Introduction
Understanding the mechanisms that drive the carbon dioxide (CO2)
exchange of terrestrial ecosystems is one of the main challenges for
ecologists working on climate change (Beer et al., 2010). Plant gross
photosynthesis, also referred to as gross primary productivity (GPP), is one
of the major components of the global carbon cycle. It interacts in complex
ways with environmental factors such as radiation, nutrients, soil moisture,
vapour pressure deficit, air temperature and soil temperature (Drolet et al.,
2005). Plant biochemistry and structure determine many fundamental ecosystem
patterns, processes and dynamics (Lambers et al., 1998; Waring and Running,
1998). The canopy nitrogen content regulates the canopy photosynthetic
capacity and the canopy light use efficiency (ε) (Ollinger et
al., 2008). In addition, the canopy chlorophyll content plays an important
role in controlling ecosystem photosynthesis and carbon gain (Peng et al.,
2011; Gitelson et al., 2006).
Optical remote sensing can help ecologists in qualitatively and
quantitatively assessing plant and canopy properties (e.g. biomass – Vescovo et
al., 2012; water content – Clevers et al., 2010; nitrogen content – Ollinger et
al., 2008, Knyazikhin et al., 2012; chlorophylls – Gitelson et al., 2006; and
photosynthetic rate – Inoue et al., 2008) that drive ecosystem processes related
to the carbon cycle.
Empirically and physically based methods have been proposed by several authors to
interpret optical plant and canopy properties. Empirical methods consist of,
for example, linear regression analysis between plant or canopy properties
and optical data. The most widely used empirical methods are hyperspectral index
methods (Peñuelas et al., 1993; Sims and Gamon, 2002; Inoue et al., 2008)
and multivariable statistical methods, e.g. stepwise linear regression,
genetic algorithm and neural network (Grossman et al., 1996; Riaño et al.,
2005a; Li et al., 2007). Physical methods are based on the use of radiative
transfer models (RTMs) to simulate light absorption and scattering through
the canopy as a function of canopy structure and leaf biochemical composition
(Jacquemoud et al., 2000; Zarco-Tejada et al., 2003). Therefore, RTMs help in
quantifying the contribution of canopy biophysical and biochemical variables
to canopy reflectance. Such models can be used for identifying regions of the
light spectrum that are of particular importance for specific biophysical
properties of vegetation. For example, it was demonstrated that the red-edge
region (between 680 to 730 nm) of the spectrum is sensitive to the leaf
chlorophyll content and leaf area index (LAI) (Baret et al., 1992). It is
also well-accepted that an increase in LAI includes a decrease in reflectance
in the red and an increase in the near-infrared (NIR) region (Jacquemoud,
1993). In the NIR region, effects of LAI and the leaf angle distribution
equally contribute to the reflectance response (Bacour et al., 2002a). NIR
reflectance between 800 and 850 nm is also related to canopy N
content (Ollinger et al., 2008; Knyazikhin et al., 2012). In addition, the
combination of the reflectance in NIR and in the short-wave infrared region
(SWIR) is correlated with canopy water content (Colombo et al., 2008), but
the reflectance between 1000 and 1400 nm is also highly sensitive to LAI.
Therefore, some attention is needed when these spectral regions are used to retrieve
water content, considering that the canopy properties in a given ecosystem
often covary (Bacour et al., 2002b).
The drawback of such an approach consists in the fact that the process of
building a model implies approximations and assumptions. For this reason we
opted for a purely data-based approach such as the hyperspectral index
approach. This method consists of the use of spectral vegetation indices
(VIs) defined as spectral band ratios or normalized band ratios between the
reflectance in the visible (VIS)-vs.-NIR, VIS-vs.-VIS or NIR-vs.-NIR region.
Description of the study sites and period.
Site characteristics
Amplero
Neustift
Monte Bondone
Latitude
41.9041
47.1162
46.0296
Longitude
13.6052
11.3204
11.0829
Elevation (m)
884
970
1550
Mean annual temperature (∘C)
10.0
6.5
5.5
Mean annual precipitation (mm)
1365
852
1189
Vegetation type
Seslerietum
Pastinaco–
Nardetum
apenninae
Arrhenatheretum
Alpigenum
Study period*
111–170, 2006 (9)
122–303, 2006 (16)
129–201, 2005 (13)
124–192, 2006 (12)
* Range refers to days of the year (DOY); this is followed by the year; the number of hyperspectral measurement dates is given in brackets.
The typical optical sampling approach, which consists of linking spectral
observations with CO2 fluxes, is based on the Monteith (1972,
1977) equation:
GPP=ε×PAR×fAPAR,
where ε is the light use efficiency and fAPAR is the fraction of
absorbed photosynthetically active radiation); both ε and
fAPAR
can be retrieved by remote optical observations. A large number of VIs that
can potentially be used to model the productivity of terrestrial ecosystems
(as a proxy of ε and fAPAR) have been suggested (Inoue et al.,
2008; Coops et al., 2010; Peñuelas et al., 2011; Rossini et al., 2012).
The various VIs differ in their sensitivity to changes in photosynthetic
status. “Greenness indices” – such the widely used normalized difference
vegetation index (NDVI) – have been demonstrated to be a good proxy for fAPAR but are
not sensitive to rapid changes in plant photosynthesis which are induced by
common environmental and anthropogenic stressors (Gitelson et al., 2008;
Hmimina et al., 2014; Soudani et al., 2014). However, in ecosystems
characterized by strong dynamics (e.g. grasslands and crops with a strong
green-up and senescence), other VIs are able to effectively monitor seasonal
changes in biophysical parameters controlling canopy photosynthesis, such as
fAPAR and chlorophyll content, and, consequently, can be adopted to monitor the
seasonal and spatial variability of carbon fluxes (Gitelson et al., 2012;
Sakowska et al., 2014). Short-term changes in ε can be remotely
detected through a spectral proxy of the xanthophyll cycle (photochemical
reflectance index, PRI; Gamon et al., 1992). The PRI is one of the most
promising VIs for a direct estimation of photosynthetic light use efficiency
and of its seasonal and diurnal variations (Nichol et al., 2002). Latest
developments in the sun-induced fluorescence method may allow even more
direct remote sensing of plant photosynthesis in the near future (Meroni et
al., 2009; Rossini et al., 2010; Frankenberg et al., 2011). On the canopy scale,
the relationship between PRI and ε was shown to be
site-dependent (Garbulsky et al., 2011; Goerner et al., 2011) and strongly
affected by environmental conditions (Soudani et al., 2014).
Whereas previous studies have demonstrated the ability of remote-sensing data
to allow modelling ecosystem GPP on the ecosystem scale (e.g. Gianelle et al.,
2009; Wohlfahrt et al., 2010; Rossini et al., 2012; Sakowska et al., 2014), a
universal model for GPP estimation applicable across different ecosystems and
a wide range of environmental conditions is still missing. Previous studies
often focussed on single sites with specific characteristics (e.g. climate,
vegetation composition, soil type; see Wohlfahrt et al., 2010). In addition,
different studies often used different sensors, platforms and protocols
(Balzarolo et al., 2011), making generalization difficult. Moreover, most of
the studies have either relied on reflectance measurements in a few spectral
wavebands (e.g. Wohlfahrt et al., 2010; Sakowska et al., 2014), missing
potentially important information in under-sampled spectral regions. In order
to overcome such heterogeneity in spectrometry measurements, SpecNet
(http://specnet.info; Gamon et al., 2006), the European COST Action
ES0903 (COST – Cooperation in Science and Technology; EUROSPEC; http://cost-es0903.fem-environment.eu/) and the COST Action ES1309 (OPTIMISE;
http://www.cost.eu/domains_actions/essem/Actions/ES1309) focused on the
definition of a standardized protocol for making optical measurements at the
eddy covariance CO2 flux towers (Gamon et al., 2010).
The overarching objective of the present paper is thus to develop a common
framework for predicting grassland carbon fluxes and ecophysiological
parameters based on optical remote-sensing data across measurement sites
exposed to diverse natural (climate) and anthropogenic (management) factors.
To this end we combine eddy covariance CO2 flux measurements with
ground-based hyperspectral reflectance measurements for six different
grasslands in Europe. In order to make the optical and fluxes measurements
comparable, these were acquired at the six sites following a common protocol
resulting in a unique standardized data set. We focused on European
grasslands, which cover roughly 22 % (80 million ha) of the EU-25 land
area and are thus among the dominating ecosystem types in Europe (EEA, 2005).
Accordingly, their role in the European carbon balance has received a lot of
scientific interest (Soussana et al., 2007; Gilmanov et al., 2007; Wohlfahrt
et al., 2008; Ciais et al., 2010). Direct measurements of the grassland
carbon exchange have been carried out and are still ongoing at a number of
different grassland sites in Europe (e.g. Soussana et al., 2007; Cernusca et
al., 2008). Scaling up these plot-level measurements to the continental scale
requires a modelling approach typically based on or supported by remotely
sensed data. Therefore, we believe that this study will improve the current
knowledge on modelling the carbon dynamics of European grasslands.
Materials and methods
Experimental site description
This study was carried out at six experimental mountain grassland sites in
Europe covering different climatic and grassland management conditions
existing in the mountain regions of Europe, which were already part of the
preceding study by Vescovo et al. (2012). This data set combined in situ
hyperspectral, biophysical and flux measurements based on common protocol
(for more details, see Sects. 2.2, 2.3 and 2.4). This data set is unique since
no common protocol for hyperspectral measurements exists in the various eddy
covariance networks (e.g. FLUXNET). In this study, three of these sites
(Amplero, Neustift and Monte Bondone, see Tables 1 and S1 in the Supplement)
composed the main data set used in the analysis, while the other three sites
(Table S2) were used to independently validate the models obtained with the
main data set.
Main study sites (Tables 1 and S1):
Amplero
The Amplero site is situated in the Mediterranean Apennine mountain region
of Italy (41.90409∘ N, 13.60516∘ E) at 884 m a.s.l. This
site is characterized by mild, rainy winters and by an intense drought in
summer. Amplero is managed as a hay meadow, with one cut in late June and
extensive grazing during summer and autumn (Balzarolo, 2008).
Monte Bondone
The Monte Bondone site is situated in the Italian Alps (46.01468∘ N,
11.04583∘ E) at 1550 m a.s.l. This site is characterized by a
typical subcontinental climate, with mild summers and precipitation peaks in
spring and autumn. Monte Bondone is managed as an extensive meadow, with one
cut in mid-July.
Neustift
The Neustift grassland site is located in the Austrian Alps
(47.11620∘ N, 11.32034∘ E) at 970 m a.s.l. The climate of
this area is continental and Alpine, with precipitation peaks during the summer
(July). This site is intensively managed as a hay meadow, with three cuts in
mid-June, at the beginning of August and at the end of September.
Validation sites:
Längenfeld
The site Längenfeld is located in the Austrian Alps (47.0612∘ N,
10.9634∘ E) at 1180 m a.s.l. The climate of this area is
continental/Alpine; however, compared to the other Alpine sites in this study,
the site receives comparably less precipitation due to rain shadowing effects
from both the north and south. The site is intensively managed as a hay
meadow, with three cuts in mid-June, mid-August and mid-October.
Leutasch
The site Leutasch is located in the Austrian Alps (47.3780∘ N,
11.1627∘ E) at 1115 m a.s.l. The climate of this area is Alpine,
with substantial precipitation due to its position on the north range of the
Alps. The site is extensively managed as a hay meadow, with two cuts at the
end of June and beginning of September.
Scharnitz
The site Scharnitz is located very close to Leutasch (47.3873∘ N,
11.2479∘ E) at 964 m a.s.l., and the climate is thus very similar
to Leutasch. The site is extensively managed as a hay meadow, with two cuts at
the beginning of July and beginning of September.
Hyperspectral reflectance measurements
The canopy hyperspectral reflectance measurements were collected at each site
under clear-sky conditions close to solar noon (between 11:00 to 14:00
Central European Time) using the same model of a portable spectroradiometer
(ASD FieldSpec HandHeld, Inc., Boulder, CO, USA) at all sites. The
spectroradiometer acquires reflectance values between 350 and 1075 nm, with a
full-width half maximum (FWHM) of 3.5 nm and a spectral resolution of 1 nm.
In order to achieve a better match between the eddy covariance flux footprint
and optical measurements, a cosine diffuser foreoptic (ASD Remote Cosine
Receptor, Inc., Boulder, CO, USA), calibrated by the manufacturer, was used
for nadir and zenith measurements (Gianelle et al., 2009; Fava et al., 2009;
Meroni et al., 2011). The ASD's cosine receptor is designed with a geometry
and material that provides a hemispherical field of view (FOV) of
180∘ and optimizes the cosine response. To reduce the nadir FOV
contamination (i.e. sky irradiance and canopy irradiance) due to the
hemispherical view of the sensor the instrument was placed on a 1.5 m long
horizontal arm at a height of 1.5 m above the ground. To avoid the zenithal
FOV contamination, the measurements were taken at least at a 15 m distance
from the eddy covariance tower (maximum height of the tower was 6 m). The
vegetation irradiance (sensor-pointing nadir) and sky irradiance (sensor-pointing zenith) were measured by rotating the spectroradiometer alternately
to acquire spectra from the vegetation and from the sky. Hemispherical
reflectance was derived as the ratio of reflected to incident radiance. Each
reflectance spectrum was automatically calculated and stored by the
spectroradiometer as an average of 20 readings. Before starting each spectral
sampling, a dark current measurement was done. For more details on the experimental set-up, see Vescovo et al. (2012). Spectral measurements were
collected from spring until the cutting date at Amplero and Monte Bondone,
while at the site in Neustift, which is cut three times during the season,
spectral measurements were taken about once per week throughout the growing
season of 2006.
Biophysical and biochemical canopy properties
Samples for dry phytomass, nitrogen and water content measurements were
collected at the time of the hyperspectral measurements in the field of view
of the hyperspectral sensor (see Vescovo et al., 2012 for more details). A
similar data set was collected in 2013 at Monte Bondone by combining
hyperspectral data with chlorophyll measurements. Chlorophyll samples were
collected in the field of view of the hyperspectral sensor and chlorophyll
content was detected by UV–VI spectroscopy. First, the samples were ground
with liquid nitrogen and then immersed in 80 % acetone solution
(0.1 g per 10 mL), shaken for 10 min in an automatic shaker at
250 rpm (Universal Table Shaker 709) and centrifuged at 4000 rpm for 10 min
(Eppendorf 5810 R) in order to remove particles from the solution. The
absorbance of extracted solutions was measured at 470, 646.8 and 663.2 nm by
a UV–VIS spectrophotometer (Shimadzu UV-1601), and the concentrations of
chlorophyll a (Ca), chlorophyll b (Cb) and carotenoids
(Cx+c) were calculated as proposed by Lichtenthaler (1987). The weight
of sampled sediment was used to calculate pigment concentrations per unit
leaf mass (mg g-1), and the weight of green biomass per ground area was
used to obtain the total chlorophyll content (mg m-2).
CO2 flux measurements
Continuous measurements of the net ecosystem CO2 exchange (NEE) were
made by the eddy covariance (EC) technique (Baldocchi et al., 1996; Aubinet
et at., 2012) at the six study sites using identical instrumentation. The
three wind components and the speed of sound were measured using ultrasonic
anemometers, and CO2 molar densities were measured using open-path infrared gas
analysers (IRGAs), as detailed in Tables S1 and S2. Raw data were acquired at
20 Hz and averaged over 30 min time windows in post-processing. Turbulent
fluxes were obtained from raw data by applying block averaging (Monte
Bondone, Neustift, validation sites) or linear de-trending (Amplero) methods
with a time window of 30 min. A 3-D coordinate correction was performed
according to Wilczak et al. (2001). The CO2 fluxes were corrected for,
the effect of air density fluctuations as proposed by Webb et al. (1980).
Low- and high-pass filtering was corrected for following Aubinet et
al. (2000) (Amplero, Monte Bondone) or Moore (1986) (Neustift, validation
sites). Data gaps due to sensors malfunctioning or violation of the
assumptions underlying the EC method were removed and filled using the
gap-filling and flux-partitioning techniques as proposed in Wohlfahrt et
al. (2008). Ecosystem respiration (Reco) was calculated from the y intercept
of the light response model (see Eq. 4). Gross primary productivity (GPP) was
calculated as the difference between NEE and Reco. Half-hourly NEE and GPP
values were averaged between 11:00 to 14:00 solar local time (at the time
window of optical measurements) to allow for direct comparison with the
hyperspectral data, and daily sums were also computed. At each site the
following supporting environmental measurements were acquired:
photosynthetically active radiation (PAR; quantum sensors), air temperature
(Ta; PT100, thermistor and thermoelement sensors) and humidity (RH;
capacitance sensors) at some reference height above the canopy and soil
temperature (Ts; PT100, thermistor and thermoelement sensors) and volumetric
water content (SWC; dielectric and time-domain reflectometry sensors) in the
main rooting zone. In this study we used CO2 flux and meteorological
data of the years 2005 and 2006 for Monte Bondone and of 2006 for the other
sites.
Estimation of grassland ecophysiological parameters
Canopy light use efficiency (ε) was derived from
photosynthetically active radiation (PAR) absorbed by the canopy (APAR) as
ε=GPPAPAR=GPPPAR×fAPAR
and was estimated both at midday and daily time resolutions. We estimated the
fraction of PAR absorbed by the canopy (fAPAR) from measured values of the
leaf area index (LAI) using the Lambert–Beer law:
fAPAR=0.951-e(-kLAI),
where k is the canopy extinction coefficient (fixed at k=0.4 as defined
for a southern mixed-grass prairie in Texas; Kiniry et al., 2007) and 0.95 is
the proportion of intercepted PAR that is absorbed by plants (Schwalm et al.,
2006). LAI was quantified non-destructively by an indirect method based on
canopy PAR transmission using line PAR sensors (SunScan, Delta-T, UK) and an inversion of an RTM (Wohlfahrt et al., 2001). These measurements were done
within the footprint area of the spectroradiometer simultaneously with the
hyperspectral measurements.
Three additional key parameters of the response of NEE to PAR were extracted
by fitting measured NEE and PAR to a simple Michaelis–Menten-type model:
NEE=-αPARFsatαPAR+Fsat+Reco,
where α represents the apparent quantum yield (µmol
CO2 µmol-1 photons), Fsat the asymptotic
value of GPP (µmol CO2 m-2 s-1), PAR the
photosynthetically active radiation (µmol photons
m-2 s-1) and Reco the ecosystem respiration
(µmol CO2 m-2 s-1). For all sites, using the
Levenberg–Marquardt (Marquardt, 1963) algorithm the parameters of Eq. () were
estimated by fitting Eq. () to both day and nighttime data, which
were pooled into 3-day blocks centred on the date of the hyperspectral data
acquisition. For each acquisition date, we then used Eq. () to
derive GPP at an incident PAR of 1500 µmol m-2 s-1,
referred to as GPPmax in the following.
Hyperspectral data analysis
In order to explore the information content of the hyperspectral data for
estimating CO2 fluxes (i.e. midday and daily average of NEE and GPP) and
ecophysiological parameters (i.e. α, ε and
GPPmax), we performed a correlation analysis between spectral
reflectance indices (independent variables) and these (dependent) variables.
To this end, we derived spectral ratio (SR; Eq. ), spectral
difference (SD; Eq. ) and normalized spectral difference (NSD;
Eq. ) indices using all possible two-band (i and j) reflectance
(ρ) combinations between 400 and 1000 nm (180 600 combinations).
These three formulations were selected since they represent the most common
equations used to compute vegetation indices (see Table 2).
SRi,j=ρiρjSDi,j=ρi-ρjNSDi,j=ρi-ρjρi+ρj
Linear regression analysis was performed among all possible
wavelength combinations for all three index types (SR, NSD and SD) and the
investigated dependent variables.
The performance of linear models in predicting dependent variables (i.e.
carbon fluxes and ecophysiological parameters) was evaluated by the
coefficient of determination (R2) and root mean square error (RMSE). The
coefficients of determination (R2) resulting from the linear models were
visualized in correlograms, as in the example in Fig. 1.
A selected example of a correlogram between NSD-type indices and
midday average GPP for all sites pooled. The correlogram shows all R2
values, the white dot indicates the two-band combination with the highest
R2 value and the red dots indicate the location of the reference VIs
reported in Table 2 (SR: simple ratio; GRI: green ratio index; WI: water
index; SRPI: simple ratio pigment index; NDVI: normalized difference
vegetation index; NPQI: normalized phaeophytinization index; NPCI: normalized
pigment chlorophyll index; CI: chlorophyll index; red-edge NDVI; SIPI:
structural independent pigment index; PRI: photochemical reflectance index).
We also calculated four SR and seven NSD indices which are commonly used in
relation to vegetation activity and CO2 fluxes (Table 2). Figure 1 shows
the location of these indices in the waveband space of the correlograms. In
this analysis, we also considered the enhanced vegetation index (EVI), which
is one of the most frequently used vegetation indices to predict CO2
fluxes. In Fig. 1 the location of the EVI is not shown since this index is
computed by the combination of three spectral bands as shown in Table 2.
Summary of the vegetation indices characteristics used in this
study.
Index name
Formula
Use
Reference
and acronym
Simple spectral ratio indices
Simple ratio (SR or RVI)
SR = R830/R660
Greenness
Jordan (1969)
Green ratio index (GRI)
GRI = R830/R550
Greenness
Peñuelas and Filella (1998)
Water index (WI)
WI = R900/R970
Water content, leaf waterpotential, canopy water content
Peñuelas et al. (1993)
Simple ratio pigmentindex (SRPI)
SRPI = (R430)/(R680)
Carotenoid/clorophyll ratio
Peñuelas et al. (1995)
Chlorophyll index (CI)
CI = (R750/R720)-1
Chlorophyll content
Gitelson et al. (2005)
Normalized spectral difference vegetation indices
Normalized differencevegetation index (NDVI)
NDVI = (R830-R660) / (R830+ R660)
Greenness
Rouse et al. (1973)
Normalized phaeophytinization index (NPQI)
NPQI = (R415-R435) / (R415+ R435)
Carotenoid/chlorophyll ratio
Barnes et al. (1992)
Normalized pigment chlorophyll index (NPCI)
NPCI = (R680-R430) / (R680+ R430)
Chlorophyll ratio
Peñuelas et al. (1994)
Red-edge NDVI
NDVIRedEdge= (R750-R720) / (R750+ R720)
Chlorophyll content
Gitelson andMerzlyak (1994)
Structural independentpigment index (SIPI)
SIPI = (R800-R445) / (R800+ R445)
Chlorophyll content
Peñuelas et al. (1995)
Photochemical reflectanceindex (PRI)
PRI = (R531-R570) / (R531+ R570)
Photosynthetic light useefficiency (and leaf pigmentcontents)
Gamon et al. (1992)
Improved for soil and atmospheric effects
Enhanced vegetationindex (EVI)
EVI = 2.5 (R830-R660) / (1+R830+6 R660-7.5 R460)
Vegetation index improved for soil and atmospheric effects
Huete et al. (1997)
The robustness of the model selected on the basis of the best band
combinations for all ecophysiological parameters for each site and all sites
pooled was tested by the leave-one-out cross-validation technique. The
predictive performance was expressed as the cross-validated coefficient of
determination (RCV2) and the cross-validated root mean square
error (RMSECV). In addition, the capability of the selected
models in predicting different ecophysiological parameters was tested by
applying the selected models to the validation data set (Table S2) composed
of three different grasslands not used in the previous analysis. This data set
was selected because the hyperspectral and flux data were collected by using
exactly the same protocol applied for the main data set (see Sect. 2.1).
In order to explore the basis of the correlation between the selected band
combinations and ecophysiological variables (e.g. α,
GPPmax, GPP, ε), the relationship between the
selected bands and biophysical parameters such as dry phytomass, nitrogen and
water content collected during the field campaign in the same footprint of
the hyperspectral measurements was examined.
Band selection based on the combination of random forests and genetic
algorithm (GA–rF)
In order to complement the more conventional analysis described in the
previous section, we also explored the use of a hybrid feature selection
strategy based on a genetic algorithm and random forests (GA–rF). The first
method was used for the feature selection and the second one as regression
for predicting the target variables. First of all, the original data set was
aggregated to 10 nm bands in order to reduce the effects of autocorrelation
in frequency space. The algorithm generates a number of possible model
solutions (chromosomes) and uses these to evolve towards an approximation of
the best solution of the model. In our case the genes of each chromosome
correspond to the wavebands. We made use of five genes for each chromosome in
order to overcome overfitting. Each population of 1000 chromosomes evolved
for 200 generations. The mutation chance was set to the inverse of a population
size increase of 1. The fitness of each chromosome was measured by
applying the random forest algorithm (Breiman, 2001). This was used as an
ensemble method for regression that is based on the uncontrolled development
of decision trees (n=100). We opted for this method because of its
demonstrated efficiency with large data sets. In combining the two methods, we
choose the mean squared error as the target variable to be minimized.
Seasonal variation of meteorological variables, LAI, CO2 fluxes
and ecophysiological parameters for the period of the hyperspectral
measurements at the three investigated grasslands. (a) Midday
average photosynthetically active radiation
(PAR; µmol m-2 s-1; solid black line) and daily average
air temperature (∘C; dotted grey line); (b) daily
precipitation (Rain; mm; solid black line) and daily average soil water
content (SWC; m3 m-3; dotted grey line); (c) leaf area
index (LAI; m2 m-2; solid black line) and light use efficiency
(ε; µmol photons µmol-1 CO2; dotted
grey line); (d) apparent quantum yield (α; µmol
CO2 µmol-1 photons; solid black line) and gross primary
production at saturating light (GPPmax;
µmol m-2 s-1; dotted grey line); (e) midday
average net ecosystem CO2 exchange (NEE;
µmol m-2 s-1; solid black line) and gross primary
production (GPP; µmol m-2 s-1; grey dotted line);
vertical lines in the lowermost panels indicate the dates of
mowing.
Results
Seasonal variation of meteorological variables, LAI and CO2
fluxes
Environmental conditions and the seasonal development of LAI, NEE, GPP
α, ε and GPPmax during the study period are
shown in Fig. 2. A strong influence of the typical climatic conditions at the
three study sites was evident: Amplero was characterized by a Mediterranean
climate, with highest incoming radiation and temperatures and the lowest
amount of precipitation, which translated into a substantial seasonal drawdown
of soil moisture. Monte Bondone and Neustift, more influenced by a continental
Alpine climate, experienced comparably lower temperatures with higher
precipitation and soil moisture with respect to Amplero (Fig. 2).
Maximum LAI values were similar at Monte Bondone and Amplero
(2.8–3.4 m2 m-2), while twice as much leaf area developed at
the more intensively managed study site Neustift. The latter was also
characterized by higher NEE and GPP (i.e. more photosynthesis and net uptake
of CO2). The reductions in leaf area related to the cuts of the
grasslands were associated, as expected, with marked increases and reductions
in NEE and GPP, respectively. The canopy light use efficiency, ε, was inversely related to GPP and LAI, peaking at the beginning of the
season at Amplero and Monte Bondone (0.01–0.10 µmol photons
µmol-1 CO2), while for Neustift ε showed the
highest values after the cuts (0.01–0.20 µmol photons
µmol-1 CO2). At Amplero, α and GPPmax
peaked in spring and then decreased during the summer drought period, while
at Neustift and Monte Bondone, temporal patterns of α and
GPPmax were more strongly affected by management.
Correlograms of R2 values for α, GPPmax, midday averaged GPP, ε and NEE, on the one hand, and NSD-type indices for
Amplero, Neustift, Monte Bondone (both study years pooled) and all sites
pooled on the other. The white dots indicate the position of paired band combinations
corresponding to the maximum R2.
Correlograms of R2 values for α, GPPmax, midday averaged GPP, ε and NEE, on the one hand, and SR-type indices for
Amplero, Neustift, Monte Bondone (both study years pooled) and all sites
pooled on the other. The white dots indicate the position of paired band combinations
corresponding to the maximum R2.
Correlograms of R2 values for α, GPPmax, midday averaged GPP, ε and NEE, on the one hand, and SD-type indices for
Amplero, Neustift, Monte Bondone (both study years pooled) and all sites
pooled on the other. The white dots indicate the position of paired band combinations
corresponding to the maximum R2.
Results of linear correlation analysis for α,
GPPmax, midday averaged GPP, ε and NEE, on the one hand, and
selected best NSD-type indices for (a) Amplero, (b) Neustift, (c) Monte
Bondone (both study years pooled) and (d) all sites pooled on the other.
R2 – coefficient of determination; RMSE – root mean square error;
R2cv – cross-validated coefficient of determination; RMSEcv –
cross-validated root mean square error. The solid red lines indicate the
fitted models, and the dotted red lines represent the 95 % upper and lower
confidence bounds.
Results of validation of linear regression models between VIs ((a)
NSD-type; (b) SR-type; (c) SD-type) and ecophysiological parameters: α, ε (midday average), GPPmax and midday average
CO2 fluxes (NEE and GPP). R2 – coefficient of determination.
Different colours represent results of the validation performed applying to
the three new sites the model for Amplero (in magenta), Neustift (in red) and
Monte Bondone (in blue) and a model-parameterized grouping Monte Bondone and Neustift (M.Bondone&Neustift; in black). Statistical significance is
indicated as * (p<0.05), ** (p<0.01) and *** (p<0.001). The black lines are 1:1 lines.
Correlation between selected (a) NSD-, (b) SR- and (c) SD-type
indices for the α, GPPmax, midday GPP, midday
ε and midday NEE (plots in the columns) and the total
chlorophyll content for Monte Bondone in 2013. R2 – coefficient of
correlation; RMSE – root mean square error; R2cv – cross-validated
coefficient of correlation; RMSEcv – cross-validated root mean square error.
The solid red lines indicate the fitted models, and the dotted red lines
represent the 95 % upper and lower confidence bounds. The selected bands
for computing NSD-, SR- and SD-type indices are reported in brackets.
Results of the GA–rF method for band selection for Amplero,
Neustift, Monte Bondone and all sites pooled for (a) α and
GPPmax, (b) midday average ε and CO2 fluxes (NEE
and GPP); (c) daily average ε and CO2 fluxes (NEE and GPP).
Hyperspectral data and their relation to CO2 fluxes and
ecophysiological parameters
Figures 3–5 show correlograms between NSD-, SR- and SD-type indices,
respectively, and the investigated dependent midday ecophysiological
parameters and fluxes. The correlograms for daily data can be found in the
Supplement (Figs. S2–S4 in Supplement). Selected examples of key spectral
signatures of the investigated grasslands are shown in Fig. S1 in the
Supplement.
A number of interesting insights may be gained from Figs. 3–5 and S2–S4,
which we summarize in the following:
The correlograms exhibited quite different patterns – some correlograms showed that a wide range
of band combinations was able to explain the simulated quantities (e.g. GPP
at Amplero; Figs. 3 and S2), while some correlograms exhibited very
pronounced patterns, with the R2 value changing greatly with subtle
changes in band combinations (e.g. ε at Neustift; Figs. 3 and
S2).
Maximum R2 values were often clearly higher than the surrounding areas of high predictive power
(e.g. ε at Amplero; Fig. 3).
The different types of indices (compare Figs. 3–5) yielded similarly high correlations with the same
dependent variable at the same site in similar spectral regions. This
indicates that band selection is more important for explanatory power than
the mathematical formulation of the VI (i.e. ratio vs. difference,
with/without normalization). SR and NSD indices (Figs. 3 and 4) yielded
similar results compared to SD indices (Fig. 5).
The highest correlations for all dependent variables were found either for indices combining bands in the
visible range (VIS: < 700 nm) or the red edge and NIR (NIR:
> 700 nm), corresponding to spectral regions used by indices such as the
SRPI, NPCI, PRI and NPQI, and the CI and WI, respectively. Spectral regions
of well-known indices, such as NDVI, SR, SIPI or GRI, which exploit the
contrasting reflectance magnitudes in the visible and NIR (Fig. S1), resulted
in comparably lower correlations.
For midday and daily time resolutions different band combinations were selected (e.g. NEE at Amplero; compare
Figs. 4 and S3). For similar selected regions, daily averages were
characterized by higher explanatory power compared to midday averages (e.g.
ε at Neustift).
Figure 6 shows the performance of linear regression models for the best
NSD-type indices for midday ecophysiological parameters for each site and all
sites pooled (Fig. S7 shows the results of the same analysis for daily
averages). Large differences existed between the study sites regarding the
explanatory power of the same index for the same dependent variable. The
highest R2cv and values were generally obtained for Amplero, followed by
Neustift and then Monte Bondone, and the lowest R2cv values resulted when
data from all three sites were pooled, confirming the difficulties in finding
a general relation valid among sites.
Results of statistic of linear regression models between VIs and
ecophysiological parameters: α,ε (midday average) and
GPPmax. R2 – coefficient of determination; RMSE – root
mean square error. Bold numbers indicate the best-fitting model.
α
ε
GPPmax
Amplero
Neustift
Monte Bondone
All
Amplero
Neustift
Monte Bondone
All
Amplero
Neustift
Monte Bondone
All
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
–
µmolCO2µmolphot
SR
0.57
0.01
0.04
0.07
0.13
0.01
0.06
0.04
0.50
0.01
0.33
0.03
0.35
0.04
0.18
0.04
0.89
1.58
0.01
4.31
0.78
2.76
0.28
6.71
GRI
0.29
0.01
0.00
0.07
0.13
0.01
0.00
0.05
0.26
0.01
0.67
0.02
0.44
0.04
0.47
0.03
0.69
2.66
0.00
4.35
0.81
2.53
0.09
7.51
WI
0.50
0.01
0.01
0.07
0.08
0.01
0.03
0.04
0.41
0.01
0.22
0.03
0.36
0.04
0.25
0.04
0.86
1.82
0.16
3.99
0.54
3.95
0.24
6.87
NDVI
0.44
0.01
0.04
0.07
0.06
0.01
0.06
0.04
0.40
0.01
0.30
0.03
0.53
0.04
0.43
0.03
0.79
2.21
0.03
4.28
0.82
2.50
0.37
6.27
SIPI
0.37
0.01
0.07
0.07
0.02
0.01
0.18
0.04
0.35
0.01
0.29
0.03
0.64
0.03
0.44
0.03
0.66
2.80
0.06
4.21
0.74
2.96
0.47
5.74
CI
0.49
0.01
0.00
0.07
0.09
0.01
0.01
0.05
0.41
0.01
0.65
0.02
0.43
0.04
0.34
0.04
0.81
2.08
0.01
4.34
0.80
2.62
0.16
7.24
PRI
0.71
0.01
0.02
0.07
0.02
0.01
0.14
0.04
0.50
0.01
0.19
0.03
0.28
0.05
0.40
0.04
0.41
3.68
0.26
3.75
0.11
5.50
0.14
7.33
EVI
0.47
0.01
0.03
0.07
0.03
0.01
0.14
0.04
0.46
0.01
0.43
0.03
0.53
0.04
0.38
0.04
0.78
2.25
0.01
4.33
0.70
3.21
0.32
6.50
NPQI
0.06
0.01
0.06
0.07
0.05
0.01
0.31
0.04
0.04
0.01
0.30
0.03
0.17
0.05
0.11
0.04
0.00
4.78
0.07
4.20
0.21
5.17
0.14
7.31
NPCI
0.50
0.01
0.07
0.07
0.03
0.01
0.37
0.04
0.51
0.01
0.17
0.03
0.00
0.05
0.00
0.05
0.53
3.28
0.17
3.97
0.17
5.33
0.32
6.52
SRPI
0.51
0.01
0.06
0.07
0.03
0.01
0.36
0.04
0.56
0.01
0.15
0.04
0.00
0.05
0.00
0.05
0.50
3.38
0.17
3.97
0.17
5.31
0.28
6.69
Red-edge NDVI
0.48
0.01
0.00
0.07
0.07
0.01
0.01
0.05
0.40
0.01
0.65
0.02
0.47
0.04
0.40
0.04
0.79
2.16
0.00
4.34
0.80
2.58
0.19
7.09
For Amplero and Neustift the NIR-vs.-NIR combinations showed a positive
linear regression model with α, GPPmax and GPP, while for
Monte Bondone a negative linear correlation was observed. For Amplero the VIS-vs.-VIS combination showed a good performance in predicting ε;
the NIR-vs.-NIR combinations showed good performance for Neustift, as did the VIS-vs.-NIR combination for Monte Bondone. The linear models for NEE were
site-specific. In fact, Amplero and Monte Bondone showed a positive linear
regression model for NEE, but the VIS-vs.-VIS band combination was selected
for Amplero and the NIR-vs.-NIR combination for Monte Bondone. Neustift performed
well with NEE for NIR-vs.-NIR combinations, but with an inverse relationship.
The different types of indices (compare Figs. 6, S5 and S6) resulted in
similar models. The different time resolutions gave different models (e.g.
GPP, ε and NEE at Monte Bondone; compare Figs. 6 and S7, Figs. S5 and S8, or Figs. S6 and S9).
Correlation between conventional VIs, ecophysiological variables and
CO2 fluxes
The correlation analysis between the conventional VIs, the midday CO2
fluxes (Table 3) and ecophysiological parameters (Table 4), generally
confirmed the results obtained with the hyperspectral data.
For the same dependent variable (α, GPPmax, GPP,
ε and NEE), the performance of the various VIs showed large
differences between sites. For example, for GPPmax all of the
investigated indices except NPQI resulted in significant linear correlations
at Amplero, explaining 41–89 % of the variability in GPPmax.
In contrast, only NDVI, PRI, NPCI and SRPI showed a slightly significant
linear performance (17–26 %) for GPPmax at Neustift.
The different VIs performed differently in predicting the same dependent
variable at the different study sites. For all dependent variables (Tables 3,
4 and S3), the VI resulting in the highest R2 values was never the same
at all sites. Often the best-fitting VI at one site resulted in a
non-significant correlation at another site. Therefore, none of the dependent
variables clearly emerged as the one best predicted (Tables 3, 4 and S3).
When data from all sites were pooled, models showed the same performance for
the same VI and dependent variable except for GPP and NEE. The best-performing VI for GPP and NEE was SIPI; NPCI performed best for α,
GRI for ε and SIPI for GPPmax.
The choice of the averaging period (midday vs. daily) applied to
ε, NEE and GPP generally did not modify the ranking of the VIs,
but the R2 values tended to be similar or somewhat higher on the daily
timescale (compare Tables 3 and 4 with Table S3).
Evaluation of the model performance
Figure 7 shows the results of the validation for each ecophysiological
parameter and midday averaged fluxes and NSD-, SR- and SD-type indices
against data from the validation sites. The models used in the validation are
based on the best models determined for each site (i.e. Amplero, Neustift and
Monte Bondone) and on pooling together the two alpine grasslands of Monte
Bondone and Neustift (referred to as M.Bondone&Neustift).
Overall, the results of the validation were mixed. Good performance was
observed mainly for the Neustift, Monte Bondone and pooled
M.Bondone&Neustift models (see Table S4). In particular, the best
performance values were obtained for (i) α for the pooled Monte
Bondone and Neustift model for SR-type indices; (ii) GPPmax for
NSD-, SR- and SD-type indices and models except for Amplero; and (iii) for
midday GPP for all NSD-, SR- and SD-type indices and models. It is
interesting to note that lower performances were generally found for the
models based on the Amplero parameterization. This is understandable as Monte
Bondone and in particular Neustift were structurally and functionally much
more similar to the validation sites compared to Amplero (Tables 1 and S2).
Considerably poorer performance was observed for ε and NEE
across all model–index type combinations (Table S4). The validation on a daily
timescale always resulted in a poorer performance compared to the midday
average timescale (Fig. S10 and Table S5).
Effects of canopy structure on selected band combinations
Tables 5 and S6 show the results of the correlation analysis between the
selected NSD-, SR- and SD-type indices for ecophysiological variables and
fluxes and biophysical properties of vegetation, such as dry phytomass,
nitrogen and water content. Overall, the spectral response in the selected
band combinations for NSD-, SR- and SD-type indices was strongly related to
vegetation properties of the three grasslands (e.g. nitrogen and dry
phytomass), which impacted on their spectral response in the NIR and VIS
regions. For the Mediterranean site (Amplero) and for all ecophysiological
parameters (e.g. α, GPPmax), dry phytomass was the main
driving factor of the spectral response in the selected bands, while nitrogen
content drove the spectral response in the NIR region for Neustift. For Monte
Bondone, both dry phytomass and nitrogen content affected the spectral
response of the grassland. Similar results were obtained for SR- and SD-type
indices.
Results of statistic of linear regression models between VIs and
midday average CO2 fluxes: NEE and GPP. R2 – coefficient of
determination; RMSE – root mean square error. Bold numbers indicate the
best-fitting model.
GPP
NEE
Amplero
Neustift
Monte Bondone
All
Amplero
Neustift
Monte Bondone
All
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
–
µmolCO2m2s
–
µmolCO2m2s
–
µmolCO2m2s
–
µmolCO2m2s
–
µmolCO2m2s
–
µmolCO2m2s
–
µmolCO2m2s
–
µmolCO2m2s
SR
0.86
1.59
0.08
4.56
0.75
3.12
0.27
7.09
0.36
2.76
0.08
4.77
0.68
3.19
0.18
6.35
GRI
0.85
1.67
0.01
4.44
0.80
2.78
0.10
7.85
0.54
2.32
0.01
4.96
0.68
3.21
0.08
6.73
WI
0.92
1.23
0.05
3.25
0.50
4.41
0.24
7.20
0.44
2.57
0.05
4.87
0.43
4.28
0.17
6.42
NDVI
0.82
1.79
0.14
4.58
0.80
2.82
0.36
6.60
0.42
2.63
0.14
4.63
0.72
3.01
0.29
5.94
SIPI
0.65
2.50
0.08
4.57
0.72
3.32
0.46
6.08
0.33
2.82
0.08
4.79
0.65
3.34
0.39
5.51
CI
0.88
1.44
0.00
4.31
0.81
2.69
0.17
7.56
0.43
2.59
0.00
4.98
0.75
2.82
0.12
6.59
PRI
0.25
3.69
0.05
4.34
0.14
5.79
0.10
7.84
0.00
3.44
0.05
4.87
0.15
5.20
0.05
6.84
EVI
0.75
2.11
0.01
4.31
0.68
3.51
0.33
6.79
0.36
2.74
0.01
4.97
0.71
3.03
0.26
6.05
NPQI
0.04
4.17
0.08
4.27
0.14
5.78
0.16
7.57
0.24
2.99
0.08
4.78
0.19
5.08
0.12
6.60
NPCI
0.40
3.29
0.01
4.45
0.14
5.76
0.30
6.92
0.11
3.25
0.01
4.95
0.21
5.03
0.25
6.09
SRPI
0.35
3.42
0.01
4.44
0.15
5.74
0.27
7.08
0.08
3.30
0.01
4.95
0.22
5.01
0.22
6.19
Red-edge NDVI
0.87
1.51
0.00
4.35
0.81
2.68
0.20
7.40
0.43
2.60
0.00
4.98
0.75
2.84
0.15
6.47
Figures 8 and S11 show the correlation analysis between the selected NSD-,
SR- and SD-type indices for all ecophysiological parameters (e.g. α,
GPPmax) and chlorophyll content for Monte Bondone in 2013. The chlorophyll
content showed a very good correlation for all selected models and for all
indices. The values of R2 were always higher than the values of R2
obtained for the other biophysical variables (Tables 5 and S6). In Fig. 9, it
is possible to see that NSD- and SR-type indices for the selected bands for
estimating GPP (i.e. 996 and 710 nm) are strongly correlated with canopy-total chlorophyll content (R2>0.80).
Band selection using the GA–rF method
Figure 9 shows the results of the band selection based on the GA–rF method. In
particular, each plot represents the frequency of the occurrence of each
band in the genetic algorithm.
Results of the correlation (R2 – coefficient of determination)
between the best NDS-, SR- and SD-type indices selected for α,
GPPmax, midday GPP, midday ε and midday NEE and dry
phytomass, nitrogen and water content for Amplero, Neustift, Monte Bondone
and all sites pooled. The selected bands for computing NSD-, SR- and SD-type
indices are reported in brackets. Statistical significance is indicated as
*
(p<0.05), ** (p<0.01) and *** (p<0.001).
α
GPPmax
GPP
ε
NEE
Index
Site
Parameter
R2
R2
R2
R2
R2
(–)
(–)
(–)
(–)
(–)
NSD-type
Amplero
Dry phytomass (g m-2)
0.66**
0.72**
0.58*
0.76**
0.36
Amplero
Nitrogen content (%)
0.30
0.32
0.19
0.49*
0.15
Amplero
Water content (%)
0.28
0.53*
0.56*
0.44
0.55*
Neustift
Dry phytomass (g m-2)
0.00
0.26
0.35
0.44*
0.02
Neustift
Nitrogen content (%)
0.16
0.15
0.21
0.77**
0.03
Neustift
Water content (%)
0.00
0.00
0.03
0.59*
0.10
Monte Bondone
Dry phytomass (g m-2)
0.02
0.59***
0.49***
0.55***
0.49***
Monte Bondone
Nitrogen content (%)
0.08
0.52***
0.38**
0.48***
0.38**
Monte Bondone
Water content (%)
0.09
0.48***
0.35**
0.42***
0.35**
All
Dry phytomass (g m-2)
0.05
0.02
0.02
0. 05
0.00
All
Nitrogen content (%)
0.26***
0.04
0.02
0.41***
0.09
All
Water content (%)
0.00
0.01
0.01
0.10*
0.00
SR-type
Amplero
Dry phytomass (g m-2)
0.66**
0.72**
0.58*
0.76*
0.36
Amplero
Nitrogen content (%)
0.30
0.32
0.20
0.49*
0.15
Amplero
Water content (%)
0.28
0.53*
0.56*
0.44
0.55*
Neustift
Dry phytomass (g m-2)
0.00
0.26
0.35
0.44*
0.02
Neustift
Nitrogen content (%)
0.16
0.15
0.21
0.77**
0.03
Neustift
Water content (%)
0.01
0.00
0.03
0.59*
0.10
Monte Bondone
Dry phytomass (g m-2)
0.02
0.50***
0.50***
0.55***
0.45*
Monte Bondone
Nitrogen content (%)
0.08
0.48***
0.38**
0.48***
0.35
Monte Bondone
Water content (%)
0.09
0.44***
0.34**
0.41***
0.30*
All
Dry phytomass (g m-2)
0.08
0.01
0.01
0.05
0.00
All
Nitrogen content (%)
0.26***
0.03
0.02
0.40***
0.09
All
Water content (%)
0.00
0.01
0.01
0.11*
0.00
SD-type
Amplero
Dry phytomass (g m-2)
0.64***
0.81**
0.59*
0.58*
0.25
Amplero
Nitrogen content (%)
0.22
0.30
0.22
0.30
0.03
Amplero
Water content (%)
0.16
0.45*
0.59*
0.19
0.49*
Neustift
Dry phytomass (g m-2)
0.21
0.04
0.37
0.20
0.00
Neustift
Nitrogen content (%)
0.11
0.01
0.12
0.81**
0.08
Neustift
Water content (%)
0.02
0.26
0.00
0.64*
0.52*
Monte Bondone
Dry phytomass (g m-2)
0.15
0.42***
0.36**
0.45***
0.36**
Monte Bondone
Nitrogen content (%)
0.28***
0.34**
0.34**
0.38**
0.35**
Monte Bondone
Water content (%)
0.27***
0.34**
0.34
0.31**
0.30**
All
Dry phytomass (g m-2)
0.20***
0.01
0.00
0.30***
0.01
All
Nitrogen content (%)
0.01
0.01
0.02
0.02
0.11*
All
Water content (%)
0.01
0.03
0.04
0.28***
0.06
Overall, using the GA–rF method it was possible to identify portions of the
spectrum that were of particular significance for estimating specific
properties of the different ecosystems. For example, for predicting midday
GPP (Fig. 9b) for all sites pooled together, the bands at 430, 630, 660 and
710 nm showed the best results. The bands at 505 nm, 660 nm and 710 nm played an important
role in predicting midday GPP for Amplero, Neustift and
Monte Bondone, respectively. Some differences were found for the
different time resolutions (compare Fig. 9b and c). For example, the bands at
580 and 800 nm showed the best results for Amplero, and bands at 530 nm showed the best results for
Neustift.
Figure 10 shows the results for the band selection by GA–rF methods for
biophysical variables (i.e. dry phytomass, nitrogen and water content). For
the variables related to slow processes, the GA–rF method highlighted
different bands for different sites; a much higher between-site variability
for the variables related to ecophysiological processes (e.g. ε,
α and GPPmax) was detected, and we weren't able to
identify common “hot spots”.
Discussion
This study aimed at evaluating the potential of hyperspectral reflectance
measurements to simulate CO2 fluxes and ecophysiological variables of
European mountain grasslands over a range of climatic conditions and
management practices (grazing, harvest). To this end, we combined eddy
covariance CO2 flux measurements with ground-based hyperspectral
measurements at six mountain grassland sites in Europe.
Results of the GA–rF method for band selection for Amplero,
Neustift, Monte Bondone and all sites pooled for dry phytomass, water and
nitrogen content.
Upscaling of in situ relationships between VI indices and CO2 fluxes and ecophysiological parameters
Despite the fact that we focused on a single type of ecosystem, our results
showed that large differences existed among the investigated sites in the
relationships between hyperspectral reflectance data and CO2 fluxes and
ecophysiological parameters. For all study sites pooled, hyperspectral
reflectance data explained 40–68 % of the variability in the dependent
variables (Figs. 3–5). The conventional VIs yielded a maximum of 47 % of
explained variability in the data (Tables 3–4).
This is the first study comparing different grasslands characterized by
different plant species and environmental conditions. The use of simple
models based on a linear relationship between GPP and VIs, related to canopy
greenness, has proven to be a good proxy for the GPP of ecosystems with strong
green-up and senescence (Peng et al., 2011; Rossini et al., 2012). The loss
of this relationship may be related to low ε variability due to
abiotic and biotic stressors, the dependency of PRI on LAI, leaf and canopy
biochemical structure (e.g. leaf orientation), and xanthophyll cycle
inhibition or saturation, and zeaxanthin-independent quenching (Gamon et al.,
2001; Filella et al., 2004; Rahimzadeh-Bajgiran et al., 2012; Hmimina et al.,
2014). For alpine grasslands, a key meteorological variable that played a
relevant role in stimulating ε was high soil water content
associated with low temperatures (Polley et al., 2011). Low soil water
contents also triggered a decrease in leaf conductance as well as in ε and in α for two oak and beech ecosystems (Hmimina et al.,
2014). However, no significant differences in leaf biochemical and structural
properties of the canopy at the lowest and highest water content were found. In
addition, in this special issue, Sakowska et al. (2014) showed that
ε is also strongly affected by the directional distribution of
incident PAR, i.e. the ratio of direct to diffuse PAR.
Considering all sites pooled together (Figs. 3 and S2), NSD-type indices
showed a very poor correlation in the VIS-vs.-NIR band combinations (i.e.
traditional greenness indices; see Table 2) with GPP. It is well-known in
the literature (Rossini et al., 2010, 2012; Peng et al., 2010; Sakowska et
al., 2014) that greenness indices, for grasslands and crops, are often
good proxies of fAPARgreen (and thus carbon fluxes). Interestingly, in our
study their performance was considerably poorer than expected. The NSD-type
index showed a better performance in VIS-vs.-VIS band combinations than VIS-vs.-NIR ones. VIS-vs.-VIS band combinations for NSD-type indices (e.g. green
vs. blue or red and green vs. green wavelengths; see, e.g., Inoue et al., 2008)
are defined as greenness indices (Fig. 1), although their performance is
generally much poorer than NSD VIS-vs.-NIR indices. These results are likely
due to the confounding effects of the different canopy structures and
consequently of the different NIR response of the investigated grasslands
(see Fig. S1). In fact, the different grassland structures (spatial
distribution of photosynthetic and also non-photosynthetic material, leaf
angles, etc.) affect our ability to use traditional indices to estimate
fAPARgreen (and fluxes) when we consider different grasslands together
because the structural effects on scattering are very complex in the NIR
(Jacquemoud et al., 2009; Knyazikhin et al., 2012). These results are of
importance for the community, which still relies on these
relationships a lot, also favoured by the availability of affordable narrow-band
sensors that allow continuous monitoring of, e.g., NDVI. These results suggest
that waveband combinations not exploited by presently used (conventional) VIs
may offer considerable potential for predicting grassland CO2 fluxes;
this has implications for the design and capabilities of future
space, airborne or ground-based low-cost sensors. In particular, these results
also have a strong impact on our ability to upscale grassland fAPARgreen and
carbon fluxes using future sensors (e.g. Sentinel 2).
The evaluation of the models for the main data set against three new
sites showed that at least some of these models can be transferred to predict
carbon fluxes and ecophysiological parameters for similar grasslands
(Fig. 7). However, for some parameters (e.g. ε), the independent
validation indicated a poor performance; it challenges the current practice
in upscaling to larger regions by grouping all grasslands into a single
plant functional type (PFT). We advocate more studies to be conducted merging
CO2 flux with hyperspectral data by means of models which use a more
process-oriented and coupled approach to simulating canopy CO2 exchange
and reflectance in order to explore the causes underlying the observed
differences between seemingly closely related study sites.
Grassland structural characteristics and their spectral response
Although we considered similar ecosystems (belonging to the same vegetation
type), the investigated canopies were very different and included
Mediterranean, extensive alpine and intensive alpine grasslands with very
different canopy structures in terms of leaf orientation, amount and spatial
distribution of green and non-photosynthetic components, leaf nitrogen, and
water content, as detailed in Vescovo et al. (2012).
For Amplero and Neustift, NSD-type indices performed well for NIR-vs.-NIR band
combinations for all investigated parameters, while Monte Bondone showed best
performances in the VIS-vs.-NIR band combinations for GPPmax and
ε (Fig. 6). The dry phytomass was the main driving factor in the
spectral response in NIR-vs.-NIR band combinations for Amplero, while
nitrogen content drove the spectral response in NIR-vs.-NIR band combinations
of Neustift for all parameters except for α (Tables 5 and S6).
Interestingly, for Monte Bondone both dry phytomass and nitrogen content
explained the spectral response of the grassland in VIS-vs.-NIR band
combinations for GPPmax and ε, while no significant
relationships with biophysical variables were found for α, GPP or NEE.
These results partially confirm the findings of Vescovo et al. (2012), who
highlighted a strong relationship, for several grassland types, between an
NSD-type index and phytomass.
For Monte Bondone, NSD- and SR-type indices for the selected bands for
estimating all variables except α were strongly correlated with
canopy-total chlorophyll content (R2>0.85).
The chlorophyll indices (e.g. red-edge NDVI and CI; see Tables 3 and 4) –
which are considered the best indices for estimating carbon fluxes in
grasslands and crops – showed a good performance for Amplero
and Monte Bondone in our data set but performed poorly for Neustift.
It was demonstrated by many authors that the red-edge domain, where
reflectance changes from very low in the absorption region to high in the
NIR, is one of the best descriptors of chlorophyll concentration. On the
other hand, it is well-known that the canopy structure can be a very strong
confounding factor. Our results confirm that this topic needs to be further
investigated, as this finding has a relevant impact concerning the use of
Sentinel 2 to upscale fAPAR and carbon flux observations.
It is interesting to see that the NSD-type indices in the NIR-vs.-NIR band
combinations appeared to be the best proxy for GPP fluxes when all the
grasslands were pooled together. These results can be linked to the
controversial paper focused on the strong impact of structure on the ability
to estimate canopy nitrogen content (Knyazikhin et al., 2012) and confirm the
need for more studies in this direction. Good relationships were found
between the NIR-vs.-NIR band combinations (> 750 nm wavelengths) and
fluxes; the physical basis of these relationships needs to be further
investigated. In fact, it is important to highlight that the literature
indicates that the wavelengths in the NIR (> 750 nm) are not sensitive to
chlorophyll content, but they are related to leaf, canopy structure and
– around the 970 nm area – to water.
As confirmed by comparing the correlation matrix approach with the GA–rF
approach, we could not find a universal relationship between reflectance in
specific wavelengths of the light spectrum and biophysical properties of
vegetation. We think that this is strongly linked to vegetation structure
effects. For this reason we believe that further research is necessary to
disentangle the impact of factors such as bidirectional reflectance distribution function
and scaling effects.