Introduction
Due to the increasing demand of a growing population for food and fibre, as
well as for bioenergy, greater anthropogenic pressures on the global land
area are expected. Today, carbon dioxide (CO2) emissions resulting from
land-use and land-cover (LUC) change are the second largest contributor to
anthropogenic emissions to the atmosphere after fossil fuel combustion
(Le Quéré et al., 2015a), and they are a factor that is associated with
large uncertainties. LUC and its changes include the processes when land
is converted from one land-cover type to another (e.g. the conversion of
forest to cropland or grasslands to pastures), and the effects of LUC
related to the management of the land, such as cropping practices,
fertiliser use, irrigation and different types of tillage. LUC changes
affect the cycling of carbon (C), energy, water and other nutrients
(phosphorous, nitrogen), in many cases enhancing greenhouse gas (e.g.
CO2, N2O, CH4) emissions from agricultural soils and pastures
when compared to natural vegetation and altering species composition. These
changes go hand in hand with altered characteristics such as surface albedo,
surface aerodynamic roughness and rooting depth
(Pongratz et al., 2010).
Conversions from natural vegetation to croplands and pastures generally
reduce C stored in vegetation (Baccini et al., 2012),
decrease soil C stocks in croplands but not in pastures
(Guo and Gifford, 2002; McLauchlan, 2006) and, unless fertilised, reduce soil
nitrogen (N) pools (McLauchlan, 2006). The alteration of C and N pools are mainly a result of initial deforestation and
of the decreased litter input due to biomass extraction upon harvest and
accelerated soil decomposition rates, the latter being stimulated through
management practices such as tillage or a changed microclimate at the soil
surface. However, in some regions croplands show increased C sequestration
potential compared to the natural vegetation owing to enhanced growth under
improved agricultural practices including fertilisation and irrigation
(Ciais
et al., 2010; Schulze et al., 2010). Legacy fluxes can change C due to, e.g.,
an imbalance between reduced litter input and decomposing dead biomass and
affect LUC emissions over decades or more
(Gasser
and Ciais, 2013; Houghton, 2010; Krause et al., 2016; Pugh et al., 2015).
During vegetation recovery on abandoned agricultural land, secondary land
ecosystems sequester C due to regrowing vegetation and reaccumulation of C
in soils. These LUC-related processes determining regional sources and sinks
of C entailed a global total net C flux to the atmosphere over the past
centuries
(deB
Richter and Houghton, 2011; Houghton et al., 2012; McGuire et al., 2001; Le
Quéré et al., 2015a).
A number of studies have recently highlighted the importance of different
definitions when assessing the net carbon flux from LUC (ELUC)
related to the fact that in different studies different LUC component fluxes
are included in the overall ELUC calculation (Gasser and Ciais,
2013; Pongratz et al., 2014). Likewise, it is important to consider whether
or not historical effects of environmental change are included in assessments
of cleared C stocks as part of ELUC. Less focus so far has been
put on the explicit datasets of historical land use employed (Le
Quéré et al., 2014). A limited number of historic LUC reconstructions
are available on a global scale, mostly at 0.5∘ spatial resolution
(Hurtt et al., 2011, 2016; Kaplan et al., 2012; Klein Goldewijk, 2016; Klein
Goldewijk et al., 2011; Olofsson and Hickler, 2008; Pongratz et al., 2008;
Ramankutty and Foley, 1999), two of which are very similar (datasets of Hurtt
et al., 2011 and Klein Goldewijk et al., 2011 are consistent when the corresponding versions are compared). On a continental scale some higher-resolution reconstructions exist, for instance
for Europe (Fuchs et al., 2015b; Kaplan et al., 2009; Williams, 2000). Most
reconstruction approaches combine information on current and recent
historical LUC from national statistics with estimates of global population
distribution and growth as the main driver of historical LUC. Model
assumptions are made to fill data gaps and extrapolate the available
information to create subnational patterns, and therefore large uncertainties
arise both from the original data sources and the modelling assumptions (see,
e.g., Klein Goldewijk and Verburg, 2013). However, reconstructions on continental
scales are able to use a more data-driven approach (e.g. Fuchs et al., 2013,
2015c) compared to global land reconstructions, since the data availability
is often better for these study areas.
Most historical LUC reconstructions focus on the difference in net area
under natural, cropland or pasture vegetation cover at a grid location
between two time steps (net land changes) instead of explicitly showing the
sum of the absolute value of all land transitions occurring on a sub-grid
scale (gross land changes). In particular over coarser grid resolutions,
gross land-cover changes allow a deeper view of LUC, tracking land
conversion events such as the parallel expansion and abandonment of
agricultural land, e.g. as in shifting cultivation (cycle of cutting forest
for agriculture and abandoning it after some years of usage, followed by a
period of fallow with regrowing forests). This entails altered
biogeochemical dynamics within different subsections of a grid cell; e.g.
secondary land acts as a C sink during vegetation regrowth, while additional
land clearing leads to a relatively rapid loss of C stocks in vegetation and
soils, along with other changes in vegetation composition, nutrients and
biogeophysical properties (e.g. Houghton et
al., 2012). Accounting in ecosystem models separately for the effects of
individual transitions (e.g. 10 % of an area converted from natural
vegetation to cropland while another 10 % of cropland is abandoned for
regrowth over the same time period) will therefore lead to a very different
response of ecosystem states and fluxes compared to the effects of net
changes, which in this case would be zero.
The availability of land-use information including gross land transitions is
limited due to a lack of reliable historical information to determine it.
However, a few datasets exist representing gross land transitions, such as
the global dataset by Hurtt et al. (2011)
(recently updated; see Hurtt et al., 2016), who
provide model estimates of shifting cultivation in certain tropical areas
based on a map of Butler (1980) and assumed land rotation
rates. Fuchs et al. (2015b)
recently estimated gross land changes for Europe over the 20th century
based on empirical evidence. As the number of gross land transitions can
greatly exceed the number of net transitions at spatial resolutions
typically employed for global studies, neglecting these can lead to a
serious underestimation of LUC dynamics with implications for
biogeochemical, ecological and environmental assessments
(Fuchs
et al., 2015a, b; Stocker et al., 2014; Wilkenskjeld et al., 2014).
Earlier studies revealed significant differences when ecosystem C dynamics
were simulated when accounting for gross land changes in areas of shifting
cultivation in addition to net changes as specified by
Hurtt et al. (2011) (e.g.
Shevliakova et al.,
2013; Stocker et al., 2014; Wilkenskjeld et al., 2014). Others have
implemented their own assumptions on spatial distribution and the rotation
scheme under shifting cultivation and combined these with C-cycle
calculations (Olofsson and Hickler, 2008; Stocker et
al., 2014).
In this study we use a state-of-the-art dynamic global vegetation model
(DGVM) to calculate ecosystem C stocks and fluxes in response to different
LUC reconstructions (1) to explore the consistency of different LUC
representations and to quantify the uncertainty in ecosystem C stocks and
fluxes, including ELUC, resulting from the different reconstructions
and (2) to quantify the effect of accounting for gross land transitions in
addition to net changes in LUC. We use five historical LUC reconstructions,
four of which are global and one of which is only for the European domain. One of the
global datasets represents gross transitions due to shifting cultivation in
certain tropical regions, and the European dataset represents gross
transitions from all sources in Europe. We apply a model with a representation
of LUC and changes therein, including a number of crop functional types and
C–N dynamics in natural vegetation and crops. We exclude wood harvest as a
form of forest management that can be represented as gross land transitions
from our analysis as, although national data on wood harvest are available,
its parameterisation in models is poorly constrained on a global scale; e.g.
the effects strongly depend on assumptions on the harvest type (clear cut,
selective logging, or a mixture of both) or assumptions regarding turnover
times of harvested C (Wilkenskjeld et al., 2014).
Raw characteristics of LUC datasets used in this
studya.
Land-use model
Time period
Model version and
From
to
Time steps
Representation of
Spatial
Original spatial
Land-cover classes
reference
LUC transitions
coverage
resolution
LUH1
1500–2005: LUH1 (Hurtt et al., 2011); extension until 2014: Le Quéré et al. (2015b)b
AD 1500
2014c
annual
net/gross
global
0.5∘
cropland, pasture, primary natural vegetation, secondary natural vegetation, urban
RAMA
1700–1992: Ramankutty and Foley (1999); extension until 2007: Ramankutty (2012)
AD 1700
2007
annual
net
global
0.5∘
cropland, pasture, primary natural vegetation, secondary natural vegetation, urban
HYDE
10 000 BC to AD 2000: HYDE3.1.1 (Klein Goldewijk et al., 2010, 2011); extension until 2005: see Klein Goldewijk et al. (2015); extension until 2013: see Le Quéré et al. (2015a)
10 000 BC
2013
annuald
net
global
5 arcmin
cropland, pasture, natural vegetatione
LUH2
850–2015: LUH2 v2, release 14 Oct 16 (Hurtt et al., 2016)
AD 850
2015
annual
netf
global
0.25∘
C3/C4 annual/perennial crops, C3 N-fixing crops, managed pasture, rangeland, primary natural vegetation, secondary natural vegetation, forest, non-forest, urban
HILDA
HILDA v2.0 (Fuchs et al., 2015b)
AD 1900
2010
decadal
netg/gross
EU27h plus Switzer-land
1 km
cropland, grassland (incl. managed pastures and shrubland), forest, settlements, water, other land (glaciers, sparsely vegetated areas, beaches and water bodies)
a Note that these datasets are preprocessed into more
consistent characteristic sets before carrying out simulations – see the “Methods” section
and Table 2. b Note that this version of LUH1 is based on an
early version of HYDE 3.2, which is different from HYDE version 3.1.1 as used
below; see the “Methods” section. c End date of historical land-use dataset.
d Decadal data until 2000. e Natural vegetation is calculated
as a remainder. f Only the net dataset was used in this study
(see the “Methods” section). g The HILDA net dataset used in this study was
derived from the gross dataset (see the “Methods” section). h European Union
2007–2013.
Methods
Land-use datasets
For the global scale, three historical LUC datasets were selected that are
frequently used for ecosystem modelling studies, as well as one recently
released dataset that is expected to be frequently used in the future. These
datasets run at least from 1700 to present and are also the basis for future
LUC scenarios (e.g. van Asselen and
Verburg, 2013; Hurtt et al., 2011). For Europe we additionally considered
one recently published dataset running from 1900 to 2010.
Table 1 provides an overview of the LUC datasets
and their characteristics.
Ramankutty and Foley (1999) (RAMA) published changes in
cropland area for the period 1700 to 1992 at a 5 min global resolution. The
dataset was built based on historical cropland inventory data at national
and subnational levels in combination with a remote-sensing-derived cropland
map for 1992. The algorithm to hindcast LUC distributed the historical
cropland area within political units, i.e. the 1992 cropland areas are
scaled for each political unit so that the cropland total matches historical
inventory data. Therefore, the reconstructed changes in historical croplands
are consistent with the history of human settlement and patterns of economic
developments, although they do not resolve changes in LUC dynamics below the
smallest spatial entity in the inventory data. The analysis was revised in
2012 so that it also accounts for pasture areas, and the dataset was extended
until 2007 at a 0.5∘ × 0.5∘ spatial resolution
(Ramankutty, 2012). Natural areas are calculated as the
remainder of cropland and pasture areas. Here we apply the revised
0.5∘ × 0.5∘ version.
The History Database of the Global Environment (HYDE) 3.1.1 (Klein Goldewijk et al.,
2010, 2011) provides spatially gridded maps of cropland and pastures at a 5 min resolution for the period 10 000 BC to AD 2000. Here, historical
population data and national and subnational statistics of change in
agricultural area (mainly the United Nations Food and Agriculture
Organization, FAO, supplemented with numerous other statistics) were
combined to a land-use per capita relationship. Land use was then allocated
for the present day based on satellite-derived land cover for 2000 and for the
past by a combination of this base map with a number of weighting and
suitability factors such as population density and habitat information on
soil suitability, distance to rivers, slopes, etc. The temporal resolution is 10 years for the historical period after 1700 and annual after 2000. The HYDE
dataset used here was extended until 2005 (Klein
Goldewijk et al., 2015) and later until 2013 in the 2014 global carbon
budget analysis
(Le
Quéré et al., 2015a).
Land-use types and transitions in global and European (EU27 + CH)
LUC reconstructions. The evolution of absolute land area of croplands, pastures
and natural vegetation (including barren land) in different (a) global
historical land-use reconstructions (LUH1: solid line; RAMA: dash-dotted
line; HYDE: dotted line; LUH2: dashed line) and (b) European land-use
reconstructions (HILDA: dashed line; LUH1: solid line). Land area
experiencing gross and net land transitions on a global scale (c) and for
Europe (d). Note the change to annual resolution in the LUH1 reconstruction
after the year 2000.
The Hurtt et al. (2011) Land Use Harmonization v1 (LUH1)
database is based on the land-use data of HYDE (Klein Goldewijk et al.,
2010, 2011) for the historical period (1500–2005) and combines these with
national statistics of historical wood harvest and assumptions regarding
shifting cultivation in certain tropical regions. Additional assumptions
were made regarding the prioritisation of land for conversion and logging,
the wood harvest spatial patterns and the residence time of land in
agricultural use in shifting cultivation areas. LUH1 data provide fractional
data on cropland, pasture, primary and secondary vegetation as well as gross
transitions between land-use states based on shifting cultivation at a 0.5∘ × 0.5∘ spatial resolution. Secondary land is
defined as natural land that was previously used for agriculture and is
recovering from this disturbance. Shifting cultivation is implemented as
bidirectional LUC change with an assumed rotation period of 15 years,
corresponding to an annual turnover rate of 6.7 % of the area
(Hurtt et al., 2011). Although the history of shifting
cultivation is not known, it is today present mainly in tropical regions; therefore, in the LUH1 dataset it is limited to certain tropical regions for
the historical period (Fig. S1 in the Supplement) Olofsson and Hickler, 2008). The LUH1
dataset was extended until 2014 for the 2015 global carbon budget analysis
(Le
Quéré et al., 2015b), using an early version of HYDE 3.2 as the
basis (now published in final version as Klein Goldewijk,
2016) and following the same methodology as described in
Hurtt et al. (2011). The version of LUH1 used in this
study is therefore a more recent development than that used for CMIP5
experiments (Taylor et al., 2012), but an earlier
version than the very recent LUH2 release (Hurtt et al.,
2016). As LUH1 is a modelled product that is based on the underlying HYDE
database, these products are very similar when the corresponding versions of
each dataset are considered (Hurtt et al., 2011). Note that
the version of HYDE used for our study (version 3.1.1, see above) is not the
same as the version of HYDE that underlies the LUH1 data used here (early
version of HYDE 3.2). HYDE 3.1.1 and 3.2 differ in terms of driving
population data and the algorithms used (Klein Goldewijk et al., 2016). The
HYDE and LUH1 data used in this study therefore differ in both their spatial
and temporal distribution of land-use fractions (Fig. 1; Klein Goldewijk et
al., 2016).
In addition, we used the very recent release of the Land Use Harmonization
v2 (LUH2; Hurtt et al., 2016) that was developed for
CMIP6 intercomparison project
(Eyring et al., 2016). This
global dataset covers the period 850–2015 and follows the same methodology
as LUH1 described above, but uses HYDE version 3.2 (Klein
Goldewijk, 2016) as a basis, along with updated attributes on wood harvest
and shifting cultivation, a higher spatial resolution of 0.25∘,
more detailed land-use transitions and additional land management
information (such as irrigation and fertiliser use). From the LUH2 dataset,
only LUC states and net land transitions aggregated to a spatial resolution
of 0.5∘ × 0.5∘ were considered in this study, as a very
new dataset that very likely will frequently be applied for modelling studies
in the future. As LUC patterns of LUH2 are directly based on HYDE3.2, these
two datasets are very similar in their land-use information (data not shown), and therefore resulting ecosystem C stocks and fluxes are expected to be
very similar. The detailed LUC gross transitions provided by LUH2 could not
be preprocessed in time to be used here.
The only non-global LUC dataset considered here, the HIstoric Land Dynamics
Assessment
(HILDA; Fuchs et
al., 2013, 2015b), reconstructs gross and net land changes for the EU27 (EU
from 2007 to 2013) plus Switzerland at a 1 km spatial resolution (Fig. S1). Net
land conversions are based on national statistics. To account for gross
changes, empirical evidence (mostly time series of large-area LUC maps and
national surveys) on historic gross LUC changes was aggregated to derive an
overall gross / net ratio per LUC class and a relative weighted land
conversion matrix, both of which were applied to national net change data. The
allocation of LUC fractions to grid cells was done based on probability maps
for each LUC class, forest volume stock maps and large-scale historic LUC
maps (Fuchs et al., 2015c). An aggregated
version of the CORINE2000 land-cover dataset was used as base map for the
year 2000. The initial LUC dataset that was built based on empirical
evidence covers 1950–2010 in decadal steps but was extrapolated back to 1900
to assess the long-term impacts of changes in LUC. For each time step the
1 km grid cells were classified as being dominated by settlement, cropland,
forest, grassland (including managed pastures), other land (glaciers, sparsely
vegetated areas, beaches and water bodies) or water. Here, we consider only
the gross LUC reconstruction of HILDA
(Fuchs et al., 2015b) and
derive net LUC changes from gross land transitions. In the original HILDA
net LUC reconstruction (Fuchs
et al., 2015b), the results differ spatially from our net reconstruction due
to the use of different land allocation mechanisms under net and gross
changes in their analysis.
LPJ-GUESS ecosystem model
We use the LPJ-GUESS (Lund–Potsdam–Jena General Ecosystem Simulator) DGVM (Sitch et al., 2003; Smith
et al., 2001) with updates for land-use change
(Lindeskog et al., 2013) and C–N
coupling in natural vegetation and crops
(Olin et al., 2015; Smith et al.,
2014) allowing for the simulation of nitrogen limitation on plant and crop
development. Three distinct land-use types are used (natural vegetation,
pasture and cropland) with natural vegetation modelled by nine woody and two grass
plant functional types (PFTs) (as in Smith et al.,
2014), which are distinguished in terms of their bioclimatic preferences,
photosynthetic pathways and growth strategies. Vegetation structure,
dynamics and competition between age cohorts of a PFT population are
explicitly represented in the LPJ-GUESS model. Croplands are represented by three crop functional types (CFTs) that are parameterised using information
on summer wheat, winter wheat and maize, with crop-specific processes including dedicated carbon allocation and phenology;
explicit sowing and harvest representation; irrigation; fertilisation; and cover
crops represented by grass growing in between cropping periods (Olin et al., 2015). Pastures are modelled using competing C3 and C4 grass PFTs, where each
year 50 % of the C and 12.5 % of the N in above-ground biomass was
removed as a representation of grazing
(Krause
et al., 2016; Lindeskog et al., 2013).
In LPJ-GUESS, upon conversion of natural land to cropland and pastures, the
natural vegetation is cleared and 97 % of wood (of which stem wood is
65 % and branches and coarse roots are 32 %), and 71 % of leaf biomass is harvested.
Out of the harvested stem wood, one-third goes to a product pool with a
turnover time of 25 years. The rest of the harvested biomass is oxidised and
released to the atmosphere, while the remaining biomass enters the litter
pool (see Lindeskog et al.,
2013). In reductions of the natural vegetation area, young stands (but older
than 15 years, the rotation period in shifting cultivation assumed by
Hurtt et al., 2011) are converted before older stands.
Following agricultural abandonment, natural vegetation recolonises the land
in a typical succession from herbaceous to woody plants if environmental
conditions are suitable for tree growth. Competition for resources and light
among age cohorts of woody PFTs is simulated directly through gap dynamics
(see, e.g., Bugmann, 2001).
The model has been evaluated extensively and has demonstrated skills in
capturing large-scale vegetation patterns (Hickler et
al., 2006, 2012) and dynamics of the terrestrial carbon cycle
(Ahlström
et al., 2012; Morales et al., 2005; Olin et al., 2015; Piao et al., 2013;
Pugh et al., 2015; Smith et al., 2014). The carbon flux response was close
to the ensemble mean in a recent intercomparison of nine dynamic global
vegetation models
(Sitch
et al., 2015).
Overview of LPJ-GUESS simulations carried out as part of this
study.
Abbreviation
First year
Last year
Representation of LUC transitions
Spatial coverage
LUH1
1700
2014
gross
global
1700
2014
net
global
1700
2014
LUC fixed to 1700
global
RAMA
1700
2007
net
global
1700
2007
LUC fixed to 1700
global
HYDE
1700
2013
net
global
1700
2013
LUC fixed to 1700
global
LUH2
1700
2015
net
global
1700
2015
LUC fixed to 1700
global
HILDA
1900
2010
gross
EU27 + CH
1900
2010
net
EU27 + CH
1900
2010
LUC fixed to 1900
EU27 + CH
LUH1
1900
2014
net
EU27 + CH
1900
2014
LUC fixed to 1900
EU27 + CH
Simulation protocol
LPJ-GUESS was run at a 0.5∘ × 0.5∘ resolution with
simulations beginning in the year 1700. CRU TS 3.21 historical global climate
data (University of East Anglia Climatic
Research Unit, CRU, 2013) was used for the period 1901–2014. Climate data for 1700–1900
were provided by repeating 1901–1930 CRU climate with de-trended
temperature data. Climate data for 2014 were repeated for the year 2015.
Atmospheric CO2 forcing was provided from observations from ice cores
and, later in the 20th century, atmospheric measurements
(Tans and Keeling, 2015), with a value of 286.4 ppmv used
from 1700 until the beginning of this dataset in 1860 and a final
atmospheric mixing ratio of 399.0 ppmv in 2015. Simulations were spun-up for
500 years using land-use fractions and the CO2 mixing ratio from the first
simulation year and de-trended climate data of the first 30 simulation
years, with a longer spin-up for soil carbon pools
(see Smith et al., 2014). Model spin-up was
therefore identical for net and gross land changes. In order to assign
cropland areas to crop functional types, global crop cover fraction was
partitioned based on Portmann et al. (2010) and mapped to
LPJ-GUESS crop types, as described in Olin et al. (2015). Crop type fractions were held constant throughout the simulations.
Where cropland was expanded into a hitherto uncropped grid cell, average
CFT fractions of the nearest neighbouring cropland cells were used to
populate it. Past values of global nitrogen deposition were taken from
simulations from
Lamarque
et al. (2010, 2011) and nitrogen fertilisation of crops was estimated as
in Zaehle et al. (2011). LPJ-GUESS simulations are
summarised in Table 2.
Global simulations starting in 1700 were carried out using the four net and
one gross LUC dataset (LUH1 net, RAMA net, HYDE net, LUH2 net, LUH1 gross),
and those for Europe starting in 1900 used two net and one gross LUC dataset
(HILDA net, LUH1 net, HILDA gross). For these, all LUC input data were
aggregated to a 0.5∘ spatial resolution and decadal HILDA, HYDE and
LUH1 LUC data were interpolated linearly to annual time steps. LPJ-GUESS
uses annual land-use states of the classes “cropland”, “pasture”, “natural
vegetation” and “barren land” (no vegetation; e.g. water- or ice-covered) as
input for net LUC runs, which are complemented for gross LUC runs by annual
gross transitions for each combination of two land-use classes. Land-use
states of RAMA, HYDE and LUH2 were used directly. To generate net
transitions from LUH1, annual land-use states were derived from land-use
states in 1700 and gross transitions from 1700 to 2014. HILDA land-use
matrices providing land-use states and transitions together in the form of an
integer land-use category were translated to annual land-use states and
gross transitions for each combination of two land-use classes. Although
some of these LUC products represent changes between forested and
non-forested land, in the simulations done here, only the changes between
the classes croplands, pastures (FAO category “permanent pasture”), natural
vegetation and barren land (available for LUH1, LUH2 and HILDA) were
considered; the composition of natural vegetation was simulated directly by
LPJ-GUESS. Primary and secondary vegetation as in LUH1 and LUH2 (wood
harvest) and the forest class in HILDA and LUH2 were represented by natural
vegetation. The HILDA LUC classes of settlements, water and other land were
aggregated to the LUC class barren. The grassland class in HILDA comprises both meadows and natural grasslands, which were both represented using the pasture class in LPJ-GUESS, a reasonable assumption for Europe due to the small area of truly natural,
unmanaged grassland in Europe (Wilkins et al., 2003). For
global simulations a land mask was used that includes only cells of the
ice-free land area for which all four global LUC datasets provide data
(58 790 cells). For Europe, all 0.5∘ grids that contained at
least one HILDA cell were used (2486 cells).
We examine differences caused by the different LUC reconstructions on net
land-use flux (ELUC), deforestation flux, net primary productivity
(NPP) and ecosystem C stocks in vegetation and soils. ELUC is
calculated as the difference between a model simulation with transient
historical LUC change (gross or net LUC change) and a simulation with
constant LUC distribution as in the first simulation year
(Table 2, LUC fixed to 1700/1900). All simulations
are driven by varying climate, atmospheric CO2 mixing ratio, N
deposition and N fertilisation
(see, e.g., Le
Quéré et al., 2015; Pongratz et al., 2014). This method includes
the effects of LUC and changes therein interacting with climate and atmospheric
CO2 and is consistent with definition 1 of
Gasser and Ciais (2013) and D3 of
Pongratz et al. (2014). In the same
way, the effect of accounting for LUC on NPP and C stocks in vegetation and
soils was calculated as the difference between the simulation with gross or
net LUC changes and the respective reference simulation. Soil C includes
both C in soils and litter. The deforestation flux is the C released upon
land conversion only. In the calculation of net cumulative ELUC for
global simulations the first 50 simulation years were ignored because of
high carbon fluxes in the first decades of gross simulations, which
reflected a re-equilibration under LUC including gross transition rates
(i.e. shifting cultivation) that were not part of model spin-up and
effectively reflect emissions from shifting cultivation that occurred before
1700 (see, e.g., Stocker et al., 2014). Because gross
transitions in Europe do not follow a systematic rotation such as shifting
cultivation areas in the global simulations, this effect is not so directly
applicable or apparent here and cumulative ELUC was calculated starting
in 1900. We restrict our analysis of the effects of using different LUC
reconstructions to the global scale, which has direct relevance for the
global carbon budget; a detailed analysis of regional differences and
processes is beyond the scope of this study.
Changes in C stocks and fluxes in four global reconstructions of net
and gross LUC changes. Land-use change flux (ELUC) and cumulative
land-use flux ELUC; net primary productivity (NPP); C stocks in
vegetation and soils. Values are always given as 10-year averages (except LUC
areas and cumulative flux).
Averaging period
Calculation
Unit
LUH1
RAMA
HYDE
LUH2
Average and
LUH1
difference
net
net
net
net
uncertainty for
gross
LUH1
four net LUC models
(gross–net)
Natural area
1700
Anat
106 km2
122.48
124.35
126.54
123.23
124.15 ± 1.77
122.48
0
Natural area
2007
Anat
106 km2
85.39
90.31
84.97
85.01
86.42 ± 2.60
85.39
0
Pasture area
1700
Apas
106 km2
7.45
4.89
3.32
6.61
5.57 ± 1.84
7.45
0
Pasture area
2007
Apas
106 km2
32.38
27.04
32.81
32.73
31.24 ± 2.81
32.38
0
Cropland area
1700
Acrop
106 km2
2.80
3.49
2.86
2.89
3.01 ± 0.32
2.80
0
Cropland area
2007
Acrop
106 km2
14.96
15.38
14.95
14.99
15.07 ± 0.21
14.96
0
Change in natural area
1700–2007
Anat
106 km2
-37.09
-34.04
-41.57
-38.22
-37.73 ± 3.11
-37.09
0
Change in pasture area
1700–2007
Apas
106 km2
+24.93
+22.15
+29.49
+26.12
+25.67±3.04
+24.93
0
Change in cropland area
1700–2007
Acrop
106 km2
+12.16
+11.89
+12.09
+12.10
12.06 ± 0.12
+12.16
0
Total area under transition
1700–2007
ΣAtrans
106 km2
51.92
78.59
64.62
57.28
63.10 ± 11.56
243.80
+191.88
Change in area under transition
1850–1960
δAtrans
km2 a-1
+5137
+7259
+6703
+5177
6069 ± 1077
+5549
+412
ELUC
1750–2007
ELUCNet/Gross
Pg C a-1
0.81
0.87
0.89
0.77
0.84 ± 0.05
0.94
+0.13
1750–2014
ELUCNet/Gross
Pg C a-1
0.84
–
–
0.80
–
0.96
+0.12
1980–1989
ELUCNet/Gross
Pg C a-1
1.10
1.40
1.55
1.31
1.34 ± 0.19
1.28
+0.18
1990–1999
ELUCNet/Gross
Pg C a-1
1.18
1.57
2.65
1.36
1.69 ± 0.66
1.41
+0.23
1998–2007
ELUCNet/Gross
Pg C a-1
1.17
2.00
2.06
1.26
1.62 ± 0.47
1.38
+0.20
2005–2014
ELUCNet/Gross
Pg C a-1
1.50
–
–
1.67
–
1.64
+0.14
Cumulative ELUC from 1750
2007
ΣELUCNet/Gross
Pg C
210.02
225.18
228.95
199.12
215.82 ± 13.82
242.04
+32.02
2014
ΣELUCNet/Gross
Pg C
222.29
–
–
212.70
–
255.27
+32.98
NPP
1700–1709
NPPNet/Gross
Pg C a-1
50.18
52.10
52.21
50.43
51.23 ± 1.07
50.04
-0.14
1998–2007
NPPNet/Gross
Pg C a-1
58.90
59.79
60.85
59.14
59.67 ± 0.87
57.49
-1.42
2005–2014
NPPNet/Gross
Pg C a-1
59.95
–
–
60.18
–
58.46
-1.49
Change in NPP due to LUC
1998–2007
NPPNet/Gross - NPPRef
Pg C a-1
-1.92
-3.32
-2.44
-1.96
-2.41 ± 0.65
-3.33
-1.42
2005–2014
NPPNet/Gross - NPPRef
Pg C a-1
-2.26
–
–
-2.30
–
-3.75
-1.49
Vegetation C
1700–2014
VegCNet/Gross
Pg C
435
–
–
444
–
419
–16
1700–1709
VegCNet/Gross
Pg C
464.18
495.70
497.19
467.59
481. 17 ± 17.71
461.35
-2.83
1998–2007
VegCNet/Gross
Pg C
414.62
439.10
435.22
424.53
428.37 ± 11.04
380.43
-34.20
2005–2014
VegCNet/Gross
Pg C
421.48
–
–
431.41
–
386.64
-34.84
Change in vegetation C due to LUC
1960–1969
VegCNet/Gross - VegCRef
Pg C
-109.49
-114.12
-108.21
-98.56
-107.60 ± 6.54
-137.06
-27.58
1998–2007
VegCNet/Gross - VegCRef
Pg C
-140.36
-153.41
-159.42
-134.36
-146.89 ± 11.54
-174.56
-34.20
2005–2014
VegCNet/Gross - VegCRef
Pg C
-148.43
–
–
-142.58
–
-183.27
-34.27
Soil C
1700–2014
SoilCNet/Gross
PgC
1425.85
–
–
1438.39
–
1420.81
-5.04
1700–1709
SoilCNet/Gross
PgC
1445.96
1511.65
1516.06
1453.68
1481.84 ± 37.15
1445.58
-0.38
1998–2007
SoilCNet/Gross
Pg C
1404.20
1471.83
1478.08
1419.90
1443.50 ± 36.97
1393.43
-10.77
2005–2014
SoilCNet/Gross
Pg C
1406.78
–
–
1421.70
–
1395.56
-11.22
Change in soil C due to LUC
1998–2007
SoilCNet/Gross - SoilCRef
Pg C
-75.80
-75.04
-73.54
-67.95
-73.08 ± 3.55
-86.57
-10.77
2005–2014
SoilCNet/Gross - SoilCRef
Pg C
-77.74
–
–
-70.59
–
-88.96
-11.22
Results
Historical land transitions
The most pronounced change in global vegetation cover over the historical
period is the deforestation of natural areas for conversion into croplands
and pastures (Fig. 1a), progressing at fairly low rates during the first
decades after 1700, followed by a steadily increasing trend from about 1860
until a slow-down and stabilisation sets in after about 1960. Total land
area transitions before 1850 (Fig. 1c) are below 100 000 km2 a-1,
from which they steadily increase with a rate of an additional ca. 6100 ± 1100 km2 a-1 under transition each year (average of four
LUC datasets 1850–1960, Table 3). Transitions in
all four LUC reconstructions peak between 1950 and 1960 when deforestation
due to expansion of agriculture in the tropics and pasture expansion in
grass- and shrub-dominated biomes was highest in the LUC reconstructions
(Fig. S2). After the 1960s all four LUC reconstructions assume continued
deforestation in the tropics at a lower rate and reforestation in Europe and
some parts of Northern America following the abandonment of agricultural
areas. Transitions around 1960 are believed to be influenced by the LUC
reconstruction process, when model assumptions for the historical period
before 1960 are merged with the records of the Food and Agriculture (FAO)
records available thereafter.
The four global net LUC datasets applied here differ primarily in the total
area of pasture and natural vegetation and in the regions in which these are
located; RAMA and LUH2, in terms of global absolute area under natural LUC,
lie in between LUH1 (lowest area under natural vegetation before 1960s) and
HYDE (highest area under natural vegetation before 1960s, see Fig. 1a).
Spatial patterns are generally more similar between LUH1, HYDE and LUH2 than
with RAMA. However, the lower amount of natural areas and higher amount of
pastures of LUH1 in the southern parts of Russia compared to the other three
reconstructions is noteworthy. Other major differences between all four
reconstructions occur in eastern Africa, eastern Europe and parts of the US
and Canada (maps not shown). After 1960, LUH1, HYDE and LUH2 are very
similar (showing major differences only in Australia). While the
deforestation trend is shown by all four global LUC reconstructions, the
absolute loss rates of natural vegetation differ, with HYDE being 14 %
above the average of the other three LUC models
(Table 3). For the present day, differences are largest
in natural areas and pastures, with RAMA reporting about 6 % more natural
areas and about 18 % less pasture areas in 2007 compared to the other
three reconstructions (Fig. 1a; Table 3).
Differences in pasture areas occur worldwide, but are somewhat higher in
Saudi Arabia, western China, Mongolia and Australia (maps not shown). Before
1950, differences in natural and pasture area between LUH1, LUH2 and HYDE on the one hand and
RAMA on the other exist predominantly in eastern Europe, southern parts of Russia and
eastern Africa.
Changes in C stocks and fluxes in reconstructions of net and gross
LUC changes for Europe (EU27 + CH). Values are always given as 10-year
averages (except LUC areas and cumulative flux).
Averaging period
Calculation
Unit
HILDA
LUH1
Average and
HILDA
difference
(net)
(net)
uncertainty∗ for
(gross)
HILDA
the LUC models
(gross–net)
Natural area
1900
Anat
106 km2
1.49
2.16
1.83 ± 0.47
1.49
0
Natural area
2010
Anat
106 km2
2.06
2.79
2.43 ± 0.52
2.06
0
Pasture area
1900
Apas
106 km2
1.62
0.97
1.30 ± 0.46
1.62
0
Pasture area
2010
Apas
106 km2
1.35
0.68
1.02 ± 0.47
1.35
0
Cropland area
1900
Acrop
106 km2
1.59
1.57
1.58 ± 0.01
1.59
0
Cropland area
2010
Acrop
106 km2
1.29
1.23
1.26 ± 0.04
1.29
0
Total change in natural area
1900–2010
Anat
106 km2
+0.57
+0.63
+0.60±0.04
+0.57
0
Total change in pasture area
1900–2010
Apas
106 km2
-0.29
-0.34
-0.28 ± 0.01
-0.29
0
Total change in cropland area
1900–2010
Acrop
106 km2
-0.28
-0.29
-0.32 ± 0.03
-0.28
0
Total area under transition
1900–2010
ΣAtrans
106 km2
1.47
1.87
1.67 ± 0.29
2.64
+1.17
ELUC
1900–1909
ELUCNet/Gross
Tg C a-1
19
38
29 ± 13
29
+9
1980–1989
ELUCNet/Gross
Tg C a-1
-25
-78
-51 ± 37
-26
-1
1990–1999
ELUCNet/Gross
Tg C a-1
-38
-84
-61 ± 33
-38
0
2001–2010
ELUCNet/Gross
Tg C a-1
-52
-80
-66 ± 20
-51
+1
Cumulative ELUC from 1900
1951–1960
ΣELUCNet/Gross
Tg C
586
1338
962 ± 532
915
+329
2010
ΣELUCNet/Gross
Tg C
-936
-1890
674 ± 48
-531
+406
NPP
1900–1909
NPPNet/Gross
Tg C a-1
1464
1517
1490 ± 38
1462
-2
2001–2010
NPPNet/Gross
Tg C a-1
2261
2361
2311 ± 71
2243
-18
Change in NPP due to LUC
1900–2010
NPPNet/Gross - NPPRef
Tg C a-1
-30
-10
-20 ± 14
-44
-14
2001–2010
NPPNet/Gross - NPPRef
Tg C a-1
-57
+10
-23 ± 47
-74
-18
VegC
1900–1909
VegCNet/Gross
Tg C
7755
9634
8694 ± 1328
7715
-40
2001–2010
VegCNet/Gross
Tg C
10 518
13 484
12 001 ± 2097
10 360
-159
Change in vegetation C due to LUC
1900–2010
VegCNet/Gross - VegCRef
Tg C
-58
-234
-146 ± 125
-199
-141
2001–2010
VegCNet/Gross - VegCRef
Tg C
+709
+1217
+963±359
+551
-159
Soil C
1900–1909
SoilCNet/Gross
Tg C
58 786
60 672
59 729 ± 1334
58 775
-11
2001–2010
SoilCNet/Gross
Tg C
60 016
62 188
59 761 ± 1536
59 761
-254
Change in soil C due to LUC
1900–2010
SoilCNet/Gross - SoilCRef
Tg C
-199
-314
-256 ± 81
-368
-169
2001–2010
SoilCNet/Gross - SoilCRef
Tg C
-29
+291
+131±226
-283
-254
∗ Note that the uncertainty given here is calculated
between two values.
In Europe the two historical LUC reconstructions show the expansion of areas
with regrowing natural vegetation after 1900 following land abandonment as a
result of intensification on high-production cropland (Fig. 1b). Net gain in
natural regrowth area from 1900 to 2010 is about 60 000 km2 (average of
two LUC datasets, Table 4). The rates of total land
conversion in Europe over the first half of the 20th century (Fig. 1d)
remain at a fairly constant level, with between 10 000 and 15 000 km2
being converted each year. Rates of land conversion are only higher between
1950 and 1970 and after 1990.
The European land-use reconstruction HILDA shows the same trend in LUC over
the historical period when compared with the same domain extracted from the
global LUH1 product (Fig. 1b), but the two datasets disagree notably with
respect to the absolute area of natural vegetation and pastures
(Table 4). HILDA shows substantially larger
fractions of pasture than LUH1 especially in Scandinavia and southern
Europe, while LUH1 shows higher pasture fractions than HILDA in central
Europe and the Baltic area. The higher areas of pasture in HILDA may result
from the fact that natural grasslands are included in the pasture class in
HILDA but not in the pasture class of LUH1 (see the “Methods” section). The peak in total
land conversion rates around the middle of the century is shown in HILDA in two
steps, with slightly higher rates in the 1950s and a maximum in the 1960s (13 700
and 19 100 km2 per year) and in LUH1 in one step with a more than
doubled rate in the 1950s compared to the previous decades (on average, about
42 600 km2 per year). From 1950 to 1960, the LUH1 dataset shows a rapid
decrease in pasture area of 1.5 × 105 km2 that is mainly reflected
in a significant gain in natural areas.
Effects of different land-use representations on global ecosystem C
stocks and fluxes: land-use flux (a), cumulative land-use
flux (b), deforestation emissions (c), net primary
productivity (NPP) (d), vegetation (e) and soil carbon
stocks (f). Flux values in (a) and (d) are given
as 15-year averages with original values in the background. NPP, vegetation
and soil C is shown for simulations experiencing net (or gross) LUC and in
addition for simulations where LUC was kept at 1700 levels (see the
“Methods” section; red lines in d, e, f).
Effects of different net LUC changes on carbon pools and fluxes
In LPJ-GUESS simulations all four global net LUC reconstructions resulted in
similar projections of the land-use change flux ELUC as a source of C
on the global scale, with the rate of emission accelerating from the early
1800s (Fig. 2a). Reflecting the time series of the land transitions (Fig. 1c), ELUC peaked in the 1950s with emissions of about 2.0 to
2.6 Pg C a-1. Mean ELUC was 1.2, 2.0, 2.1 and 1.3 Pg C a-1
for LUH1, RAMA, HYDE and LUH2, respectively, at the end of the historical
period (1998–2007, Table 3), and cumulative
ELUC since 1750 was between 199 Pg C for LUH2 and 229 Pg C for HYDE in 2007 (Fig. 2b; Table 3). From the
four reconstructions, projections based on HYDE (with LUH2 being very close)
resulted in the lowest emissions until the early 1900s, probably because of
the lowest conversion of natural areas to pastures until this period
compared to the other reconstructions (Fig. 1a). Also, when using the HYDE
product, a second peak of ELUC of around 2.7 Pg C a-1 occurred in
the late 1990s (15-year average) that is not seen in LUH1, RAMA and LUH2
reconstructions (between 1.2 and 1.6 Pg C a-1 in this period).
Global average NPP was simulated to increase strongly over the last century
due to the effect of higher vegetation productivity under an increased
atmospheric CO2 mixing ratio and (in cool areas) climate warming.
Compared to reference simulations with LUC fixed in 1700 (red lines in Fig. 2d), all four LUC representations showed a reduced increase in NPP over the
duration of the simulations (Table 3; Fig. S3),
with the reduction due to LUC at the end of the historical period (averages
1998–2007) being 1.9 Pg C for LUH1, 2.4 Pg C for HYDE, 3.3 Pg C for RAMA and
2.0 Pg C for LUH2 LUC reconstructions (Table 3).
The global total and the time series of NPP simulated with the four LUC
reconstructions were very similar for RAMA and HYDE and were about 2 Pg C
lower for LUH1 and LUH2 simulations for the entire simulation period (Fig. 2d), possibly as a result of a higher pasture area instead of natural
(woody) vegetation in the LUH1 and LUH2 data in the extratropical regions
of Africa and areas in eastern Europe.
For global C stocks, both in vegetation and soils, the positive trend
induced by CO2 fertilised vegetation growth (red lines in Fig. 2e and f) was counteracted by the effects of LUC change. A minimum of vegetation C
stocks during the simulation period was simulated for all LUC
reconstructions in the 1960s when LUC reduced vegetation C stocks on average
by 108 Pg C (Table 3) compared to the reference
simulations. Following the decline in conversion rates of natural into
managed land thereafter, vegetation C stocks increased in response to
vegetation productivity responding to the fertilising effect of increasing
atmospheric CO2 concentration. Vegetation C stocks at the beginning of
the simulation period were again similar for RAMA and HYDE on the one hand and for LUH1 and LUH2
reconstructions on the other with about 482 Pg C but on average about 31 Pg C lower for
LUH1 and LUH2 simulations because of the lower natural area in these datasets
around 1700 (Table 3).
Time trends in soil C stocks over the simulation period followed similar
trends as vegetation C stocks, albeit with a 5- to 10-year time lag (Fig. 2e, f). Loss in soil C as a result of accounting for LUC was a direct effect
of the removal of biomass upon harvest, reducing the litter input in the
following years, and the increase in soil C decomposition rates associated
with tillage. On average, 73 Pg soil C was lost due to changes in LUC in the
2000s, with only a small variation of ±1 Pg C between LUH1, RAMA and
HYDE reconstructions and about 7 Pg C less for LUH2
(Table 3; Table S3 in the Supplement). Overall soil C stocks were
again more similar for RAMA and HYDE (average 1514 Pg C) and about 64 Pg C
lower for LUH1 and LUH2 at the beginning of the simulation period
(1700–1709; Fig. 2e; Table 3).
Effects of different land-use representations on ecosystem C stocks
and fluxes for Europe (EU27 + CH): land-use flux (a), cumulative land-use
flux (b), deforestation emissions (c), net primary productivity (NPP) (d),
vegetation (e) and soil C (f). Flux values in (a) and (d) are given as
15-year averages with original values in the background. NPP, vegetation and
soil C is shown for simulations experiencing net (or gross) LUC and in
addition for simulations where LUC was kept at 1700 levels (see the “Methods” section; red
lines in d, e, f).
In Europe, LUC caused the emission of C until the 1960s but turned into a sink
thereafter (Fig. 3a, negative ELUC) as a result of the reduction in
pastures and also croplands in the second half of the last century and the
regrowth of natural (woody) vegetation (Fig. 1b). This development is shown
in simulations with both HILDA and LUH1; however, the magnitude of the effect
of LUC on C stocks and fluxes was somewhat stronger in simulations applying
the LUH1 dataset, due to a higher deforestation rate in LUH1 until the 1950s
(Fig. 3c; absolute land transitions were similar for both LUC datasets, Fig. 1d). In this
sense, ELUC decreases from about 19 Tg C a-1 (HILDA)
and 38 Tg C a-1 (LUH1) in the 1900s to about -52 Tg C a-1 for
HILDA and -80 Tg C a-1 for LUH1 in the 2000s, and cumulative ELUC
from 1900 to 2010 was -936 Tg C for HILDA and -1890 Tg C for LUH1
(Table 4).
As in the global simulations, a positive trend in NPP from 1900 was also
simulated for Europe, which is linked to increasing CO2 fertilisation
(Fig. 3d). Accounting for changes in LUC reduced NPP in simulations applying
HILDA by 57 Tg C a-1 but only slightly changed NPP in the 2000s when
applying LUH1 (increase of 10 Tg C a-1, Table 4) because of highly productive croplands in central Europe (Fig. S4a).
NPP simulated with HILDA was between 50 and 100 Tg C a-1 lower over the
simulation period than simulated with LUH1 (Table 4). This also resulted from the lower share of natural areas in HILDA, as
opposed to pastures, and thus lower productivity.
Similar trends in vegetation and soil C were simulated with both datasets,
with changes dominated by deforestation over the first decades and
reforestation thereafter (see Fig. 1b). In the 2000s vegetation C stocks
were even higher under net LUC changes compared to the respective reference
simulations (Table 4). C stocks in the vegetation of
simulations using LUH1 were on average about 2000 Tg C and 25 % higher
than simulations applying HILDA (Fig. 3e, Table 4).
Differences in soil C stocks between the two LUC representations were small,
with soil C being about 1900 Tg C higher in LUH1 simulations (3 % of
soil C stocks projected with HILDA) at the beginning of the simulation
period (Table 4). In comparison to the trend in C
stored in vegetation, stocks in soils only increased from the 1950s on.
Effects of LUC on C stocks in vegetation and soils were stronger for
simulations applying LUH1 (Fig. 3e), showing increases in C stocks in both central and eastern Europe but decreases in southern countries (Fig. S4b, c).
Global and European effects of accounting for gross land
transitions
The global land area undergoing LUC when considering gross land transitions
based on the LUH1 dataset (Fig. 1c) was 4.7 times the net area converted
(total transitions 1700–2014, Table 3), with all
additional land transitions in this dataset being generated by shifting
cultivation in certain tropical regions (Fig. S1a). This increased the
global land-use flux ELUC by about 0.14 to 1.64 Pg C a-1 at the end of the historical period (2005–2014 average flux)
and resulted in cumulative ELUC being 33 Pg C higher in 2014 for gross
compared to net LUC simulations (Fig. 2a, b; Table 3). Global NPP was 1.5 Pg C a-1 (i.e. 3 %) lower in simulations of
gross land changes compared to the net simulations (Fig. 2d,
Table 3), which was an effect of lower mean levels
of forest canopy closure in the tropical areas subject to shifting
cultivation. For the same reason, the reduction in vegetation C stocks as an
effect of accounting for gross effects was high, with 35 Pg C, i.e. -8 %,
at the end of the simulation period. For soils, however, the effect was
relatively low compared to the absolute size of soil C stocks, with an 11 Pg C reduction (-0.8 %, 2005–2014; Fig. 2e
and f; Table 3). The reduction in vegetation C stocks by
the effects of LUC changes further increased by 24 % when accounting for
gross LUC and for soil C stocks by 14 % (2005–2014,
Table 3).
For Europe, the HILDA gross dataset predicted land transitions (Fig. 1d)
that were about 1.4 times higher when accounting for gross transitions
relative to net LUC changes (total transitions 1900–2010,
Table 4) (see also
Fuchs et al., 2015b) with
significant gross land changes occurring all over Europe (Fig. S1b). As a
result, gross ELUC was enhanced by about 11 Tg C a-1 in the beginning of the simulation period (1901–1910) compared to net ELUC.
Cumulative gross ELUC was -531 Tg C in 2010, or 406 Tg C higher than
ELUC under net LUC transitions (-936 Tg C), representing a reduced
cumulative sink as the result of higher previous emissions from LUC when
considering gross transitions (Fig. 3b; Table 4).
NPP was also lower in gross simulations; however, differences were small (-18 Tg C a-1; see Table 4) and the gross
simulation followed the same trend as the net LUC simulation. For C stocks
on the European level, the difference between net and gross simulations was
-158 Tg C for vegetation carbon and -254 Tg C for soil C stocks at the end
of the simulation period (2001–2010; Fig. 3e and f; Table 4).
Discussion
Uncertainties in carbon stocks and fluxes due to the choice of historical
LUC reconstruction
Resulting from the fact that historical reconstructions of land use and its
changes are inherently uncertain because of the limited existing data base
that needs complementary assumptions (e.g. on land rotation times), it is
widely acknowledged that a key uncertainty in estimating changes in C stocks
and fluxes as a response to LUC change stems from the choice of the LUC
dataset (e.g. Houghton et al., 2012;
Jain et al., 2013). With a detailed representation of succession when
natural vegetation recolonises abandoned agricultural lands, the
representation of croplands by a number of crop functional types and the
consideration of C–N interaction in natural vegetation and crops, the
LPJ-GUESS model considers key processes and interactions that are crucial
when accurate estimates of C stocks and fluxes are derived based on detailed
dynamics of LUC (see, e.g.,
Hickler
et al., 2004; Lindeskog et al., 2013; Olin et al., 2015; Pugh et al., 2015;
Zaehle, 2013).
Uncertainties in ELUC resulting from the choice of LUC reconstruction
(expressed as 1 standard deviation around decadal means), as quantified
with the LPJ-GUESS model and the four global net LUC datasets, were ±0.19, ±0.66 and ±0.47 Pg C a-1 (14, 39 and
29 % of ELUC) in the 1980s, 1990s and 2000s respectively (see
Table 3 for exact periods). Among the four
datasets, before the 1960s, ELUC using LUH1 and RAMA was similar, with
the HYDE and LUH2 datasets differing somewhat from this as a result of
regional differences in croplands and pastures, although global totals
remained similar (e.g. HYDE showed less croplands in the northeastern US around
1900 than LUH1 or RAMA; LUH2 showed a similar cropland distribution to LUH1 but
big differences in pastures in the southern parts of Russia; and LUH2
showed less pastures in southern Argentina and Kazakhstan than RAMA). These
uncertainty calculations are consistent with estimates from previous studies
in which a subset of the three LUC hindcasts (LUH1, RAMA and HYDE, partially
as earlier versions) were applied, sometimes in combination with a
bookkeeping method (Table S1). Uncertainties were also estimated for the
1980s as ±0.30 Pg C a-1 from a synthesis experiment of
Houghton et al. (2012) and as
±0.20 Pg C a-1 for the 1990s as determined by Shevliakova et al. (2009) when using HYDE and RAMA cropland data (in earlier versions including wood harvest). For the 2000s, Jain et al. (2013) found an uncertainty of ±0.21 Pg C a-1 when quantifying
ELUC with HYDE and RAMA datasets with both a coupled C–N DGVM and the
bookkeeping approach from Houghton (2008). Our
uncertainty of ELUC due to the choice of LUC reconstruction was higher
in the 1990s, where ELUC derived using HYDE data was significantly
higher due to a strong increase in pastures (Fig. 2a; Table S1). The
uncertainty in cumulative ELUC accumulated to ±14 Pg C for
1750–2007. For vegetation and soil carbon, the uncertainty introduced due to
the choice of LUC reconstruction was ±11 and ±37 Pg C,
respectively, in 1998–2007, translating into a change of 3 % of the
average size of both vegetation and soil C stocks
(Table 3; see Fig. S3 for regional differences for
the entire simulation period). These uncertainties for vegetation C stocks are
higher compared to the ones found by Arora and Boer (2010), who used only two of the LUC models applied here, and about the same
size for C stocks in soils. This implies non-linear interactions between
the DGVM structure and LUC dataset. We would expect uncertainty in C stocks and
fluxes to increase, at least during the pre-1900 period, if LUC
reconstructions applying a non-linear development of per capita land use
were also considered, such as the KK10 dataset does for the period 8000 BC
to AD 1850 (Kaplan et al., 2010).
For Europe, the uncertainty in ELUC is also large with ±37,
±33 and ±20 Tg C a-1 for the 1980s, 1990s and 2000s,
values which are between 30 and 72 % of the average flux
(Table 4). Differences result mainly from a
disagreement in the amount of pastures between the LUC reconstructions,
where the comparison is impaired by different definitions of the pasture
class (see the “Methods” section). This highlights the problem of fundamentally different
structures and assumptions between LUC models and DGVMs, recently identified as a major uncertainty in model estimates of ELUC
(Pongratz et al., 2014). Although
forests, natural grasslands and pastures can show similar gross primary
productivity (GPP), they significantly vary in their C sequestration
potential in vegetation and soils also depending on their location within
Europe
(Ciais
et al., 2008; Schulze et al., 2010). For instance, larger pasture areas in
HILDA in southern Europe compared to LUH1 lead to an increase in vegetation
C under LUC, whereas a decrease was simulated with LUH1 (Fig. S4). The
productivity and carbon dynamics of croplands in LPJ-GUESS is mainly
governed by the crop selection, the bioclimatic conditions of the land where
the crop is planted and the degree of fertilisation and irrigation
(for
productivity of croplands under different degrees of fertilisation see also
Ciais et al., 2010; Schulze et al., 2010). For instance, in Poland, high
cropland fractions in LUH1 that were only marginally fertilised compared to
croplands in western Europe decreased NPP under LUC, but increased
vegetation and soil C (Fig. S4). Differences between the LUC reconstructions
and therefore uncertainties in C stocks and fluxes converge during the first
half of the 19th century (maps not shown). Vegetation C stocks derived
with HILDA are lower than estimates of Fuchs et al. (2015a; R. Fuchs, personal communication, 2016), using the same datasets (minor difference in net data; see the “Methods” section) and a
bookkeeping method, but are within the uncertainty spanned by using HILDA
and LUH1 net LUC datasets (Table S2).
It is important to consider that the relative uncertainties in LUC
transitions between datasets are not constant through time, although the
absolute uncertainties remain remarkably constant over the simulation
period. The relative deviation of pasture area for the three global datasets
was about ±29 % before the 1850s, decreasing to about ±9 % in the 2000s (Fig. S5c). Global cropland areas had a deviation of
about ±11 % in the 1700s, decreasing to below ±2 % in the
2000s (Fig. S5c). For Europe the agreement of the two LUC reconstructions is
high for croplands (average deviation ±2 % for 1900–2010) but
lower for pastures and natural areas (1900–2010 on average ±36 %
for pastures and ±21 % for natural areas), with the deviation
increasing until 2010 for pastures (Fig. S5d). The general agreement on the
fractional coverage of natural land, pastures and croplands is higher for
the periods after 1960, when FAO statistical data and later improved data
from satellites became available (see also
Houghton, 2010; Verburg et al.,
2011). Before this period, the extrapolation of historical LUC information
was very much dependent on the applied model algorithms in combination with
census data. LUC reconstructions also differ in the resolution of past LUC
changes, providing annual time steps (RAMA for the entire historical period; LUH1, HYDE, LUH2 after 2000) or originating from decadal aggregations (HILDA
for the entire historical period; LUH1, HYDE, LUH2 until 2000). Such
methodological discrepancies and artefacts from the LUC modelling
significantly overlay the observed trends in the LUC reconstructions
(compare Fig. 1c, d) and are included in simulated C stocks and fluxes (Figs. 2, 3). It is noted that LUH1, HYDE and LUH2 net LUC data have LUC
features in common and cannot be regarded as completely independent
datasets, as LUH data are generally built based on HYDE input data. Here,
two subsequent (HYDE as HYDE3.1.1 and LUH2 as based on HYDE3.2) and one
intermediate (LUH1 as based on an early version of HYDE3.2) developments of
LUH and HYDE were used (see the “Methods” section); however, no versions directly building up
on each other were considered (e.g. dataset 1 is a version of HYDE that
was used as direct basis for a version of LUH that was considered as dataset 2).
Uncertainties in carbon stocks and fluxes due to accounting for gross land
transitions
The consideration of detailed gross land conversions in our simulations
increased the effects of LUC on carbon storage and fluxes due to larger
areas transitioning between land-use types (Fig. 1c, d). The increase of
about 16 % in average net annual ELUC and 15 % in cumulative
ELUC (Table 3; change due to gross relative to
net, 1750–2014) compared well with previous estimates (Table S3). The effect
of shifting cultivation on cumulative ELUC was quantified by
Olofsson and Hickler (2008) by using the LPJ model
(Sitch et al., 2003), which has similarities in the way plant
and soil physiological processes are calculated to the model used here but
has a simpler representation of vegetation dynamics and croplands and no C–N
dynamics. Olofsson and Hickler (2008) found an increase in cumulative ELUC by 28 and
29 % for 1700–1990 and 1850–1990, whilst Stocker et al. (2014), using a model with coupled C–N dynamics at a 1∘ × 1∘
spatial resolution, reported an increase by 15 % (despite
the coarser spatial resolution acting to accentuate the gross–net
differences). Using a bookkeeping model, Hansis et al. (2015) found an
enhancement of ELUC by 22–24 % over the period 1500–2012 at a
0.5∘ × 0.5∘ spatial resolution. The combined effect of
shifting cultivation and wood harvest on cumulative ELUC was summarised
by Houghton et al. (2012) as an increase by
25–35 %, and Wilkenskjeld et al. (2014) found an
increase in cumulative ELUC of 61 % (51 % without the effect of
wood harvest). Shevliakova et al. (2013) provide an estimate of ELUC under gross transitions including
wood harvest for the period 1860–2005 using a combination of modelled C
fluxes and a bookkeeping method to derive ELUC that is fairly close to
the value calculated in this study. For total land C stocks,
Shevliakova et al. (2009) reported an increase in
ELUC of 49 % due to shifting cultivation and wood harvest and
concluded from this that the effect on land carbon losses was comparable in
magnitude to the effect of cropland and pasture expansion. In our study we
quantified an increase in ELUC of 39 % total C globally due to
shifting cultivation alone (Table S3). Of these studies, only the model of
Stocker et al. (2014) (Table S3), accounts for C–N
interactions. C–N interactions have previously been found to enhance LUC
emissions (Jain et al., 2013). In addition, all these
studies differ from our study in the DGVM used, the LUC datasets and climate
model data applied, and the process representations in the models. All
studies except Olofsson and Hickler (2008) applied spatial
resolutions coarser than the 0.5∘ applied here (1∘ or
∼ 2∘; see Table S3).
Annual area transitions for major land change processes for global (a, data from LUH1) and European (b, EU27 + CH, data from HILDA) gross
(solid lines) and net (dashed lines) datasets. The class “other”
in (b) includes transitions between natural vegetation and barren land represented
in the HILDA dataset.
In our global experiment, the only contribution of gross land changes came
from shifting cultivation in certain tropical areas. These gross changes
were implemented in the LUH1 dataset based on assumed spatial extension and
transition rates and were reflected in significantly increased rates of
deforestation and reforestation in gross simulations (Fig. 4a). As would be
expected, removing forest material from the system through harvest or the burning of cleared vegetation, instead of it entering the soil pool through
litter decomposition, reduces soil C content. However, the soil C losses are
much less marked than those in vegetation, perhaps reflecting a dominance of
vegetation carbon turnover by leaves and fine roots in LPJ-GUESS, which are
inputs of C to the litter pool which are less affected by the harvesting of
vegetation on multi-annual rotation periods than woody inputs.
In comparison to the global gross land transitions, which accounted only for
shifting cultivation with uniform assumptions regarding cultivation cycles,
the European gross dataset accounted for irregular land-use and land-cover
changes based on regionally available empirical evidence (national
statistics and maps). ELUC when accounting for gross land changes was
about 53 % higher in the first simulation decade and converged to minor
differences from about 1980 onwards (Fig. 3a). Since LUC in Europe developed from
being a source of C in the beginning of the 1900s to being a sink after
about 1960, the small difference in C stocks and fluxes as a result of
accounting for gross transitions, in addition to net land changes, delayed
the year in which cumulated ELUC switched from a source to a sink.
Thus, the sink capacity of cumulative ELUC in the 2000s was reduced, as were
increases in vegetation and soil C stocks (Fig. 3b, e, f). To our knowledge,
no previous studies are available in which ELUC for Europe was derived
under net and gross LUC. The effect of accounting for gross land transitions
on vegetation C was negative in our simulations, with vegetation C under
gross LUC about 2 % lower than under net LUC, because the increased
number of regrowing stands under higher land transitions lowered mean
forest canopy closure, which also lowered NPP and, ultimately, soil C (see
Table 4). In contrast Fuchs et al. (2015a and unpublished results), using
a bookkeeping method, derived about 1 % higher vegetation C stocks under
gross LUC for the same area (Table S2). Discrepancies result from major
methodological differences between the bookkeeping and process-based
approach and also from Fuchs et al. (2015a) not accounting for C stocks in
croplands and pastures.
Gross LUC transitions in Europe over the entire simulation period were
dominated by conversions between pastures and croplands, i.e. cropland
expansion into pasture areas and abandonment of croplands and their
conversion into pastures (Fig. 4b), that were direct adjustments to market
demands and changes in land-use-related policies. Apart from this, LUC in
Europe was characterised by the abandonment of agricultural land and
reforestation peaking in the 1970s. Reforestation of European grasslands was
reported to entail a reduction in soil C stocks and an increase in
vegetation C
(Schulze
et al., 2010; Fig. 3e, f) and therefore a positive ELUC (Fig. 3a).
After a first period of regrowth, the additional tree biomass and increased
litter inputs to soils balanced soil C losses, so that vegetation and soil
stocks increased in the second half of the 20th century because wood
harvest was lower than growth (e.g.
Ciais et al., 2008), thereby contributing to the LUC sink capacity. Because
land abandonment and reforestation are one-directional LUC changes which are
represented in the same way in net and gross HILDA data
(see Fuchs et al., 2015b),
this did not lead to major differences between net and gross simulations. It
should be noted that with its current four LUC classes, the LPJ-GUESS model
was not able to make full use of the HILDA LUC dataset, as not all HILDA
land-cover classes were represented, e.g. urbanisation (urban areas were
assigned to the barren LUC class; see the “Methods” section), causing ∼ 18 % of gross land-use changes between 1900 and 2010
(Fuchs et al., 2015b).
Uncertainties in the modelling approach
The effects and uncertainties discussed above must be considered in relation
to other uncertainties arising in the modelling process (e.g. different model
implementations in respect of the representation of LUC and changes therein,
treatment of environmental change). A meta-analysis by
Houghton et al. (2012) estimated the
uncertainty in ELUC arising from the applied modelling approach and
method to be in the range of ±0.2 Pg C a-1 and that due to
data-related uncertainty and incomplete process understanding to be in the
range of ±0.5 Pg C a-1. The complex linkages between the
contributing factors, however, made it difficult to attribute uncertainties
to their sources (see also Jain et al., 2013).
A consideration of C–N interaction in vegetation and soils, as was done in
this study, is important when studying the effects of environmental drivers
such as LUC on carbon emissions; however, with the exception of a very few
models (e.g.
Jain
et al., 2013; Smith et al., 2014; Xu-Ri and Prentice, 2008; Zaehle and
Friend, 2010), most models do not represent N cycling. We did not quantify
the effect of C–N interactions in this study, but we note that our estimates
of cumulative ELUC from 1850 to 2005 with net HYDE (Table S3) are about
2 % lower than those of Pugh et al. (2015) using a
version of LPJ-GUESS without C–N interactions and the same LUC data. This is
in opposition to the findings of Jain et al. (2013), who
found about 40 % higher ELUC globally when accounting for N dynamics
and N limitation.
The implementation of gross transitions in a DGVM framework is subject to
considerable uncertainty. By varying the minimum age upon which a regrowing
natural stand becomes eligible for clearance again between 5 and 30 years
(recalling that natural stands are removed in order of age when natural
vegetation is reduced; see the “Methods” section), cumulative ELUC was found to
differ by ±20 Pg C (10 %) (1900–2014, LUH1 only; results not
shown). Hence, choosing too young a clearance age would lead to considerable
underestimates of ELUC which highlights an important, and hitherto
unremarked upon, implementation uncertainty for including gross transitions
in DGVM simulations. In the same way, other land-use and land-cover-change-related processes (fate of harvested wood, residue management,
occurrence of disturbances, including fire, etc.) might differ under
repeated land transitions such as shifting cultivation, and the realistic
representation of these interactions in process-based models remains a
subject for further research. Wood harvest mainly in extratropical regions
was assessed to account for an increase in ELUC of 0.2–0.3 Pg C a-1 (Houghton et al., 2012). We did
not account for wood harvest in this study as the uncertainty in the actual
spatial pattern of wood harvest in combination with the ways wood harvest is
done in practice over the globe (clear cut, selective harvesting of specific
age classes or a mixture of both) introduces many possibilities as to how
this process can be implemented in DGVMs. In a model such as LPJ-GUESS,
where forest ecosystem and wood parameters vary significantly over tree age
classes, this would result in a wide span of possible solutions depending on
the parameters used for the implementation of wood harvest that would be better
addressed in a dedicated sensitivity study investigating a variety of
possible implementations rather than with a single representation.
Simulation results of biogeochemical cycles with a DGVM such as LPJ-GUESS
depend critically on the year when simulations are started. In this study we
tested the effect of starting LUH1 simulations in either 1700 or 1900 (Fig. S6). ELUC cumulated over 1950–2014 was 17 % lower and soil C was
about 33 Pg C (2 %) lower when simulations were started in 1900, while
vegetation C stocks were similar (net LUC, Table S4). This emphasises the
impact on simulation results of soil legacy emissions resulting from
previous LUC changes (see, e.g., Gasser and Ciais,
2013, and Sentman et al., 2011, for legacy fluxes, and
Pugh et al., 2015, for breakdown of LUC emissions). Hansis et al. (2015) in contrast found a 28 % higher ELUC over the period
2008–2012 when legacy emissions predating 2008 were excluded. This resulted
from previous LUC transitions generating more uptake from regrowth than loss
from soil decomposition during the 2008–2012 period, and it further exemplifies
the considerable importance of legacy effects for the calculation of
instantaneous emissions. They also emphasise that the change in soil and
vegetation stocks induced by previous LUC can substantially modify ELUC
calculations compared to an assumption of equilibrium in the initial stocks.
Often such changes in initial conditions tend to lower vegetation C stocks
and thus subsequent deforestation emissions. Our 17 % lower ELUC
excluding legacy emissions accounts for both these effects. To exclude this
effect from the analysis of uncertainties due to LUC dataset selection and
the effect of accounting for gross LUC transitions, all global simulations
were started in 1700. For Europe, where HILDA reconstructions were not
available before 1900, simulations were started in 1900 for both HILDA and
LUH1 to ensure comparability.
Conclusions
Global and European carbon stocks and fluxes and the effects of changes in
LUC were shown to be subject to significant uncertainties resulting from the
choice of historical LUC reconstruction. In our global simulations, data from HYDE and RAMA and from LUH1 and LUH2 often lead to similar results in ecosystem C stocks
and fluxes. Therefore, LUH1 and LUH2 as the more recent developments of the four considered reconstructions (both based on HYDE version 3.2; however, LUH1 is based on an intermediate version and LUH2 on the final version) differ more
from older developments than those do from each other (see the “Methods” section for model
versions used in this study). For Europe, variables predicted based on the
newly available HILDA dataset were similar to those resulting from using
LUH1 for Europe; however, LUH1 predicted larger changes under LUC.
Differences in the effects of both global and European LUC on C stocks and
fluxes were found to be mainly based on the total area and spatial
distribution of pastures in the datasets; however, note that the area of
pastures is impaired by different classifications used by the LUC models. To
account for the uncertainty arising from different reconstructions of
historical LUC in the dynamic modelling of C stocks and fluxes and to provide
realistic estimates of this uncertainty for the land-use C flux, the
consideration of multiple LUC reconstructions exploring the full range of
reasonable assumptions is needed, as well as efforts to narrow the
uncertainty in constructions of historical land use. Multiple LUC
reconstructions were calculated by Hurtt et al. (2011),
but the consequences of uncertainty in land-use transitions are not
routinely explored by the carbon cycle community. This goes along with the
reduction in uncertainties in the implementation of these datasets and
different forms of LUC in DGVMs which has recently been the focus of
discussion
(Pongratz et
al., 2014; Pugh et al., 2015; Stocker and Joos, 2015).
The results herein show that considering gross land conversions
significantly increased the effect of LUC change on C stocks and fluxes.
Most noticeably, the land-use C flux was enhanced by about 15 % of carbon
released in addition to when only accounting for net land changes (cumulated
1750–2014), primarily resulting from a reduction in vegetation C storage.
Note that for DGVMs operating at a lower spatial resolution than the
0.5∘ × 0.5∘ used here, the underestimation of ELUC
would be even larger, as shown by Wilkenskjeld et al. (2014). Given the
large percentage enhancements in ELUC found by considering gross
transitions, this should be the preferred method whenever possible.
The implementation of gross land transitions, however, poses technical and
parameterisation challenges to the process-based models. It also relies on
extensive information on historical land-use transitions, which is largely
lacking; at present, only a few LUC models are able to represent gross land
changes on larger spatial scales, providing a limited basis to characterise
the uncertainty. The LUH datasets are the only global-scale reconstructions
representing gross land changes by explicitly implementing shifting
cultivation in certain tropical areas with assuming a fixed period of 15 years for which land is cultivated before abandonment (for LUH1). For
Europe, the HILDA dataset is the first reconstruction representing gross
land transitions which are based on actual LUC inventory data and
complementary model assumptions. The reconstruction of detailed regional
sub-grid land transitions, and possibly more realistic patterns of shifting
cultivation today, is restricted by the lack of reliable information on
continental and global-scale historical land transitions. New datasets based
on archived LUC data and remote-sensing sources are currently becoming
available with high spatial resolution for the global to continental scale
(Chen
et al., 2015; European Environment Agency, 2014; Wang et al., 2015) and on a national to regional level
(Homer
et al., 2015; MOFOR, 2016; RCMRD, 2016; Roy et al., 2015; TerraClass, 2016).
New promising efforts also provide LUC change data globally derived from
remote sensing with a 250 m spatial resolution
(Wang et al., 2015) and with a 30 m
spatial resolution (Chen et
al., 2015). In the coming years, new high-resolution LUC datasets can be
expected from the Landsat archives (http://earthexplorer.usgs.gov) and the
new Sentinel missions (Aschbacher and
Milagro-Pérez, 2012). These will contribute to closing the information gap
and in this way improve the calibration of LUC models to represent the
underlying processes, reduce the uncertainty in ecosystem functions such as
the present-day land-use flux, and provide enhanced information for, e.g.,
the assessment of ecosystem services and biodiversity indicators in the
future.