2.1 Remote sensing data

For our analysis, we rescaled all data to 0.25° spatial resolution by calculating the mean value within each cell, with the exception of land cover type where we used the dominant cover type within each 0.25° grid cell.

2.1.3 Burned area drivers

Monthly rainfall data were obtained from the Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) dataset (Funk et al., 2015) for the extended study period between 2002 and 2016. We used rainfall data to calculate mean annual rainfall (MAR, in mm/year, Figure 1b) over the calendar year and estimate the fuel moisture and fuel build-up effects on interannual variability in burned area. We defined the fuel moisture effect as the accumulated rainfall during the 6 months prior to the end of the burning season. We assumed that rainfall occurring during, or just before the burning season determines the probability of ignition and fire spread. The fuel build-up effect was defined as the accumulated rainfall during 24 months prior to the end of the burning season, as previous rainfall is an important control on the amount of biomass produced. We selected the 6- and 24-month cut-offs as, on average, the strongest negative response in fuel moisture limited landscapes was found around 6–7 months of antecedent rainfall (Supporting Information Figure S1), while across fuel build-up limited landscapes the antecedent rainfall over two wet seasons (24 months) had a slightly higher explanatory power than over a single wet season (12 months; Supporting Information Figure S1).

We considered three explanatory variables for our initial analysis of the drivers of observed spatial patterns in fuel build-up and fuel moisture limited fire regimes. First, we focus on spatial differences in dry season duration. Following Hulme and Viner (1998), we define the dry season duration (in months) as the average number of months with rainfall below 50 mm/month during the 2002–2016 calendar years (Supporting Information Figure S2b). This intermediate (50 mm/month) rainfall threshold assures reasonable sensitivity to dry season duration across both arid and more humid tropical environments. Second, to investigate how humans affect fire occurrence and climate–fire interactions, we used the Wildlife and Conservation Society (WCS) human influence index (HII, Supporting Information Figure S2a; WCS & CIESIN, 2005), a measure, varying between 0 and 64 (for no human and maximum influence, respectively), of the direct human influence on ecosystems based on eight different measures of human presence: population density (people per km²), land cover type, and a measure of the presence of railroads, major roads, navigable rivers, coastlines, nighttime stable lights, and urban polygons. Third, because vegetation structure can affect fire activity and varies across continents (Lasslop, Moeller, D'Onofrio, Hantson, & Kloster, 2018), we also considered tree cover as an explanatory variable for observed patterns of fuel and moisture limited fire regimes. Tree cover data were obtained from MODIS vegetation continuous fields (MOD44B collection 5, Supporting Information Figure S2c) for 2010 (DiMiceli et al., 2011). Because in our definition of tropical savannas and grasslands we already excluded areas with tree cover > 40%, this variable ranged from 0% to 40%.

2.2 Methods

2.2.2 Drivers of burned area response

We used two different approaches to explore the drivers of spatial differences in the relationship between annual burned area and antecedent rainfall. First, to keep annual rainfall constant, we binned all grid cells based on 200-mm/year MAR increments; within each rainfall bin, we further subdivided the grid cells based on bins of dry season duration (DS; increments of months with rainfall below 50 mm), human influence index (HII; increments of 5 units of HII, HII ranged from 0 to 40 across the study area) and tree cover (TC; 5% increments from 0% to 40%). Based on this subdivision along rainfall gradients, we explored how DS, HII and TC modified patterns of (a) mean annual burned area, (b) interannual variability in burned area (measured as the coefficient of variation), and (c) the correlation coefficient between antecedent rainfall and annual BA.

Second, we compared two multiple linear regression models to understand how the fuel moisture effect (Rain6, 6 months of accumulated rainfall) and the fuel build-up effect (Rain24, 24 months of accumulated rainfall) influenced the interannual variability of BA (Equation 1) and investigate if BA’s response to rainfall varied at continental scales (Equation 2). In order to increase model sensitivity to temporal variability in burned area, we used burned area anomalies rather than absolute burned area time series:

(1)

(2)

where BA_i,j anomaly is the burned area anomaly for each pixel (i) and year (j) calculated as BA_i,j – mean BA_i, β parameters represent the slope of the linear regression between the BA anomaly and the explanatory variables (Rain6_i,j and Rain24_i,j), α is the intercept and ε the residual error term. The BA anomaly includes both negative and positive values, where negative values indicate that the BA for the year j was lower than the mean BA and positive values indicate that annual BA was higher than the mean. Thus, β₁ and β₂ indicate the rate of BA change per unit of accumulated rainfall (% year/mm).

Next, we explored how other variables, including MAR, DS, HII and/or TC modify the influence of antecedent rainfall on burned area anomalies by analysing how the β₁ and β₂ values changed when introducing each driver into the model. In addition, when including a new variable, we compared the model with and without the effect in question, using an ANOVA likelihood-ratio test and Akaike’s information criterion (AIC) to confirm the selection of the best model (Burnham & Anderson, 2004). Finally, we constructed the same models, but now based on full burned area time series instead of anomalies. This analysis helped to understand how each variable contributes to both spatial and temporal patterns of biome wide burned area. All analyses were done using the “raster” and “rgdal” packages in R version 2.5.1 (R Core Team, 2018).

3.1 Burned area response to antecedent rainfall

We observed the strongest correlations between antecedent rainfall and annual BA in frequently burning savannas and grasslands across the tropics (Figure 1). Here we considered grid cells with a negative correlation between burned area and rainfall (6-month lead time) to be fuel moisture limited and grid cells with a positive correlation (24-month lead time) to be fuel build-up limited. Interestingly, we found that savannas with fuel build-up limited fire regimes (41.4%) in more arid regions covered less area than savannas with fuel moisture limited fire regimes (58.6%, Figure 1b,e) in more humid systems. MAR varied widely across tropical savannas on the three continents, resulting in predominantly fuel moisture limited fire regimes across the relatively humid savannas of South America, and fuel build-up limited fire regimes across Australian savannas that were more arid on average (Figure 1b–e). Africa showed a mix of fuel build-up and moisture limited savannas across a large rainfall gradient. For example, we observed strong positive correlations for arid regions (e.g., Namibia, Botswana and Zimbabwe) and strong negative correlations for humid regions (e.g., the north of Mozambique and the south of Tanzania and the Democratic Republic of Congo; Figure 1e).

3.2 Continental differences in the switch from fuel build-up to fuel moisture limitation

African tropical savanna and grassland fire regimes switched from a predominantly positive (fuel build-up effect) to negative (fuel moisture effect) response to antecedent rainfall around 800 mm annual rainfall (Figure 2), while fire regimes in South America switched around 500 mm/year, and in Australia around 1,000 mm/year (Figure 2). For all three continents, MAR bins that contained a low number of grid cells often showed a more variable response (cf. Figure 2 and Supporting Information Figure S3). We also observed large spatial variability in burned area–rainfall responses (Figures 1 and 2), indicating that the switch from fuel build-up to fuel moisture limited fire regimes occurred gradually. On each continent, there was a transition zone in MAR levels rather than a clear threshold, where the strength of the dominant correlation weakened before switching to a different dominant driver. Only 24% of South American savannas were fuel build-up limited against 61% of Australian savannas and 47% of African savannas.

3.3 Drivers of continental differences

In addition to MAR, we explored how rainfall seasonality influences median annual BA and interannual variability in BA, both important indicators of the strength of fire–climate interactions. Globally, longer dry season durations tended to increase median annual BA, particularly in intermediate productivity savannas and grasslands (MAR between 900 and 1,500 mm, Figure 3a). Interestingly, when dry season length exceeded 9 months, annual burned area typically declined again, likely because very short growing seasons may limit ecosystem productivity and thus fuel availability. In addition to burned area, we also analysed its coefficient of variation, because we hypothesize that the strength of the burned area response to antecedent rainfall partly depends on the variability of both variables. For bins of comparable rainfall and dry season duration, Australia showed the lowest coefficients of variation, potentially weakening correlation coefficients between antecedent rainfall and burned area, seen as a more variable response of positive and negative correlations (Figure 3b,c). In contrast, large on average coefficients of variation across South America may be responsible for the relatively strong negative correlation observed across productive savannas. We also observed a reduction in the coefficient of variation in areas with a high fraction of annual burned area (Figure 3b), possibly reducing the strength of the correlation between annual burned area and fuel conditions (Figure 3c). Although dry season duration clearly affected burned area and its variability, patterns were not uniform, suggesting other factors also played a role.

Human impact strongly reduced burned area across continents (Figure 4a), while Australian savannas and grasslands were generally characterized by low human impact values (HII < 10) and African and South American savannas were characterized by higher impact values (HII = 10–25; Figure 4b). The coefficient of variation was clearly reduced in natural areas with a large fraction of burned area and low human influence (Figure 4b). Despite the large impact of HII on absolute burned area, impacts on the interannual variability were more limited and complex. Globally, a small decline in the strength of the burned area response to rainfall variability was observed with decreasing HII and increasing burned area in the peak biomass burning regions (MAR ranging from 900 to 1,500 mm/year; Figure 4c). In contrast, at continental scales sometimes the opposite pattern was observed. For example, in productive savannas (MAR 1,300–2,100 mm/year) of South America the negative correlation between antecedent rainfall and burned area strengthened with decreasing HII. The global pattern in the response of BA to the fuel build-up and fuel moisture effects was mainly determined by South America and Africa, with a dominant negative response in savannas with MAR above 900 mm/year and positive response in savannas with MAR below 900 mm/year, independently of the HII (Figure 4c).

Vegetation structure also influenced biome wide patterns of burned area and the strength and sign of correlation coefficients between antecedent rainfall and burned area (Figure 5). As expected, we observed that higher tree cover was often associated with reduced burned area, particularly in the humid tropics (Figure 5a). In productive savannas (MAR ranging from 900 to 2,000 mm/year), the fuel moisture effect tended to strengthen with increasing tree cover, although relationships were often weak (Figure 5c). In fuel limited ecosystems of Australia, there was a weak increase in the strength of the fuel build-up effect with tree cover, opposite to the global pattern, where the strength of the fuel build-up effect weakened with increasing tree cover. As noted earlier, coefficients of variation varied widely across continents, possibly strengthening or weakening regional correlation coefficients. In contrast to Africa and Australia, South America showed high coefficients of variation for savannas of intermediate productivity, likely contributing to the exceptionally strong moisture limitation on regional burned area.

Despite the clear biome wide patterns of fuel moisture and fuel build-up limited fire regimes, we could not establish a single global model to explain the interannual variability in BA based on fuel build-up effect and fuel moisture effect alone (Table 1). Here we used multiple linear regression models to test the effect of the antecedent rainfall on BA anomalies (Table 1) per pixel across the time series, and the spatial and temporal pattern in BA time series (Supporting Information Table S1), across tropical savannas and grasslands. We expect that the model based on BA anomalies is better able to capture interannual variability, while the second model captures both temporal variability and spatial patterns of burned area. The global model of BA anomalies that included only the fuel build-up effect and fuel moisture effect as explanatory variables explained less than 1% of BA variation (Table 1), while the same model for absolute BA explained 3% of the variance (Supporting Information Table S1). Surprisingly, BA did not vary as expected when an interaction term between MAR and fuel build-up effect and fuel moisture effect was added to the model and its performance did not improve. In contrast, the inclusion of “continent” as an interaction term with the 6- and 24-month accumulated rainfall increased the percent of explained variance and reduced AIC for both BA (from 3.6 to 19%, Supporting Information Table S1) and its interannual variability (from 0.0019 to 0.0029%, Table 1). The two models supported different slopes between the fuel moisture effect and fuel build-up effect and BA across continents (p < .001), confirming continental scale differences in burned area–rainfall response (Figure 2, Table 1 and Supporting Information Table S1).

The models with the highest explanatory power (lowest AIC) explained 0.4% of the variance in the interannual variability of BA (Table 1) and 29% of spatial occurrence (Supporting Information Table S1). These models included tree cover, dry season duration and HII all in interaction with Continent as additional factors to fuel build-up effect and fuel moisture effect. The response of BA anomalies to the fuel-moisture effect and the fuel build-up effect was strongest in Australia (β₁ = −0.0036 and β₂ = 0.0019%/mm, respectively, Table 1) and weakest in Africa (β₁ = −0.0012 and β₂ = 0.00035%/mm respectively, Table 1). Statistical analysis confirmed the expected response, with negative slope coefficients for fuel moisture effect and positive coefficients for fuel build-up effect (Table 1). The inclusion of tree cover, dry season duration or HII in the models modified the slopes of fuel build-up effect across all the three continents, while the slopes for fuel moisture effect remained more similar (Table 1). When we included all three variables in the model, we detected a slight decrease in the slope of fuel build-up effect for African (from 0.00041 to 0.00035%/mm) and South American savannas (from 0.00058 to 0.00046%/mm) and a larger increase for Australian savannas (from 0.00088 to 0.0019%/mm, Table 1). The inclusion of these three factors also modified the intercept sign from negative to positive indicating a positive anomaly (BA > mean BA) when these factors are zero. When considering the inclusion of each of these explanatory variables (DS, HII and TC) separately, the inclusion of dry season length had the strongest effect on the response of BA to the fuel build-up effect; in Africa β₂ decreased from 0.00041 to 0.00023%/mm and the smallest effect was observed in South America from 0.00058 to 0.00051%/mm. In contrast, the inclusion of HII had the strongest effect on the response of BA to the fuel build-up effect in Australia, with β₂ increasing from 0.00088 to 0.0011%/mm, while in Africa and South America we observed a slight decrease (from 0.00041 to 0.00040%/mm and from 0.00058 to 0.00054%/mm, respectively). The inclusion of tree cover had the strongest effect on the response of BA to the fuel build-up effect for South America, with β₂ increasing from 0.00058 to 0.010%/mm. When we analysed the absolute BA (Supporting Information Table S1), HII coefficients were also negative for the three continents, in line with lower BA in human dominated landscapes (Andela et al., 2017; Archibald, Staver, & Levin, 2012), with the highest decrease in BA variation when human influence increased in Australia (β₄ = −3.10%), and a similar lower variation observed in Africa and South America (β₄ = −1.29 and −1.10% respectively, Supporting Information Table S1).

4.1 Fire–climate threshold

Here we explore the extent of fuel build-up and fuel moisture limited fire regimes across tropical savannas based on a per-pixel temporal correlation between burned area and antecedent rainfall. Savanna fire–climate interactions changed along gradients of mean annual precipitation, with burned area in xeric savannas being primarily limited by fuel build-up and in mesic savannas by fuel moisture (Figure 1; Kahiu & Hanan, 2018; Krawchuk & Moritz, 2011). In line with previous work, we find that fire activity in humid savannas and grasslands primarily responds to drought conditions during the fire season (Alvarado, Fornazari, Cóstola, Morellato, & Silva, 2017; Archibald et al., 2010; Lehsten, Harmand, Palumbo, & Arneth, 2010) similar to tropical rain forests (Aragão et al., 2008). We find that fuel moisture was the dominant control on fire activity over 58.6% of tropical savannas and grasslands. These systems currently account for 59.1% of the tropical area burned, and the remaining 40.9% is in systems that are fuel build-up limited.

Striking differences were observed across continents, with large areas of fuel build-up limited fire regimes occurring across more arid grasslands and savannas of southern Africa and northern Australia, and a near-absence of fuel build-up limited systems in tropical South America (Figures 1 and 2). Fuel moisture formed the key control on burned area across South America's savannas, except for more arid grasslands along the eastern edge of the Brazilian Cerrado. Burned area in arid regions of Africa and Australia responded strongly to antecedent rainfall, highlighting the importance of fuel build-up and connectivity in these regions (Archibald et al., 2010; Krawchuk & Moritz, 2011; Price et al., 2015; Whitlock et al., 2010). Continental scale differences were partly driven by differences in climate, for example, the extent of semi-arid and arid savannas with MAR < 1,000 mm/year was largest across Africa and Australia, resulting in an overall larger fraction of ecosystems where fire occurrence was limited by fuel build-up (Figure 1; Archibald et al., 2010). However, savanna fire regimes also switched from being dominantly fuel build-up limited to fuel moisture limited at different thresholds, around 500 mm/year in South America, 800 mm/year in Africa and 1,000 mm/year in Australia (Figure 2). Together, these two factors resulted in continental scale differences in fire regimes, and fire activity was limited by fuel build-up in only 24% of South American savannas, against 47% of African savannas and 61% of Australian savannas. Interestingly, these continental differences in fire regimes are in line with previous work showing similar differences in controls on savanna distribution and structure (Lehmann et al., 2014; Lehmann, Archibald, Hoffmann, & Bond, 2011). In the transition zones, where fire regimes switched from being predominantly fuel-build up limited to fuel moisture limited, the relationship between burned area and fuel dryness or fuel availability was often weak, and likely further modified by other climatic, ecological and anthropogenic factors influencing fuel conditions.

4.2 Drivers of fire response

Seasonal rainfall distribution varied considerably across continents and had a strong effect on annual burned area (Figure 3a). Previous analyses have shown that rainfall amount during the dry and wet seasons contributes to explaining the spatial patterns of tropical fire activity (Bowman et al., 2014; Chen et al., 2017; van der Werf et al., 2008), and that climate seasonality can explain observed differences in fire activity across regions with similar MAR (Saha, Scanlon, & D'Odorico, 2019). We found that a minimum dry season duration of 6 to 8 months was required for frequent fires to occur in productive and humid savannas, but we only detected a weak relationship between annual burned area and increasing dry season lengths longer than 6 months. A possible explanation for this weak relationship could be that dry season duration longer than 6 months may limit herbaceous productivity by shortening the growing season in spite of MAR. In addition, our results suggest that observed differences in rainfall seasonality may also modify the response of burned area to antecedent rainfall across different regions (Figure 3b,c). Although the relatively long and pronounced African dry season is one of the factors contributing to high fire frequencies across the continent (Archibald et al., 2009), African savannas were characterized by relatively low variability in burned area. In contrast, South American savannas were characterized by lower fire frequencies, but showed higher interannual variability in burned area driven by climate anomalies (cf. Figure 3b,c; Alvarado et al., 2017; Chen et al., 2017; Mataveli et al., 2018).

Several analyses have shown that human land management, and therefore population density, has a significant impact on global burned area (Bistinas et al., 2013). In line with these findings, we found that higher human influence significantly reduced burned area across continents, with larger consequences for more densely populated continents like Africa and South America compared to Australia (Figure 4a and Supporting Information Table S1; Andela et al., 2017; Archibald et al., 2012). Previous work has also shown that human land management may reduce the sensitivity of fire regimes to climate extremes (Bird et al., 2016). We found that the observed biogeographic differences in fire responses to antecedent rainfall could be related to human land management to some extent, but this factor alone could not explain the differences observed across continents (Figure 4c). In general, areas with large annual mean burned area and low population densities showed a relatively strong burned area response to rainfall variability. Nevertheless, this pattern did not hold everywhere, and particularly in savannas of intermediate productivity we observed an overall increase in the strength of the fuel-moisture effect on burned area in human dominated landscapes.

Continental scale differences in tree cover also explained part of the observed differences in fire–climate interactions. Previous work has shown that tree cover may limit fire activity in savannas (Archibald et al., 2009), although these effects may be partly masked out in our study, which focuses on more open cover types with tree cover lower than or equal to 40%. Across areas with fuel moisture limited fire regimes, we observed a slight increase in the strength of the responses of BA to the antecedent rainfall with the increase of tree cover percentage at similar MAR. While all three variables (DS, HII and TC) modified the response of burned area to antecedent rainfall, none of these variables could explain the differences in thresholds observed across the continents (Figures 3-5). For example, when controlling for TC, continental scale differences in rainfall thresholds at which savannas switched from fuel build-up to fuel moisture limited fire regimes remained different.

To confirm these findings, we used a range of multiple linear regression models to explore if the continental scale differences could be explained by differences in DS, HII and TC. Allowing the burned area to respond differently to antecedent rainfall across continents caused a considerable model improvement both when modelling absolute burned area (Supporting Information Table S1) and its variability (Table 1). While the introduction of DS, HII and TC as additional explanatory variables further improved model performance, they only marginally affected continental scale differences in burned area response to antecedent rainfall (compare slopes in Table 1). Nevertheless, our linear model explained just 29% of absolute burned area and about 1% of the burned area anomalies even when considering continental scale differences in DS, HII and TC as additional drivers. Improving model representation of fire response to antecedent rainfall therefore remains a topic of future investigation. While we explored the role of dry season duration, it is possible that other indicators of vegetation and fuel conditions, like evapotranspiration, also play an important role (Boer et al., 2016). Similarly, regional differences in herbivory and human fire management, as well as the different composition and structure of grass and tree communities across continents may also be important (Lehmann et al., 2011).

Understanding the distribution of fuel build-up and fuel moisture limited fire regimes is critical for fire management now and in the future, as changes in land management or climate may result in contrasting responses across fuel and moisture limited systems. In contrast to earlier studies that have suggested that fire activity in savannas was mostly limited by fuel availability (Krawchuk & Moritz, 2011; Whitlock et al., 2010), we found that fuel moisture controlled burned area variability in more than half (58.6%) of the tropical savannas and grasslands, accounting for 59.1% of total burned area. Striking differences in burned area response to rainfall variability across continents highlighted that South American savannas were particularly sensitive to fuel moisture conditions, suggesting that rising temperatures may increase fire activity across the continent, and explaining the extraordinarily strong response of fire activity across the continent to drought conditions driven by sea surface temperature anomalies (Chen et al., 2011). In contrast, a reduction of moisture availability would likely decrease burned area over most of Australia, where fire activity was mainly controlled by fuel build-up. In African savannas and grasslands, the area where burned area was primarily controlled by fuel build-up was about equal to the area where fuel moisture conditions were most important. Although we could not conclusively attribute the continental scale differences to a single driver, we found that rainfall seasonality, human land management and tree cover all modified fire–climate interactions regionally through their effects on fuel availability and moisture status. Our work demonstrates that one single “global model” for savanna fires will not be enough to predict future fire regimes and fire regimes across different continents will likely respond differently to the same drivers of global change.