Global search for temporal shifts in fire activity: potential human influence on southwest Russia and north Australia fire seasons

Decadal trends in fire activity can reveal important human and climate-driven influences across a multitude of landscapes from croplands to savannas. We use 16 years of daily satellite observations from 2003 to 2018 to search globally for stationary temporal shifts in fire activity during the primary burning season. We focus on southwest Russia and north Australia as case study regions; both regions experienced nearly 40 d shifts over a 16 year period but in opposite directions. In southwest Russia, a major wheat-growing region, we trace the delay in post-harvest fires to several potential drivers: modernization in the agricultural system and recent droughts, followed by government restrictions on wheat exports. In north Australia, prescribed burns in the early dry season are a key practice in Aboriginal fire management of savannas, and the increasing trend of such fires has limited the size and extent of fast-spreading late dry season fires, thereby shifting overall fire activity earlier. In both regions, human action, through controlling fire ignition and extent, is an important driver of the temporal shifts in fire activity with climate as both a harbinger and an amplifier of human-induced changes.


Introduction
Two decades of daily satellite data from NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) have made analysis of long-term trends in global fire activity possible at daily to monthly scale . While climate variability is a dominant control on fire activity, humans play an important role in modulating fire behavior through ignition, fire suppression and management, and land use change (Pechony and Shindell 2010, Andela and Van Der Werf 2014, Chen et al 2017, Riley et al 2019, Bowman et al 2020. Andela et al (2017) mainly attribute the 24% decline in global burned area from 1998-2015 to human activity, through expansion in populated areas, croplands, and livestock, rather than variability in precipitation or temperature. Understanding trends in fire activity can (a) shed light on the consequences of changes in local land management practices and agricultural systems and (b) lead to efforts to mitigate fire-related impacts on air quality, public health, ecosystem services, and the carbon budget (Johnston et al 2012, Zhang et al 2014a, Koplitz et al 2016, Liu et al 2016, Bowman et al 2020. While most studies to date have investigated global trends in the magnitude of fire activity (e.g. Andela et al 2017), here we focus on regional shifts in the timing of peak fire activity. Regions with large, sustained temporal shifts over a multi-year period may reveal human influence on fire regimes through management or land use. In forest management, controlled burning of fuel-abundant areas in advance of the fire season peak can reduce the number and size of large fires (Fernandes and Botelho 2003). An earlier fire season suggests better managed forests with fewer large and destructive fires, while a later fire season would indicate the opposite. In croplands, recent work using satellite fire data identified a primarily human-driven delay in rice residue burning in north India, where a ∼2 week delay in the post-monsoon fire season (October-November) from 2003 to 2018 can be partially attributed to a 2009 groundwater policy that delayed rice planting dates closer to the summer monsoon onset (Balwinder-Singh et  The policy, which arose from an effort to combat the severe groundwater decline (Singh 2009, Tripathi et al 2016, unleashed a cascade of effects that delayed the rice growing season and subsequent burning of rice residues, potentially with unintentional consequences for air quality (Balwinder et al 2019, Jethva et al 2019, Liu et al 2021. This is because burning later in the year coincides with meteorological conditions that favor the accumulation of pollutants-i.e. weak ventilation from temperature inversions, stagnant winds, and low mixing layer heights (Ojha et al 2020).
To fully understand the potential of temporal shifts in the fire season to signal human imprints and influence regional air quality and human health, we must first identify such regions. In this study, we search globally and uncover additional regions with notable temporal shifts in fire activity from 2003 to 2018. We further investigate two case study regions, southwest Russia and north Australia, where we observe particularly large temporal shifts of nearly 40 d during the primary fire season in opposite directions-later in Russia and earlier in Australia. Finally, we probe and discuss possible human-and to a lesser extent, climatic-influences on temporal shifts in the two case study regions, and briefly, on a global scale.

Satellite fire datasets
As two measures of fire activity, we use the Level 3 MODIS Collection 6 burned area (MCD64A1) and active fire (MOD14A1/MYD14A1) products (table S1). MCD64A1 is available in monthly timesteps from 2000 to present at 500 m spatial resolution (Giglio et al 2018). We disaggregate MCD64A1 burned area to daily timesteps according to the Julian burn date. The MOD14A1 and MYD14A1 gridded active fire products are available daily at 1 km spatial resolution for the Terra (10:30 am/pm) and Aqua (1:30 am/pm) satellite overpasses, respectively . For each day, we combine MOD14A1 and MYD14A1 active fires (hereafter referred to as MxD14A1) into a binary fire mask (0 = no fire, 1 = fire), a custom product for this analysis that is different than the MCD14ML global monthly fire location product . To reduce noise at the pixel level, we sum daily MCD64A1 burned area and MxD14A1 active fire area on a 0.25 • × 0.25 • grid. More details about the satellite fire datasets are provided in supplementary section S1 (available online at stacks.iop.org/ERL/16/044023/mmedia).

Global search for temporal shifts in the primary fire season
As an initial search for temporal shifts, we focus on a 16 year time period from 2003 to 2018 but include observations in 2002 if the primary fire season for a particular grid cell extends from December to January of the following year. In brief, our approach, following Liu et al (2020Liu et al ( , 2021, is to track the changes in the midpoint of the primary fire season, k midpoint , defined as the mean of days t, weighted by daily burned area (BA t ): (1) For each grid cell, we define the primary fire season as a 7 month time window centered on the month with the highest burned area but exclude those months with <1% annual burned area; we use t = 1, 2, 3 . . . n to represent days of the fire season. For each grid cell, we take the slope of a simple linear regression of the yearly k midpoint timeseries as the magnitude of the temporal shift in the fire season, in days per year. In this initial step, we define a lax threshold of p < 0.1 as statistically significant to more easily identify clusters of grid cells with temporal shifts and define regions of interest (ROIs). We further exclude grid cells with <5 years of nonzero annual burned area to focus on stationary temporal shifts.
Because of this focus on areas where fires recur on an annual basis at similar intensity, potential ROIs consist primarily of cropland and savanna, grassland, and shrubland areas rather than forested areas. We use the MODIS MCD12Q1 land cover product with the University of Maryland classification to assign the primary land cover for each ROI (Sulla-Menashe et al 2019). We further investigate two ROIs as our case studies: southwest Russia and north Australia (figure 1, table 1), where the temporal shifts in the primary fire season are particularly large (>1 month from 2003 to 2018) and spatially cohesive. For each ROI, we refine the start and end dates of the primary fire season based on the time series of daily regional stacked by Julian day from 2003 to 2018. As described above, our initial global search to identify ROIs uses a generalized 7 month window centered on the month with the highest burned area.
Additionally, the very different geographical context of these two ROIs-croplands in southwest Russia and savannas in north Australia-allows us to juxtapose the potential land cover and region-specific mechanisms that led to the temporal shifts. Fires in both regions are heavily impacted by human activity, either through agricultural practices or fire management. For case study 1, we define southwest Russia as Krasnodar Krai, Republic of Adygea, Rostov Oblast, and Stavropol Krai, which are major wheatgrowing areas located in the North Caucasus and Southern Federal Districts to the east of the Black Sea ( figure S2(a)). This region, with its fertile black soils and favorable climate, produces over half of Russia's winter wheat, with ∼25% higher yields than the For our two case study regions, southwest Russia and north Australia, we use two methods to estimate k midpoint : (a) weighted mean (as described in section 2.1) and (a) sigmoid-smoothed partial sums (Zhang et al 2014b, Liu et al 2020, 2021. Both methods have been used to characterize the temporal progression of the fire season; moreover, the two methods show minimal differences in the derived temporal shifts in the fire season (Liu et al 2020).
In method 2, k midpoint is defined as the first day by which 50% of total fire activity as indicated by burned area has occurred. To reduce the sensitivity of this method to large spikes in fires, we apply sigmoid smoothing to the partial sums before calculating k midpoint . Hereafter, we abbreviate k midpoint from method 1 as k midpoint, ω and that from method 2 as k midpoint, ρ . We find that applying sigmoid smoothing in method 2 increases agreement between k midpoint, ω and k midpoint, ρ . While method 1 is computationally efficient for our initial global search, method 2 allows us to more accurately characterize the start and end of the fire season, k start and k end , for fire seasons with asymmetrical temporal distributions. We define k start and k end as the day by which 10% and 90%, respectively, of total burned area for that season has accumulated. In method 2, k can generalized as k β , where β represents various breakpoints in the fire season: whereŷ k is the sigmoid-smoothed partial sums of the sequence of daily burned area over day 1 to k, n is the total number of days in the primary fire season, and k β is the first day by whenŷ k , normalized by the sum of burned areaŷ n during the fire season, has surpassed breakpoint β. As in the global search, we use a simple linear regression to assess the trends in k midpoint (β = 0.5), k start (β = 0.1), and k end (β = 0.9). Here for our case study regions, we define p < 0.05 as statistically significant. Ancillary datasets and methods used to provide context on the potential links between fires, human activity, and meteorology are described in supplementary section S2. Here we focus on human influences on fire activity rather than meteorological changes, as land management practices and policies are often documented and occur on a much finer spatial scale, i.e. within designated areas or bounded by administrative borders.

Global temporal shifts in fire activity
We detect temporal shifts in 22% of global burned area using MCD64A1 and 25% using MxD14A1 active fire area at a statistical significance threshold of p < 0.1 (figure 1). However, we note that some grid cells with temporal shifts are spatially fragmented or opposite in sign relative to their neighbors. This 'speckle' may be characteristic of a mismatched grid cell size for analysis where it is incompatible with the process being observed (i.e. non-stationary) or statistical noise. Instead, we focus on large clusters of grid cells with similar magnitudes and the same sign in temporal shifts, so that when burned area is aggregated over the defined ROI, the overall temporal shift is likely statistically significant as well. In figure 1, we highlight select ROIs with coherent temporal shifts for croplands (northwest India, southwest Russia, south Russia, and Thailand) and for savannas, grasslands, and shrublands (north Australia, central Madagascar, Zimbabwe, and north Brazil). However, other clusters may be present, such as within the African Sahel, where the spatial patterns of temporal shifts are complex.
In the following sections, we delve into the two case study regions, southwest Russia (section 3.2) and north Australia (section 3.3), which see large temporal shifts of 2.41 d yr −1 , or a total of 39 d from 2003 to 2018 (  figure S1). We conclude with a discussion of the drivers of temporal shifts in fire activity (section 3.4).

Case study 1: southwest Russia
The primary fire season in southwest Russia croplands extends from summer to autumn, mainly from late June to October, after the winter wheat harvests (Hall et al 2016, Korontzi et al 2006 figures 3(a) and (b)). We estimate that the midpoint of the fire season has shifted later by 2.41 d yr −1 , or 39 d from 2003 to 2018 (table 1, figures 2(a) and S2). This shift is consistent across breakpoints in the fire season, varying Redder colors denote later years. Right panels: Estimated breakpoints (e.g. start, end, midpoint) in primary fire season. Inset shows the temporal shift at the midpoint of the fire season using the weighted mean method (black line) and the total 16 year shift. Inset right shows the temporal shift for each breakpoint using the partial sums method (gray lines), and asterisks denote statistical significance at p < 0.05. from 2.3 to 2.82 d yr −1 , and driven largely by a 95% decrease in fires from June to July and a 41% increase in September.
Here the example of northwest India described above can inform our case study on southwest Russia, as both involve water-stressed agricultural systems. In northwest India, state government policies delayed rice sowing to alleviate heavy groundwater usage prior to the monsoon onset; in turn, post-monsoon rice harvests and subsequent burning of rice residues shifted later. For southwest Russia, we discuss two potential drivers of the delay in fire activity: (a) improved management, investment, and productivity in the agricultural sector and (b) groundwater depletion and drought, followed by federal government policies, such as wheat export restrictions and bans.
First, since the early 2000s, the region has experienced the rise of 'agroholdings,' a new type of agricultural enterprise that acquires corporate farms and efficiently manages end-to-end production, processing, and distribution (Rada et al 2017). Rada et al (2017) attributes the high productivity in southwest Russia to these agroholdings, as well as to the favorable climate and soils and proximity to Black Sea ports. Puzzlingly, the 132% increase in winter wheat production in the region, and by extension the volume of crop residues, coincides with a 77% decrease in burned area since 2005 ( figure S3). This is opposite of the trend in northwest India, where the rise in rice production and adoption of mechanization (i.e. use of combine harvesters) led to an increase in rice residue burning (Liu et al 2019, Shyamsundar et al 2019. This discrepancy between the regions may exist for two reasons. First, wheat residue is more suitable for livestock feed than rice residue, which has high silica content, so there may be greater utility for residue in southwest Russia than in north India (Sidhu et al 2015). Second, modernized agriculture systems in southwest Russia may have driven changes in crop residue management, wherein farmers use fire less to remove wheat residues immediately postharvest (around July) and more to clear residual waste on fields just prior to sowing (around September) (figures 2(a), 3(a) and (b)). In contrast, the short turnaround period (on the order of weeks) between crops in the rice-wheat rotation of northwest India (Vadrevu et al 2011), means that Indian farmers depend on fire to rapidly clear their fields (figure 3(a); Liu et al 2020). Consistent with this hypothesis, the average fire intensity per active fire detection peaks in July and decreases by 41% by September ( figure S4).
In addition, Rostov farmers surveyed by the U.S. Department of Agriculture (USDA) claim that only small enterprises still burn to reduce risk of soil-based diseases affecting crop growth (Hall et al 2016). Greater alternative uses for wheat residue, as mentioned above, and the availability of direct seeding or no-till technology may explain the declining use of fire as a primary tool for crop residue management. For example, Krasnodar farmers typically embed loose wheat residues into soil as a fertilizer and to prevent soil degradation (Sidorenko et al 2017). Additionally, wheat residues may be baled for livestock feed, in view of the resurgence in the livestock sector in the 2000s due to government policies and investment (Liefert and Liefert 2012, Fellmann et al 2014, Rada et al 2017. As early as 2014, civil society organizations invested in agricultural outreach and training to eliminate agricultural burning in Krasnodar and Rostov (ICCI 2020). Evidence of a link between the temporal shift in the fire season in southwest Russia and modernization in the agricultural system may be found in the record of wheat yields, which are correlated to fertilizer use and farm machinery efficiency (Liefert et al 2013, Deppermann et al 2018. Indeed, we see higher winter wheat yields and temporal shifts in Krasnodar and Adygea (2.69-2.8 d yr −1 , p < 0.05) than in Rostov and Stavropol (1.59-1.93 d yr −1 , p < 0.05) (figure S3, table S4).
Second, high water withdrawal for competing uses in residential, industrial, and agricultural sectors has led to moderate to severe water stress in southwest Russia (Alcamo et al 2007). Rodell et al (2018) attributes the downward trend in freshwater storage from 2003 to 2016 to groundwater depletion and drought. Within our study region, GRACE observations indicate a decrease in freshwater availability of 1.43 cm yr −1 (figure 3(d)), which is comparable to that in northwest India (−1.44 cm yr −1 ). In particular, changes in vapor pressure deficit (VPD) are most pronounced during summer and early autumn days, with increases up to ∼30-70 Pa yr −1 from 2003 to 2018 (figure 3(c)). Dryness can delay winter wheat sowing (Wegren 2011, USDA 2020, which may partly explain the increase in late season fires. Additionally, drought in tandem with federal government policies may have exacerbated delays in winter wheat sowing in this region. In August 2010, after intense heatwaves, drought, and wildfires devastated crop production elsewhere in Russia, the government banned wheat exports in an effort to stabilize domestic food prices (Wegren 2011;figure 3(e)). While southwest Russia was spared damage to its winter wheat, the export ban disrupted the supply chain in this region, and wheat prices plummeted (Wegren 2011, ROSSTAT 2020a. Anecdotally, some farmers in Krasnodar waited to sell their wheat crop to return a higher profit (De Sousa 2019), and the wheat price then strongly rebounded. In contrast, no export restrictions were imposed in 2012, a severe drought year with low wheat production and exports similar to those in 2010 (Götz et al 2016). Nevertheless, like in 2010, wheat producer prices in 2012 sharply rose starting in July, indicating that farmers may have again waited to sell their wheat crop ( figure 3(e)). These price swings and related delays in selling wheat may have also delayed sowing if farmers relied on these profits to buy seeds and fertilizer for the following year's crop.
In summary, factors that may have driven the 39 d delay in fires in southwest Russia include improvements in agricultural management and technology, drought and groundwater depletion, and fluctuations in wheat prices due to export restrictions and bans.

Case study 2: north Australia
The primary fire season in the savannas of north Australia occurs during the dry season primarily from May through November. We estimate that the midpoint of the fire season has shifted by −2.41 d yr −1 , or 39 d earlier from 2003 to 2018 (table 1, figures 2(b) and S6). This shift is consistent across breakpoints in the fire season, varying from −3.06 to −2.29 d yr −1 . We find the trend is induced by a 40% increase during the early dry season (EDS) from May to July and a 67% decrease in fires during the late dry season (LDS) from September to November, with the transition point from the EDS to LDS defined as July 31 (Price et al 2012).
During the LDS, fires tend to burn out of control due to favorable fire weather, such as high VPD values. The resurgence of Aboriginal fire management practices in north Australia, which began in the 1990s, has spurred the decrease in fires during the LDS, most notably in Arnhem Land (Yibarbuk et al 2001, Whitehead et figure S6). The economic incentive for Aboriginal people involved in these carbon abatement projects adds to inherent environmental, social, and cultural co-benefits of such projects (Andersen et al 2005, Ansell et al 2020. Our results provide quantitative evidence of the success of these policy efforts to reduce fire size, especially later in the dry season when fires are more likely to burn out of control. We find that fire management within the savanna burning project areas in the Top End and Cape York drives the temporal shifts toward an earlier fire season, with spatially cohesive shifts  (table S5). Moreover, ALFA project areas, which have a long history of fire management (Price et al 2012, Price 2015, Ansell et al 2020, exhibit a larger overall temporal shift of −4.3 d yr −1 (p < 0.05), particularly in WALFA at −6.13 d yr −1 (p < 0.05), than the region overall.
In north Australia fire management, prescribed burning creates a mosaic of burned and unburned areas during the EDS, decreasing overall fuel availability across the landscape (Andersen et al 2005, Trauernicht et al 2015, Russell-Smith et al 2019. The fuel-limited areas, where prescribed burning has occurred, act like a patchwork of endpoints for fastspreading wildfires in the LDS, thus limiting the maximum extent of the final burn perimeters (Petty and Bowman 2007, Price et al 2012, Petty et al 2015. This phenomenon is widespread across Arnhem Land, where small fires have become more prevalent during the EDS and widely dispersed across the landscape (figures 3(a)-(d)). During the EDS, we see a 60% decline in the average distance to the closest fire and a 50%-60% increase in the number of distinct fire perimeters from 2003 to 2018 (figure 4(e)). The increase in EDS fires and their spatial coverage suggest fewer open spaces for fires later in the season to spread across, and indeed, we see a 49%-54% decrease in the average area per fire during the LDS (figure 4(f)). The decrease in LDS fire activity is consistent across burned area and fire emissions datasets, as described in supplementary section S5.
Finally, we investigate the influence of meteorology on the 59% decrease in LDS burned area (figures S9(b) and (c)). If meteorology played a major role, we would expect wetter and cooler conditions, or decreases in VPD and FWI over the 2003-2018 timeframe. In fact, we see the opposite-a 6% increase in VPD and 3% increase in FWI, although these trends are not statistically significant (figures S9(b) and (c)). When detrended, however, VPD and FWI are moderately to strongly correlated with burned area, with the correlation r = 0.57 for VPD and r = 0.79 for FWI (p < 0.05). Thus, meteorology appears to drive the interannual variability but not the decrease in LDS burned area over this time period that caused the fire season to shift earlier. Our hypothesis is consistent with Andela et al (2017), who concluded on a global scale that precipitation is strongly related to the interannual variability in burned area, yet human activity drives the overall decrease in burned area, inferred from proxies such as GDP, population growth, and widespread wildland to agricultural transition.
In summary, our approach using temporal shifts and distance to nearest fire strengthens the conclusion of previous studies that EDS prescribed burning, particularly in Arnhem Land, has shifted the fire season earlier and lowered LDS burned area and emissions.

Drivers of temporal shifts
In north India and the two case study regionssouthwest Russia and north Australia-humans play an integral part in shaping the landscape by using prescribed burns, either to clear fields in an agricultural context or as a tool to prevent larger, more destructive fires in savannas. Although we do not find large changes in meteorological variability in line with the 39 d temporal shifts in the fire season, climate shocks, such as droughts and heatwaves, or concerns about rising greenhouse gas emissions may spur human actions that affect the use of fire, either through fire management or changes in government policy and agricultural practices. For example, in response to groundwater depletion and drought, state governments in north India delayed rice sowing closer to the monsoon onset, whereas in Russia, the federal government enacted wheat export bans and restrictions to protect domestic supply. In north Australia, Indigenous fire managers use prescribed burns to limit greenhouse gas emissions from LDS fires. Our probe of the potential human and climatic drivers is not exhaustive, and further study, such as through household surveys, is needed to systematically detail and confirm the mechanisms for the temporal shifts in each region described here, especially in southwest Russia.
Through the lens of the two case study regions, we briefly discuss five other ROIs, grouped by the two broad land cover types: croplands and savannas, grasslands, and shrublands. For croplands, the southwest Russia case study and studies on north India (Balwinder et al 2019, Jethva et al 2019, Liu et al 2021 suggest that water stress, as well as increasing mechanization, modernization, and productivity, in agricultural systems are likely causes of changes in the timing of peak burning. In both south Russia and Thailand, the primary fire season coincides with the postharvest period for spring wheat and rice, respectively (Vadrevu et al 2015, Hall et al 2016. GRACE observations of freshwater availability show an increase in south Russia of +0.52 cm yr −1 from 2003 to 2016, coinciding with earlier peak burning. Thailand, on the other hand, reveals a decline in freshwater availability of −0.53 cm yr −1 -that is, drought and/or groundwater depletion-coinciding with a delay in the fire season. For savannas, grasslands, and shrublands, the north Australia case study suggests that proactive fire management practices are key to limiting the fires with fast spread rates and high fuel consumption that occur later in the season. In contrast, poor fire management, in combination with drought, may drive fire seasons later. Fire is mainly used pasture management for grazing cattle in central Madagascar, whereas deforestation is common in Brazil, and careless use of fire, both deliberate and accidental, has led to destructive 'veld' fires in Zimbabwe (Kull 2012, Nyamadzawo et al 2013, Aragão et al 2018. In fact, for Zimbabwe, Nyzmadzawo et al (2013) recommends a shift from fire suppression to fire management, with early season prescribed burns to curb the fire activity later in the season. Globally, Lipsett-Moore et al (2018) find widespread potential in EDS prescribed burning to both mitigate fire emissions and improve livelihoods.

Conclusion
In conclusion, we use 16 years of daily satellitederived burned area to quantify the temporal shifts in the primary fire season at global scale. While earlier studies have diagnosed trends in annual fire activity (Giglio et al 2010, Andela and Van Der Werf 2014, this is, to our knowledge, the first global study of trends in the timing of peak fire activity. We focus on regions where fires consistently recur year after year, and in particular, we select southwest Russia and north Australia as two case study regions, where the fire season has shifted by a total of nearly 40 d. No matter the setting, either croplands in southwest Russia or savannas in north Australia, we find human action as likely an important contributor of these temporal shifts. Pinpointing regions where such shifts in the fire season occur is critical for contextualizing broader changes in agricultural systems, fire management practices, and government policies and for understanding the consequences of these changes. In India, government efforts to conserve water in spring have delayed the onset of rice residue burning during the post-monsoon season and inadvertently led to degraded air quality. In southwest Russia, human innovations and investment in agricultural technology and management have limited the need for fire to clear wheat residue. Meanwhile in north Australia, widespread adoption of Aboriginal fire management with increased prescribed burning have curbed the frequency of large fires over a broad region and potentially limited greenhouse gas emissions.

Data availability statement
All data used in this study are publicly available. MODIS burned area and active fire datasets are distributed by NASA Earthdata (https:// earthdata.nasa.gov/). The ERA5 reanalysis dataset is provided by ECWMF (www.ecmwf.int/). GRACE data are provided by NASA/JPL, University of Texas/CSR, and GFZ-Potsdam. These datasets are also available from the Google Earth Engine Data Catalog (http://earthengine.google.com/).
The data that support the findings of this study are available upon reasonable request from the authors.