Remote Sensing of Environment 203 (2017) 71–89

Contents lists available at ScienceDirect

Remote Sensing of Environment

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

Uncertainty in plant functional type distributions and its impact on land T

surface models
A.J. Hartleya,⁎, N. MacBeanc,d, G. Georgievskib, S. Bontempse
a
Met Office Hadley Centre, FitzRoy Road, Exeter, EX1 3PB, United Kingdom
b
Max Planck Institute for Meteorology, Bundesstrasse 53, 20146 Hamburg, Germany
c
Laboratoire des Sciences du Climat et de l'Environnement, Université Paris-Saclay, F-91191 Gif-sur-Yvette, France
d
School of Natural Resources and the Environment, University of Arizona, 1064 E Lowell St, Tucson, AZ 85721, USA
e
Université catholique de Louvain, Place de l'Université 1, 1348 Louvain-la-Neuve, Belgium

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

Keywords: The spatial distribution and fractional cover of plant functional types (PFTs) is a key uncertainty in land surface
Land cover models (LSMs) that is closely linked to uncertainties in global carbon, hydrology and energy budgets. Land cover
Plant functional type is considered to be an Essential Climate Variable because changes in it can result in local, regional or global scale
Uncertainty impacts on climate. In LSMs, land cover (LC) class maps are converted to PFT fractional maps using a cross-
Vegetation distribution
walking (CW) table by prescribing the fraction of each PFT that occurs within each LC class. In this study we
Land surface model
Terrestrial biosphere
assess the largest plausible range of PFT uncertainty derived from remotely sensed LC maps produced under the
Biogeochemical cycles European Space Agency Land Cover Climate Change Initiative on simulations of land surface fluxes using 3
Carbon cycle leading LSMs. We evaluate the impact of uncertainty due to both LC classification algorithms, and CW proce-
Hydrological cycle dure, on energy, moisture and carbon fluxes in LSMs. We investigate the maximum plausible range of un-
Energy budget certainty deriving from both LC and CW within the context of a potential biomass scale (bare ground-grass-
shrub-tree), representing a gradient from low to high biomass PFTs. More specifically, plausible alternative land
cover maps and associated PFT fractional distributions were produced to prioritise low or high biomass vege-
tation in the LC classification (uncertainty in LC), and subsequently in the assignment of PFT fractions for each
LC class (uncertainty in CW), relative to a reference PFT distribution.
We examined the impact of PFT uncertainty on 3 key variables in the carbon, water and energy cycles (gross
primary production (GPP), evapo-transpiration (ET), and albedo), for 3 LSMs (JSBACH, JULES and ORCHIDEE)
at global scale. Results showed a greater uncertainty in PFT fraction due to CW as opposed to LC uncertainty, for
all three variables. CW uncertainty in tree fraction was found to be particularly important in the northern boreal
forests for simulated LSM albedo. Uncertainty in the balance between grass and bare soil fraction in arid parts of
Africa, central Asia, and central Australia was also found to influence albedo and ET in all models. The spread
due to PFT uncertainty for albedo was between 30 and 105% of inter-model uncertainty, for GPP between 20 and
90%, and for ET 0–30%. Each model had a different sensitivity to PFT uncertainty, for example, GPP in JSBACH
was found to have a much higher sensitivity to PFT uncertainty in the tropics than JULES and ORCHIDEE,
whereas the inverse was true for ET.
These results show that inter-model uncertainty for key variables in LSMs can be reduced by more accurate
representation of PFT distributions. Future efforts in land cover mapping should therefore be focused on re-
ducing CW uncertainty through better understanding of the fractional cover of PFTs within a land cover class.
Efforts to reduce LC uncertainty should particularly be focused on more accurate mapping of grass and bare soil
fractions in arid areas. In the context of Land Surface Models, these results demonstrate that prescribed vege-
tation distribution in models is a key source of uncertainty that is comparable to the spread between models for
key model state variables.

⁎
Corresponding author.
E-mail address: andrew.hartley@metoffice.gov.uk (A.J. Hartley).

http://dx.doi.org/10.1016/j.rse.2017.07.037
Received 3 October 2016; Received in revised form 10 July 2017; Accepted 31 July 2017
Available online 20 September 2017
0034-4257/ Crown Copyright © 2017 Published by Elsevier Inc. All rights reserved.
A.J. Hartley et al.
1. Introduction
Land cover (LC) is considered by the Global Climate Observing
System (GCOS) as an Essential Climate Variable that is used to monitor
natural and anthropogenic changes to the land surface. It is a key
component of the earth system that influences carbon, moisture, energy
and momentum exchanges between the land surface and the atmo-
sphere (Jung et al., 2007; Poulter et al., 2011; Sterling et al., 2013).
Land Surface Models (LSMs) are the land component of numerical
weather prediction, climate and Earth System Models. Most LSMs as-
sign fixed vegetation, derived from global land cover maps (e.g.
(Loveland and Belward, 1997; Wilson and Henderson-Sellers, 1985)),
while some also simulate dynamic vegetation that responds to and in-
teracts with climate, anthropogenic land use, carbon dioxide, hydrology
and other aspects of the earth system (e.g. (Cox, 2001; Cramer et al.,
2001; Prentice et al., 1992; Sitch et al., 2003)). Commonly, LSMs re-
present global vegetation in terms of a small set of Plant Functional
Types (PFTs; the exact number of which differ between LSMs) and need
as input the spatial coverage of each PFT in each model grid cell, ex-
pressed as a grid cell fraction. The LSMs that can simulate dynamic
vegetation can also be configured with fixed vegetation. LSMs represent
processes in the earth system (such as photosynthesis and transpiration)
via equations that are common for all PFTs which have fixed parameter
values that differ according to PFT (e.g. the relationship between leaf
stomatal closure and vapour pressure deficit). LC information is com-
monly used by LSM modelling groups to determine the spatial dis-
tribution of PFTs via a cross-walking(CW) procedure (described in more
detail in Section 1.2) that assigns PFT fractions for each LC class. While
mapping the global fractional coverage of PFTs directly from satellite
radiances would be preferable (because the uncertain CW procedure
would become unnecessary), this approach is not currently used by any
LSM modelling group because of difficulties to distinguish the spectral
properties of different functional types and mixing of different func-
tional types within a pixel (see Section 4 for further discussion).
However, different groups have constructed their own PFT maps, based
on the unique set of PFTs required by each model and on different
underlying land cover map products, potentially leading to incon-
sistencies between LSMs.
The spatial distribution of PFTs is associated with uncertainties in
three important aspects of LSMs: budgets of carbon (Ballantyne et al.,
2015), moisture (Boisier et al., 2014) and energy (Hoffmann and
Jackson, 2000; Mahmood et al., 2014). Annual reporting of the global
carbon budget by (Le Quéré et al., 2015) using both a book keeping
method and LSMs has shown that uncertainty in the amount of carbon
released by land use change was 0.5 PgC/year (1σ) in 2014, with the
land carbon uptake varying by an additional 0.9 PgC/year. These
ranges are influenced by uncertainties in the reporting or detection of
land use change, and by uncertainties in the vegetation type and carbon
stored in the vegetation before the change occurred (Anav et al., 2013;
Houghton et al., 2012). Furthermore, uncertainties in the global land
carbon uptake are related to PFT distributions via uncertainties in the
rate of primary production (Quaife et al., 2008), soil and vegetation
carbon storage (Anav et al., 2013; Brovkin et al., 2013), and plant and
soil CO2 respiration.
Moisture budgets are also sensitive to uncertainties in PFT dis-
tribution. For example, (Boisier et al., 2014) showed that LSM simu-
lations of evapotranspiration (ET) are poorly constrained by observa-
tions, and concluded that reductions in historical simulated ET
uncertainty can be made by improving historical land cover re-
constructions. In global terms, changes in land-atmosphere moisture
fluxes are governed by two competing anthropogenic processes. Firstly,
the location and magnitude of forest conversion to agriculture reduces
global ET as a result of reducing leaf area and increases surface runoff
due to reduced interception of water by vegetation (Findell et al., 2007;
Gordon et al., 2005; Sterling et al., 2013). Secondly, the global ex-
pansion of irrigated agriculture during the 20th Century has been
72
Remote Sensing of Environment 203 (2017) 71–89
shown to increase the amount of ET, due to greater moisture avail-
ability for photosynthesis and surface evaporation (Gordon et al., 2005;
Puma and Cook, 2010). Regionally, it has also been shown in land-
atmosphere coupled simulations that these human-induced changes to
the moisture budget may have an impact on the variability (Zeng,
1999), location (Hartley et al., 2016; Knox et al., 2011), and strength
(Feddema et al., 2005b) of tropical monsoon systems in South America,
Africa and South East Asia.
Energy budgets can also be directly influenced by the spatial dis-
tribution of PFTs. Forests, in comparison to cropland and grasslands,
tend to exert a cooling effect on regional climate in the tropics and
temperate regions through evaporative cooling, whereas boreal forests
tend to exert a warming effect due to lower surface albedo (Bonan,
2008; Luyssaert et al., 2014; Zeng and Neelin, 1999). This has been
shown by studies that have used both satellite observations (Alkama
et al., 2016) and coupled land-atmosphere models (Berbet and Costa,
2003; Betts, 2001; Boisier et al., 2012) to show the strong local positive
radiative effects of replacing forest cover with cropland or pasture.
Conversely, (Betts, 2000) showed that boreal afforestation reduced
surface albedo by 0.1 to 0.3, leading to a positive radiative forcing of
10–20 Wm− 2, which is higher than the equivalent radiative cooling
due to increasing carbon sequestration. Despite a clear biogeophysical
sensitivity of LSMs to LULCC at regional scales, very little work exists
on the impact of uncertainty in present-day PFT distributions on the
land surfaces fluxes of energy, moisture and carbon in LSMs. One ex-
ception to this is the work by (Feddema et al., 2005a) who show that
while average global temperature model sensitivity to present day ve-
getation uncertainty is only 0.21 K, a much larger uncertainty range of
up to 5 K can be found at regional scales.
The accuracy of PFT fractional coverage in each model grid cell is
therefore an important component of LSMs that can have a significant
impact on simulations of carbon, water, and energy fluxes.
Understanding and reducing the uncertainty in land cover-derived PFT
spatial distributions should lead to more confident predictions of how
ecosystem services have responded, and will respond in the future, to
the combined impacts of climate change and land use and land cover
changes (LULCC).
A principle aim of the European Space Agency (ESA) Land Cover
Climate Change Initiative (LC_CCI) is to reduce LSM uncertainty
through the use of spatially and temporally consistent LC maps that are
created from satellite-derived surface reflectance and ground-truth
observations. Part of this initiative has involved interaction between
land surface modellers and LC mapping experts in the earth observation
community in order to provide PFT maps for each individual LSM,
based on the same underlying LC data. This process of expert interac-
tion concluded that while it is currently not possible to accurately map
PFTs directly from satellite observations, the approach of deriving PFT
fractions from LC maps via a LC-to-PFT conversion (“cross-walking”)
table was a viable approach. In the process of creating PFT fractions for
use in LSMs, there are two key sources of uncertainty, explained in
detail below.
1.1. Land cover mapping uncertainty
In the context of LSMs, the requirements of LC maps are very pre-
cisely defined and sometimes are not consistent with other applications
for LC maps. LSMs usually consider only 1 vegetation level, with no
understory vegetation, meaning that the PFT fraction refers to the
crown cover of each PFT at the point in the season when maximum leaf
area occurs. This is because LSMs commonly use either seasonally
varying leaf area information or a phenological model to define vege-
tation seasonality, therefore, for climate applications, it is not necessary
to incorporate this temporal information into land cover classes. This
ambiguity and lack of communication between climate and LC scien-
tists can lead to considerable errors and uncertainties in land cover-
derived PFT fractions.
A.J. Hartley et al.
Uncertainty also exists in the classification approaches used to dis-
criminate, identify and categorise LC classes from satellite-derived
surface reflectance products. This uncertainty may come from a pro-
pagation of the uncertainty of the input satellite products, but can also
be related to the performance of the classification approaches and the
natural intra- and inter-annual variability of the earth system. An as-
sessment of 3 commonly used global LC maps revealed a mapping ac-
curacy of between 67% and 75% (Fritz et al., 2011), although dis-
agreement between maps is also high (Fritz and See, 2008). While
various attempts have been made to combine such maps into hybrid
products that reduce uncertainty (Fritz and See, 2005; Jung et al.,
2006), there currently haven't been any efforts to produce LC maps of
plausible alternative classes, with associated likelihoods.
1.2. Cross-walking uncertainty
The process of converting LC classes to PFTs, called the “cross-
walking” (CW) procedure, uses the LC class legend descriptions, expert
knowledge, and ancillary information to determine the fraction of each
PFT that occurs within a given LC class (Poulter et al., 2011, 2015). For
example, in (Poulter et al., 2015), where the LC class “Tree cover,
broadleaf, evergreen, closed to open” occurred in a 300 m pixel, 90% of
that area was assigned to Broadleaf Evergreen Tree PFT, 5% to
Broadleaf Evergreen Shrub PFT, and 5% to Broadleaf Deciduous Shrub
PFT. LC classes typically follow a precisely-defined thematic descrip-
tion, such as the Food and Agricultural Organisation of the United
Nations (FAO) Land Cover Classification System (LCCS) (Di Gregorio
and Jansen, 2005). The LCCS allows the user to develop a LC class le-
gend that describes the fraction of major cover types such as tree, shrub,
grass, bare soil and water. These cover types have important physiog-
nomic differences that are important for LSMs, for example, differences
in vegetation structure and density affects the roughness of the surface
and the surface albedo, which can drive the exchange of energy, mo-
mentum and carbon between the land surface and the atmosphere.
Furthermore, the LCCS allows the definition of a legend that describes
the fractional cover of a vegetation type as a broad range, such as open
cover (10–40%), closed-to-open (30–70%), or closed (60–100%).
Clearly, these ranges are designed to allow a certain amount of flex-
ibility from a LC mapping perspective, but in LSMs these wide ranges
may create considerable uncertainty in key aspects of land-atmosphere
interactions. Consequently, when populating the CW matrix by in-
ferring PFT fractions for a given land cover class, the climate user must
choose a fraction from a wide range that may span up to 40%. This CW
uncertainty may be further complicated by regional variations in PFT
fractions within a LC class that occurs in many places around the world,
such as broadleaf deciduous tree cover that could have different tree
fractions depending on continent and bioclimatic zone. Clarification of
which fraction to use for which region of the world is therefore the
responsibility of expert judgement.
Ancillary information is used for defining PFT distributions that
cannot be accurately defined using LC maps. It is commonly used to
make the distinction between grasses that use the C3 and C4 photo-
synthetic pathways, or between PFTs that are specific to a particular
bioclimatic zone. Currently, it is not possible to distinguish between C3
and C4 grasses using existing remote sensing techniques, therefore
climate users typically create a grassland fraction from LC, and use an
ancillary dataset to split this fraction into C3 and C4 grass fractions.
This is done either by using a pre-existing map of C4 grass fraction such
as (Still et al., 2003) or by using temperature thresholds, such as in
(Poulter et al., 2011), derived from the Koeppen-Geiger climate clas-
sification (Kottek et al., 2006). Such choices bring additional un-
certainty, for example, the (Still et al., 2003) data uses agricultural
statistics and crop fractional cover (Ramankutty and Foley, 1998) to
determine the fraction of C4 grasses. This highlights the need for more
ground-based observation to improve verification, and raises questions
of poor agreement between global cropland maps (Fritz et al., 2011),
73
Remote Sensing of Environment 203 (2017) 71–89
and the validity of using agricultural statistics for differing time per-
iods. Furthermore, dependence on climate thresholds introduces un-
certainty in choice of climate data, which may be particularly uncertain
in sub-tropical areas (Harris et al., 2014), and in choice of threshold.
While the use of the 22 °C crossover temperature threshold to distin-
guish between C3 and C4 grasses has been well studied (Collatz et al.,
1998) and used in the ORCHIDEE LSM (Poulter et al., 2011), it is not
universally used, with the LPJ LSM using a threshold of 15.5 °C (Sitch
et al., 2003) to identify C4 grass. In summary, at each step in the
process of generating PFT maps for LSMs, the individual LSM modeller,
often not an expert in LC mapping, has had to make decisions based on
limited knowledge and availability of information sources to the in-
dividual. It is therefore no surprise that LC and fractional coverage
maps utilised across modelling groups can vary significantly.
In this study, we investigate the sensitivity of three offline LSMs to
the maximum plausible range of PFT uncertainty derived from two
explicitly quantified sources of error as described above: LC class
mapping uncertainty and uncertainty in the cross-walking (CW) pro-
cedure used to prescribe PFT fractions within each land cover class. To
our knowledge, this is the first systematic assessment of LSM sensitivity
to the maximum plausible range of uncertainties in observed vegetation
cover.
The aim is firstly understand the impact of LC mapping and CW
uncertainty on the PFT fractional distributions that are used in LSMs.
Secondly, we aim to quantify the impact of this PFT uncertainty on the
spatio-temporal patterns of carbon, water and energy fluxes simulated
by three LSMs. This includes inter-comparison of model sensitivity to
PFT uncertainty, and comparison of PFT uncertainty impacts in the
context of uncertainty between models.
2. Data and methods
2.1. Reference dataset
The procedure for creating PFT fractional distributions from sa-
tellite-derived surface reflectance and ground truth observations is
shown in detail in Fig. 1. A classification algorithm is used to categorise
satellite-derived radiances according to the Land Cover Classification
System (LCCS; (Di Gregorio and Jansen, 2005)) legend categories. This
is done using both machine learning and K-Means clustering algo-
rithms, that rely on a large set of existing national and regional land
cover maps which were selected as the most accurate ones available for
a given region, with the highest spatial resolution and with a suitable
legend (LC_CCI ATBD, 2013). This approach provides a confidence level
that a given 300 m pixel should be assigned to each land cover class,
with the most likely class chosen. This approach was used to create the
LC map produced by the European Space Agency (ESA) Land Cover
Climate Change Initiative (LC_CCI) for 2010, at 300 m spatial resolution
(version 1.4; available at http://maps.elie.ucl.ac.be/CCI/viewer/),
which was used in the present study as a reference dataset (henceforth
refLC). The LC dataset is converted to PFT fractional distributions by
assigning pre-defined, LC class-specific, PFT fractions to each 300 m
pixel, following the cross-walking table approach that was developed
for this LC dataset by (Poulter et al., 2015). This CW table is also used as
a reference in the present study (see “ref” columns in Table 2). The
resulting PFT fractions are then aggregated from 300x300m to 2 × 2°
(approximately 200 km at equator) spatial resolution, using the LC_CCI
spatial averaging tool documented in (Poulter et al., 2015). Model-
specific post-processing was then conducted, which is described in de-
tail in Section 2.3.
2.2. Quantification of PFT uncertainty using a biomass scale
The key areas of uncertainty within the processing chain of the re-
ference dataset (as discussed above, and shown by dotted lines in Fig. 1)
are firstly the classification algorithm, which allocates LC class labels to
A.J. Hartley et al.
pixels with varying levels of reliability, meaning that the probability of
an alternative class to that chosen in refLC may be high for some pixels.
A second key area of uncertainty is the cross-walking table, where the
fraction of a PFT in a given LC class may be uncertain, particularly for
mixed vegetation classes. We therefore explored the possible cause and
range of uncertainty in the LC classification and CW table conversion
factors in order to develop plausible alternative PFT fractional dis-
tributions.
To do this, the concept of biomass was used because it is closely
linked to the magnitude of many aspects of land-atmosphere interac-
tions, such as carbon exchange via leaf area, photosynthesis and auto-
trophic respiration, water flux via transpiration and canopy evapora-
tion, and energy exchange via surface albedo and sensible and latent
heat flux. A simple conceptual biomass scale is established, where the
highest biomass category is defined as “tree”, then followed by “shrub”,
“grass”, “moss and lichen” and “bare” (Fig. 2). Each of these levels is
associated with a set of LC classes from the Land Cover CCI legend
(Table 1) and this grouping is used to decide whether to replace a LC
class with an alternative class, or to decide whether to alter the fraction
of a PFT that occurs in a LC class in the CW table. To be clear, actual
biomass values were not specifically quantified; the biomass concept
was used purely with the aim of providing a simple framework to
quantify the minimum and maximum plausible range of LC class and
CW uncertainty. The following two sub-sections describe how the
conceptual biomass scale is used to achieve this.
2.2.1. Land cover class uncertainty
In order to develop plausible alternative LC maps arising from LC
uncertainty, we use the biomass scale shown in Fig. 2 to develop
‘minimum biomass’ (minLC) and ‘maximum biomass’ (maxLC) LC maps.
These maps are derived from the reference LC map (refLC) in which an
alternative LC class is available for all pixels that are classified with a
confidence level lower than 85%(LC_CCI ATBD, 2013). Where this
criterion is met, we deem an alternative LC class to be plausible if it has
a confidence level higher than 5%. The confidence level used here is an
output of the CCI LC classification chain which is based on two different
algorithms. The confidence level is derived from the classification
probability in the case of the machine learning algorithm and from the
distance of a pixel to the cluster mean in the case of the K-Mean clus-
tering algorithm (CECR).
If a plausible alternative class is available, and it is lower down the
biomass scale (when making the minLC map), or further up the scale
(when making the maxLC map) shown in Fig. 2 and Table 1, than the
reference LC class (refLC), then this alternative is used in the resulting
minLC or maxLC map. For example, if a grid cell in the refLC map is
74
Remote Sensing of Environment 203 (2017) 71–89
Fig. 1. Data flow involved in producing the 5
cases of PFT fractional uncertainty. Land Cover
uncertainty is accounted for at the stage of the
classification algorithm, and subsequently cross-
walking uncertainty is accounted for before the
resulting PFT fractions are aggregated to the
model grid resolution. At this stage, post-proces-
sing may involve using ancillary data such as C4
grass distribution to separate grass cover between
functional types, or thermal thresholds to sepa-
rate tree cover between biomes.
classified as class 130 (grass) with a confidence level of 55%, but an
alternative class 120 (shrub) has a confidence level of 35%, then, if we
are making a maxLC map, we would choose class 120. If we are making
a minLC map, then the pixel remains as class 130. In the case of a grid
cell where the refLC map is classified as class 130 (grass) with a con-
fidence level of 92%, no plausible alternative LC class is selected and
the refLC class is retained.This approach results in land cover maps that
identify the minimum and maximum plausible extent of the main cover
types (e.g. forests, shrub, grass).
2.2.2. Cross walking uncertainty
In order to develop alternative PFT fractional distributions arising
from CW uncertainty, we again use the biomass scale to develop
plausible alternative CW tables (see Table 2) for ‘minimum biomass’
(minCW) and ‘maximum biomass’ (maxCW). The biomass scale was
used to alter the fractions of PFTs that occur in each LC class in the
reference CW table (refCW) within plausible ranges. A plausible range is
derived either from the LCCS legend description (Di Gregorio and
Jansen, 2005), or by further constraining this range with expert
knowledge of earth observation scientists. Changes to the refCW table
are achieved for a given LC class by increasing the fractions of PFTs
lower down the biomass scale (minCW), or by increasing the fractions of
PFTs further up the biomass scale (maxCW), always keeping within
plausible ranges (see Table 2 for minCW, refCW and maxCW fractions).
For example, class 61 (Broadleaf deciduous tree cover, closed (> 40%))
has a plausible range of 40–100% according to the LCCS, and in refCW
it is converted to 70% broadleaf deciduous tree (BDT) PFT, 15%
broadleaf deciduous shrub PFT and 15% natural grass (NG). In minCW,
the BDT PFT fraction is reduced to 40%, the minimum fractional cover
permitted by the LCCS description, and NG fraction increased to 45%,
whereas in maxCW, 100% of the grid cell is BDT PFT.
2.3. LSM modelling protocol
For the LSM simulations, these two sources of uncertainty in PFT
fractional distributions were subsequently combined, using the meth-
odology shown in Fig. 1, in order to characterise the extremes of LC and
CW uncertainty. This was done by using the minLC map with the minCW
table (to create the PFT fractional distribution we term as min-
LC_minCW), and similarly using maxLC with maxCW (to create
maxLC_maxCW). In order to allow the separation of the effects of LC and
CW uncertainty, we also created two scenarios that combined the re-
ference CW table the two extremes of LC (minLC_refCW and maxL-
C_refCW). Finally, a control dataset was created using the reference LC
and CW (refLC_refCW). The cases refLC_maxCW and refLC_minCW were
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

Fig. 2. Cartoon to show how PFT uncertainty is quantified in this study using a conceptual biomass scale. A classification algorithm uses satellite-derived radiances (1) to assign
probabilities of plausible LC classes for each pixel. If making the reference LC map (refLC), the most likely class is assigned to the pixel. However, if making a ‘maximum (minimum)
biomass’ LC map, and an alternative class is higher up (lower down) the conceptual biomass scale, the class would be changed. For example, a pixel classified as shrubland in refLC may
have plausible but lower probability alternative classes of grassland or open tree (2). However, if the opposite occurs, or there is no alternative LC class, the refLC class is again used in the
alternative map. Uncertainty in the CW fraction of PFTs assigned to a LC class is shown in blue (3) for three LC classes within the conceptual biomass scale (Grassland, Shrubland and
Broadleaf Evergreen Tree Closed-Open). Here, to make a ‘maximum (minimum) biomass’ map, the PFT fractions for a class were altered within plausible ranges by favouring PFTs that are
further up (lower down) the conceptual biomass scale. For example, the LC class Broadleaf Evergreen Tree Closed-Open contains 90% tree PFT, and 10% grass PFT in refCW (see Table 2
for more detail). Under minCW (maxCW), this changes to 70% (100%) tree and 30% (0%) grass. (Image of satellite ©AMT4SentinelFRM).

not explored in order to reduce complexity of the study and because the
impact of maxCW (or minCW) would be possible to derive by sub-
tracting results for maxLC_refCW from maxLC_maxCW (or minLC_refCW
from minLC_minCW).
We use three offline LSMs (JSBACH v3.10, JULES v4.2 and
ORCHIDEE r3576), each of which form the land surface component of
three major Earth System Models (ESMs; MPI-ESM, UK-ESM, and IPSL-
ESM respectively) that will be used in the 6th Coupled Model Inter-
comparison Project (CMIP6). In using offline LSMs, rather than coupled
ESMs, we avoid additional uncertainty associated with other model
components, such as the atmosphere and oceans, making it easier to
relate PFT uncertainties to impacts on surface fluxes. While each LSM
uses a different number of PFT classes, they each can be related back to
the CW table of (Poulter et al., 2015), as shown in Table S1 (Supple-
mentary Information). Each LSM was run for the historical period
1979–2010, using the WFDEI (WATCH Forcing Data methodology ap-
plied to ERA-Interim data) meteorological forcing data developed by
(Weedon et al., 2011, 2014). WFDEI forcing data were resampled from
the native 0.5° spatial resolution to 2° using a conservative remapping
method (Jones, 1999) to aggregate the forcing fields (precipitation,
temperature, wind, downwards shortwave and longwave radiation
fluxes and pressure). Special care was taken for the treatment of specific
humidity. Like (Weedon et al., 2011) we followed the methods of
(Cosgrove et al., 2003) in order to conserve moisture in the air and
avoid supersaturation. Relative humidity was calculated at 0.5°, and
then aggregated at 2°. Specific humidity at 2° was then calculated from
relative humidity. Both JULES and ORCHIDEE use the data in their
original 3 h temporal resolution, while JSBACH makes use of the data
accumulated to daily intervals.
A spin-up of the terrestrial carbon cycle was conducted by cycling
over WFDEI forcing repeatedly for the years 1979–1983, with 1979 CO2
concentrations, until net ecosystem exchange reached an equilibrium
(fluctuating around 0). For JSBACH, the spin-up period was 3500 years,
ORCHIDEE 3300 years, and JULES 1000 years. This ensures that the
carbon cycle is in equilibrium with the 1979 climate at the start of the
simulation period. Simulations are then run for the period 1979–2010
with transient CO2 concentrations taken from global historical ob-
servations.
In order to evaluate the impact of the 5 different PFT fractional
distributions (minLC_minCW, minLC_refCW, refLC_refCW, maxLC_refCW,
maxLC_maxCW), we show results for model state variables that are
closely related to exchanges of carbon, water, and energy between the
land and the atmosphere. For the carbon cycle, we analyse gross pri-
mary production (GPP); for the water cycle, we analyse surface eva-
potranspiration (ET) flux; and for energy fluxes, we analyse the impact
of PFT uncertainty on surface shortwave albedo. Observational studies
have previously shown these variables to be sensitive to changes in land
use (e.g. (Jung et al., 2010; Wang et al., 2009; Zhao and Jackson,
2014)).
3. Results and discussion
3.1. Spatial distribution of PFT uncertainty
Uncertainty in PFT distribution is not globally uniform, and the
locations of LC uncertainty are not necessarily spatially consistent with
CW uncertainty. Tree cover uncertainty was found to be greatest in the
northern boreal zone (Fig. 3), north of 50°N, where tree cover un-
certainty in PFT fraction ranged from − 40% to + 60% in relation to
the reference (Fig. 3 middle row). This represented a large variability in
land area of tree fraction (Fig. 4), ranging from 600 to 2000 × 109 m2,
and was found to be mostly related to CW uncertainty in the boreal
forests of Eurasia and North America. In this region, PFT uncertainty in
tree cover is also directly related to CW uncertainty in grass cover and

75
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

Table 1 reference PFT distributions. Under PFT distributions that maximise

Each land cover class within the Land Cover CCI legend is assigned a class within a biomass, LC and CW uncertainty was found to contribute in equal
biomass hierarchy for use in assessing classification uncertainty. The highest biomass is
measure to this uncertainty, but in PFT distributions that minimise
tree, followed by shrub, grass, moss and lichen and then bare.
biomass, LC uncertainty was found to be the main cause (Fig. 4).
LC class Description Biomass In tropical forests (between 20°S and 10°N), CW uncertainty was
category found to be the main contributing factor to tree fraction uncertainty.
Here, the reference PFT distribution assigned a fraction of 90% tree
10 Cropland, rainfed Grass
11 Cropland, Herbaceous cover Grass cover for the evergreen broadleaved tree cover class, however, CW
12 Cropland, Tree or shrub cover Shrub uncertainty derived from the LCCS legend description ranged from 70%
20 Cropland, irrigated or post-flooding Grass (minCW) to 100% (maxCW) tree cover. In comparison to this range, LC
30 Mosaic cropland (> 50%)/natural vegetation (tree, Grass uncertainty was relatively low. This highlights that efforts to improve
shrub, herbaceous cover) (< 50%)
land cover information for use in climate science should have a greater
40 Mosaic natural vegetation (tree, shrub, herbaceous Grass
cover) (> 50%)/cropland (< 50%) focus on reducing uncertainty in the fractional coverage of PFTs within
50 Tree cover, broadleaved, evergreen, closed to open Tree a land cover class in the CW table, and on identifying regional varia-
(> 15%) tions in these assumptions, rather than on improving the mapping ac-
60 Tree cover, broadleaved, deciduous, closed to open Tree
curacy of LC classes.
(> 15%)
61 Tree cover, broadleaved, deciduous, closed Tree Finally, we also found uncertainty in PFT distributions to be high for
(> 40%) grass cover in tropical savannahs, particularly where mosaic land cover
62 Tree cover, broadleaved, deciduous, open (15–40%) Shrub classes exist. Mosaic land cover classes tend to have a broad mix of all
70 Tree cover, needleleaved, evergreen, closed to open Tree major natural or anthropogenic vegetation types (tree, shrub or grass,
(> 15%)
possibly mixed with agriculture). Between 15°S and 5°N, uncertainty is
71 Tree cover, needleleaved, evergreen, closed Tree
(> 40%) mostly due to LC uncertainty, which is perhaps an indication of diffi-
72 Tree cover, needleleaved, evergreen, open Shrub culties in separating the spectral signature of subsistence agriculture
(15–40%) from that of natural vegetation in this region, a factor which may
80 Tree cover, needleleaved, deciduous, closed to open Tree
contribute towards the high frequency of mosaic classes in these areas.
(> 15%)
81 Tree cover, needleleaved, deciduous, closed Tree
However, notably between 5°N and 15°N, particularly in West Africa
(> 40%) and India, the minLC PFT distribution increases grass fraction (Fig. 4),
82 Tree cover, needleleaved, deciduous, open Shrub whereas minCW reduces it. We suggest that this variation is also related
(15–40%) to mosaic classes, and is therefore an example of the uncertainties in-
90 Tree cover, mixed leaf type (broadleaved and Tree
volved in converting land cover classes with broad descriptions (such as
needleleaved)
100 Mosaic tree and shrub (> 50%)/herbaceous cover Tree mosaic classes) into PFT fractions for use in LSMs using a CW table.
(< 50%)
110 Mosaic herbaceous cover (> 50%)/tree and shrub Grass 3.2. Impacts of PFT uncertainty on model state variables
(< 50%)
120 Shrubland Shrub
121 Shrubland evergreen Shrub
3.2.1. Short wave albedo
122 Shrubland deciduous Shrub Globally, the main impact of uncertainty in PFT distribution on
130 Grassland Grass short-wave albedo of the land surface was found to occur in the
140 Lichens and mosses Moss/Lichens northern boreal zone for all models (Fig. 5). In these locations, an in-
150 Sparse vegetation (tree, shrub, herbaceous cover) Bare
crease (decrease) in tree fractional cover under a maxLC_maxCW
(< 15%)
151 Sparse tree (< 15%) Bare (minLC_minCW) PFT distribution lead to a decrease (increase) in
152 Sparse shrub (< 15%) Bare shortwave albedo in all models. The spatial location of albedo impacts
153 Sparse herbaceous cover (< 15%) Bare across the boreal zone (Fig. 5) was also found to vary depending on the
160 Tree cover, flooded, fresh or brakish water Tree source of PFT distribution uncertainty, and depending on whether
170 Tree cover, flooded, saline water Tree
180 Shrub or herbaceous cover, flooded, fresh/saline/ Grass
biomass was minimised or maximised. For example, in JSBACH and
brakish water JULES, minLC leads to albedo increases between 60°E and 90°E in
190 Urban areas Bare Western Boreal Russia, whereas under maxLC, albedo decreases were
200 Bare areas Bare most strongly found in Eastern Boreal Russia between 90°E and 140°E.
201 Consolidated bare areas Bare
The impacts of CW uncertainty on albedo were found to occur over a
202 Unconsolidated bare areas Bare
210 Water bodies n/a larger area in the Boreal zone than for LC uncertainty, but the main
220 Permanent snow and ice n/a regions of change for minCW and maxCW matched those for minLC
(Eastern Boreal Russia) and for maxLC (Western Boreal Russia).
The variability in global mean annual albedo due to PFT distribu-
LC uncertainty in bare cover, especially to the north of 60°N, where less tion uncertainty (i.e. the spread across all scenarios) was found to be
tree cover under a minCW PFT distribution leads to increases in grass much greater than the inter-annual variability for each scenario for all
cover, and increases in bare soil fraction. three models (Fig. 6). We also find that for both JULES and JSBACH the
Sparsely vegetated areas in semi-arid and arid regions such as albedo response due to minCW, minLC and maxLC changes in equal
Central Asia, Central Australia, Sahel, Namib desert, Nevada desert and amounts, whereas maxCW PFT uncertainty had a greater impact on
Patagonia were found to have large uncertainty in grass and bare soil albedo (Table 3 and Fig. 6). This may be accounted for by high un-
fractions. This is shown in Fig. 4 between 35°S and 15°S, and between certainty in the maximum tree cover area (maxLC_maxCW) around
30°N and 45°N where uncertainty in PFT distributions is largely driven 60°N, as shown in Fig. 4.
by CW uncertainty in grass and bare soil fractions. In these locations, Uncertainty in PFT distributions was also found to have an impact
the maxLC_maxCW PFT distribution results in a larger grass area, and on the seasonal cycle of albedo in the boreal zone during the winter
lower bare soil area. (Fig. 7) in JSBACH, JULES and ORCHIDEE, which was related to un-
Tree fraction uncertainty in continental SE Asia between 20°N and certainty in the tree fraction in this region. All three models have a
40°N (Southern China, Vietnam, Cambodia, Laos) was found to be re- multi-layered snow scheme, meaning that snow can settle on both the
latively large, ranging from −30% to + 50% (Fig. 3) relative to the tree canopy and on the ground, with the snow in the canopy melting

76
 Table 2
Cross-walking table of (Poulter et al., 2015) modified to account for uncertainty in PFT fractions. BLE = Broadleaf Evergreen, BLD = Broadleaf Deciduous, NLE = Needleleaf Evergreen, NLD = Needleleaf Deciduous, Wtr = Inland Water,
Sn = Permanent Snow and Ice.

ID Description BLE tree BLD tree NLE tree NLD tree BLE shrub BLD shrub NLE shrub NLD shrub Natural grass Managed grass Bare soil Wtr Sn
A.J. Hartley et al.

Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Ref Ref
Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm

0 No data 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
10 Cropland, 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 60 100 100 40 0 0 0 0
rainfed
11 Herbaceous 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 60 100 100 40 0 0 0 0
cover
12 Tree or 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 30 50 50 0 0 0 0 0 0 0 0 0 30 50 50 40 0 0 0 0
shrub cover
20 Cropland, 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 100 100 100 0 0 0 0 0
irrigated or
post-
flooding
30 Mosaic 0 5 10 0 5 10 0 0 0 0 0 0 3.3 5 5 3.3 5 5 3.4 5 5 0 0 0 15 15 15 75 60 50 0 0 0 0 0
cropland
(> 50%)/
natural
vegetation
(tree, shrub,
herbaceous
cover)
(< 50%)
40 Mosaic 0 5 15 0 5 15 0 0 0 0 0 0 7.5 7.5 7.5 10 10 10 7.5 7.5 7.5 0 0 0 25 25 25 50 40 20 0 0 0 0 0

77
natural
vegetation
(tree, shrub,
herbaceous
cover)
(> 50%)/
cropland
(< 50%)
50 Tree cover, 70 90 100 0 0 0 0 0 0 0 0 0 7.5 5 0 7.5 5 0 0 0 0 0 0 0 15 0 0 0 0 0 0 0 0 0 0
broad-
leaved,
evergreen,
closed to
open
(> 15%)
60 Tree cover, 0 0 0 40 70 100 0 0 0 0 0 0 0 0 0 15 15 0 0 0 0 0 0 0 45 15 0 0 0 0 0 0 0 0 0
broad-
leaved,
deciduous,
closed to
open
(> 15%)
61 Tree cover, 0 0 0 40 70 100 0 0 0 0 0 0 0 0 0 15 15 0 0 0 0 0 0 0 45 15 0 0 0 0 0 0 0 0 0
broad-
leaved,
deciduous,
closed
(> 40%)
62 Tree cover, 0 0 0 15 30 40 0 0 0 0 0 0 0 0 0 25 25 25 0 0 0 0 0 0 50 35 35 0 0 0 10 10 0 0 0
(continued on next page)
Remote Sensing of Environment 203 (2017) 71–89
 Table 2 (continued)

ID Description BLE tree BLD tree NLE tree NLD tree BLE shrub BLD shrub NLE shrub NLD shrub Natural grass Managed grass Bare soil Wtr Sn

Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Ref Ref
A.J. Hartley et al.

Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm

broad-
leaved,
deciduous,
open
(15–40%)
70 Tree cover, 0 0 0 0 0 0 40 70 100 0 0 0 5 5 0 5 5 0 5 5 0 0 0 0 45 15 0 0 0 0 0 0 0 0 0
needle-
leaved,
evergreen,
closed to
open
(> 15%)
71 Tree cover, 0 0 0 0 0 0 40 70 100 0 0 0 5 5 0 5 5 0 5 5 0 0 0 0 45 15 0 0 0 0 0 0 0 0 0
needle-
leaved,
evergreen,
closed
(> 40%)
72 Tree cover, 0 0 0 0 0 0 15 30 40 0 0 0 0 0 0 5 5 5 5 5 5 0 0 0 45 30 30 0 0 0 30 30 20 0 0
needle-
leaved,
evergreen,
open

78
(15–40%)
80 Tree cover, 0 0 0 0 0 0 0 0 0 40 70 100 5 5 0 5 5 0 5 5 0 0 0 0 45 15 0 0 0 0 0 0 0 0 0
needle-
leaved,
deciduous,
closed to
open
(> 15%)
81 Tree cover, 0 0 0 0 0 0 0 0 0 40 70 100 5 5 0 5 5 0 5 5 0 0 0 0 45 15 0 0 0 0 0 0 0 0 0
needle-
leaved,
deciduous,
closed
(> 40%)
82 Tree cover, 0 0 0 0 0 0 0 0 0 15 30 40 0 0 0 5 5 5 5 5 5 0 0 0 45 30 30 0 0 0 30 30 20 0 0
needle-
leaved,
deciduous,
open
(15–40%)
90 Tree cover, 0 0 0 20 30 50 13.- 20 33.- 6.7 10 16.- 5 5 0 5 5 0 5 5 0 0 0 0 35 15 0 0 0 0 10 10 0 0 0
mixed leaf 3 3 7
type
(broad-
leaved and
needle-
leaved)
100 Mosaic tree 7.5 10 15 15 20 30 3.8 5 7.5 3.7 5 7.5 5 5 5 10 10 10 5 5 5 0 0 0 50 40 20 0 0 0 0 0 0 0 0
and shrub
(continued on next page)
Remote Sensing of Environment 203 (2017) 71–89
 Table 2 (continued)

ID Description BLE tree BLD tree NLE tree NLD tree BLE shrub BLD shrub NLE shrub NLD shrub Natural grass Managed grass Bare soil Wtr Sn

Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Min Ref Mx Ref Ref
A.J. Hartley et al.

Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm Bm

(> 50%)/
herbaceous
cover
(< 50%)
110 Mosaic 1.3 5 7.5 2.4 10 15 1.3 5 7.5 0 0 0 5 5 5 10 10 10 5 5 5 0 0 0 75 60 50 0 0 0 0 0 0 0 0
herbaceous
cover
(> 50%)/
tree and
shrub
(< 50%)
120 Shrubland 0 0 0 0 0 0 0 0 0 0 0 0 13.- 20 26.- 13.- 20 26.- 13.- 20 26.- 0 0 0 30 20 20 0 0 0 30 20 0 0 0
3 7 3 7 4 6
121 Shrubland 0 0 0 0 0 0 0 0 0 0 0 0 20 30 40 0 0 0 20 30 40 0 0 0 30 20 20 0 0 0 30 20 0 0 0
evergreen
122 Shrubland 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 40 60 80 0 0 0 0 0 0 30 20 20 0 0 0 30 20 0 0 0
deciduous
130 Grassland 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 51 60 95 0 0 0 49 40 5 0 0
140 Lichens and 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 60 60 60 0 0 0 40 40 40 0 0
mosses
150 Sparse 0 1 1 0 3 3 0 1 1 0 0 0 1 1 1 3 3 3 1 1 1 0 0 0 5 5 5 0 0 0 90 85 85 0 0
vegetation
(tree, shrub,

79
herbaceous
cover)
(< 15%)
151 Sparse tree 1 2 3 3 6 9 1 2 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5 5 0 0 0 0 90 85 85 0 0
(< 15%)
152 Sparse shrub 0 0 0 0 0 0 0 0 0 0 0 0 1 2 3 3 6 9 1 2 3 0 0 0 5 5 0 0 0 0 90 85 85 0 0
(< 15%)
153 Sparse 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 15 15 0 0 0 90 85 85 0 0
herbaceous
cover
(< 15%)
160 Tree cover, 20 30 37.- 20 30 37.- 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 30 20 12.- 0 0 0 0 0 0 20 0
flooded, 5 5 5
fresh or
brakish
water
170 Tree cover, 40 60 75 0 0 0 0 0 0 0 0 0 30 20 12.- 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 20 0
flooded, 5
saline water
180 Shrub/ 0 0 0 0 5 5 0 10 10 0 0 0 0 0 0 12.- 10 10 6.4 5 5 0 0 0 50.- 40 40 0 0 0 0 0 0 30 0
herbaceous 7 9
cover,
flooded,
fresh/
saline/
brakish
water
190 Urban areas 0 0 0 2.5 2.5 2.5 2.5 2.5 2.5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 15 15 15 0 0 0 75 75 75 5 0
200 Bare areas 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 100 100 100 0 0
(continued on next page)
Remote Sensing of Environment 203 (2017) 71–89
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

more quickly than snow on the ground. Therefore, more tree cover

Ref

10-
Sn

0
under the maxLC and maxCW PFT distributions (Fig. 4), means that

100
Wtr

Ref
snow cover melts faster than PFT distributions with lower tree cover

0
(minLC and minCW), causing the surface to have a lower reflectance
100

100
Bm
Mx

under greater tree cover.

0
100

100
Ref
Bare soil

0
3.2.2. GPP
Min

100

100
Bm

All models showed an increase in annual total GPP with increasing

0

0
biomass for both the maxLC and maxCW PFT distributions, although
Bm
Mx
Managed grass

JSBACH had a greater sensitivity to maxLC_maxCW than ORCHIDEE

0

0 and JULES (Table 4 and Fig. 6). The main influence on GPP is from CW
Ref

uncertainty in tree cover, with increases (decreases) in GPP where tree

Min

cover increases (decreases) relative to the reference, particularly in the

Bm

tropics. Globally, the effect of PFT uncertainty on mean annual GPP

Bm

uncertainty ranges from − 12.7% to + 11.2% for JSBACH, − 7.5% to

Mx

0
Natural grass

2.6% for JULES and − 9.5% to 4.5% for ORCHIDEE (Table 4), in-
Ref

dicating a higher sensitivity of all models to minLC_minCW PFT dis-

0

tributions. Generally, the three models have similar spatial and tem-
Min
Bm

poral responses of GPP to PFT uncertainty (Fig. 8). However, changes in

Bm
Mx

GPP are found to occur in more locations globally in JSBACH compared

0

to ORCHIDEE and JULES, which showed more localised differences.

NLD shrub

Ref

The minimum biomass PFT distribution, minLC_minCW, was found

0

to result in a strong reduction in GPP across all models. Interestingly

Min
Bm

however, the minLC_refCW PFT distribution produced a higher annual

total GPP than the reference (refLC_refCW; Table 4), in all three models,
Bm
Mx

mainly due to increases in C4 grassland fraction in dry tropical/semi-

NLE shrub

Ref

arid regions in Africa (Fig. 3 and Fig. 8) that are more productive than
0

their tree/shrub counterparts. While increases in grassland fraction

Min
Bm

under the minLC_minCW PFT distribution were also seen in boreal re-
0

gions (Fig. 3), concomitant large increases in bare soil fraction in

Bm
Mx

minLC_minCW resulted in a predictable strong decrease in GPP, thus

BLD shrub

Ref

masking any increase in GPP due to higher grass fractions. The min-
0

LC_minCW PFT distribution results in the highest change in global total

Min
Bm

mean annual GPP in all regions and for all models (Fig. 6), which as
0

previously discussed was due to changes in tropical regions (Fig. 8 and

Bm
Mx

Fig. 7). It is also noteworthy that if minLC_minCW were not considered,

BLE shrub

the PFT uncertainty range for JULES and ORCHIDEE would have been
Ref

much smaller than the range of GPP due to inter-annual variability.

Min
Bm

Theuncertainty in PFT distributions did not dramatically alter the

0

timing (phase) of the GPP mean seasonal cycle in any of the model
Bm
Mx

simulations (Fig. 7); however, the magnitude of the change typically

0

scaled with the magnitude of GPP, and therefore greater differences

Ref
NLD tree

were seen during the growing season in all regions (Fig. 7). Overall
Min
Bm

JSBACH had a higher sensitivity to changes in biomass than JULES and

0

ORCHIDEE, especially at high latitudes (40–70°N - Fig. 7).

Bm
Mx

3.2.3. Evapotranspiration
Ref
NLE tree

In general, all models show an increase in annual total evapo-

Min

transpiration (ET) for both the maxLC_refCW and maxLC_maxCW PFT

Bm

distributions, but the response to both the minLC_refCW and

Bm
Mx

minLC_minCW PFT distributions was not consistent across models

0

(Table 5). JULES shows a consistent increase in ET with increasing

Ref
BLD tree

biomass scale, from minLC_minCW to maxLC_maxCW (Fig. 6), with

0

changes ranging from 71.2 to 76.2 × 103 km 3 yr− 1 (−4.1% to 2.7%)

Min
Bm

of evapotranspired water in the reference simulation (Table 5). ORC-

Bm
Mx

HIDEE shows a similar magnitude of ET change (− 4.1% to 2.9%) as

0

JULES from the lower end to the top of this scale, however minLC_refCW
Ref
BLE tree

uncertainty has no effect on ET in ORCHIDEE and JSBACH in com-

0

parison to the reference simulation. Similar to GPP, this anomaly is

Min
Bm

coincident with increases in C4 grass fraction in tropical grasslands

Consolidate-

(Fig. 3). JSBACH shows very small increases in ET with increasing

Unconsolid-
d bare area
Description
Table 2 (continued)

Permanent
snow and
ated bare

biomass scale, although slightly larger increases are found in the

bodies
Water
areas

maxLC_maxCW simulations due to increases in boreal zone ET, a feature

ice

which is not seen in JULES or ORCHIDEE (Fig. 9).

210

220
201

202

Comparison of annual mean maps of ET (Fig. 9) reveals that changes

ID

in JSBACH ET due to PFT uncertainty generally followed an opposite

80
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

Fig. 3. Maps showing the difference (compared to the reference case) in fraction of vegetation for Trees (1st column), Shrubs (2nd column), Natural and Managed Grasses (3rd column)
and Bare soil (4th column) for each of the uncertainty simulations (rows). The actual vegetation fractions for the reference case is shown in the middle row (refLC_refCW). The difference
in vegetation fraction for the minimum biomass vegetation distributions are shown below the reference case, first with minimum LC maps and the reference cross-walking table
(minLC_refCW – 4th row) and with both the minimum biomass LC maps and minimum cross-walking table (minLC_minCW – 5th row). The rows above the reference map are the
equivalent differences in vegetation fraction for the maximum biomass vegetation distributions (maxLC_refCW – 2nd row, maxLC_maxCW – 1st row).

Fig. 4. The latitudinal distribution of uncertainty in the

area covered by the major vegetation types (tree, shrub and
grass) and bare soil. Vegetation areas (y-axis) were calcu-
lated using area-weighting for each latitudinal zone with a
spatial resolution of 2° latitude.

81
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

Fig. 5. Global maps of the monthly mean albedo for the reference (refLC_refCW) PFT distribution (middle row) for JSBACH, JULES and ORCHIDEE for the minLC_minCW, minLC_refCW,
maxLC_refCW and maxLC_maxCW PFT distributions from bottom to top respectively.

spatial pattern in the tropics in comparison to JULES and ORCHIDEE. is shown by maxLC and maxCW leading to a decrease in ET from May to
JSBACH showed increases in ET in the Amazon and Congo basins under October in mid-latitudes and from September to June in the tropical
lower tree fractions (minLC and minCW), and small reductions in ET zone. In Fig. 7 it can be seen that these months correspond to the
under high tree fractions (maxLC and maxCW). This had the effect of growing seasons (when maximum leaf area occurs) in mid-latitudes and
cancelling out changes in the boreal zone, resulting in apparent low tropical zone.
sensitivity of annual mean global ET (−0.1% to 1.1% change with
regard to the reference; Table 5). This highlights a known issue in
3.3. LSM sensitivity to PFT uncertainty
JSBACH, that is related to transpiration in tropical forests (Hagemann,
2016). The seasonal cycle of ET (Fig. 7) for different latitudinal zones
In order to understand the significance of these results, and there-
(high-northern: 40–70 N, mid-latitude: 10–40 N and tropical:
fore the sensitivity of LSMs to PFT uncertainty, it is important to put
20Se10N) also shows that JSBACH has an inverted signal for the mid-
these changes in the context of other commonly used measures of inter-
latitudes and tropical zone with respect to JULES and ORHCIDEE. This
model uncertainty. Previous studies have shown that large inter-model

82
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

Fig. 6. Sensitivity of global mean albedo, evapotranspira-

tion (ET) and gross primary productivity (GPP) in JSBACH,
JULES and ORCHIDEE due to inter-annual variability and
PFT uncertainty. Each point represents the annual value of
that variable for each model, therefore each cluster of
points for each scenario shows the inter-annual variability
for a given model and PFT distribution. The difference be-
tween clusters of a certain colour shows the individual
model response to PFT uncertainty. Boxplots on the left of
each plot show the range of PFT uncertainty for that model.
Note that boxplots summarise the information shown to the
right by the coloured points for a given model.

Table 3 Table 4
Global 30-year mean annual albedo for each model and PFT distribution. Values in Global 30-year mean annual GPP (PgC yr− 1) for each model and PFT distribution. Values
brackets show the percentage difference to the reference simulation for each model. in brackets show the percentage difference to the reference simulation for each model.

PFT distribution JSBACH JULES ORCHIDEE Realisations JSBACH JULES ORCHIDEE

maxLC_maxCW 0.222 (−6.7) 0.244 (− 8.3) 0.206 (− 4.6) maxLC_maxCW 172.4 (11.2) 132.1 (2.6) 102.4 (4.5)
maxLC_refCW 0.232 (−2.5) 0.258 (− 3) 0.211 (− 2.3) maxLC_refCW 159.5 (2.9) 130 (1) 100.1 (2.2)
refLC_refCW 0.238 (0) 0.266 (0) 0.216 (0) refLC_refCW 155 (0) 128.7 (0) 98 (0)
minLC_refCW 0.242 (1.7) 0.274 (3) 0.22 (1.9) minLC_refCW 155.8 (0.5) 130.7 (1.6) 100.5 (2.6)
minLC_minCW 0.249 (4.6) 0.282 (6) 0.225 (4.2) minLC_minCW 135.4 (− 12.7) 119 (− 7.5) 88.6 (− 9.5)

uncertainty is a dominant feature of multi-model studies (Friedlingstein extent GPP and ET (Fig. 10). Table 6 shows that the range in simulated
et al., 2006; Jung et al., 2007; Sitch et al., 2015). We find that un- albedo due to uncertainty in PFT distributions (JSBACH: 0.027; JULES:
certainty in PFT distribution has a potentially larger effect on global 0.038; ORCHIDEE: 0.019) is of a similar magnitude for JULES, to the
albedo relative to the between model uncertainty range, and to a lesser inter-model uncertainty range. When inter-annual variability is also

Fig. 7. Mean seasonal cycles of GPP (top), ET (middle) and albedo (bottom) for all PFT distributions for JSBACH, JULES and ORCHIDEE for three different latitudinal zones (left: 40°N to
70°N; middle: 10°N to 40°N; right: 20°S to 10°N).

83
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

Fig. 8. Global maps of the annual mean GPP (gCm− 2 month− 1) for the reference (refLC_refCW) PFT distribution (middle row) for JSBACH, JULES and ORCHIDEE and the difference in
annual mean GPP for the minLC_minCW, minLC_refCW, maxLC_refCW and maxLC_maxCW PFT distributions from bottom to top respectively.

accounted for (Fig. 10), JULES (JSBACH) PFT distribution-based un-

Table 5
Global 30-year mean annual ET (103 km3 yr− 1 of water) for each model and PFT dis- certainty in albedo was found to range from 65% to 105% (45% to
tribution. Values in brackets show the percentage difference to the reference simulation 75%) of the albedo uncertainty between models. As previously dis-
for each model. cussed, much of this albedo uncertainty derives from CW uncertainty in
the fractional cover of tree and bare soil in the boreal zone.
Realisations JSBACH JULES ORCHIDEE
For global GPP, JSBACH was found to be most sensitive to un-
maxLC_maxCW 60.1 (1.1) 76.2 (2.7) 93.4 (2.9) certainty in PFT distributions, with 30-year means ranging from
maxLC_refCW 59.6 (0.3) 75.2 (1.4) 92.2 (1.6) 135.4 PgC yr− 1 under minLC_minCW to 172.4 PgC yr− 1 under
refLC_refCW 59.4 (0) 74.2 (0) 90.8 (0) maxLC_maxCW (Table 4). This equates to a range of 38 PgC yr− 1
minLC_refCW 59.5 (0) 73.1 (− 1.4) 90.7 (0)
(Table 6), which is between 55 and 90% of the inter- model uncertainty
minLC_minCW 59.4 (−0.1) 71.2 (− 4.1) 87.1 (− 4.1)
range (Fig. 10). JULES and ORCHIDEE were found to show slightly
lower sensitivity of GPP to PFT uncertainty. While the magnitude of

84
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

Fig. 9. Global maps of annual mean ET (mm/m2/year) for the reference (refLC_refCW) PFT distribution (middle row) for JSBACH, JULES and ORCHIDEE and the difference in monthly
mean ET for the minLC_minCW, minLC_refCW, maxLC_refCW and maxLC_maxCW PFT distributions from bottom to top respectively.

global 30-year mean GPP (Table 4) was found to be higher in JULES
(119–132.1 PgC yr− 1) than in ORHCIDEE (88.6–102.4 PgC yr− 1), it
was found that ORCHIDEE was slightly more sensitive to PFT un-
certainty, varying for − 9.5% to +4.5% of the reference compared to
− 7.5% to 2.6% in JULES. The inter-model uncertainty range of GPP
was found to vary considerably according to PFT distribution, with
relatively low GPP uncertainty due to LC (minLC: 44.8 PgC yr− 1 and
maxLC: 48.6 PgC yr− 1), but relatively high GPP uncertainty when CW
uncertainty was considered (minCW: 38 PgC yr− 1 and maxCW:
59.1 PgC yr− 1). This sensitivity of GPP to CW uncertainty is related to
uncertainty in tree fractional cover in the tropics, with all models
showing GPP increases (Fig. 9) where tree fraction increased (Fig. 3)
under maxCW. The opposite was true for minCW, with GPP reductions
where tree fraction reduced, but even larger GPP reductions were found
where bare soil fractions increased due to minCW uncertainty in the
tropics.
In contrast to GPP, JSBACH had the lowest absolute totals of ET
(Table 5), and the lowest sensitivity to PFT uncertainty, with ET ran-
ging from 59.4 to 60.1 × 103 km 3 yr− 1 for all PFT distributions, al-
though sensitivity of ET inter-annual variability to PFT uncertainty was
comparable for all models (See Supplementary Information Fig. S1).
PFT uncertainty had a greater impact on JULES ET, with an uncertainty
of 5 × 103 km 3 yr− 1, which equates to 15.8% of the between model
uncertainty in the reference simulation (Table 6). The largest impact of

85
A.J. Hartley et al. Remote Sensing of Environment 203 (2017) 71–89

efforts for future LC mapping exercises. For example, a small increase in

Fig. 10. Each point shows, for a given year and model, the PFT-based uncertainty in
global annual mean albedo, ET and GPP expressed as a percentage of between model
uncertainty. Values > 100% mean that the PFT-based uncertainty is greater than the
between-model uncertainty. The between model uncertainty range (Rmod) was calculated
annually for each of the 30 years of the simulations, for each of the 5 PFT distributions.
The PFT uncertainty range (Rpft) was calculated for each of the 3 models (shown in red,
blue and green). For each year of the simulations and for each model (shown as a co-
loured point in the plot), 100 ∗ Rpft / Rmod was calculated for each of the 5 different model
uncertainty ranges. Therefore, for each model and variable, we show 150 (30 years ∗ 5
model uncertainty ranges) different measures of PFT uncertainty as a percentage of the
model uncertainty range. Boxplots (showing the median, interquartile range, minimum
and maximum of the 3 model ensemble) are displayed in the background to allow visual
comparison.
grass fraction, replacing shrub cover under minLC, was found to lead to
relatively large impacts on GPP in West and East African savannahs.
From the remote sensing point of view, different options could be
identified to tackle these issues. Firstly, specific classifications could be
run for these critical classes and locations (such as grass and bare soil
fractions in arid areas). This approach could provide improvements
compared to the classification methodology used here that has been
developed to be valid at global scale. Different strategies have been
used to better represent the regional heterogeneity of the land cover
characteristics such as the use of an a priori stratification, or the in-
corporation of external datasets for specific classes, although these will
never replace specific classifications. A second approach could be the
use of continuous fields information (e.g. Hansen et al., 2003) for cer-
tain classes and locations where vegetation gradients exist instead of a
categorical classification, which could contribute towards more realistic
PFT distributions. This approach has the advantage of removing the
need for a CW step in locations where LSMs are potentially sensitive,
however careful attention would be needed to accurately map PFT
fractions where more than one PFT exists. Finally, the use of higher
spatial resolution data, provided by missions such as Sentinel 1 and 2,
will be very important to resolve some of the issues related to mixed
pixels/classes, as well as helping to better constrain CW uncertainty in
PFT fractional cover presented in this study. However, higher spatial
resolution LC maps might also raise new challenges, particularly in
relation to ensuring high resolution LC legends are easily related to the
legends designed for medium-resolution mapping, such as the one
presented here.

4.2. Implications for earth system modellers

Table 6
Uncertainty ranges for albedo, ET (103 km3 yr− 1) and GPP (PgC yr− 1) due to either inter-
model uncertainty or inter-PFT distribution uncertainty. Values in brackets show the This study has shown that PFT uncertainty is an important aspect of
percentage of between model uncertainty (for the reference simulation) that was ac- uncertainty in land surface models that has not been fully acknowl-
counted for by inter-PFT uncertainty. edged or accounted for in previous studies. We have shown that PFT
Uncertainty type PFT scenario Albedo ET GPP
uncertainty has a larger impact on GPP than previously reported by
(Jung et al., 2007; Quaife et al., 2008), who estimated it to be in the
Inter-Model (n = 3) maxLC_maxCW 0.038 33.3 70.1 region of 10–15% of GPP. In this study, by perturbing the key elements
maxLC_refCW 0.047 32.6 59.3 of the processing chain of one land cover dataset to produce the max-
refLC_refCW 0.05 31.3 57
minLC_refCW 0.054 31.3 55.3
imum plausible range of uncertainty in PFT distribution, we show that
minLC_minCW 0.057 27.7 46.7 sensitivity of GPP was 23.9% for JSBACH, 10.1% for JULES and 14%
Inter-PFT (n = 5) JSBACH 0.027 (54) 0.7 (2.3) 37.1 (65) for ORCHIDEE (Table 4) compared to the reference simulation. In terms
JULES 0.038 (76) 5 (16.1) 13.1 (23) of global GPP estimates, the uncertainty ranges found here were much
ORCHIDEE 0.019 (38) 6.3 (20.2) 13.7 (24.1)
wider than those found by upscaling flux station observations to the
global scale (Beer et al., 2010; Jung et al., 2011) (.
PFT uncertainty on ET was in ORCHIDEE, which was also found to have Table 7). These results were more comparable to other approaches
the largest absolute ET values (Table 5), and the highest inter-annual for quantifying GPP using atmospheric constraints that suggest
variability (Fig. 10). PFT distribution uncertainty lead to an uncertainty 150–175 PgC yr− 1 (Welp et al., 2011) or 146 ± 19 PgC yr− 1 (Koffi
range in ORCHIDEE ET of 7.6 × 103 km 3 yr− 1 (23.8% of the between et al., 2012) may be better estimates. In this study, given that PFT
model range for the reference simulation), the largest of all three distribution uncertainty contributes up to 100% of inter-model un-
models. ORCHIDEE sensitivity to individual PFT distributions was certainty (Fig. 10), these results would suggest that LSM uncertainty
found to differ from JSBACH and JULES, particularly for the min- can be reduced by better constraining the PFT fractional distributions
LC_refCW distribution where virtually no change was found compared through more accurate land cover mapping, and a more accurate CW
to the reference simulation (Table 5). This was found to coincide with table.
increases in tropical grass cover in particular in Africa (Fig. 3), and is In percentage terms, PFT uncertainty had less impact on global ET
consistent with increases in GPP in ORCHIDEE in similar areas. While than GPP and albedo in this study. However, the impact on total annual
JULES and JSBACH showed a similar sensitivity in GPP to the min- ET land to atmosphere flux is significant for JULES
LC_refCW PFT distribution, such an impact on ET was not found in these (5.0 × 103 km 3 yr− 1) and ORCHIDEE (7.6 × 103 km 3 yr− 1) in the
models. context of estimates of Anthropocene ET change given by [Gordon et al.,
2005 (3.0 × 103 km 3 yr− 1); Sterling et al., 2013
(3.5 × 103 km 3 yr− 1); Boisier et al., 2014 (1.26 ± 0.85 ×
4. Summary and perspectives 103 km 3 yr− 1)] (Table 7). This would suggest, as noted by (Boisier
et al., 2014) that improving PFT fractional distributions would result in
4.1. Implications for remote sensing of land cover improved estimates of global ET. The impact of PFT distribution un-
certainty on albedo was found to be as much as inter-model uncertainty
This study has highlighted the locations where LSMs are most sen- for JULES in this study (Fig. 10). Forest cover uncertainty in the boreal
sitive to LC uncertainty. This is essential information that can focus zone impacted the monthly mean albedo in JULES by up to 0.2, in

86
A.J. Hartley et al.
Table 7
Other modelling and observation-based estimates of global ET and GPP.
Variable Method
GPP Data-oriented models using FLUXNET observations
GPP Machine learning technique based on FLUXNET observations
GPP Land Surface Models driven by observed climate forcing
GPP Stable isotope ratios of atmospheric CO2
GPP Atmospheric transport model, LSM and atmospheric CO2 observations
ET Machine learning technique based on FLUXNET observations
ET Observation-based estimates
ET LSMs driven by observed climate
ET LSMs driven by observed climate
JSBACH by 0.15 and in ORCHIDEE by 0.1 during the boreal winter, a
figure which matches the magnitude of that found by (Betts, 2000),
using the Met Office Surface Exchange Scheme, a pre-curser of JULES.
Given the potential impact of this albedo uncertainty on the earth's
radiation budget, this highlights the need for improved understanding
of the fractional cover of all vegetation types in the boreal zone.
The aim of this study was to investigate the impact of the widest
plausible range of PFT uncertainty on LSMs, and to identify ways in
which future work on land cover mapping of the land cover essential
climate variable can reduce LSM uncertainty. We note that if the ex-
periment were to include coupled land-atmosphere simulations, we
may expect an even wider range of impacts due to feedback processes.
Whilst this study was not an exhaustive investigation of all un-
certainties related to either LC mapping or CW, it has revealed im-
portant implications for the regions and classes on which future land
cover mapping efforts should be focussed. We have shown that there is
still considerable uncertainty in the cross-walking procedure used to
convert LC to the PFT distributions used by LSMs (e.g. (Poulter et al.,
2015)). The uncertainty related to the CW table will likely be higher
when considering any differences between regions (e.g. sparse vegeta-
tion in high latitude tundra regions may have a different % grass cover
than in tropical dryland regions). Furthermore, uncertainty in the un-
derlying LC mapping algorithm for assigning a LC class to a spectral
signature has been shown to be of an equal magnitude to cross-walking
uncertainty. For example, uncertainty in bare soil fraction in northern
latitudes (north of 60°N) is all related to choice of land cover class, not
cross-walking uncertainty (Fig. 4). Further south however, grass cover
uncertainty is all related to cross-walking uncertainty rather than land
cover uncertainty. In the tropics (15°S to 15°N), tree cover uncertainty
is equally related to LC and CW uncertainty, whereas in the other major
forest belt (45°N to 65°N), uncertainty is more related to CW un-
certainty. This is also seen in northern Asia and eastern North America
regions. Uncertainty in grass and shrub fractions is more locally de-
pendent. It appears this is due to regional differences in LC uncertainty
in grass classes, and CW uncertainty affecting different classes in dif-
ferent ways. For example, in tropical regions, maximising biomass in
the CW approach leads to a reduction in grass cover relative to the
reference, whereas in temperate regions (30°N to 50°N and 25°S to
35°S), maximising biomass leads to an increase in grass cover.
In this study we defined the range of CW uncertainty to correspond
closely to the LC class descriptions where possible. For example, for the
sparse vegetation class (LC class 150), which is defined as < 15% ve-
getation, the minimum bare soil fraction that we assigned was 85%.
Similarly, in a class with an open canopy (15–40%) we defined the tree
fraction within this range. However, the interpretation of the LC de-
scriptions is not always easy, especially for mosaic classes (e.g. Mosaic
natural vegetation (tree, shrub, herbaceous cover) (> 50%)/cropland
(< 50%)), therefore these classes are strong sources of CW uncertainty.
87
Remote Sensing of Environment 203 (2017) 71–89
Values Period Reference
123 ± 8 PgCyr− 1 1993–2006 (Beer et al., 2010)
119 ± 6 PgCyr− 1 1982–2008 (Jung et al., 2011)
135.4–172.4 PgCyr− 1 (JSBACH) 1980–2009 Present study
119–132.1 PgCyr− 1 (JULES)
97.4–113.4 PgCyr− 1 (ORCHIDEE)
150–175 PgCyr− 1 – (Welp et al., 2011)
146 ± 19 PgCyr− 1 – (Koffi et al., 2012)
65 ± 3 × 103 km 3 yr− 1 1982–2008 (Jung et al., 2010)
65.5 × 103 km 3 yr− 1 – (Oki and Kanae, 2006)
58–85 × 103 km 3 yr− 1 1986–1995 (Dirmeyer et al., 2006)
59.4–60.1 × 103 km 3 yr− 1 (JSBACH) 1980–2009 Present study
71.2–76.2 × 103 km 3 yr− 1 (JULES)
87–94.6 × 103 km 3 yr− 1 (ORCHIDEE)
Given the large impact of these uncertainty issues on LSMs, we are
lead to question whether defining PFT fractions from LC maps is indeed
the most appropriate method. Some studies have suggested mapping
PFT distributions directly from satellite data, therefore bypassing the
classification into LC classes (e.g. the “optical functional types” ap-
proach proposed by (Ustin and Gamon, 2010)). However, further re-
search is needed by both the remote sensing and modelling commu-
nities to determine how accurately different PFTs can be distinguished
from satellite data and which level of classification is required by the
models. Future developments of some LSMs are moving towards spe-
cies-level vegetation groupings for some processes and parameters.
With the increased spatial resolution and frequency of observations that
are now available from missions such as ESA Sentinel 2 and 3, it may be
possible to distinguish between species with differing structural and
temporal characteristics, thereby reducing uncertainty particularly
where pixels were classified as mixed vegetation in coarser resolution
maps. However, direct mapping of functional characteristics such as the
distinction between C3 and C4 photosynthesis (a key requirement for
LSMs) will be much more challenging. It will be imperative that the two
communities continue to discuss such issues as the models and satellite
data products continue to be developed.
If we continue to use the cross-walking from LC approach, there is
still an open question as to what the optimal CW table should be.
Ongoing investigations by the authors are addressing this by comparing
maps produced with several different versions of the CW table with
independent estimates, such as the Land Use Harmonisation data (Hurtt
et al., 2011). Note that this is not a like-for-like comparison, because the
latter describes land use and not land cover. However, a comparison of
major classes such as cropland and forest spatial distribution can give us
information on whether we have an appropriate split between trees and
grasses for mixed classes. It may be the case that the optimal solution
would be to have a different CW table for different regions, or to blend a
continuous vegetation map (such as (Hansen et al., 2003)) with land
cover classes in areas where there is a particularly strong vegetation
gradient.
Finally, historical land use and land cover change (LULCC) is the
second biggest source of CO2 to the atmosphere after fossil fuel emis-
sions between 1750 and 2011 (~ 1.1 ± 0.8PgCyr− 1 - (Stocker et al.,
2013)). Reconstructions of historical LC time series maps that are used
in earth system model simulations propagate LC transitions backwards
starting from recent LC/PFT maps, with an increase in uncertainty back
in time (Eberenz et al., 2015; Fuchs et al., 2015). Therefore it is not only
important to define accurate LC or PFT distributions for present and
future simulations (Brovkin et al., 2013), but imperative for de-
termining historical emissions (Goll et al., 2015) and the relative con-
tribution of climate change, CO2 increase and LULCC to changes in
terrestrial biosphere feedbacks with the atmosphere.
A.J. Hartley et al.
Acknowledgements
This study was supported by funding from the European Space
Agency (4000109875) under the Climate Change Initiative Land Cover
Project Phase 2. The authors would especially like to thank Philippe
Peylin, Stefan Hagemann, Stephen Sitch, Pierre Defourny, Catherine
Ottlé and Debbie Hemming for their advice and comments that un-
doubtedly helped to improve the quality of this manuscript.
Appendix A. Supplementary data
Supplementary data associated with this article can be found in the
online version, at http://dx.doi.org/10.1016/j.rse.2017.07.037. These
data included the Google map of the most important areas describe in
this article.
References
Alkama, R., Cescatti, A., Quéré, C. Le, Peters, G.P., Andres, R.J., Andrew, R.M., Boden,
T.A., Ciais, P., Friedlingstein, P., Houghton, R.A., et al., 2016. Biophysical climate
impacts of recent changes in global forest cover. Science 351, 600–604. http://dx.doi.
org/10.1126/science.aac8083.
Anav, A., Friedlingstein, P., Kidston, M., Bopp, L., Ciais, P., Cox, P., Jones, C., Jung, M.,
Myneni, R., Zhu, Z., 2013. Evaluating the land and ocean components of the global
carbon cycle in the CMIP5 earth system models. J. Clim. 26, 6801–6843. http://dx.
doi.org/10.1175/JCLI-D-12-00417.1.
Ballantyne, A.P., Andres, R., Houghton, R., Stocker, B.D., Wanninkhof, R., Anderegg, W.,
Cooper, L.A., DeGrandpre, M., Tans, P.P., Miller, J.B., et al., 2015. Audit of the global
carbon budget: estimate errors and their impact on uptake uncertainty.
Biogeosciences 12, 2565–2584. http://dx.doi.org/10.5194/bg-12-2565-2015.
Beer, C., Reichstein, M., Tomelleri, E., Ciais, P., Jung, M., Carvalhais, N., Rödenbeck, C.,
Arain, M.A., Baldocchi, D., Bonan, G.B., 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.
Berbet, M.L.C., Costa, M.H., 2003. Climate change after tropical deforestation: seasonal
variability of surface albedo and its effects on precipitation change. J. Clim. 16,
2099–2104.
Betts, R.A., 2001. Biogeophysical impacts of land use on present-day climate: near-surface
temperature change and radiative forcing. Atmos. Sci. Lett. 2, 39–51. http://dx.doi.
org/10.1006/asle.2001.0023.
Betts, R.A., 2000. Offset of the potential carbon sink from boreal forestation by decreases
in surface albedo. Nature 408, 187–190. http://dx.doi.org/10.1038/35041545.
Boisier, J.P., de Noblet-Ducoudré, N., Ciais, P., 2014. Historical land-use-induced eva-
potranspiration changes estimated from present-day observations and reconstructed
land-cover maps. Hydrol. Earth Syst. Sci. 18, 3571–3590. http://dx.doi.org/10.5194/
hess-18-3571-2014.
Boisier, J.P., de Noblet-Ducoudré, N., Pitman, A.J., Cruz, F.T., Delire, C., van den Hurk,
B.J.J.M., van der Molen, M.K., Müller, C., Voldoire, A., 2012. Attributing the impacts
of land-cover changes in temperate regions on surface temperature and heat fluxes to
specific causes: results from the first LUCID set of simulations. J. Geophys. Res.
Atmos. 117. http://dx.doi.org/10.1029/2011JD017106.
Bonan, G.B., 2008. Forests and climate change: forcings, feedbacks, and the climate
benefits of forests. Science 320, 1444–1449. http://dx.doi.org/10.1126/science.
1155121.
Brovkin, V., Boysen, L., Arora, V.K., Boisier, J.P., Cadule, P., Chini, L., Claussen, M.,
Friedlingstein, P., Gayler, V., Van den hurk, B.J.J.M., et al., 2013. Effect of anthro-
pogenic land-use and land-cover changes on climate and land carbon storage in
CMIP5 projections for the twenty-first century. J. Clim. 26, 6859–6881. http://dx.
doi.org/10.1175/JCLI-D-12-00623.1.
Collatz, G.J., Berry, J.a., Clark, J.S., 1998. Effects of climate and atmospheric CO 2 partial
pressure on the global distribution of C 4 grasses: present, past, and future. Oecologia
114, 441–454. http://dx.doi.org/10.1007/s004420050468.
Cosgrove, B.A., Lohmann, D., Mitchell, K.E., Houser, P.R., Wood, E.F., Schaake, J.C.,
Robock, A., Marshall, C., Sheffield, J., Duan, Q., et al., 2003. Real-time and retro-
spective forcing in the North American Land Data Assimilation System (NLDAS)
project. J. Geophys. Res. 108, 8842. http://dx.doi.org/10.1029/2002JD003118.
Cox, P.M., 2001. Description of the TRIFFID Dynamic Global Vegetation Model, Hadley
Centre Technical Note. Bracknel, UK.
Cramer, W., Bondeau, A., Woodward, F.I., Prentice, I.C., Betts, R.a., Brovkin, V., Cox,
P.M., Fisher, V., Foley, J.a., Friend, A.D., et al., 2001. Global response of terrestrial
ecosystem structure and function to CO 2 and climate change: results from six dy-
namic global vegetation models. Glob. Chang. Biol. 7, 357–373. http://dx.doi.org/
10.1046/j.1365-2486.2001.00383.x.
Di Gregorio, A., Jansen, L.J.M., 2005. Land Cover Classification System (LCCS), Version 2:
Classification Concepts and User Manual, Environmental and Natural Resources
Series. FAO, Rome, Rome, Italy.
Dirmeyer, P.A., Gao, X., Zhao, M., Guo, Z., Oki, T., Hanasaki, N., 2006. GSWP-2: multi-
model analysis and implications for our perception of the land surface. Bull. Am.
Meteorol. Soc. 87, 1381–1397. http://dx.doi.org/10.1175/BAMS-87-10-1381.
Eberenz, J., Herold, M., Verbesselt, J., Wijaya, A., Linquist, E., Defourny, P., Gibbs, H.,
88
Remote Sensing of Environment 203 (2017) 71–89
Arino, O., Achard, F., 2015. Consistent forest change maps 1981-2000 from the
AVHRR time series: case studies for South America and Indonesia. In: 2015 8th
International Workshop on the Analysis of Multitemporal Remote Sensing Images
(Multi-Temp). IEEE, pp. 1–4. http://dx.doi.org/10.1109/Multi-Temp.2015.7245799.
Feddema, J., Oleson, K., Bonan, G., Mearns, L., Washington, W., Meehl, G., Nychka, D.,
2005b. A comparison of a GCM response to historical anthropogenic land cover
change and model sensitivity to uncertainty in present-day land cover representa-
tions. Clim. Dyn. 25, 581–609. http://dx.doi.org/10.1007/s00382-005-0038-z.
Feddema, J.J., Oleson, K.W., Bonan, G.B., Mearns, L.O., Buja, L.E., Meehl, G.A.,
Washington, W.M., 2005a. The importance of land-cover change in simulating future
climates. Science 310, 1674–1678. http://dx.doi.org/10.1126/science.1118160.
Findell, K.L., Shevliakova, E., Milly, P.C.D., Stouffer, R.J., Findell, K.L., Shevliakova, E.,
Milly, P.C.D., Stouffer, R.J., 2007. Modeled impact of anthropogenic land cover
change on climate. J. Clim. 20, 3621–3634. http://dx.doi.org/10.1175/JCLI4185.1.
Friedlingstein, P., Cox, P., Betts, R., Bopp, L., von Bloh, W., Brovkin, V., Cadule, P.,
Doney, S., Eby, M., Fung, I., et al., 2006. Climate–carbon cycle feedback analysis:
results from the C4MIP model Intercomparison. J. Clim. 19, 3337–3353. http://dx.
doi.org/10.1175/JCLI3800.1.
Fritz, S., See, L., 2008. Identifying and quantifying uncertainty and spatial disagreement
in the comparison of global land cover for different applications. Glob. Chang. Biol.
14, 1057–1075. http://dx.doi.org/10.1111/j.1365-2486.2007.01519.x.
Fritz, S., See, L., 2005. Comparison of land cover maps using fuzzy agreement. Int. J.
Geogr. Inf. Sci. 19, 787–807. http://dx.doi.org/10.1080/13658810500072020.
Fritz, S., See, L., McCallum, I., Schill, C., Obersteiner, M., van der Velde, M., Boettcher, H.,
Havlík, P., Achard, F., 2011. Highlighting continued uncertainty in global land cover
maps for the user community. Environ. Res. Lett. 6, 44005. http://dx.doi.org/10.
1088/1748-9326/6/4/044005.
Fuchs, R., Herold, M., Verburg, P.H., Clevers, J.G.P.W., Eberle, J., 2015. Gross changes in
reconstructions of historic land cover/use for Europe between 1900 and 2010. Glob.
Chang. Biol. 21, 299–313. http://dx.doi.org/10.1111/gcb.12714.
Goll, D.S., Brovkin, V., Liski, J., Raddatz, T., Thum, T., Todd-Brown, K.E.O., 2015. Strong
dependence of CO2 emissions from anthropogenic land cover change on initial land
cover and soil carbon parametrization. Glob. Biogeochem. Cycles 29, 1511–1523.
http://dx.doi.org/10.1002/2014GB004988.
Gordon, L.J., Steffen, W., Jonsson, B.F., Folke, C., Falkenmark, M., Johannessen, A., 2005.
Human modification of global water vapor flows from the land surface. Proc. Natl.
Acad. Sci. U. S. A. 102, 7612–7617. http://dx.doi.org/10.1073/pnas.0500208102.
Hagemann, S., 2016. Personal Communication.
Hansen, M.C., DeFries, R.S., Townshend, J.R.G., Carroll, M., Dimiceli, C., Sohlberg, R.A.,
2003. Global percent tree cover at a spatial resolution of 500 meters: first results of
the MODIS vegetation continuous fields algorithm. Earth Interact. 7, 1–15. http://dx.
doi.org/10.1175/1087-3562(2003)007<0001:GPTCAA>2.0.CO;2.
Harris, I., Jones, P.D., Osborn, T.J., Lister, D.H., 2014. Updated high-resolution grids of
monthly climatic observations - the CRU TS3.10 dataset. Int. J. Climatol. 34. http://
dx.doi.org/10.1002/joc.3711.
Hartley, A.J., Parker, D.J., Garcia-Carreras, L., Webster, S., 2016. Simulation of vegeta-
tion feedbacks on local and regional scale precipitation in West Africa. Agric. For.
Meteorol. 222, 59–70. http://dx.doi.org/10.1016/j.agrformet.2016.03.001.
Hoffmann, W.A., Jackson, R.B., 2000. Vegetation-climate feedbacks in the conversion of
tropical savanna to grassland. J. Clim. 13, 1593–1602.
Houghton, R.A., House, J.I., Pongratz, J., van der Werf, G.R., DeFries, R.S., Hansen, M.C.,
Le Quéré, C., Ramankutty, N., 2012. Carbon emissions from land use and land-cover
change. Biogeosciences 9, 5125–5142. http://dx.doi.org/10.5194/bg-9-5125-2012.
Hurtt, G.C., Chini, L.P., Frolking, S., Betts, R.A., Feddema, J., Fischer, G., Fisk, J.P.,
Hibbard, K., Houghton, R.A., Janetos, A., et al., 2011. Harmonization of land-use
scenarios for the period 1500–2100: 600 years of global gridded annual land-use
transitions, wood harvest, and resulting secondary lands. Clim. Chang. 109, 117–161.
http://dx.doi.org/10.1007/s10584-011-0153-2.
Jones, P.W., 1999. First- and second-order conservative remapping schemes for grids in
spherical coordinates. Mon. Weather Rev. 127, 2204–2210. http://dx.doi.org/10.
1175/1520-0493(1999)127<2204:FASOCR>2.0.CO;2.
Jung, M., Henkel, K., Herold, M., Churkina, G., 2006. Exploiting synergies of global land
cover products for carbon cycle modeling. Remote Sens. Environ. 101, 534–553.
http://dx.doi.org/10.1016/j.rse.2006.01.020.
Jung, M., Reichstein, M., Ciais, P., Seneviratne, S.I., Sheffield, J., Goulden, M.L., Bonan,
G., Cescatti, A., Chen, J., de Jeu, R., et al., 2010. Recent decline in the global land
evapotranspiration trend due to limited moisture supply. Nature 467, 951–954.
http://dx.doi.org/10.1038/nature09396.
Jung, M., Reichstein, M., Margolis, H.A., Cescatti, A., Richardson, A.D., Arain, M.A.,
Arneth, A., Bernhofer, C., Bonal, D., Chen, J., 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. J. Geophys. Res. 116. G00J07.
doi:http://dx.doi.org/10.1029/2010JG001566.
Jung, M., Vetter, M., Herold, M., Churkina, G., Reichstein, M., Zaehle, S., Ciais, P., Viovy,
N., Bondeau, A., Chen, Y., et al., 2007. Uncertainties of modeling gross primary
productivity over Europe: a systematic study on the effects of using different drivers
and terrestrial biosphere models. Glob. Biogeochem. Cycles 21. http://dx.doi.org/10.
1029/2006GB002915.
Knox, R., Bisht, G., Wang, J., Bras, R., 2011. Precipitation variability over the forest-to-
nonforest transition in Southwestern Amazonia. J. Clim. 24, 2368–2377. http://dx.
doi.org/10.1175/2010JCLI3815.1.
Koffi, E.N., Rayner, P.J., Scholze, M., Beer, C., 2012. Atmospheric constraints on gross
primary productivity and net ecosystem productivity: results from a carbon-cycle
data assimilation system. Glob. Biogeochem. Cycles 26. http://dx.doi.org/10.1029/
2010GB003900.
Kottek, M., Grieser, J., Beck, C., Rudolf, B., Rubel, F., 2006. World map of the Köppen-
A.J. Hartley et al.
Geiger climate classification updated. Meteorol. Z. 15, 259–263. http://dx.doi.org/
10.1127/0941-2948/2006/0130.
LC_CCI ATBD, 2013. Algorithm Theoretical Basis Document for Land Cover CCI. Frascati
(Rome).
Le Quéré, C., Moriarty, R., Andrew, R.M., Canadell, J.G., Sitch, S., Korsbakken, J.I.,
Friedlingstein, P., Peters, G.P., Andres, R.J., Boden, T.A., et al., 2015. Global carbon
budget 2015. Earth Syst. Sci. Data 7, 349–396. http://dx.doi.org/10.5194/essd-7-
349-2015.
Loveland, T.R., Belward, A.S., 1997. The IGBP-DIS global 1 km land cover data set,
DISCover: First results. Int. J. Remote Sens. 18, 3289–3295. http://dx.doi.org/10.
1080/014311697217099.
Luyssaert, S., Jammet, M., Stoy, P.C., Estel, S., Pongratz, J., Ceschia, E., Churkina, G.,
Don, A., Erb, K., Ferlicoq, M., et al., 2014. Land management and land-cover change
have impacts of similar magnitude on surface temperature. Nat. Clim. Chang. 4,
389–393. http://dx.doi.org/10.1038/nclimate2196.
Mahmood, R., Pielke, R.a., Hubbard, K.G., Niyogi, D., Dirmeyer, P.a., Mcalpine, C.,
Carleton, A.M., Hale, R., Gameda, S., Beltrán-Przekurat, A., et al., 2014. Land cover
changes and their biogeophysical effects on climate. Int. J. Climatol. 34, 929–953.
http://dx.doi.org/10.1002/joc.3736.
Oki, T., Kanae, S., 2006. Global hydrological cycles and world water resources. Science
313, 1068–1072. http://dx.doi.org/10.1126/science.1128845.
Poulter, B., Ciais, P., Hodson, E., Lischke, H., Maignan, F., Plummer, S., Zimmermann,
N.E., 2011. Plant functional type mapping for earth system models. Geosci. Model
Dev. 4, 2081–2121. http://dx.doi.org/10.5194/gmdd-4-2081-2011.
Poulter, B., MacBean, N., Hartley, A., Khlystova, I., Arino, O., Betts, R., Bontemps, S.,
Boettcher, M., Brockmann, C., Defourny, P., et al., 2015. Plant functional type clas-
sification for earth system models: results from the European Space Agency's land
cover climate change initiative. Geosci. Model Dev. 8, 2315–2328. http://dx.doi.org/
10.5194/gmd-8-2315-2015.
Prentice, I., Cramer, W., Harrison, S., Leemans, R., Monserud, R., Solomon, A., 1992. A
global biome model based on plant physiology and dominance, soil properties and
climate. J. Biogeogr. 19, 117–134.
Puma, M.J., Cook, B.I., 2010. Effects of irrigation on global climate during the 20th
century. J. Geophys. Res. 115, D16120. http://dx.doi.org/10.1029/2010JD014122.
Quaife, T., Quegan, S., Disney, M., Lewis, P., Lomas, M., Woodward, F.I., 2008. Impact of
land cover uncertainties on estimates of biospheric carbon fluxes. Glob. Biogeochem.
Cycles 22. http://dx.doi.org/10.1029/2007GB003097.
Ramankutty, N., Foley, J.A., 1998. Characterizing patterns of global land use: an analysis
of global croplands data. Glob. Biogeochem. Cycles 12, 667–685. http://dx.doi.org/
10.1029/98GB02512.
Sitch, S., Friedlingstein, P., Gruber, N., Jones, S.D., Murray-Tortarolo, G., Ahlström, A.,
Doney, S.C., Graven, H., Heinze, C., Huntingford, C., et al., 2015. Recent trends and
drivers of regional sources and sinks of carbon dioxide. Biogeosciences 12, 653–679.
http://dx.doi.org/10.5194/bg-12-653-2015.
89
Remote Sensing of Environment 203 (2017) 71–89
Sitch, S., Smith, B., Prentice, I.C., Arneth, A., Bondeau, A., Cramer, W., Kaplan, J.O.,
Levis, S., Lucht, W., Sykes, M.T., et al., 2003. Evaluation of ecosystem dynamics,
plant geography and terrestrial carbon cycling in the LPJ dynamic global vegetation
model. Glob. Chang. Biol. 9, 161–185. http://dx.doi.org/10.1046/j.1365-2486.2003.
00569.x.
Sterling, S.M., Ducharne, A., Polcher, J., 2013. The impact of global land-cover change on
the terrestrial water cycle. Nat. Clim. Chang. 3, 385–390. http://dx.doi.org/10.1038/
nclimate1690.
Still, C.J., Berry, J.A., Collatz, G.J., DeFries, R.S., 2003. Global distribution of C 3 and C 4
vegetation: carbon cycle implications. Glob. Biogeochem. Cycles 17, 1–14. http://dx.
doi.org/10.1029/2001GB001807.
Stocker, B.D., Roth, R., Joos, F., Spahni, R., Steinacher, M., Zaehle, S., Bouwman, L.,
Prentice, I.C., 2013. Multiple greenhouse-gas feedbacks from the land biosphere
under future climate change scenarios. Nat. Clim. Chang. 1–7. http://dx.doi.org/10.
1038/nclimate1864.
Ustin, S.L., Gamon, J.A., 2010. Remote sensing of plant functional types. New Phytol.
186, 795–816. http://dx.doi.org/10.1111/j.1469-8137.2010.03284.x.
Wang, J., Chagnon, F.J.F., Williams, E.R., Betts, A.K., Renno, N.O., Machado, L.A.T.,
Bisht, G., Knox, R., Bras, R.L., 2009. Impact of deforestation in the Amazon basin on
cloud climatology. Proc. Natl. Acad. Sci. U. S. A. 106, 3670–3674. http://dx.doi.org/
10.1073/pnas.0810156106.
Weedon, G.P., Balsamo, G., Bellouin, N., Gomes, S., Best, M.J., Viterbo, P., 2014. The
WFDEI meteorological forcing data set: WATCH forcing data methodology applied to
ERA-interim reanalysis data. Water Resour. Res. 50, 7505–7514. http://dx.doi.org/
10.1002/2014WR015638.
Weedon, G.P., Gomes, S., Viterbo, P., Shuttleworth, W.J., Blyth, E., Österle, H., Adam,
J.C., Bellouin, N., Boucher, O., Best, M., 2011. Creation of the WATCH Forcing Data
and Its Use to Assess Global and Regional Reference Crop Evaporation over Land
during the Twentieth Century. http://dx.doi.org/10.1175/2011JHM1369.1.
Welp, L.R., Keeling, R.F., Meijer, H.A.J., Bollenbacher, A.F., Piper, S.C., Yoshimura, K.,
Francey, R.J., Allison, C.E., Wahlen, M., 2011. Interannual variability in the oxygen
isotopes of atmospheric CO2 driven by El Niño. Nature 477, 579–582. http://dx.doi.
org/10.1038/nature10421.
Wilson, M.F., Henderson-Sellers, A., 1985. A global archive of land cover and soils data
for use in general circulation climate models. J. Climatol. 5, 119–143. http://dx.doi.
org/10.1002/joc.3370050202.
Zeng, N., 1999. Enhancement of Interdecadal climate variability in the Sahel by vege-
tation interaction. Science 286, 1537–1540. http://dx.doi.org/10.1126/science.286.
5444.1537.
Zeng, N., Neelin, J.D., 1999. A land–atmosphere interaction theory for the tropical de-
forestation problem. J. Clim. 12, 857–872.
Zhao, K., Jackson, R.B., 2014. Biophysical forcings of land-use changes from potential
forestry activities in North America. Ecol. Monogr. 84, 329–353. http://dx.doi.org/
10.1890/12-1705.1.