Remote Sensing of Environment 168 (2015) 360–373
Contents lists available at ScienceDirect
Remote Sensing of Environment
journal homepage: www.elsevier.com/locate/rse
Uncertainty analysis of gross primary production upscaling using
Random Forests, remote sensing and eddy covariance data
Gianluca Tramontana a,⁎, Kazuito Ichii b, Gustau Camps-Valls c, Enrico Tomelleri d, Dario Papale a
a
DIBAF, Department for Innovation in Biological, Agro-food and Forest Systems, University of Tuscia, Via San camillo de Lellis snc., 01100 Viterbo, Italy
Department of Environmental Geochemical Cycle, Japan Agency for Marine-Earth Science and Technology 3173-25, Showa-machi, Kanazawa-ku, Yokohama 236-0001, Japan
Image and Signal Processing Group (ISP), Image Processing Laboratory (IPL), Parc Científic - Campus de Paterna, Universitat de València, C/ Catedrático José Beltran, 2, 46980 Paterna,
Valencia, Spain
d
Institute for Applied Remote Sensing, European Academy of Bozen (EURAC), Viale Druso, 1, 39100 Bolzano, Italy
b
c
a r t i c l e
i n f o
Article history:
Received 22 January 2015
Received in revised form 6 July 2015
Accepted 15 July 2015
Available online xxxx
Keywords:
Upscaling
Uncertainty
Random Forest
MOD17
GPP
Vegetation indices
a b s t r a c t
The accurate quantification of carbon fluxes at continental spatial scale is important for future policy decisions in
the context of global climate change. However, many elements contribute to the uncertainty of such estimate. In
this study, the uncertainties of eight days gross primary production (GPP) predicted by Random Forest (RF) machine learning models were analysed at the site, ecosystem and European spatial scales. At the site level, the uncertainties caused by the missing of key drivers were evaluated. The most accurate predictions of eight days GPP
were obtained when all available drivers were used (Pearson's correlation coefficient, ρ ~ 0.84; Root Mean Square
Error (RMSE) ~ 1.8 g C m−2 d−1). However, when predictions were based on only remotely sensed data the
accuracy was close to the optimum (ρ ~ 0.8; RMSE ~ 1.9 g C m−2 d−1) and to a commonly used light use efficiency
model (MOD17) with parameters optimised for the applied study sites (the MOD17+, ρ ~ 0.79; RMSE ~ 2.04 g C
m−2 d−1). Remotely sensed data were key drivers for the accurate prediction of GPP in ecosystems with high variability of green biomass over the phenological cycle (e.g., deciduous broad-leaved forests) or highly affected by
the human management (e.g. croplands). In contrast, in the ecosystems with low variability of greenness
(e.g., evergreen broad-leaved forests), the predictions were poor when meteorological information were not
used. At a European spatial scale, when modelled grids of meteorological, land cover and fPAR data were used
as inputs, the propagation of their uncertainty, not accounted in the models training, had significant effects on
the uncertainty of the mean annual GPP. At this scale, the effects of meteorological uncertainty were higher
than the misclassification error. These findings suggested that a strategy based on satellite-measured data
could be a favourable improvement for the spatial upscaling of GPP, because avoiding the propagation of the uncertainties of the modelled grids.
© 2015 Elsevier Inc. All rights reserved.
1. Introduction
The accurate estimation of spatially explicit carbon fluxes is an
important goal to improve the understanding of the feedbacks between
the terrestrial biosphere and the atmosphere in the context of global
change and facilitation of climate policy decisions (Running et al.,
1999).
The carbon, water, and energy fluxes of land ecosystems are intimately connected (Beer, Reichstein, Ciais, Farquhar, & Papale, 2007;
Beer et al., 2009; Schimel, Braswell, & Parton, 1997). The in-situ estimations of carbon, water and energy fluxes can be obtained by the eddy
⁎ Corresponding author.
E-mail addresses: [email protected] (G. Tramontana), [email protected]
(K. Ichii), [email protected] (G. Camps-Valls), [email protected]
(E. Tomelleri), [email protected] (D. Papale).
http://dx.doi.org/10.1016/j.rse.2015.07.015
0034-4257/© 2015 Elsevier Inc. All rights reserved.
covariance technique (Aubinet, Vesala, & Papale, 2012), a welldeveloped method for measuring trace flux quantities between the biosphere and the atmosphere (Running et al., 1999). Using this technique,
net ecosystem carbon exchange (NEE) is directly measured, whereas
gross primary production (GPP) and total ecosystem respiration are
estimated using different partitioning methods (Desai et al., 2008;
Lasslop et al., 2010; Reichstein et al., 2005).
From the site level measurements, the regional, continental and
global estimates of carbon fluxes are obtained by spatial extrapolation conducted with models in which the spatial variability is mostly
driven by earth observation data (Jung, Le Maire, et al., 2007; Jung,
Reichstein, & Bondeau, 2009; Jung, Vetter, et al., 2007; Running
et al., 1999).
Both process-based and empirical approaches are commonly used to
estimate spatially explicit carbon fluxes. Process-based models such as
ORCHIDEE (Krinner et al., 2005), BIOME3 (Haxeltine & Prentice, 1996)
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
and LPJ-DGVM (Sitch et al., 2003) explicitly describe the physical processes that regulate energy, carbon and water cycles. These models
are useful for predicting future scenarios under global climate change.
However, the use of these models has limitations due to inherent assumptions such as the complexity of the model structure and ad-hoc
parameters. The empirical models are established differently and use
statistics to find the best possible relation between a set of explanatory
variables (inputs) and one or more target (outputs) without including
an explicit parametric description of the physical processes relating
them. In general, machine learning (ML) techniques are applied for
data-driven models that use empirical data (measured examples) to develop quantitative predictive models (Hastie, Tibshirani, & Friedman,
2001). Several ML algorithms that are based on different statistical or
computational principles, such as the Artificial Neural Networks (ANN,
Papale & Valentini, 2003), the Model Tree Ensemble (MTE, Jung et al.,
2009) and the Support Vector Machine (SVM, Yang et al., 2007) are applied to upscale fluxes.
Because of the basic premise, the application of empirical models is
strictly dependent on the variables used as drivers and on the representativeness of all primary ecosystem characteristics that affect carbon
fluxes (i.e., vegetation type, age, health, abiotic and biotic stress, seasonality, and phenology). Additionally, empirical models generally predict
outcomes for samples that have similar characteristics to training data,
but typically fail when they are applied to situations not observed
during the training phase (extrapolation). The ability of the model to
correctly estimate the output when applied to new examples is the
“generalisation” and it is affected by many factors, such as model complexity, missing of important drivers, data quality and the representativeness of the training examples.
In the spatial upscaling by empirical models, the choice of the drivers
is crucial and often it is a compromise between usefulness for the
upscaling purposes and availability in gridded format with sufficient
quality. As an example, the importance of soil characteristics in ecosystem carbon flux dynamics is well known, but these data are generally not used because their limited availability as spatially explicit
databases and high uncertainty. In contrast, meteorological data are
often used as drivers because these key variables are measured at the
sites and are available as spatially explicit fields from reanalysis
products. Moreover, meteorological data provide information both
for seasonal conditions and for daily stress factors but not for the
green biomass and the vegetation health, which can be inferred
from earth observation data.
Remote sensing variables, particularly vegetation indices, do not directly represent carbon fluxes processes (Jung et al., 2008), but as
shown previously, they are statistically related to ecosystem fluxes
(Olofsson et al., 2008; Rahman, Sims, Cordova, & El-Masri, 2005). Vegetation indices are calculated using measured reflectances in specific
spectral bands that are related to some chemical and physical properties
of the vegetation. For example, greenness indices such as the Normalised Difference Vegetation Index (NDVI) or the Enhanced difference
Vegetation Index (EVI) (Olofsson et al., 2008; Sims et al., 2008) are related to the amount of green biomass (e.g., leaf area index, LAI), whereas
water indices such as the Normalised Difference Water Index (NDWI)
(Gao, 1996) provide information on the canopy water content. Remote
sensing data are also used as the basis to derive the land cover maps that
are used in modelling exercises when the model parameterisation is
specific for a Plant Functional Type (PFT).
Generally, ML methods use both meteorological data and measured
or derived remote sensing data as inputs to estimate carbon fluxes
(Jung et al., 2011). At the site level, this strategy provided satisfactory
results (Moffat, Beckstein, Churkina, Mund, & Heimann, 2010), though
the model parameters could be affected by the uncertainty of the measurements. When the models are applied at larger spatial scales,
gridded versions of the inputs are necessary, and the uncertainties
must be considered an additional source of errors that affect simulated
outputs.
361
The spatially gridded inputs necessary to apply the models can
be measured (e.g., the remotely sensed spectral reflectances), obtained by other models or interpolation techniques (e.g., the
gridded meteorological data) or be obtained from classification
schemes such as the land cover or PFT maps. If a ML model only
uses spatially explicit variables that are directly measured as inputs
(e.g., vegetation indices or spectral reflectances), the uncertainty
associated with the production of the derived spatial data is removed. Moreover, although remotely sensed spectral reflectance
and land surface temperature provide a great amount of useful information, if the modelling exercise is performed without meteorological or land cover data, important information may be missing.
For example, during drought, an immediate effect occurs on the
fluxes caused by stomata closure, but reflectance is generally
affected later when the stress conditions persist (e.g., when the
leaf tissue chlorophyll contents change).
In this study, a diagnostic machine learning method called Random
Forest (RF) (Breiman, 2001), was used to predict the eight days GPP
and the mean European annual carbon budget, with the aim of
analysing the impacts of different sources of uncertainty on the
predictions. RF methods were used with the GPP derived from the
eddy-covariance measurements of NEE. At site and ecosystem levels,
the effects of the missing key drivers on the accuracy of GPP predictions
were evaluated. At European scale it has been analysed the effects of the
uncertainty in gridded drivers that are obtained by other models (meteorological variables and land cover maps) on the mean European annual
GPP.
2. Materials and methods
2.1. Site level data
In this study, the time series of meteorological variables, GPP, and remote sensing measured and derived data coming from 44 European
study sites were used (Table 1). GPP and meteorological in-situ data,
in particular the incoming solar radiation, air temperature, vapour
pressure deficit (VPD) and precipitation, were obtained by the
European database of flux data (www.europe-fluxdata.eu), while
the satellite data were obtained by the MODIS sensor on board of
the TERRA satellite.
The measurements of net CO2 fluxes (NEE) were quality filtered
using standard methods (Papale et al., 2006) and gap-filled using Artificial Neural Network (ANN) method (Papale & Valentini, 2003). The GPP
was calculated from NEE measurements using the partitioning method
described by Reichstein et al. (2005). Half hourly estimations of GPP
were aggregated to eight days time resolution to be consistent with
the satellite data (Sims et al., 2008; Xiao et al., 2004, 2010). An additional quality filtering was applied to the data using only eight days periods
in which the GPP was calculated starting with at least 80% of original or
high quality gap filled NEE measurements (quality flag in the data equal
to 0.8). This last filtering method was also applied to each meteorological variable.
The remotely sensed data were obtained by the MODIS sensor on
board of the TERRA satellite. For the site level analysis, the MODIS cutout
data, (available at the FLUXNET website (http://daac.ornl.gov/cgi-bin/
MODIS/GR_col5_1/mod_viz.html)), were used. These data cover
7 × 7 km2 centred on each site. The time series of spectral reflectances
(MOD09A1; Xiao et al., 2010; Vermote & Vermeulen, 1999), daily land
surface temperature (LST; Xiao et al., 2010; Wan, Zhang, Zhang, & Li,
2002), fraction of absorbed photosynthetically active radiation (fPAR,
MOD15A2 product; Xiao et al., 2010; Myneni et al., 2002), and GPP
(MOD17A2 product; Xiao et al., 2010; Heinsch et al., 2006; Zhao,
Heinsch, Nemani, & Running, 2005) were acquired. The MOD09A1 spectral reflectances were further used to calculate NDVI, EVI and NDWI
vegetation indices. The MODIS products had specific quality flags that
362
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
Table 1
Eddy covariance sites used in this study. Acronyms of PFT are DBF (Deciduous Broadleaf Forest), ENF (Evergreen Needleleaf Forest), EBF (Evergreen Broadleaf Forest), CRO (Cropland), GRA
(Grassland), and MF (Mixed Forest), SAV (Savannah).
Site code
Site name
PFT*
Lat (°N)
Lon (°E)
Altitude (m)
T (a) [°C]
Prec (a) [mm]
Years
AT-Neu
BW-Jal
BW-Lon
BW-Vie
CH-Lae
CZ-BK1
DE-Hai
DE-Geb
DE-Gri
DE-Kli
DE-Meh
DE-Tha
DE-Wet
DK-Ris
DK-Sor
ES-ES2
ES-LMa
ES-LJu
FR-Fon
FR-Gri
FR-Hes
FI-Hyy
FI-Lom
FR-Pue
HU-Bug
IT-Amp
IT-Bon
IT-Col
IT-Lav
IT-Lec
IT-MBo
IT-Ren
IT-Ro1
IT-Ro2
IT-SRo
NL-Ca1
NL-Ca2
NL-Loo
PL-wet
PT-Esp
PT-Mi1
UK-AMo
UK-EBu
UK-Gri
Neustift
Jalhay
Lonzee
Vielsalm
Laegern
Bily Kriz forest
Hainich
Gebesee
Grillenburg
Klingenberg
Mehrstedt
Tharandt
Wetzstein
Risbyholm
Soroe
El Saler–Sueca (València)
Las Majadas del Tietar (Caceres)
Llano de los Juanes
Fontainbleu
Grignon
Hesse
Hyytiala
Lompolojankka
Puechabon
Bugac
Amplero
Bonis
Collelongo
Lavarone
Lecceto
Monte Bondone
Renon
Roccarespampani 1
Roccarespampani 2
San Rossore
Cabauw
Cabauw extension
Loobos
Rzecin (PolWet)
Espirra
Mitra II (Evora)
Auchencorth Moss
Easter Bush
Griffin
GRA
MF
CRO
MF
MF
ENF
DBF
CRO
GRA
CRO
MF
ENF
ENF
CRO
DBF
CRO
SAV
M
DBF
CRO
DBF
ENF
WET
EBF
GRA
GRA
ENF
DBF
ENF
EBF
GRA
ENF
DBF
DBF
ENF
GRA
GRA
ENF
WET
EBF
EBF
GRA
GRA
ENF
47.12
50.56
50.55
50.31
47.48
49.50
51.08
51.10
50.95
50.89
51.28
50.96
50.45
55.53
55.49
39.28
39.94
36.93
48.48
48.84
48.67
61.85
68.00
43.74
46.69
41.90
39.48
41.85
45.96
43.30
46.01
46.59
42.41
42.39
43.73
51.97
51.95
52.17
52.76
38.64
38.54
55.79
55.87
56.61
11.32
6.07
4.74
6.00
8.37
18.54
10.45
10.91
13.51
13.52
10.66
13.57
11.46
12.10
11.64
−0.32
−5.77
−2.75
2.78
1.95
7.07
24.30
24.21
3.60
19.60
13.61
16.53
13.59
11.28
11.27
11.05
11.43
11.93
11.92
10.28
4.93
4.90
5.74
16.31
−8.60
−8.00
−3.24
−3.21
−3.80
975
485
166
485
705
814
458
159
375
479
297
385
789
17
45
0
266
1607
102
120
313
179
273
272
106
833
1160
1532
1352
311
1554
1750
173
160
10
0
−1
33
56
79
213
265
188
350
6.3
6
10
7
9.3
4.9
7.5
8.5
7.2
9
7.8
7.7
5.9
8
8.2
18
18.5
16
10.2
11.5
9.2
3.8
−1.4
13.5
11
10
8.9
6.3
7.8
12
5.5
4.7
15.15
15.15
14.2
10
10
9.8
8.5
16
15.4
7.4
8
8.2
852
1200
750
1150
1100
1100
800
470
853
750
570
820
840
580
660
550
572
400
720
700
820
709
484
883
500
1360
1170
1180
1150
870
1189
809.3
876.2
876.2
920
800
780
786
526
709
665
900
890
1200
2003–2009
2006–2007
2004–2010
2000–2010
2004–2009
2004–2009
2000–2007
2001–2009
2004–2010
2004–2010
2003–2006
2000–2010
2002–2008
2004–2007
2000–2009
2004–2008
2004–2009
2005–2009
2005–2008
2004–2007
2000–2008
2000–2008
2007–2009
2000–2008
2002–2008
2002–2008
2007–2009
2003–2009
2004–2010
2005–2009
2003–2010
2003–2004
2000–2008
2004–2010
2000–2010
2003–2008
2005
2000–2010
2004–2007
2002–2008
2003–2005
2006–2009
2004–2009
2000–2006
were used to select only high or good quality pixel values at the eddy
tower positions.
2.2. Spatial data
Two meteorological gridded products, two land use cover maps and
one fPAR gridded dataset were used to evaluate the variability of the
mean annual GPP due to the uncertainty of the modelled spatially explicit inputs. The analysis of the spatial grid was conducted for the period 2006–2008.
The meteorological grids were provided by the Modern-Era Retrospective Analysis For Research and Applications (MERRA; http://disc.
gsfc.nasa.gov/SSW/) and the European Centre for Medium-Range
Weather Forecasts (ECMWF) Interim Reanalysis (ERA-Interim; http://
data-portal.ecmwf.int/data/d/interim_full_daily). These products were
easily available and were frequently used for simulations in models at
global scales. The MERRA products provided data as hourly averages
and at a horizontal resolution of 2/3° longitude by 1/2° latitude. The
ERA-Interim provided data at three or six hourly time steps (depending
on the variable) and at a spatial resolution of approximately 0.75°. The
spatial grids obtained by MERRA and the ERA-Interim were aggregated
at an eight days time step and resampled at a 0.5° spatial resolution by
applying the nearest neighbour algorithm.
The land cover maps used were the MCD12Q1 product and the Global Land Cover 2000 (GLC 2000). The MCD12Q1 was land cover data with
a spatial resolution of 500 m and a yearly time step obtained by the
MODIS sensor on board of the TERRA and AQUA satellites (Friedl et al.,
2010). Among the available classification schemes, the 17-class International Geosphere Biosphere Programme (IGBP) was adopted. The GLC2000 project was implemented by the Joint Research Centre (JRC) of
the European Commission (EC) in partnership with more than 30 partner institutions using the Satellite Pour l'Observation de la Terre (SPOT)
VEGETATION one km satellite data (Giri, Zhu, & Reed, 2005). The two
land cover maps (GLC2000 and MODIS) were reclassified following
Giri et al. (2005) for comparison with the categories in the eddy covariance sites ancillary data.
The gridded fPAR (eight days time step and one km spatial resolution)
was obtained with MCD15A2, in which the fPAR values were derived
through the inversion of a radiative transfer model using the MODIS
data acquired by satellites TERRA and AQUA as inputs. For each MODIS
product (MCD15A2 and MCD12), quality control was conducted with
specific quality flag layers, and low-quality pixels were discarded.
The fPAR and land cover data were aggregated to a 0.5 degree spatial
resolution per PFT to maintain information related to each land cover
type. More specifically, for each land cover class, the fractional cover
of that class, the mean value of fPAR and its variability determined by
the standard deviation, were derived at a 0.5 degree spatial resolution.
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
363
2.3. Models
(Heinsch et al., 2006; Sjöström et al., 2011).
2.3.1. Random Forest
The modelling approach used in this study was based on the Random Forest (RF) concept. The RF concept is an improved ensemble
method of decorrelated tree models (e.g., Breiman, 2001; Hastie et al.,
2001; Ho, 1998). Roughly speaking, RF combines several decision
trees built on different combinations of input explanatory variables
(drivers), and provide as output the mean prediction of the individual trees. This strategy is very beneficial to alleviate the often reported overfitting problem of simple decision trees. Let us review the
main aspects of RF. The primary property of tree models is a
partitioning of the space into smaller regions to manage phenomena
characterised by very complex interactions among drivers. In particular, in tree models, partitioning is recursive. This phenomenon occurs when the sub-divisions are divided again until the partitioning
reduces the appropriate cost function. Recursive partitioning is terminated when the cost function cannot be further minimised;
hence, a simple model, usable only for the partitioned sub-region,
can be estimated.
The tree models use a tree scheme for representing recursive
partitioning. Each node of the tree is a splitting rule for the inputs, except for the terminals ones (the leaves), to which a simple model is attached and applied in that cell only. For regression trees, a mean value of
the target is in each leaf.
Different techniques are used to realise an ensemble of tree
models. For the RF method, tree models are independently realised
by “bagging”, a bootstrap aggregating technique (Hastie et al.,
2001) by which m training sets of samples are randomly extracted
from the overall dataset and used to parameterise m tree models.
The fraction of observations not used for the parameterisation,
“Out-Of-Bag” (OOB), is used to evaluate the ensemble. The
decorrelation of trees is achieved in the tree-growing process
through the random selection of the input variables (Hastie et al.,
2001) by bootstrap methods. Thus, many of the issues with regression trees, such as overfitting (Breiman, 2001; Cutler et al., 2007),
are corrected, and the risk of overfitting is reduced.
For each observation, the output of RF is the average of the outputs of
the trees, and hence RFs typically yield a reduced bias of the estimations
and in general good accuracies.
The Random Forest method has been applied in different areas of
concern in forest ecology, such as modelling the gradient of coniferous
species (Evans & Cushman, 2009), the occurrence of fire in Mediterranean regions (Oliveira, Oehler, San-Miguel-Ayanz, Camia, & Pereira,
2012), the classification of species or land cover type (Cutler et al.,
2007; Gislason, Benediktsson, & Sveinsson, 2006) and the analysis of
the relative importance of the proposed drivers (Cutler et al., 2007) or
for the drivers selection (Genuer, Poggi, & Tuleau-Malot, 2010; Gislason
et al., 2006; Jung & Zscheischler, 2013).
ε ¼ εmax T mins VPDs
2.3.2. Semiempirical GPP model: MOD17 and MOD17+
The results provided by the RF were compared to those obtained
by the well-known radiation use efficiency model MOD17 (Heinsch
et al., 2006; Zhao et al., 2005). In MOD17, (1) the GPP is calculated
as the product of fPAR, incoming PAR and actual light use efficiency
(ε):
GPP ¼ PAR fPAR ε
ð1Þ
The actual light use efficiency was obtained multiplying a maximum
light use efficiency (εmax), specific to each plant functional type (PFT)
with two reducing factors (2) that were calculated using the minimum
air temperature (Tmins) and the vapour pressure deficit (VPDs)
ð2Þ
For this study, we used two types of MOD17 outputs. The first was
the standard distributed product MOD17A2 collection 5 (Zhao et al.,
2005), and the second was the output produced by a version of the
model in which parameters were optimised for the site level data
used in this study (MOD17 +), following the example of Beer et al.
(2010).
The optimisation of MOD17+ was conducted in the R environment
by “R Based Genetic Algorithm” (rbga) on “genalg” package. The optimisation procedure selected a population of parameters to minimise the
sum of squares of the model residuals. The MOD17 + parameters
were estimated per PFT; more specifically, after 50 iterations, 100 best
vectors of MOD17+ parameters were selected, and the median values
were calculated.
Other models based on remote sensing input were available such for
example the Vegetation Photosynthesis Model (VPM, Xiao et al., 2004)
and the Temperature Greenness model (TG, Sims et al., 2008). For the
scope of this paper, it was sufficient to select only one model to be
used as reference; however, a cross-validation activity based on
optimised version of these additional models was performed and results
showed very similar performances (data not shown).
2.4. Experimental setup
Ten versions of RF were trained with different combinations of four
groups of drivers: vegetation indices (VIS), meteorological data
(METEO), fraction of absorbed photosynthetically active radiation
(fPAR) and plant functional type (PFT). The VIS dataset was composed
of vegetation indices NDVI, EVI, and NDWI, daily land surface temperature and the top of atmosphere (TOA) potential radiation, which was
calculated based on the geographical position of each site. The METEO
dataset was composed of the measured incoming solar radiation, air
temperature, VPD and precipitation.
A binary code was associated to each version of RF to indicate which
group of drivers was used. The order of the drivers in the binary code
was bit four (‘1000’) for VIs, bit three (‘0100’) for METEO, bit two
(‘0010’) for fPAR and bit one (‘0001’) for PFT, with ‘1’ indicating the
use of the drivers. For example, the model RF1000 considered as input
only vegetation indices, whereas the model RF0111 used meteorological data, fPAR and PFT. The models codes and the input data groups
are presented in Table 2.
2.4.1. Analysis of model performances and uncertainties at the site level
The site level cross-comparison of RF was conducted with the leaveone-out strategy: more specifically one site at time was excluded by the
training of RF and then used for the models evaluations by predictions.
The capability of the models to predict the overall time series, the differences among sites (i.e., the mean GPP of each site), the seasonal patterns
Table 2
Models codes and datasets used as inputs.
Model code
VIS
METEO
fPAR
PFT
Input datasets
RF1000
RF0100
RF0110
RF1100
RF1110
RF1001
RF0101
RF0111
RF1101
RF1111
Yes
No
No
Yes
Yes
Yes
No
No
Yes
Yes
No
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
Yes
No
No
Yes
No
Yes
No
No
Yes
No
Yes
No
No
No
No
No
Yes
Yes
Yes
Yes
Yes
VIS
METEO
METEO, fPAR
VIS, METEO
VIS, METEO, fPAR
VIS, PFT
METEO, PFT
METEO, fPAR, PFT
VIS, METEO, PFT
VIS, METEO, fPAR, PFT
364
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
and the anomalies was analysed following Jung et al. (2011). More specifically, the mean seasonal cycle (MSC) per site was calculated with the
averaged values for each eight days period across all available years, but
only when at least two values (i.e., years) for each eight days period
were available. To assess the among sites variability, the mean value
for each site was calculated as the mean of MSC if at least 23 of the 46
values of the MSC were present, whereas the anomalies were calculated
as the deviation of a flux value from the MSC. Finally, the mean site
values were removed from the MSC.
For each analysed time series component (overall, among sites, seasonality and anomalies), a ranking of the different RF approaches
(different input datasets) was realised using Pearson's linear correlation coefficient between the predicted values and the ones estimated by eddy covariance. Then, RF performances using only
remotely sensed data were compared with the best RF result from
the ranking analysis and with the ones of MOD17, both the standard product and that optimised at the site level (MOD17 +). This
specific focus was useful to estimate the distance between the accuracy of the best predictions and that obtained using only remote
sensing measurements.
The evaluation of eight days GPP predicted by RF at the site level was
completed with the analysis of the accuracy of the prediction per
ecosystem type. This evaluation was conducted because the capability
to predict GPP for some ecosystems could be affected by the absence
of important drivers, particularly if they were representative of specific
limiting factors for carbon uptake. For this analysis, the RMSE, as calculated per model and PFT, were estimated by recursive random sampling
of predictions stratified for seasonality. This sampling strategy was chosen because the RMSE was affected by the magnitude of the change in
GPP in relation to ecosystem type and with seasonal dynamics. More
specifically, 100 sets of 46 predicted values per ecosystem type, equally
representing the seasons, were sampled, and the RMSE was calculated.
The estimated values of RMSE per model and PFT were used as the input
to a one-way analysis-of-variance (ANOVA) with a blocking effect (the
driver datasets were considered “treatments”, and the PFT was considered a “blocking factor”) to establish if the differences between models
and PFT were statistically significant and to determine which models
were the best for each ecosystem type.
2.4.2. Analysis of uncertainties at a European spatial scale
When an empirical or statistical model such as the one applied in
this work uses data that are not directly measured but coming from
another modelling exercise (e.g., meteorological data or land cover
maps), the uncertainty associated with the production should be
considered. This effect is typically referred in statistics and remote
sensing and geosciences as uncertainty propagation (O'Hagan,
2012).
The uncertainty of the up-scaled mean annual GPP of Europe was
analysed for the uncertainties introduced by meteorological grids and
land cover maps. RF was applied to different meteorological and land
cover gridded products combinations (Table 3). The prediction of GPP
was performed for the period 2006–2008 at an eight days time step
then aggregated at yearly time step; the yearly outputs, produced by
different gridded product combinations, were used to estimate both
the mean GPP over the European domain and the uncertainties originated from the modelled grids. A two-way analysis of variance (ANOVA)
Table 3
Input combinations applied to produce a European scale gridded output.
Combination
Meteorological grid
Land cover grid
MERRA MCD12
ERA MCD12
MERRA GLC2000
ERA GLC2000
MERRA (GMAO)
ERA-Interim
MERRA (GMAO)
ERA-Interim
MCD12
MCD12
GLC2000
GLC2000
was used to find the pixels where the use of different land cover maps
and meteorological gridded products determined statistically significant differences in predictions.
The results were also compared with those obtained by nonparametric Friedman's and Kruskal–Wallis statistical tests but not important
differences were observed (see Appendix A).
Finally, the differences between the European GPP predicted by each
one-product combination and their mean (used as reference), were
analysed and compared with the differences between modelled drivers
(meteorological and land-use) in order to explore their roles and relationships with the uncertainty of output.
3. Results and discussions
In this section we analysed the main findings of our work. We first
provided an analysis of the accuracy at site level and at ecosystem
type. Then, we investigated the uncertainty of prediction at European
spatial scale and the relationships between uncertainty of predictions
and the ones of the modelled drivers' dataset.
3.1. Analysis of model performances and uncertainties at the site level
The ranking of Pearson's correlation coefficients for the different
tests (among site comparisons, anomalies, seasonality and overall
time series) is presented in Fig. 1. The models that used VIS as input
were among the best (Fig. 1, red bars) when they were also associated
with meteorological data.
In contrast, when VIS were not used, the models were among those
with a lower ρ (Fig. 1, yellow bars), particularly RF0100 and RF0101,
in which the fPAR was not used (Fig. 1, green bar). The RF driven by
only remotely sensed measured data (Fig. 1, blue bars) were ranked in
intermediate positions, with ρ closer to the best models. This result confirmed not only the critical role of remotely sensed data in the GPP empirical upscaling, but also the importance of meteorological data in
describing the dynamics both of the driving forces (solar radiation
e.g.) and the seasonal limiting factors (e.g., drought and water
availability).
The contribution of PFT as an input to a model was minor, and the
performance of models that differed only in the use of this additional
input (RF1000 vs RF1001, RF0100 vs RF0101, RF0110 vs RF0111,
RF1100 vs RF1101 and RF1110 vs RF1111) was very similar, with slightly better results obtained when PFT was used. This result demonstrated
that the PFT information improved the performance of models, though
this improvement was limited.
In Fig. 2 were shown the cross correlation between the eight days
GPP derived from eddy covariance and predictions by RF (the best predictions and RF with only measured remote sensing variables) and
MOD17 (standard and optimised).
From Fig. 2, time series components were predicted well by RF, except for the anomalies. In fact the RMSE of the eight days time step estimated from the best RF predictions was 1.8 g C m−2 d−1 (Fig. 2) and ρ
was approximately 0.84 for the overall time series. High accuracy
were also obtained for the among sites evaluation (ρ of 0.76 and
RMSE of 0.83 g C m−2 d−1) and for the seasonality with a RMSE of
1.03 g C m−2 d−1 and ρ of 0.93. However, the anomalies were not predicted well, with a RMSE of approximately 1.09 g C m−2 d−1 and a ρ of
0.54. The low ρ confirmed that prediction of anomalies was a difficult
task because the prediction was largely dependent on factors
(e.g., management, site history, soil carbon pool) that were not directly
represented in the drivers. The importance of measured remote sensing
variables in the spatial up-scaling of GPP was confirmed by the model
based only on that variable (RF1000), which also had high performance,
close to that of the best model. For the overall time series, the RMSE of
RF1000 was 1.9 g C m−2 d−1, with a ρ of 0.81, whereas for seasonality,
the RMSE was 1.12 g C m−2 d−1 and the ρ was 0.91. Additionally, similar
to the best predictions were the performances for the among-sites (ρ of
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
365
Fig. 1. Performances of RF at the site level. Each bar indicates a ρ value between GPP predicted by RF and estimated by eddy-covariance towers. Among site comparison, seasonality, anomalies and the overall time series were explored. RF1100, RF1110, RF1101 and RF1111 (red bars) used both meteorological data and VIS as input, RF0100 and RF0101 (yellow bars) used
meteorological data as a driver, RF0110 and RF0111 (green bars) used meteorological data and fPAR from a satellite as drivers, and RF1000 and RF1001 (blue bars) used only VIS as
input. The continuous black lines identify models in which PFT was used (RF1001, RF0101, RF111, RF01101 and RF1111). Vertical continuous black lines indicate the low and high ρ values
across models. Vertical dotted black line indicates ρ value for RF1000 (driven by only remote sensing measured data). (For interpretation of the references to colour in this figure legend,
the reader is referred to the web version of this article.)
0.75, RMSE of 1.01 g C m−2 d−1) and anomalies (ρ of 0.42, RMSE of
1.23 g C m−2 d−1). About the anomalies, the performance of RF1000
was very close to the best predictions and was significantly better
than models operating primarily with meteorological drivers. These results suggest a minor role of the meteorological drivers at least in the
simulation of the average eight days aggregated GPP. In fact, it is possible that the fast or instantaneous cause–effect relationships between
photosynthetic activity and meteorological drivers become less important when the data are analysed at eight days time resolution, suggesting that other slow responding drivers gain importance and that those
drivers are better represented by remotely sensed data such the vegetation indexes. In addition, it is important to consider that some of the meteorological drivers of GPP are also important drivers for phenology (e.g.
incoming solar radiation or temperature) and for this reason in a
Fig. 2. Cross comparison scatter plots between model output (x-axis, g C m−2 d−1) and reference GPP derived by eddy covariance (y-axis, g C m−2 d−1). First column: the results of the best
RF in accordance with the ranking in Fig. 1. Second column: the results of the RF driven by only measured remotely sensed data (RF1000). Third column: standard MOD17A2 product
extracted at the site location. Fourth column: MOD17+ with parameters optimised using the eddy covariance data.
366
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
machine learning tool the information is already included in the remotely sensed input.
The GPP anomalies were affected both by physiological reactions of
green ecosystems to meteorological drivers and by the interannual variability of seasonal greenness. The results suggested that this last component, representing interannual variability of photosynthetically
active vegetation density, might be more important than the meteorological drivers at the time resolution of the study. However, the generally poor results in the anomalies simulations confirm the difficulties in
the prediction of year-by-year GPP dynamic respect to the multi-year
average. This could be due, on the one hand, to some key driver that
could be missing (e.g. related to land use history, soil dynamic or management) and, on the other hand, to the uncertainty in the eddycovariance derived GPP that could be larger when analysed year by
year respect to when it is averaged to multi-annual means.
The performances of the RF were very close to the performance of
MOD17 operating with parameters optimised using the same eddy covariance data (MOD17 +); this model showed similar but slightly
worse results compared with both RF in Fig. 2 (the best model and the
one driven only by VIS). For MOD17 +, the ρ values were 0.79, 0.89
and 0.64 for overall, seasonality and among site comparisons, respectively, whereas the RMSE was approximately 2.04, 1.24 and 1.05 g C
m−2 d−1, respectively. Similar to the RF, the anomalies were not predicted well by MOD17 +; the ρ was 0.4 and the RMSE was 1.27 g C
m−2 d−1.
In particular, the RF1000 predictions very close to the best RF and
comparable with the performances of MOD17 + optimised at the site
level, should encourage future efforts with the application of machine
learning methods such as RF, with only measured remote sensing data
to upscale GPP. The MOD17 standard product had the highest RMSE
and the lowest correlation. The comparative results of MOD17+ with
the best RF suggest the requirement of MOD17 parameter optimization
for better GPP prediction using MOD17.
3.1.1. Analysis of accuracy assessment per ecosystem type
A more detailed investigation of the accuracy assessment of GPP by
RF was conducted to analyse the uncertainties of output produced at
the site level for each model and ecosystem type. Fig. 3 summarises
the performance of the RFS evaluated per PFT; ρ and RMSE were used
as metrics. Notably, the primary differences in the accuracy of predictions were among PFTS and not among models, which highlighted that
some ecosystems were less predictable than others. However, there
were also differences among the models, with generally better results
for those using remote sensing inputs.
The correlation coefficients were generally high for all models and
PFTS, with the exception of evergreen broadleaf forests (EBF), for
which the ρ was significantly lower and the relative RMSE significantly
higher. The models with the lowest performance in this PFT were the
ones driven by VIS (RF1000 and RF1001); this low performance was
most likely because of insufficient information from the remote sensing
data on the typical summer reduction of GPP caused by drought and
warming stress.
In contrast, models driven by only VIS dataset as input had good
performances in PFTS characterised by high variability of green biomass density during the vegetative period (e.g., deciduous broadleaf
forests) or by human management (e.g. croplands). For croplands
(CRO), all models produced high RMSEs, most likely because of the
heterogeneity in management practices, which were different
among sites and not described by the meteorological inputs. Nevertheless significant improvements there were for the models operating with VIS because carrying the information of the changed
amount of green vegetation.
These results confirm that at the time resolution used in this study,
most of the information about GPP and its dynamic are included in the
remotely sensed data that are also directly or indirectly correlated to
the meteorological drivers. In fact, in vegetation types where the greenness seasonality is not following any of the meteorological drivers
(Evergreen broadleaf forests EBF) errors were significantly higher
The one-way ANOVA confirmed that the significant variability of
RMSEs at the eight days time step was related both to the blocking effect
by PFTS (p b 0.001) and to the dataset used as a driver of GPP (p b 0.05).
However, the multiple comparisons of means (the Tukey test with a significance level of 95%) indicated that RF0100 (a model that used only
meteorological information) was significantly less accurate than other
models. If the RF0100 was excluded, significant differences among
models were not found.
3.2. Analysis of uncertainties at a European spatial scale
To estimate the effects of the uncertainties of the modelled inputs on
the European GPP, the RF0111 was applied. This model used as input
only modelled data as a grid: meteorological re-analysis, PFT classification and fPAR derived from remote sensing. The uncertainty of RF0111
estimated at the site level with a yearly time step was not significantly
different from the best models, thus justifying the use of the model in
the analysis.
In the following sections, the uncertainty of the GPP predicted by
RF0111 at a European spatial scale was analysed. In particular, in
Section 3.2.1, we analysed both the mean annual European GPP predicted by RF0111 and its uncertainty because of the modelled inputs. In
Section 3.2.2, we analysed the importance of the two kinds of modelled
input (meteorological grids and land cover maps) on the uncertainty of
predictions. Finally, in Section 3.2.3, we analysed how the divergence
between products has conditioned the uncertainty of predictions.
3.2.1. GPP predicted at a European spatial scale and uncertainty
constrained by modelled drivers
The annual European GPP predicted by RF0111 is shown in Fig. 4,
which was calculated as the average of the results obtained using as
input the four combinations of the available meteorological and land
Fig. 3. Maps of accuracy estimation per PFT as indicated by the Pearson linear correlation coefficients (ρ, left panel) and RMSE normalised for the mean values of GPP per PFT (rRMSE, right
panel).
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
367
Fig. 4. the mean annual GPP predicted by RF0111 at European spatial scale using four different combinations of meteorological and land cover grids (products), then calculating the mean
(left). The spatial distribution of the relative uncertainty (%), estimated as the ratio between the standard deviation of the mean annual predictions (each one of them obtained from a
specific combination of land cover and meteorological products) and their mean value (right).
Fig. 5. Maps of the intrinsic uncertainties of the models at European spatial scale for RF1000 (uncRF1000, a), RF0111 (uncRF0111, b), and their difference uncRF0111–uncRF1000 (c). Additionally, map for the uncertainties of the RF0111 predictions, introduced by the modelled drivers (uncRF0111*, d), and the boxplot of the intrinsic uncertainties of the models (RF1000
and RF0111) compared with the uncertainties of the predictions (by RF0111) introduced by the use of the modelled drivers (e).
368
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
cover products. Primarily, the west of Europe and in particular
along the Atlantic coast of the Iberian Peninsula, the west of
France and the Italian Peninsula were the areas with highest GPP.
High annual GPP was estimated for the southwest of the United
Kingdom (UK), the middle of France, in the central Europe (particularly in Germany) and along the Black Sea coast of Turkey. In contrast, low annual GPP was predicted for the North Africa, the
southeast of the Iberian Peninsula, particularly in Mediterranean
regions, the Anatolian Plateau, and the Scandinavian regions. This
spatial pattern was similar to the maps produced by Jung et al.
(2008), who applied the process model LPJ and four different diagnostic methods: ANN (Papale & Valentini, 2003), MOD17 + (Jung
et al., 2008), FAPAR-based productivity assessment (FPA), and
FPA including land cover (FPA + LC) (Jung et al., 2008). The spatial
pattern was also similar to the spatial pattern of annual GPP
produced by Beer et al. (2010) and Jung et al. (2011) for the
European countries.
The mean European GPP was 1293 (±136–202) g C m−2 y−1, which
was comparable with estimations by other studies. Jung et al. (2008)
found that GPP from different data driven approaches ranged from
900 g C m−2 y−1 to 1100 g C m−2 y−1 and Jung, Vetter, et al., 2007,
with the process models BIOME-BGC, LPJ and ORCHIDEE, reported a
mean yearly GPP between 783 and 1042 g C m−2 y−1. If compared
with those studies, RF0111 overestimated European GPP, but the low
productivity regions of the north and the east of Europe were not
included in the RF0111 simulations. The GPP predicted by RF0111
varied between 440 and 2480 g C m−2 y−1, results that were comparable with the estimation reported by Beer et al. (2007), who applied
a semiempirical watershed derived water use efficiency (WUE)
model to estimate a mean GPP that ranged between 1393 and
2071 g C m−2 y−1.
The uncertainty of annual GPP due to modelled drivers (Figs. 4 and
5d) and estimated pixel-by-pixel as the standard deviation of the
mean annual GPP predicted with different gridded products combinations, varied between 0.3 and 437 g C m−2 y−1, from 4% to 44%
of the predicted GPP while the average value was 72 g C m−2 y−1,
6% of the predicted GPP. The effects of the meteorological and land
cover uncertainty were particularly noticeable in the south of
Europe, in the Mediterranean region that included the North Africa,
the Italian Peninsula, the Hellenic and the Iberian Peninsulas, and
the northwest of Turkey, often in correspondence with regions
characterised by low predicted GPP (Figs. 4b and 5d). Other regions
with relatively high uncertainties were found in the southwest of
Scandinavia. In contrast, large areas of France and central Europe,
characterised by high estimated productivity, were less sensitive to
the drivers used.
Because the RMSE of RF0111 at the site level was 165 g C m−2 y−1
(136–202 g C m−2 y−1), very close to the best RF model (RF1111) in
which the site level RMSE at yearly time step was 152 g C m−2 y−1
(127–189 g C m−2 y−1), the importance of the uncertainty caused by
the modelled input was clear. This additional uncertainty could be reduced by using as input only directly measured data such as remote
sensing reflectances, vegetation indices, or land surface temperatures
because the input uncertainties were similar both for training and for
upscaling phases. To explore this hypothesis, the uncertainties of the
models RF1000 (uncRF1000) and RF0111 (uncRF0111) were upscaled
from the site level to the European spatial scale and were compared
for the uncertainty of predictions caused by the use of different
modelled gridded products (uncRF0111*). The upscaling of models uncertainty was conducted by calculating for each pixel the weighted
mean RMSE on the basis of the PFT fractional cover. The results are
shown in Fig. 5a and b. The spatial pattern of uncRF0111 and
uncRF1000 maps was similar, with values of uncRF0111 slightly lower
than for uncRF1000. The spatial pattern of the two maps (Fig. 5a and
b) highlighted low uncertainty in Scandinavia and on the Anatolian
Peninsula, which were also regions with low estimated GPP. High values
of model uncertainties occurred on the Italian and Iberian Peninsulas, in
France, in Germany, and in Central and East Europe. The latitudinal
trend of models uncertainties was similar to the one for the GPP predicted by RF0111, with values higher at southern and middle latitudes than
at northern latitudes. By contrast, in the east–west direction, no trend of
models uncertainties was discernable, however, the predicted annual
GPP was significantly lower at eastern than at western longitudes. The
differences between uncertainties of the two models (uncRF0111 and
uncRF1000, Fig. 5c) were in a narrow range of values, with a slightly
negative mean, and were uniformly distributed across Europe. However, two regions with high differences between the uncertainties of the
two models were found. The first, characterised by higher uncertainty
of RF0111 than RF1000, was the Ireland and the west of the UK, and
the second was the northeast of Europe where RF1000 produced clearly
higher uncertainty. The prevalence of slightly negative values in Fig. 5c
indicated that RF0111 was somewhat more accurate than the RF driven
only by measured remotely sensed data (RF1000). Moreover, the
mean value of unRF1000 across Europe was 138 g C m−2 y−1, with a
range of 85–188 g C m−2 y−1, whereas for uncRF0111, the value was
119 g C m−2 y−1, with a range of 72–147 g C m−2 y−1 (Fig. 5e). Nevertheless, when the error caused by uncertainty in the modelled inputs
was included (Fig. 5d, uncRF0111*) in the predictions, the final uncertainty of the output of RF0111 might change. The uncertainty of
RF0111 significantly increased, particularly in correspondence with
the Atlantic coast, Mediterranean regions, southern Europe, southwestern
coast of Scandinavia, and western Ireland and UK, where the uncertainty
Fig. 6. Map of pixels in which the ANOVA test has identified a statistically significant variability of the outputs (a) and the distribution of the uncertainties among the recognised factors (b).
NS was used for the pixels not significant at the ANOVA test, LC for the pixels significant only for the use of different land cover products, MT for the pixels significant for the use of different
meteorological data, LC + MT, for the pixels significant both for the use of different land cover and meteorological data but with an additive effect, LC ∗ MT for the pixels significant for the
interactions of the two factors.
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
369
Fig. 7. Maps of differences between mean annual GPP (g C m−2 y−1) predicted by each product combination and their mean.
caused by different versions of the drivers was high (orange and red regions in Fig. 5d). By contrast, the effects were small in the middle and central Europe, on the Scandinavian Peninsula and in the middle-East.
The distributions of the uncertainties of the two models at a
European spatial scale (uncRF1000 and uncRF0111) and the uncertainties of the outputs of RF0111 caused by different modelled drivers
Fig. 8. Maps of differences between meteorological products, calculated as the mean of the day-by-day differences ERA-Interim — MERRA. The green areas indicate where the ERA-Interim
product was lower than MERRA, whereas red regions indicate the opposite. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)
370
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
Table 4
Mean representativeness of PFTS in MCD12 and GLC2000 across European lands and their
differences.
PFT
MCD12 (%)
GLC2000 (%)
Difference (%)
DBF
ENF
CRO
MF
GRA
EBF
Others
2.41
9.20
21.67
7.77
3.22
0.02
55.70
8.64
10.85
18.45
4.89
3.43
0.05
53.69
−6.23
−1.65
3.23
2.88
−0.21
−0.03
2.01
(uncRF0111*) are shown in Fig. 5e. The boxes show that although the
uncertainties caused by the drivers were generally lower than the
ones caused by the models, their contributions were still relevant and
in some cases were very high (Fig. 5e). This finding should encourage
future efforts for the use of only remotely sensed measured data for
upscaling of carbon fluxes. This solution might have an impact on the
model capacity to reproduce GPP dynamics because of the missing of
key information, but the uncertainty of modelled drivers would be
avoided.
3.2.2. Relative importance of meteorological and land cover uncertainties
To disentangle the role of the two drivers in the uncertainty of predicted GPP, the RF0111 simulations were submitted to a pairwise comparison. The mean absolute uncertainty caused by the use of different
meteorological products and the same land cover was 100.05 g C m−2
y−1 (range 0.32–753.13 g C m−2 y−1), whereas the mean absolute uncertainty caused by the different land cover products was 47.15 g C
m−2 y−1 (range 0–754.41 g C m−2 y−1). Although the ranges of uncertainties were approximately the same, if these uncertainties were
expressed in relative terms as the ratio between uncertainty and
mean annual European GPP, they were 8% (range 0.0003–76%) for the
meteorological data and 3.5% (range 0–41.7%) for the land cover
maps. Thus, the areas with higher absolute uncertainty caused by the
land cover products occurred in pixels with high GPP, whereas the uncertainty of the meteorological drivers was more homogeneously distributed, covering also areas with low GPP.
The important role of the quality of meteorological grids on
modelled GPP was also highlighted by Jung, Vetter, et al. (2007), who
found a difference of 20% between overall GPP predicted by a reference
setup and a change in the meteorological gridded data product. Zhao,
Running, and Nemani (2006) found similar results in a study on the effects of the reanalysed meteorological data on the global GPP and NPP
estimated by MOD17. In that study, the largest difference occurred between predictions based on the National Centres for Environmental Prediction/National Centre for Atmospheric Research (NCEP/NCAR) dataset
and the European Centre for Medium-Range Weather Forecasts
(ECMWF ERA-40) dataset and was approximately 23 GtC yr−1, whereas
the relative error for the GPP ranged from 16% (ECMWF ERA-40) to 24%
(NCEP/NCAR).
Thus, the possible conclusion was that uncertainties in land cover
products had a secondary importance compared with the meteorological data. Jung, Vetter, et al. (2007) also found that uncertainty in the GPP
predicted by a process model driven with two different land cover classification schemes was less than 10%, approximately.
Regions where the use of different meteorological or land cover products led to a significantly different mean annual GPP were identified by
the two-way ANOVA test, conducted pixel by pixel. The results (Fig. 6) indicated that the regions where the differences in GPP among products
were statistically significant were concentrated in the south of Europe,
in the middle-east and in the north of Europe (Scandinavia and UK).
Additionally, for the spatial extent, the most important factor in the
uncertainty of GPP was the meteorological dataset, which was statistically significant for approximately 3.0 ∗ 106 km2 (Fig. 6, yellow area),
whereas land cover misclassification was significant for an area of
6.9 ∗ 105 km2 (Fig. 6, cyan area). The magnitude of the effects of land
cover and meteorological data were similar (Fig. 6b), with a mean induced uncertainty of 89 g C m−2 y−1 for land cover divergence
(Fig. 6b, LC) and 94 g C m−2 y−1 for differences in meteorological products (Fig. 6b, MT), and ranges that varied from 13 to 335 g C m−2 y−1
and from 12 to 434 g C m−2 y−1, respectively. Jointly, meteorological
and land cover factors were statistically significant for an area of
1.23 ∗ 106 km2 of which 1 ∗ 106 km2 was by an additive effect (Fig. 6, orange area) and 2.3 ∗ 105 km2 was by their interaction (Fig. 6, red area).
The boxes in Fig. 6b show that where both land cover and meteorological uncertainties were statistically significant there was an increase
in the estimated uncertainties. For the pixels in which the combined
effect was additive (Fig. 6b, LC + M), the mean value of uncertainties
was 122 g C m−2 y−1. For the pixels in which the interactions between
meteorological and land cover factors was statistically significant
(Fig. 6b, LC ∗ M), the mean yearly uncertainty was approximately
193 g C m−2 y−1.
We also used alternative nonparametric tests such as Friedman's
and Kruskal–Wallis tests of significance, and the results were similar.
(see Supplementary material, Appendix A).
3.2.3. Analysis of roles and relationships between drivers and GPP
uncertainties
The effects of the different driver combinations on the upscaled GPP
are shown in Fig. 7 and significant differences between predictions provided by each driver combination and the mean annual GPP were found.
The spatial pattern of the uncertainties was highlighted in the Mediterranean regions and middle-eastern Europe. The mean annual GPP produced using meteorological data from the ERA-Interim was higher
than the one obtained using MERRA in Southern (e.g. Mediterranean regions) and Middle-East Europe. In Central Europe, the west of the UK
Fig. 9. Maps showing for each 0.5 × 0.5 degree aggregated pixel the PFT where the two land cover products MCD12 and GLC2000 show higher differences in the % of coverage (a) and the
values of the difference calculated as %_MCD–%_GLC (b).
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
and Ireland, at northern latitudes and in cold areas at lower latitudes
(e.g., Alps Mountains), the output produced with ERA-Interim grids
led to lower GPP compared with the average and MERRA.
The only area where the land cover product was likely responsible
for the difference was the Iberian Peninsula, where the GPP estimated
using GLC2000 was greater than the GPP estimated using the MCD12
product.
This pattern could be explained by the analysis of the primary differences between the drivers.
By Fig. 8, showing the mean differences among meteorological products, it was noticeable that the shortwave global incoming radiation and
VPD were systematically lower in the ERA-Interim product compared
with MERRA, whereas for precipitation, the values were generally greater
in the ERA-Interim product than in MERRA. For air temperature, the distribution of divergences was complex, with greater values for MERRA in
North Africa, along the coast of Turkey, on the Iberian Peninsula and in
north Scandinavia. By contrast, air temperatures of the ERA-Interim
were greater than MERRA in Italy, near the Carpathian Mountains, in
the north east of Europe and in the southern half of Scandinavia.
The systematic bias between ERA-Interim and MERRA, particularly
for VPD and shortwave incoming solar radiation, could explain the differences in GPP products. The higher GPP in the west UK, cold regions
and in northern Europe when MERRA was used, could be related to
the higher incoming radiation, whereas the higher values of VPD in
the same areas did not have any effect because water availability was
not a limiting factors in those environments. The opposite result was
the higher productivity estimated by ERA-Interim at lower latitudes
(water limited environments such as Italy, North Africa and middleEast), which could be explained by the systematically lower value of
VPD in the ERA-Interim products than in MERRA.
The differences in the land cover classification that led to the uncertainty when this input was used are summarised in Table 4 and Fig. 9.
Table 4 summarises the presence of the different PFTS in Europe as extracted by the applied products and the differences. The most important
differences in the mean value of fractional land cover across Europe
were found for DBFs, which were more represented in GLC2000 than
MCD12 and, in contrast, for CROs and MFs that were more present in
MCD12 than GLC2000. Less important was the difference for ENF,
which was present more in GLC2000, and though very small could account for the differences with GRA and EBF.
The data reported in Table 4 confirmed the findings of Giri et al.
(2005), who described a larger degree of presence of forests for
GLC2000 at global scales. This relative abundance was also associated
with a different definition of forests in the land cover definition scheme
adopted by the two land cover products. Fig. 9a shows that for each
0.5 × 0.5 degree pixel, the PFT in which the two land cover products
had the higher difference; in Fig. 9b, its value was calculated as
%PFT_MCD12 − %PFT_GLC2000.
Correspondence between differences in classification and uncertainty in GPP was found on the Iberian Peninsula, where the GPP predicted
by GLC2000 was significantly higher than MCD12 (Fig. 7), and the
fractional land cover of CRO, ENF and MF estimated by GLC2000
was higher than MCD12 (Fig. 9, green areas). The higher GPP for
the predictions by GLC2000 was also affected by the absence of a
savannah-like land cover class in GLC2000. Because savannahs
(present in the training set and used as a separate PFT) were
characterised by relatively low GPP, this property could explain the
higher GPP when the GLC2000 was used. Cases in different regions
were also found where the effects on the mean annual GPP was
low, although the differences in the classification were important.
Cases such as these were found at north western France, in northern
UK, and in eastern Europe that involved primarily CRO and MF,
which was most likely because the two PFTS yielded similar total annual GPP values. Clearly, an analysis of the annual trend or interannual variability would most likely show stronger differences, but
such an analysis was beyond the scope of this paper.
371
4. Conclusions
In this paper, it was presented the application of the Random Forests
algorithm to estimate eight days GPP (at the site level) and the mean
annual European budget. The results showed that RF methods were
promising and comparable with other machine learning approaches
published before, including MTE (Beer et al., 2010; Jung et al., 2011)
and ANN (Papale & Valentini, 2003, Beer et al., 2010), or semiempirical
LUE models such as MOD17 (Beer et al., 2010; Running, Thornton,
Nemani, & Glassy, 2000; Sjöström et al., 2011).
The application at a European spatial scale confirmed the potential
applicability of the RF method for GPP prediction; the mean annual
values of upscaled product was in the range of estimations reported
by other authors, both at the site level and at the European spatial
scale (Beer et al., 2007, 2010; Jung, Vetter et al., 2007; Jung et al., 2007).
At the site level, using a leave-one-out cross-comparison strategy,
the best results were obtained by the combination of meteorological
and remotely sensed data as input (RF1111); however the differences
with results obtained by the other models (that have used in input a reduced drivers dataset) were not significant, except for the RF trained
with only meteorological inputs (RF0100). The models driven by only
measured remotely sensed data (RF1000) gave very good results,
close to the best model, which confirmed the importance of remote
sensing data in the spatial upscaling exercises. However, the accuracy
of predictions of the different models was variable for different PFTS.
Particularly for EBF sites, the models parameterised using only remotely
sensed data were unable to capture the effects of limiting factors such as
drought and water stress, which were instead better identified when
VPD was used as driver. Notably, when PFT was used as a driver, significant improvements in the accuracy of predictions did not occur, while
the variables describing the seasonality and vegetation health (fPAR
and remotely sensed vegetation indices) were fundamental to predict
GPP across PFTS.
These findings were the basis for the second part of the analysis in
which the role of differences in the modelled meteorological spatially
explicit drivers and misclassification of land cover maps were analysed
in relation to the GPP uncertainty at a European scale.
About the uncertainty of the upscaled GPP, a significant role of the
uncertainties of the land cover and meteorological drivers was found,
with the differences in land cover classification having a less significant
role compared with those in the meteorological fields. However, a significant effect of land cover misclassification was observed where the
PFTS having different capabilities (and magnitudes) of GPP were
involved.
The uncertainty caused by the differences between the modelled
input grids was very high, and in some regions it was greater than the
intrinsic uncertainty of the models.
The modelled inputs could be avoided by using only measured remotely sensed data as drivers. However, in some PFTS, where the meteorological conditions and induced stress factors not affect the canopy
greenness, the effects on averaged GPP were not well captured by the
available remote sensing data. Additionally, the uncertainties caused
by specific disturbances that affected the quality of remotely sensed
data such as acquisition geometry, atmospheric conditions, snow,
clouds, and unmixed pixels that were not addressed in this paper
could also play an important role.
With the availability of new sensors in the near future (e.g., the Sentinel constellation and the fluorescence bands currently available at
coarse resolution and under evaluation for future missions like
EarthExplorer 8 candidate's mission FLEX) and the expansion of the network of site level measurements in FLUXNET, further exploration of the
use empirical models based only on directly measured drivers will continue to capture fast responses of vegetation to stress factors (e.g., with
fluorescence or non-photochemical quenching using PRI), to monitor
disturbance and management history through multitemporal images
and to quantify ecosystem stocks using radar information.
372
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
Acknowledgement
The MODIS data products were obtained from the Oak Ridge National Laboratory (ORNL) Distributed Active Archive Center (DAAC) and the
Earth Observing System Data and Information System (EOSDIS). MERRA
data have been provided by the Global Modeling and Assimilation Office
(GMAO) at NASA Goddard Space Flight Center through the NASA GES
DISC online archive. ECMWF ERA-Interim data have been provided by
ECMWF data server. GLC2000 data have been provided by EU-JRC.
This work used eddy covariance data acquired by the FLUXNET community and in particular by the following networks: CarboEuropeIP,
CarboItaly, CARBOMONT, IMECC, ICOS-PP. The authors thank all these
data providers for the high quality of measurements/products made
available to the scientific community.
The authors thank Martin Jung, Markus Reichstein and Max Planck
Institute (MPI) for Biogeochemistry for their support, discussions and
hospitality.
Gianluca Tramontana has been supported by the EU project
GEOCARBON. Kazuhito Ichii acknowledges the support by the Environment Research and Technology Development Fund (2-1401)
from the Ministry of the Environment of Japan. Gustau Camps-Valls
acknowledges the support of the Spanish Ministry of Economy and
Competitiveness (MINECO) under project LIFE-VISION TIN201238102-C03-01. Enrico Tomelleri acknowledges the financial support
of the GHG-Europe project. Dario Papale thanks ICOS-INWIRE for the
support.
Appendix A. Analysis of the importance of meteorological and land
cover uncertainties by nonparametric tests and comparison with
ANOVA results
Fig. A1 and Table A1 show the primary findings concerning the identification of regions where the use of different meteorological or land
cover products as drivers led to significantly different estimates of
mean annual GPP, as determined by Friedman's and Kruskal–Wallis
nonparametric statistical tests.
This was performed as confirmation of the ANOVA results
(Section 3.2.2) with the use of statistical tests based on different
assumptions.
Similar to Section 2.4.2, the nonparametric tests were conducted
pixel-by-pixel with a significance level of 0.05.
The spatial pattern of pixels in which significant effects of meteorological and land cover products were recognised is shown in Fig. A1. The
results obtained with nonparametric analysis were similar to the findings by ANOVA, as shown in Fig. 6a.
In Fig. A1, the cyan pixels represent the regions in which the differences in land cover products had a significant role in the uncertainty
Table A1
Estimated surfaces in which mean annual predictions of GPP were significantly affected by
differences between land cover products, meteorological products or both. Surfaces are estimated by results derived from the two-way ANOVA (see Section 3.2.2) and two nonparametric methods (Friedman's and Kruskal–Wallis). The significance level was 0.05.
Factors
ANOVA
(km2 × 106)
Friedman's
(km2 × 106)
Kruskal–Wallis
km2 × 106)
Land cover
Meteo
Both
Not affected
0.69
3.04
1.23
4
0.73
3.07
1.11
4.06
0.88
2.28
1.11
4.65
of the GPP, and the yellow pixels represent the regions in which the differences in meteorological products had a significant role in the uncertainty of the GPP, whereas the red pixels represent the regions where
both factors were statistically significant. The patterns of the two maps
were similar to the pattern obtained by the ANOVA. Larger effects of divergence among meteorological products were found than for land
cover products. In particular, in the Mediterranean region, in middleeastern Europe, in regions with a continental climate, in the cold regions
of western UK, in Ireland and in southwestern Scandinavia, meteorological uncertainties were significant. In contrast, pixels in which the divergence in land cover products caused statistically significant differences
in GPP predictions were located on the Iberian and Hellenic Peninsulas,
in western Anatolia, and in sparse and fragmented regions of Scandinavia
and northern Europe. In these regions, and also on the Italian Peninsula
and in middle-eastern Europe, pixels were identified in which both factors were statistically significant.
In Table A1, it was compared the estimated surfaces for the factors recognised by the applied statistical tests. As expected, the estimated surface per category was consistent among the methods, with
results of the Kruskal–Wallis tests slightly different from the other
methods.
These results also supported our expectation concerning the importance of the meteorological reanalysis uncertainty on the uncertainty of
GPP, which could be considered more relevant than the land cover
maps.
References
Aubinet, M., Vesala, T., & Papale, D. (2012). Eddy covariance: A practical guide to measurement and data analysis. Dordrecht Heidelberg London New York: Springer (460 pp.).
Beer, C., Ciais, P., Reichstein, M., Baldocchi, D., Law, B. E., Papale, D., et al. (2009). Temporal
and among-site variability of inherent water use efficiency at the ecosystem level.
Global Biogeochemical Cycles, 23, 1–13, http://dx.doi.org/10.1029/2008GB003233,
2009 (GB2018).
Beer, C., Reichstein, M., Ciais, P., Farquhar, G. D., & Papale, D. (2007). Mean annual GPP of
Europe derived from its water balance. Geophysical Research Letters, 34, 1–4, http://
dx.doi.org/10.1029/2006GL029006 (L05401).
Fig. A1. Significant factors (land cover or meteorological products) affecting mean annual GPP variability recognised by nonparametric tests: Friedman's (left) and Kruskal–Wallis (right).
Cyan zones are significantly affected by the uncertainty of land cover products, yellow are zones significantly affected by differences between meteorological products, and red regions are
significantly affected both by differences in meteorological and in land cover products. Tests were conducted at the significance level of 0.05.
G. Tramontana et al. / Remote Sensing of Environment 168 (2015) 360–373
Beer, C., Reichstein, M., Tomelleri, E., Ciais, P., Jung, M., Carvalhais, N., et al. (2010). Terrestrial gross carbon dioxide uptake: Global distribution and covariation with climate.
Science, 329, 834–838, http://dx.doi.org/10.1126/science.1184984.
Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32, http://dx.doi.org/10.
1023/A:1010933404324.
Cutler, D. R., Edwards, T. C., Jr., Beard, K. H., Cutler, A., Hess, K. T., Gibson, J., et al. (2007).
Random forests for classification in ecology. Ecology, 88(11), 2783–2792, http://dx.
doi.org/10.1890/07-0539.1.
Desai, A. R., Richardson, A. D., Moffat, A. M., Kattge, J., Hollinger, D. Y., Barr, A., et al. (2008).
Cross-site evaluation of eddy covariance GPP and RE decomposition techniques.
Agricultural and Forest Meteorology, 148, 821–838, http://dx.doi.org/10.1016/j.
agrformet.2007.11.012.
Evans, J. S., & Cushman, S. A. (2009). Gradient modeling of conifer species using random forests. Landscape Ecology, 24, 673–683, http://dx.doi.org/10.1007/s10980-009-9341-0.
Friedl, M. A., Sulla-Menashe, D., Tan, B., Schneider, A., Ramankutty, N., Sibley, A., et al.
(2010). MODIS collection 5 global land cover: Algorithm refinements and characterization of new datasets. Remote Sensing of Environment, 114, 168–182, http://dx.doi.
org/10.1016/j.rse.2009.08.016.
Gao, B. C. (1996). NDWI — A Normalized Difference Water Index for remote sensing of
vegetation liquid water from space. Remote Sensing of Environment, 58, 257–266.
Genuer, R., Poggi, J. M., & Tuleau-Malot, C. (2010). Variable selection using random forests.
Pattern Recognition Letters, 31, 2225–2236, http://dx.doi.org/10.1016/j.patrec.2010.
03.014.
Giri, C., Zhu, Z., & Reed, B. (2005). A comparative analysis of the Global Land Cover 2000
and MODIS land cover data sets. Remote Sensing of Environment, 94, 123–132, http://
dx.doi.org/10.1016/j.rse.2004.09.005.
Gislason, P. O., Benediktsson, J. A., & Sveinsson, J. R. (2006). Random Forests for land cover
classification. Pattern Recognition Letters, 27, 294–300, http://dx.doi.org/10.1016/j.
patrec.2005.08.011.
Hastie, T., Tibshirani, R., & Friedman, J. (2001). The Elements of Statistical learning. New
York, NY, USA: Springer Series in Statistics Springer New York Inc, 587–604.
Haxeltine, A., & Prentice, I. C. (1996). BIOME3: an equilibrium terrestrial biosphere model
based on ecophysiological constraints, resource availability and competition among
plant functional types. Global Biogeochemical Cycles, 10, 693–710, http://dx.doi.org/
10.1029/96GB02344.
Heinsch, F. A., Zhao, M., Running, S. W., Kimball, J. S., Nemani, R. R., Davis, K. J., et al.
(2006). Evaluation of remote sensing based terrestrial productivity from MODIS
using regional tower eddy flux network observations. IEEE Transactions on
Geoscience and Remote Sensing, 44(7), 1908–1925, http://dx.doi.org/10.1109/TGRS.
2005.853936.
Ho, T. K. (1998). The random subspace method for constructing decision forests. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 20(8), 832–844, http://dx.
doi.org/10.1109/34.709601.
Jung, M., Le Maire, G., Zaehle, S., Luyssaert, S., Vetter, M., Churkina, G., et al. (2007).
Assessing the ability of three land ecosystem models to simulate gross carbon uptake
of forests from boreal to Mediterranean climate in Europe. Biogeosciences, 4, 647–656,
http://dx.doi.org/10.5194/bg-4-647-2007.
Jung, M., Reichstein, M., & Bondeau, A. (2009). Towards global empirical upscaling of
FLUXNET Eddy Covariance observations: Validation of a model tree ensemble approach using a biosphere model. Biogeosciences, 6, 2001–2013, http://dx.doi.org/10.
5194/bg-6-2001-2009.
Jung, M., Reichstein, M., Margolis, H. A., Cescatti, A., Richardson, A. D., Arain, M. A., et al.
(2011). Global patterns of land–atmosphere fluxes of carbon dioxide, latent heat,
and sensible heat derived from eddy covariance, satellite, and meteorological observations. Journal of Geophysical Research, 116, 1–16, http://dx.doi.org/10.1029/
2010JG001566 (G00J07).
Jung, M., Verstraete, M., Gobronz, N., Reichstein, M., Papale, D., Bondeau, A., et al. (2008).
Diagnostic assessment of European gross primary production. Global Change Biology,
14, 1–16, http://dx.doi.org/10.1111/j.1365-2486.2008.01647.x.
Jung, M., Vetter, M., Herold, M., et al. (2007). Uncertainties of modelling gross primary
productivity over Europe: A systematic study on the effects of using different drivers
and terrestrial biosphere models. Global Biogeochemical Cycles, 21, GB4021, http://dx.
doi.org/10.1029/2006GB002915.
Jung, M., & Zscheischler, J. (2013). A guided hybrid genetic algorithm for feature selection
with expensive cost functions. Procedia Computer Science, 18, 2337–2346, http://dx.
doi.org/10.1016/j.procs.2013.05.405.
Krinner, G., Viovy, N., de Noblet-Ducoudre, N., Ogee, J., Polcher, J., Friedlingstein, P., et al.
(2005). A dynamic global vegetation model for studies of the coupled atmosphere
biosphere system. Global Biogeochemical Cycles, 19, GB1015, http://dx.doi.org/10.
1029/2003GB002199.
Lasslop, G., Reichstein, M., Papale, D., Richardson, A. D., Arneth, A., Barr, A., et al. (2010).
Separation of net ecosystem exchange into assimilation and respiration using a
light response curve approach: critical issues and global evaluation. Global Change
Biology, 16, 187–208, http://dx.doi.org/10.1111/j.1365-2486.2009.02041.x.
Moffat, A. M., Beckstein, C., Churkina, G., Mund, M., & Heimann, M. (2010). Characterization of ecosystem responses to climatic controls using artificial neural networks.
Global Change Biology, 16, 2737–2749, http://dx.doi.org/10.1111/j.1365-2486.2010.
02171.x.
373
Myneni, R. B., Hoffman, S., Knyazikhin, Y., Privette, J. L., Glassy, J., Tian, Y., et al. (2002).
Global products of vegetation leaf area and fraction absorbed PAR from year one of
MODIS data. Remote Sensing of Environment, 83, 214–231.
O'Hagan, A. (2012). Probabilistic uncertainty specification: Overview, elaboration
techniques and their application to a mechanistic model of carbon flux.
Environmental Modelling and Software, 36, 35–48, http://dx.doi.org/10.1016/j.
envsoft.2011.03.003.
Oliveira, S., Oehler, F., San-Miguel-Ayanz, J., Camia, A., & Pereira, J. M. C. (2012). Modeling
spatial patterns of fire occurrence in Mediterranean Europe using Multiple Regression and Random Forest. Forest Ecology and Management, 275, 117–129, http://dx.
doi.org/10.1016/j.foreco.2012.03.003.
Olofsson, P., Lagergren, F., Lindroth, A., Lindström, J., Klemedtsson, L., Kutsch, W., et al.
(2008). Towards operational remote sensing of forest carbon balance across Northern
Europe. Biogeosciences, 5, 817–832, http://dx.doi.org/10.5194/bg-5-817-2008.
Papale, D., Reichstein, M., Aubinet, M., Canfora, E., Bernhofer, C., Longdoz, B., et al. (2006).
Towards a standardized processing of Net Ecosystem Exchange measured with eddy
covariance technique: Algorithms and uncertainty estimation. Biogeosciences, 3,
571–583, http://dx.doi.org/10.5194/bg-3-571-2006.
Papale, D., & Valentini, R. (2003). A new assessment of European forests carbon exchanges by eddy fluxes and artificial neural network spatialization. Global Change
Biology, 9, 525–535, http://dx.doi.org/10.1046/j.1365-2486.2003.00609.x.
Rahman, A. F., Sims, D. A., Cordova, V. D., & El-Masri, B. Z. (2005). Potential of MODIS EVI
and surface temperature for directly estimating per-pixel ecosystem C fluxes.
Geophysical Research Letters, 32, l19404, http://dx.doi.org/10.1029/2005GL024127.
Reichstein, M., Falge, E., Baldocchi, D., Papale, D., Aubinet, M., Berbigier, P., et al. (2005).
On the separation of net ecosystem exchange into assimilation and ecosystem respiration: Review and improved algorithm. Global Change Biology, 11, 1424–1439,
http://dx.doi.org/10.1111/j.1365-2486.2005.001002.x.
Running, S. W., Baldocchi, D. D., Turner, D. P., Gower, S. T., Bakwin, P. S., & Hibbard, K. A.
(1999). A global terrestrial monitoring network integrating tower fluxes, flask sampling, ecosystem modeling and EOS satellite data. Remote Sensing of Environment,
70, 108–127.
Running, S. W., Thornton, P. E., Nemani, R., & Glassy, J. M. (2000). Global terrestrial gross
and net primary productivity from the Earth Observing System. In O. E. Sala, R. B.
Jackson, H. A. Mooney, & R. W. Howarth (Eds.), Methods in Ecosystem Science
(pp. 44–57). New York: Springer-Verlag.
Schimel, S. D., Braswell, B. H., & Parton, W. J. (1997, Aug 5). Equilibration of the terrestrial
water, nitrogen, and carbon cycles. Proceedings of the National Academy of Sciences of
the United States of America, 94(16), 8280–8283.
Sims, D. A., Rahman, A. F., Cordova, V. D., El-Masri, B. Z., Baldocchi, D. D., Bolstad, P. V., et al.
(2008). A new model of gross primary productivity for North American ecosystems
based solely on the enhanced vegetation index and land surface temperature from
MODIS. Remote Sensing of Environment, 12, 1633–1646, http://dx.doi.org/10.1016/j.
rse.2007.08.004.
Sitch, S., Smith, B., Prentice, I. C., Arneth, A., Bondeau, A., Cramer, W., et al. (2003). Evaluation of ecosystem dynamics plant geography and terrestrial carbon cycling in the LPJ
dynamic global vegetation model. Global Change Biology, 9, 161–185, http://dx.doi.
org/10.1046/j.1365-2486.2003.00569.x.
Sjöström, M., Zhao, M., Archibald, S., Arneth, A., Cappelaere, B., Falk, U., et al. (2011). Evaluation of MODIS gross primary productivity for Africa using eddy covariance data.
Remote Sensing of Environment, 131, 275–286, http://dx.doi.org/10.1016/j.rse.2010.
12.013.
Vermote, E. F., & Vermeulen, A. (1999). MODIS algorithm technical background document — Atmospheric correction algorithm: Spectral reflectance (MOD09), version
4.0. http://modis.gsfc.nasa.gov/data/atbd/atbd_mod08.pdf
Wan, Z., Zhang, Y., Zhang, Q., & Li, Z. L. (2002). Validation of the land-surface temperature
products retrieved from Terra Moderate Resolution Imaging Spectroradiometer data.
Remote Sensing of Environment, 83, 163–180, http://dx.doi.org/10.1016/S00344257(02)00093-7.
Xiao, X., Zhang, Q., Braswell, B., Urbanski, S., Boles, S., Wofsy, S., et al. (2004). Modeling
gross primary production of temperate deciduous broadleaf forest using satellite images and climate data. Remote Sensing of Environment, 91, 256–270, http://dx.doi.org/
10.1016/j.rse.2004.03.010.
Xiao, J., Zhuang, Q., Law, B. E., Chen, J., Baldocchi, D. D., Cook, D. R., et al. (2010). A continuous measure of gross primary production for the conterminous United States derived from MODIS and AmeriFlux data. Remote Sensing of Environment, 114,
576–591, http://dx.doi.org/10.1016/j.rse.2009.10.013.
Yang, F., Ichii, K., White, M. A., Hashimoto, H., Michaelis, A. R., Votava, P., et al. (2007). Developing a continental-scale measure of gross primary production by combining
MODIS and AmeriFlux data through Support Vector Machine approach. Remote
Sensing of Environment, 110, 109–122, http://dx.doi.org/10.1016/j.rse.2007.02.016.
Zhao, M., Heinsch, F. A., Nemani, R. R., & Running, S. W. (2005). Improvements of the
MODIS terrestrial gross and net primary production global data set. Remote Sensing
of Environment, 95, 164–176, http://dx.doi.org/10.1016/j.rse.2004.12.011.
Zhao, M., Running, S. W., & Nemani, R. (2006). Sensitivity of Moderate Resolution Imaging
Spectroradiometer (MODIS) terrestrial primary production to the accuracy of meteorological reanalyses. Journal of Geophysical Research, 111, 1–13, http://dx.doi.org/10.
1029/2004JG000004 (G01002).