Recent global and regional trends in burned area and their compensating environmental controls

The apparent decline in the global incidence of fire between 1996 and 2015, as measured by satellite-observations of burned area, has been related to socioeconomic and land use changes. However, recent decades have also seen changes in climate and vegetation that influence fire and fire-enabled vegetation models do not reproduce the apparent decline. Given that the satellite-derived burned area datasets are still relatively short (<20 years), this raises questions both about the robustness of the apparent decline and what causes it. We use two global satellite-derived burned area datasets and a data-driven fire model to (1) assess the spatio-temporal robustness of the burned area trends and (2) to relate the trends to underlying changes in temperature, precipitation, human population density and vegetation conditions. Although the satellite datasets and simulation all show a decline in global burned area over ~20 years, the trend is not significant and is strongly affected by the start and end year chosen for trend analysis and the year-to-year variability in burned area. The global and regional trends shown by the two satellite datasets are poorly correlated for the common overlapping period (2001–2015) and the fire model simulates changes in global and regional burned area that lie within the uncertainties of the satellite datasets. The model simulations show that recent increases in temperature would lead to increased burned area but this effect is compensated by increasing wetness or increases in population, both of which lead to declining burned area. Increases in vegetation cover and density associated with recent greening trends lead to increased burned area in fuel-limited regions. Our analyses show that global and regional burned area trends result from the interaction of compensating trends in controls of wildfire at regional scales.


Introduction
for the period 1996-2000 with active fire hotspots from VIRS (Visible and Infrared Scanner) and ATSR (Along-Track Scanning Radiometer). The merging of the burned area estimates from different sensors likely affects the computation of burned area trends (Giglio et al 2013). Other burned area datasets have also been derived from MODIS but using different retrieval algorithms and spatial resolutions (Chuvieco et al 2018, Giglio et al 2018. Comparisons of these datasets show similar spatial patterns of burning but some large differences in global and regional total burned area (Chuvieco et al 2016, 2018, Humber et al 2019. Generally, the relatively short period covered (15-20 years) makes it difficult to achieve a robust quantification of burned area trends.
The apparent recent decline in global burned area has been associated to human activities, specifically population growth, agricultural expansion and land-use changes mostly in northern-hemisphere Africa (Andela et al 2017). However, the incidence of wildfires is influenced by many factors including climate, ignition sources, and vegetation properties (Bowman et al 2009, Krawchuk et al 2009, Moritz et al 2012, Bistinas et al 2014, Knorr et al 2014. The impact of changes in climate on fire are obviously regionally specific, although high temperatures and increasing summer drought have been invoked as the cause of recent extreme fire seasons (Holden et al 2018, Turco et al 2018 and climate change has led to an increase in wildfire season length over large parts of the land area (Jolly et al 2015). On the other hand, large parts of the world such as the African Sahel have experienced widespread increases in vegetation cover and above-ground biomass (Liu et al 2015, Zhu et al 2016, Brandt et al 2017 which affects fuel availability and thus likely fire incidence and burned area. The influence of such changes on recent trends in global burned area has yet to be adequately assessed.
Fire-enabled dynamic global vegetation models (DGVMs) explicitly account for the effects of climate, humans and vegetation on fire occurrence and could potentially be used to assess controls on burned area trends (Hantson et al 2016). However, state-of-the-art DGVMs do not reproduce the observed decline in global burned area: half of the DGVMs from the Fire Model Intercomparion Project (FireMIP) underestimate the apparent decline in global burned area, while the other half show an increase in global burned area (Andela et al 2017). These differences in behaviour suggest that some functional relationships and associated parameterisations are poorly constrained in these models. Indeed, analyses of the FireMIP simulations suggest that, while they represent the climate controls on burned area reasonably well, they underestimate the sensitivity to previousseason plant productivity and have over-simplistic representations of the influence of human activities on burned area (Forkel et al 2019). Empirical fire models (Moritz et al 2012, Bistinas et al 2014, Forkel et al 2017 provide an alternative and arguably better approach to reproduce observed fire dynamics and to quantify the relative importance of climate, vegetation and human factors on temporal changes in burned area.
In this paper, we first assess the robustness of trends in global burned area using two data sets and taking into consideration the impact of the sampling period and time series length. We then use a recently developed empirical fire model to analyse the relative importance of recent climate, human population, and vegetation changes on these trends.

Burned area data
We analysed two global burned area satellite datasets: We simulated burned area using an empirical fire model based on the SOFIA (Satellite Observations for FIre Activity) approach, which estimates monthly burned area from observed time series of land cover, vegetation, climate variables and human population density (Forkel et al 2017).
All burned area and ancillary datasets were either obtained at or aggregated to 0.25°x 0.25°spatial resolution.

Trend analysis
We performed a multi-temporal trend analysis to assess the effects of sampling period, time series length, and year-to-year variability on estimated trends (McKinley et al 2011). We computed trends for different periods within the time series, where the periods are all possible combinations of first and last years with time lengths 8 years within the time series (R package greenbrown version 2.4.3). Trends for each time period were computed from annually aggregated total burned area based on linear quantile regression to the median. Quantile regression is more robust than ordinary least squares regression because it reduces the effect of single extreme years (i.e. years with extreme high burned area) on the estimated trend. Trend slopes were expressed as percentage of change relative to the multiyear mean of the time series. The two-tailed Mann-Kendall trend test was used to estimate the significance of the trends (Mann 1945, Kendall 1975. All analyses were done in the R software (R Core Team 2018) (an overview of the packages and functions used is provided in supplementary table 3).

Predictor data
Previous studies have identified a number of climatic variables, vegetation properties and socio-economic factors that either directly control or are surrogates for known mechanistic controls on fire occurrence and spread ( where T, S, H, and C are the fractional coverage of trees, shrubs, herbaceous vegetation, and croplands per 0.25°g rid cells, respectively. FAPAR and VOD were used as average values over the 12 precedent months to account for the year-to-year variability in vegetation conditions. The controlling functions f(x) are logistic functions: where the parameters mx, x0, and sl vary for each dataset x and by land cover type (Forkel et al 2017). Values of f(x) were trimmed to the range between zero (i.e. complete restriction of fire) and unity (complete allowance), allowing the response functions to be represented as an exponential (supplementary figure 1 is available online at stacks.iop.org/ERC/1/051005/mmedia). The parameters of the controlling functions were estimated by optimizing the simulated monthly burned area from the SOFIA model against the GFED4 dataset for selected grid cells (supplement 1, supplementary table 1).
The SOFIA model simulations cover the period from 1994 (12 months after the beginning of the VOD dataset) until the end of 2011 (end of the FAPAR dataset). There were no harmonized long-term FAPAR and VOD datasets available that would allow us to run SOFIA simulations for the full periods of the GFED4 (1996GFED4 ( -2015 or FireCCI50 (2001FireCCI50 ( -2015 datasets. The SOFIA model can be used to assess the relative contribution of a control x on the dynamics of burned area [BA(x)]: This approach assumes that only a single factor x (e.g. FAPAR) controls burned area while all other factors have no effect on burning. In the following we refer to BA(x) as the 'marginal burned area'. We summed monthly BA and BA(x) time series to annual totals for the computation of trends. BA(x) time series are shown as anomalies relative to the average multiyear (1994-2011) total BA(x) for visual purposes in figure 2.
SOFIA reproduced global spatial patterns of burned area in comparison with the GFED4 and FireCCI50 burned area datasets but simulated global total burned area (295 Mha yr −1 ) that is lower than the range from satellite datasets (supplementary figure 2). However, SOFIA reproduced the year-to-year variability and burned area trends within the uncertainties of the datasets in most regions (supplementary figures 3 and 4).
The satellite data sets have a limited temporal coverage, and the identification of trends is strongly affected by years with abnormally high or low burned area at the beginning and end of the time series. In particular, the last three years 2013-2015 have very low global burned area and are the main cause of the reported decline over 1996/2001-2015 ( figure 1(c)). On the other hand, 2011 had the largest global burned area in the FireCCI50 dataset and the third largest burned area in the GFED4 dataset and this would result in a plateau or a positive trend in burned area if the analysis were terminated in this year. The GFED4 dataset showed mostly significant declining trends (p£0.05, Mann-Kendall trend test) if the first analysis year is between 1998 and 2002 and the last year is 2014 or 2015. The FireCCI50 dataset did not show a significant trend in global burned area for any sampled period (supplementary figure 3(b)). The estimated trend slopes for different time periods are not correlated (r=0.13) between the two datasets ( figure 1(d)).
Most of the overall decline in burned area occurs in northern-hemisphere Africa, southern-hemisphere South America, central Asia, and Australia, whereas large parts of southern-hemisphere Africa had increasing burned area ( figure 1(a), supplementary figure 4). In general, the sign of the regional trends are consistent between the two sets of observations even though the slopes differ (figure 2). However, very few of the observed regional trends are significant.
The SOFIA model also shows a (non-significant) decline in global burned area between 1994 and 2011. Overall, the SOFIA model had mostly declining trends ( figure 1(e)), especially if the first year of the analysis is before 2000 (supplementary figure 3(c)). In the period of overlap between the satellite observations and the simulation (2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011), there is agreement between the SOFIA model and the FireCCI50 dataset that there is no significant trend (or a slight positive tendency) in global burned area, whereas the GFED4 dataset shows a significant negative trend ( figure 1(e), left panel). For the full period of each dataset and the common overlapping period, trends in simulated burned area from the SOFIA model are within the uncertainties of the satellite datasets.
The satellite datasets and the SOFIA model all show significant declining burned area in northern hemisphere Africa (figures 2(a), (f)). This region dominates the global decline in burned area. Southernhemisphere Africa showed a tendency towards increasing burned area (figures 2(b), (g)). The FireCCI50 dataset and the SOFIA model had significant declining burned area in northern hemisphere South America ( figure 2(h)). The satellite datasets suggest a non-significant decline and the SOFIA model a stable burned area in southern hemisphere South America (figures 2(d), (i)). The satellite datasets show a significant decline in burned area in central Asia and Australia whereas the SOFIA model shows no trend in these regions (supplementary figure 4). In the following, controls on regional burned area trends will be assessed only for northern-and southern-hemisphere Africa and South America where trends are registered in satellite datasets and reproduced by SOFIA. The satellite datasets and the SOFIA model show a decline in global burned area overall but differences between the datasets prevent an accurate quantification of this decline.

Controls on regional burned area trends
The SOFIA simulations suggest that changes in climate, vegetation and human population had regionally diverse effects on marginal burned area and together shape the overall regional burned area trends (figure 2).
In northern-hemisphere Africa, the decline in burned area was driven by significant declines in the marginal burned area associated with population density, number of wet days and precedent FAPAR while maximum temperature caused a non-significant increase in the marginal burned area ( figure 2(a)). According to the SOFIA model, increases in population density, wet days and FAPAR in tree-covered regions all cause a decline in burned area ( supplementary figures 1(a), (c) and (d)). Thus, the positive trends in FAPAR ('greening') with a simultaneous increase in tree cover at the expense of herbaceous and shrub cover in northern-hemisphere Africa (supplementary figures 5 and 6), coupled with an increase in the number of wet days per month, and with increasing population density all contribute to reducing the amount of fire.
In southern-hemisphere Africa, however, the controls on burned area counteracted one another, resulting in no trend in simulated burned area. The marginal burned area declined because of increases in population density and number of wet days but increases in maximum temperature and VOD contributed to increasing marginal burned area ( figure 2(b)). Strong positive trends in VOD occurred especially in regions dominated by herbaceous vegetation (supplementary figures 5 and 6). The fire-suppressing effects of increasing wetness and increasing population density compensate the tendency for increased burned area due to an increase in herbaceous fuels and temperature.
In both northern and southern hemisphere South America, population density contributed to declining marginal burned area and increasing temperatures and FAPAR contributed to increasing burned area (figures 2(c)-(d)). FAPAR had widespread positive trends especially in shrub or herbaceous-dominated regions (supplementary figure 5). Increasing FAPAR in non-tree-covered regions causes increasing marginal burned area in the SOFIA model (supplementary figure 1). In the arc of deforestation at the southern edge of the Amazon rainforests, VOD decreased because of decreasing tree cover (supplementary figure 5 and 6) (Andela et al 2013). This decrease in VOD caused an increase in the simulated marginal burned area ( figure 2(d)). Hence, the SOFIA model results suggest that deforestation contribute to increased flammability in tropical forests. In summary, vegetation changes in South America are diverse but overall promote increasing burned area, which is then counteracted by changes in other factors.

Discussion and conclusions
Our results provide a more nuanced view of recent changes in global burned area. The multi-temporal trend analysis approach allows us to assess the robustness of trends by accounting for the effects of individual years. Although the satellite datasets do show a decline globally, we found that the decline is not significant, strongly affected by the year-to-year variability in burned area, mostly caused by the last three years (2013-2015) with low Figure 2. Trends and controls in regional burned area. Panels (a)-(d) show regional annual totals and trends of burned area (BA) and the marginal burned area, which is controlled by a single factor (BA(x) anomaly relative to mean BA(x) in 1994-2011). Trends denoted with a * star symbol are significant (p £0.05, Mann-Kendall trend test). Panel (e) shows the spatial extent of the regions (NHAF=Northern-hemisphere Africa, SHAF=Southern-hemisphere Africa, NHSA=Northern-hemisphere South America, SHSA=Southern-hemisphere South America). Panels (f)-(i) show distributions of the regional burned area trend slope from the GFED4, FireCCI50 datasets and the SOFIA model for the overlapping period global burned area, and shows major differences between satellite datasets. The detection of trends in (satellite) time series is generally hampered by changes in the underlying sensor (Tian et al 2015, Hammond et al 2018 and by the year-to-year variability and noise (Forkel et al 2013), which increases the number of years required to detect trends (Weatherhead et al 1998). Our results suggest that trends in burned area need to be examined with (1) longer and MODIS-independent satellite time series (e.g. by exploiting the AVHRR or Landsat archives); (2) with long-term regional observations (Kasischke and Turetsky 2006, Müller et al 2015, Doerr and Santín 2016; and (3) with independent fire-related variables (Kaiser et al 2012, Santín et al 2016. Even with longer records, the interplay of climate, vegetation and human controls on fire will pose challenges for the attribution of the trends. Interpretations based on bivariate relationships only reflect the emergent controls on fire trends; attribution to fundamental underlying controls requires a more sophisticated approach using multivariate statistical or machine learning approaches (Bistinas et al 2014, Forkel et al 2019). Only by combining long time series and advanced analytical methods will it be possible to make robust statements about changes in fire regimes and the underlying controls.
The ability to reproduce recent trends in global burned area has been used as an indication of the poor performance of fire-enabled DGVMs (Andela et al 2017). This conclusion needs to be carefully re-assessed. DGVMs have regionally various performances in simulating burned area (Forkel et al 2019) and hence comparison of global trends might be misleading. Uncertainties in the satellite datasets also challenge the statement that 'fire models were unable to reproduce the pattern and magnitude of observed declines' (Andela et al 2017). The SOFIA model, in common with some of the fire-enabled DGVMs, has a weaker decline in global burned area (1994-2011) than the GFED4 dataset (1996-2015) but globally and in northern-hemisphere Africa it has comparable trends with the FireCCI50 dataset in the overlapping period.
As fire is influenced by the interplay of climate, vegetation, and human activities, changes in these controls can cause diverse changes in fire and burned area. The SOFIA model indicates that the decline in northernhemisphere Africa is associated with increasing population density project. SPH acknowledges the support from the ERC-funded project GC2.0 (Global Change 2.0: Unlocking the past for a clearer future, grant number 694481).