This study used CLM version 4.5, which is the land model of the NCAR CESM version 1.2. The CESM is maintained by NCAR's Climate Global Dynamics Laboratory (CGD) and comprises different components such as land, atmosphere, ocean, land ice, and ocean ice (Worley at el., 2011; Kay et al., 2012). Each component utilizes various formulae to represent the complex interplay of physical, chemical, and biological processes, and each can be used either independently or coupled (Smith et al., 2010; Neale et al., 2012; Bonan et al., 2013). Land surface in the CLM is represented by sub-grid land cover (glacier, lake, wetland, urban, or vegetated), and vegetation coverage is represented by 17 plant functional types (PFTs) comprising 11 tree PFTs, 2 crop PFTs, 3 grass PFTs, and bare ground. For a detailed description of the model, please refer to Lawrence et al. (2011).

CLM can be run by including different levels of vegetation processes. In the satellite phenology (SP) option, vegetation coverage of different PFTs is prescribed using satellite-based land cover data (Lawrence and Chase, 2007), derived from a variety of satellite products including MODIS and Advanced Very High-Resolution Radiometer data. Land fractions are divided into bare ground, grass, shrub, and evergreen/deciduous trees. In addition, grass, shrub, and tree PFTs are classified as tropical, temperate, and boreal types, based on the physiology and climate rules of Nemani et al. (1996). Vegetation is further divided into C3 or C4 plants based on MODIS-derived leaf area index (LAI) values and the mapping methods of Still et al. (2003). Crop is also prescribed based on the merged dataset of the MODIS-derived land cover product and the global land cover in 2000 (GLC2000) (Ramankutty et al., 2008). Furthermore, the vegetation state (i.e., LAI) of different PFTs on land surface can be set based on the satellite-derived climatological data (Lawrence and Chase, 2007), which differ between months but not between years.

In addition to the SP option, CLM 4.5 can be extended using the biogeochemistry model (BGC) and dynamic vegetation model (DV); CLM simulations with BGC without DV (BGConly) and BGC with DV (BGC-DV) can be configured. BGConly simulates the carbon and nitrogen cycles in addition to biophysics and hydrology in a given distribution of vegetation PFTs (Paudel et al., 2016). In BGConly, phenological variations of LAI are simulated and whole-plant mortality is assumed as an annual mortality rate of 2 % without biogeographical changes in the vegetation distribution. In contrast, BGC-DV simulates biogeographical changes in the natural vegetation distribution and mortality as well as seasonal changes in LAI (Castillo et al., 2012; Castillo and Gurney, 2013). A PFT can occupy a region or degenerate by competing with other PFTs, or PFTs can coexist under various environmental factors, such as light, soil moisture, temperature, and fire (Zeng, 2010; Song and Zeng, 2014). Plant mortality in BGC-DV is determined by heat stress, fire, and growth efficiency (Rauscher et al., 2015). Note that BGC-DV does not simulate the crop PFTs, which is included in BGConly, because it simulates the changes in the natural vegetation only.

In the fire model (Li et al., 2012, 2013; Bonan et al., 2013), fire types are divided into four groups: non-peat fires outside cropland and tropical closed forests, agricultural fires, deforestation fires in tropical closed forests, and peat fires. Fire counts are determined based on natural and artificial ignition, fuel availability, fuel combustibility, and anthropogenic and unsuppressed natural fires related to socioeconomic conditions. The burned area is calculated by multiplying the fire count by the average fire spread, which is considered to be driven by wind speed, PFT, fuel wetness, and socioeconomic factors. In other words, the burning and spread of fire are related to the CLM input parameters of climate and weather conditions, vegetation conditions, socioeconomic conditions, and population density. After biomass and peat burning are calculated, trace gas and aerosol emissions as well as carbon emissions, which are the by-products of fires, are estimated.

Once the burned area is identified, impacts of the fire on vegetation mortality, peat burning, and the carbon cycle can be addressed. The amount of carbon emitted from the fire (E) is calculated as follows: E=A⋅C⋅CC, where A is the burned area; C is a vector of elements including carbon density of the leaf stem and the root and transfer and storage of carbon; CC is the corresponding combustion completeness factor vector.

Burned area also impacts the carbon and nitrogen pools of the vegetation, which are related to leaf, stem, and root; fire changes the vegetation state (e.g., LAI) and vegetation height during the burning period in both BGConly and BGC-DV runs. However, the number of individual PFTs does not change in BGConly but decreases by biomass burning in BGC-DV. In other words, individual plants are killed by fire only when the DV option is included in the model. The number of PFTs killed by fire (Pdistrub) is calculated using Eq. (2). Pdistrub=AbfAgPξ, where P is the population density for each PFT, ξ is the whole-plant mortality factor for each PFT, Ag is the grid cell area, Ab is the burned area of each PFT, and f is the fraction of coverage of each PFT. The whole-plant mortality, the rate at which plants die completely by fire, is a calibrated PFT-dependent parameter, which is 0.1 for broadleaf evergreen trees, 0.13 for needleleaf evergreen trees, 0.07 for deciduous trees, 0.15 for shrubs, and 0.2 for grass (Li et al., 2012).

The terrestrial carbon balance is affected when biomass is burned. The net ecosystem exchange (NEE) can be estimated using NEP (NEP=NPP–heterotrophic respiration (Rh)) and carbon loss due to biomass burning (Cfe). NEE=-NEP+Cfe.

A series of global numerical experiments were conducted in this study using a spatial resolution of 1.9∘ longitude ×2.5∘ latitude. Global climate data from the Climate Research Unit (CRU)-National Centers for Environmental Prediction (NCEP) reanalysis were used for atmospheric driving forcing of CLM. Data from 1901 to 2000 included 6 h precipitation, air temperature, wind speed, specific humidity, longwave radiation, and shortwave radiation. Figure 1 and Table 1 summarize the experimental process used in this study. The BGC run for the year of 1850 was initialized with the PFT distribution from the Land Use Harmonization (LUH) transient dataset for 1850 to 2005 (Hurtt et al., 2006) to simulate the year 1850 equilibrium state, used to initialize the 20th-century transient run. In the transient run, the amount of atmospheric carbon dioxide is increased since the onset of the Industrial Revolution in 1850 and the composition of land cover and vegetation is changed with the LUH dataset of Hurtt et al. (2006) (Vitousek et al., 1997; Pitman et al., 2004). The final surface conditions should represent those of the year 2000 after running the transient simulation using the CLM-BGC model.

Using the simulated surface conditions for the year 2000, four different 200-year equilibrium CLM simulations (BGConly and BGC-DV simulations with and without the fire model) were conducted (Table 1). For BGConly runs, a restart file from the transient run was used with and without the fire model (hereafter, BGConly-F and BGConly-NF, respectively). Similarly, the BGC-DV runs were performed using the same restart file to simulate the equilibrium vegetation in 200-year offline BGC-DV runs both with and without the fire model (hereafter, BGC-DV-F and BGC-DV-NF, respectively; Erfanian et al., 2016). In BGC-DV runs, the initial land surface state was bare ground with the vegetation previously in the system being entirely removed while soil conditions were adjusted with a restart file from the transient run (i.e., BGC run for the 20th century in Table 1) (Castillo et al., 2012; Rauscher et al., 2015; Qiu and Liu, 2016; Wang et al., 2016). Therefore, the vegetation state is quickly stabilized for 200 years of the BGC-DV runs since the runs restart from the spun-up soil carbon condition (i.e., after decomposition spin-up). Furthermore, the last 30-year results of the 200-year runs are analyzed to focus on the equilibrium states of both BGConly and BGC-DV runs. While the fire model is optional when using CLM with BGC, it is always run when using CLM with BGC-DV. Hence, the model was modified when conducting the BGC-DV-NF run, and the burned area was set to zero to neglect any fire incidences.

A comparison between the BGConly-F and BGConly-NF runs enables the isolation of the impact of fire on land surface, regardless of DV. In addition, the impact of fires and the interactive impacts of fires and vegetation distribution on the Earth system can be identified by comparing the BGC-DV-F and BGC-DV-NF runs. Note that this study focuses on the impact of fires and vegetation dynamics on land carbon and water fluxes by forcing the CLM with the CRU-NCEP climate data (1961–2000) without considering the land–atmosphere feedbacks. Simulations were run for 200 years from the initial surface conditions of the year 2000 to derive equilibrium land surface conditions. In addition, the average surface conditions of the last 30 years were compared with the simulation results.

In this section, we evaluate how the simulated burned areas differ between the runs with and without vegetation dynamics, i.e., BGC-DV-F and BGConly-F runs. On average, the BGC-DV-F and BGConly-F runs show burned areas of 320 and 487 Mha yr-1, respectively. These results are similar to those of previous studies that applied CLM (i.e., Li et al., 2012; Li and Lawrence, 2017). The fire model of Li et al. (2012) was originally developed by comparing the BGC-DV-F-type CLM simulations and resulted in 322 Mha yr-1 for 1997–2004. The BGC-DV-F simulation, under the equilibrium condition driven by the 1961–2000 CRU-NCEP data in this study, estimates a similar burned area (320 Mha yr-1) to that of Li et al. (2012). Li and Lawrence (2017) estimated the annual burned area as 489 Mha, which is similar to that of BGConly-F (487 Mha), using a BGC-F-type simulation coupled with Community Atmosphere Model (CAM).

In comparison to the burned area of BGConly-F, BGC-DV-F simulates a relatively small burned area (1) because agricultural fires are excluded in BGC-DV-F and only natural vegetation is simulated (Castillo et al., 2012) and (2) because fewer trees and thus less fuel, fed back from fire, are simulated in BGC-DV-F than in BGConly-F. Furthermore, the spatial distribution of burned areas in Fig. 2 shows that BGC-DV-F particularly underestimates the burned area in Africa and Oceania compared to BGConly-F. The differences in vegetation distribution between BGC-DV-F and BGConly-F in Fig. 3, where PFTs, excluding two crop PFTs, are simplified into six vegetation groups (broadleaf evergreen trees, needleleaf evergreen trees, deciduous trees, shrubs, grasses, and bare ground) (Rauscher et al., 2015), may impact the size of the burned area. In BGC-DV-F (Fig. 3a), evergreen and deciduous trees show limited growth whereas grass and bare ground are dominant in some regions such as southern Africa. Overall, BGC-DV-F simulates trees on 37.5 % of the global land area while BGConly-F, which is derived from observations (Fig. 3b), indicates that trees cover 41.46 % of the global land area (Table 2). More trees provide increased fuel for the occurrence and spread of fires in BGConly-F than in BGC-DV-F, consistent with the larger burned area in BGConly-F than in BGC-DV-F.

We also compare the model estimates to the satellite-based observational datasets of GFED (van der Werf et al., 2010, 2017; Giglio et al., 2013) (Fig. 3). Although the model simulations are not intended to reflect the reality but rather to understand the model mechanisms under the equilibrium states under the 1961–2000 climate forcing, it is still valuable to assess the model results using the observations. Different versions of GFED datasets provided different sized burned areas: GFED3 (van der Werf et al., 2010), GFED4 (Giglio et al., 2013), and GFED4 with small fires, i.e., GFED4s (van der Werf et al., 2017), suggest burned areas of 371 Mha yr-1 for 1997–2009, 348 Mha yr-1 for 1997–2011 and 513 Mha yr-1 for 1997–2016, respectively. In comparison to the most recent data, i.e., GFED4s, both BGConly-F and BGC-DV-F runs, especially BGC-DV-F, underestimate the burned area. Possible reasons for this underestimation in BGC-DV-F include the exclusion of agricultural fires and relatively small tree-dominated land coverage. The initial model development with a BGC-DV-F-type simulation (Li et al., 2012) was carried out in comparison to GFED3 (van der Werf et al., 2010), and BGC-DV-F estimated a burned area (320 Mha yr-1) similar to that of GFED3 (i.e., 371 Mha yr-1).

The impact of fires on vegetation distribution is assessed by comparing BGC-DV-F and BGC-DV-NF simulations (Table 2 and Figs. 4 and 5). Figure 4 shows the vegetation distribution of BGC-DV-NF (Fig. 4a) and BGC-DV-F minus BGC-DV-NF (Fig. 4b: Fig. 4a minus 3a). The plots clearly indicate large differences in vegetation cover in areas of high fire frequency (i.e., southern Africa, South America, western North America, India, and a portion of China) (Table 2), whereas areas with relatively low fire occurrence (i.e., the Arctic and desert regions) show small differences.

We estimated the fraction of burned areas, where fractions are grouped into four categories (>10 %, 10 %–1 %, 1 %–0.1 %, and <0.1 %) for each vegetation type and investigated the relationship between vegetation distribution and fire occurrence. Differences in the vegetation distribution between BGC-DV-F and BGC-DV-NF in Fig. 5 illustrate a nonlinear change in vegetation distribution in response to post-fire area. The changes are small in areas with minimal fire occurrence or where the burned area fraction is small (0.1 %–1 %). However, relatively large changes in vegetation distribution occur when the burned area fraction exceeds 1 %. Furthermore, there are large changes in the vegetation distribution in areas with burned area fractions above 10 %, including increases in bare ground, grass, and shrubs (31.19 %, 52.28 %, and 7.91 %, respectively) but decreases in deciduous, needleleaf evergreen, and broadleaf evergreen trees (8.85 %, 79.22 %, and 91.17 %, respectively).

In ecosystems, plants die in regions where fires occur and grass with rapid growth rates occupies those regions. Therefore, fire increases the ratio of bare ground and grassland but reduces the number of trees. However, there are no significant changes in the global fraction of shrubs and deciduous trees in the middle of the ecological succession process with respect to the presence or absence of fires (Table 2). When a fire occurs in a region where shrubs grow, the ratio of shrubland is diminished (e.g., in the middle of North America in Fig. 4b), but fire increases the ratio of shrubland in regions where trees grow (e.g., in the southwestern Asia in Fig. 4b). Similarly, the number of deciduous trees increases or decreases due to fires. Thus, the role of fires in areas of shrubland and deciduous trees varies with the region, and the actual vegetation distribution is a result of many factors including fire, climate, topography, and soil conditions (He et al., 2007; Cimalová and Lososová, 2009).

The direct and indirect impacts of fires on carbon balance were investigated for static and dynamic vegetation cover (Fig. 6 and Table 3). The impact of fires in BGConly was estimated by calculating the difference between BGConly-F and BGConly-NF, averaged over the final 30 years of each 200-year simulation. Similarly, the impact of fires in BGC-DV was estimated by calculating the difference between BGC-DV-F and BGC-DV-NF.

Carbon emissions from fires (direct impacts) are shown in Fig. 6. The spatial distribution of the BGConly and BGC-DV runs is similar, but average annual emissions are higher in BGConly (3.5 Pg) than in BGC-DV (3.0 Pg) because trees are less dominant in BGC-DV than in BGConly, which causes a reduced fuel load.

Carbon emission estimates from both BGConly and BGC-DV simulations are relatively high; however, they do fall within the range of previous findings. For example, 1997–2014 GFED4s data estimated annual direct carbon emissions as 2.3 Pg. Mouillot et al. (2006) estimated annual carbon emissions as 3.0 Pg for the end of the 20th century and the 20th-century average as 2.5 Pg. Li et al. (2012) estimated the 20th-century emissions as 3.5 Pg C yr-1 using the CLM3-DGVM and Li et al. (2014) and Yue et al. (2015) both estimated the 20th-century emissions as 1.9 Pg C yr-1 using the CLM4.5 and Organising Carbon and Hydrology In Dynamic Ecosystems (ORCHIDEE) land surface models, respectively.

In addition to direct carbon emissions from fires, fire influences terrestrial carbon sinks by impacting ecosystem processes (Fig. 6). Fire increases the NEP in post-fire regions in BGConly simulations (i.e., difference between BGConly-F and BGConly-NF, Fig. 6a), which is consistent with the findings of the previous studies (Li et al., 2014). The overall NEP increase is 2.5 Pg C yr-1 in this study, which is greater than the value of 1.9 Pg C yr-1 calculated by Li et al. (2014). However, Li et al. (2014) performed a transient simulation from 1850 to 2004, whereas the BGConly runs in our study were conducted following an equilibrium simulation using the year 2000 as the reference year, which means that no fire exchanges are caused by land cover changes.

Simulations that ignore vegetation dynamics (i.e., the BGConly runs in this study; Li et al., 2014; Yue et al., 2015) show a global fire-induced NEP increase when comparing fire-on and fire-off runs. However, a decrease in fire-induced NEP is apparent in some regions in BGC-DV simulations (i.e., differences between BGC-DV-F and BGC-DV-NF, Fig. 6b). This carbon sink reduction occurs in regions where dominant PFTs change from broadleaf and needleleaf evergreen trees to grass (Table 3 and Fig. 6). Table 4 shows the correlation coefficients between percent changes in vegetation types and changes in carbon fluxes (NEP, NPP, and Rh) for six different PFTs in each grid cell, and Fig. 7 shows the broadleaf evergreen tree, needleleaf evergreen tree, and grass PFTs. NEP changes are strongly linked to changes in dominant PFTs, for example, decreases in broadleaf evergreen and needleleaf evergreen trees and increases in grass. Furthermore, the changes in NEP and PFTs are related to the changes in NPP and Rh to some extent. Our results differ from those of previous studies that did not consider vegetation dynamics (e.g., Amiro et al., 2010) because the inclusion of vegetation dynamics enables the model to capture NEP decreases in post-fire regions at the beginning of the post-fire succession.

Since land use changes are not considered in this study, the overall impact of fires was estimated by the sum of direct carbon emissions from fires and terrestrial carbon sinks, i.e., NEP (Eq. 3). Both simulations resulted in net carbon sources in the post-fire regions, even though different processes were involved. Direct carbon emissions from fires (Cfe in Eq. 3) were partly negated by the increased NEP in the BGConly runs, but they were enhanced by the reduction in NEP in BGC-DV runs.

The impact of fires on water balance was examined by estimating the changes in runoff, evapotranspiration, and soil moisture between cases with and without fire. The differences between BGConly-F and BGConly-NF were assessed for the case without considering the vegetation dynamics, and differences between BGC-DV-F and BGC-DV-NF were assessed for the case considering the vegetation dynamics (Table 5 and Fig. 8). Increases in runoff and decreases in evapotranspiration (ET) were observed in post-fire regions to a different degree, which is consistent with the results of the previous studies (Neary et al., 2005; Li and Lawrence, 2017). Our study used CLM as a stand-alone model without coupling it with atmospheric or ice models, whereas Li and Lawrence (2017) examined the impact of fires on global water budget using CLM-BGC coupled with the CAM and Community Ice CodE (CICE) models and showed that the impact of fires on global annual precipitation was limited.

Li and Lawrence (2017) demonstrated that a reduction in vegetation canopy (LAI; Table 6) is a critical pathway for fires that decrease ET. Fire events lower the leaf area, which decreases vegetation transpiration and canopy evaporation; however, they also expose more of the soil to the air and sunlight, which increases soil evaporation. Post-fire decreases in vegetation height (Table 6) can increase and decrease ET because the resulting decrease in land surface roughness potentially reduces water and energy exchange and leads to higher leaf temperatures and wind speeds. In this study, both BGConly and BGC-DV runs show that the vegetation canopy is the main pathway leading to a decrease in ET, which is similar to the findings of Li and Lawrence (2017). In addition, an examination of the changes in the vegetation composition in post-fire regions shows that the overall impact of those changes in ET and runoff does not differ greatly when dynamic vegetation is employed in the model.

The results show that fire-induced vegetation changes (from trees to grass or bare ground) in BGC-DV lead to a significant decrease in canopy transpiration and increase in soil evaporation relative to BGConly runs. Fire destroys plant roots and leaves; changes in the dominant vegetation types in BGC-DV lead to changes in the soil moisture profile through reduced transpiration (Fig. 9 and Table 7). Consequently, there is less water stress in each soil layer in the burned areas than in unburned areas. Grasslands dominate the post-fire regions in BGC-DV runs, and they absorb and transpire more water from the top soil layer than trees (Mazzacavallo and Kulmatiski, 2015). Therefore, there is less moisture in the top soil layers in fire-affected regions than in unburned regions, although the overall transpiration is diminished. In summary, fire has an impact on vegetation distribution, which in turn impacts the soil water profile.

Despite the differences in soil moisture and vegetation canopy and height, changes in ET and runoff do not vary significantly between BGConly and BGC-DV. Thus, including dynamic vegetation does not impact the physiological and physical processes of evapotranspiration and runoff, respectively. However, changes in ET and runoff can be amplified in BGC-DV than in BGConly by modeling the land–atmosphere interactions with a coupled land–atmosphere model (e.g., CLM–CAM) because changes in land characteristics in BGC-DV would feed back to the changes in precipitation. Therefore, the limited impact of fires on precipitation in Li and Lawrence (2017) with the coupled model would be increased by including dynamic vegetation in the model.