Verrelst 2016 SIF-NPC RSE

Remote Sensing of Environment 176 (2016) 139–151
Contents lists available at ScienceDirect
Remote Sensing of Environment
journal homepage: www.elsevier.com/locate/rse
Evaluating the predictive power of sun-induced chlorophyll fluorescence

to estimate net photosynthesis of vegetation canopies: A SCOPE
modeling study
Jochem Verrelst a,⁎, Christiaan van der Tol b, Federico Magnani c, Neus Sabater a, Juan Pablo Rivera a,
Gina Mohammed d, Jose Moreno a
a
Image Processing Laboratory, Department of Earth Physics and Thermodynamics, University of Valencia, C/Catedrático José Beltrán 2, E-46980 Paterna, Valencia, Spain
b
Department of Water Resources, Faculty ITC, University of Twente, P.O. Box 217, 7500 AE Enschede, The Netherlands
c
Department of Agricultural Sciences, University of Bologna, Bologna, Italy
d
P&M Technologies, 66 Millwood Street, Sault Ste. Marie, Ontario, Canada
a r t i c l e i n f o a b s t r a c t
Article history: Progress in imaging spectroscopy technology and data processing can enable derivation of the complete sun-
Received 27 July 2015 induced chlorophyll fluorescence (SIF) emission spectrum. This opens up opportunities to fully exploit the use
Received in revised form 13 January 2016 of the SIF spectrum as an indicator of photosynthetic activity. Simulations performed with the coupled
Accepted 20 January 2016
fluorescence–photosynthesis model SCOPE were used to determine how strongly canopy-leaving SIF can be re-
Available online xxxx
lated to net photosynthesis of the canopy (NPC) for various canopy configurations. Regression analysis between
Keywords:
SIF retrievals and NPC values produced the following general findings: (1) individual SIF bands that were most
Sun-induced fluorescence sensitive to NPC were located around the first emission peak (SIFred) for heterogeneous canopy configurations
Photosynthesis (i.e., varying biochemistry, leaf, canopy variables); (2) using two SIF retrieval bands, e.g. O2-B at 687 nm and
SCOPE O2-A at 760 nm, or the red and NIR emission peaks at 685 nm and 740 nm, led to stronger correlations than
Canopy using only one band; (3) using the O2-B and the O2-A SIF retrieval bands was at least as effective as using the
Regression two emission peaks; (4) superior correlations were achieved by using the four main SIF retrieval bands (Hα,
Band analysis O2-B, water vapor, O2-A); and (5) further improvements may be obtained by exploiting the full SIF profile and
FLEX
by using an adaptive, nonlinear regression algorithm such as Gaussian processes regression (GPR). Relationships
can be due to variation in photosynthetic capacity (Vcmo), but also from variation in leaf optical and canopy struc-
tural variables such as chlorophyll content and leaf area index. Overall, modeling results suggest that sampling
the SIF profile in at least both O2-B and O2-A bands enables quantification photosynthetic activity of vegetation
with high accuracy.
© 2016 Elsevier Inc. All rights reserved.
1. Introduction (but which can also extend somewhat below or above that range,
e.g., 640–850 nm), and which contains useful information on the photo-
Sun-induced fluorescence (SIF) emitted by chlorophyll molecules is synthetic process (Franck, Juneau, & Popovic, 2002; Lichtenthaler &
one of three main de-excitation mechanisms for energy captured by Rinderle, 1988). Even though the canopy will produce different SIF
light harvesting pigments in plants. SIF emitted by vegetation is seen spectra under different environmental and structural conditions, the
as a meaningful indicator of plant stress (Van Wittenberghe et al., shape of SIF spectra preserves typical features. In general, the SIF spec-
2013), instantaneous plant photosynthetic function (e.g., carbon fixa- trum is composed of two peaks, one located in the red (SIFred) spectral
tion), and possibly gross primary productivity (GPP) at the ecosystem region with a maximum around 685 nm that is mainly attributed to
scale (Porcar-Castell et al., 2014). the fluorescence emission of Photosystem II (PSII), and the other located
Although the SIF flux emitted from the canopy is relatively small in the NIR (SIFNIR) with a maximum around 740 nm that is attributable
compared to reflected sunlight (about 1–5% in the near infrared; NIR), to both Photosystem I (PSI) and PSII (Baker, 2008; Papageorgiou &
it is a broadband spectrum that typically spans about 650–800 nm Govindjee, 2004).
With the advent of imaging spectrometers, the retrieval of SIF using
remote sensing technologies has become a novel area of research
⁎ Corresponding author at: Laboratory of Earth Observation (LEO), Image Processing
Laboratory (IPL), Cientific Parc, BOX 22085, University of Valencia, 46071, Paterna
(Alonso et al., 2007; Guanter et al., 2010; Meroni et al., 2009, 2010)
(Valencia), Spain. aimed primarily at mapping SIF from the site-specific (Damm et al.,
E-mail address: jochem.verrelst@uv.es (J. Verrelst). 2014; Daumard et al., 2012; Moya, Daumard, Moise, Ounis, & Goulas,
http://dx.doi.org/10.1016/j.rse.2016.01.018
0034-4257/© 2016 Elsevier Inc. All rights reserved.
140 J. Verrelst et al. / Remote Sensing of Environment 176 (2016) 139–151
2006; Perez-Priego, Zarco-Tejada, Miller, Sepulcre-Canto, & Fereres, The FLEX scientific studies have also investigated radiative transfer
2005; Zarco-Tejada, Gonzalez-Dugo, & Berni, 2012; Zarco-Tejada, Mo- modeling of the SIF signal through the leaf and canopy based on explicit
rales, Testi, & Villalobos, 2013) to the global scale (Frankenberg et al., leaf physiological descriptions. The SCOPE (Soil-Canopy Observation,
2011; Joiner et al., 2011). In general, two strategies to extract SIF from Photosynthesis and Energy Balance) model (Van der Tol, Berry,
passive detection methods have been pursued in recent years, Campbell, & Rascher, 2014; Van der Tol, Verhoef, Timmermans,
exploiting either atmospheric (telluric) absorption features due to oxy- Verhoef, & Su, 2009) has been coupled with new leaf fluorescence mod-
gen in the O2-A absorption region at 760 nm (Alonso et al., 2007; ules (resulting in Version 1.53). SCOPE combines the functionality of a
Guanter et al., 2010; Meroni et al., 2010), or solar Fraunhofer lines, Soil–Vegetation–Atmosphere-Transfer (SVAT) model with radiative
which are narrow dark lines (absorption features) in the solar spectrum transfer of reflected and emitted (thermal and fluorescent) radiation
in which irradiance is strongly reduced (e.g. in the NIR 740–770 nm and enables the theoretical quantification of the canopy-leaving SIF
spectral window) (Frankenberg et al., 2011; Guanter et al., 2012). Re- broadband spectrum and canopy fluxes, such as the net photosynthesis
gardless of the retrieval strategy, current remote sensing approaches of the canopy (NPC).
have mostly emphasized the second emission peak region (SIFNIR) for Now that derivation of the full SIF spectrum is possible, it creates op-
photosynthetic quantification, especially from satellite-based atmo- portunities to utilize more effectively the spectral information content
spheric sensors. For example, Frankenberg et al. (2011) and Guanter related to NPC. An imminent requirement is to identify which SIF wave-
et al. (2012) linked monthly-aggregated global SIFNIR observations lengths are most sensitive to NPC. This leads to the main objective of this
with global GPP products using (biome-specific) linear relationships. work: to analyze the sensitivity of single wavelengths, as well as combi-
More recently, at the local scale, Damm, Guanter, Paul-Limoges, et al. nations of SIF retrieval bands in estimation of NPC for various canopy
(2015) linked airborne SIFNIR measurements with eddy covariance configurations.
flux tower GPP data and found that relationships were not linear but For this purpose, a SCOPE modeling study was applied. Simulations
asymptotic when instantaneous rather than temporally aggregated of canopy-leaving SIF and NPC outputs were conducted for different
measured data were used, and relationships were also ecosystem- combinations of biochemical, leaf, canopy and micrometeorological
specific. variables. Simulated SIF spectra were subsequently analyzed with
One reason for the focus on the SIFNIR has been the absence of respect to their predictive power in estimating NPC. Specifically,
spaceborne sensors spectrally optimized to capture the full SIF emission the following aspects were investigated; (1) linear regression analy-
spectrum. To fill this gap, the European Space Agency (ESA) has been sis between individual SIF bands and NPC outputs; (2) linear regres-
conducting Phase A/B1 evaluations of a candidate Earth Explorer mission analysis between combined SIF bands and NPC outputs; and
sion dedicated to measurement of SIF in terrestrial vegetation. The Fluo- (3) adaptive, nonlinear machine learning regression between com-
rescence Explorer (FLEX) satellite, equipped with a Fluorescence bined SIF bands and NPC outputs.
Imaging Spectrometer (FLORIS) onboard, has recently been approved
as ESA's Earth Explorer 8 mission (ESA, 2015). FLEX will operate in a 2. Materials and methods
tandem mission with ESA's Sentinel-3 satellite, the latter to provide at-
mospheric and land surface data needed for atmospheric corrections 2.1. SCOPE
and accurate SIF characterizations. FLORIS will measure the radiance be-
tween 500 and 780 nm with a bandwidth between 0.3 nm and 2 nm The coupled fluorescence–photosynthesis model SCOPE simu-
(depending on wavelength), providing images with a 150 km swath lates photosynthesis, radiative transfer in the leaf and canopy, and
and 300 m pixel size (Kraft et al., 2013; Moreno, Asner, Bach, et al., surface energy balance (Van der Tol et al., 2009, 2014). SCOPE re-
2006). Such finely resolved spectral sampling will allow retrieval of cently became a virtual laboratory for studies on surface energy bal-
the full broadband fluorescence emission spectrum and related prod- ance (Timmermans, Su, Van der Tol, Verhoef, & Verhoef, 2013),
ucts such as Ftotal (i.e., integral of the fluorescence broadband spectrum). remote sensing thermal infrared measurements (Duffour, Olioso,
In addition, a novel airborne imaging spectrometer HyPlant has become Demarty, Van der Tol, & Lagouarde, 2015), and SIF-photosynthesis
available recently which demonstrates the potential of a FLORIS-type studies (Damm, Guanter, Paul-Limoges, et al., 2015; Verrelst et al.,
sensor (Rascher et al., 2015). HyPlant has an ultra-high spectral resolu- 2015; Zhang, Guanter, et al., 2014).
tion in the red and near-infrared spectral region (0.26 nm FWHM (Full For photosynthesis, SCOPE uses either the model of Von Caemmerer
Width at Half Maximum) in the 670–780 nm spectral range). This al- (2000, 2013) or Collatz, Ball, Grivet, and Berry (1991), Collatz, Ribas-
lows quantification of sun-induced fluorescence fluxes in physical Carbo, and Berry (1992). These physiological models originally were
units for SIFred and SIFNIR (Rossini et al., 2015), and eventually over developed to interpret measurements of leaf gas exchange. The
the full SIF spectral region at a local scale. Another airborne experiment main boundary conditions for photosynthesis are energy supply
demonstrated that GPP is most strongly related to SIFred at the O2-B ab- (light) and carbon dioxide diffusion into the leaf. The models calcu-
sorption band (Cheng et al., 2013), possibly due to the relevance of the late photosynthesis under the condition that these two aspects, the
red band to photosystem II processes (Baker, 2008). energy supply and the carbon dioxide flux, are in equilibrium.
Additional advances in signal retrieval and data processing are evi- Electron transport (i.e., transfer of the energy supply) is calculated
dent. Developments include the use of multiple absorption lines in classically from active fluorescence measurements (e.g. Genty,
both emission peak regions for SIF retrievals, and simulations of the in- Wonders, & Baker, 1990; Maxwell & Johnson, 2000; Weis & Berry,
fluences of different atmospheric conditions (Liu & Liu, 2014). For ex- 1987). However, active techniques typically are only feasible in a
ample, Zhao et al. (2014) examined SIF retrieval in five absorption laboratory or small scale field study due to requirements for saturat-
lines to allow reconstruction of the full SIF emission between 650 and ing light flashes, and for modulation of the measuring light beam
850 nm based on simulated data. Also within the FLEX scientific studies, when fluorescence is assessed in natural outdoor conditions
a spectral fitting method has been developed that, when combined with (Maxwell & Johnson, 2000; Van der Tol et al., 2014). For this reason,
an atmospheric correction algorithm, is able to reconstruct the full SIF the photosynthesis model in SCOPE is complemented with these al-
spectrum directly from top of atmosphere (TOA) radiance data ternative predictive models for fluorescence that serve to simulate
(Cogliati et al., 2015). Reconstruction of the full SIF spectrum will the fluorescence leaf emission efficiency, ε, as a function of weather
allow calculation of some other meaningful parameters relevant to de- conditions and photosynthesis parameters, normalized by the leaf
tection of plant stress status, such as the spectral positions and FWHM fluorescence emission efficiency in (near) dark or pre-dawn condi-
of the SIFred and SIFNIR peaks, and the area under the SIF emission tions. The model of Van der Tol et al. (2014), is a semi-empirical
curve (Subhash & Mohanan, 1997; Zhao et al., 2014). model based upon field and laboratory experiments of unstressed
J. Verrelst et al. / Remote Sensing of Environment 176 (2016) 139–151 141
and drought-stressed vegetation and hereon referred to as TB12 and is crucial as it determines the wavelength dependence of the sensitivity
TB12-D models, respectively. The model of Magnani et al. (2009), of SIF to micrometeorological conditions and photosynthesis parame-
Dayyoub, 2011), hereon referred to as MD12 model, has a more explicit ters. PSI emits SIF only in the near-infrared part of the spectrum, while
parameterization of fluorescence quenching mechanisms. In its most PSII emits over a wide spectrum (640–800 nm) and peaks in the red
recent development by Magnani (summarized in Mohammed et al., SIF. Assuming a constant PSI fluorescence implies that SIFred is more
2014), the module also incorporates effects on photosynthesis and fluo- sensitive to photosynthetic parameters and weather conditions than is
rescence of seasonal changes in PSII photoinhibition and sustained, SIFNIR (Porcar-Castell et al., 2014).
dark-adapted non-photochemical quenching (NPQ) (Porcar-Castell, Both the incident light on the individual leaves (E) and the prop-
2011). The MD12 module is not limited to empirical calibration (com- agation of SIF throughout the canopy are calculated with radiative
pared to the TB12 models) and is therefore able to reproduce interme- transfer models based on the Scattering of Arbitrarily Inclined Leaves
diate conditions using two additional variables, those being the rate (SAIL) model (Verhoef, 1984). SAIL is a 1-D vertical model that dis-
constant of sustained thermal dissipation (kNPQs) and the fraction of tributes the canopy into 60 horizontal layers with an optical thick-
functional reaction centers (qLs) (Porcar-Castell, 2011). ness of LAI/60 (where LAI is leaf area index). Within each layer,
The radiative transfer of incident light and SIF emission in the leaf is leaves are described by their optical properties (from Fluspect) and
handled with a separate sub-model, Fluspect. The Kubelka–Munk theo- their inclination. The zenith inclination distribution of the leaves is
ry is applied to the mesophyll layer of the leaf to calculate diffuse scat- described stochastically with user defined parameter values, and dis-
tering and absorption of both incident light, as in the model PROSPECT tribution of leaf orientations in the azimuthal (horizontal) direction
(Jacquemoud & Baret, 1990), and SIF. Two SIF spectra as published by is uniform. SCOPE also simulates thermal radiation, net radiation,
Franck et al. (2002), one for PSI and another for PSII, are used to convert soil heat flux and the turbulent heat fluxes through an aerodynamic
incident light at any depth in the leaf into fluorescence emission spectra resistance scheme. Leaf temperature and leaf boundary layer gas
for both photosystems. The output of Fluspect consists of reflectance concentrations are simulated as well.
and transmittance, and four matrixes M that quantify the probability SCOPE also simulates a diversity of fluxes, one of which is NPC. Net
of incident light (between 400 and 750 nm) to return as SIF (between photosynthesis is the total gross photosynthesis minus the flux of CO2
640 and 850 nm) from PSI and PSII at the illuminated and shaded side associated with foliage respiration (photorespiration and ‘dark’ respira-
of the leaf. In the calculation of M, values for the emission efficiencies tion), or gross primary productivity less the dark respiration of the fo-
of 0.002 for PSI and 0.01 for PSII (the latter in dark adapted conditions) liage. NPC is calculated here by simply aggregating the photosynthesis
have been assumed. over the leaf area of the canopy, because photosynthesis in SCOPE is
The output of Fluspect is combined with the fluorescence emission the gas exchange flux of CO2 between atmosphere and leaf, which is a
efficiency scale factor ε to obtain the leaf SIF spectra: scalar quantity. NPC from SCOPE may be used to compute GPP for ap-
proximate comparisons with that derived from eddy covariance (EC)

F f ¼ εM f ;PSII þ M f ;PSI E flux measurements over canopies, for example by setting the respira-
ð1Þ tion parameter to zero, or by summing the net photosynthesis and
F g ¼ εMg;PSII þ Mg;PSI E
leaf dark respiration; however, the GPP from both approaches will not
be precisely identical because of differences in specific assumptions
Where E is the incident irradiance vector (400–750 nm) on the leaf and methodological approaches used in EC-derived GPP.
(W m−2 sr− 1 μm−1). The variable ε scales the PSII SIF spectrum as a An example of bidirectional, canopy-leaving SIF and NPC simulations
function of micrometeorological conditions and photosynthetic param- for ranging leaf chlorophyll content (Cab) and LAI is shown in Fig. 1. Cab
eters. The separation between the biochemical model on the one hand and LAI were earlier identified as key driving variables in governing
(for ε) and Fluspect on the other hand (for M) has the advantage that canopy-leaving SIF spectra (Verrelst et al., 2015). That can also be ob-
the effects of scattering and re-absorption due to leaf structure (the ma- served in the plotted SIF spectra. SIF is especially governed by LAI,
trices M) can be calculated in advance, because they are functions of leaf with a higher LAI (i.e., more leaves within a given area) leading to a
structure only and do not depend on micrometeorological conditions. more pronounced SIF signal. The same is true for NPC, where the occur-
In Eq. (1) it is assumed that the micrometeorologically induced var- rence of a higher LAI is linked with a higher NPC, i.e. greater photosyn-
iations of the fluorescence emission (ε) affect only PSII. This assumption thetic capacity at the canopy scale.
Fig. 1. Example of full SIF profiles (left) and net photosynthesis of the canopy (NPC) (right) simulated by SCOPE as a function of leaf area index (LAI; 14 samples) and leaf chlorophyll
content (Cab; 8 samples).
2.2. SIF retrieval bands Table 1

Single absorption lines and features tested in regression analysis.
2.2.1. Single bands for SIF retrievals Index Element Central Spectral range
Current SIF retrieval methods may be considered as single-line wavelength (nm) absorption
in-filling approaches that are applicable where absorption lines are (nm) lines
present. Four commonly used absorption lines were chosen: the tel- 1 Hα absorption line 656 653–662
luric atmospheric absorption lines of O 2 -A (centered at approxi- 2 Red peak (attributed to SIF emission 685
of Photosystem II)
mately 760 nm), O 2 -B (687 nm), and water vapor (719 nm), and
3 O2-B absorption line 687 683–692
the solar Fraunhofer line of Hα (656 nm) (Zhao et al., 2014). Other 4 Mid-valley between red and NIR peaks 699
absorption lines such as solar Fraunhofer lines around 755–759 nm 5 Water vapor absorption line 719 714–722
were not selected because they were considered too narrow and of 6 Near-infrared peak (attributed to SIF 740
insufficient depth for capturing the subtle SIF signal with sufficient emission of Photosystem I and to PSII)
7 O2-A absorption line 760 757–771
accuracy. Additionally, the peaks of the red emission (centered at
685 nm) and the near-infrared emission (centered at 740 nm), and
the mid-valley (centered at 699 nm) between the peaks were also
selected (Fig. 2, Table 1). In Verrelst et al. (2015), the method of Saltelli et al. (2010) was used to
identify the driving variables that shape the variability of canopy-
2.2.2. Combined bands leaving SIF spectrum across its full spectral range. The method has
Assuming the successful retrieval of single bidirectional SIF been demonstrated to be effective in identifying both the main sensitiv-
bands, the following combinations were analyzed for their predictive ity effects (first-order effects, i.e., the contribution to the variance of the
value: (1) O2-B and O2-A absorption lines, (2) the four main absorp- model output by each input variables, Si) and total sensitivity effects
tion lines (Hα, O2 -B, water vapor, O2-A), (3) the two peaks (red, (the first-order effects plus interactions with other input variables, STi)
NIR), (4) the SIF peak-ratio (red:NIR), which is an indicator of chlo- of input variables. In this study, the GSA analysis was extended to in-
rophyll content and plant status (Hak, Lichtenthaler, & Rinderle, clude the integrated SIF (Ftotal) from 641 to 850 nm. As such, the relative
1990; Pedrós, Goulas, Jacquemoud, Louis, & Moya, 2010), and contribution of each input variable to SIF can be disentangled and quan-
(5) the two peaks and mid-valley. Additionally, (6) given the possi- tified. The MD12 biochemical sub-model was used here, based on earli-
bility to reconstruct the full SIF profile, it is possible to calculate Ftotal er performance comparisons to the TB12 modules, and keeping KNPQs
(taken here as the integrated hemispherical SIF from 641 to 850 nm). and qLs constant (Mohammed et al., 2014). Further, the full variable
Finally, (7) the reconstruction of the full SIF profile also enables in- space of SCOPE was analyzed for a spherical leaf angle distribution
clusion of all single bands (from 650 to 790 nm was used here) in and without varying soil variables. See Table 3 and Verrelst et al.
the regression analysis. An overview is provided in Table 2. (2015) for details on the model variable boundaries. In order to catch
the full model variability, it must be noted that some variable bound-
2.3. Canopy configurations aries likely exceed normal real-world situations (for instance, the varia-
tion of CO2 concentration in the air is considerably higher than would be
SIF–NPC relationships were assessed for various canopy configura- expected under normal circumstances). Variables were sampled
tions based on the variables that are known to be drivers of SIF emission according to Sobol's quasi-random sequence generator. In total,
characteristics. In order to identify driving variables a global sensitivity (N(k + 2)) model simulations were run, where N is the sample size
(GSA) analysis on SCOPE SIF simulations was conducted in a related and equals 2000, and k is the number of input variables and equals 25.
study (Verrelst et al., 2015). Variance-based GSA explores the full This produced 54,000 simulations. Only total order sensitivity effects
input variable space and evaluates the relative importance of each (STi) expressed as percentages were considered.
input variable in a model (Saltelli, Tarantola, & Chan, 1999). The method Fig. 3 provides the STi results of the SCOPE v.1.53 input variables.
can be used to identify the most influential variables affecting model Most variables exerted a negligible effect on Ftotal. The driving variables
outputs. In variance-based GSA the contribution of each input variable were: maximum carboxylation capacity at optimum temperature
to the variation in outputs is averaged over the variation of all input var- (Vcmo) (sometimes referred to as Vcmax,25 in the literature, where 25
iables, i.e., all input variables are changed together (Saltelli et al., 1999). stands for an optimum temperature of 25 °C), dry matter content
Fig. 2. A typical incident solar irradiance spectrum at top of canopy simulated by MODTRAN-5 for a standard mid-latitude summer model atmosphere, and the default rural aerosol model.
Both the spectral resolution and spectral sampling interval are 1 nm. Four main absorption lines used for SIF retrieval are shown in the figure. Their spectral wavelengths (spectral positions
of the maximum absorption) and spectral ranges are marked with vertical lines and gray shade, respectively, which are further specified in Table 1. SIF radiance leaving the canopy and
simulated by SCOPE has been added.
Table 2 varying all variables at all scales. The driving variables taken as a
Combined SIF absorption lines and features used in regression analysis. starting point were Vcmo at the biochemical scale, Cab at leaf scale,
Index Combined wavelengths Wavelengths (nm) and LAI at canopy scale. The number of ranging variables was then
1 O2-B and O2-A absorption lines 687, 760
increased to eventually produce a total of 12 canopy configurations,
2 Hα, O2-B, water vapor 656, 687, 719, 760 and then lastly all SCOPE variables were ranged (Table 4). Each var-
absorption lines, and O2-A iable was randomly sampled 2000 times within their minimum–
3 Two SIF emission peaks 685, 740 maximum boundaries according to Table 3 (see also Verrelst et al.,
4 Peak ratio 685/740
2015). From the combined variable space, a uniform random set of
5 Two SIF emission peaks and 685, 699, 740
mid-valley 2000 simulations was selected. In the unlikely event that the radia-
6 Ftotal Hemispherically and spectrally tive transfer equations were unsuccessfully resolved, e.g., due to un-
integrated SIF at the TOC (from 641 realistic variable combinations, another random simulation was
to 850 nm) taken. Finally, from all SCOPE output variables the NPC output flux,
7 Fall All individual bidirectional SIF
wavelengths (from 650 to 790 nm)
the bidirectional SIF spectra and Ftotal were collected.
2.4. Regression analysis
(Cdm), Cab, LAI, canopy height (hc), within-canopy-layer resistance To enable estimation of NPC from SIF retrievals, a regression anal-
(rwc), air pressure (P), atmospheric vapor pressure (ea), atmospheric ysis was used. For each canopy configuration a random subset of
CO2 concentration (Ca), air temperature (Ta), and broadband incoming 2000 simulations was selected and then split into 50% for regression
shortwave radiation (Rin). Altogether these variables explained 97.5% model calibration and 50% for validation. An ordinary least squares
of the total variance (taking interactions into account). linear regression (LR) and an adaptive, nonlinear regression algo-
Because SCOPE is a SVAT model, analysis of input variables to the rithm regression called Gaussian processes regression (GPR) were
SIF–NPC relationships can be undertaken at various scales and con- applied. LR was used to analyze all bands individually and in combi-
texts, e.g., biochemical (physiological), leaf, canopy, geometrical, nations. GPR was applied to determine whether and how much im-
and/or micrometeorological. A balance must be struck between in- provement could be achieved for the combined band analyses.
cluding a sufficient number of ranging variables to achieve good rep- GPR is a machine learning regression technique in a Bayesian frame-
resentation of reality but not so many variables that the escalating work equivalent to kernel ridge regression, least squares support vector
heterogeneity becomes uninterpretable. Various canopy configura- machine and kriging (Rasmussen & Williams, 2006). In a previous study
tions with increasing heterogeneity were generated. First, only the (Verrelst, Muñoz, et al., 2012), GPR outperformed neural networks, sup-
main driving biochemical variable was ranged, then more variables port vector regression and kernel ridge regression for the majority of
were ranged at biochemical, leaf, and canopy scales, eventually biophysical parameter retrievals. The GPR model establishes a relation
Table 3
Input variables and their boundaries and default values of the SCOPE model. Soil and aerodynamic variables and kNPQs, qLs and leaf angle distribution have been kept to their default
values. See also Verrelst et al., 2015.
Input Definition Unit Min Max Default
Leaf biochemistry
Vcmo Maximum carboxylation capacity (at optimum temperature) μmol m−1 s−1 0 200 30
m Ball-Berry stomatal conductance parameter [−] 2 20 8
Rdparam Parameter for dark respiration (Rd = Rdparam ∗ Vcmo) [−] 0.001 0.03 0.015
Extinction coefficient for a vertical profile of Vcmo (maximum value of Vcmo
kV [−] 0 0.8 0.64
occurs at the top of the canopy).
Leaf optical
N Mesophyll structural parameter in PROSPECT [−] 1 2.5 1.4
Cw Water content in PROSPECT g cm−2 0 0.1 0.009
Cdm Dry matter content in PROSPECT g cm−2 0 0.05 0.012
Cs Senescence factor in PROSPECT [−] 0 0.9 0
Cab Chlorophyll content in PROSPECT μg cm−2 0 80 40
Canopy
lw Leaf width m 0.01 0.1 0.1
LIDFa LIDF parameter a, which controls the average leaf slope [−] −1 1 −0.35
LIDFb LIDF parameter b, which controls the distribution's bimodality [−] −1 1 −0.15
LAI Leaf area index m2 m−2 0 7 3
hc Canopy height m 0.1 2 1
Micrometeorological
p Air pressure hPa 300 1090 970
u Wind speed m s−1 0 50 2
Oa O2 concentration in the air ppm 0 220 209
ea Atmospheric vapor pressure hPa 0 150 15
Ca CO2 concentration in the air ppm 50 1000 380
Ta Air temperature °C −10 50 20
Rin Incoming shortwave radiation W m−2 0 1400 600
Rli Incoming longwave radiation W m−2 0 400 300
Geometry
VZA Viewing zenith angle Degree 0 10 0
RAA Relative azimuth angle Degree 0 180 0
SZA Sun zenith angle Degree 0 60 30
Fig. 3. Driving variables of Ftotal as identified by Global Sensitivity Analysis for SCOPE biochemistry (without KNPQs, qLs), leaf, canopy (without leaf angle distribution) and
micrometeorology variables.
between the input (B-bands spectra) x∈RB and the output variable (leaf be automatically optimized by maximizing the marginal likelihood in
parameter) y∈R of the form: the training set (Rasmussen & Williams, 2006).
The accuracies of LR and GPR models were validated using the vali-
N
dation dataset with coefficient of determination (R2) and the root-
ŷ ¼ f ðxÞ ¼ ∑ α i Kðxi ; xÞ; ð2Þ
i¼1 mean-square error (RMSE) calculated between retrieved and “true”
(simulated) NPC values.
where {xi}N
i =1 are the spectra used in the training phase, αi is the weight
assigned to each one of them, and K is a sophisticated kernel function 3. Results
evaluating the similarity between the test spectrum and all N training
spectra (Verrelst, Alonso, Camps-Valls, Delegido, & Moreno, 2012; 3.1. Single band analysis
Verrelst, Muñoz, et al., 2012). A scaled Gaussian kernel function was
used, The sensitivity to NPC of each wavelength within the SIF emission
0 1 spectrum was first assessed. LR was used and the R2 of validation data
ðbÞ ðbÞ 2 plotted in Fig. 4. Results are organized according to ranging variables
B B xi x j C
K xi ; x j ¼ υ exp@ ∑ A; ð3Þ at the scales of biochemistry, leaf, canopy, and at all SCOPE scales. The
b¼1 2σ 2b
following trends were observed:
When ranging only variables at biochemical scale (Fig. 4a), Vcmo was
where υ is a scaling factor, B is the number of bands, and σb is a dedicat- the main variable driving NPC and produced a very strong relationship
ed parameter controlling the spread of the relations for each particular (R2 of 0.99). Relationships with NPC weakened when the other bio-
spectral band b. Model parameters (υ, σb) and model weights αi can chemical variables (m, Rparam, kV) were also varied, but impacts on
Table 4
SCOPE canopy configurations with ranging variables. (See Table 3 for definitions of the variables.)
Index Ranging variables Justification
Vcmo is the main biochemical driver of photosynthesis. Hence, this is the theoretical baseline when SIF is not influenced
1 Vcmo
by any other variable.
2 Biochemistry All biochemical variables (Vcmo, m, Rdparam, kV). Represents the most heterogeneous situation at the biochemical scale.
3 Vcmo, Cab Driving biochemical and leaf variables.
4 Vcmo, leaf Driving biochemical variable and all leaf variables (N, Cw, Cdm, Cs, Cab).
All biochemical and leaf variables. Represents the most heterogeneous situation at biochemical and leaf scales (Vcmo, m,
5 Biochemistry, leaf
Rdparam, kV, N, Cw, Cdm, Cs, Cab).
6 Cab, LAI Driving leaf and canopy variables.
7 Vcmo, LAI Driving biochemical variable (Vcmo) with driving canopy variable (LAI)
8 Vcmo, canopy Driving biochemical variable (Vcmo) with all varying canopy variables (LAI, lw, hc).
Vcmo, N, Cw, Cdm, Cs, Cab, LAI, hw, hc
9 Driving biochemical variable (Vcmo) with all leaf and all canopy (N, Cw, Cdm, Cs, Cab, LAI, lw, hc).
(spherical LIDF)
Al biochemical, leaf and canopy variables (Vcmo, m, Rdparam, kV, N, Cw, Cdm, Cs, Cab, LAI, lw, hc). Represents the most
10 Biochemistry, leaf, canopy
heterogeneous situation at the canopy scale
11 Key SCOPE variables driving SIF Vcmo, Cdm, Cab, LAI, hc, rwc, P, ea., Ca, Ta, Rin. These variables and their interactions explain 97.5% of the variability in Ftotal.
All SCOPE variables (Vcmo, m, Rdparam, kV, N, Cw, Cdm, Cs, Cab, LAI, lw, hc, rwc, rb, P, u, Oa, ea., Ca, Ta, Rin, Rli, VZA, RAA,
12 All SCOPE variables
SZA). Represents the most heterogeneous configuration.
Fig. 4. Strength of relationships between SIF and NPC for single wavelengths in the spectral emission profile. Regression analysis results (R2) of single wavelengths across the 650–790 nm
spectral range were analyzed for different canopy configurations as outlined in Table 3.
R2 were spectrally invariant. Although it should be noted that even fully heterogeneous canopies, and also ranging other biochemical vari-
though R2 was spectrally invariant, the slope of the relationship varied ables then correlations degraded to such an extent that meaningful re-
with wavelength (not shown) as SIFred is affected more strongly by bio- lationships could not be derived.
chemical variables than is SIFNIR. Combining input variables at all SCOPE scales (Fig. 4d), i.e., including
When SCOPE variables were ranged at the leaf scale (Fig. 4b), varying also micrometeorological variables, caused relationships to deteriorate,
the driving variables Cab and Vcmo produced strong relationships with especially for the second peak. When considering the driving SIF var-
NPC. However, relationships degraded at wavelengths beyond the red iables as identified by the GSA exercise (see Fig. 3), the first peak
emission peak. The relationship degraded further when additionally achieved a maximal R2 of about 0.5, whereas for the second peak it
varying other leaf variables, and a clear distinction between the first did not exceed about 0.3. Given that this dataset was mainly generat-
and second peak was observed. After 780 nm relationship degraded ed from varying the key micrometeorological variables (P, ea, Ca, Ta,
until zero approaching the end of PSII spectrum. It should be empha- Rin), results suggest that these factors can weaken SIF–NPC relation-
sized that the effect of the biochemical parameters (Vcmo, m, Rparam, ships. Varying all SCOPE variables no longer produced meaningful
kV) on SIF is through parameter ε in Eq. (1) (Van der Tol et al., 2014). relationships. This finding underlines the utility of applying a GSA
Due to the fact that ε applies only to PSII, which has its peak in the red study in order to constrain the number of input variables in an opti-
fluorescence, the effect of biochemical parameters is the strongest in mized way.
the red fluorescence. When also ranging all biochemical variables, cor- Table 5 summarizes the predictive strength of the most important
relations weakened further and the difference between the two SIF SIF spectral bands (as identified in Table 1) for estimating NPC. Addi-
emission peaks was less prominent. tionally the wavelength that produced the strongest correlation is
When SCOPE vegetation variables were subsequently ranged at the shown. The following observations can be made. Overall, the red peak,
canopy scale (Fig. 4c), strong spectrally invariant relationships were ob- O2-B, and Hα line showed similar predictive strength. The NIR peak
tained only in the case of Vcmo plus LAI and other canopy variables. The and O2-A were also similar in performance. In most instances the red
relationship broke down for the NIR emission peak when leaf variables peak or O2-B band were better predictors than the NIR peak or O2-A.
were ranged. Also strong relationships could be obtained when varying The water vapor band was not as strong a predictor as the three red
only the driving leaf and canopy variables Cab and LAI, but these rela- bands or the mid-valley, though it was superior to the NIR peak and
tionships also broke down for the second peak region. When simulating O2-A. The best performing wavelength differed by scenario and spanned
Table 5
Sensitivity of single bands and spectral features under different canopy configurations using linear regression. Best performing R2 per canopy configuration is shown in bold-face text. The
R2 cells that fall within 0.01 interval of the best performing R2 are highlighted in gray. R2 of the best performing wavelength is also indicated (far right column).
NIR Best
Ranging SCOPE Hα Redpeak O2–B Mid–valley Water vapor O2–A
peak wavelength R2
variables (656 nm) (685 nm) (687 nm) (699 nm) (719 nm) (760 nm)
(740 nm) (nm)
R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE
1 Vcmo 0.9970 0.7484 0.9966 0.7948 0.9966 0.7890 0.9971 0.7361 0.9975 0.6760 0.9977 0.6491 0.9978 0.6426 790 0.9978
2 biochemistry 0.7070 8.5148 0.7065 8.5221 0.7066 8.5207 0.7072 8.5118 0.7079 8.5031 0.7082 8.4986 0.7082 8.4975 790 0.7087
3 Vcmo, Cab 0.9830 1.7376 0.9801 1.8811 0.9819 1.7934 0.9911 1.2605 0.9652 2.4903 0.9110 3.9809 0.8883 4.4594 703 0.9922
4 Vcmo, leaf 0.9026 4.3075 0.9092 4.1596 0.9040 4.2746 0.8371 5.5674 0.6887 7.6947 0.6415 8.2583 0.6175 8.5301 676 0.9159
biochemistry,
5 leaf 0.6275 9.9114 0.6309 9.8653 0.6288 9.8938 0.5980 10.2940 0.5178 11.2759 0.4863 11.6382 0.4720 11.8004 677 0.6337
6 Cab, LAI 0.9208 1.4306 0.9197 1.4411 0.9257 1.3864 0.9438 1.2047 0.7789 2.3754 0.6297 3.0730 0.5772 3.2838 696 0.9459
7 Vcmo, LAI 0.9744 2.3695 0.9760 2.2963 0.9766 2.2664 0.9829 1.9356 0.9880 1.6217 0.9869 1.6966 0.9875 1.6581 777 0.9895
8 Vcmo, canopy 0.9199 3.4569 0.9211 3.4316 0.9215 3.4221 0.9227 3.4001 0.9132 3.6086 0.9179 3.5075 0.9166 3.5353 696 0.9232
Vcmo, leaf,
9 canopy 0.8879 4.2744 0.8947 4.1411 0.8925 4.1835 0.8540 4.8753 0.7377 6.5325 0.6985 7.0273 0.6768 7.2514 678 0.8974
Biochemistry,
10 leaf, canopy 0.2453 29.3064 0.2429 29.3539 0.2411 29.3867 0.2168 29.8472 0.1773 30.5901 0.1782 30.5714 0.1727 30.6767 650 0.2462
Key variables
11 driving SIF 0.5153 15.2639 0.5030 15.4557 0.4973 15.5454 0.4109 16.8229 0.2785 18.6249 0.3020 18.3188 0.2866 18.5130 650 0.5190
All SCOPE
12 variables 0.2260 39.4620 0.2249 39.4886 0.2241 39.5090 0.2120 39.8133 0.1896 40.3767 0.1902 40.3663 0.1869 40.4500 650 0.2263
Table 6
The sensitivity of combined bands and spectral features under different canopy configurations using linear regression. Best performing R2 per canopy configuration is shown in bold-faced
text. The R2 cells that fall within 0.01 interval of the best performing R2 are highlighted in gray.
Hα, O2–B, water Two peaks and Fall:

Ftotal:
O2–B, O2–A: vapor, O2–A: Two peaks: Peak ratio: valley: All individual SIF
Ranging SCOPE variables Integrated SIF (from
687, 760 nm 656, 687, 719, 685, 740 nm 685/740 685, 699, 740 wavelengths (from
641 to 850 nm)
760 nm nm 650 to 790)
1 Vcmo 0.9965 0.8255 0.9970 0.7412 0.9965 0.7869 0.8686 4.9370 0.9966 0.7831 0.9983 0.5665 1 0.0843
2 biochemistry 0.7234 8.0229 0.7111 8.1849 0.7195 8.1690 0.5874 9.8065 0.7141 8.3908 0.7201 8.2419 0.7011 8.4288
3 Vcmo, Cab 0.9965 0.7792 0.9982 0.5657 0.9963 0.8123 0.4721 9.6652 0.9966 0.7816 0.9900 1.3493 0.9991 0.4056
4 Vcmo, leaf 0.9333 3.5056 0.9462 3.0807 0.9305 3.6105 0.3062 11.1958 0.9573 2.8784 0.6467 8.0626 0.9783 2.0221
5 biochemistry, leaf 0.6356 9.7534 0.6205 10.2080 0.6337 9.8866 0.4056 12.3210 0.6423 9.3908 0.5176 10.9974 0.6828 8.9835
6 Cab, LAI 0.9529 1.1083 0.9354 1.2673 0.9547 1.0751 0.1955 4.5583 0.9434 1.2164 0.9426 2.0248 0.9728 0.8098
7 Vcmo, LAI 0.9907 1.4723 0.9853 1.7952 0.9898 1.5035 0.7240 7.7957 0.9770 2.2615 0.9923 1.3046 0.9968 0.8415
8 Vcmo, canopy 0.9315 3.2390 0.9333 3.2096 0.9321 3.2583 0.6404 7.3998 0.9244 3.4306 0.9105 3.7510 0.9490 2.8608
9 Vcmo, leaf, canopy 0.8950 4.1334 0.9023 4.0438 0.9077 3.9263 0.3495 10.2994 0.8982 4.1313 0.7157 6.8358 0.9169 3.7433
10 Biochemistry, leaf, canopy 0.2356 31.8478 0.2388 30.8678 0.2348 30.8049 0.1152 33.2900 0.2144 30.2922 0.1292 32.3184 0.2805 30.4726
1 Key variables driving SIF 0.4581 13.7917 0.5068 13.6618 0.5342 12.5735 0.2484 15.8487 0.5407 12.6540 0.3078 16.1429 0.5693 12.4038
12 All SCOPE variables 0.2234 41.6271 0.2112 45.0060 0.2446 40.2953 0.1260 41.4454 0.2091 41.6138 0.2278 35.7368 0.2275 40.8113
the entire SIF emission spectrum. In realistic canopy scenarios (i.e., with 3.3. Nonlinear Gaussian processes regression
ranging variables at scales of biochemistry, leaf and canopy; scenario
11) the best performing wavelength was situated on the slope before The nonlinear GPR (Table 7) produced stronger relationships with
the first peak. NPC than did linear regression for the majority of cases, although im-
provements were generally modest. Considering the best two-band
combinations from Table 6 (i.e., the two peaks or the O2-A and O2-B),
3.2. Combined bands analysis R2 values were higher in scenarios 2, 4–6, and 8–12. The strongest im-
provements were under conditions of increasing canopy and environ-
Table 6 shows the predictive power of combined SIF wavelengths. mental heterogeneity. Again, including all individual wavelengths in
Combining the O2-A and O2-B bands or the red and NIR peaks produced the regression led to strongest relationships for the majority of scenari-
stronger relationships with NPC than those that were obtained when os, although improvements as compared to using the SIF absorption
the O2-A band or the NIR peak was used individually (scenarios 3–6 bands were generally modest. From a pragmatic perspective, using an
and 8–12), but in contrast combinations produced small improvements adaptive, nonlinear regression method and retrieving SIF in the two
over using only the O2-B band or the red peak. Hence, the O2-A and the deepest absorption lines could be sufficient to derive NPC with sufficient
NIR peak benefited the most from band combinations. The combination accuracy.
of O2-B and O2-A bands produced similar results as when combining the
two peaks. Hence, these two combinations could be considered essen- 4. Discussion
tially equivalent from this analysis. In all cases, the peak ratio (F685/
F740) produced considerably poorer correlations than using the two 4.1. Potentials and limitations of the applied approach
bands individually. Combining the mid-valley with the two peaks pro-
duced only marginal improvements over the combined peaks. Also Progress in imaging spectroscopy technology and data processing
small improvements were obtained when combining SIF retrievals at will soon make it possible to derive and exploit the full SIF spectrum
the four absorption lines (Hα, O2-B, water vapor, O2-A). The Ftotal (inte- emitted from vegetation canopies. However, to date no imaging spec-
grated SIF) generally did not yield a predictive advantage and in several troscopy broadband SIF spectra are available for the canopy scale.
instances produced weaker correlations than other features. Converse- Hence, simulation studies are required to predict the information con-
ly, further improvements were achieved for most of the scenarios tent of this unique source of information. This study used SCOPE simu-
when including all individual wavelengths in the regression analysis, lations to conduct a theoretical examination of SIF band sensitivity to
but gains in explaining NPC variance were modest. Overall, selection net photosynthesis of the canopy (NPC). This type of modeling study
of single SIF bands such as the red peak or the O2-B band, or a combina- can be useful in helping to disentangle the complex relationships be-
tion of the two peaks or the O2-A and O2-B bands appeared sufficient to tween canopy-leaving SIF (as estimated by an imaging spectrometer)
estimate NPC. and vegetation photosynthetic activity. Importantly, it helps to establish
Table 7
The sensitivity of combined bands and spectral features under different canopy configurations using GPR. Best performing R2 per canopy configuration is shown in bold-faced text. The R2
cells that fall within 0.01 interval of the best performing R2 are highlighted in gray.
Hα, O2–B, water Two peaks and Fall:

Ftotal:
O2–B, O2–A: vapor, O2–A: Two peaks: Peak ratio: valley: All individual SIF
Ranging SCOPE variables Integrated SIF (from
687, 760 nm 656, 687, 719, 685, 740 nm 685/740 685, 699, 740 wavelengths
641 to 850 nm)
760 nm nm (from 650 to 790)
1 Vcmo 1 0.0040 1 0.0053 1 0.0038 0.9900 1.4283 1 0.0039 1 0.0089 1 0.0028
2 biochemistry 0.7863 7.0925 0.7901 6.9813 0.7332 7.8987 0.7181 8.2289 0.7389 8.1721 0.7268 35.8424 0.7996 6.7456
3 Vcmo, Cab 0.9996 0.2598 0.9998 0.1948 0.9996 0.2548 0.6668 7.6360 0.9987 0.4841 0.9905 1.3127 1 0.0811
4 Vcmo, leaf 0.9695 2.3618 0.9883 1.4783 0.9739 2.1710 0.6382 8.0869 0.9830 1.7833 0.6934 7.5117 0.9899 1.3645
5 biochemistry, leaf 0.6797 9.2628 0.6640 9.2910 0.6833 9.0908 0.5779 10.5707 0.6777 9.1064 0.5459 10.6994 0.7633 7.9247
6 Cab, LAI 1 0.0101 0.9919 0.4640 1 0.0066 0.4044 3.8472 0.9879 0.5552 0.9589 1.7122 1 0.0040
7 Vcmo, LAI 0.9967 0.8520 0.9988 0.5203 0.9993 0.3837 0.8926 4.8546 0.9791 2.1539 0.9945 1.1048 1 0.0070
8 Vcmo, canopy 0.9456 2.9284 0.9431 2.9300 0.9447 2.9073 0.8274 5.2367 0.9397 3.0608 0.9416 3.0289 0.9945 0.9428
9 Vcmo, leaf, canopy 0.9164 3.7109 0.9174 3.7290 0.9177 3.7213 0.5272 8.9396 0.8982 4.1313 0.7250 6.7224 0.9401 3.1440
10 Biochemistry, leaf, canopy 0.3180 27.7851 0.3573 27.6869 0.3392 30.4935 0.1938 33.3582 0.3280 27.6240 0.1935 31.0465 0.3819 28.0068
11 Key variables driving SIF 0.5881 12.1416 0.6610 10.9384 0.6000 11.2757 0.3184 15.4214 0.5374 13.0736 0.4053 14.7770 0.6411 11.7772
12 All SCOPE variables 0.3131 37.2426 0.2676 41.4844 0.3204 35.4161 0.1350 43.8470 0.2870 38.1309 0.2292 39.3470 0.2873 42.9957
a theoretical foundation upon which to build hypotheses and formulate computations also have their limitations: for example, flux partitioning
future field studies, including identification of the sorts of ancillary in- methods conventionally estimate GPP from net ecosystem exchange of
formation required to interpret and apply SIF findings. CO2 by using night-time CO2 flux measurements to estimate day-time
Although the strengths of a theoretical study are well-known, respiration (e.g. Reichstein et al., 2005). There are significant uncer-
i.e., full control of all variables, virtually unlimited capability to ana- tainties with that approach and with other assumptions used in EC-
lyze canopy configurations and their interactions with incoming ra- derived GPP as discussed by various authors (e.g. Hilker et al., 2014;
diation (e.g., Verrelst et al., 2015), relying on simulated data also Wohlfahrt & Gu, 2015), hence, those computations should be also con-
has its drawbacks. The model representativeness of actual field sidered as estimates of GPP. In addition, there is a certain inconsistency
data may be limited by (1) random errors and bias in actual retrieved over the usage of the term ‘GPP’ across disciplines (Wohlfahrt & Gu,
SIF that are not included in the simulations, (2) model representa- 2015), which contributes to confusion when trying to compare outputs.
tion errors and (3) representation errors in the selected input data. Nonetheless, SCOPE NPC simulations (although directly called GPP in
These errors are briefly discussed below. some publications) have demonstrated utility in inversion schemes
First, we have assumed a perfectly known TOC signal, but it is recog- (Zhang, Guanter, et al., 2014) or comparative studies against flux
nized that signals acquired by airborne or space sensors must be atmo- tower GPP data (Damm, Guanter, Paul-Limoges, et al., 2015).
spherically corrected. Currently, such correction can be achieved using Finally, representation in input variables affects results. In this study
sophisticated techniques such as inversion of the atmospheric radiative full ranges from minimum to maximum values of variables were used. It
transfer codes from MODTRAN (Berk et al., 1999). However, atmospher- may be expected that in actual field situations not all of the variables
ic correction algorithms generally ignore the contribution of the atmo- might vary independently over the full ranges. Also, some variables
spheric spherical albedo and assume a Lambertian surface behavior may be readily quantifiable from actual measured data, such as short-
(Guanter, Gonzalez-Sanpedro, & Moreno, 2007), a known temperature wave incoming radiation (Rin) and LAI (Baret et al., 2013; Zhang,
vertical profile and a ‘guesstimate’ of the aerosols nature (Chavez, Liang, Zhou, Wu, & Zhao, 2014), and thus would not need to be consid-
1996). Even with sophisticated algorithms, propagated errors could po- ered as unknown. Therefore, the actual set of ranging variables may dif-
tentially degrade the quality of SIF retrievals. Errors can arise due to e.g. fer in each field situation.
instrumental noise, atmospheric interferences, and surface anisotropy
effects (e.g., Damm, Guanter, Verhoef, et al., 2015). Thus, for a full sensi- 4.2. Implications regarding exploitation of retrieved SIF data
tivity analysis of SIF retrievals from space, in principle all these aspects
should be considered. Also, spectral shifts in band central wavelengths This work underlines the complexity of SIF–NPC relationships in
or a non-accurate characterization of the instrument spectral response heterogeneous vegetation environments, simulated here through
function (ISRF) can become crucial when working at very high spectral the ranging of SCOPE variables to represent varying levels of biolog-
resolutions (b 1 nm). Hence, for results to be valid for a specific sensor, ical and environmental complexity. Results led to the following three
sensor-specific signal-to-noise ratios plus sensor ISRF should be consid- key observations.
ered to determine if the technical capabilities are sufficient to success- A first key observation is that with increasing heterogeneity, i.e.
fully acquire SIF measurements. For instance, Zhao et al. (2014) more ranging variables, poorer relationships were achieved, until a
showed that poor signal-to-noise ratio will degrade SIF retrievals. Con- point is reached where no meaningful relationships with NPC could be
sequently, we could expect concomitant degradation of relationships derived. It indicates the difficulty of interpreting canopy-leaving SIF
with photosynthesis. Another point is that not all SIF absorption bands emitted by heterogenous vegetation. However extreme situations
equally enable accurate retrieval of SIF. SIF is typically more easily re- with all SCOPE variables fully ranging seem unlikely in reality. In con-
trieved in the O2-A region than in the other absorption lines due to a trast, when varying only SCOPE's key variables or vegetation variables,
deeper and wider absorption at 760 nm (see also Fig. 2) and the TOC re- both linear and advanced nonlinear regression methods were able to
flectance is a smooth profile in the NIR shoulder (Cogliati et al., 2015). In produce robust relationships between SIF and NPC. But varying only
order to quantify and mitigate these limitations when moving towards some variables also seems unlikely in reality. In this respect, combining
spaceborne SIF retrievals, an end-to-end mission performance simula- modeling studies with experimental studies may be better suited for de-
tor (FLEX-E) has been developed within the framework of FLEX where fining the type of relationships. For instance, Damm, Guanter, Paul-
most of these limitations are explicitly taken into account. FLEX-E com- Limoges, et al. (2015) reported an asymptotic relationship using instan-
bines the forward simulation of complex synthetic scenes using coupled taneous experimental SIF760 and eddy covariance GPP flux tower data.
SCOPE and MODTRAN-5 (Berk et al., 2006) with modeling of the satel- In this respect, our modeling findings confirmed that nonlinear regres-
lite and instrument behavior and the full processing scheme up to the sion (i.e., GPR) is better able than linear regression to deal with hetero-
retrieval of the final SIF products (details in Vicent et al., submitted for geneous situations. Moreover, GPR possesses several more interesting
publication). properties not exploited in this study. It provides (1) an indication of
Second, SCOPE, like any model, has representation errors. For in- band relevancy for each variable; (2) a weight for the most relevant
stance, the radiative transfer in SCOPE is described with a 1D vertical spectra contained in the training data set; and (3) probabilistic outputs,
model that assumes horizontal and vertical uniformity. The model is i.e. a mean estimate and an associated uncertainty interval (Verrelst,
valid for vegetation in which scattering of radiation by adjacent objects Alonso, et al., 2012; Verrelst, Muñoz, et al., 2012). Particularly the latter
of different leaf composition or canopy structure plays no significant may be of interest in future analyses, as it is an elegant mathematical in-
role. Tree crowns surrounded by open space are not represented by dicator of the reliability of the NPC prediction on a per-pixel basis. Typ-
SCOPE, because in that case illumination is mostly from the side of the ically, an increase in pixel heterogeneity provokes an increase in
crowns, and tree crowns cast shadows on other crowns. Also vertically prediction uncertainty. Similarly, the associated uncertainties enable
varying leaf properties within the crown are not considered in SCOPE. quantification of the portability of regression model in space and time
Other representation errors are possible in the models for photosynthe- (Verrelst, Rivera, Moreno, & Camps-Valls, 2013).
sis, stomatal regulation and turbulent atmospheric transport that con- A second key observation is that SIF–NPC relationships likely are
tain semi-empirical relationships with coefficients that cannot be influenced, not only by variation in photosynthetic activity (Vcmo),
determined from physics, but require calibration to measurements but also of other key variables such as Cab and LAI. This is not surpris-
(Van der Tol et al., 2014). ing since Vcmo is a rather modest key driver of SIF variability in natu-
Third, the modeled NPC only approximates GPP as derived from ral field situations. For instance, LAI has a stronger influence. The
eddy covariance (EC) flux measurements over canopies; in particular, relative influence of the various drivers can also be derived through
SCOPE scales up net rather than gross photosynthesis. But EC-flux GPP a global sensitivity analysis, as shown in Fig. 3 or in Verrelst et al.,
2015. Consequently, this suggests that empirically obtained GPP of the full broadband SIF signal, all individual bands could be directly
maps, as has been assembled at the global scale (e.g. Frankenberg inserted into a full-spectrum (nonlinear) regression analysis, as demon-
et al., 2011; Guanter et al., 2012; Joiner et al., 2011), might actually strated successfully here.
be more of a representation of spatial variations of chlorophyll con-
tent and canopy structural variables than of photosynthetic activity. 4.3. Towards unbiased global estimation of photosynthesis
Such aspects will require consideration in future SIF retrievals and
photosynthesis mapping in order to account for possible confound- Given the biochemical, leaf, structural and micrometeorological influ-
ing factors. ences on the canopy-leaving SIF signal, it will be necessary to disentangle
A third key observation is that a clear difference in performance of the information content related to photosynthetic activity from that due
SIFred and SIFNIR can be observed for quantification of canopy photosyn- to vegetation structure in order to achieve unbiased estimations of photo-
thesis. SIFNIR was a significantly worse predictor than SIFred, suggesting synthetic carbon uptake. Two complementary strategies to address these
that exploitation of the second emission peak is not an optimal choice issues may be envisaged: (1) exploitation of the full SIF emission,
for the estimation of photosynthesis. Although current practices rely on i.e., including both SIFred and SIFNIR emission features; and (2) use of joint-
SIFNIR to quantify GPP, experimental data over a maize field (Cheng ly derived biophysical variables in order to account for structural effects.
et al., 2013) confirmed that SIFred led to stronger relationships than SIFNIR. Firstly, SIFred and SIFNIR have different behavioral features: SIFred is less
Two reasons could help explain the superior linkage of SIFred to photosyn- scattered and more sensitive to PSII photochemistry, whereas SIFNIR is
thesis at the canopy scale. First, the red emission peak has a strong sensi- less re-absorbed. Detection of these two signals in combination with re-
tivity to PSII processes such as NPQ (Porcar-Castell et al., 2014). Second, flectance based estimates of pigments could reveal phenological and
this peak, due to re-absorption in the red, is coming mostly from the physiological changes in canopies. Secondly, through simultaneous re-
upper leaves of the canopy and therefore the canopy-leaving SIFred is trieval of key biophysical variables (e.g. LAI, Cab), such a priori informa-
less susceptible to multiple scattering effects. In turn, although the tion could be used to constrain the regression models or be
canopy-leaving SIFNIR emission is typically more pronounced due to incorporated into assimilation schemes, potentially leading to stronger
lower incidence of re-absorption as compared to SIFred, the SIFNIR is highly SIF-photosynthesis relationships. Similarly, exploiting covariance linkages
subject to scattering, implying that observations of spatial variations of among key retrievable vegetation attributes (e.g. Cab, LAI) and biochem-
the SIFNIR flux may be more subject to misinterpretation due to the influ- istry (Vcmo) in space and time could lead to improved understanding of
ence of leaf and canopy structure (Van Wittenberghe, Alonso, Verrelst, plant physiological responses to environmental drivers and stresses.
Moreno, & Samson, 2015; Van Wittenberghe et al., 2013). In order to de- Atmospheric platforms recently or currently operational in space
velop links with photosynthetic activity, exploiting the first emission peak (e.g. GOSAT, SCIAMACHY, GOME-2, OCO-2) or anticipated in the near
may therefore be more successful than focusing on the currently empha- future (e.g. GOSAT-2, Sentinels 4, 5 & 5P) are not optimized to realize
sized second emission peak. The best strategy, as results here suggest, multiple SIF retrievals along with estimations of biophysical variables.
would be to retrieve SIF from both emission peaks or the O2-B and O2-A Such platforms either are too restricted spectrally to permit derivation
bands, and combining them into a regression model. Further, if SIF can of both SIFred and SIFNIR or their spatial resolution is too coarse to
be retrieved in other absorption lines such as Hα and water vapor, theo- allow characterization of heterogeneity at canopy or stand scales. For
retically that could lead to even stronger and stable correlations. But these example, the TROPOMI sensor, which will be on-board ESA's Sentinel-
regions are more difficult to retrieve with high accuracies. The Hα absorp- 5 Precursor satellite, starts sampling at 675 nm, but there are outstand-
tion line is at the edge of the SIF signal (656 nm) and may be too weak to ing issues to be resolved for reliable discrimination of the red peak, and
retrieve meaningful values. Also SIF retrievals at the water vapor absorp- the spatial footprint at 7 km × 7 km is well beyond the canopy scale
tion line (719 nm) may be perturbed due to spatial variably of columnar (Guanter et al., 2014). Similarly, GOME-2, used in recent retrievals of
atmospheric water vapor. Considering these various aspects, it is recom- red and NIR fluorescence, samples at an even coarser spatial resolution
mended to focus predominantly on SIF retrievals in the O2-B and O2-A re- of 40 km × 40 km or 40 km × 80 km (Joiner et al., 2013; Wolanin et al.,
gions or in the two peaks. 2015). At the present time, the strongest prospect for acquisitions is the
Interestingly, using the absorption lines at O2-B and O2-A, which Fluorescence Explorer (FLEX) mission, now approved as ESA's Earth Ex-
are located on the slopes just beyond the red and NIR peaks respec- plorer 8. FLEX, a small satellite flying in tandem with ESA's Sentinel-3,
tively, was as successful as the actual emission peaks. Similarly, but would be unique in the sense that it would be equipped with an imaging
at the leaf scale, Van Wittenberghe et al. (2014) analyzed the sensi- spectrometer (FLORIS) specifically designed to capture the full broad-
tivity of the SIF spectral bands using GPR by correlating it to chloro- band SIF signal, including both emission peaks. The full spectral range
phyll content. GPR ranking of most relevant bands revealed that the would be 500 to 780 nm at high spectral resolution (up to 0.3 nm) in
most sensitive SIF bands were found on the slopes. By subsequently the regions of the fluorescence peaks and the PRI region. With a contin-
plotting the first derivative of the SIF profiles of the leaf dataset it uous spatial resolution of 300 m, FLEX would provide both high resolu-
was illustrated that largest range of variation is effectively to be tion and bi-weekly global coverage (Kraft et al., 2013). The visible part
found on the slopes. of the spectrum (500–677 nm) can allow for the concurrent estimation
While the use of a combination of SIF retrieval bands improved rela- of chlorophyll absorption and xanthophyll-related NPQ, using informa-
tionships, using the ratio of red to NIR peaks in the regression analysis tion from, for example, the PRI (also noting precautions discussed by
caused a degradation of accuracies in comparison to using the peak Porcar-Castell et al., 2012; Wong & Gamon, 2015a, 2015b) while the
values independently. We suggest the reason might be that the (nonlin- complete FLORIS spectrum allows for the estimation of LAI. These
ear) regression algorithm can exploit spectral data more efficiently sources of information will facilitate interpretations of SIF data. Relevant
when provided with multiple data layers than when that data is trans- meteorological variables would be available from Sentinel-3. Apart from
formed with a simple function to one data layer. The same phenomenon supporting the interpretation of SIF, these meteorological variables
was also reported when comparing the performance of vegetation indi- would also serve for signal correction (Kraft et al., 2013). Such a plat-
ces as opposed to individual reflectance data into GPR to estimate bio- form should improve the capacity to monitor terrestrial vegetation vi-
physical variables (Verrelst, Alonso, et al., 2012). tality and to link SIF with photosynthetic carbon uptake.
We are not aware of imaging spectroscopy studies that exploited the
information content of the full SIF broadband spectrum, but this should 5. Conclusions
soon become possible, e.g. with the HyPlant airborne sensor (Rascher
et al., 2015). Based on findings presented here, retrieving and utilizing A SCOPE modeling simulation study was conducted to examine how
multiple SIF observations is strongly encouraged. Upon reconstruction successfully top-of-canopy sun-induced fluorescence (SIF) can be
related to net photosynthesis of the canopy (NPC). To analyze the influ- Cheng, Y., Middleton, E., Zhang, Q., Huemmrich, K., Campbell, P., Corp, L., ... Daughtry, C.
(2013). Integrating solar induced fluorescence and the photochemical reflectance
ence of biochemistry, leaf and canopy mechanisms impacting the SIF index for estimating gross primary production in a cornfield. Remote Sensing, 5(12),
signal, multiple canopy configurations were simulated. Regression anal- 6857–6879.
yses between simulated SIF retrievals and NPC outputs indicated that Cogliati, S., Verhoef, W., Kraft, S., Sabater, N., Alonso, L., Vicent, J., ... Colombo, R.
(2015). Retrieval of sun-induced fluorescence by advanced spectral fitting
robust relationships can be achieved. However, it should be emphasized methods. Remote Sensing of Environment, 169, 344–357.
that in heterogeneous conditions, the main biochemical variable regu- Collatz, G. J., Ball, J. T., Grivet, C., & Berry, J. A. (1991). Physiological and environmental regu-
lating photosynthetic capacity (Vcmo) is only one of the key variables af- lation of stomatal conductance, photosynthesis and transpiration: a model that includes
a laminar boundary layer. Agricultural and Forest Meteorology, 54, 107–136.
fecting SIF–NPC relationships. Other key vegetation variables that drive
Collatz, G. J., Ribas-Carbo, M., & Berry, J. A. (1992). Coupled photosynthesis-stomatal con-
SIF–NPC relationships include chlorophyll content (Cab) and canopy ductance model for leaves of C4 plants. Australian Journal of Plant Physiology, 19,
structural variables (e.g., LAI). Regarding the predictive strengths of 519–538.
Damm, A., Guanter, L., Paul-Limoges, E., Van der Tol, C., Hueni, A., Buchmann, N., ...
SIF features the following conclusions may be drawn:
Schaepman, M. E. (2015a). Far-red sun-induced chlorophyll fluorescence shows
ecosystem-specific relationships to gross primary production: An assessment based
(1) The most sensitive SIF bands to NPC were located around the first on observational and modeling approaches. Remote Sensing of Environment, 166,
emission peak for heterogeneous canopy configurations. 91–105.
(2) Combining two SIF retrieval bands (e.g., O2-B and O2-A) led to Damm, A., Guanter, L., Verhoef, W., Schläpfer, D., Garbari, S., & Schaepman, M. E. (2015b).
Impact of varying irradiance on vegetation indices and chlorophyll fluorescence de-
stronger correlations than using only one SIF band, especially if
rived from spectroscopy data. Remote Sensing of Environment, 156, 202–215.
that single band was the O2-A. Damm, A., Guanter, L., Laurent, V. C. E., Schaepman, M. E., Schickling, U., & Rascher, U.
(3) Using a combination of O2-B and O2-A bands produced similar (2014). FLD-based retrieval of sun-induced chlorophyll fluorescence from medium
performances as using the two emission peaks. Using the peak spectral resolution airborne spectroscopy data. Remote Sensing of Environment, 147,
256–266.
ratio produced poorer relationships than when both bands Daumard, F., Goulas, Y., Champagne, S., Fournier, A., Ounis, A., Olioso, A., & Moya, I. (2012).
were individually entered into the regression model. Continuous monitoring of canopy level sun-induced chlorophyll fluorescence during
(4) Even stronger correlations were achieved using four main SIF re- the growth of a sorghum field. IEEE Transactions on Geoscience and Remote Sensing, 50,
4292–4300.
trieval bands (Hα, O2-B, water vapor, O2-A). Dayyoub, A. (2011). Novel Techniques for the Remote Sensing of Photosynthetic Processes.
(5) Superior correlations were obtained when all the individual (PhD Thesis) Italy: University of Bologna.
bands of the full SIF emission spectrum (650–790 nm) were in- Duffour, C., Olioso, A., Demarty, J., Van der Tol, C., & Lagouarde, J. P. (2015). An evaluation
of SCOPE: a tool to simulate the directional anisotropy of satellite-measured surface
cluded in the analysis. temperatures. Remote Sensing of Environment, 158, 362–375.
(6) Using an adaptive, nonlinear regression algorithm (GPR) further ESA (2015). http://www.esa.int/Our_Activities/Observing_the_Earth/New_satellite_to_
improved the correlations for the great majority of cases. This measure_plant_health
Franck, F., Juneau, P., & Popovic, R. (2002). Resolution of the photosystem I and photosys-
suggests that SIF–NPC correlations are typically nonlinear.
tem II contributions to chlorophyll fluorescence of intact leaves at room temperature.
Biochimica et Biophysica Acta (BBA) – Bioenergetics, 1556, 239–246.
Frankenberg, C., Fisher, J. B., Worden, J., Badgley, G., Saatchi, S. S., Lee, J. E., ... Yokota, T. (2011).
Overall, it is recommended to sample the SIF profile in at least both New global observations of the terrestrial carbon cycle from GOSAT: patterns of plant
O2-B and O2-A regions to enable accurate quantification of vegetation fluorescence with gross primary productivity. Geophysical Research Letters, 38. http://
dx.doi.org/10.1029/2011GL048738.
photosynthetic activity from SIF retrievals. Genty, B., Wonders, J., & Baker, N. R. (1990). Non-photochemical quenching of Fo in leaves
When advancing from a purely statistical approach towards a is emission wavelength dependent: consequences for quenching analysis and its in-
more process-based approach, it is also recommended to retrieve terpretation. Photosynthesis Research, 26, 133–139.
Guanter, L., Aben, I., Tol, P., Krijger, J. M., Hollstein, A., Köhler, P., ... Landgraf, J. (2014). Poten-
in parallel with SIF retrievals key biophysical and meteorological tial of the TROPOspheric monitoring instrument (TROPOMI) onboard the sentinel-5
variables, in order to disentangle their influence on the SIF signal at precursor for the monitoring of terrestrial chlorophyll fluorescence. Atmospheric
the pixel scale. Such an assimilation approach would provide unique Measurement Techniques Discussions, 7, 12545–12588.
Guanter, L., Alonso, L., Gomez-Chova, L., Meroni, M., Preusker, R., Fischer, J., & Moreno, J.
opportunities to derive unbiased global estimates of photosynthetic (2010). Developments for vegetation fluorescence retrieval from spaceborne high-
carbon uptake. resolution spectrometry in the O(2)-A and O(2)-B absorption bands. Journal of
Geophysical Research-Atmospheres, 115. http://dx.doi.org/10.1029/2009JD013716.
Guanter, L., Frankenberg, C., Dudhia, A., Lewis, P. E., Gomez-Dans, J., Kuze, A., ... Grainger, R. G.
Acknowledgment (2012). Retrieval and global assessment of terrestrial chlorophyll fluorescence from
GOSAT space measurements. Remote Sensing of Environment, 121, 236–251.
Funding for this work was provided by the European Space Agency Guanter, L., Gonzalez-Sanpedro, M. C., & Moreno, J. (2007). A method for atmospheric cor-
rection of ENVISAT/MERIS data over land targets. International Journal of Remote
(ESA), ‘2012 FLEX/Sentinel-3 Tandem Mission Photosynthesis Study’
Sensing, 28, 709–728.
(ESA/ESTEC Contract No. 4000106396/12/NL/AF). Three anonymous re- Hak, R., Lichtenthaler, H. K., & Rinderle, U. (1990). Decrease of the fluorescence ratio F690/
viewers are thanked for providing comments that helped to improve F730 during greening and development of leaves. Radiation and Environmental
the quality of the original manuscript. Biophysics, 29, 329–336.
Hilker, T., Hall, F. G., Coops, N. C., Black, A. T., Jassal, R., Mathys, A., & Grant, N. (2014). Poten-
tials and limitations for estimating daytime ecosystem respiration by combining tower-
References based remote sensing and carbon flux measurements. Remote Sensing of Environment,
150, 44–52.
Alonso, L., Gómez-Chova, L., Vila-Francés, J., Amorós-López, J., Guanter, L., Calpe, J., & Moreno, Jacquemoud, S., & Baret, F. (1990). PROSPECT: a model of leaf optical properties spectra.
J. (2007). Sensitivity analysis of the Fraunhofer Line Discrimination method for the mea- Remote Sensing of Environment, 34, 75–91.
surement of chlorophyll fluorescence using a field spectroradiometer. Geoscience and Joiner, J., Guanter, L., Lindstrot, R., Voigt, M., Vasilkov, A. P., Middleton, E. M., ...
Remote Sensing Symposium 3756–3759. IGARSS 2007. Frankenberg, C. (2013). Global monitoring of terrestrial chlorophyll fluorescence
Baker, N. R. (2008). Chlorophyll fluorescence: a probe of photosynthesis in vivo. Annual from moderate-spectral-resolution near-infrared satellite measurements: methodol-
Review of Plant Biology, 59, 89–113. ogy, simulations, and application to GOME-2. Atmospheric Measurement Techniques, 6,
Baret, F., Weiss, M., Lacaze, R., Camacho, F., Makhmara, H., Pacholcyzk, P., & Smets, B. (2013). 2803–2823.
GEOV1: LAI and FAPAR essential climate variables and FCOVER global time series capi- Joiner, J., Yoshida, Y., Vasilkov, A. P., Yoshida, Y., Corp, L. A., & Middleton, E. M. (2011). First
talizing over existing products. Part1: principles of development and production. observations of global and seasonal terrestrial chlorophyll fluorescence from space.
Remote Sensing of Environment, 137, 299–309. Biogeosciences, 8, 637–651.
Berk, A., Anderson, G. P., Acharya, P. K., Bernstein, L. S., Muratov, L., Lee, J., ... Shettle, E. P. Kraft, S., Del Bello, U., Drusch, M., Gabriele, A., Harnisch, B., & Moreno, J. (2013). On the
(2006). MODTRAN™ 5: 2006 update. Proceedings of SPIE - The International Society for demands on imaging spectrometry for the monitoring of global vegetation fluores-
Optical Engineering, 6233 (II). cence from space. SPIE Proceedings, 8870, 88700N–887012.
Berk, A., Anderson, G. P., Bernstein, L. S., Acharya, P. K., Dothe, H., Matthew, M. W., ... Lichtenthaler, H. K., & Rinderle, U. (1988). The role of chlorophyll fluorescence in the de-
Hoke, M. L. (1999). MODTRAN4 radiative transfer modeling for atmospheric cor- tection of stress conditions in plants. CRC Critical Reviews in Analytical Chemistry,
rection. Proceedings of SPIE - The International Society for Optical Engineering, 19(Suppl. 1), S29–S85.
3756, 348–353. Liu, X., & Liu, L. (2014). Assessing band sensitivity to atmospheric radiation transfer for
Chavez, P. S., Jr. (1996). Image-based atmospheric corrections-revisited and improved. space-based retrieval of solar-induced chlorophyll fluorescence. Remote Sensing, 6(11),
Photogrammetric Engineering and Remote Sensing, 62, 1025–1036. 10656–10675.
Magnani, F., Olioso, A., Demarty, J., Germain, V., Verhoef, W., Moya, I., ... Van der Tol, C. Van Wittenberghe, S., Alonso, L., Verrelst, J., Hermans, I., Delegido, J., Veroustraete, F., ...
(2009). Assessment of vegetation photosynthesis through observation of solar in- Samson, R. (2013). Upward and downward solar-induced chlorophyll fluorescence
duced fluorescence from space. ESTEC Contract No. 20678/07/NL/HE, Final Report. yield indices of four tree species as indicators of traffic pollution in Valencia.
Maxwell, K., & Johnson, G. N. (2000). Chlorophyll fluorescence: a practical guide. Journal Environmental Pollution, 173, 29–37.
of Experimental Botany, 51, 659–668. Van Wittenberghe, S., Alonso, L., Verrelst, J., Moreno, J., & Samson, R. (2015). Bidirectional
Meroni, M., Busetto, L., Colombo, R., Guanter, L., Moreno, J., & Verhoef, W. (2010). Perfor- sun-induced chlorophyll fluorescence emission is influenced by leaf structure and
mance of spectral fitting methods for vegetation fluorescence quantification. Remote light scattering properties — a bottom-up approach. Remote Sensing of Environment,
Sensing of Environment, 114, 363–374. 158, 169–179.
Meroni, M., Rossini, M., Guanter, L., Alonso, L., Rascher, U., Colombo, R., & Moreno, J. Van Wittenberghe, S., Verrelst, J., Rivera, J. P., Alonso, L., Moreno, J., & Samson, R. (2014).
(2009). Remote sensing of solar-induced chlorophyll fluorescence: review of Gaussian processes retrieval of leaf parameters from a multi-species reflectance, ab-
methods and applications. Remote Sensing of Environment, 113, 2037–2051. sorbance and fluorescence dataset. Journal of Photochemistry and Photobiology B:
Mohammed, G. H., Goulas, Y., Magnani, F., Moreno, J., Olejníčková, J., Rascher, U., ... Drusch, Biology, 134, 37–48.
M. (2014). 2012 FLEX/Sentinel-3 Tandem Mission Photosynthesis Study. Final report. Verhoef, W. (1984). Light scattering by leaf layers with application to canopy reflectance
ESA/ESTEC Contract No. 4000106396/12/NL/AF. modeling: the SAIL model. Remote Sensing of Environment, 16, 125–141.
Moreno, J., Asner, G. P., Bach, H., et al. (2006). Fluorescence explorer (FLEX): an optimized Verrelst, J., Alonso, L., Camps-Valls, G., Delegido, J., & Moreno, J. (2012a). Retrieval of veg-
payload to map vegetation photosynthesis from space. AIAA 57th International etation biophysical parameters using Gaussian process techniques. IEEE Transactions
astronautical Congress, Vol. 3. (pp. 2065–2074). Valencia: IAC. on Geoscience and Remote Sensing, 50(5 PART 2), 1832–1843 (6071004).
Moya, I., Daumard, F., Moise, N., Ounis, A., & Goulas, Y. (2006). First airborne multiwave- Verrelst, J., Muñoz, J., Alonso, L., Delegido, J., Rivera, J. P., Camps-Valls, G., & Moreno, J.
length passive chlorophyll fluorescence measurements over La Mancha (Spain) (2012b). Machine learning regression algorithms for biophysical parameter retrieval:
fields. 2nd International Symposium on Recent Advances in Quantitative Remote Sensing opportunities for sentinel-2 and −3. Remote Sensing of Environment, 118, 127–139.
(pp. 820–825). Verrelst, J., Rivera, J. P., Moreno, J., & Camps-Valls, G. (2013). Gaussian processes uncer-
Papageorgiou, G. C., & Govindjee (2004). Chlorophyll a fluorescence - a signature of photo- tainty estimates in experimental sentinel-2 LAI and leaf chlorophyll content retrieval.
synthesis. Springer, 818. ISPRS Journal of Photogrammetry and Remote Sensing, 86, 157–167.
Pedrós, R., Goulas, Y., Jacquemoud, S., Louis, J., & Moya, I. (2010). FluorMODleaf: A new Verrelst, J., Rivera, J. P., Van der Tol, C., Magnani, F., Mohammed, G., & Moreno, J. (2015).
leaf fluorescence emission model based on the PROSPECT model. Remote Sensing of Global sensitivity analysis of the SCOPE model: what drives simulated canopy-
Environment, 114, 155–167. leaving sun-induced fluorescence? Remote Sensing of Environment, 166, 8–21.
Perez-Priego, O., Zarco-Tejada, P. J., Miller, J. R., Sepulcre-Canto, G., & Fereres, E. (2005). Vicent, J., Sabater, N., Tenjo, C., Acarreta, J. R., Manzano, M., Rivera, J. P., ... Moreno, J.
Detection of water stress in orchard trees with a high-resolution spectrometer (2015). FLEX end-to-end mission performance simulator. IEEE Transactions on
through chlorophyll fluorescence in-filling of the O-2-A band. IEEE Transactions on Geoscience and Remote Sensing (submitted for publication).
Geoscience and Remote Sensing, 43, 2860–2869. Von Caemmerer, S. (2000). Biochemical models of Leaf Photosynthesis. Techniques in
Porcar-Castell, A. (2011). A high-resolution portrait of the annual dynamics of photo- plant science no 2. Collingwood: CSIRO Publishing.
chemical and non-photochemical quenching in needles of Pinus sylvestris. Von Caemmerer, S. (2013). Steady-state models of photosynthesis. Plant, Cell &
Physiologia Plantarum, 143(2), 139–153. Environment, 36(9), 1617–1630.
Porcar-Castell, A., Garcia-Plazaola, J. I., Nichol, C. J., Kolari, P., Olascoaga, B., Kuusinen, N., ... Weis, E., & Berry, J. A. (1987). Quantum efficiency of photosystem II in relation to energy
Nikinmaa, E. (2012). Physiology of the seasonal relationship between the photo- dependent quenching of chlorophyll fluorescence. Biochimica et Biophysica Acta, 894,
chemical reflectance index and photosynthetic light use efficiency. Oecologia, 198–208.
170(2), 313–323. Wohlfahrt, G., & Gu, L. (2015). Opinion: the many meanings of gross photosynthesis and
Porcar-Castell, A., Tyystjärvi, E., Atherton, J., Van der Tol, C., Flexas, J., Pfündel, E. E., ... their implication for photosynthesis research from leaf to globe. Plant, Cell &
Berry, J. A. (2014). Linking chlorophyll a fluorescence to photosynthesis for remote Environment. http://dx.doi.org/10.1111/pce.12569.
sensing applications: mechanisms and challenges. Journal of Experimental Botany, Wolanin, A., Rozanov, V. V., Dinter, T., Noël, S., Vountas, M., Burrows, J. P., & Bracher, A.
65(15), 4065–4095. (2015). Global retrieval of marine and terrestrial chlorophyll fluorescence at its red
Rascher, U., Alonso, L., Burkart, A., Cilia, C., Cogliati, S., et al. (2015). Sun-induced fluores- peak using hyperspectral top of atmosphere radiance measurements: feasibility
cence – a new probe of photosynthesis: first maps from the imaging spectrometer study and first results. Remote Sensing of Environment, 166, 243–261.
HyPlant. Global Change Biology. http://dx.doi.org/10.1111/gcb.13017. Wong, C. Y. S., & Gamon, J. A. (2015a). Three causes of variation in the photochemical re-
Rasmussen, C. E., & Williams, C. K. I. (2006). Gaussian processes for machine learning. MIT flectance index (PRI) in evergreen conifers. New Phytologist, 206, 187–195.
Press. Wong, C. Y. S., & Gamon, J. A. (2015b). The photochemical reflectance index provides an
Reichstein, M., Falge, E., Baldocchi, D., Papale, D., Aubinet, M., Berbigier, P., et al. (2005). optical indicator of spring photosynthetic activation in evergreen conifers. New
On the separation of net ecosystem exchange into assimilation and ecosystem respi- Phytologist, 206, 196–208.
ration: review and improved algorithm. Global Change Biology, 11(9), 1424–1439. Zarco-Tejada, P. J., Gonzalez-Dugo, V., & Berni, J. A. J. (2012). Fluorescence, temperature
Rossini, M., Nedbal, L., Guanter, L., Ač, A., Alonso, L., Burkart, A., ... Rascher, U. (2015). Red and narrow-band indices acquired from a UAV platform for water stress detection
and far red Sun-induced chlorophyll fluorescence as a measure of plant photosynthe- using a micro-hyperspectral imager and a thermal camera. Remote Sensing of
sis. Geophysical Research Letters, 42, 1632–1639. Environment, 117, 322–337.
Saltelli, A., Annoni, P., Azzini, I., Campolongo, F., Ratto, M., & Tarantola, S. (2010). Variance Zarco-Tejada, P. J., Morales, A., Testi, L., & Villalobos, F. J. (2013). Spatio-temporal patterns
based sensitivity analysis of model output. Design and estimator for the total sensitiv- of chlorophyll fluorescence and physiological and structural indices acquired from
ity index. Computer Physics Communications, 181, 259–270. hyperspectral imagery as compared with carbon fluxes measured with eddy covari-
Saltelli, A., Tarantola, S., & Chan, K. P. S. (1999). A quantitative model-independent meth- ance. Remote Sensing of Environment, 133, 102–115.
od for global sensitivity analysis of model output. Technometrics, 41, 39–56. Zhang, X., Liang, S., Zhou, G., Wu, H., & Zhao, X. (2014a). Generating global LAnd surface
Subhash, N., & Mohanan, C. N. (1997). Curve-fit analysis of chlorophyll fluorescence spec- satellite incident shortwave radiation and photosynthetically active radiation prod-
tra: application to nutrient stress detection in sunflower. Remote Sensing of ucts from multiple satellite data. Remote Sensing of Environment, 152, 318–332.
Environment, 60, 347–356. Zhang, Y., Guanter, L., Berry, J. A., Joiner, J., Van der Tol, C., Huete, A., ... Köhler, P. (2014b).
Timmermans, J., Su, Z., Van der Tol, C., Verhoef, A., & Verhoef, W. (2013). Quantifying the Estimation of vegetation photosynthetic capacity from space-based measurements of
uncertainty in estimates of surface-atmosphere fluxes through joint evaluation of the chlorophyll fluorescence for terrestrial biosphere models. Global Change Biology, 20,
SEBS and SCOPE models. Hydrology and Earth System Sciences, 17, 1561–1573. 3727–3742.
Van der Tol, C., Berry, J. A., Campbell, P. K. E., & Rascher, U. (2014). Models of fluorescence Zhao, F., Guo, Y., Verhoef, W., Gu, X., Liu, L., & Yang, G. (2014). A method to reconstruct the
and photosynthesis for interpreting measurements of solar-induced chlorophyll fluo- solar-induced canopy fluorescence spectrum from hyperspectral measurements.
rescence. Journal of Geophysical Research — Biogeosciences, 119, 2312–2327. Remote Sensing, 6(10), 10171–10192.
Van der Tol, C., Verhoef, W., Timmermans, J., Verhoef, A., & Su, Z. (2009). An integrated
model of soil-canopy spectral radiances, photosynthesis, fluorescence, temperature
and energy balance. Biogeosciences, 6, 3109–3129.

Verrelst 2016 SIF-NPC RSE

Uploaded by

Copyright:

Available Formats

Verrelst 2016 SIF-NPC RSE

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Verrelst 2016 SIF-NPC RSE

Uploaded by

Copyright:

Available Formats

Remote Sensing of Environment 176 (2016) 139–151

Contents lists available at ScienceDirect

Remote Sensing of Environment

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

Evaluating the predictive power of sun-induced chlorophyll ﬂuorescence

2.2. SIF retrieval bands Table 1

2.4. Regression analysis

Input Deﬁnition Unit Min Max Default

Index Ranging variables Justiﬁcation

R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE

Hα, O2–B, water Two peaks and Fall:

R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE

Hα, O2–B, water Two peaks and Fall:

R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE R2 RMSE

1 Vcmo 1 0.0040 1 0.0053 1 0.0038 0.9900 1.4283 1 0.0039 1 0.0089 1 0.0028

You might also like