Biomass burning is an important environmental process
with a strong influence on vegetation and on the atmospheric composition. It
competes with microbes and herbivores to convert biomass to
We present SEVER-FIRE v1.0 (Socio-Economic and natural Vegetation ExpeRimental global fire model version 1.0), which is incorporated into the SEVER DGVM. One of the major focuses of SEVER-FIRE is an implementation of pyrogenic behavior of humans (timing of their activities and their willingness and necessity to ignite or suppress fire), related to socioeconomic and demographic conditions in a geographical domain of the model application. Burned areas and emissions from the SEVER model are compared to the Global Fire Emission Database version 2 (GFED), derived from satellite observations, while number of fires is compared with regional historical fire statistics. We focus on both the model output accuracy and its assumptions regarding fire drivers and perform (1) an evaluation of the predicted spatial and temporal patterns, focusing on fire incidence, seasonality and interannual variability; (2) analysis to evaluate the assumptions concerning the etiology, or causation, of fire, including climatic and anthropogenic drivers, as well as the type and amount of vegetation.
SEVER reproduces the main features of climate-driven interannual fire variability at a regional scale, for example the large fires associated with the 1997–1998 El Niño event in Indonesia and Central and South America, which had critical ecological and atmospheric impacts. Spatial and seasonal patterns of fire incidence reveal some model inaccuracies, and we discuss the implications of the distribution of vegetation types inferred by the DGVM and of assumed proxies of human fire practices. We further suggest possible development directions to enable such models to better project future fire activity.
The biosphere is affected by fires through physical and chemical pathways, involving interactions between the terrestrial and atmospheric components of carbon, water and nutrient cycles. As a natural phenomenon, fires are an integral part of the majority of ecosystems, influencing soil fertility, stand regeneration, vegetation composition and succession (Le Page et al., 2015; Levine et al., 1999). However, through its anthropogenic use for land management (agriculture, pasture, deforestation), fire incidence is considerably higher than under natural conditions in many regions, including savannas in Africa and Australia or tropical forests in South America and Southeast Asia (Bond et al., 2005).
Abundant literature points to a variety of impacts, roles, and drivers of fires
and an extended range of spatial scales and timescales involved. It is estimated
that, on average, an area equivalent to that of India burns every year,
predominantly in savannas and grasslands (Tansey et al., 2004). Burned areas
in tropical and boreal forests are smaller, but their high productivity and
carbon storage capacity results in significant emissions of numerous
greenhouse gases (e.g.,
The strong integration of fires with the biosphere system is also emphasized by their dependence on a complex system of interactive drivers, designated as the fire triangle (Schoennagel et al., 2004), dominated by climate, vegetation and human activities. Precipitation rates and temperature partly control the amount of fuel available to burn, its moisture content and fire behavior in case of ignition (Crevoisier et al., 2007; Turner et al., 2008). Fire incidence, fire severity and ensuing emissions are also dependent on the vegetation types, structure and productivity of the ecosystem (Andreae and Merlet, 2001; Hammill and Bradstock, 2006). Finally, anthropogenic activities, as mentioned above, greatly bias the natural occurrence of fires, increased in many regions as a land management tool or decreased through fire suppression strategies (firefighting, preventive fires; Veblen et al., 2000). Other factors are involved (topography, natural landscape breaks, grazing), but most important is the interaction among those drivers, which needs to be considered to yield relevant information about fire risk (Dwyer et al., 2000a).
Dynamic global vegetation models (DGVMs) and Earth system models (ESMs)
simulate vegetation dynamics at a global scale; fire is included as an explicit
process in some of these models (Arora and Boer, 2005; Bachelet et al., 2001;
Li et al., 2013; Rabin et al., 2018; Thonicke et al., 2010, 2001; Venevsky et
al., 2002; Wu et al., 2017; Yue et al., 2014). Given the importance of fires
and their dependence on various model inputs or simulated processes, the
development of fire modules is of great interest to understand and evaluate
the fire-related coupling and feedback assumptions. Comprehensive review of
global fire modeling activity is given by Hantson et al. (2016) and an
overview of recent global fire models participating in the Fire Modeling
Intercomparison Project (FireMIP) is presented by Rabin et al. (2017).
Hantson et al. (2016) distinguish four levels of complexity for global fire
models incorporated into DGVMs (see Fig. 2 in their study) depending on
processes included in models.
The simplest statistical model relates burnt areas with climate and/or
vegetation (Glob-FIRM; Thonicke et al., 2001) and/or human activities (Knorr
et al., 2014). Models statistically estimate the number of fires and expected size of fires
(Pechony and Shindell, 2009). Process-based quasi-mechanistic models use functional relationships
among climate, vegetation and socioeconomic drivers of wildfires (MC-FIRE
(Lenihan and Bachelet, 2015), CTEM (Arora and Boer, 2005), CLM-Li (Li et al.,
2013), LM3-FINAL (Rabin et al., 2018), etc.). This approach was first
introduced by the Reg-FIRM model (Venevsky et al., 2002) and further
developed by the follow-up SPITFIRE (Thonicke et al., 2010) model and its
derivatives JSBACH–SPITFIRE (Lasslop et al., 2014), LPJ-LMfire (Pfeiffer et al., 2013),
LPJ-GUESS–SPITFIRE (Lehsten et al., 2009), ORCHIDEE–SPITFIRE (Yue et al.,
2014) and LPX-Mv1 (Kelley and Harrison, 2014). There is a complete representation of all processes in space and time (first-principle
approach model).
Nine from the 11 global models participating in FireMIP experiment are
process-oriented quasi-mechanistic models (Rabin et al., 2017); however,
mainly due to complexity of the processes involved, all these models are
still not at level 4. The closest to the complete representation of all fire
related processes in time is the SPITFIRE model (see Table 1 in Hantson et
al., 2016) and its modifications. The SPITFIRE modeling community achieved
significant results in global and regional fire modeling describing dynamics
of wildfires in a savanna–forest transition zone (Baudena et al., 2015),
contemporary dynamics of burnt areas in Europe (Wu et al., 2015), global fire
regimes in the preindustrial zone (Pfeiffer
et al., 2013) and changes in global carbon balance (Prentice et al., 2011).
While complete representation of all processes which determine wildfire dynamics in space and time is still underway, quasi-mechanistic models use different parameterizations of ignitions and spread of wildfire. Parameterizations are based on either long-term fire statistics or on remote-sensing data, which are a valuable data source due to their availability and global coverage. SPITFIRE, for example, uses lightning frequency as an input for calculation of the number of lightning fires. We argue that it would be advantageous if one can produce long-term fire relationships without depending on remote sensing, which is only available for a relatively short period of time (a few decades). Fire return intervals can be of the order of hundreds of years, whereas remote sensing is available for several decades. Therefore, using remote sensing to derive relationships implicitly assumes a space-for-time substitution, which may or may not hold. Conversely, parameterizations based on ground-based measurements or laboratory-based experiments have their own problems, like insufficient accuracy and low representativeness in space, but are considered to be more robust in time and thus very useful in DGVMs or ESMs for investigation of future global change impacts or past global fire regime reconstruction.
We present in this study SEVER-FIRE v1.0 (Socio-Economic and natural Vegetation ExpeRimental global fire model version 1.0; simplified as SEVER-FIRE in the following text) incorporated into the SEVER DGVM (Venevsky and Maksyutov, 2007; Wu et al., 2017), which is a modification of LPJ-DGVM (Sitch et al., 2003) for daily time step computation. SEVER-FIRE is a quasi-mechanistic model, which is a follow-up of Reg-FIRM for the globe, with several new features aiming for complete representation of wildfire processes. We improve earlier algorithms of Reg-FIRM and introduce new functionality with respect (1) to estimating the number of lightning fires from data on convective activity in the atmosphere, (2) to estimating the number of human fires from urban vs. rural population (timing of their appearance in natural landscapes and their ratio) and regional wealth index, and (3) to more realistically estimating fire duration, which in the new model depends on human suppression and weather situations and can last for up to 2 days. One of the major focuses of SEVER-FIRE is an implementation of pyrogenic behavior of humans (timing of their activities and their willingness and necessity to ignite or suppress fire), related to socioeconomic and demographic conditions in a geographical domain of the model application. The importance of description of pyrogenic behavior of humans is confirmed by recent findings of bimodal fire regimes, reflecting the human fingerprint in global fire dynamics (Benali et al., 2017), as well as by differences in timing of ignitions determined by a religious background in sub-Sahara Africa (Pereira et al., 2015). Fire weather regimes set by climate dynamics and fuel state set by vegetation dynamics are other important drivers in SEVER-FIRE. The SEVER DGVM fire module, based on climate observations, external anthropogenic parameters and SEVER-DGVM-derived vegetation, estimates fire incidence and emissions. The resulting vegetation disturbance feeds back to the DGVM, ensuring a fully coupled system (see model description).
We perform a comparison of SEVER outputs with fire data derived from satellite sources, the Global Fire Emission Database version 2 (GFED) (van der Werf et al., 2006) and historical fire data (number of lightning and human fires and their burnt area) with two objectives. First, we aim for a global evaluation of a DGVM fire model, focusing on crucial and simple features, namely fire incidence, seasonality, interannual variability and emissions. Second, (the most important) by identifying the reasons for large inconsistencies, we propose further modifications to SEVER-FIRE. The work presented in this paper is partly based on the PhD thesis by Yannick Le Page. We therefore inform the reader that significant parts of the text in Sects. 3 and 4 already appeared in Le Page (2009).
We are making an effort to create a first-principle global mechanistic fire model. We have named our model “experimental” in order to show that some processes are included in SEVER-FIRE ad hoc (timing of ignition activity of rural versus urban population, others) as mechanisms are still not described or studied, some processes are simplified (e.g., setting maximum time of fire to 2 days but this may be updated and modified in the future by introducing the latest global fire duration datasets; Andela et al., 2018) and some processes are based on statistical descriptions from satellite data (number of on-ground flashes), as they wait their nearest time to be substituted by physically based mechanistic models.
Five of the 35 parameters defined for each of the 10 SEVER PFTs.
The SEVER DGVM is a coupled vegetation–fire mechanistic model designed to
run at a range of temporal (daily to monthly) and spatial (10 km to
2.5
Gridded climate, demographic and socioeconomic data comprise external input
for the fire module. Minimum and maximum daily temperature
The model separates human-induced (indexed as hum) and lightning fires
(indexed as nat) by sources of ignition and all output variables of fire
models can be obtained by either these two classes of fires or in total (not
indexed). We omit the mentioned indexes in description of output variables
further on for simplicity. The output of the model includes number of fires
(
Thus, the DGVM and fire module work in interactive mode, incorporating a representation of fire–vegetation feedbacks.
The SEVER-FIRE model consists of six related components described below:
estimation of fire weather danger index and fire probability, simulation of lightning ignition events and number of lightning fires, simulation of human ignition events and number of human fires, simulation of fire spread after ignition, fire termination, estimation of fire effects (burnt areas, pyrogenic emissions, number of each
PFT's individuals killed). Mathematical expectation of number of ignition produced by one person for
1 million ha The total number of human fires in a grid cell is calculated as Eq. (10): The description of human ignitions in SEVER-FIRE is very simplistic and does
not have the intention of describing, to a major extent, complex economic,
cultural and social practices of people (agricultural, hunting or pastoral,
other) resulting in pyrogenic activities. We left out (or oversimplified,
like in the timing function and mathematical expectation of the number of
ignitions produced by one person) description of an influence of land use on
the number of human ignitions in the fire model because the SEVER
DGVM does not include description
of human land use and/or its influence on natural vegetation. By application
of the SEVER DGVM, we aim to describe relatively human-free global
vegetation, which was given an additional control regulator, namely external
human and/or lightning ignitions. This limitation of the SEVER DGVM implies
certain constraints on our results in both vegetation distribution and burnt
areas, but it also gives us an opportunity to identify and locate the areas
where interaction among land use, fire regimes and vegetation should be
described explicitly and accurately. However, the limitation of maximum fire duration to 2 days was set due to
range in the fire duration of the EFFIS database, which covers mainly European
domain. Globally this limitation may be not valid for remote high-latitude
areas, but even in these regions mathematical expectation of fire duration
will be close to 1 day (see Korovin, 1996). Daily burnt area in the DGVM grid cell is calculated as Eq. (12): Daily burned area estimates are aggregated annually to estimate fire effects.
Percentage of vegetation individuals killed depends on area burned and on
resistance of each PFT to fires (Table 1), taken directly from Glob-FIRM (Thonicke et al.,
2001). The percentages are then converted to emissions, based on vegetation
carbon content (dead PFT individuals are considered to be entirely burned),
and redistributed daily following the profile of fire probability. The model outlined above should be considered an approach to design a global
comprehensive process-oriented fire model based mainly on field observations
and physically based assumptions. Still more analysis needs to be done for
representation of fire processes within the model and calibration of
parameters used in the model. For instance, the study of Scott and
Burgan (2005) indicated that moisture of extinction, used in SEVER-FIRE (see
Table 1) may vary from 12 % to 40 % for different fuel types, i.e.,
has a larger range than in our model. We plan to make sensitivity and
optimization tests to improve the SEVER-FIRE model parameters and
modifications of equations when necessary.
All six components are controlled by PFT-dependent fire parameters (see list
in Table 1).
Observed and simulated number of lightning strikes in the central
cordillera of Canada
Total number of lightning fires observed (Wierzchowski et al., 2002)
and simulated
Registered and simulated number of fires in Canada.
For this study, precipitation data from the National Centers for
Environmental Prediction (NCEP climate data;
minimum and maximum temperature,
precipitation and convective precipitation, shortwave radiation and wind
speed;
Distance from a city was precalculated from population density and the ratio
of rural (urban) population. For this, areas where urban population density
exceeds 400 persons km
GFED is a global 1
The active fire to burned area calibration step and the use of three different sensors to build this dataset generate significant uncertainties on burned area estimates, which are considered to be about 50 % at regional scales, although not quantified in the version of GFED we used (van der Werf et al., 2006). The version of GFED used also does not contain small fires. Emission uncertainties are consequently higher, taking into account their further dependence on the CASA model and on fuel loads and emission factor values.
We chose to focus primarily on burned area to evaluate the model at a global scale, as this is a prerequisite to estimate carbon emissions. However, carbon emission being an essential aspect of biomass burning, its representation is briefly evaluated.
Fire incidence, seasonality and interannual variability from SEVER are compared to GFED data over the 1997–2006 period. The SEVER DGVM considers grid cells to be 100 % land or water. This required a few adjustments on both datasets (regridding of GFED data to SEVER lat–long grid and overlay of two datasets), causing minor changes in the original GFED statistics (less than 3 % for total global area burnt and global fire emissions). We consider burned fraction (BF) rather than burned areas, a latitudinal unbiased indicator of fire density given the use of a lat–long grid.
Fire incidence is mostly dependent on three key factors, conceptualized by
the fire triangle (Schoennagel et al., 2004): fuel availability, readiness
of fuel to burn and ignition source. SEVER spatial patterns of fire
incidence are first compared to GFED, through the mean annual grid cell
BF. BF drivers are then explored with a selection of
relevant environmental variables, based on the fire triangle concept.
Annual amount of precipitation, from the CPC merged analysis of
precipitation (CMAP, Xie and Arkin, 1997), provided by the NOAA/OAR/ESRL PSD,
Boulder, Colorado, USA ( An indicator of dry season severity (DSS), which was constructed from
precipitation (CMAP) and temperature data (NCEP/NCAR reanalysis project; Kalnay et al., 1996) is used. The indicator (Breckle, 2002), representing a
rainfall deficit or a temperature excess, is computed as indicated by Fig. 4. Here we consider it as a rainfall deficit (unit: mm). Net primary productivity (NPP) is used. Its influence on fires is estimated with NPP
estimates from Imhoff et al. (2004) and from SEVER. Land cover spatial distribution is used. The SEVER DGVM vegetation distribution and its
impacts on BF patterns are evaluated with the global land cover for the year
2000 (GLC2000; Bartholomé and Belward, 2005). Human rural and urban population density from global demographic data
collection (Vorosmarty et al., 2000), provided by the University of New
Hampshire, EOS-WEBSTER Earth Science Information Partner (ESIP) is used. An
indicator of the rural predominance of the population was defined (Eq. 13): Gross domestic product (GDP) gridded data (van Vuuren et al., 2007),
provided by the Netherlands Environmental Assessment Agency is used.
We left aside wind speed, which significantly affects readiness of fuel to
burn and fire to spread, as an analyzed environmental variable, due to
constrains put on it in the presented SEVER-FIRE simulations (see
description of input data in Sect. 2.2). However, Lasslop et al. (2015)
demonstrated that modification of rate-of-spread dependence on wind speed
may sufficiently influence burnt areas, so we plan to explore wind speed as a
BF driver in the future.
Definition of the dry season indicator on a climatic diagram as the
yellow patch area. On the
We used CMAP precipitation data (extracted mainly from remote-sensing data) in analysis to obtain a more realistic relationship between observed fire regimes and precipitation. We, however, could not use CMAP precipitation as climate input for the SEVER DGVM due to too short of a period of observations (CMAP started from 1979) and used instead the NCEP reanalysis precipitation data, which are longer and provide bigger ratios between lengths of transient and spin-up simulation periods in DGVM important for realistic description of vegetation and fires. Thus, discrepancies in relationships between fire and precipitation in our analysis for GFED and SEVER-FIRE cases can be, to some extent, explained by differences between NCEP and CMAP precipitation fields. These differences, however, have only a regional character and do not change our general conclusion.
The relationship of chosen variables with fire incidence is not linear, and it involves multivariable interactions. A more in-depth analysis of fire drivers would thus benefit from the use of multivariate statistics. We chose to avoid this level of complexity since the most important conclusions are likely to be drawn from straightforward analysis, as a first evaluation of a global fire model. We thus analyze fire incidence through simple bidimensional plots.
Seasonality is evaluated via the fire season peak, i.e., the month with
maximum fire activity for each grid cell. Interannual variability is
compared to GFED both globally and regionally to identify how the model
performs on specific fire events and for different ecosystems. Again, in a
way similar to fire incidence, fire interannual variability has been shown
to depend on climatic and vegetation conditions. Meyn et al. (2007) highlight
three types of fire ecosystems, depending on their annual fire limitation by
fuel amount, readiness of fuel to burn or both, considering that the
availability of ignition sources is relatively constant in time. Here, we
further explore the climate impact on the readiness of fuel to burn,
analyzing the implications of both fire season precipitation and fire season
maximum temperature for fire interannual variability, along three ecosystem
types (boreal, tropical humid and dry/semidry). To extract those variables,
the extent of the fire season in a grid cell was defined as the months with
more than
Figure 5 shows the spatial distribution of the average annual BF for GFED and SEVER. GFED clearly depicts the most extensively burned continents, i.e., Africa and Australia. It also indicates high fire activity at the edges of the tropical forest, due to land clearing and pasture management, in Central and South America and Southeast Asia (Langner et al., 2007; Morton et al., 2006). Fire incidence is much lower in most temperate and boreal ecosystems, except for the northwestern Iberian Peninsula and Kazakhstan, both regularly affected by fires. A few other regions display high BF values, for example eastern Siberia and Alaska. Note, however, that for ecosystems with a long fire return interval, as is the case in boreal regions, the statistics computed over 10 years are very sensitive to the occurrence of important fire events during that period, and they cannot be considered representative of the long-term regional fire regime. Eastern Siberia, for example, was highly affected by fires in 1998, boosting the 10-year average (Kajii et al., 2002; Le Page et al., 2008).
Mean annual burned fraction (percentage) over 1997–2006.
SEVER accurately reproduces some of the main spatial patterns of fire
incidence, i.e., high BF values over Africa and Australia and very limited fire
activity in the tropical evergreen forest and in most temperate and boreal
regions. For a better emphasis of the discrepancies, Fig. 6 illustrates
the mismatch between GFED and SEVER through a normalized difference burned
fraction index (NDBF) computed as Eq. (14):
Discrepancies in the model outputs relative to GFED observation-derived data, as represented by the normalized difference burned fraction index (see text). Black/grey colors represent grid cells in which fires only occur in GFED or SEVER.
where Regions with low observed fire incidence and the presence of grass in the
model display fire overestimation, regardless of the GLC2000 land cover, and
the more grass, the higher the overestimation. This is the case for example
in North America, India, South America and Papua New Guinea. The
overestimation in these areas can also be caused by high fractional coverage
of croplands, not included in SEVER-FIRE model. Regions with dominant tree cover, or with a large overestimation of trees
in the model, display underestimation of fire incidence. This is the case
in a large strip covering Kazakhstan and eastern Europe and in most of
Southeast Asia, for example. The model underestimates the very high fire incidence observed in
sub-Saharan Africa.
Considering drivers of BF spatial distribution, Fig. 8 illustrates the
interactive influence of paired combinations of the previously described
variables. In GFED, the most affected regions are clearly constrained by
annual precipitation between 500 and 1500 mm yr
SEVER DGVM land cover distribution, grouped in three broad classes: bare
soil, grass (C
Mean annual burned fraction over 1997–2006 (
Finally, Fig. 9 displays the mean annual carbon emissions for GFED and SEVER. Emissions are mainly dependent on fire incidence, the type and moisture content of the affected vegetation, and fire severity. In SEVER, dead PFT individuals are entirely emitted to the atmosphere, while GFED takes into consideration combustion completeness. Consequently, the absolute level of emissions cannot be compared, being much higher in SEVER, as expected. However, the spatial patterns reveal the importance of tropical savannas and forests in the global partitioning of carbon emissions in both GFED and SEVER, as well as a significant contribution from boreal regions. We are planning to correct SEVER for combustion completeness as well as for post-fire mortality processes.
Figure 10 shows the spatial patterns of the month with maximum fire activity for each grid cell and the mismatch between GFED and SEVER. SEVER roughly reproduces the observed spatial patterns, with 73 % of the grid cells with a mismatch lower than or equal to 2 months. Significant discrepancies occur in sub-Saharan Africa, which peaks over March to June in the model, while GFED, along with other observation sources, indicate October to February (Barbosa et al., 1999a; Clerici et al., 2004; Dwyer et al., 2000b).
Mean annual emissions (g C m
Sub-Saharan Africa is a major fire region (Dwyer et al., 2000c; Tansey et al., 2004), contributing to a large fraction of global fire activity from October to February, a period when most other regions experience little or no fire activity. As such, the inability of SEVER to reproduce fire seasonality in sub-Saharan Africa is one of its major current limitations. Delayed fire season is also significant in central North America and southeastern Australia.
Averaged correspondence of fire season with dry season anomalies over regions of sub-Saharan Africa with a delay in peak month superior or equal to 4.
The fire seasonal cycle is partially driven by climate, but it can also be
strongly influenced by human activities. Figure 11 illustrates the averaged
profile of the fire season and the dry season over sub-Saharan Africa for
those grid cells with a SEVER fire peak discrepancy larger than or equal to
4 months. For each of these cells, we computed the monthly fire season,
centered the peak month on the
In regions with lower use of fire as a management tool, as in boreal forests, the model performs much better and, along with the observations, tends to place the peak month in the middle or late dry season (not shown). The implication of these findings for model improvement is detailed in the discussion section.
Figure 12 shows the grid cell correlation between annual BF time series from GFED and SEVER. Equatorial Asia, Mexico and the majority of boreal regions, along with part of South America, are in good agreement. As discussed later, those regions are characterized by their sensitivity to climate variability, especially to the El Niño of 1997/98 (Le Page et al., 2008). The poorest agreement is found in Africa, India, China, western Russia, south of the US Great Lakes and in parts of South America.
Interannual variability is further analyzed using a set of 13 regions, originally created for GFED analysis (Giglio et al., 2006) as represented in Fig. 13. Globally, and for each of those regions, Fig. 14 shows the BF interannual anomalies from GFED and SEVER, along with the monthly distribution of fire activity as a further indicator of the timing of specific fire events and of fire seasonality. The very poor agreement in the global plot was to be expected, given the discrepancies in mean spatial fire incidence (Fig. 5), resulting in different contributions from regions to the total fire anomalies. This is clearly revealed by the monthly plot, showing that total fire activity in December–February, peaking in GFED with the large contribution of sub-Saharan Africa, is very low in SEVER. Consequently, a given fire anomaly in Africa has a much bigger global impact in GFED than in SEVER.
Correlation of annual BF from GFED and SEVER, over 1997–2006.
Regional partitioning allows identification and comparison of specific fire events more easily, especially the ones driven by large-scale climatic variability. The El Niño episode of 1997–1998 appears clearly in the BONA, CEAM, BOAS and EQAS regions in the observations and is generally captured by the model with precise timing. Annually, the importance of those events is also reproduced for EQAS and BOAS, with, respectively, 1997 and 1998 being the peaking years in GFED and SEVER. Generally, fire patterns in the other regions are not properly represented. The monthly resolution plots also give further insight into the regional scale seasonal cycle, which is generally reproduced very well, except for Northern Hemisphere Africa and Australia.
Regions used for interannual variability analysis. BONA: boreal North America; TENA: temperate North America; CEAM: Central America; SOAM: South America; EURO: Europe; NHAF: Northern Hemisphere Africa; SHAF: Southern Hemisphere Africa; BOAS: boreal Asia; CEAS: Central Asia; SEAS: Southeast Asia; EQAS: equatorial Asia; AUST: Australia.
Figure 15 displays the dependence of fire anomalies on precipitation and
temperature anomalies over the fire season, through their effect on soil and
vegetation moisture status. Drought conditions are the main prerequisite for
fire occurrence within all vegetation types, although in low NPP ecosystems
low vegetation amount can be a limiting factor, resulting in a dependence of
fire anomalies on growing season precipitation (Holmgren et al., 2006; van
der Werf et al., 2008). The relationship is first pictured globally
(Fig. 15), showing that both precipitation and temperature anomalies are
strong drivers, constraining positive fire anomalies almost exclusively to
precipitation deficits and towards positive temperature anomalies. This
relationship is then analyzed in GFED for three types of ecosystems.
Boreal ecosystems, a spatial aggregation of the BONA and BOAS regions, are analyzed. Boreal fires are shown to be strongly dependent on
temperature, at a level comparable to precipitation. Tropical humid regions, selected within South America, Africa and Equatorial
Asia, with the pixels with annual precipitation above
1500 mm, are analyzed. Their fire anomalies are also strongly related to
precipitation, while temperature is a weak driver. Semidry and dry African and Australian regions (annual precipitation below
500 mm), which are characterized by high anthropogenic fire
activity, are analyzed. For those regions, both fire season precipitation
and temperature anomalies are poor predictors of fire anomalies.
Those patterns are well reproduced on a global scale, such that the patterns
of dependence on both climatic variables are similar in the model and in
observations (Fig. 15). In boreal or tropical humid ecosystems, SEVER shows the
same trends towards more or less dependence on temperature, although not as
neatly as in GFED. In the case of semidry and dry African and Australian
regions, the model also shows a weaker dependence on precipitation and
temperature, but stronger than in the observations.
Perhaps one of the most important achievements of SEVER, as revealed by this study, is the realistic modeling of strong climate-driven fire anomalies, such as the large biomass burning events resulting from El Niño-induced droughts in various regions of the world (Figs. 12 and 14). This climate-induced variability is known to be considerable and has important consequences for atmospheric composition, the terrestrial carbon cycle and biodiversity, as discussed in the introduction. As such its accurate representation in DGVMs and ESMs is essential.
Regional comparison of fire variability over 1997–2006. For each region subplot the top shows annual anomalies and the bottom shows monthly time series constrained to [0, 1]. The region name is indicated at the top left corner and the average fire incidence at the top right.
Dependence of fire anomalies on temperature and precipitation.
The in-depth analysis of this climatic influence highlights the variability in temperature–precipiation dependence patterns (Fig. 15). Boreal regions are characterized by great annual amplitudes of precipitation and temperature. As such, both play an important role in the dynamics of soil and vegetation moisture status, through rainfall and evaporation, and thus the strong fire dependence on both variables. In tropical humid regions, temperature variability is much lower, and only a major and prolonged precipitation deficit will result in fire-prone conditions (van der Werf et al., 2008).
Finally, semidry and dry regions of Africa and Australia are characterized by a low dependence on both parameters. Those regions are under specific climatic conditions, characterized by a rather short and irregular wet season for vegetation growth, followed by a long dry season (Peel et al., 2007). Under those conditions, fuel availability, rather than its readiness to burn, limits the occurrence of fires (Meyn et al., 2007). Under low wet season precipitation, vegetation buildup may be too low to sustain a fire. Under high wet season precipitation, vegetation growth leads to less patchy vegetation, which will dry out over the following dry season, becoming highly susceptible to fires. This scheme is very specific of those hot dry and semidry regions dominated by annual herbaceous vegetation. In the case of middle- to high-productivity ecosystems with the presence of woody vegetation, the relationship is generally reversed: enhanced wet season precipitation leads to a higher soil and vegetation moisture status, delaying desiccation over the dry season and thus reducing fire susceptibility. The contrast between those two distinct vegetation–climate–fire relationships is most evident in Australia (Fig. 16). The SEVER vegetation scheme did not perform very well over Australia, and so the role of wet season precipitation is not properly represented (not shown).
At a global scale, SEVER is shown to be fairly realistic regarding this temperature–precipitation dependence, which was to be expected since both variables are involved in the fire weather danger and fire spread calculations. However, the variability in the relationship along ecosystem types (boreal, tropical humid, semidry/dry), resulting from complex interactions among fire drivers, is not as straightforward to capture. The realistic results for such an interactive system suggest that the feedback mechanisms as defined in the SEVER DGVM–SEVER-FIRE coupled scheme do reach a reasonable level of complexity and accuracy, especially in the case of boreal and tropical ecosystems.
The mean BF (Fig. 5) is a more challenging feature for the model to
replicate. Key associations represented in the fire triangle (Schoennagel et
al., 2004) are, however, reproduced (Fig. 8), i.e., fire occurrence
limitation by moisture in very humid ecosystems or by low fuel amount in arid
regions. Unfortunately, SEVER models potential – not actual – vegetation
cover, hampering an in-depth diagnostic of the fire incidence estimates.
However, grass–trees appear to be over- or under-sensitive to fires, with
the exception of highest-fire-incidence regions (Africa, northern Australia),
where SEVER underestimates fire activity, independent from vegetation cover
(Figs. 6 and 7). The main PFT parameters controlling fire incidence are bulk
density (fire ignition and spread; see Table 1) and flammability (fire danger
index computation). Flammability takes the same value for all tree PFTs and a
distinct value for both C
Dependence of fire anomalies on wet season precipitation and land cover type in Australia for GFED data.
It is also essential to improve our understanding of anthropogenic impacts on
fire incidence. The initial assumptions of the model, with population and
wealth status as the most important human proxies, are to be reassessed
carefully in regional studies, given the implication of other factors. In
particular, the most evident cases of human-induced increased or decreased
fire activity are related to land use type and agricultural practices, more
than to economic and social status. For example, Pfeiffer et al. (2013)
divided population into three according to their dominating land use types:
farmers, pastorals and hunter–gathers. Kaplan et al. (2016) showed that this
division determined structure of burnt areas during the Last Glacial Maximum.
Thus, a simple timing function for rural population implemented into the
SEVER DGVM may not work properly in
Africa. Relating those ignitions to low wealth status, as in SEVER, is
certainly functional after a few adjustments but seems less robust to other
regions than an association of land use with timing of human pyrogenic
activities and number of human ignitions. As an illustration, wealth status
is not well adapted to account for high fire incidence induced by humans in
northern Australia (Russell-Smith et al., 2007). Additional proxies for human
pyrogenic activities implemented in SEVER-FIRE could include deforestation
activities (Zhan et al., 2002) and land use and land cover data (Thenkabail
et al., 2006). A fire management factor should be added to the model in the
regions where a coordinated wildfire controlling program is in place (e.g.,
existence and actions of European Commission Emergency Response Coordination
Centre in Europe;
Advantages of including the relationship between land use and timing of pyrogenic activities in SEVER would possibly also extend to a better representation of fire seasonality. In sub-Saharan Africa for example, Fig. 11 reveals that the fire season (October–February, Fig. 10) is shifted towards early months of the dry season, which mainly results from the use of fires for agricultural and land management practices (Clerici et al., 2004). For the whole Southern Hemisphere, however, human pyrogenic activity in SEVER is set to reach a maximum from March to May and September to November, which is not realistic in the case of sub-Saharan Africa, a major fire region. Timing of pyrogenic activities in sub-Saharan Africa may be rather challenging as even implementation of land use in a global fire model (Le Page et al., 2015) still brings a 1- to 3-month delay in fire peak. Furthermore, it was demonstrated that religious affiliation modulates agricultural burning activities in the area (Pereira et al., 2015), which was not taken into account by global fire modelers at the time. It is seen that a set of regional case studies with an active use of available historical data is necessary to implement more realistic features of human pyrogenic activities in global fire models. Study and parameterization of fire duration in remote areas is necessary for improvement of burnt area calculation in these areas.
Description of lightning fires also needs improvements, starting from estimation of the number of lightning strikes effective for fire ignition. Despite lightning strike being considered, to a major extent, a stochastic event, there is visible room for better description of the number of cloud-to-ground flashes based on recent findings of the role of aerosols in electrification of thunder clouds (Stolz et al., 2015; Venevsky, 2014). In addition, a sensitivity study for critical newly implemented features timing and duration and further formal optimization for parameters of SEVER-FIRE using a teaching subset of remote-sensing data for observed burnt areas (Khvostikov et al., 2015; Rabin et al., 2015, 2018) can further improve performance of the presented global fire model.
This paper analysis results from a DGVM that includes an interactive, dynamically linked fire module. It reveals that the most important climate-driven fire features are reproduced by the model, while the dependence on vegetation characteristics and, especially, human pyrogenic activities prevents the further development of realistic estimates of fire incidence and of regional to global interannual variability. Regional adjustments of global fire models based on analysis of both historical fire statistics and records and recent satellite observations are necessary for further understanding of global fire dynamics in the past, present and future.
SEVER-FIRE is presented in its 1.0 version, which is
realized in the FORTRAN language. It is open-use scientific software. The
source code of SEVER-FIRE and the socioeconomic input data can be accessed
freely from
SV developed equations of the fire model. SV and CW worked on the code. YLP and JMCP suggested the scheme of the model validation and made the validation. SV, YLP, JMCP and CW equally contributed to writing the paper.
The authors declare that they have no conflict of interest.
We thank Guido van der Werf for providing the GFED data and for helpful
comments on the paper. This study is funded by the Marie Curie Research
Training Network GREENCYCLES, contract number MRTN-CT-2004-512464
(