Fire Regions as Environmental Niches: A New Paradigm to Define Potential Fire Regimes in Africa and Australia

Despite the widespread use of the “fire regime” concept for describing spatial and temporal patterns and ecosystem impacts of fire, this concept lacks an unambiguous, quantitative definition. By adopting from the ecological literature the concept of climate niche, that is, the environmental conditions that allow a specie to exist, we propose a new framework where variables that promote fuel accumulation and desiccation were used to define the environmental space at the continental level, later divided into regions (“fire regions”) with distinct fire potential. Our proposed approach emphasizes climate controls on fire patterns, distinct from the controls that humans exert on observed fire activity. By applying this framework, we identified nine fire regions in Africa and eight in Australia, distinguishing differences in fire patterns between continents as a result of changes in environmental gradient. Not only did we find that fire size and intensity varied significantly between continents, but biomes at a continental level were also found to be heterogeneous in terms of fire frequency, size, and intensity. For example, within African tropical savannas, the total annual rainfall and tree cover change drastically North and South of the equator, resulting in fire regions with significantly different fire characteristics. Meanwhile, in Australia, a strong gradient of annual temperature and precipitation seasonality was observed within tropical savannas and xeric shrublands, which was recognized by dividing those biomes into five regions with statistically different fire activity. Additionally, human presence led to some heterogeneity of fire patterns within delineated fire regions that also varied across biomes.

A common approach for delineating fire regimes is using several fire metrics in an empirical, data-driven framework (e.g., Archibald et al., 2013;Conedera et al., 2018;Faivre et al., 2011;Le Page et al., 2008). In recent decades, virtually all databases of fire extent and fire characteristics rely on satellite data. Multi-decadal higher resolution (30-250 m) databases are available for North America (e.g., MTBS (Picotte et al., 2020), ABoVE (Loboda et al., 2017)), Australia (e.g., NAFI (NAFI, 2022)), and part of Europe (e.g., EFFIS (San-Miguel-Ayanz et al., 2012)), whereas coarse resolution (500 m) remotely sensed imagery is, for most of the globe, the only available data source that provides coherent multi-temporal fire spatial information since the year 2000 (Benali et al., 2017). The use of global satellite products plays a key role in characterizing fire regimes from the local to the global scale by allowing the systematic computation of fire-related summary metrics such as fire extent, inter-annual variability, seasonality, and intensity . The choice of fire metrics is constrained to which variables can be quantified at the chosen scale from the selected source of fire data. Fire regime studies typically delineate fire regimes geographically by identifying spatial clusters by way of principal component analysis (PCA, Jiménez-Ruano et al., 2017;Whitman et al., 2015), clustering (e.g., hierarchical (Conedera et al., 2018), k-means ) or a combination of both (Archibald et al., 2013;Le Page et al., 2008).
While the above-mentioned approach is appealing due to its simplicity, it does not take into consideration the underlying biophysical processes that regulate fire activity, and it is inherently dependent on a number of implicit assumptions. The resulting "fire regime" classification is sensitive to the choice of fire metrics taken into consideration, the mathematical formulation of the metrics, and the spatial scale at which the metrics are computed. Delineating fire regimes based solely on a small number of fire metrics can result in attributing the same fire-regime label to regions where the set of considered fire metrics happens to assume similar values, but the environmental conditions and the ecological implications of fires are very different. Additionally, studies that classify fire regimes from remotely sensed data are fundamentally constrained by the length of the available data archive, which go back at most to 2000 for consistent, well-calibrated global satellite data (Mouillot et al., 2014). Together these issues can lead to a misinterpretation of the processes that drive current fire activity, as well as a misunderstanding of how fire activity might change in the future.
Because of the inherent ambiguity of the concept of fire regime, we propose in this study a new analysis framework based on fire regions, drawing from the concept of climate niche. This concept is commonly employed in the ecological literature and broadly defined as "the states of the environment which would permit a species to exist" (Hutchinson, 1957). More specifically, in terrestrial ecology, climate niches are defined in terms of geographic and climatic factors that define the habitat of a species (Bonetti & Wiens, 2014;Holt, 2009). We thus define "fire regions" as areas that are homogeneous in terms of environmental conditions that promote fuel accumulation and desiccation, adopting a conceptual model that links specific climate metrics to the key drivers of fire proposed by Bradstock (2010).
The proposed framework was applied in Africa and Australia, the two fire-prone continents responsible for over 75% of the global burned area (Giglio et al., 2018). Tropical savannas are the dominant biome affected by fire on both continents, followed by Xeric shrublands and Mediterranean forest and woodlands (Giglio et al., 2018;Melchiorre & Boschetti, 2018;van der Werf et al., 2017). Similarly, a global study by Archibald et al. (2013) that mapped fire regimes on both continents found that Africa and Australia shared the same fire regimes except for the parts of tropical savannas in Africa that experienced fires smaller in size and with lower intensity, which was primarily attributed to differences in anthropogenic pressure. However, studies by Hantson et al. (2017) and Lehmann et al. (2014) reported dissimilarities in the overall relationships between fire-regime metrics, moisture, and vegetation cover, with the implication that savannas in these two continents might respond differently to future climatic changes.
This study aims to define continental-scale maps of major regions with similar environmental conditions that drive fire activity in Africa and Australia. The hypothesis is that the fire regions delineated solely from environmental drivers would have distinct values of the fire activity metrics, whereas variations of fire activity within a region would be an indicator of anthropogenic pressure. To this end, we first selected the most important drivers of fire activity in each continent and then used these drivers to define the multi-dimensional environmental space, which was subsequently clustered into distinct fire regions.
Fire-regime metrics (fire frequency, intensity, and size) were utilized to verify the hypothesis that the set of key environmental drivers of fire activity can provide sufficient information to separate regions with different fire characteristics. Finally, we located areas with various degrees of human pressure and explored the gradient of anthropogenic influences on fire activity across the environmental space. We intend that aggregated fire regions can be used as a scale unit for computing fire statistics and assessing current and future trends as a more ecologically meaningful alternative to current fire-regime maps.

Potential Environmental Drivers of Fire Activity
Our proposed approach for defining fire regions is based on the conceptual model of key drivers of pyrogeography proposed by Bradstock (2010), where fire activity is restricted by four "switches": biomass, availability to burn, fire spread, and ignition. Based on previous studies, we identified 12 environmental variables that potentially influence Bradstock's "switches" on a broad scale, including total precipitation, temperature, their intra-and inter-annual variability, water deficit, evapotranspiration, vegetation type, and biomass (Table 1). All variables were calculated/extracted within individual grid cells of a global 0.5° analysis grid, using the same temporal interval (March 2002-February 2019) used to generate the fire activity metrics (Section 2.3). The March-February optimal fire year definition (Boschetti & Roy, 2008) was used instead of the January-December calendar year to avoid artifacts in the time series of annual fire metrics due to the timing of the fire season in Northern Hemisphere Africa (November-February), which would be split between calendar years. Because several of these variables are highly correlated, we retained only six based on the selection procedure described in Section 3.1. For brevity, only these six are described in detail here ( Figure S1 in Supporting Information S1): • Mean annual temperature (Temp, °C) was defined as the annual average over 17 years. Temp was calculated from the Famine Early Warning Systems Network Land Data Assimilation System (FLDAS, McNally et al., 2017) that used outputs from Modern-Era Retrospective Analysis for Research and Applications (MERRA-2, Rienecker et al., 2011) and Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS, Funk et al., 2015). This data set is available as monthly means at 0.1° resolution. Many authors consider Temp an important driver of global fire activity (Aldersley et al., 2011;Gillett et al., 2004;Krawchuk & Moritz, 2011). Lehmann et al. (2014) depicted the strong relationship between temperature and fire frequency in Africa and Australia due to temperature's ability to induce plant growth, that is, fuel accumulation, and reduce fuel moisture, which creates favorable conditions for fire ignition and propagation (Daniau et al., 2012;Flannigan et al., 2009). • Precipitation Interannual Variability (P_Var) is calculated as the interannual coefficient of rainfall variation, that is, the standard deviation of the time series of annual rainfall divided by its mean based on FLDAS data; the annual precipitation totals were computed using the fire year definition. Moisture fluctuation, specifically in semi-arid regions, affects fuel accumulation rates and fuel connectivity, which enables bigger, more intense fires (Bradstock, 2010;Murphy et al., 2013). Consequently, the extent of area burned varies substantially between years with a "wet" dry season and with a "dry" dry season (Mulqueeny et al., 2011). Additionally, in forested regions where high levels of moisture usually prevent fire ignition and propagation, prolonged droughts can create suitable conditions for fire (Cochrane, 2003;van der Werf et al., 2008). • Precipitation amplitude (P_Amp, mm) was defined as the difference in average precipitation between the wet and dry seasons following the Saha et al. (2019) approach (see SI for calculation) based on FLDAS data. It is widely accepted that the temporal fluctuation in moisture availability is one of the main drivers of fire activity Markham, 1970;Murphy & Bowman, 2007;Staver et al., 2011), and here P_Amp was used to characterize the weather conditions that favor desiccation, which is necessary for fire ignition and propagation once sufficient fuel is accumulated. • Seasonality index (P_SI) characterizes the extent to which a yearly precipitation cycle can be partitioned into well-defined wet and dry seasons (Saha et al., 2019, SI), thus complementing P_Amp in describing the intra-annual precipitation variability. P_SI assumes values close to 1 if a particular region has two distinct 10.1029/2021JG006694 4 of 22 precipitation seasons (i.e., one wet and one dry season) and close to 0 otherwise. Prolonged dry seasons in mesic savannas desiccate fuels and contribute to high burning rates (Leys et al., 2018;van der Werf et al., 2008). In wet environments, the absence of a highly defined dry season promotes tree cover growth, where an increase in the duration of the dry season reduces tree cover and increases the chances of fire occurrence (Staver et al., 2011). • Percent tree cover (Tree, %) calculated from the Collection 6 MODIS MOD44B Global Vegetation Continuous Fields product (DiMiceli et al., 2017). Fire behavior varies strongly across the savanna-forest boundaries (Hoffmann et al., 2012;Zubkova et al., 2019). The type of vegetation influences the likelihood of fire (Lehmann et al., 2011), andArchibald et al. (2009) identified tree cover as one of the main drivers of the amount of burned area in Southern Africa due to its negative impact on grass biomass and, therefore, the fire spread. • Mean annual actual evapotranspiration (AET, mm) was acquired from TerraClimate (Abatzoglou, Dobrowski, et al., 2018) based on a one-dimensional water-balance model (WBM, Willmott et al., 1985). Fuel productivity and dryness are a function of climatic water balance (Boer et al., 2016). At annual timescales, AET can be considered a reliable predictor of continental patterns of primary productivity (Rosenzweig, 1968;Yang et al., 2013). Actual evapotranspiration (*) AET, mm Calculated using a Thornthwaite-Mather climatic water-balance model (Willmott et al., 1985) TerraClimate (Abatzoglou, Williams, et al., 2018) Potential evapotranspiration PET, mm Calculated using Penman-Monteith approach (Allen et al., 1998) TerraClimate (Abatzoglou, Williams, et al., 2018) Climatic water deficit DEF, mm Defined as the difference between PET and AET TerraClimate (Abatzoglou, Williams, et al., 2018) Vegetation Note. An asterisk indicates the six climate and vegetation variables retained after the initial screening (Section 3.1). All climate variables are calculated from 17 years of data (2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019).

Anthropogenic Pressure
Human presence profoundly alters fire activity in any ecosystem, with multiple direct and indirect effects on the number, extent, and intensity of fire events. Humans increase the number of ignitions but also lead to fire suppression, as well as alterations to vegetation composition, biomass availability, and landscape connectivity, that is, the human population density was found to impact three of the four "switches" of the Bradstock (2010) model (Aldersley et al., 2011;Archibald, 2016;Archibald et al., 2009;Kahiu & Hanan, 2018). In addition to population density, we considered the proportion of natural vegetation and the livestock density as variables reflecting different aspects of anthropogenic pressure relevant to fire (Archibald, 2016;Archibald et al., 2009Archibald et al., , 2013Guyette et al., 2002;Sanderson et al., 2002;Syphard et al., 2017). The proportion of natural vegetation is a proxy for fuel connectivity and landscape fragmentation, a decrease of which negatively impacts fire spread, whereas grazers remove biomass and compete with fire for fuel, thus having a negative effect on the area burned and on fire intensity.
The three gridded anthropogenic pressure variables are defined as follows: The anthropogenic pressure variables were used in the evaluation of fire regions (Section 3.3).

Fire Activity Metrics
Three metrics of fire activity, commonly used in published local to global fire studies, were calculated on 0.5° grid cells from the available MODIS fire products ( Figure S3 in Supporting Information S1): • Percent mean annual burned area (BA, %), calculated as the annual average over 17 years from the Collection 6 MODIS global burned area product (MDC64A1). MCD64A1 is a monthly gridded product, mapping the burned area location and approximate day of burning at 500 m resolution (Giglio et al., 2009). • Maximum fire size (FSize, km 2 ) was defined as the 95th percentile of the fire size distribution. The fire size distribution was calculated by extracting individual fire events from the MCD64A1 product through a novel flood-fill algorithm that uses the spatial and temporal burned area information and the uncertainty layer to define a locally adjusted temporal threshold (Humber, 2019). Due to the reduced accuracy of small fire detection , fires covering less than five cells (<1.07 km 2 ) on the 500-m MODIS sinusoidal grid were excluded from the analysis. • Maximum fire radiative power as a proxy for fire intensity (FRP, MW) was defined as the 95th percentile of the FRP distribution extracted from the Collection 6 MODIS MCD14ML fire location product (Giglio et al., 2016). MCD14ML is a summary product reporting all the active fires detected as actively burning at the satellite overpass under relatively cloud-free conditions with a nominal spatial resolution of 1 km. FRP was calculated by taking into account the variable pixel size in the across-track direction (Giglio, 2013). Only fire points with "nominal" or "high" confidence, detected within natural land, were included in the analysis.
As with the environmental variables, fire metrics were computed using the fire year definition utilizing data from March 2002 to February 2019. The first two years of MODIS data (2000)(2001) were not used due to sensor calibration problems and acquisition outages (Giglio et al., 2016).
The fire activity metrics were used to inform the selection of the environmental variables (Section 3.1) and evaluate the resulting fire regions (Section 3.3).

Burnable Land Mask
To avoid assigning fire regions to areas that do not experience fire activity due to the lack of vegetation, we created a burnable land mask. First, the proportion of burnable vegetation (i.e., all land cover types except urban, bare soil, and water from the CCI LC product) was calculated within each 0.5° cell. The relationship between BA and the proportion of vegetation was subsequently analyzed within each continent. We found that BA increases with the proportion of burnable land following an exponential distribution, with nominal fire activity in areas of low vegetation cover (<0.74 in Africa, <0.88 in Australia; Figure S4 in Supporting Information S1). These thresholds were calculated using the Unit Invariant Knee point as an objective estimator for the elbow criterion (Christopoulos, 2016). All grid cells with the proportion of burnable vegetation below estimated thresholds were excluded from further analysis.

Methods
The fire regions were mapped by means of a statistical analysis based on principal component (PCA) and clustering of gridded environmental driver variables, following established approaches for identifying regions with similar spatiotemporal characteristics. PCA is used to identify patterns in the data which might not be obvious when a data set contains a large number of variables that are not statistically independent (Le Page et al., 2008), which is common in ecological and climatological studies (e.g., Abatzoglou et al., 2009;Archibald et al., 2013;Faivre et al., 2011;Whitman et al., 2015). While PCA can handle some degree of multicollinearity, previous research has suggested setting aside highly correlated variables to reduce the bias of PCA results (Jiang & Eskridge, 2000).
Including the nearly-redundant variables can cause the PCA to overemphasize their contribution. For this reason, the first processing step ( Figure 1, green dotted box) was the selection of the input variables, removing those with high collinearity (Section 3.1). The fire regions were then identified through hierarchical clustering of the principal components (PC) of the inputs, a method commonly used in fire studies since (a) it does not require the prior definition of the number of target clusters and (b) it is suitable for the definition of biogeographic regions which are hierarchically arranged (Conedera et al., 2018;Faivre et al., 2011;Le Page et al., 2008). The clusters were aggregated using the WWF Terrestrial Ecoregions of the World (Olson et al., 2001), resulting in the final fire region map ( Figure 1 red dotted box, Section 3.2). Finally, the fire region map was evaluated by analyzing the difference in fire characteristics across regions and the impact of anthropogenic factors within regions ( Figure 1 blue dotted box, Section 3.3).

Selection of the Environmental Drivers
Collinearity among variables was assessed by calculating pair-wise Pearson's correlation coefficients separately for Africa ( Figure S5 in Supporting Information S1) and Australia ( Figure S6 in Supporting Information S1). Correlation coefficients higher than 0.65 or lower than −0.65 indicated high collinearity, adopting the threshold value frequently utilized in the literature (Monjarás-Vega et al., 2020;Oliveira et al., 2014). Uncorrelated variables (P_Amp and P_SI in Africa, P_Var and P_SI in Australia) were directly included in the PCA, while the rest had to undergo additional exploratory analysis based on bivariate generalized additive models (GAMs, Wood & Wood, 2017). GAMs were generated to investigate the relationship between each of the remaining highly correlated variables and the three fire metrics in Africa ( Figure S7 in Supporting Information S1) and Australia ( Figure  S8 in Supporting Information S1). Among sets of highly correlated variables, the one which is a better predictor for a majority of the fire metrics (i.e., higher explained deviance, reported in the upper right corner of each plot in Figures S7 and S8 in Supporting Information S1) was selected as follows.
In Africa: • Tree was selected from the highly correlated set {Tree, Biomass, P} because it was a better predictor of BA and FRP. • AET was selected from {PET, DEF, AET} because it was a better predictor for BA and FRP. • Temp was selected from {Temp, T_Max} because it was a better predictor for FSize, and substantially equivalent for BA and FRP. • P_Var was included because it correlates highly only with P, which was excluded as detailed above.
10.1029/2021JG006694 7 of 22 In Australia: • P_Amp was selected from {P_Amp, P, AET, PET, DEF, T_Amp} because it was a better predictor of BA and FSize. • Temp was selected from {Temp, T_Max} because it was a better predictor of all three fire metrics. • Tree was selected from {Tree, Biomass} because it was a better predictor of all three fire metrics.
The two steps of the exploratory analysis resulted in the selection of six environmental drivers in Africa (Temp, P_Amp, P_SI, P_Var, Tree, AET) and five in Australia (Temp, P_Amp, P_SI, P_Var, Tree).

Fire Region Mapping
PCA and clustering were performed separately for Africa and Australia since these two continents vary in terms of vegetation cover and level of anthropogenic pressure, which can alter the climate-fire relationships among continents (Archibald et al., 2010;Lehmann et al., 2014;Stevens et al., 2017). To ensure that all variables have equal weights, all data were normalized by rescaling the values between 0 and 1 and centered (Mardia et al., 1979). Clustering was based on a Euclidean distance matrix computed using all PCs with eigenvalue >1 (Girden, 2010). Grid cells were incrementally merged into clusters using Ward's linkage method (Ward, 1963), according to which the distance between two clusters is equivalent to an increase of the sum of squares when they are being merged. The number of clusters was determined by inspecting the dendrograms. Following the recommendations by Wilks (2005), moving from the bottom to the top of the dendrogram, at each step, the distance between merged clusters is computed as defined above, and the merging process is stopped one step before the distance between merged clusters jumps markedly (i.e., before a discontinuity in the derivative of the distance). The selection of the number of clusters was a compromise between the highest separation between clusters based on the dendrograms and selecting enough clusters to represent the full range of environmental gradients at the continental level.
Due to the continuous nature of environmental variables, clustering of grid cells can lead to a noisy output map, making it difficult to interpret or use for ecological applications. For that reason, once each 0.5° grid cell was assigned to a cluster, the results were aggregated at the ecoregion scale using a majority overlap criterion. The Olson et al. (2001) ecoregions, defined as large areas of relatively uniform climate and vegetation that are refuge to a specific set of species and ecological communities, were used because of commonalities in vegetation that mediate climate-fire relationships (Abatzoglou, Williams, et al., 2018;Pausas & Fernández-Muñoz, 2012;Zubkova et al., 2019).

Evaluation of the Fire Regions
To evaluate the separability of the proposed fire regions in terms of fire activity, we compared fire metrics within each region using the Wilcoxon signed-rank test. The Wilcoxon test is a nonparametric test that can be used to test the significance of the differences between the means of two samples (Wilcoxon, 1945). Fire regions were analyzed within each continent and between continents to test if similar environmental conditions lead to similar fire activity in different geographical regions.
Additionally, within each fire region, we tested if different levels of anthropogenic pressure led to significant variation in fire activity, thus separating environmental drivers of fire from human influences to evaluate human drivers within different climate niches. We ran the PCA and clustering procedure on the three anthropogenic pressure variables for each continent, resulting in continental maps of clusters homogeneous in terms of anthropogenic pressure. We then stratified each region by the level of anthropogenic pressure to verify whether it significantly affected the fire metrics.

Fire Regions
The PCA analysis was performed to reduce the dimensionality of the set of environmental drivers; in both continents, only the first three PCs had eigenvalues >1, and were therefore used in the subsequent hierarchical clustering. Like Faivre et al. (2011), we cut the dendrograms into several clusters of increasingly complex biophysical patterns. A simple pattern would lead to 5 clusters in Africa and 4 clusters in Australia, while a more complex pattern included 9 clusters in Africa and 8 in Australia ( Figure S9 in Supporting Information S1). Beyond this, the distance between clusters was negligible, meaning similarities between clusters were drastically increasing; therefore, further splitting of the data was not justified. Figure 2 shows the fire regions as scatterplots in the space of the first three PCs, whereas Figure 3a shows the gridded fire region map at 0.5° resolution, which were subsequently aggregated at the ecoregion scale through the application of a majority filter (Figure 3b).
To allow for unambiguous labeling of the fire regions, each is identified by a three-digit numerical code, where the first digit indicates the realm (1 -Australasian, 3 -Afrotropical), the second digit indicates the label from the first split (simple pattern), which was based on the mean value of the first PC, ranked from high to low, while the third digit is based on the further split ( Figure S9 in Supporting Information S1). It follows that regions whose codes have the same first two digits share more similarities than regions with entirely different codes. For example, the environmental space between clusters 321, 322, and 323 is more similar than between clusters 321 and 311. The digit "1" in the numeric code "321" means that cluster 321 had the highest mean value of the first PC compared to clusters 322 and 323 (Figures 2 and 3). In both continents, the first two PCs were driven by the Tree, P_Amp, and Temp gradient (Figure 2). In Africa, the primary gradient was related to P_Var, P_Amp, and Tree, splitting the continent into regions with low P_Var and high Tree (Figure 3, region 311-323, P_Var = 0.1 and Tree = 32.5% on average) and high P_Var and low Tree (regions 331-351, P_Var = 0.2, Tree = 6.6%). The second PC was driven by Temp and AET that further split high P_Var regions into hot, Sahel (region 331-332, Temp = 27.8 °C), medium, Horn of Africa (region 351, Temp = 25.0 °C), and cold, deserts and xeric shrublands (region 341-342, Temp = 20.7 °C), while the third PC was related to P_SI that helped to differentiate between tropical forest (region 311, P_SI = 0.7) and savannas (regions 321-323, P_SI = 0.9). In Australia, Temp was a primary gradient that split the continent into hot (region 111-123, Temp = 24.2°C) and cold (region 131-141, Temp = 17.8°C) regions, where the second PC attributed to P_Amp and Tree split hot regions further into high P_Amp, tropical savannas (region 111-112, P_Amp = 192 mm) and low, deserts and xeric shrublands (region 121-123, P_Amp = 48 mm) and cold regions into high Tree, temperate forest (region 141, Tree = 48.4%) and low, Mediterranean forest and temperate grasslands (region 131-132, Tree = 18.3%).

Across-Region Variations in Fire Activity
The overall relationship between BA and the environmental drivers was similar in the two continents: the highest BA was detected in regions with strong precipitation seasonality (i.e., P_Amp, and P_SI), while regions with high Tree had the lowest BA (Figures 4 and 5). Similarly, the results of the GAMs (Figures S7 and S8 in Supporting Information S1) showed that both rainfall seasonality variables had a positive relationship with BA, albeit not for the entire range of the two variables. In particular, BA increases with P_Amp until it reaches about 200 mm, and with P_SI > 0.5.
In Africa, BA initially increases with Tree and P, reaching a maximum at approximately Tree = 20% or P = 1,200 mm, and then decreases, confirming the intermediate fire-productivity hypothesis (Pausas & Ribeiro, 2013;van der Werf et al., 2008) and would indicate that in xeric environments P-which drives the availability of fuel-is the main fire limiting factor (i.e., water-limited systems). In wetter environments, however, fuel is not a limiting factor, but precipitation seasonality (measured by P_Amp), with the presence of distinct wet and dry seasons, is (i.e., energy-limited systems). High P and low P_Amp create favorable conditions for tree growth, and at Tree >60% we observed a drastic reduction of BA in Africa ( Figure S7 in Supporting Information S1). We note that although P was not explicitly used in the fire region clustering due to strong collinearity with Tree (Africa) and P_Amp (Australia), changes in fire activity as a function of the P gradients were reported to acknowledge that-on the physical level-P remains one of the key drivers of fire activity in the tropics.
In Australia, the relationship between BA and P shows a similar pattern but significantly less pronounced, with a much wider spread of the confidence interval at high P values (>1500 mm). This pattern reflects the fact that, in Australia, some of the areas with the highest P also have high P_Amp and consequently experience high fire activity (i.e., Northern Australia). In contrast, other areas have high P and low P_Amp, and therefore experience low BA (i.e., the South East of the continent and Tasmania).
Large fires in both continents were observed within areas with similar environmental conditions. In general, FSize tends to increase with P_SI and P_Var and decreases with Tree (Figures 4 and 5) and P > 1000 mm in Australia and P > 1500 mm in Africa ( Figures S7 and S8 in Supporting Information S1). However, for P < 400 mm, FSize has a strong positive relationship with P, which can be explained by the low biomass accumulation rate in dry regions. Infrequent years with higher than usual rainfall can promote large fires in the following year (Bradstock, 2010), specifically in Australia, where P_Var and FSize had a positive linear relationship. The single most important predictor of FSize in Africa was P_SI, whereas Temp explained the most variation in FSize distribution in Australia. A drastic increase in the size of fires was observed when Temp >19°C.
In contrast, the relationship between FRP and the environmental variables was distinctly different in Africa and Australia. In Australia, similar to Murphy et al. (2013), interrelations between FRP and BA were identified. Very intense fires were observed under a variety of climate conditions, with the exception of regions experiencing high P_Amp, where in general high BA and low FRP were observed ( Figure 5). In contrast, this relationship was not observed in Africa: intense fires could be found in most regions except tropical forests (i.e., the region with low P_SI and high Tree). Arguably, the lack of a long, distinct season with conditions conducive to fire in the African tropical forests limits not only BA but also FRP. Additionally, a strong positive relationship was observed between FRP and P at the lower end of the precipitation gradient (P < 1000 mm). P_Var and P_SI showed the opposite pattern; while P_Var negatively affects FRP, P_SI promoted intense fires in Africa.
It should be noticed that, while the overall relationships between fire metrics and environmental variables were similar, the domain of the environmental space of fire regions differs between the two continents, as evident when comparing the range of the x-axes of the plots in Figure S7 in Supporting Information S1 with those of Figure S8 in Supporting Information S1. Likewise, fire regions were significantly different in terms of summary values of at least one of the fire metrics, as confirmed by Wilcoxon signed-rank tests ( Table 2). As an illustrative example, we compared fire activity and environmental space of Tropical Savannas and Deserts and Xeric Shrublands, that is, two biomes with significant fire activity both in Africa and in Australia. Australian Tropical Savannas support lower tree cover (Tree = 8.5% vs. Tree = 20% in Africa), receive a lower annual rainfall (P = 976 and 1,208 mm in Australia and Africa, respectively) and experience higher precipitation variability (P_Var = 0.26 and 0.11 respectively). And while Tropical Savannas in the two continents experience similarly very large annual burned area (BA = 25% in regions 321-331 and 111-112), FRP and especially FSize were significantly higher  in Australia (Table 2). Differences in FRP and FSize were even more drastic between regions of the Desert and Xeric Shrublands biome (342 and 121-131), where fire activity in Australia is up to three times higher in terms of FSize and up to 1.5 times in terms of FRP. As in Tropical Savannas, Deserts and Xeric Shrublands in Australia occupy a portion of the environmental space with lower P (324 mm), higher P_Var (0.32), and higher Temp (22.7°C) compared to Africa (355 mm, 0.27, and 21.4°C, respectively).

Within-Region Anthropogenic Influence on Fire Activity
While the primary goal of this work was to delineate broad fire regions based on similar environmental conditions that shape fire activity, a simplistic analysis of human factors was included to demonstrate the potential use of the defined fire regions to study more local drivers of fire activity.
Overall, the analysis of the effect of anthropogenic pressure on fire activity metrics within the fire regions (Figure 2c, Table 3) confirmed the findings of previous studies: except for tropical forests where fire metrics increased with human pressure-consistent with fires in this region being tied with deforestation (Hoffmann et al., 2003)-we observed a negative relationship between fire activity and anthropogenic pressure (Archibald, 2016;Le Page et al., 2010;van Langevelde et al., 2003). Interestingly, our analysis showed that humans could exacerbate or minimize the variation in fire activity between two continents. For example, the drastic difference in FSize between the Deserts and Xeric Shrublands regions in Africa and Australia could be attributed to differences in the magnitude of anthropogenic pressure between the two continents since areas with a negligible level of human impact within the two fire regions show similar FSize (regions 342 and 122). In contrast, the difference in average BA in Tropical Savannas regions between two continents, which are overall small, substantially increased when   only low human pressure areas were considered. In other words, in the absence of humans, Tropical Savannas in Africa support more frequent fires than in Australia. It is also important to state that while, based on previous research, it was expected that humans drastically alter fire activity, our analysis showed that the magnitude of human influences varied substantially across biomes. For instance, both regions 132 (mostly Mediterranean and Temperate Forests in Southwestern and Southeastern Australia) and 341 (mostly Montane Grassland and Tropical Savannas, both North and South of the Equator in Africa) encompass mostly human-managed lands and therefore experience a very high level of anthropogenic pressure ( Figure S2 in Supporting Information S1). The high proportion of croplands and high livestock density has, however, a different impact on fire metrics in the two regions. High levels of human pressure substantially decreased BA in region 341 but had a low impact on Fsize and FRP, whereas, in region 132, high human pressure had a more severe impact on FSize and FRP than BA.

Correspondence Between Fire Regions, Biomes and Climate Zones
The spatial distribution of the fire regions (Figure 3) shares some similarities with the terrestrial biomes defined by Olson et al. (2001) and the Köpper-Geiger climate classification (KG, Kottek et al., 2006), which is not surprising since climate data are one of the main inputs of all three maps. However, biomes were delineated based on similar climate and vegetation cover that support the distribution of selected groups of plants and animals. In contrast, fire regions were defined as areas homogeneous in terms of environmental conditions that promote fuel accumulation and desiccation (i.e., fire potential). As a result, a number of dissimilarities can be recognized  between the spatial distribution of biomes and fire regions. First, similar to the KG map, the tropical savannas biome was split into several regions to represent the full spectrum of biophysical conditions and fire activity. While the P in Australia has a distinct Northeast to Southwest gradient (evident in the KG map), the gradient of BA follows the North-South direction and FSize the West-East one. The patterns are captured by the fire regions: the Tropical Savannas biome is split into a Northern tropical region (112) with the highest BA, a Northern arid region (111) with high BA and high FSize, and an Eastern region (123) with drastically lower BA. The Tropical Savannas biome in Africa was split into two regions North of the equator (323, 331) and two South of the equator (321, 322). This split is driven by the spatial pattern of Tree, which negatively affects BA, as reported above. Additionally, while the Sahel and the Horn of Africa also belong to the Tropical Savannas biome, they experience different climates. Consequently, they were separated into region 332 (Sahel) with the lowest Tree in Africa, lower BA and higher FSize compared to other tropical savannas regions (321-331), and region 351 (Horn of Africa), which had the lowest P_SI in both continents and the lowest BA in Africa.
In Australia, the Deserts and Xeric Shrublands biome is also assigned to a single climatic zone by KG, but it experiences a large gradient of BA and FSize. This biome was split into three fire regions, roughly along the North-South direction, corresponding to progressively higher FSize and lower BA. The Northern region (121) Figure S2 in Supporting Information S1), which showed to substantially decrease the rate of fire spread (Archibald, 2016;Archibald et al., 2012;Le Page et al., 2010;Mouillot & Field, 2005). Therefore, both biomes were assigned to one fire region (132) except for Tasmania and a small southwest part of the mainland with a temperate, dry and warm summer climate (region 141) that showed higher FSize and FRP than in region 132.

Comparison of the Fire Regions to Published Fire Regimes Maps
Despite the importance of fire regimes delineation, only two global fire regimes maps are currently available in the peer revied literature. Chuvieco et al. (2008) defined eight fire regimes based on levels of fire frequency and its inter-and intra-annual variability, whereas Archibald et al. (2013) used the clustering of BA, FSize, and FRP to define five key global fire regimes. Similar to our results, both studies distinguished some variation in fire activity between the Tropical Savannas in Africa and Australia ( Figure S12 in Supporting Information S1). In the Desert and Xeric Shrublands biome, however, the two global fire regimes map were not able to distinguish drastic differences in fire activity between Africa and Australia, which we found to vary significantly in terms of FSize and FRP. Additionally, Archibald et al. (2013) identified similar fire regimes in temperate grasslands in Australia and tropical forests in Africa, while our results indicate that those two regions differ significantly, especially in terms of FRP.
In short, previous efforts to map global fire regimes often failed to differentiate between regions with significant variation in one or more fire-regime metrics and combined several biomes with different environmental conditions (e.g., fuel-limited and energy-limited) into one fire regime. One possible explanation is the a priori choice of the number of global fire regimes, which might be insufficient to represent the variety of fire activity between and within continents. For example, in contrast to our findings, Archibald et al. (2013) failed to observe any clear relationship between delineated fire regimes and climate variables or biomes.
It is also important to acknowledge that the opposite of our fire regions, previously mentioned maps of fire regimes were restricted by the length of the fire record, which was limited to 6 years for Chuvieco et al. (2008) and 14 years for Archibald et al. (2013). The lack of fire detections in places with moderate to low-frequency fire return intervals impeded assigning a fire regime to many parts of the world, which can be misinterpreted as an absence of fire activity. Therefore, existing fire regime maps based on a few fire metrics cannot be utilized for their main purpose, such as understanding what processes shape current fire activity.
To incorporate fire-climate relationships into fire regime mapping, similar to our approach, several previous studies analyzed changes in fire metrics across environmental gradients. However, by selecting one rather than a set of environmental variables, those studies could not recognize the variety of fire regimes across biomes. For example, Hantson et al. (2017) detected a drastic reduction in FSize in the tropics once P reaches 550 m due to an increase in tree cover and concluded that dry tropical regions experience larger, hotter, and more rare fires compared to wet tropics. While overall, it is a fair statement, we showed that FSize and FRP are not always higher in dry tropics compared to the wet (e.g., the Horn of Africa vs. tropical savannas), and not all wet tropical regions experience higher BA than dry regions (e.g., tropical forest vs. Sahel or desert and xeric shrublands in Africa). Additionally, based on our GAMs, we found that P threshold for FSize reduction was much higher, 1000 mm in Australia and 1500 mm in Africa ( Figures S7 and S8 in Supporting Information S1).
Recognizing a variety of fire regimes within one continent and their strong connection with moisture, Russell-Smith et al. (2007) defined 10 regions in Australia using seasonal rainfall distribution. We similarly found that P_Amp and P_SI were important predictors of fire activity in Australia, especially for BA, and a certain resemblance between our fire regions and Russell-Smith's rainfall classes can be recognized. However, by including other environmental variables in our classification, we were able to define regions statistically different not only in terms of BA but also FSize and FRP. We found that FSize and FRP do not follow a precipitation gradient as closely as BA; therefore, rainfall seasonality alone cannot explain the complexity of the fire phenomenon and account for multiple processes that drive fire (Bradstock, 2010).
We finish our comparison by acknowledging the work of Murphy et al. (2013) that described in great detail broad fire regimes in Australia also using environmental controls of fire activity, such as major fuel types and climatic domains, as a baseline. The comparison between the fire regions and Murphy's map revealed that our regions somewhat captured the overall spatial pattern of fire frequency and intensity in Australia. However, instead of using statistical analysis to delineate fire regions, Murphy et al. (2013) used experts' opinions and local literature. While we do not dispute that local knowledge can enhance our understanding of the spatial distribution of historical and modern fire activity, it can also bring a certain degree of subjectivity, bias, and a lack of reproducibility to the method.

Limitations of the Study
Some limitations must be considered in interpreting the results of the proposed approach. First, while the potential drivers of fire activity used in this study were identified through a thorough literature review, not all potential fire climatic and environmental drivers were considered. Other variables such as wind and fire-weather indices that can reflect short-term dry spells can potentially explain some discrepancies between climate domains and fire metrics (Abatzoglou & Kolden, 2013;Archibald et al., 2009). Moreover, while elevation, aspect, slope, soil properties, fuel types, ignitions, and fire management practices can have a profound effect on different fire characteristics (Archibald et al., 2009;Russell-Smith et al., 2007;Stevens et al., 2017), their impact might be too local for the scope of this paper. It is also worth mentioning that while only a few variables (6 in Africa, 5 in Australia) were used in the PCA and clustering analysis, strong collinearity between many of the environmental variables might preclude the identification of the single most important driver-caution is therefore recommended if the results of this study were to be used in physically-based fire models.
Second, this study is limited by the quality of currently available remotely sensed data. In particular, the MODIS fire products do not capture small, low-intensity fires Kumar et al., 2011), which could, in some fire regions, introduce biases in the fire activity metrics. This issue is particularly relevant in frequently burned environments and within highly managed lands (Laris, 2005), thus fire-data limitations could affect the results of the anthropogenic pressure analysis, spuriously restricting it to bigger fires. Recent studies by Roy et al. (2019) and Ramo et al. (2021) demonstrated the severe underestimation of small fires in Africa by the MCD64A1 product, especially within cropland and savannas. Additionally, coarse resolution land cover products fail to detect small agricultural fields, which are prevalent in Africa (Inglada et al., 2015;Midekisa et al., 2017) and can be used as fire breaks, increasing landscape fragmentation. Meanwhile, precipitation datasets are less accurate in regions with inadequate rain gauge network (Dinku et al., 2014;Toté et al., 2015).
Third, while remotely sensed data is the only source for global analysis to capture spatial and temporal patterns of fire activity (Benali et al., 2017), there is no agreement on the minimum fire record length required to adequately capture the inter-annual variability in fire activity, especially in places with moderate to low-frequency fire return interval (Bowman et al., 2009;Daniau et al., 2012;Krawchuk & Moritz, 2014). And only a limited number of fire metrics can be currently estimated from remote-sensed data. Some essential components of the fire regime concept, for example, fire type and severity, were omitted in this study.
Fourth, some aspects of the statistical analysis add uncertainty. For instance, the selection of the number of clusters through inspection of the dendrogram is rather arbitrary; a lower or higher number of fire regions might more preferable depending on the scale of analysis and on the intended use of the ecoregions. We note, however, that the hierarchical clustering provides a straightforward aggregation of the fire regions: using the first two digits of the fire region ID results in a rigorously defined coarser map. Additionally, there are issues with the separability of some of the clusters: Figure 2 shows a certain degree of overlap between clusters and a number of cells close to the decision boundary.
Finally, the aggregation of 0.5° cells into ecoregions also resulted in the loss of some local-scale features. This is evident in Ethiopia, a highly diverse country that encompasses 11 different ecoregions, where cells belonging to 7 different fire regions were primarily merged into two major regions (341, 351) with a few cells merged into regions 311, 331, and 332.
These limitations notwithstanding, the methods presented in this paper are general ones. We would advocate that, for local studies, smaller than a country level, a user can adopt the proposed approach to define more refine fire regions using data from local weather stations, field data, and higher resolution imagery to avoid some of the biases and limitations of coarser resolution data.

Conclusions
This study proposed a new approach for mapping the diversity of potential fire activity at the continental level, based on key processes that control fuel accumulation, desiccation, and fire propagation. One of the advantages of the proposed method is that the fire regions were identified through the statistical analysis of climate and environmental variables reflecting broad biophysical processes rather than using observed fire activity metrics as input data -unlike previous studies that have used a limited set of metrics estimated over short time periods (e.g., Archibald et al., 2013;Chuvieco et al., 2008;Le Page et al., 2008). Furthermore, the analysis was conducted separately for each continent, without the implicit assumption that the same "fire regime" exists in different continents. Reflecting the heterogeneity of fire characteristics within biomes and between continents, the proposed methods resulted in 9 and 8 regions in Africa and Australia, with distinct climate, vegetation, and fire activity.
Fire activity metrics computed from the 2002-2019 record of the global MODIS satellite fire record were used to characterize the fire activity level of each region, thus evaluating the effectiveness of the proposed fire region map. We found that different environmental variables act as key limiting factors for fire activity in Africa and Australia. Consistently with the current fire ecology literature Lehmann et al., 2014), temperature was found to be a more important predictor of fire activity (especially for the average annual burned area and the fire size) in Australia than in Africa, whereas precipitation and the highly correlated tree cover were the primary predictors of BA and fire intensity (FRP) in Africa.
We believe that delineated fire regions can be used not only as a spatial unit for calculating fire statistics but also to study how changes in fire drivers could reshape fire activity in the future and to analyze anthropogenic drivers of fire within regions homogeneous in terms of climate controls. In this study, we demonstrated that humans' potential to transform a different aspect of fire activity varied within climate space; therefore, any attempt to generalize the human-fire relationship at the continental or global scale will severely oversimplify the role of humans in shaping modern fire activity. Ultimately, while our understanding of broad-scale fire-climate relationships is improving, more research is needed to evaluate the impacts of socio-economic factors on those relationships, particularly in tropics where climate-or human-related changes in fire activity could also drive long-term changes in vegetation composition since fire is one of the main controls of grass-forest coexistence (Lehmann et al., 2014;Sankaran et al., 2005).

Data Availability Statement
The observational datasets used in this study are freely available.