Fire emissions are a critical component of carbon and nutrient cycles and strongly affect climate and air quality. Dynamic global vegetation models (DGVMs) with interactive fire modeling provide important estimates for long-term and large-scale changes in fire emissions. Here we present the first multi-model estimates of global gridded historical fire emissions for 1700–2012, including carbon and 33 species of trace gases and aerosols. The dataset is based on simulations of nine DGVMs with different state-of-the-art global fire models that participated in the Fire Modeling Intercomparison Project (FireMIP), using the same and standardized protocols and forcing data, and the most up-to-date fire emission factor table based on field and laboratory studies in various land cover types. We evaluate the simulations of present-day fire emissions by comparing them with satellite-based products. The evaluation results show that most DGVMs simulate present-day global fire emission totals within the range of satellite-based products. They can capture the high emissions over the tropical savannas and low emissions over the arid and sparsely vegetated regions, and the main features of seasonality. However, most models fail to simulate the interannual variability, partly due to a lack of modeling peat fires and tropical deforestation fires. Before the 1850s, all models show only a weak trend in global fire emissions, which is consistent with the multi-source merged historical reconstructions used as input data for CMIP6. On the other hand, the trends are quite different among DGVMs for the 20th century, with some models showing an increase and others a decrease in fire emissions, mainly as a result of the discrepancy in their simulated responses to human population density change and land use and land cover change (LULCC). Our study provides an important dataset for further development of regional and global multi-source merged historical reconstructions, analyses of the historical changes in fire emissions and their uncertainties, and quantification of the role of fire emissions in the Earth system. It also highlights the importance of accurately modeling the responses of fire emissions to LULCC and population density change in reducing uncertainties in historical reconstructions of fire emissions and providing more reliable future projections.
Fire is an intrinsic feature of terrestrial ecosystem ecology, occurring in
all major biomes of the world soon after the appearance of terrestrial
plants over 400 million years ago (Scott and Glasspool, 2006; Bowman et al.,
2009). Fire emissions affect the Earth system in several important ways.
First, chemical species emitted from fires are a key component of the global
and regional carbon budgets (Bond-Lamberty et al., 2007; Ciais et al., 2013;
Kondo et al., 2018), a major source of greenhouse gases (Tian et al., 2016),
and the largest contributor of primary carbonaceous aerosols globally
(Andreae and Rosenfeld, 2008; Jiang et al., 2016). Second, by changing the
atmospheric composition, fire emissions affect the global and regional
radiation balance and climate (Ward et al., 2012; Tosca et al., 2013; Jiang
et al., 2016; Grandey et al., 2016; McKendry et al., 2019; Hamilton et al.,
2018; Thornhill et al., 2018). Third, fire emissions change the terrestrial
nutrient and carbon cycles by altering the deposition of nutrients
(e.g., nitrogen, phosphorus), surface ozone concentration, and
meteorological conditions (Mahowald et al., 2008; Chen et al., 2010;
McKendry et al., 2019; Yue and Unger, 2018). In addition, they degrade the
air quality (Val Martin et al., 2015; Knorr et al., 2017), which poses a
significant risk to human health and has been estimated to result in at
least
Summary description of the dynamic global vegetation models (DGVMs) that participated in FireMIP.
Abbreviations: CLM4.5 and CLM5: Community Land Model version 4.5 and 5; CTEM: Canadian Terrestrial Ecosystem Model; JSBACH: Jena Scheme for Biosphere- Atmosphere Coupling in Hamburg; SPITFIRE: Spread and InTensity fire model; JULES: Joint UK Land Environment Simulator; INFERNO: Interactive Fire And Emission Algorithm For Natural Environments; GlobFIRM: fire module Global FIRe Model; SMIFIRE: SIMple FIRE model; BLAZE: Blaze-Induced Land-Atmosphere Flux Estimator; ORCHIDEE: Organizing Carbon Hydrology In Dynamic Ecosystems; PFT: plant functional type; P: prescribed; M: modeled.
To date, only emissions from individual fires or small-scale fire complexes
can be directly measured from field campaigns and laboratory experiments
(Andreae and Merlet, 2001; Yokelson et al., 2013; Stockwell et al., 2016;
Andreae, 2019). Regionally and globally, fire emissions are often estimated
based on satellite observations, fire proxy records, and numerical models,
even though some attempts have been made to bridge the gap between local
observations and regional estimations using combinations of aircraft- and
ground-based measurements from field campaigns (e.g., SAMBBA, ARCTAS),
satellite-based inventories, and chemical transport models (e.g., Fisher et
al., 2010; Reddington et al., 2019; Konovalov et al., 2018). Satellite-based
fire emission estimates are primarily derived from satellite observations of
burned area, active fire counts, and/or fire radiative power, and are
sometimes constrained by satellite observations of aerosol optical depth
(AOD), CO, or
Historical change in fire emissions has been inferred from a variety of
proxies, such as ice-core records of
Dynamic global vegetation models (DGVMs) that include fire modeling are indispensable for estimating fire carbon emissions at local to global scales for past, present, and future periods (Hantson et al., 2016). These models represent interactions among fire dynamics, biogeochemistry, biogeophysics, and vegetation dynamics at the land surface within a physically and chemically consistent modeling framework. DGVMs are often used as the terrestrial ecosystem component of Earth system models (ESMs) and have been widely applied in global change research (Levis et al., 2004; Li et al., 2013; Kloster and Lasslop, 2017). Fire emissions of trace gases and aerosols can be derived from the product of fire carbon emissions simulated by DGVMs and fire emission factors (Li et al., 2012; Knorr et al., 2016).
Modeling fire and fire emissions within DGVMs started in the early 2000s (Thonicke et al., 2001) and has rapidly progressed during the past decade (Hantson et al., 2016). The Fire Model Intercomparison Project (FireMIP) initiated in 2014 was the first international collaborative effort to better understand the behavior of global fire models (Hantson et al., 2016). A set of common fire modeling experiments driven by the same forcing data were performed (Rabin et al., 2017). Nine DGVMs with different state-of-the-art global fire models participated in FireMIP. All global fire models used in the upcoming 6th Coupled Model Intercomparison Project (CMIP6) and IPCC AR6 are included in FireMIP, except for the fire scheme in GFDL-ESM (Rabin et al., 2018; Ward et al., 2018), which is similar to that of CLM4.5 (Li et al., 2012) in FireMIP. Note that GlobFIRM (Thonicke et al., 2001) in FireMIP is the most commonly used fire scheme in CMIP5 (Kloster and Lasslop, 2017) and is still used by some models in CMIP6.
Earlier studies provided only one single time series of fire emissions for global grids or regions (Schultz et al., 2008; Mieville et al., 2010; Lamarque et al., 2010; Marlon et al., 2016; van Marle et al., 2017b; and references therein). This limits their utility for quantifying the uncertainties in global and regional reconstructions of fire emissions and the corresponding impacts on estimated historical changes in carbon cycle, climate, and air pollution. A small number of studies also investigated the drivers of fire carbon emission trends (Kloster et al., 2010; Yang et al., 2014; Li et al., 2018; Ward et al., 2018). However, these studies could not identify the uncertainty source in recent model-based reconstructions or help understand the inter-model discrepancy in projections of future fire emissions because only a single DGVM was used in each.
This study provides a new dataset of global gridded fire emissions, including carbon and 33 species of trace gases and aerosols, over the 1700–2012 time period, based on nine DGVMs with different state-of-the-art global fire models that participated in FireMIP. The dataset provides a basis for developing multi-source (e.g., satellite-based products, model simulations, and/or fire proxy records) merged fire emission reconstructions and methods. It also, for the first time, allows end users to select all or a subset of model-based reconstructions that best suit their regional or global research needs. Importantly, it enables the quantification of the uncertainty range of past fire emissions and their impacts. In addition, the model-based estimates of fire emissions are comprehensively evaluated through comparison with satellite-based products, including amounts, spatial distribution, seasonality, and interannual variability, thus providing information on the limitations of recent model-based reconstructions. We also analyze the simulated long-term changes and the drivers for each DGVM and inter-model differences.
Nine DGVMs with different fire modules participated in FireMIP: CLM4.5 with the CLM5 fire module, CTEM, JSBACH-SPITFIRE, JULES-INFERNO, LPJ-GUESS-GlobFIRM, LPJ-GUESS-SIMFIRE-BLAZE, LPJ-GUESS-SPITFIRE, MC2, and ORCHIDEE-SPITFIRE (Table 1; see Rabin et al., 2017, for a detailed description of each model). JSBACH, ORCHIDEE, and LPJ-GUESS used the variants of SPITFIRE (Thonicke et al., 2010) with updated representation of human ignition and suppression, fuel moisture, combustion completeness, and the relationship between spread rate and wind speed for JSBACH (Lasslop et al., 2014), combustion completeness for ORCHIDEE (Yue et al., 2014, 2015), and human ignition, post-fire mortality factors, and modifications for matching tree age/size structure for LPJ-GUESS (Lehsten et al., 2009; Rabin et al., 2017).
The global fire models in the nine DGVMs have diverse levels of complexity (Rabin et al., 2017). SIMFIRE is a statistical model based on present-day satellite-based fire products (Knorr et al., 2016). In CLM4.5, crop, peat, and tropical deforestation fires are empirically/statistically modeled (Li et al., 2013). The scheme for fires outside the tropical closed forests and croplands in CLM4.5 (Li et al., 2012; Li and Lawrence, 2017), fire modules in CTEM (Arora and Boer, 2005; Melton and Arora, 2016), GlobFIRM (Thonicke, 2001), and INFERNO (Mangeon et al., 2016) are process-based and of intermediate complexity. That is, area burned is determined by two processes: fire occurrence and fire spread, but with simple empirical/statistical equations for each process. Fire modules in MC2 (Bachelet et al., 2015; Sheehan et al., 2015) and SPITFIRE variants are more complex, which use the Rothermel equations (Rothermel, 1972) to model fire spread and consider the impact of fuel composition on fire behavior.
Summary description of global fire modules in FireMIP DGVMs.
FireMIP experiment design. Note that CTEM and MC2 start at 1861
and 1901 and spin-up using 1861 and 1901
How humans affect fires differs among these global fire models (Table 2), which influences their estimates of fire emissions. GlobFIRM does not consider any direct human effect on fires and the MC2 fire model only considers human suppression on fire. CLM4.5 models fires in croplands, human deforestation and degradation fires in tropical closed forests, and human ignition and suppression for both occurrence and spread of fires outside of tropical closed forests and croplands. Burned area in SIMFIRE and human influence on fire occurrence in other models are a nonlinear function of population density. CTEM and JSBACH-SPITFIRE also consider human suppression on fire duration. JULES-INFERNO treats croplands and crop fires as natural grasslands and grassland fires. All models, except for CLM4.5 and INFERNO, set burned area to zero in croplands. FireMIP models treat pasture fires as natural grassland fires by using the same parameter values if they have pasture plant functional types (PFTs) or lumping pastures with natural grasslands otherwise. Biomass harvest is considered in pastures in LPJ-GUESS-GlobFIRM and LPJ-GUESS-SIMFIRE-BLAZE, which decreases fuel availability for fires, and that JSBACH-SPITFIRE sets high fuel bulk density for pasture PFTs.
Only CLM4.5 simulates peat fires, although only emissions from burning of vegetation tissues and litter are included in outputs for FireMIP; i.e., burning of soil organic matter is not included (Table 2).
In the FireMIP models, fire carbon emissions are calculated as the product of burned area, fuel load, and combustion completeness. Combustion completeness is the fraction of live plant tissues and ground litter burned (0 %–100 %). It depends on PFT and plant tissue type in GlobFIRM and in the fire modules of CLM4.5 and CTEM, and is also a function of soil moisture in INFERNO. Combustion completeness depends on plant tissue type and surface fire intensity in SIMFIRE, fuel type and wetness in the SPITFIRE family models, and fuel type, load, and moisture in the MC2 fire module.
The nine DGVMs in FireMIP are driven with the same forcing data (Rabin et
al., 2017). The atmospheric forcing is from CRU-NCEP v5.3.2 with a spatial
resolution of 0.5
Six FireMIP models (CLM4.5, JSBACH-SPITFIRE, JULES-INFERNO,
LPJ-GUESS-SPITFIRE, LPJ-GUESS-SIMFIRE-BLAZE, and ORCHIDEE-SPITFIRE) also
provide outputs of five sensitivity simulations: constant climate, constant
atmospheric
Based on fire carbon emissions and vegetation characteristics from DGVMs and
fire emission factors, fire emissions of trace gas and aerosol species
Emission factors (g species (kg DM
The EFs used in this study (Table 3) are based on Andreae and Merlet (2001), with updates from field and laboratory studies over various land cover types published during 2001–2018 (Andreae, 2019). All FireMIP model simulations used the same EFs from Table 3.
DGVMs generally simulate vegetation as a mixture of PFTs in a given grid location to represent plant function at global scale, instead of land cover types. In Table 4, we associate the PFTs from each DGVM with the land cover types shown in Table 3. Grass, shrub, savannas, woodland, pasture, and tundra PFTs are classified as grassland/savannas. Tree PFTs and crop PFTs are classified as forests and croplands, respectively, similarly to Li et al. (2012), Mangeon et al. (2016), and Melton and Arora (2016). PFTs of evergreen and other broadleaf deciduous tree in CTEM, extra-tropical evergreen and deciduous tree in JSBACH, and broadleaf deciduous tree and needleleaf evergreen tree in JULES are divided into tropical, temperate, and boreal groups following Nemani and Running (1996).
Attribution of plant function types (PFTs) in FireMIP DGVMs to land cover types (LCTs) for emission factors described in Table 2.
Abbreviations: T: tree; S: shrub; W: woodland; F: forest; G: grass; P: pasture;
Sava: savanna; N: needleleaf; E: evergreen; B: broadleaf; D: deciduous; R:
rain-green; SI: shaded-intolerant; SG: summer-green; M: mixed; I: irrigated;
RF: rainfed; C/W: cool or warm; S/W: spring or winter, Tro: tropical; Tem:
temperate; Bor: boreal; Sub-Tro: subtropical; Ex-Tro: extratropical; A:
Arctic.
We provide two versions of fire emission products with different spatial
resolutions: the original spatial resolution for each FireMIP DGVM output
(Table 1) and a
Satellite-based products are commonly used as benchmarks to evaluate present-day fire emission simulations (Rabin et al., 2017, and references therein). In the present study, six satellite-based products are used (Table 5). Fire emissions in GFED4/GFED4s (small fires included in GFED4s) (van der Werf et al., 2017), GFAS1.2 (Kaiser et al., 2012) and FINN1.5 (Wiedinmyer et al., 2011) are based on emission factor (EF) and fire carbon emission (CE) (Eq. 1). CE is estimated from MODIS burned area and VIRS/ATSR active fire products in the GFED family, MODIS active fire detection in FINN1.5, and MODIS fire radiative power (FRP) in GFAS1. Fire emissions from FEER1 (Ichoku and Ellison, 2014) and QFEDv2.5 (Darmenov and da Silva, 2015) are derived using FRP and constrained with satellite AOD observations. Satellite-based present-day fire emissions for the same region can differ by a factor of 2–4 on an annual basis (van der Werf et al., 2010) and up to 12 on a monthly basis (Zhang et al., 2014). The discrepancy among satellite-based estimates of present-day fire emissions mainly comes from the satellite observations used, the methods applied for deriving fire emissions, and the emission factors.
We also compared the simulated historical changes with historical reconstructions merged from multiple sources used as forcing data for CMIPs. Fire emission estimates for CMIP5 and CMIP6 were merged from different sources (Table 5). For CMIP5 (Lamarque et al., 2010), the decadal fire emissions are available from 1850 to 2000, estimated using GFED2 fire emissions (van der Werf et al., 2006) for 1997 onwards, RETRO (Schultz et al., 2008) for 1960–1900, and GICC (Mieville et al., 2010) for 1900–1950, and kept constant at the 1900 level for 1850–1900. RETRO combined literature reviews with satellite-based fire products and the GlobFIRM fire model. GICC is based on a burned area reconstruction from literature review and sparse tree ring records (Mouillot and Field, 2005), satellite-based fire counts, land cover map, and representative biomass density and burning efficiency of each land cover type.
Summary description of satellite-based products and historical constructions merged from multiple sources.
Abbreviations: GFED4: Global Fire Emissions Dataset version 4; GFED4s: GFED4 with small fires; GFAS1.2: Global Fire Assimilation System version 1.2; FINN1.5: Fire Inventory from NCAR version 1.5; FRP: fire radiative power; FEER1: fire emissions from the Fire Energetics and Emissions Research version1; QFED2.5: Quick Fire Emissions Dataset version 2.5; AOD: aerosol optical depth; GFED2: GFED version 2; RETRO: REanalysis of the TROpospheric chemical composition; GICC: Global Inventory for Chemistry-Climate studies; GCDv3: Global Charcoal Database version 3.
For CMIP6, monthly fire emission estimates are available from 1750 to 2015 (van Marle et al., 2017b). The CMIP6 estimates are merged from GFED4s fire carbon emissions for 1997 onwards, charcoal records GCDv3 (Marlon et al., 2016) for North America and Europe, visibility records for equatorial Asia (Field et al., 2009) and the central Amazon (van Marle et al., 2017b), and the median of simulations of six FireMIP models (CLM4.5, JSBACH-SPITFIRE, JULES-INFERNO, LPJ-GUESS-SPITFIRE, LPJ-GUESS-SIMFIRE-BLAZE, and ORCHIDEE-SPITFIRE) for all other regions. Then, based on the merged fire carbon emissions, CMIP6 fire trace gas and aerosol emissions are derived using EF from Andreae and Merlet (2001) with updates to 2013 and Akagi et al. (2011) with updates for temperate forests to 2014, and a present-day land cover map.
The spatial pattern and temporal variability of different fire emission species are similar, with some slight differences resulting from the estimated fire carbon emissions from the land cover types that have different emission factors (Table 3). Therefore, we focus on several important species as examples to exhibit the performance of FireMIP models in the simulations of present-day fire emissions.
Global total of fire emissions from 2003 to 2008 for DGVMs in
FireMIP and benchmarks. Unit: Pg (Pg
As shown in Table 6, FireMIP models, except for MC2 and LPJ-GUESS-GlobFIRM,
estimate present-day fire carbon,
Spatial distribution of annual fire black carbon (BC) emissions (g BC m
FireMIP DGVMs, except for MC2, represent the general spatial distribution of fire emissions evident in satellite-based products, with high fire BC emissions over tropical savannas and low emissions over the arid and sparsely vegetated regions (Fig. 2). Among the nine models, CLM4.5, JULES-INFERNO, and LPJ-GUESS-SIMFIRE-BLAZE have higher global spatial pattern correlation with satellite-based products than the other models, indicating higher skill in their spatial-pattern simulations. It should also be noted that, on a regional scale, CTEM, JULES-INFERNO, LPJ-GUESS-SPITFIRE, and ORCHIDEE-SPITFIRE underestimate fire emissions over boreal forests in Asia and North America. LPJ-GUESS-GlobFIRM and LPJ-GUESS-SIMFIRE-BLAZE overestimate fire emissions over the Amazon and African rainforests. CLM4.5 and LPJ-GUESS-GlobFIRM overestimate fire emissions over eastern China. JSBACH-SPITFIRE underestimates fire emissions in most tropical forests. MC2 underestimates fire emissions over most regions, partly because it allows only one ignition per year per grid cell and thus underestimates the burned area.
We further analyze the spatial distribution of inter-model differences. As
shown in Fig. 3, the main disagreement among FireMIP models occurs in the
tropics, especially over the tropical savannas in Africa, South America, and
northern Australia. This is mainly driven by the MC2, CTEM, JSBACH-SPITFIRE,
and ORCHIDEE-SPITFIRE simulations (Fig. 2). Differences among the
satellite-based estimates have a similar spatial pattern, but higher than
the inter-model spread in savannas over southern Africa and lower in the
temperate arid and semi-arid regions and north of 60
The FireMIP models reproduce similar seasonality features of fire emissions to satellite-based products; that is, peak month is varied from the dry season in the tropics to the warm season in the extra-tropics (Fig. 4).
For the tropics in the Southern Hemisphere, fire PM
For the tropics in the Northern Hemisphere, most FireMIP models exhibit larger fire emissions in the northern winter, consistent with the satellite-based products.
In the northern extra-tropical regions, satellite-based products show two periods of high values: April–May resulting mainly from fires in croplands and grasslands and July mainly due to fires in the boreal evergreen forests. Most FireMIP models can reproduce the second one, except for LPJ-GUESS-SPITFIRE, which peaks in October. CLM4.5 is the only model that can capture both peak periods, partly because it is the only one to consider unique seasonality of crop fires.
Global fire PM
Inter-model standard deviation of 2003–2008 averaged fire BC
emissions
(g BC m
Seasonal cycle of fire PM
Temporal correlation of annual global fire PM
Temporal change in annual global fire PM
We use the coefficient of variation (CV, the standard deviation divided by
the mean, %) to represent the amplitude of interannual variability of
fire emissions. As shown in Fig. 5, for 1997–2012, all FireMIP models
underestimate the variation as a result of (at least) partially missing the
1997–1998 fire emission peak. For 2003–2012 (the common period of all
satellite-based products and models), interannual variation of annual fire
PM
Figure 6 shows historical simulations of the FireMIP models as well as the
CMIP5 and CMIP6 reconstructions for fire carbon,
Long-term temporal change in fire emissions from DGVMs in FireMIP and CMIP forcing. A 21-year running mean is used.
Long-term trends in simulated global fire emissions for all models are weak
before the 1850s (relative trend < 0.015 % yr
Earlier reconstructions based on fire proxies also show a big difference in
long-term changes after the 1850s. The reconstruction based on the Global
Charcoal Database version 3 (GCDv3, Marlon et al., 2016) exhibits a decline
from the late 19th century to the 1920s and then an upward trend until
Spatial patterns of inter-model spread of fire emissions for 1700–1850 and 1900–2000 (Fig. S1b–c) are similar to the present-day patterns as shown in Fig. 3.
Six FireMIP models also conducted sensitivity experiments, which can be used
to isolate the role of individual forcing factors in long-term trends of
fire emissions during the 20th century. The medians of the six models are
also used for building CMIP6 fire emission estimates (van Marle et al.,
2017b). The 20th century changes in driving forces used in FireMIP are
characterized by an increase in the global land temperature, precipitation,
lightning frequency, atmospheric
As shown in Figs. 6 and 7, the downward trend of global fire emissions in
LPJ-GUESS-SIMFIRE-BLAZE is mainly caused by LULCC and increasing population
density. Upward trends in LPJ-GUESS-SPITFIRE and ORCHIDEE-SPITFIRE are
dominated by LULCC and rising population density and
As shown in Fig. 7, the inter-model spread in long-term trends mainly arises
from the simulated anthropogenic influence (LULCC and population density
change) on fire emissions, as the standard deviation in simulated responses
to LULCC (0.27 Pg C yr
LULCC decreases global fire emissions sharply in LPJ-GUESS-SIMFIRE-BLAZE
during the 20th century but increases global fire emissions for the other
models except for JSBACH-SPITFIRE. The response to LULCC in
LPJ-GUESS-SIMFIRE-BLAZE is because it assumes no fire in croplands and
accounts for biomass harvest (thus reducing fuel availability) in pastures
(Table 2), the area of which expanded over the 20th century. The
LULCC-induced increases in fire emissions for ORCHIDEE-SPITFIRE,
LPJ-GUESS-SPITFIRE, and JULES-INFERNO are partly caused by increased burned
area due to the expansion of grasslands (pastures are lumped in natural
grasslands in these models) where fuels are easier to burn than woody
vegetation in the model setups (Rabin et al., 2017). CLM4.5 models crop
fires and tropical deforestation and degradation fires. Crop fire emissions
in CLM4.5 are estimated to increase during the 20th century due to expansion
of croplands and increased fuel loads over time (Fig. S2). Emissions of
tropical deforestation and degradation fires in CLM4.5 are increased before
Change in global annual fire carbon emissions (Pg C yr
Rising population density throughout the 20th century decreases fire emissions in CLM4.5 and LPJ-GUESS-SIMFIRE-BLAZE because they include human suppression on both fire occurrence and fire spread. Fire suppression increases with rising population density and is simulated explicitly in CLM4.5 and implicitly in LPJ-GUESS-SIMFIRE-BLAZE. In contrast, rising population density increases fire emissions in LPJ-GUESS-SPITFIRE and ORCHIDEE-SPITFIRE because observed human suppression on fire spread found in Li et al. (2013), Hantson et al. (2015), and Andela et al. (2017) is not taken into account in the two models. The response to population density change for the other models is small, reflecting the compensating effects of human ignition and human suppression on fire occurrence (strongest in JULES-INFERNO in FireMIP models) and also human suppression on fire duration (JSBACH-SPITFIRE).
All models simulate increased fire emissions with increased atmospheric
We divided the global map into 14 regions following the definition of the GFED family (Fig. 8a). As shown in Fig. 8b, inter-model discrepancies in long-term changes are largest in Southern Hemisphere South America (SHSA), southern and northern Africa (NHAF and SHAF), and central Asia (CEAS).
Most FireMIP models reproduce the upward trends of fire CO emissions found
also in the CMIP5 or CMIP6 estimates since the 1950s in SHSA and till
Long-term changes in annual regional fire CO emissions (Tg CO yr
In other regions, the difference in long-term changes among models is
smaller (Fig. 8b). Emissions of most models and CMIP5 estimates exhibit a
significant decline in temperate North America (TENA) from
As shown in Figs. S3–S5, long-term changes in regional fire emissions for other species are similar to those of fire CO emissions.
The long-term changes in regional fire emissions and inter-model
disagreement are mainly caused by simulated responses to LULCC and/or
population density change for the 20th century (Figs. S6–S19). Besides,
climate change also plays an important role in North America, northern South
America, Europe, northern Africa, boreal and central Asia, and Australia.
FireMIP models generally simulate increased regional fire emissions with
increased
Our study provides the first multi-model reconstructions of global
historical fire emissions for 1700–2012, including carbon and 33 species of
trace gases and aerosols. Two versions of the fire emission product are
available, at the original spatial resolution for outputs of each FireMIP
model and on a unified
Our study provides an important dataset with wide-ranging applications for the Earth science research community. First, it is the first multi-model-based reconstruction of fire emissions and can serve as a basis for further development of multi-source merged products of global and regional fire emissions and of the merging methodology itself. van Marle et al. (2017b) presented an example of using part of the dataset to develop a multi-source merged fire emission product as a forcing dataset for CMIP6. In van Marle et al. (2017b), the median of fire carbon emissions from six FireMIP models was used to determine historical changes over most regions of the world. The merging method and merged product in van Marle et al. (2017b) are still preliminary and need to be improved in the future, e.g., by weighting the different models depending on their global or regional simulation skills. Secondly, our dataset includes global gridded reconstructions for 300 years. It can thus be used for analyzing global and regional historical changes in fire emissions on interannual to multi-decadal timescales and their interplay with climate variability and human activities. Third, the fire emission reconstructions based on multiple models provide, for the first time, a chance to quantify and understand the uncertainties in historical changes in fire emissions and their subsequent impacts on carbon cycle, radiative balance, air quality, and climate. Hamilton et al. (2018), for example, used fire emission simulations from two global fire models and the CMIP6 estimates to drive an aerosol model. This allowed for quantification of the impact of uncertainties in pre-industrial fire emissions on estimated pre-industrial aerosol concentrations and historical radiative forcing.
This study also provides significant information on the recent state of fire model performance by evaluating the present-day estimates based on FireMIP fire models (also those used in the upcoming CMIP6). Our results show that most FireMIP models can overall reproduce the amount, spatial pattern, and seasonality of fire emissions shown by satellite-based fire products. Yet they fail to simulate the interannual variability partly due to a lack of modeling peat and tropical deforestation fires. In addition, Teckentrup et al. (2019) found that climate was the main driver of interannual variability for the FireMIP models. A good representation of fire duration may be important to get the response of fire emissions to climate right. However, all FireMIP models limit the fire duration of individual fire events no more than 1 day in natural vegetation regions, so they cannot skillfully model the drought-induced large fires that last multiple days (Le Page et al., 2015; Ward et al., 2018). Recently, Andela et al. (2019) derived a dataset of fire duration from MODIS satellite observations, which provides a valuable dataset for developing parameterization of fire duration in global fire models.
This study also identifies population density and LULCC as the primary uncertainty sources in fire emission estimates. Therefore, accurately modeling the responses to these remains a top priority for reducing uncertainty in historical reconstructions and future projections of fire emissions, especially given that modeling is the only way for future projections. For the response to changes in population density, many FireMIP models have not included the observed relationship between population density and fire spread (Table 2). Moreover, Bistinas et al. (2014) and Parisien et al. (2016) reported obvious spatial heterogeneity of the population density–burned area relationship that is poorly represented in FireMIP models.
For the response to LULCC, improving the modeling of crop fires, pasture
fires, deforestation and degradation fires, and human indirect effect on
fires (e.g., fragmentation of the landscape) and reducing the uncertainty in
the interpretation of land use datasets in models are critical. Fire has
been widely used in agricultural management during the harvesting,
post-harvesting, or pre-planting periods (Korontzi et al., 2006; Magi et
al., 2012). Crop fire emissions are an important source of greenhouse gases
and air pollutants (Tian et al., 2016; Wu et al., 2017; Andreae, 2019).
GFED4s reported that fires in croplands can contribute 5 % of burned area
and 6 % of fire carbon emissions globally in the present day (Randerson et
al., 2012; van der Werf et al., 2017). In FireMIP, only CLM4.5 simulates
crop fires, whereas the other models assume no fire in croplands or treat
croplands as natural grasslands. In CLM4.5, crop fires contribute 5 % of
the global burned area in 2001–2010, similar to GFED4s estimates. However,
CLM4.5 estimates a total of 260 Tg C yr
Le Page et al. (2017) and Li et al. (2018) highlighted the importance of tropical deforestation and degradation fires in the long-term changes in reconstructed and projected global fire emissions, but in FireMIP only CLM4.5 estimates the tropical deforestation and degradation fires. For pasture fires, all FireMIP models assume that they behave like natural grassland fires, which needs to be verified by, for example, satellite-based products. If fires over pastures and natural grasslands are significantly different, adding the gridded coverage of pasture as a new input field in DGVMs without pasture PFTs and developing a parameterization of pasture fires will be necessary. Furthermore, Archibald (2016) and Andela et al. (2017) found that expansion of croplands and pastures decreased fuel continuity and thus reduced burned area and fire emissions. However, no FireMIP model parameterizes this indirect human effect on fires. In addition, DGVMs generalize the global vegetation using different sets of PFTs (Table 4) and represent land use data in a different way. This may lead to different responses of fire emissions to LULCC and thus different long-term changes in fire emissions among model simulations, given that many parameters and functions in global fire models are PFT-dependent. LUH2 used in LUMIP and ongoing CMIP6 provide information on forest/non-forest coverage changes (Lawrence et al., 2016), which can reduce the misinterpretation of the land use data in models and thus the inter-model spread of fire emission changes.
As discussed above, most FireMIP models do not consider the human suppression of fire spread, decreased fuel continuity from expanding croplands and pastures, human deforestation and degradation fires, and crop fires. Therefore, these models, and hence the CMIP6 estimates that are mainly based on them, may have some uncertainties in estimating historical fire emissions and long-term trends. This may further affect the estimates of the radiative forcing of fire emissions and the historical response of trace gas and aerosol concentrations, temperature, precipitation, and energy, water, and biogeochemical cycles to fire emissions based on Earth/climate system models that include these fire models or are driven by such fire emissions. It may also influence future projections of climate and Earth system responses to various population density and land use scenarios.
Data of FireMIP fire emissions are freely available from
The supplement related to this article is available online at:
FL contributed to the processing and analyses of the fire emission dataset. SS and AA designed the FireMIP experiments and LF, SH, GL, CY, DB, SM, MF, JM, and TH performed FireMIP simulations. MA compiled the EF table. JK, AD, CI, GvdW, and CW provided satellite-based and CMIP estimates of fire emissions. FL prepared the first draft of manuscript and revised it with contributions from MVM and other co-authors.
The authors declare that they have no conflict of interest.
We are grateful to Robert J. Yokelson, Zhong-Da Lin, Samuel Levis, Silvia Kloster, Margreet J. E. van Marle, Ben Bond-Lamberty, Jennifer R. Marlon, and Xu Yue for helpful discussions. We also thank two anonymous reviewers for their valuable comments and suggestions, and the editor Qiang Zhang for handling this paper.
This research has been supported by the National Key R&D Program of China (grant nos. 2017YFA0604302 and 2017YFA0604804), the National Natural Science Foundation of China (grant nos. 41475099 and 41875137), and the CAS Key Research Program of Frontier Sciences (grant no. QYZDY-SSW-DQC002). Maria Val Martin is supported by the US Joint Fire Science Program (13-1-01-4) and the UK Leverhulme Trust through a Leverhulme Research Centre Award (RC-2015-029). Almut Arneth acknowledges support from the Helmholtz Association, its ATMO program, and the Impulse and Networking fund, which funded initial FireMIP activities. Almut Arneth and Stijn Hantson also acknowledge EU FP7 project BACCHUS (603445). Gitta Lasslop is funded by the German Research Foundation (338130981). Brian I. Magi is supported by the NSF (BCS-1436496). Charles Ichoku is supported by NASA (NNH12ZDA001N-IDS).
This paper was edited by Qiang Zhang and reviewed by Douglas Hamilton and two anonymous referees.