 2020 research and innovation
The relative simplicity of these models makes them particularly suitable programme (grant agreement No 771049). The text reflects only the
when using computationally intensive (inversion) techniques (e.g., authors’ view and the Agency is not responsible for any use that may be
Schraik et al., 2019). For example, the PARAS model was used for made of the information it contains. This study was also funded by the
sensitivity analyses quantifying the driving factors of forest reflectance Academy of Finland grants BOREALITY (286390) and DIMEBO
(Hadi et al., 2017; Hovi and Rautiainen, 2020). We hope that the results (323004). This study was also supported by the Ministry of Education,
presented here will further help to facilitate the use of these models in Youth and Sports of the Czech Republic within the CzeCOS program,
practice, including (but not limited to) the above mentioned grant number LM2018123. Petr Lukeš acknowledges support by the
applications. project LTAUSA18154 awarded by the Ministry of Education, Youth and
Our study also revealed new aspects about the role of leaf (and Sports of the Czech Republic. Jan Pisek acknowledges support by the
needle) spectra in forest reflectance and land surface modeling. Based on Eesti Teadusagentuur (grant no. PUT1355).
our results, it appears that using a leaf or needle spectrum as input is not
sufficient for modeling a forest reflectance spectrum accurately. Rather,
Appendix A. Supplementary data
one should be able to determine, perhaps with the help of measurements
or spectral libraries, the effective plant element albedos that would be
Supplementary data to this article can be found online at https://doi.
used as model input. Probably, if the effective plant element albedos are
org/10.1016/j.rse.2021.112804.
species- or biome-specific, this kind of model calibration would not be
needed for each and every dataset or study area. This would make the
References
model widely applicable. Overall, we would like to emphasize the
importance of empirical measurements in evaluation and development Baret, F., Buis, S., 2008. Estimating canopy characteristics from remote sensing
of forest reflectance models. The finding that the default model did not observations: Review of methods and associated problems. In: Liang, S. (Ed.),

18
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Advances in Land Remote Sensing: System. Modeling, Inversion and Application. Myneni, R.B., Hoffman, S., Knyazikhin, Y., et al., 2002. Global products of vegetation leaf
Springer, pp. 147–172. area and fraction absorbed PAR from year one of MODIS data. Remote Sens. Environ.
Branch, M.A., Coleman, T.F., Li, Y., 1999. A subspace, interior, and conjugate gradient 83, 214–231. https://doi.org/10.1016/S0034-4257(02)00074-3.
method for large-scale bound-constrained minimization problems. SIAM J. Sci. Nobis, M., Hunziker, U., 2005. Automatic thresholding for hemispherical canopy-
Comput. 21, 1–23. https://doi.org/10.1137/S1064827595289108. photographs based on edge detection. Agric. For. Meteorol. 128, 243–250. https://
Deering, D.W., Eck, T.F., Banerjee, B., 1999. Characterization of the reflectance doi.org/10.1016/j.agrformet.2004.10.002.
anisotropy of three boreal forest canopies in spring-summer. Remote Sens. Environ. Panferov, O., Knyazikhin, Y., Myneni, R.B., Szarzynski, J., Engwald, S., Schnitzler, K.G.,
67, 205–229. https://doi.org/10.1016/S0034-4257(98)00087-X. Gravenhorst, G., 2001. The role of canopy structure in the spectral variation of
Disney, M., Lewis, P., Quaife, T., Nichol, C., 2005. A spectral invariant approach to transmission and absorption of solar radiation in vegetation canopies. IEEE Trans.
modelling canopy and leaf scattering. In: ISPMSRS 2005 Conference Proceedings, Geosci. Remote Sens. 39, 241–253. https://doi.org/10.1109/36.905232.
Beijing, China, pp. 318–320. Rast, M., Painter, T.H., 2019. Earth observation imaging spectroscopy for terrestrial
Ganguly, S., Nemani, R.R., Zhang, G., et al., 2012. Generating global Leaf Area Index systems: an overview of its history, techniques, and applications of its missions. Surv.
from Landsat: algorithm formulation and demonstration. Remote Sens. Environ. 122, Geophys. 40, 303–331. https://doi.org/10.1007/s10712-019-09517-z.
185–202. https://doi.org/10.1016/j.rse.2011.10.032. Rautiainen, M., 2005. Retrieval of leaf area index for a coniferous forest by inverting a
Gerard, F.F., North, P.R.J., 1997. Analyzing the effect of structural variability and canopy forest reflectance model. Remote Sens. Environ. 99, 295–303. https://doi.org/
gaps on forest BRDF using a geometric-optical model. Remote Sens. Environ. 62, 10.1016/j.rse.2005.09.004.
46–62. https://doi.org/10.1016/S0034-4257(97)00070-9. Rautiainen, M., Stenberg, P., 2005. Application of photon recollision probability in
Gower, S.T., Kucharik, C.J., Norman, J.M., 1999. Direct and Indirect Estimation of Leaf coniferous canopy reflectance simulations. Remote Sens. Environ. 96, 98–107.
Area Index, fAPAR, end net primary production of terrestrial ecosystems. Remote https://doi.org/10.1016/j.rse.2005.02.009.
Sens. Environ. 70, 29–51. https://doi.org/10.1016/S0034-4257(99)00056-5. Rautiainen, M., Stenberg, P., Nilson, T., Kuusk, A., 2004. The effect of crown shape on
Hadi, Rautiainen, M., 2018. A study on the drivers of canopy reflectance variability in a the reflectance of coniferous stands. Remote Sens. Environ. 89, 41–52. https://doi.
boreal forest. Remote Sens. Lett. 9, 666–675. https://doi.org/10.1080/ org/10.1016/j.rse.2003.10.001.
2150704X.2018.1458344. Rautiainen, M., Mõttus, M., Yáñez-Rausell, L., Homolová, L., Malenovský, Z.,
Hadi, Pfeifer, M., Korhonen, L., Wheeler, C., Rautiainen, M., 2017. Forest canopy Schaepman, M.E., 2012. A note on upscaling coniferous needle spectra to shoot
structure and reflectance in humid tropical Borneo: a physically-based interpretation spectral albedo. Remote Sens. Environ. 117, 469–474. https://doi.org/10.1016/j.
using spectral invariants. Remote Sens. Environ. 201, 314–330. https://doi.org/ rse.2011.10.019.
10.1016/j.rse.2017.09.018. Richter, R., Schläpfer, D., 2018. Atmospheric / topographic correction for airborne
Hovi, A., Rautiainen, M., 2020. Spectral composition of shortwave radiation transmitted imagery. ATCOR-4 User Guide, Version 7.2.0 (March 2018).
by forest canopies. Trees 34, 1499–1506. https://doi.org/10.1007/s00468-020- Ryu, Y., Nilson, T., Kobayashi, H., Sonnentag, O., Law, B.E., Baldocchi, D.D., 2010. On
02005-7. the correct estimation of effective leaf area index: does it reveal information on
Hovi, A., Lukeš, P., Rautiainen, M., 2017. Seasonality of albedo and FAPAR in a boreal clumping effects? Agric. For. Meteorol. 150, 463–472. https://doi.org/10.1016/j.
forest. Agric. For. Meteorol. 247, 331–342. https://doi.org/10.1016/j. agrformet.2010.01.009.
agrformet.2017.08.021. Ryu, Y., Berry, J.A., Baldocchi, D.D., 2019. What is global photosynthesis? History,
Hovi, A., Forsström, P., Ghielmetti, G., Schaepman, M.E., Rautiainen, M., 2020. uncertainties and opportunities. Remote Sens. Environ. 223, 95–114. https://doi.
Empirical validation of photon recollision probability in single crowns of tree org/10.1016/j.rse.2019.01.016.
seedlings. ISPRS J. Photogramm. Remote Sens. 169, 57–72. https://doi.org/ Savitzky, A., Golay, M.J.E., 1964. Smoothing and differentiation of data by simplified
10.1016/j.isprsjprs.2020.08.027. least squares procedures. Anal. Chem. 36, 1627–1639. https://doi.org/10.1021/
Huang, D., Knyazikhin, Y., Dickinson, R.E., et al., 2007. Canopy spectral invariants for ac60214a047.
remote sensing and model applications. Remote Sens. Environ. 106, 106–122. Schneider, F.D., Morsdorf, F., Schmid, B., Petchey, O.L., Hueni, A., Schimel, D.S.,
https://doi.org/10.1016/j.rse.2006.08.001. Schaepman, M.E., 2017. Mapping functional diversity from remotely sensed
Juola, J., Hovi, A., Rautiainen, M., 2021. A dataset of stem bark reflectance spectra for morphological and physiological forest traits. Nat. Commun. 8 https://doi.org/
boreal and temperate tree species. Mendeley Data V1. https://doi.org/10.17632/ 10.1038/s41467-017-01530-3.
pwfxgzz5fj.1. Schraik, D., Varvia, P., Korhonen, L., Rautiainen, M., 2019. Bayesian inversion of a forest
Knyazikhin, Y., Martonchik, J.V., Myneni, R.B., Diner, D.J., Running, S.W., 1998. reflectance model using Sentinel-2 and Landsat 8 satellite images. J. Quant.
Synergistic algorithm for estimating vegetation canopy leaf area index and fraction Spectrosc. Radiat. Transf. 233, 1–12. https://doi.org/10.1016/j.jqsrt.2019.05.013.
of absorbed photosynthetically active radiation from MODIS and MISR data. Shabanov, N.V., Wang, Y., Buermann, W., et al., 2003. Effect of foliage spatial
J. Geophys. Res. 103, 32257–32275. https://doi.org/10.1029/98JD02462. heterogeneity in the MODIS LAI and FPAR algorithm over broadleaf forests. Remote
Kuusinen, N., Lukeš, P., Stenberg, P., Levula, J., Nikinmaa, E., Berninger, F., 2014. Sens. Environ. 85, 410–423. https://doi.org/10.1016/S0034-4257(03)00017-8.
Measured and modelled albedos in Finnish boreal forest stands of different species, Smolander, S., Stenberg, P., 2003. A method to account for shoot scale clumping in
structure and understory. Ecol. Model. 284, 10–18. https://doi.org/10.1016/j. coniferous canopy reflectance models. Remote Sens. Environ. 88, 363–373. https://
ecolmodel.2014.04.007. doi.org/10.1016/j.rse.2003.06.003.
Kuusinen, N., Hovi, A., Rautiainen, M., 2021. Contribution of woody elements to tree Smolander, S., Stenberg, P., 2005. Simple parameterizations of the radiation budget of
level reflectance in boreal forests. Silva Fennica 55, 10600. https://doi.org/ uniform broadleaved and coniferous canopies. Remote Sens. Environ. 94, 355–363.
10.14214/sf.10600. https://doi.org/10.1016/j.rse.2004.10.010.
Kuusk, A., 1991. The hot spot effect in plant canopy reflectance. In: Myneni, R.B., Ross, J. Spencer, S., Rock, B.N., 1999. BOREAS TE-08 Aspen Bark Spectral Reflectance Data. Data
(Eds.), 1991. Photon-Vegetation Interactions: Applications in Optical Remote Set. Oak Ridge National Laboratory Distributed Active Archive Center, Oak Ridge,
Sensing and Plant Ecology. Springer-Verlag, Berlin Heidelberg. Tennessee, U.S.A. https://doi.org/10.3334/ORNLDAAC/336 [WWW Document].
Kuusk, A., Kuusk, J., Lang, M., 2014. Measured spectral bidirectional reflection (Accessed 31 March 2021).
properties of three mature hemiboreal forests. Agric. For. Meteorol. 185, 14–19. Stenberg, P., 2007. Simple analytical formula for calculating average photon recollision
https://doi.org/10.1016/j.agrformet.2013.10.011. probability in vegetation canopies. Remote Sens. Environ. 109, 221–224. https://
Lang, M., Kuusk, A., Nilson, T., Lükk, T., Pehk, M., Alm, G., 2002. Reflectance spectra of doi.org/10.1016/j.rse.2006.12.014.
ground vegetation in sub-boreal forests. Tartu Observatory, Estonia [WWW Stenberg, P., Nilson, T., Smolander, H., Voipio, P., 2003. Gap fraction based estimation of
Document]. http://www.aai.ee/bgf/ger2600/ (Accessed 31 March 2021). LAI in Scots pine stands subjected to experimental removal of branches and stems.
LI-COR, 2012. LAI-2200 plant canopy analyzer instruction manual. LI.COR, Inc., Can. J. Remote. Sens. 29, 363–370. https://doi.org/10.5589/m03-007.
publication number 984–10633 rev 2. https://www.licor.com/documents/6 Stenberg, P., Rautiainen, M., Manninen, T., Voipio, P., Mõttus, M., 2008. Boreal forest
n3conpja6uj9aq1ruyn. leaf area index from optical satellite images: model simulations and empirical
Lukeš, P., Rautiainen, M., Stenberg, P., Malenovský, Z., 2011. Empirical test of the analyses using data from central Finland. Boreal Environ. Res. 13, 433–443.
spectral invariants theory using imaging spectroscopy data from a coniferous forest. Stenberg, P., Lukeš, P., Rautiainen, M., Manninen, T., 2013. A new approach for
Int. J. Appl. Earth Obs. Geoinf. 13, 668–675. https://doi.org/10.1016/j. simulating forest albedo based on spectral invariants. Remote Sens. Environ. 137,
jag.2011.04.003. 12–16. https://doi.org/10.1016/j.rse.2013.05.030.
Majasalmi, T., Rautiainen, M., Stenberg, P., 2014. Modeled and measured fPAR in a Stenberg, P., Mõttus, M., Rautiainen, M., 2016. Photon recollision probability in
boreal forest: validation and application of a new model. Agric. For. Meteorol. modelling the radiation regime of canopies - a review. Remote Sens. Environ. 183,
189–190, 118–124. https://doi.org/10.1016/j.agrformet.2014.01.015. 98–108. https://doi.org/10.1016/j.rse.2016.05.013.
Malenovský, Z., Martin, E., Homolová, L., Gastellu-Etchegorry, J.P., Zurita-Milla, R., Thérézien, M., Palmroth, S., Brady, R., Oren, R., 2007. Estimation of light interception
Schaepman, M.E., Pokorný, R., Clevers, J.G.P.W., Cudlín, P., 2008. Influence of properties of conifer shoots by an improved photographic method and a 3D model of
woody elements of a Norway spruce canopy on nadir reflectance simulated by the shoot structure. Tree Physiol. 27, 1375–1387. https://doi.org/10.1093/treephys/
DART model at very high spatial resolution. Remote Sens. Environ. 112, 1–18. 27.10.1375.
https://doi.org/10.1016/j.rse.2006.02.028. Varvia, P., Rautiainen, M., Seppänen, A., 2018. Bayesian estimation of seasonal course of
Manninen, T., Stenberg, P., 2009. Simulation of the effect of snow covered forest floor on canopy leaf area index from hyperspectral satellite data. J. Quant. Spectrosc. Radiat.
the total forest albedo. Agric. For. Meteorol. 149, 303–319. https://doi.org/ Transf. 208, 19–28. https://doi.org/10.1016/j.jqsrt.2018.01.008.
10.1016/j.agrformet.2008.08.016. Verrelst, J., Schaepman, M.E., Malenovský, Z., Clevers, J.G.P.W., 2010. Effects of woody
Mõttus, M., Stenberg, P., 2008. A simple parameterization of canopy reflectance using elements on simulated canopy reflectance: implications for forest chlorophyll
photon recollision probability. Remote Sens. Environ. 112, 1545–1551. https://doi. content retrieval. Remote Sens. Environ. 114, 647–656. https://doi.org/10.1016/j.
org/10.1016/j.rse.2007.08.002. rse.2009.11.004.

19
A. Hovi et al. Remote Sensing of Environment 269 (2022) 112804

Virtanen, P., Gommers, R., Oliphant, T.E., et al., 2020. SciPy 1.0: fundamental Woodcock, C.E., Collins, J.B., Gopal, S., et al., 1994. Mapping forest vegetation using
algorithms for scientific computing in Python. Nat. Methods 17, 261–272. https:// Landsat TM imagery and a canopy reflectance model. Remote Sens. Environ. 50,
doi.org/10.1038/s41592-019-0686-2. 240–254. https://doi.org/10.1016/0034-4257(94)90074-4.
Wang, W., Buermann, W., Stenberg, P., Smolander, H., Häme, T., Tian, Y., Hu, J., Woodcock, C.E., Collins, J.B., Jakabhazy, V.D., Li, X., Macomber, S.A., Wu, Y., 1997.
Knyazikhin, Y., Myneni, R.B., 2003. A new parameterization of canopy spectral Inversion of the Li-Strahler canopy reflectance model for mapping forest structure.
response to incident solar radiation: case study with hyperspectral data from pine IEEE Trans. Geosci. Remote Sens. 35, 405–414. https://doi.org/10.1109/36.563279.
dominant forest. Remote Sens. Environ. 85 (3), 304–315. https://doi.org/10.1016/ Wulder, M.A., Loveland, T.R., Roy, D., et al., 2019. Current status of Landsat program,
S0034-4257(03)00009-9. science, and applications. Remote Sens. Environ. 225, 127–147. https://doi.org/
Wang, W., Nemani, R., Hashimoto, H., Ganguly, S., Huang, D., Knyazikhin, Y., 10.1016/j.rse.2019.02.015.
Myneni, R., Bala, G., 2018. An interplay between photons, canopy structure, and Yan, K., Park, T., Chen, C., et al., 2018. Generating global products of LAI and FPAR from
recollision probability: a review of the spectral invariants theory of 3D canopy SNPP-VIIRS data: theoretical background and implementation. IEEE Trans. Geosci.
radiative transfer processes. Remote Sens. 10, 1805. https://doi.org/10.3390/ Remote Sens. 56, 2119–2137. https://doi.org/10.1109/TGRS.2017.2775247.
rs10111805. Yang, B., Knyazikhin, Y., Mõttus, M., et al., 2017. Estimation of leaf area index and its
Widlowski, J., Mio, C., Disney, M., et al., 2015. The fourth phase of the radiative transfer sunlit portion from DSCOVR EPIC: theoretical basis. Remote Sens. Environ. 198,
model intercomparison (RAMI) exercise: actual canopy scenarios and conformity 69–84. https://doi.org/10.1016/j.rse.2017.05.033.
testing. Remote Sens. Environ. 169, 418–437. https://doi.org/10.1016/j.
rse.2015.08.016.

20