Spatial and temporal patterns of global burned area in response to anthropogenic and environmental factors: Reconstructing global fire history for the 20th and early 21st centuries

Authors


Abstract

Fire is a critical component of the Earth system, and substantially influences land surface, climate change, and ecosystem dynamics. To accurately predict the fire regimes in the 21st century, it is essential to understand the historical fire patterns and recognize the interaction among fire, human, and environment factors. Until now, few efforts are put on the studies regarding to the long-term fire reconstruction and the attribution analysis of anthropogenic and environmental factors to fire regimes at global scale. To fill this knowledge gap, we developed a 0.5° × 0.5° data set of global burned area from 1901 to 2007 by coupling Global Fire Emission Database version 3 with a process-based fire model and conducted factorial simulation experiments to evaluate the impacts of human, climate, and atmospheric components. The average global burned area is ~442 × 104 km2 yr−1 during 1901–2007 and our results suggest a notable declining rate of burned area globally (1.28 × 104 km2 yr−1). Burned area in tropics and extratropics exhibited a significant declining trend, with no significant trend detected at high latitudes. Factorial experiments indicated that human activities were the dominant factor in determining the declining trend of burned area in tropics and extratropics, and climate variation was the primary factor controlling the decadal variation of burned area at high latitudes. Elevated CO2 and nitrogen deposition enhanced burned area in tropics and southern extratropics but suppressed fire occurrence at high latitudes. Rising temperature and frequent droughts are becoming increasingly important and expected to increase wildfire activity in many regions of the world.

1 Introduction

Fire plays a critical role in shaping biosphere and atmospheric patterns. Fire regimes are largely regulated by climate [Morton et al., 2013] and human activities [Marlon et al., 2008]; meanwhile, burning of biomass can speed up climate change through altering atmospheric radiative characteristics and land surface albedo [Andreae, 1991; Langmann et al., 2009; Levine et al., 1995; Randerson et al., 2006; Y. Liu et al., 2013]. In the future, the global fire regimes may be quite different from the present pattern due to rapid climate change [Bowman et al., 2009], and anthropogenic effects on fire might become less important than climatic influences [Pechony and Shindell, 2010]. A better understanding of the interaction among fire, climate, and human activities is helpful to enhance our capability of predicting future fire pattern and provide scientific information for fire management policy in the 21st century.

To date, a variety of methods have been used to retrieve fire history at local and regional scales, including charcoal records of biomass burning [Marlon et al., 2008], tree fire scars [Baisan and Swetnam, 1990; Baker and Ehle, 2001; Kitzberger et al., 2007; Niklasson and Granström, 2000; Payette et al., 1989; Wallenius et al., 2004; Westerling and Swetnam, 2003], official fire records [Stocks et al., 2002], and satellite observations [Eidenshink et al., 2007]. At the global scale, gridded burned area has also been estimated through methods grouped into three categories: satellite observation, fire modeling, and hybrid approach. Since the early 1980s, satellite imagery has been widely used to retrieve global fire pattern [Bowman et al., 2009]. Numerous satellite-based global fire products have become available in the recent decade, including Global Burnt Area 2000 (GBA2000) [Grégoire et al., 2003], L3JRC [Tansey et al., 2008], Global Fire Emission Database version 3 (GFED3) [Giglio et al., 2010], and so on. GFED has the longest global fire observation, spanning 15 years from 1997 to 2012, and was verified by ground fire records in various regions [Giglio et al., 2013; Giglio et al., 2010]. However, studies of global fire patterns prior to the 1980s are lacking.

Recently, process-based fire modeling has provided an effective tool to estimate large-scale fire patterns. It simulates fire activities (e.g., ignition, spread, and extinguishment) while considering critical environmental factors such as climate, land cover, and land management practices [Pechony and Shindell, 2009]. Process-based fire models have been applied in reconstructing global fire history and investigating fire impacts on ecosystems. These models include MC-FIRE [Lenihan et al., 1998], Glob-FIRM [Thonicke et al., 2001], Reg-FIRM [Venevsky et al., 2002], SPITFIRE [Thonicke et al., 2010], Community Land Model (CLM)-Fire [Li et al., 2012], and Canadian Terrestrial Ecosystem Model (CTEM)-Fire [Arora and Boer, 2005]. Model simulated fire patterns are not often consistent with each other. For example, Kloster et al. [2010] estimated global fire patterns based on the modified CTEM-Fire model and found a declining trend in burned areas from the 1900s to the 1960s and an increasing trend from the 1970s to the 1990s. However, the estimation by Li et al. [2013] based on CLM-Fire model presented a declining trend from the 1870s to the 1990s.

Mouillot and Field [2005] developed a global burned area data set for the twentieth century through a hybrid approach by incorporating satellite information, official fire records, and some fire trend assumptions. Their estimated global burned area was much larger than satellite observations. For example, their estimated global burned area was 722 × 104 km2 in 1999, while the GFED3 observed 339 × 104 km2. Current studies regarding century-scale spatial-explicit global fire history have shortcomings that hinder their application to studying fire-climate-human interactions. First, there is considerable disagreement in estimated global burned area with satellite observations in terms of temporal variation and spatial distribution; second, the effects of anthropogenic and environmental factors on fire patterns are difficult to interpret; third, temporal resolutions are relatively low, which is insufficient to study the seasonality of fire activity and emissions. Therefore, an urgent need is to develop a long-term, high-resolution, gridded global burned area database consistent with satellite observations, with a consideration of both long-term climate change and human impacts.

Global fire regimes have experienced extensive anthropogenic and environmental stresses in the twentieth century due to the rapid climate and land use changes [Prentice, 2010]. Anthropogenic impacts have been recognized as the primary factor in controlling the global burned area [Flannigan et al., 2009; Marlon et al., 2008; Pechony and Shindell, 2010]. The impacts of climate variation have also been studied at local, regional, and global levels [e.g., Westerling et al., 2006]. A key message is “Climate variability and fire weather influence wildfire behavior and account for the variability in fire severity at various time scales” [Liu et al., 2010; Liu et al., 2012]. Most previous studies focused on either climate or anthropogenic factors; only a few quantitatively evaluated the impacts of both climate and human activities. Other environmental factors, such as CO2 and nitrogen deposition, also play important roles in influencing fire regimes, however, are often neglected. To understand the mechanism of fire regimes at global scale, it is essential to quantify the relative contribution of human and multiple environmental factors simultaneously.

The Dynamic Land Ecosystem Model (DLEM) [Tian et al., 2010] is a highly integrated ecosystem model that incorporates biogeochemical cycle and water cycle to estimate the interactions and feedbacks among multiple ecosystem components. In this study, we improved the DLEM by coupling a process-based fire model to estimate global burned area and the feedback of fire to the ecosystem. The objectives of this study are (1) to develop a global historical fire data set at 0.5° resolution at monthly intervals from 1901 to 2007 through integrating the GFED3 burned area information, (2) to present the global and regional fire patterns, and (3) to discuss the determinative factors controlling fire trend and quantify the relative contributions of anthropogenic and multiple environmental factors to global and regional burned areas.

2 Materials and Methods

2.1 The Fire Module of Dynamic Land Ecosystem Model (DLEM-Fire)

The DLEM has been well validated and applied to studying the fluxes and storage of water, carbon, nitrogen, and multiple greenhouse gases at regional and global levels [e.g., Chen et al., 2012; M. Liu et al., 2013; Lu and Tian, 2013; Ren et al., 2012; Tao et al., 2013; Tian et al., 2010; Tian et al., 2011; Tian et al., 2012; Xu et al., 2010; Zhang et al., 2010]. The DLEM-Fire is a process-based fire module built in DLEM. It is capable of estimating burned area, fire emissions, and fire impacts on ecosystem function and structure by simulating fuel characteristics, number of fires, and fire behavior (Figure 1). Since this study aims to reconstruct the historical burned area, we focus on the algorithm of estimating burned area, rather than fire emissions and fire impacts. For each grid, the daily burned fraction BF is calculated as

display math(1)

where BF′ is DLEM simulated burned fraction prior to coupling with satellite information, which is equal to the sum of DLEM simulated wildfire burned fraction (BFwd′ ), deforestation fire burned fraction (BFdef ′ ), and agricultural fire burned fraction (BFag ′ ); and v is the grid-based satellite-adjusted scalar to regulate the magnitude and spatial distribution of fires and is calculated as the quotient of satellite-observed fire climatology and DLEM simulated fire climatology prior to coupling with satellite information. The detail process to estimate v can be found in section 'Model Implementation and Experimental Design'.

Figure 1.

Conceptual diagram presenting inputs, outputs, and the major processes of DLEM-Fire.

The following briefly describes the wildfire, agricultural fire, and deforestation fire schemes adapted from previous fire models [Arora and Boer, 2005; Pechony and Shindell, 2009; Li et al., 2012, 2013].

[10] Daily burned fraction in each grid as a result of wildfire, BFwd ′  (day−1), is calculated as a function of the number of daily fire events, Nf (count d−1), average burned area of each fire event, Af (km2 count−1), and grid area, Agrid (km2):

display math(2)

The number of daily fire events, Nf (count d−1), is expressed as

display math(3)

where Ia and In are the anthropogenic and natural ignition (count km−2 d−1), respectively; fs is the fraction of suppressed fires; fcrop is the fraction of cropland; Pm is reduction factor on ignition probability caused by fuel moisture; and Pb is the reduction factor on ignition probability caused by the insufficient fuel loading.

At the global level, cloud-to-ground lightning is the major natural fire ignition source. The natural ignition, In (count km−2 d−1), is estimated through the frequency of lightning flashes, Il (flash km−2 d−1), and latitude, λ [Prentice and Mackerras, 1977]:

display math(4)

Anthropogenic ignition, Ia (count km−2 d−1), is parameterized as

display math(5)

where PD is population density (person km−2); α is the daily potential anthropogenic ignitions (1.3 × 10−4, count person−1 d−1); and k(PD) is a population density-related parameter for adjusting anthropogenic ignition potential (=6.8PD− 0.6) [Venevsky et al., 2002].

Fire suppression, fs, is estimated as a function of population density:

display math(6)

The reduction factor, Pm, is calculated as a function of the available soil water:

display math(7)

where θ is the volumetric soil water content in the top 20 cm; and θwp and θfc are the volumetric soil water content at wilting point and field capacity, respectively.

The reduction factor, Pb, is determined by the abundance of fuel loading and is calculated as a function of aboveground biomass [Arora and Boer, 2005]:

display math(8)

where bag is the aboveground biomass (g C m−2), including leaf, stem, woody debris, and litter; bup is the upper limit of aboveground biomass (1000 g C m−2) where fuel load is not a limiting factor; and blow is the lower limit of aboveground biomass (150 g C m−2), below which ignition probability is 0.

The post-fire area, i.e., fire spread from starting point, is postulated to be elliptical in shape. The equation to calculate burned area of single fire event is

display math(9)

where udown is fire spread rate in downwind direction (km h−1); τ is the fire duration (h); and LB and HB are the length-to-breadth ratio and head-to-back ratio of the post-fire area, respectively. Fire spread rate, udown (km h−1), is computed as

display math(10)

where umax is the maximum spread rate in optimum condition (0.6 km h−1); and fm, fvpd, and fw are the limiting scalars of fuel moisture, vapor pressure deficit (VPD), and wind speed to adjust spread rate, respectively. The fuel moisture scalar, fm, is estimated as

display math(11)

VPD scalar, fvpd, is calculated as [Pechony and Shindell, 2009]

display math(12)

and

display math(13)

where RH is relative humidity (%), T is air temperature (K), and Ts is water boiling point temperature (373.15 K).

Wind speed scalar, fw, is parameterized as [Li et al., 2012]

display math(14)

Fire duration, τ (h), is the time that each fire lasts, which is estimated as a function of the global average fire persistence τave (24 h) and the slope of landscape (γ):

display math(15)

where γmin (2°) and γmax (10°) are the lower and upper limits of the slope effect on fire duration, respectively. Terrain with steep slopes indicates more watercourse or rocks, which could break down fuel continuity and reduce fire persistence time [Pfeiffer and Kaplan, 2012].

Length-to-breadth ratio (LB) and head-to-back ratio (HB) are computed according to wind speed, w (m/s):

display math(16)
display math(17)

The burned area of agricultural fire, BFag′  (month−1), is calculated as

display math(18)

where fcrop is the fraction of cropland, fharv is the monthly harvest index (1 means harvest happens, and 0 stands for no harvest happens), Pb is the reduction factor of fuel loading which is the same as that used for wildfire, and fGDP is the socioeconomic factor indicated by gross domestic production (GDP), which is set by Li et al. [2013].

The burned area of deforestation fire, BFdef' (month−1), is calculated as

display math(19)

where fdef is the annual deforestation fraction of tropical forests, Pb is the same reduction factor of fuel loading as used for wildfires, and fdry is monthly dryness function indicating the influence of dry season, which is computed by the ratio of monthly precipitation to potential evapotranspiration (P/PET)

display math(20)

2.2 Data Sets

Gridded (0.5° × 0.5°), georeferenced data sets were compiled to drive DLEM-Fire, including satellite-observed fire climatology, climate (temperature, precipitation, relative humidity, wind speed, and lightning frequency), atmospheric CO2 concentration, population density, nitrogen deposition, land use/cover change and land management practices (cropping system, fertilization and irrigation), GDP, and topography data (elevation, slope, and aspect) for the entire globe (Table 1). Satellite-observed fire climatology was derived from the GFED3 burned area from 1997 to 2007, which was developed based on multiple satellite sensors (Moderate Resolution Imaging Spectroradiometer (MODIS), Tropical Rainfall Measuring Mission (TRMM), Visible and Infrared Scanner (VIRS), and Along Track Scanning Radiometer (ASTR)) [Giglio et al., 2010]. Another satellite burned area product, MCD45A1 [Roy et al., 2008], was used for validation. Daily climate variables (1901–2007) were generated based on Climatic Research Unit/National Centers for Environmental Prediction (CRU/NCEP) data set (http://nacp.ornl.gov/thredds/fileServer/reccapDriver/cru_ncep/analysis/readme.htm). Daily lightning frequency was collected from NASA Lightning Imaging Sensor/Optical Transient Detector (LIS/OTD) (ftp://ghrc.nsstc.nasa.gov/pub/lis/climatology/HRAC/), which is the climatology of daily lightning frequency for the period 1995–2011. Due to the lack of historical lightning data, the 17 year climatology was applied to each year across the study period. Population density (1901–2005) was extracted from the History Database of the Global Environment (HYDE3.1) [Goldewijk et al., 2010] and assumed unchanged since 2005 due to the lack of data in 2006 and 2007. The CO2 concentration was obtained from the Carbon Dioxide Information Analysis Center (CDIAC). Land use/cover change was developed based on Synergetic Land Cover Product (SYNMAP) [Jung et al., 2006] and land use conversion data [Hurtt et al., 2011]. Annual deforestation rate was computed through the expansion of cropland and pasture within the grid where forest exists. Cropping system and crop phenology were identified based on a global crop geographic distribution map [Leff et al., 2004] and MODIS Leaf Area Index (LAI) products. Gross domestic production (GDP) data set in 2000 is from van Vuuren et al. [2007]. Elevation, slope, and aspect were derived from Global 30 Arc-Second Elevation product (https://lta.cr.usgs.gov/GTOPO30).

Table 1. The Input Data Sets for DLEM-Firea
VariablesPeriodSources
  1. a

    The variables simulated by DLEM, such as soil moisture and fuel loading, are not included in this table.

Climate1901–2007CRU/NCEP
Population density1901–2005HYDE3.1
Lightning frequencyclimatology during 1995–2011NASA LIS/OTD
GDP2000van Vuuren et al. [2007]
CO21901–2007CDIAC
Land use/cover change1901–2007SYNMAP and Hurtt et al. [2011]
TopographyStaticGTOPO30
Satellite-observed fire climatologyclimatology during 1997–2007GFED3

2.3 Model Implementation and Experimental Design

DLEM-Fire implementation is composed of three stages: (1) equilibrium run, (2) spin-up, and (3) transient run. The equilibrium run aims to determine the initial condition on 1 January 1901. In the equilibrium run, the model is fed with detrended climate data from 1901 to 1930 and atmospheric composition and land cover/use pattern in 1901 to reach an equilibrium state for carbon, nitrogen, and water (i.e., the changes in annual carbon, nitrogen, and water fluxes and pools are less than 0.1 g C m−2, 0.1 g N m−2, and 0.1 mm among consecutive years). After equilibrium state was reached, the model was run another 100 years for spin-up. In the transient mode, DLEM-Fire was driven by the time series of all input data sets and had two-type simulations: Type I was run without integrating satellite information, in which the annual satellite-adjusted scalar (v) on each grid was estimated as the quotient of the average annual burned area of GFED3 and DLEM-Fire simulated average annual burned area from 1997 to 2007; type II was run after type I simulation by coupling the satellite-adjusted scalar (v) estimated through type I simulation, and all the experiments below were based on type II simulation.

Five experiments were designed to estimate the global burned area and quantify the relative contribution of human activities, climate change, and atmospheric components (CO2 and nitrogen deposition), respectively (Table 2). In the “baseline” scenario (sim0), all the forcing data sets were kept at the level in 1901 to estimate the inherent model fluctuation. The “all-Combined” scenario (sim1) considered the historical variation of all the driven factors. The “climate only” scenario (sim2) was driven by historical climate condition but constant human activities, CO2 concentration, and nitrogen deposition. The “human only” scenario (sim3) was forced by dynamic human activities but static climate, CO2 concentration, and nitrogen deposition. The “CO2 + Ndep” scenario (sim4) was forced by historical CO2 concentration and nitrogen deposition but static climate and human activities.

Table 2. Experimental Design for This Studya
SimulationScenarioClimateLCLUC and PopulationCO2 and Ndep
  1. a

    LCLUC, land cover and land use change; Ndep, nitrogen deposition; 1901–2007 indicates the data set in 1901–2007 was used to drive model; 1901 indicates the data set in 1901 was used to drive model through the whole simulation period; 1901–2005 (L), 1901–2007 (P) indicates LCLUC data ends at 2005 and population data ends at 2007.

sim0Baseline190119011901
sim1All-Combined1901–20071901–2005 (L), 1901–2007 (P)1901–2007
sim2Climate only1901–200719011901
sim3Human only19011901–2005 (L), 1901–2007 (P)1901
sim4CO2 + Ndep190119011901–2007

2.4 Model Evaluation

The cross-validation scheme was used to evaluate DLEM-Fire simulated global fire patterns according to two satellite-based global fire products (GFED3 and MCD45A1). The DLEM-Fire simulation and GFED3 have 11 years overlap from 1997 to 2007. In the evaluation, 10 years from 11 overlapped years are selected to construct the satellite-adjusted scalar (v). And then the scalar was applied to estimate burned area of the left year to compare with GFED3 observation. This procedure was rotated 11 times to make every overlapped year to be selected.

2.4.1 Evaluation of Global Fire Distribution

The global fire distribution from 2002 to 2007 simulated by DLEM-Fire was compared with GFED3 and MCD45A1 data (Figure 2). The DLEM-Fire simulated tropical burned area was ~2% higher than GFED3 and ~4% higher than MCD45A1; DLEM-Fire simulated burned area in northern high latitudes was ~24.9% lower than GFED3 and ~11.7% higher than MCD45A1; DLEM-Fire simulated burned area in northern extratropics was ~9.5% lower than GFED3 and ~24.4% lower than MCD45A1; DLEM-Fire simulated burned area in southern extratropics was ~17.2% higher than GFED3 and ~68.1% higher than MCD45A1 (Figure 2d). Spatial correlations of the 6 year average burned area between our study and the two satellite products were 0.96 and 0.91, respectively. The result indicated the spatial patterns of DLEM-Fire simulation agreed with GFED3 and MCD45A1. Li et al. [2013] tested three fire schemes coupled in the Community Land Model (CLM) and reported the spatial correlations between three simulated global burned areas and GFED3 were 0.23, 0.44, and 0.69, respectively. The higher spatial correlations in our study suggested that our method by incorporating remote sensing data could substantially improve the accuracy of simulated fire spatial pattern.

Figure 2.

Six-year average of global burned fraction from 2002 to 2007 estimated by (a) DLEM-Fire, (b) GFED3, and (c) MCD45A1, and (d) the comparison of global and regional burned area in N. High, northern high latitudes (>55°N); N. Extra, northern extratropics (55°N to 30°N); Tropics (30°N to 20°S); S. Extra, southern extratropics (>20°S)

2.4.2 Evaluation of Interannual Variations

Interannual variation from 1997 to 2007 of the global burned area from the DLEM-Fire simulation was compared with GFED3, MCD45A1, and Mouillot and Field [2005] (Figure 3). DLEM-Fire estimated global burned area as 369.4 × 104 km2 yr−1 compared to 374.9 × 104 km2 yr−1 estimated by GFED3. During 2002–2007, MCD45A1 estimated an area of 347.9 × 104 km2 yr−1, while we estimated 366.6 × 104 km2 yr−1. In terms of interannual variation, our estimation also presented a similar pattern as GFED3 and MCD45A1. From 1997 to 2000, the annual mean global burned area estimated by Mouillot and Field [2005] was 602 × 104 km2 yr−1. Their estimate was 61% higher than our estimation and 51.3% higher than GFED3. This comparison with satellite products demonstrated that our estimates for the global burned area and temporal patterns were close to satellite observations.

Figure 3.

Interannual variations of global burned area (annual burned area normalized by mean) estimated from multiple global burned area products.

2.4.3 Evaluation of Fire Seasonality

Fire seasonality simulated by DLEM-Fire from 2002 to 2007 was compared with GFED3 and MCD45A1 products along latitude (Figure 4). All data sets demonstrated evident seasonal fire variation. At the middle and high latitudes, 40°N to 70°N, GFED3 and MCD45A1 revealed fires were likely to happen from June to September. Our simulated fire season was similar to GFED3 and MCD45A1, which also spanned from June to September. The fire season in summer was probably caused by the relatively warm climate, which dried up fuel moisture and increased fire danger potential. From 15°N to 40°N, GFED3 and MCD45A1 displayed fire season lasted from February to June, and the peak fire month shifted gradually from summer to spring toward the equator, roughly at a rate of 3.6 days per latitude degree. Our simulation also showed spring fire season shifting from June to February as it moved toward the equator. Overall, all data sets indicated fires were scarce in summer, which could be attributed to summer precipitation dampening fuel and reducing fire risk. From 15°N to the equator, satellite products revealed the fire season lasted from November to January, while the fire season of DLEM-Fire simulation was from December to March. In Southern Hemisphere from equator to 10°S, high fire activities in June and July were detected by all the data sets. From 10°S to 30°S, GFED3 and DLEM-Fire simulation demonstrated a fire peak from September to November, while MCD45A1 indicated fires were more active in August and September. From 30°S to 50°S, the fires estimated by GFED3, MCD45A1, and DLEM-Fire simulation were concentrated in December, January, and February. In general, our estimated seasonal variation was similar to GFED3 and MCD45A1.

Figure 4.

The zonal mean of monthly burned area of 6 year average (2002–2007), as normalized by the monthly average burned area of that latitude. (a) DLEM-Fire simulation, (b) GFED3, and (c) MCD45A1.

3 Results and Discussions

3.1 Spatial Patterns

During 1901–2007, our simulated global average burned area was 442.1 × 104 km2 yr−1, of which approximately 70.2% and 10.1% occurred in Africa and Oceania, respectively (Figure 5). The regions with high burned fraction included Africa, Northern Australia, and South America, where fuel was sufficient and dry season was long enough; the regions with low burned fraction were distributed in cold areas (e.g., Tibet plateau), very humid areas (e.g., interior Amazon rainforest), and regions with low fuel loading (e.g., Sahara desert) (Figure 5a). In the west-east direction, maximum burned area occurred between 10°E and 40°E, which was mainly due to fire in Africa (Figure 5c). In the north-south direction, two high burned area zones were shown clearly: One was approximately located at 10°N, and the other one was within 5°S–20°S (Figure 5b).

Figure 5.

Spatial distribution of global fire. (a) Average burned fraction from 1901 to 2007, (b) zonal sum of burned area per 0.5° latitude, and (c) meridional sum of burned area per 0.5° longitude.

Table 3 lists global burned area estimated by satellite observations, process-based fire models, and hybrid methods. Satellite observations (GFED, MCD45A1, L3JRC, GBA2000, and GlobScar) clustered the burned area around 350 × 104 km2 yr−1 in the 2000s. Yet Randerson et al. [2012] argued that satellite imagery with coarse spatial resolution failed to detect small fires and global burned area could be more than 400 × 104 km2 yr−1. Fire models estimated a similar global burned area as satellite observations, since satellite observed fires were often used as benchmarks to calibrated fire models [Kelley et al., 2012]. Annual burned areas from 1997 to 2004 estimated by Kloster et al. [2010] and Li et al. [2012] were 300 × 104 km2 and 330 × 104 km2, respectively. Our estimated global burned area was 362.6 × 104 km2 yr−1 in the 2000s, which fell within the range of satellite observations.

Table 3. Annual Global Burned Area Estimated by Different Studies
SourceStudy PeriodAnnual Burned Area (104 km2 yr−1)Study Approach
GFED4 [Giglio et al., 2013]1997–2011348multiple satellite observations
GFED3 [Giglio et al., 2010]1997–2010363multiple satellite observations
GFED2 [van der Werf et al., 2006]1997–2004329multiple satellite observations
MCD45A1 [Roy et al., 2008]2002–2010338satellite observation
L3JRC [Tansey et al., 2008]2000–2007392satellite observation
GBA2000 [Tansey et al., 2004]2000350satellite observation
GlobScar [Simon et al., 2004]2000211satellite observation
[Randerson et al., 2012]2001–2010464satellite observation
[Kloster et al., 2010]1997–2004300process-based fire model
[Li et al., 2012]1997–2004330process-based fire model
[Schultz et al., 2008]1960–2000383hybrid of official statistics and process-based fire model
[Mouillot and Field, 2005]1900–2000503hybrid official statistics and satellite observation
This study1901–2007442hybrid of process-based fire model and satellite observation

3.2 Temporal Variations

For the period from 1901 to 2007, Mann-Kendall trend test indicated a significant decreasing trend in global burned area at the rate of 1.28 × 104 km2 yr−1, with a relatively small annual variation (CV = 0.1) (Table 4). The largest burned area was in 1912 (552.1 × 104 km2), and the lowest burned area appeared in 2006 (348.9 × 104 km2). Decadal burned area from the 1900s through the 2000s (2000s in this study refers to 2001–2007) varied between 362.6 × 104 km2 yr−1 and 492.3 × 104 km2 yr−1 (Figure 6). The burned area in the tropics and extratropics presented a significant declining trend, while no significant trend was detected at high latitudes.

Table 4. The Mean, CV, and Trend of Global and Regional Burned Areas During 1901–2007a
 Mean (104 km2 yr−1)CVSlope (104 km2 yr−1)Trend Significance
  1. a

    Trends were analyzed with significance of Mann-Kendall test (S for significant, p < 0.05; NS for not significant, p ≥ 0.05) and Sen's slope; CV, coefficient of variation; in the column of “Slope,” the values within the parentheses are relative slope (slope normalized by the average burned area).

Global442.10.1−1.28 (−0.29%)S
N. High4.70.43−0.01 (−0.21%)NS
N. Extra30.40.43−0.37 (−1.21%)S
Tropics374.70.09−0.84 (−0.22%)S
S. Extra32.30.14−0.05 (−0.15%)S
Figure 6.

Decadal variation of global burned area. Error bar refers to the standard deviation of annual burned area within that decade.

The estimation of our study suggesting global fires declined in the twentieth century was supported by Antarctic ice core record of atmospheric carbon monoxide of Wang et al. [2010], charcoal records of Marlon et al. [2008], and fire model simulation of Li et al. [2013]. Prentice [2010] argued present biomass burning is at a historical low because of human-induced fuel loads reduction and fuel beds fragmentation. However, there are some studies presenting a different trend. As revealed by the data set of Mouillot and Field [2005], 503 × 104 km2 of land area experienced fire annually in the twentieth century, with an upward trend after the 1950s. However, their data set should be used with caution, since they made some assumptions and interpolations to quantify the burned area based on the scattered and qualitative fire records, which may bring in large uncertainties. Besides that, their data set may not be able to reflect the variation of burned area as influenced by driving forces, as they disregarded the underlying mechanism of fire activities. van der Werf et al. [2013] analyzed biomass burning sources and historical methane record and argued that the historical high levels of biomass burning implied by ice core record of carbon monoxide are likely overestimated. Yet ambient methane concentration may not be a good indicator of biomass burning, because methane emission from fires only contributes 5.6% of global emission [Kirschke et al., 2013] and explains only 15% of global emission anomalies, as compared with 70% explained by wetland emissions [Bousquet et al., 2006]. Kloster et al. [2010] simulated global burned area and found a downward trend from the 1900s to the 1960s, followed by an upward trend from the 1970s to the 1990s. The simulation of Pechony and Shindell [2010] showed global fire activity increases from the 1900s to the 1940s and then decreases afterward. These simulations contradict our estimated trend, which may be attributed to the different parameterization of anthropogenic impact on fire.

3.3 Relative Contribution of Human Activities, Climate, and Atmospheric Components

Previous studies have reported the impacts of climate or human activities on fire regime, while few of them quantified the relative contribution of multiple factors. The impact of atmospheric components, such as CO2 and nitrogen deposition, was basically neglected. In this study, we conducted factorial experiments to quantify the relative contribution of human activities and multiple environmental factors from 1901 to 2007. The contribution of climate on burned area was estimated by the difference between sim2 and sim0; the contribution of human activity was estimated by the difference between sim3 and sim0; and the contribution of CO2 and nitrogen deposition was estimated by the difference between sim4 and sim0 (see experimental design in section 'Model Implementation and Experimental Design').

3.3.1 Human Activities

Anthropogenic impacts are critical factors in determining global fire patterns, since they may increase or decrease wildfire activity through grazing, clearing forests, altering ignition patterns, and suppressing fires [Bowman et al., 2011]. Marlon et al. [2008] found an abrupt fire reduction after the 1870s in the tropics and extratropics as a response to cropland and pasture expansion. The model simulation of Pechony and Shindell [2010] suggested global fire is an anthropogenic-driven phenomenon in the twentieth century. In our study, human impact was identified as the primary factor accounting for the declining trend in global fire activity, which reduced global burned area by 141.2 × 104 km2 yr−1 in the 2000s (sim3 − sim0) (Figure 7a). In the tropics, cropland area increased by 76.6%, and population increased by 310% from the 1900s to 2000s. Although it has been reported that deforestation rate and burned area in Amazonia increased substantially since the 1970s [Houghton et al., 2000], human activities remarkably reduced burned area in the tropics as a whole region (Figure 7d), which is mainly contributed by the fire reduction in the tropical Africa and tropical Asia. In northern extratropics, human impact continuously increased from 1900s to 1980s (Figure 7c). Until the 1980s, burned area was reduced by 31.7 × 104 km2 yr−1 compared to the beginning of the twentieth century, in which human activities contributed 23.6 × 104 km2 yr−1. Studies have attributed the reduction of large fires in United States during this period to extensive livestock grazing and fire suppression efforts [Belsky and Blumenthal, 1997; Savage and Swetnam, 1990]. Fire models also suggested fire suppression efforts have substantially reduced the extent of wildfire in United States [Lenihan et al., 2008]. In China, burned area showed a downward trend from the 1950s to the 1990s. From 1981, fire suppression was strengthened across the forested region and abruptly reduced burned area [Lü et al., 2006]. At high latitudes, our results indicate human impact was relatively small compared with climate effects (Figure 7b), although fire suppression was an important component in forest management [Flannigan et al., 2005]. Generally, in the twentieth century, human activities acted as the dominant factor in determining the downward trend in the tropics and extratropics, but its impact was relatively minor at high latitudes.

Figure 7.

Factorial contributions to the inter-decadal variation of burned area. (a) Globe, (b) northern high latitudes (>55°N), (c) northern extratropics (55°N to 30°N), (d) tropics (30°N to 20°S), and (e) southern extratropics (>20°S).

3.3.2 Climate Variation

At high latitudes, we found fire temporal pattern was closely related to climate variation (Figure 7b). The contribution of climate variation to the burned area (sim2 − sim0) dropped down from 2.3 × 104 km2 yr−1 to −0.5 × 104 km2 yr−1 during the 1930s–1960s and increased from −0.5 × 104 km2 yr−1 to 3.41 × 104 km2 yr−1 during the 1960s–1990s. Correspondingly, the burned area decreased by 2.8 × 104 km2 yr−1 from the 1930s to 1960s and increased by 3 × 104 km2 yr−1 from the 1960s to 1990s, which indicated decadal variation of burned area at high latitudes was controlled by climate variation. Our estimated temporal pattern of burned area was comparable with other studies. The Large Fire Database, which includes large fire information in Canada starting from 1959 [Stocks et al., 2002], indicated an upward trend in the amount of area burned in Canada from 1959 to the end of the twentieth century. Wagner [1988] described burned area in Canada as a downward trend from the 1940s to the 1960s, followed by an upward trend from the 1960s to the 1990s. Krezek-Hanes et al. [2011] reported Canada burned area increased from the 1960s to the 1990s and then decreased in the 2000s. Flannigan et al. [2005] suggested fire regimes in Canada were controlled by climate factors with temperature as the most important predictor to estimate burned area.

In the extratropics, climate influence was also important for fire activities. In the western United States, wildfire frequency has increased since the mid-1980s in response to the climate warming and extended fire season [Westerling et al., 2006]. As demonstrated by Figure 7c, our simulation captured the upward trend of burned area in northern extratropics from the 1980s to the 2000s. Within this period, climate dominated the fire trend, and its contribution increased from −7.1 × 104 km2 yr−1 to 7.8 × 104 km2 yr−1. In southern Africa, either human activities or climate change could act as the primary factor in shaping fire regimes [Archibald et al., 2009]. As shown in Figure 7e, both climate and human activities were critical in southern extratropics. Although human impact was identified as the dominant factor in the extratropics, its influence was largely counteracted by climate change. If the warming trend continues in the 21st century, climate impact may outweigh anthropogenic influence and become the primary factor in the extratropics. In general, climate variation was the primacy of fire variation at high latitudes and is becoming increasingly important in extratropics in the context of global warming. However, the impact of climate variation was much less than anthropogenic influences in tropical regions and at the global level in the twentieth century.

3.3.3 Atmospheric Components

The impacts of atmospheric components on fire activity are rarely studied, because they are often covered up by the impacts of climate and human activities and their effect is difficult to detect based on short-term fire observations. The impacts of atmospheric components on ecological and hydrological processes have been verified by observations and model simulations. Theoretically, CO2 and nitrogen deposition can stimulate fire activities by boosting ecosystem productivity, vegetation biomass, and then fuel loading [Norby et al., 2010; Thomas et al., 2009]; at the same time, CO2 can suppress fire occurrence by retaining more water in the soil [Nelson et al., 2004] through reducing transpiration [Ainsworth and Rogers, 2007]. Fire models are able to provide insight into the impact of atmospheric components on burned area (sim4 − sim0). As indicated by our simulations, the impact of atmospheric components reduced burned area at high latitudes (Figure 7b). In the tropics and southern extratropics, elevated CO2 and nitrogen deposition significantly stimulated fire activity (Figures 7d and 7e). In northern extratropics, the impact of elevated CO2 and nitrogen deposition was small, which is probably because the positive effect and negative effect counteracted each other. At the global level, the impact of CO2 and nitrogen deposition enhanced the burned area significantly (Figure 7a). Therefore, it is reasonable to expect that atmospheric components would have more significant effect on global fire regimes with the continuous increasing trend in CO2 and nitrogen deposition projected in the 21st century [Lamarque et al., 2005].

3.4 Uncertainties and Future Needs

The global burned area data set of this study agrees well with satellite fire products in terms of spatial and temporal patterns, but there should be some uncertainties when interpreting the results due to the input data and fire model parameterization. The constant lightning frequency applied to each year in the study period was one source of uncertainties. Currently, the dynamic lightning data set prior to 1995 is still unavailable. To improve the spatial pattern of burned area simulation, GFED3 observed was integrated into DLEM-Fire. Thus, the reconstructed database has inherited uncertainties from GFED3 [Giglio et al., 2010]. Fire activities are a complicated phenomenon depending on numerous parameters [Pechony and Shindell, 2009]. Although DLEM-Fire has addressed most critical factors influencing fire activity, some processes, such as the impact of fuel quality, have not been well represented in the current fire module. The parameterization of fire duration in the present study only considered average fire duration and natural fire breaks, which may lead to underestimate burned area caused by megafire events that last much longer than the average fire duration. A new algorithm estimating fire duration based on fuel characteristics, fire suppression effort, and weather conditions is expected to greatly improve model performance in estimating burned area under extreme climate conditions. On the deforested land in Amazonia, trees were reburned several times within 1 year to clear the land, and the burned area is at least twice the deforested area [Nepstad et al., 1999]. Combustion completeness and climate conditions are the two major factors determining the burning times. DLEM-Fire probably underestimated the burned area of deforestation fire. Moreover, the performance of DLEM-Fire needs further evaluation as one component within the framework of DLEM. The information collected from field observations related to fuel loading, fuel consumption, fire emissions, and fire damage to vegetation would be valuable for model development, validation, and application.

4 Conclusions

In this study, we reconstructed a 0.5° × 0.5° data set of the global burned area for the 20th and early 21st centuries based on a process-based fire model and made comparisons to satellite observations. The comparisons indicated our reconstructed global fire history was capable of capturing the spatial and temporal patterns of global fire activities. As indicated by this data set, the global burned area during 1901–2007 was 442.1 × 104 km2 yr−1 and showed a significant declining trend at the rate of 1.28 × 104 km2 yr−1. At the regional level, burned area in tropics and extratropics exhibited a significant declining trend, with no significant trend at high latitudes.

This study highlighted the evaluation of the relative contributions of human activities, climate, and atmospheric components to global and regional burned areas and was the first attempt to address the importance of elevated CO2 and nitrogen deposition to fire regimes. The factorial experiments identified human activities as the dominant factor in determining the declining trend of burned area in the tropics and extratropics, and climate variation as the primary factor in shaping the decadal variation of burned area at high latitudes. The impact of climate change in extratropics is becoming increasingly important and may induce more fires in the context of global warming. Elevated CO2 and nitrogen deposition enhanced burned area in tropics and southern extratropics but suppressed fire occurrence at high latitudes. The spatial and temporal information on global burned area derived from this study can be used for ecosystem, hydrological, and climate modeling as well as by policy makers for understanding and assessing complex interactions among fire, climate, and human in a changing global environment.

Acknowledgments

This research has been supported by NSF/USDA/DOE Decadal and Regional Climate Prediction using Earth System Models (AGS-1243220 and NIFC2013-35100-20516), USDA/USDI Joint Fire Science Program (JFSP 11172), NSF Dynamics of Coupled Natural and Human Systems (1210360), and NSF Computer and Network Systems (CNS-1059376).

Ancillary