Re

MSTa

b

c

d

e

Nf

a

ARA

KGVPGEPr

1

bcpeeprvestTs

0h

International Journal of Applied Earth Observation and Geoinformation 29 (2014) 1–10

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation andGeoinformation

jo ur nal home page: www.elsev ier .com/ locate / jag

emote estimation of grassland gross primary production duringxtreme meteorological seasons

icol Rossinia,∗, Mirco Migliavaccab, Marta Galvagnoc, Michele Meronid,ergio Cogliati a, Edoardo Cremonesec, Francesco Favaa, Anatoly Gitelsone,ommaso Julittaa, Umberto Morra di Cellac, Consolata Siniscalco f, Roberto Colomboa

Remote Sensing of Environmental Dynamics Laboratory, DISAT, University of Milano-Bicocca, Piazza della Scienza 1, 20126 Milano, ItalyMax Planck Institute for Biogeochemistry, Biogeochemical Data Integration Department, Hans Knoll Str 10, Jena, GermanyAgenzia Regionale per la Protezione dell’Ambiente della Valle d’Aosta, Sez. Agenti Fisici, Aosta, ItalyEuropean Commission, DG-JRC, Institute for Environment and Sustainability, Monitoring Agricultural Resources Unit, Ispra, VA, ItalyCenter for Advanced Land Management Information Technologies, School of Natural Resources, University of Nebraska-Lincoln, Lincoln,E 68588-0973, USADepartment of Life Science and Systems Biology, University of Torino, Torino, Italy

r t i c l e i n f o

rticle history:eceived 19 July 2013ccepted 19 December 2013

eywords:ross primary productionegetation indexRI

a b s t r a c t

Different models driven by remotely sensed vegetation indexes (VIs) and incident photosyntheticallyactive radiation (PAR) were developed to estimate gross primary production (GPP) in a subalpine grass-land equipped with an eddy covariance flux tower. Hyperspectral reflectance was collected using anautomatic system designed for high temporal frequency acquisitions for three consecutive years, includ-ing one (2011) characterized by a strong reduction of the carbon sequestration rate during the vegetativeseason. Models based on remotely sensed and meteorological data were used to estimate GPP, and across-validation approach was used to compare the predictive capabilities of different model formula-

rasslandxtreme eventsotential photosynthetically activeadiation

tions. Vegetation indexes designed to be more sensitive to chlorophyll content explained most of thevariability in GPP in the ecosystem investigated, characterized by a strong seasonal dynamic. Modelperformances improved when including also PARpotential defined as the maximal value of incident PARunder clear sky conditions in model formulations. Best performing models are based entirely on remotelysensed data. This finding could contribute to the development of methods for quantifying the temporal

broa

variation of GPP also on a

. Introduction

Terrestrial plants play important roles in the global car-on sequestration as they fix carbon dioxide (CO2) as organicompounds through photosynthesis. Terrestrial gross primaryroduction (GPP), the total amount of CO2 fixed by terrestrialcosystems, is one of the determinants of land-atmosphere CO2xchange (Beer et al., 2010). Thus, accurate GPP estimates canrovide valuable information for global carbon studies. Among ter-estrial ecosystems, mountain grasslands show a high interannualariability of their productivity (Yi et al., 2012). This variability isnhanced by grassland vulnerability to extreme climate events,uch as unusual spring warm temperature which can contribute

o early snowmelt in mountain environments (Beniston, 2005).he melting of snow cover influences the start and length of theummer growing season, the amount and timing of nutrients and

∗ Corresponding author. Tel.: +39 0264482848; fax: +39 0264482895.E-mail address: micol.rossini@unimib.it (M. Rossini).

303-2434/$ – see front matter © 2013 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.jag.2013.12.008

der scale using current and future satellite sensors.© 2013 Elsevier B.V. All rights reserved.

water release from the snow pack and can contribute to changesin species composition (Wipf and Rixen, 2010). The expectedincrease of the occurrence of extreme events predicted by mod-els and confirmed by observational data (Easterling et al., 2000;Meehl and Tebaldi, 2004) could hamper our ability to quan-tify ecosystem production. Hence, the development of modelsable to estimate ecosystem carbon cycle related processes acrossyears characterized by markedly different meteorological condi-tions is necessary to increase our confidence of future modelpredictions.

Measurements of vegetation reflectance at eddy covariance (EC)sites have notably increased in recent years (Balzarolo et al., 2011)because remote sensing holds considerable potential for advancingour capabilities to estimate and monitor vegetation production atdifferent temporal and spatial scales. Several studies demonstratedthe effectiveness of empirical models driven by remote sensing

inputs to model GPP on different ecosystems (Peng et al., 2011;Wu et al., 2009), including grasslands (Rossini et al., 2012). Never-theless a general consensus about model formulations and inputvariables performing better, especially considering long term data

2 Earth

sa

ecvaocbtahcic(

PbbfAeunPatrds

cimbt

iibm

aogylcbe

ruami

2

2

aN

M. Rossini et al. / International Journal of Applied

eries characterized by high inter-annual variability, has still to bechieved.

Previous studies reported the ability to model GPP using veg-tation indexes (VIs) designed to be sensitive to chlorophyll (Chl)ontent. Chlorophyll content is a key variable for explaining GPPariability in vegetation characterized by strong seasonality, suchs crops (Gitelson et al., 2006b; Peng et al., 2011; Rossini et al., 2010)r grasslands (Rossini et al., 2012). This is not surprising since Chlontent is a main factor controlling the amount of light absorbedy green vegetation and also directly relates to the enhanced elec-ron transport activity, which governs light use efficiency (LUE). Theccuracy in GPP estimation may be improved taking into accountigh frequency changes in radiation conditions and light use effi-iency modulation, through the inclusion of incident photosynthet-cally active radiation (PARi) and surrogate of LUE, the photochemi-al reflectance index (PRI, Gamon et al., 1992) in model formulationPeng et al., 2011; Rossini et al., 2012; Sakamoto et al., 2011).

For the calibration of this kind of models an estimation ofARi is required. In the prospect of GPP monitoring from space-orne remote sensing sensors, an accurate proxy for PARi that cane measured remotely is needed. Shortwave radiation obtainedrom coarse scale meteorological data sets from the NASA Datassimilation Office was used as a substitute for PARi by Sakamotot al. (2011). However, these estimates of PARi have significantncertainties; the coefficient of variation was 23.6% and meanormalized bias was 13.9% (Sakamoto et al., 2011). AlternativelyARi can be indirectly estimated by radiative transfer modelingpproach (e.g., Liu et al., 2008) once the optical properties ofhe atmosphere have been retrieved. However, incorporating suchadiative transfer calculations into the practical generation of stan-ardized product providing regular observations of global PARi istill a challenging topic in remote sensing (Liang et al., 2006).

Gitelson et al. (2012) suggested using potential photosyntheti-ally active radiation (PARpotential), defined as the maximal value ofncident PAR under clear sky conditions, rather than PARi for esti-

ating GPP. Such approach has the advantage that PARpotential cane more easily estimated using radiative transfer models or look-upable-based algorithms (Lyapustin, 2003).

To our knowledge, such approach has been successfully appliedn soybean and maize using Landsat and daily MODIS satellitemages (Gitelson et al., 2012; Peng et al., 2013), but it has nevereen applied to simulate GPP using ground spectral reflectanceeasurements collected with daily resolution.This paper provides an evaluation of the robustness of an

pproach based solely on remotely sensed data to estimate GPPver grasslands. We monitored spectral reflectance throughout therowth period of a subalpine grassland during three consecutiveears (2009–2011), including a year (2011) characterized by theongest snow-free period on record (83 years) leading to changes inanopy structure and functioning with a strong reduction of the car-on sequestration rate during the carbon uptake period (Galvagnot al., 2013).

The specific objectives of this paper are (1) to evaluate theobustness of GPP prediction by remote sensing driven modelssing a three year dataset including one year with an exception-lly long growing season, (2) to assess the performance of the bestodel formulation, and (3) to test whether the use of PARpotential

mproves the accuracy in GPP estimation.

. Methods

.1. Site description and micrometeorological measurements

The study was conducted in an unmanaged grassland of the sub-lpine belt located in the North-Western Italian Alps (45◦50′40′′

, 7◦34′41′′ E, Torgnon, Aosta Valley) at 2160 m a.s.l. during three

Observation and Geoinformation 29 (2014) 1–10

growing seasons (2009–2011). The vegetation of the site is com-posed mainly by matgrass (Nardus stricta) and, secondarily, byTrifolium alpinum, Arnica montana, Poa alpina and Carex sem-pervirens. The area is classified as an intra-alpine region withsemi-continental climate with an annual mean temperature of3.1 ◦C and mean annual precipitation of about 880 mm (Mercalliand Berro, 2003). The snow-free period lasts generally from lateMay to early November. This site is approximately 9 ha and it isequipped with an EC flux tower which provides continuous mea-surements of net ecosystem CO2 exchange (NEE) from January2009. Detailed descriptions of the EC flux measurements, the pro-cedures used to partition the NEE to derive GPP and of the fluxfootprint are reported in Migliavacca et al. (2011) and Galvagnoet al. (2013). Along with EC fluxes, the main meteorological vari-ables (in particular PARi) are available with a time step of 30 min.

2.2. Radiometric measurements and spectral vegetation indexcomputation

Hyperspectral reflectance measurements were collected in theproximity of the EC tower using the HyperSpectral Irradiometer(HSI, Cogliati, 2011; Meroni et al., 2011), designed for unattendedhigh temporal frequency acquisitions. This instrument acquiredspectra in the visible and near-infrared region (400–1000 nm) witha full width at half maximum of 1 nm.

HSI employs a rotating arm equipped with a cosine-responseoptic to measure alternately the sky (sensor pointing zenith)and the target irradiance (sensor pointing nadir), allowing thecomputation of the Bi-Hemispherical Reflectance factor (BHR,Schaepman-Strub et al., 2006). With this configuration the 97% ofthe total signal comes from a cone limited by a zenith angle of80◦ and the maximum contribution to the measured signal comesfrom zenith angles around 45◦. This means that with an installationheight of 3.5 m above the investigated surface, the 97% of the totalsignal comes from a circular ground area of 40 m diameter (Meroniet al., 2011).

Hyperspectral reflectance measurements were acquired every5 min from sunrise to sunset during the snow-free season in 2009,2010 and 2011. The values of the VIs used in the following analysesare means of index values calculated for each reflectance spectrumcollected between 11:00 and 13:00 local solar time to minimizechanges in solar angle. For more details about spectral reflectancedata acquisition and preprocessing see Rossini et al. (2012).

As a first step for scaling the results of this study to currentand future Earth Observation systems, the VIs considered havebeen defined on the basis of MODIS (Moderate Resolution ImagingSpectrometer, NASA) and OLCI (Ocean and Land Colour Instru-ment on board of future Sentinel-3 satellite, ESA) spectral bands. Inparticular, we considered the following VIs: (i) the normalized dif-ference vegetation index (NDVI) and the photochemical reflectanceindex (PRI) computed from MODIS simulated data, (ii) the red-edgechlorophyll index (CIre), the green chlorophyll index (CIg), the OLCIterrestrial chlorophyll index (OTCI) and the normalized differencered-edge (NDRE), computed from OLCI simulated data (Clevers andGitelson, 2013). Their definitions using MODIS (M) and OLCI (O)spectral bands are provided in Table 1.

2.3. Data selection

Ground-observed PARi values were used to calculate PARpotentialas the maximal PARi in 8-day-PARi windows (Gitelson et al., 2012;Peng et al., 2013). The relationship between maximal PARi and

the day of the year (DOY) was fitted with polynomial functionsto compute PARpotential at daily step. The best-fit functions, asecond and a third order polynomial, were used to calculate mid-day and daily PARpotential respectively. Based on PARpotential, the

M. Rossini et al. / International Journal of Applied Earth Observation and Geoinformation 29 (2014) 1–10 3

Table 1Vegetation indexes evaluated in this study. M refers to the MODIS band, O refers to the OLCI band and R� refers to the reflectance factor at wavelength � in nm.

Index Formulation Formulation with satellite bands References

NDVIR858.5 − R645

R858.5 + R645

M2 − M1M2 + M1

Rouse et al. (1974)

PRIR531 − R551

R531 + R551

M11 − M12M11 + M12

Gamon et al. (1992)

CIreR779

R709− 1

O16O11

− 1 Gitelson et al. (2003, 2006a)

CIgR779

R560− 1

O16O6

− 1 Gitelson et al. (2003, 2006a)

OTCIR754 − R709 O12 − O11

O10Dash and Curran (2004)

O11 O11

fdfs

cbw

2

w(gt

(

(

NWs2deiG

�

w

a

R709 − R681 O11 −NDRE

R754 − R709

R754 + R709

O12 −O12 +

requency of occurrence of cloudy or hazy weather conditionsuring the three years was then computed as the percentage dif-erence (PARpotential − PARi)/PARpotential (hereafter �PAR) for eachampling date.

The proposed models were calibrated and validated using dataollected with �PAR lower than 60%. This threshold was indicatedy Peng et al. (2013) as the one for which the GPP estimation modelith PARpotential was more accurate than the models with PARi.

.4. Model development

The set of models proposed in Rossini et al. (2012)as used to estimate both the daily midday average GPP

GPPm, �mol CO2 m−2 s−1) and the daily cumulated GPP (GPPd,C m−2 d−1). The use of both PARi and PARpotential was tested inhe following model formulations:

(i) model 1 (MOD 1), linear relationship between GPP and a VIrelated to canopy Chl content

GPP = a0 VI + b0 (1)

(ii) model 2 (MOD 2), linear relationship between GPP and theproduct of a VI related to canopy Chl content and PAR

GPP = a1(VI × PAR) + b1 (2)

iii) model 3 (MOD 3), LUE-like model (Monteith, 1972, 1977)assuming constant light use efficiency (ε) and modeling fAPARas a linear function of a VI related to canopy Chl content

GPP = ε × (a2 VI + b2) × PAR = (a3 VI + b3) × PAR (3)

iv) model 4 (MOD 4), modeling ε and fAPAR as a linear function ofPRI and VI, respectively

GPP = (a4 PRI + b4) × (a5 VI + b5) × PAR (4)

Model coefficients (ax and bx) were estimated using the Gauss-ewton nonlinear least square optimization method (Bates andatts, 1988), implemented in the function optim() of the R

tandard package (R, version 2.6.2, R Development Core Team,011). Model accuracy was evaluated in terms of root mean squareeviation (RMSD) which represents the “mean” deviation of mod-led values (GPPmod) with respect to the observed ones (GPPobs),n the same units as the model variable under evaluation (GPPd orPPm) and the relative RMSD (rRMSD, %) computed as:

rRMSD =

√√√√1n

n∑ (GPPmodi − GPPobs

i )2

(GPPobs)2

× 100 (5)

i=1 i

here n is the number of samples.When using observed PAR, models were tested using the aver-

ge of the values measured between 11:00 and 13:00 local solar

Gitelson and Merzlyak (1994)

time for the midday analysis (�mol m−2 s−1) and the sums ofhalf-hourly measurements during the day (when global radiationexceeds 15 Watt m−2) for the daily analysis.

Out of sample performances were assessed following a train-ing/validation splitting approach: one year at a time was excludedusing the remaining subset as the training set and the excludedone as the validation set. The model was fitted against each train-ing set and the resulting parameterization was used to predict theCO2 uptake of the excluded year.

3. Results and discussion

3.1. Interannual variability of gross primary production

The seasonal variations of carbon fluxes exhibited similardynamics in 2009 and 2010 that are markedly different from 2011(Fig. 1). The year 2011 was characterized by the earliest snowmeltsince 1945 and the longest snow-free period on record (83 years)(Galvagno et al., 2013). An early snowmelt meant a lengthenedsnow-free period and consequently an expanded growing seasonfor the plants. The snow-free period varied from 159 and 160 daysin 2010 and 2009, respectively, to 234 days in 2011.

GPP increased sharply just after snowmelt in 2009 and 2010,reaching its maximum value in mid-July (18th–13th, DOY 199 and194). During the second part of the season the decrease rate of GPPwas slower compared to the increase one. On the contrary, in 2011GPP increased earlier in spring, but with a lower rate and showinga less pronounced asymmetry between increasing and decreasingphases. This resulted in a longer carbon uptake period character-ized by a reduced carbon sequestration rate, compared to previousyears, but a similar seasonal cumulative carbon uptake (Galvagnoet al., 2013).

GPPm appeared more affected than GPPd by the early snowmeltoccurred in 2011. The peak of GPPm in 2011 was 25% and 29% lowerthan those recorded in 2009 and 2010, respectively. The corre-spondent decrease in 2011 GPPd peak was 16.8 and 16.6% referredto 2009 and 2010, respectively. The snowmelt timing observed in2011 exposed plants acclimated to a narrow range of environmen-tal conditions typical of late spring/early summer, to the unusualweather conditions of the early spring. As a result, vegetation grownin 2011 manifested, throughout the whole summer, a reducedresponsiveness to high irradiance, typical of central hours of theday in this site. As GPPd integrates periods of high and low PARi, thereduced responsiveness of GPP to high PARi is likely to be maskedwhen considering GPPd (Turner et al., 2003).

3.2. Time courses of spectral vegetation indexes

The seasonal variations in midday VIs are shown in Fig. 2. AllVIs exhibited a clear seasonal trajectory, increasing from the endof May (DOY 160) due to canopy greening, reaching a peak in July

4 M. Rossini et al. / International Journal of Applied Earth Observation and Geoinformation 29 (2014) 1–10

F hite s(

(soeorc

gatoH2imoe(rtbw

ocir

Vrisw

3

2f

tac

ig. 1. Seasonal variation of 2009 (gray circles), 2010 (black triangles) and 2011 (wb, d, f) daily Gross Primary Production (GPPd , gC m−2 d−1). DOY is day of the year.

DOY 220) and decreasing more or less steeply afterwards withenescence and canopy yellowing. Most VIs peaked in the middlef July, except PRI which peaked about two weeks later. A possiblexplanation of this delay is that PRI uses bands in the green regionf the spectrum where Chl absorption is weaker compared to theed region and the tendency toward saturation for medium to highanopy greenness is reduced.

Comparing the three years analyzed in this study, the onset ofreening was anticipated in 2011 due to the earlier snowmelt andll VIs started to increase (decrease) about 15 days earlier thanhe previous years. For example, NDVI level of 0.5 was attainedn DOY 166, 164 and 151 in 2009, 2010 and 2011, respectively.owever, the slope of the VIs in the green-up phase was lower in011 and VIs peaked in the same period across the three years,

ndicating a slower green-up period with advanced snowmelt. Theaximum values reached by the VIs in 2011 were lower than those

f previous years; this reduction is more evident when consid-ring VIs computed using red-edge bands and in particular OTCImaximum values of 2.84, 2.79 and 2.55 in 2009, 2010 and 2011,espectively). NDVI showed the most similar trends between thehree years because of its saturation at medium to high canopyiomass (Huete et al., 2002). Conversely, the minimum PRI valuesere slightly lower in 2011.

The differences in 2011 VIs can be attributed to the responsef the ecosystem to an exceptionally early snowmelt followed byold temperature typical of springtime. This event caused a changen the typical trajectory of canopy development and physiologicalesponses (Galvagno et al., 2013) visible also in the VI trajectories.

The examination of the seasonal time courses of the differentIs with respect to GPP showed that there was a good general cor-espondence between GPP and VIs tracking canopy greenness andn particular canopy Chl content: both GPP and VIs showed a moreymmetric behavior between increasing and decreasing phases asell as lower peak values in 2011 compared to previous years.

.3. Analysis of the effect of the illumination conditions

Fig. 3 shows higher occurrences of “not clear sky” conditions in010, with 15% of data collected with �PAR greater than 60%. Thisrequency is higher than those recorded in 2009 (7%) and 2011 (8%).

The effect of including cloudy or hazy weather observations inhe analysis was evaluated plotting the rRMSD of the models aimedt GPP estimation versus the maximum value of �PAR used as ariterion to select the observations composing both the calibration

quares): (a, c, e) midday Gross Primary Production (GPPm , �mol CO2 m−2 s−1) and

and the validation datasets. In this paragraph results obtained forGPPd estimation in 2010 are presented as an example, as 2010 wasthe year characterized by higher values of �PAR. Fig. 4 shows that,as expected, the accuracy of all the proposed models increasedas �PAR decreased. Higher improvements in model performanceswere observed between the rRMSD calculated using the completedataset and the rRMSD computed using data collected with �PARbelow 60%. This 60% threshold was selected as further reduction ofthis threshold did not result in significant performance increases.The same threshold was obtained by Peng et al. (2013) for GPPmodeling in soybean and maize. Relative RMSD for models drivenby PARpotential varied between 36% and 84% using the completedataset and between 16% and 52% when using observations with�PAR < 60%. Similar improvements were observed applying the�PAR threshold on PARi: the rRMSD range varied from 29–83% to16–52%. Models using PARpotential had smaller rRMSD than modelsusing PARi when �PAR was below 80%.

The analysis on midday values showed smaller improvementswhen considering lower �PAR compared to the analysis on dailyvalues (data not shown). We hypothesize that GPPd estimation canbe more affected by variations in �PAR because daily GPP andPAR result from the integration of half-hourly measurements dur-ing the day. A high correlation between midday fluxes and dailyfluxes (r = 0.95) enabled the daily analysis to work well when usingmidday remote sensing acquisitions. However, these relationshipscan be lowered when observations collected with highly variableweather conditions during the day are included because they maycause a decrease in the correlation between daily and middayfluxes, and consequently between daily fluxes and midday VIs. Thisis most likely to happen during hazy or cloudy days that can be fil-tered out based on �PAR thresholds. The proposed models wereconsequently calibrated and validated using only data collectedwith �PAR lower than 60%.

3.4. GPP estimation

The performances of different models for GPPd and GPPm esti-mation are shown in terms of cross validated rRMSD (Fig. 5).

Considering the analysis of daily values, some general consid-erations can be drawn: (i) the VIs based on the red-edge region

performed better than NDVI and CIg; (ii) GPP estimation accuraciesincreased using models accounting for variations in the incidentPAR (both PARi and PARpotential) and including the PRI in modelformulation (MOD 4); (iii) the use of PARpotential instead of PARi

M. Rossini et al. / International Journal of Applied Earth Observation and Geoinformation 29 (2014) 1–10 5

Fig. 2. Seasonal temporal profiles of measured vegetation indexes in 2009 (circles, left panels), 2010 (triangles, middle panels) and 2011 (squares, right panels): (a, g, m)n g; (c, iN tances

it(beboGt

ormalized difference vegetation index, NDVI; (b, h, n) green chlorophyll index, CIDRE; (e, k, q) OLCI terrestrial chlorophyll index, OTCI; (f, l, r) photochemical reflec

olar time). DOY is day of the year.

mproved GPP estimation. The best performing model formula-ions allowed GPPd estimation with a RMSD of 0.50 gC m−2 d−1

corresponding to a rRMSD of about 15%). For most of the models,etter performances were achieved using midday values. The low-st rRMSD values were confirmed at about 15% with MOD 4 driven

y VIs based on red-edge wavelengths (e.g. CIre). Comparing resultsbtained in the three years analyzed, the lowest performances inPPm estimation were observed in 2011. A specific discussion about

he possible reasons of this is reported in the next paragraph.

, o) red-edge chlorophyll index, CIred-edge; (d, j, p) normalized difference red-edge, index, PRI. Each point indicates the average value between 11:00 and 13:00 (local

Fig. 5(d, h, l, p) shows the rRMSD between predicted andobserved values for GPPm and GPPd considering the three yearstogether. The best performing model for GPPm estimation is MOD2 driven by the product of CIre and PARpotential and MOD4 driven byCIre, PARpotential and PRI for GPPd estimation. This means that the

use of a VI related to Chl content in combination with PARpotentialalone or with PRI allowed to account for both the seasonal changein Chl and the modulation of GPP due to changes in radiation con-ditions and light use efficiency.

6 M. Rossini et al. / International Journal of Applied Earth

Fig. 3. Distribution of the sampling days used in the analysis in 5 classes of(PARpotential − PARi)/PARpotential: from �PAR between 0% and 20% (clear sky condi-tb

dtcuicem

NoCPtmeusso

the extent that ε is correlated with pigment content. Such a rela-

Fo

ions) to �PAR between 80% and 100% (overcast conditions). The percentage of dayselonging to each class is plotted for each year.

Models driven by PARpotential performed better than thoseriven by PARi. This indicates that the use of PARpotential improvedhe accuracy in GPP estimation when a decrease in PARi may notorrespond to a decrease in GPP due to the tendency toward sat-ration of the relationship between GPP and PARi. Thus, at high

rradiance loads slightly PARi fluctuations due to unstable weatheronditions may cause noise and unpredictable uncertainties in GPPstimation. As our results show, the use of PARpotential in the modelay partially reduce such noise of the model.Vegetation indexes based on the red-edge region (OTCI, CIre and

DRE) and designed to be sensitive to Chl content explained mostf the variability in GPP and performed better than NDVI and CIg.Ire was the best performing VI when used together with PRI andAR (MOD 4) or PAR alone (MOD 2 and MOD3) because togetherhey account for both the seasonal change in Chl content and the

odulation of GPP due to changes in radiation conditions (Pengt al., 2011). OTCI was instead the best performing index whensed as single driver for GPP estimation. The use of OTCI as a

ingle variable to predict GPPd was already proven successful inites dominated by a predictable seasonal cycle such as decidu-us forests and grasslands (Harris and Dash, 2010). Further works

ig. 4. Relative root mean squared difference (rRMSD) of different GPPd estimation modbtained for GPPd estimation in 2010 are shown.

Observation and Geoinformation 29 (2014) 1–10

are needed to evaluate the possibility to apply this approach alsoon vegetation characterized by weak seasonality (e.g. evergreenforests).

Although the NDVI is the most widely used spectral index atEuropean EC sites (87.5% of the sites according to Balzarolo et al.(2011)), it exhibits limitations in GPP estimation in the investigatedgrassland. Models driven by NDVI alone or NDVI and PAR underesti-mated GPP at the beginning of the growing season (green-up phase)while overestimated GPP from DOY 180 on (data not shown). Onepossible reason is that NDVI response saturates with high biomassamount (Fava et al., 2010). The inclusion of PRI in model formula-tions improves GPP estimation using NDVI. Thus, while NDVI canprovide useful information about the temporal evolution of ecosys-tem phenology (Hmimina et al., 2013), the potential of using NDVIfor GPP monitoring requires further investigations. Fig. 6 showsthe strong relationship between PRI and indexes tracking canopyChl content such as OTCI and CIre. This relationship showed pro-nounced hysteresis: for the same OTCI or CIre, sPRI in the green-upstage was much higher than in the senescence stage. The hystere-sis was more pronounced for OTCI compared to CIre. The reasonof this hysteresis can be explained considering that PRI is relatedto the carotenoid/Chl ratio (Panigada et al., 2009; Stylinski et al.,2002) since it measures the relative reflectance on either side ofthe green hump, one side being affected only by Chl absorbance andthe other by the coupled Chl and carotenoid absorbance (Sims andGamon, 2002). Carotenoid concentration increases together withChl concentration in the green-up period while during senescenceChl decreases sharper compared to carotenoids that are retained forphotoprotection (Merzlyak and Gitelson, 1995). Even if only fewauthors (Sims and Gamon, 2002; Stylinski et al., 2002) reportedsimilar relationships, PRI variation over weeks or months may bea combined function of the variation of ε and changes in the totalpools of pigments. This may help explaining why the use of PRIin model formulation increases the accuracy in GPP estimation,since varying carotenoid/Chl ratios can covary with xanthophyllpigment levels, and this may enhance the ability of PRI to predict ε.On the other hand, to the extent that pigment ratios are not closelyrelated to ε, changing pigment ratios would be a confounding vari-able. Thus, PRI may still be effective as a measure of changes in ε to

tionship appears likely in the present study but the implicationsof this relationship for the use of PRI in ε estimation needs to beinvestigated in a wider number of ecosystems.

els plotted versus the difference (PARpotential − PARi)/PARpotential (�PAR, %). Results

M. Rossini et al. / International Journal of Applied Earth Observation and Geoinformation 29 (2014) 1–10 7

F rved

u Ri , (m–

aitrww

ig. 5. Relative root mean square deviation (rRMSD, %) between predicted and obsesing PARi , (e–h) GPPd estimation using PARpotential , (i–l) GPPm estimation using PA

The modeling approach proposed here has some potential forpplication with current and future satellite data because all thenput variables can be in principle derived from satellite observa-

ions, including PARpotential that can be more easily computed withadiative transfer models as compared to PARi. However, whenorking with satellite instruments, the spatial and temporal detailould be necessarily reduced compared to the one available for

values using the 2009, 2010, 2011 and the total datasets for: (a–d) GPPd estimationp) GPPm estimation using PARpotential .

this study. As a result, GPP estimation uncertainty expected to besubstantially greater than that achieved in this study using groundmeasurements. Satellite application is expected to be particularly

challenging in ecosystems characterized by high spatial and tem-poral heterogeneity such as the one investigated here. In particular,one of the main limitations in mountain areas is that the ecosystemsize is often lower than the pixel size (from 250 m to 1 km for MODIS

8 M. Rossini et al. / International Journal of Applied Earth Observation and Geoinformation 29 (2014) 1–10

lotted

sbG

3l

ictethTab2T

Fmma

Fig. 6. Relationship between (a) PRI and OTCI and (b) PRI and CIre. PRI is p

ensor and 300 m for OLCI) causing different land cover classes toe included in the same pixel and decreasing the performances inPP estimation.

.5. Model performances in a year characterized by an unusuallyong growing season

The proposed method was tested to assess the ability in estimat-ng GPP for the year 2011 characterized by a markedly differentarbon sequestration dynamic during the vegetative season dueo an extremely long growing season caused by an exceptionallyarly snowmelt. The prediction accuracy was better for daily ratherhan midday GPP. For the analysis of daily values, better resultsave been obtained with MOD 4 driven by PARpotential, CIre and PRI.his model yielded a RMSD of 0.55 gC m−2 d−1, corresponding to

rRMSD of 15.8%. When analyzing midday values, MOD 1 driveny CIre predicted GPPm with the highest performances, rRMSD of2.1%, corresponding to an absolute RMSD of 1.9 �mol CO2 m−2 s−1.o get a better understanding of the capability of remote sensing

ig. 7. Time courses of 2011 (a) daily (gC m−2 d−1) and (b) midday GPP (�mol CO2 m−2 s−

odeled (black line) with the best performing model formulation (GPPmod): model 4 parodel 1 parameterized with CIre (rRMSD = 22.1%) for the midday analysis. Temporal chan

nd (d) midday (�GPPm) analysis. Temporal changes in �GPP normalized by GPPobs for t

as scaled PRI, computed as sPRI = (PRI + 1)/2, for clarity of representation.

driven models to represent the seasonal time courses of GPP, wecompared CO2 uptake observations and model outputs obtainedwith the best-performing model for the daily and midday values(Fig. 7).

Predicted GPPd values agreed quite well with observationsas concerns both amplitude and seasonal phase and successfullydescribed the dynamics captured by CO2 uptake fluxes. PredictedGPPm values agreed quite well with GPPobs as concerns the seasonalphase but showed overestimation in the middle of the growingseason (see Fig. 7d from DOY 174 to DOY 254).

The early start of season triggered by snow melting exposedplants to weather conditions harder than previous years. High alti-tude plants depend on highly efficient carbon assimilation sincetheir growing season is usually very short and characterized byextreme climatic conditions (Streb et al., 1998). The acclimation

of physiological processes is therefore essential to cope with suchclimatic conditions. The snowmelt observed in 2011 was the thirdearliest in the last 83 years and hence exposed the investigatedgrassland to less favorable weather occurring earlier in the season.

1) estimated from eddy covariance measurements (GPPobs) (full symbols) and GPPameterized with PRI, CIre and PARpotential (rRMSD = 15.8%) for the daily analysis andges in the difference between simulated and observed GPP for the (c) daily (�GPPd)he (e) daily (�GPPd/GPPd) and (f) midday (�GPPm/GPPm) analysis.

Earth

Ppslyadwhdpiwrdop

rntet

4

sguadatcctgvtss

tdoSts

cdmwereGd

oaTcs

M. Rossini et al. / International Journal of Applied

lants responded to such conditions through a reduction of thearameters describing the light response curve of photosynthe-is (Galvagno et al., 2013): this means that for a given irradianceevel the observed GPP was lower in 2011 compared to otherears. This effect is enhanced for high irradiance loads occurringt noon. As a consequence when the models were used to pre-ict GPPm, predicted GPPm was overestimated because the modelsere trained with the 2009 and 2010 datasets characterized byigher GPPm values for the same PARi. It is interesting to note that,espite the extraordinary conditions experienced in 2011, the pro-osed method allowed to accommodate for inter-annual changes

n the length of the growing season and predict GPPd and GPPm

ith an acceptable accuracy. Since future warming of the Alpineegion (Foppa and Seiz, 2012) will likely result in earlier snowmeltates and thus an increase of the frequency of events such thene occurred in 2011, the evaluation of the robustness of modelsroposed for GPP estimation in such conditions is crucial.

This study demonstrates the importance to collect spectraleflectance data continuously, regularly and from a worldwideetwork in connection with the well-established network of fluxowers (FLUXNET) to improve our ability to model changes incosystem production in response to interannual climatic varia-ions.

. Conclusions

This study showed that models entirely based on remote-ensing data can provide good predictions of GPP in high altituderassland ecosystems. Better results in GPPm have been obtainedsing the model driven by the product of CIre and PARpotential, whileccuracy in GPPd estimation have been obtained with the modelriven by CIre, PARpotential and PRI, confirming the highest perform-nces of VIs related to Chl content for GPP estimation. However,he use of PRI as estimate of ε over weeks or months should bearefully evaluated because at this temporal scale PRI may be aombined function of changes in ε and changes in chlorophyll con-ent as well as carotenoid/Chl ratio. So PRI is expected to provideood estimate of ε to the extent that ε is correlated with pigmentariations. Further studies are needed to explore the seasonal rela-ionship between PRI and ε in ecosystems characterized by weakeasonal variations where the correlation between pigment sea-onal variation and ε is likely to be lower.

After the application of a cloud screening criterion based onhe difference between PARi and PARpotential (�PAR < 60%), modelsriven by PAR in the form of PARpotential provided better estimatesf both GPPm and GPPd compared to the models driven by PARi.ince PARpotential is expected to be more easily available than PARi,he use of PARpotential can be preferred in GPP monitoring usingolely remotely sensed data.

The proposed technique was validated also during a year (2011)haracterized by a strong reduction of the carbon sequestration rateuring the vegetative season. Results showed accurate GPPd esti-ation with a rRMSD of 15.8%, while GPPm was estimated with aorse accuracy compared to the other years analyzed. This can be

xplained by the reduction of the parameters describing the lightesponse of photosynthesis occurred in 2011 due to the unusualarly snowmelt, affecting GPPm to a higher extent compared toPPd. As a consequence, models calibrated with the 2009 and 2010atasets overestimated the GPPm measured in 2011.

Long-term time series of hyperspectral reflectance at daily res-lution are available in a limited number of sites due to high costs

nd complexity in making the instruments running unattended.his makes difficult to test the proposed approach on similar dataollected in different vegetation types. However the approach pre-ented in this study is not limited to estimate GPP using spectral

Observation and Geoinformation 29 (2014) 1–10 9

reflectance collected by radiometers mounted on a platform closeto the canopy. It could be also applied to remotely sensed data col-lected at multiple scales from close range to satellite platforms,allowing regional or global GPP monitoring. However the largescale difference between medium to coarse resolution satellitedata and in situ local hyperspectral measurements can make thedirect extrapolation of this approach to wider area challengingwhen a single pixel contains multiple land cover classes (character-ized by different VIs-GPP relationships), as in mountain areas withhigh landscape fragmentation and heterogeneity. Consequently, itappears worthwhile to address further research to test this kind ofapproach in different vegetation types using satellite-derived VIsand PARpotential calculation using radiative transfer approaches.

Acknowledgements

This study was supported by the PhenoALP project, an Interregproject co-funded by the European Regional Development Fund,under the operational program for territorial cooperation Italy-France (ALCOTRA) 2007–2013. The authors acknowledge the staffof the Remote Sensing of Environmental Dynamics Laboratory(DISAT, UNIMIB) for the support during field campaigns.

References

Balzarolo, M., Anderson, K., Nichol, C., Rossini, M., Vescovo, L., Arriga, N., et al.,2011. Ground-based optical measurements at European flux sites: a review ofmethods, instruments and current controversies. Sensors 11, 7954–7981.

Bates, D.M., Watts, D.G., 1988. Nonlinear regression analysis and its applications.John Wiley & Sons, New York.

Beer, C., Reichstein, M., Tomelleri, E., Ciais, P., Jung, M., Carvalhais, N., et al., 2010.Terrestrial gross carbon dioxide uptake: global distribution and covariation withclimate. Science 329, 834–838.

Beniston, M., 2005. Mountain climates and climatic change: an overview of processesfocusing on the European Alps. Pure and Applied Geophysics 162, 1587–1606.

Clevers, J.G.P.W., Gitelson, A.A., 2013. Remote estimation of crop and grass chloro-phyll and nitrogen content using red-edge bands on Sentinel-2 and -3.International Journal of Applied Earth Observation and Geoinformation 23,344–351.

Cogliati, S., 2011. Development of automatic spectrometric systems for proximalsensing of photosynthetic activity of vegetation. University of Milano-Bicocca,pp. 162.

Dash, J., Curran, P.J., 2004. The MERIS terrestrial chlorophyll index. InternationalJournal of Remote Sensing 25, 5403–5413.

Easterling, D.R., Meehl, G.A., Parmesan, C., Changnon, S.A., Karl, T.R., Mearns, L.O.,2000. Climate extremes: observations, modeling, and impacts. Science 289,2068–2074.

Fava, F., Parolo, G., Colombo, R., Gusmeroli, F., Della Marianna, G., Monteiro, A.T., et al.,2010. Fine-scale assessment of hay meadow productivity and plant diversityin the European Alps using field spectrometric data. Agriculture Ecosystems &Environment 137, 151–157.

Foppa, N., Seiz, G., 2012. Inter-annual variations of snow days over Switzerland from2000–2010 derived from MODIS satellite data. Cryosphere 6, 331–342.

Galvagno, M., Wohlfahrt, G., Cremonese, E., Rossini, M., Colombo, R., Filippa, G., et al.,2013. Phenology and carbon dioxide source/sink strength of a subalpine grass-land in response to an exceptionally short snow season. Environmental ResearchLetters 8, 025008.

Gamon, J.A., Penuelas, J., Field, C.B., 1992. A narrow-waveband spectral index thattracks diurnal changes in photosynthetic efficiency. Remote Sensing of Environ-ment 41, 35–44.

Gitelson, A.A., Gritz, Y., Merzlyak, M.N., 2003. Relationships between leaf chlorophyllcontent and spectral reflectance and algorithms for non-destructive chlorophyllassessment in higher plant leaves. Journal of Plant Physiology 160, 271–282.

Gitelson, A.A., Keydan, G.P., Merzlyak, M.N., 2006a. Three-band model for noninva-sive estimation of chlorophyll, carotenoids, and anthocyanin contents in higherplant leaves. Geophysical Research Letters 33, L11402.

Gitelson, A., Merzlyak, M.N., 1994. Spectral reflectance changes associated withautumn senescence of Aesculus hippocastanum L and Acer platanoides L leaves –spectral features and relation to chlorophyll estimation. Journal of Plant Physi-ology 143, 286–292.

Gitelson, A.A., Peng, Y., Masek, J.G., Rundquist, D.C., Verma, S., Suyker, A., et al., 2012.Remote estimation of crop gross primary production with Landsat data. Remote

Sensing of Environment 121, 404–414.

Gitelson, A.A., Vina, A., Verma, S.B., Rundquist, D.C., Arkebauer, T.J., Keydan, G., et al.,2006b. Relationship between gross primary production and chlorophyll contentin crops: implications for the synoptic monitoring of vegetation productivity.Journal of Geophysical Research 111, D08S11.

1 Earth

H

H

H

L

L

L

M

M

M

M

M

M

M

P

P

tion of gross primary production in wheat using chlorophyll-related vegetation

0 M. Rossini et al. / International Journal of Applied

arris, A., Dash, J., 2010. The potential of the MERIS terrestrial chlorophyll index forcarbon flux estimation. Remote Sensing of Environment 114, 1856–1862.

mimina, G., Dufrêne, E., Pontailler, J.Y., Delpierre, N., Aubinet, M., Caquet, B., et al.,2013. Evaluation of the potential of MODIS satellite data to predict vegeta-tion phenology in different biomes: an investigation using ground-based NDVImeasurements. Remote Sensing of Environment 132, 145–158.

uete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview ofthe radiometric and biophysical performance of the MODIS vegetation indices.Remote Sensing of Environment 83, 195–213.

iang, S., Zheng, T., Liu, R., Fang, H., Tsay, S.-C., Running, S., 2006. Estimation ofincident photosynthetically active radiation from moderate resolution imagingspectrometer data. Journal of Geophysical Research 111, D15208.

iu, R.G., Liang, S.L., He, H.L., Liu, J.Y., Zheng, T., 2008. Mapping incident photo-synthetically active radiation from MODIS data over China. Remote Sensing ofEnvironment 112, 998–1009.

yapustin, A.I., 2003. Interpolation and profile correction (IPC) method for short-wave radiative transfer in spectral intervals of gaseous absorption. Journal ofthe Atmospheric Sciences 60, 865–871.

eehl, G.A., Tebaldi, C., 2004. More intense, more frequent, and longer lasting heatwaves in the 21st century. Science 305, 994–997.

ercalli, L., Berro, D.C., 2003. Atlante climatico della Valle d’Aosta. Società meteo-rologica subalpina.

eroni, M., Barducci, A., Cogliati, S., Castagnoli, F., Rossini, M., Busetto, L., et al., 2011.The hyperspectral irradiometer, a new instrument for long-term and unattendedfield spectroscopy measurements. Review of Scientific Instruments 82, 043106.

erzlyak, M.N., Gitelson, A., 1995. Why and what for the leaves are yellow in autumn– on the interpretation of optical-spectra of senescing leaves (Acer platanoidesL). Journal of Plant Physiology 145, 315–320.

igliavacca, M., Galvagno, M., Cremonese, E., Rossini, M., Meroni, M., Sonnentag,O., et al., 2011. Using digital repeat photography and eddy covariance data tomodel grassland phenology and photosynthetic CO2 uptake. Agricultural andForest Meteorology 151, 1325–1337.

onteith, J.L., 1972. Solar radiation and productivity in tropical ecosystems. Journalof Applied Ecology 9, 747–766.

onteith, J.L., 1977. Climate and efficiency of crop production in Britain. Philosophi-cal Transactions of the Royal Society of London. Series B, Biological Sciences 281,271–294.

anigada, C., Rossini, M., Meroni, M., Marzuoli, R., Gerosa, G., Colombo, R., 2009.Indicators of ozone effects on Fagus sylvatica L. by means of spectroradiometricmeasurements. Italian Journal of Remote Sensing 41, 3–20.

eng, Y., Gitelson, A.A., Keydan, G., Rundquist, D.C., Moses, W., 2011. Remote esti-mation of gross primary production in maize and support for a new paradigm

Observation and Geoinformation 29 (2014) 1–10

based on total crop chlorophyll content. Remote Sensing of Environment 115,978–989.

Peng, Y., Gitelson, A.A., Sakamoto, T., 2013. Remote estimation of gross primaryproductivity in crops using MODIS 250 m data. Remote Sensing of Environment128, 186–196.

Rossini, M., Cogliati, S., Meroni, M., Migliavacca, M., Galvagno, M., Busetto, L., et al.,2012. Remote sensing-based estimation of gross primary production in a sub-alpine grassland. Biogeosciences 9, 2565–2584.

Rossini, M., Meroni, M., Migliavacca, M., Manca, G., Cogliati, S., Busetto, L., et al.,2010. High resolution field spectroscopy measurements for estimating grossecosystem production in a rice field. Agricultural and Forest Meteorology 150,1283–1296.

Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., Harlan, J.C., 1974. Monitoring thevernal advancements and retro gradation of natural vegetation. In: NASA/GSFCFinal Report, Greenbelt, MD, USA, p. 371.

Sakamoto, T., Gitelson, A.A., Wardlow, B.D., Verma, S.B., Suyker, A.E., 2011. Esti-mating daily gross primary production of maize based only on MODIS WDRVIand shortwave radiation data. Remote Sensing of Environment 115, 3091–3101.

Schaepman-Strub, G., Schaepman, M.E., Painter, T.H., Dangel, S., Martonchik, J.V.,2006. Reflectance quantities in optical remote sensing-definitions and case stud-ies. Remote Sensing of Environment 103, 27–42.

Sims, D.A., Gamon, J.A., 2002. Relationships between leaf pigment content and spec-tral reflectance across a wide range of species, leaf structures and developmentalstages. Remote Sensing of Environment 81, 337–354.

Streb, P., Shang, W., Feierabend, J., Bligny, R., 1998. Divergent strategies of photo-protection in high-mountain plants. Planta 207, 313–324.

Stylinski, C.D., Gamon, J.A., Oechel, W.C., 2002. Seasonal patterns of reflectanceindices, carotenoid pigments and photosynthesis of evergreen chaparral species.Oecologia 131, 366–374.

Turner, D.P., Urbanski, S., Bremer, D., Wofsy, S.C., Meyers, T., Gower, S.T., et al.,2003. A cross-biome comparison of daily light use efficiency for gross primaryproduction. Global Change Biology 9, 383–395.

Wipf, S., Rixen, C., 2010. A review of snow manipulation experiments in Arctic andalpine tundra ecosystems. Polar Research 29, 95–109.

Wu, C.Y., Niu, Z., Tang, Q., Huang, W.J., Rivard, B., Feng, J.L., 2009. Remote estima-

indices. Agricultural and Forest Meteorology 149, 1015–1021.Yi, C., Rustic, G., Xu, X., Wang, J., Dookie, A., Wei, S., et al., 2012. Climate extremes

and grassland potential productivity. Environmental Research Letters 7,035703.