Wetter environment and increased grazing reduced the area burned in northern Eurasia from 2002 to 2016

Northern Eurasia is currently highly sensitive to climate change. Fires in this region can have significant impacts on regional air quality, radiative forcing and black carbon deposition in the Arctic which can accelerate ice melting. Using a MODIS-derived burned area dataset, we report that the total annual area burned in this region declined by 53 % during the 15-year period from 2002 to 2016. Grassland fires dominated this trend, accounting for 93 % of the decline in the total area burned. Grassland fires in Kazakhstan contributed 47 % of the total area burned and 84 % of the decline. A wetter climate and increased grazing are the principle driving forces for the decline. Our findings (1) highlight the importance of the complex interactions of climate–vegetation– land use in affecting fire activity and (2) reveal how the resulting impacts on fire activity in a relatively small region such as Kazakhstan can dominate the trends in burned areas across a much larger landscape of northern Eurasia.


Introduction
Fire activity worldwide is very sensitive to climate change and human actions, especially over high-latitude ecosystems (Goetz et al., 2007). Identifying and unraveling confounding drivers of fire is critical for understanding the recent and future impacts of fire activity. In northern Eurasia, fire activity impacts of chief concern include carbon cycling, boreal ecosystem dynamics, fire emissions (Hao et al., 2016a), accelerated ice melting in the Arctic (Hao et al., 2016a;Evangeliou et al., 2016), early thawing of permafrost and the hydrological cycle at high latitudes (IPCC, 2014). In addition, fire activity affects air quality in Europe, Asia and North America. An improved understanding of the region's fire dynamics can also be applied to develop climate change mitigation policy and can be incorporated into the fire modules of Earth system models to improve their predictions (Hantson et al., 2016).
The global mean surface temperature rose by approximately 0.72 • C from the year 1951 to 2012 according to the Fifth Assessment Report (AR5) of the Intergovernmental Panel on Climate Change (IPCC, 2013), but it remained relatively constant ("warming slowdown") from 1998 to 2013 (Fyfe et al., 2013(Fyfe et al., , 2016Cowtan and Way, 2014;Tren-W. M. Hao et al.: Wetter environment and increased grazing reduced the area burned in northern Eurasia berth et al., 2014). Nevertheless, extreme high-temperature events continued to occur even during the warming slowdown (Seneviratne et al., 2014;Trenberth et al., 2015). Since 2013, the global temperature has risen rapidly (NASA Global Climate Change, 2019), and there were hemispheric temperature anomalies from 1850 to 2015 (Jones et al. 2016). In the Northern Hemisphere, temperatures have increased more profoundly than in the Southern Hemisphere since the 1980s, as they are greatly affected by the sources of greenhouse gases and many other factors. High latitudes are projected to have the largest temperature increase globally by 2100 (IPCC, 2013). At the same time, however, climatic components of the fire weather index (FWI), an index of fire intensity potential, have experienced regional divergence at these latitudes with a positive FWI trend in East Asia and a negative trend in Kazakhstan (Jolly et al. 2015), suggesting divergent regional climate impacts. In northern Eurasia, current accelerated high temperatures in the summer have also been observed on the eastern European Plain and in central Siberia (Sato and Nakamura, 2019).
Over the past 20 years, a decline in total area burned in Eurasia has been observed (Giglio et al., 2013;Hao et al., 2016a, Andela et al., 2017. We will investigate trends in the spatial and temporal distribution of area burned from 2002 to 2016 across different land cover types and geographic regions in northern Eurasia, a region highly sensitive to climate change. The geographic subregion with the largest declining trend is examined, and the influence of the confounding factors of climate and human activity on burned area is explored. Our study seeks to evaluate the decline in burned area as a function of variable fuel conditions (Krawchuk and Moritz, 2011), land use and relative moisture conditions (Pausas and Ribeiro 2013). Beside these climate variables, on a global scale, abrupt changes have been observed to significantly impact long-term or recent fire history (Pausas and Keeley, 2014), along with other mechanisms such as herbivory from native and domestic ungulates and humans (e.g., fire prevention). Considerable research has been done to understand climate-fire-grazing interactions in grassland ecosystems. In grasslands, reductions in fuel availability due to decreasing net primary production, grazing or other management activities can be the key variables limiting fire spread (Moritz et al., 2005). For instance, in the western United States, research has significant implications on forest and rangeland management (e.g., Bachelet et al., 2000;Gedalof et al., 2005;Riley et al., 2013;Abatzoglou and Kolden, 2013). Similar issues have been investigated on an African savanna for maintaining sustainable grassland (e.g., Archibald et al., 2009;Koerner and Collins, 2014). In this study we closely examine the interactions of climate, fire, grazing and fuel availability in Kazakhstan, the country in northern Eurasia with the largest decline in burned area during the 2002-2016 period.

Study area
First, we study the area of northern Eurasia, a region from 35 • N to the Arctic and from the Pacific Ocean to the Atlantic Ocean. The region comprises 21 % of the Earth's land area and encompasses diverse ecosystems from the steppes of central Asia to the Arctic. Forest is the major ecosystem in this region, covering 27 % of the area, followed by grasslands, which cover 16 % (Friedl et al., 2010).
Second, to understand the forces driving the decline in burned area, we focus on the effects of drought and grazing in Kazakhstan. From 2002 to 2016, Kazakhstan had the highest rate of decline in burned area in northern Eurasia (Figs. 1, 2). In Kazakhstan, grassland is the dominant ecosystem and grazing is the major agricultural activity (Food and Agriculture Organization FAO Live Animals Database, 2016).

Burned area in northern Eurasia
Since 2000, global burned area has been mapped by remote sensing (e.g., Mouillot et al. 2014) with different sensors and detection algorithms , leading to multiple datasets with a significant uncertainty in the magnitude of spatial distribution, interannual variability and trends in burned area (Hantson et al., 2016). We used the daily NASA MODIS (Moderate Resolution Imaging Spectroradiometer) dataset at a 500 m × 500 m resolution. Our MODIS-derived burned area algorithm was validated in eastern Siberia with the Landsat-derived burned area (30 m × 30 m) (Hao et al., 2012). The ratio of these two satellite-derived burned areas was 1.0 with a standard deviation of 0.5 % over 18 754 grid cells. Among other sources of variability, surface and crown fires generate significantly different spectral signals, so that the detection algorithm depends on vegetation type classification .
The burned area data were analyzed at multiple spatial and temporal scales using frequentist statistical methods (see Sect. 2.4) to identify regional trends. Assessing burned area changes in northern Eurasia over this time period benefits from the lack of fire suppression in this region (Goldammer et al., 2013), so the impact of climate and land use on fire activity can be better understood. Our methodology for mapping daily burned area is very similar to that used by Hao et al. (2016a, b), which was specifically developed for this region. For the study of Hao et al. (2016aHao et al. ( , 2016b, the MCD12Q1 land cover map of 2015 was used for 2002-2016. For this present study, a temporally consistent and up-to-date land cover product was used for 2002-2013 and the 2013 land cover map was used for 2014-2016, as current versions were not available for present and previous studies.

Data sources of drought, livestock, annual biomass production and land cover
The following data sources for estimating the factors affecting the burned area in Kazakhstan are described below: drought, livestock, annual biomass production and land cover. All data were evaluated at the county level for 174 counties during the period from 2002 to 2016 (Fig. 1). We focused on Kazakhstan, as it was the region with the largest decline in burned area in northern Eurasia (see Sect. 3.1).

Drought
The Palmer Drought Severity Index (PDSI) from the Ter-raClimate site (http://www.climatologylab.org, last access: 9 April 2021) was used to estimate drought throughout Kazakhstan (Abatzoglou et al., 2018). The PDSI was developed by Palmer (1965) and is widely used to estimate a rough soil water budget based on monthly precipitation and potential evapotranspiration, with varying soil properties for the available water content in order to account for pedological variations and species roots access to water. We used monthly PDSI data from March to July, defined as the fire season (Roy et al., 2008), to compute a cumulative drought effect index. The gridded PDSI data were available at a spatial resolution of ∼ 4 km and were aggregated to the county within the study area ( Fig. 1). The PDSI varies from +4 for wet conditions to −4 for dry conditions.

Livestock
The annual population of livestock in each of the 14 provinces, each consisting of multiple counties, of Kazakhstan from 2002 to 2016 were compiled from the official agriculture statistics of the Ministry of National Economy of the Republic of Kazakhstan Committee on Statistics (MANE, 2019). These data included yearly numbers of large horned livestock and sheep and goats at the province level, which is coarser than the county level. Livestock populations are only available at the province level, and the population was distributed proportionally to the size of the county area so that all potential drivers of fire activity could be evaluated on a common spatial scale. Thus, the livestock density for each county is defined as the ratio of the number of animals to the area of the county.

Annual biomass production
We estimated the annual biomass production within the grassland domain of the study area ( Fig. 2) using the production subroutine of the Rangeland Vegetation Simulator (RVS) model (Reeves, 2016) which applied the methods of Reeves et al. (2020). The RVS, which was originally developed for simulating rangeland vegetation dynamics in the continental United States, models annual production based on the MODIS normalized difference vegetation index (NDVI) at a 250 m spatial resolution (MOD13Q1). The MOD13Q1 NDVI data are composited on a biweekly basis and are available at a spatial resolution of 250 m. The quality assurance/quality control (QA/QC) flags were used to isolate only the best-quality NDVI pixels. At each pixel, the highestquality maximum value composite on an annual basis was retained for further analysis. The relationships between the estimates of annual net primary production (ANPP) and maximum NDVI were divided into two groups to enable different models to be fit to the upper and lower ends of production. The upper and lower ends of production are given as follows: where y is the estimated ANPP (in kg ha −1 of dry weight), and x is the NDVI for the upper range (x ≥ 0.46); and where y is the estimated ANPP (in kg ha −1 ), and x is the NDVI for the lower range (x < 0.46). The partitioning into two groups was done, in part, because of the asymptotic nature or "saturation" feature (Santin-Janin et al., 2009) of NDVI with respect to ANPP.

Land cover
The MODIS land cover product (MCD12Q1) Version 6 was used to assess factors affecting the burned area in Kazakhstan. The product is available at a 500 m spatial resolution and describes the distribution of broad vegetation types. We screened these data to subset only those vegetation types considered to represent grassland vegetation (Class 10 in the MCD12Q1 dataset) from 2000 to 2016.
In each year of the assessment, the number of grassland pixels was summed to enable estimates of grassland area throughout the study area.

Statistical analysis
For each 0.5 • × 0.5 • pixel, the annual trend ( Fig. 2) was estimated as the robust linear slope computed from burned area on year using M-estimation as described in Huber (1981). Our objective was to present consistent grid cell trends in the presence of within-cell variation. We chose to use Mestimation to mitigate the effect of large within-cell variation due to a relatively small within-cell samples such that the map presents a consistent surface. If computed using ordinary least squares (OLS) estimates, such large within-cell variation could result in some cells having inconsistent or "outlier" trends compared with their neighbors. The trends were estimated using the R platform (R Core Team, 2019) with the rlm R function in the MASS package (Venables and Ripley, 2002). Pairwise robust rank correlations (Figs. 5, 6) were computed as described in Kendall (1938) using the cor R function.
To validate our estimates on burned areas, we compare our annual northern Eurasian burned areas from the Forest Service Fire Emission Inventory -Northern Eurasia (FEI-NE) with the latest version of the MODIS burned area product (MCD64A1, Collection 6; Giglio et al., 2018) from 2002 to 2016 (Fig. 3). The burned areas reported by FEI-NE and MODIS MCD64 were each modeled separately by year. The models each include a first-order autoregressive term on the residuals to account for the presence of temporal autocorrelation. The response was assumed to be gamma distributed (Bickel and Doksum, 2015). A generalized linear mixed model (GLMM) approach was used that was estimated with the glmmTMB R function (R Core Team, 2019) in the glmmTMB R package (Brooks et al., 2017).
The potential driving forces of burned area at the county level for 174 counties over a period of 15 years from 2002 to 2016 were modeled using the GLMM approach to interpret the effects on the extent of the area burned. The proportion of burned area per county was modeled on the effects of year, PDSI during the fire season (May-July), proportion of grass area, ANPP and livestock density along with two-way interactions. The model included a random effect that accounts for spatial correlation within each region along with a firstorder autoregressive term on the residuals within each county that accounts for temporal autocorrelation. The response was assumed to be beta distributed (Bickel and Doksum, 2015). The model was estimated with the glmmTMB R platform (R Core Team, 2019) in the glmmTMB R package (Brooks et al., 2017).

Spatial and temporal distribution of burned areas in northern Eurasia
The declining trends in the spatial distribution of the area burned from 2002 to 2016 in northern Eurasia at a 0.5 • ×0.5 • resolution are shown in Fig. 2. The majority of the area burned was grassland in Kazakhstan in central Asia. However, substantial areas were also burned in the Russian Far East along the Chinese border because of illegal logging (Vandergert and Newell, 2003) and the subsequent fires to burn the remaining forest residues. The annual areas burned according to ecosystem and geographic region are summarized in Table 1. The interannual burned area in northern Eurasia varied by about 4 times: from 5.0 × 10 5 km 2 in 2003 to 1.2 × 10 5 km 2 in 2013 with an average of (2.7 ± 1.0) × 10 5 km 2 (n = 15). Grassland accounted for 71 % of the total area burned, despite comprising only 16 % of the land cover (Friedl et al., 2010). Almost all of the grassland fires occurred in Kazakhstan in central and western Asia (Table 1). In contrast, forest is the major ecosystem that covers 27 % of northern Eurasia (Friedl et al., 2010), but it only contributes to 18 % of the total area burned. About 90 % of the forest area burned occurred in Russia.  Fig. 4 were modeled in the same way as those reported in Fig. 3, with the same response distribution. The trends in wave-like burned areas are typical for burned area trends worldwide (e.g., Andela et al., 2017). The annual total area burned over northern Eurasia during this period decreased by 53 %, from 3.3 × 10 5 km 2 in 2002 to 1.6 × 10 5 km 2 in 2016 (Table 1), or at a rate of 1.2 × 10 4 km 2 (or 3.5 %) yr −1 . The grassland area burned during the 15 years declined by 74 %, from 2.8 × 10 5 km 2 in 2002 to 7.3 × 10 4 km 2 in 2016, or at a rate of 1.3 × 10 4 km 2 (or 4.9 %) yr −1 . Grassland fires in Kazakhstan accounted for 47 % of the total areas burned but contributed to 84 % of the declining trend. The annual forest burned area varied by a factor of 5, from 21 243 km 2 in 2010 to 111 019 km 2 in 2003, but there is no trend over the 15 years (Table 1).

Regional trends in driving forces over time in Kazakhstan
One of our objectives was to evaluate trends in the primary drivers responsible for reducing area burned, especially in grasslands at the county level. Pairwise correlation results are shown in Fig. 5. Each panel in Fig. 5 illustrates the coefficient of correlation between a key variable and year (2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016) for the 174 counties of Kazakhstan. The major factors affecting the trend in area burned in Kazakhstan are wetter climate (represented as PDSI), the proportion of grassland cover, ANPP and livestock density (Table 2). Both grassland partition and ANPP enable spreading fires. The declining trends in the fraction of the area burned annually are shown in Fig. 5a. The trend in PDSI from March to July during the 15-year period is illustrated in Fig. 5b. A higher PDSI value indicates a wetter environment. Increasing wetness, i.e., higher PDSI, during the fire season reduces the probability of fire ignition and fire spread. The declining trend in the burned area (Fig. 5a) is then consistent with the increasing trend in PDSI (wet conditions), especially in central and southern Kazakhstan (e.g., East Kazakhstan, Qaraghandy, Zhambyl and Almaty) (Fig. 5b).
Through time, the proportion of grassland cover has been asymmetric with some counties having exhibited strong decreases, such as in the north-central region of Kazakhstan, while others have seen increases, such as in the northwestern region (Fig. 5c). This north-central region has also exhibited decreases in burned area (Fig. 5a). Similarly, some regions have shown increasing trends in grassland cover through time without commensurate increases in the proportion of burned area (Fig. 5a, c).  "Estimate" denotes a parameter estimate from GLMM, "Std. error" refers to the standard error of the parameter estimate, "z" is the z statistic, and "Pr(> |z|)" is the p value.
The impacts of year, PDSI, land cover, ANPP and livestock density on the extent of the area burned and the correlations of burned area with these driving forces are illustrated in Fig. 6. Area burned and PDSI were negatively correlated in most of the counties in Kazakhstan (Fig. 6b). Therefore, as Kazakhstan becomes wetter during the fire season, the area burned declined over the 2002-2016 period. At the same time, grassland cover decreased across most of Kazakhstan, with a notable exception being the north-central and southwestern regions (Fig. 6c). The ANPP decreased with time over most of Kazakhstan, with the exception being the central and southwestern counties (Fig. 6d).
Finally, we investigated livestock density as a potential non-climatic driver affecting fuel amount. The population density of livestock increased with time in all counties and was greatest in the central, northern and southern counties of Qostanay, Pavlodar and Qaraghandy (Fig. 5e). The coupling of livestock density with PDSI affected the extent of the area burned ( Fig. S1.4 in the Supplement) with p = 0.042 (Table 2). The area burned was negatively correlated with the population of livestock throughout nearly all of Kazakhstan (Fig. 6e). This observation suggests that the increasing population of grazing livestock may have reduced fuel-bed continuity, contributing to the decrease in the area burned in Kazakhstan. Since 2000, the numbers of sheep, goats and cat- tle have increased by 60 % in Kazakhstan based on MANE statistics (2019) (Figs. S2 and S3 in the Supplement). Thus, increased livestock grazing could decrease the amount of herbaceous fuel across the landscape and offset increases in fuel quantity due to expanded grassland cover. The net result would be reductions in fire spread and the area burned.

Interactions of driving forces
The driving forces (e.g., year, PDSI, proportion of grassland cover, ANPP and livestock density) for the decline in the burned areas in Kazakhstan from 2002 to 2016 are interrelated. It is, therefore, critical to evaluate their interactions. For instance, Figs. S1.1-S1.4 illustrate the proportion of burned area affected by the interactions of the driving forces in 174 counties over 15 years (see table Table 2 for more information).

Proportion of grassland cover and year
Both year and the proportion of grassland area had significant effects on burned area when they interacted (Table 2, p < 0.001). When the proportion of grassland cover in a county was very low (e.g., 0.48 %), only about 0.6 % of the area was burned annually during the period from 2002 to 2016 ( Fig. S1.1, upper left panel). On the contrary, when the grassland cover was 25 % or greater, the area burned declined steadily from 1.5 % in 2000 to 0.6 % in 2016 (Fig. S1.2 lower right panel). This observation is consistent with grassland enhancing the spread of fires in the absence of opposing factors.

PDSI and the proportion of grassland area
Both the PDSI and the proportion of grassland area had significant effects on burned area when they interacted (Table 2, p = 0.028). As in Fig. S1.2, for a PDSI ranging from −4.5 to ∼ 2, the percentage of the area burned remained at about 0.6 % for a grassland area of 0.5 % (upper left panel). On the other hand, when grassland cover was 60 %, the fraction of area burned declined from 2.2 % to 0.8 % (lower right  panel). This analysis is consistent with grassland enhancing the spread of fires, as in the previous section discussing the proportion of grassland cover through time, and illustrates that increasing wetness significantly decreases burned area, mostly when grassland cover is high.

Livestock density and year
We investigated livestock density as a potential non-climatic driver affecting fuel amount and area burned. The effects of grazing on the area burned from 2002 to 2016 are shown in Table 2 (p = 0.089). The declining trend in the area burned with time for different livestock densities are illustrated in Fig. S1.3. A higher livestock density results in less available biomass to burn and less area burned (lower right panel). This provides additional evidence that grazing could reduce the area burned in Kazakhstan.

PDSI and livestock density
The interaction between the PDSI and livestock was significant with respect to the area burned (p = 0.042). Figure S1.4 shows the decline in the proportion of burned area with PDSI at different livestock densities. As the PDSI increases (wetter landscape), less area is burned. However, the declining trends differ with livestock density. This relationship is quite different for livestock densities of 0.002 (Fig. S1.4, upper left panel) and 0.05 head km −2 (Fig. S1.4, lower right panel). For instance, for a low PDSI (−4, dry), 1.5 % of the area was burned for all livestock densities. In contrast, at a high PDSI (+2, wet), the percentage of burned area decreases with increasing livestock density. Thus, the area burned is unaffected by grazing intensity during dry years, but during wet years with high biomass (based on our RVS analysis of Reeves, 2016), high grazing intensity tends to decrease burned area.

Burned area
The spatial and temporal extent of the area burned was examined in different ecosystems in northern Eurasia from 2002 to 2016, during which the average area burned was (2.7 ± 1.0) × 10 5 km 2 yr −1 . The burned area in grasslands declined by 74 %, from ∼ 282 000 km 2 in 2002 to ∼ 73 000 km 2 in 2016, at a rate of 1.3 × 10 4 km 2 yr −1 . The area burned in forest showed no trend over time. Our burned area is higher than the MODIS MCD64 Collection 6, in which the average annual burned area was 9.7×10 4 km 2 in boreal Asia during the same period (Giglio et al., 2018). Boreal Asia in MCD64 has a similar geographic region to our northern Eurasia. Nevertheless, the interannual variability and the trends in burned area for the two datasets are consistent (Fig. 3).
Our results regarding burned area trends are consistent with other published results (Giglio et al., 2013;Hao et al., 2016a;Andela et al., 2017) which concluded that the area burned in northern Eurasia declined, contrary to the projections of increased fire frequency driven by climate change (Groisman et al., 2007;Kharuk et al., 2008). Uncertainty in global burned area remains a critical challenge, with trends and interannual variability reported by sensors and processing algorithms exhibiting large differences (Hantson et al., 2016;Chuvieco et al., 2019).

Grassland fires and grazing
Grassland fires in Kazakhstan accounted for 47 % of the total area burned but comprised 84 % of the decline in the total area burned in northern Eurasia during the 15 years from 2002 to 2016. The grassland fires are a human-induced issue caused by the production of fresh grass for grazing (Lebed et al., 2012) with a cycle of about every 2 years. A similar temporal pattern characterizes grassland fire occurrence on the African savanna (Hao and Liu, 1994;Andela and van der Werf, 2014).
Central Asia experienced tremendous socioeconomic change, with the collapse of the Soviet Union in the 1990s leading to a full restructure of the agricultural system, followed by a rapid collapse of cattle industry which has progressively recovered over the last 20 years (Figs. S2 and S3;Food and Agriculture Organization, 2016). This change has potentially altered the fuel available to burn, as observed in other ecosystems (Robinson and Milner-Gulland, 2003;Holdo et al., 2009;Vigan et al., 2017). The coincident decline in burned area with increasing livestock population suggests that changing agricultural practices may have exerted an influence on fire activity in Kazakhstan and northern Eurasia. In addition, a relationship between the livestock population and the burned area was observed in arid grassland in a small region of southern Russia from 1986 to 2006 (Dubinin et al., 2011). During this time period, the livestock population was negatively correlated with the area burned.
The fire activity data for Kazakhstan and Mongolia can be estimated from 1985 to 2017, as shown in Fig. 7, based on the recently released Advanced Very High Resolution Radiometer (AVHRR) long-term fire history (Otón et al., 2019). This new information extends the analysis to before our observed decrease during the 2002-2016 period and shows that fire activity only increased in Kazakhstan during the economic collapse and the associated reduction in livestock around the year 2000. This opposite trend supports our interpretation regarding the relationship between grazing and burned area, particularly when this variation in burned area is not clearly observed in neighboring Mongolia, where grazing collapse did not occur.
In the steppe of neighboring Mongolia, overgrazing also affected fire activity from 1988 to 2008 (Liu et al., 2013) in a manner similar to Kazakhstan. However, extreme winter freezing and inadequate preparation affected the increasing livestock trend driven by the poorly prepared feeding of hay and foliage. This led to higher livestock reductions during the colder season than the average period of 2000 to 2014 (Nandintsetseg et al., 2018), highlighting the potential impact of climate on the livestock population in addition to human management decisions and practices (Xu et al., 2019).
We investigated grazing and land use as the main drivers of changes in fuel availability in grasslands that could abruptly impact the fire regime, as observed for Africa (Holdo et al., 2009;Andela et al., 2017) or globally over long periods (Marlon et al., 2008). Political changes can be associated with additional human processes affecting fire activity or fire spread. Among others, decreasing population density (−10 % observed in Kazakhstan after 1991) could decrease fire activity or suppression effort and firefighting capacities, as mentioned for the post-Soviet period (Mouillot and Field, 2005), as well as local conflicts potentially exacerbating fire ignitions, as observed in Africa (Bromley 2010). These effects might contribute less significantly than the direct effects of grazing and land use on fuel loading and the subsequent fire activity in the region. Gathering social information remains a challenge with respect to better understanding the human impact on fire activity.

Modeling fire and grazing interactions
Accounting for confounding factors related to burned area and the subsequent effects on ecosystems, biosphereatmosphere interactions and climate have been a challenge in developing fire modules in global vegetation models (Hantson et al., 2016). Climate (drought, temperature and humidity), land cover and fuel amount are the main drivers related to fire activity in dynamic global vegetation models (DGVMs) coupled with human-related information such as population density and countries' wealth (Gross Domestic Product). Our understanding of land use dynamics (Prestele et al., 2017), especially forest management, fire prevention and grazing practices, is still developing (Rolinski et al., 2018), and better data assemblage and modeling processes are needed (Pongratz et al., 2018). In our study, we showed the strong impact of political events (here, the collapse of the political regime) on grazing intensity and the subsequent effect on fire activity. These stochastic events are hard to forecast and simulate so that DGVMs cannot fully capture longterm trends in burned area (Kloster et al., 2010;Yue et al., 2014) when compared with observed burned area reconstructions (Mouillot and Field 2005).
The Soviet economic collapse provides fruitful information on the potential amplitude and impact of grazing changes on ecosystem functioning. The 1998 Russian financial crisis led to a dramatic decrease in the consumption of livestock in neighboring countries such as in Kazakhstan. Both sheep and goats (Fig. S2) and cattle (Fig. S3) declined substantially from 1992 to 1998. As the economy improved after late 1990s, the consumption of livestock has grown steadily. Integrating grazing into DGVMs has recently emerged for global models (Chang et al., 2013;Pachzelt et al., 2015;Dangal et al. 2017) and for local studies (Bachelet et al., 2000;Caracciolo et al., 2017;Vigan et al., 2017). Grazing processes as implemented in DGVMs can capture the climate impact on livestock populations which could be affected by climate extremes (Nandintsetseg et al., 2018) and lack of forage or water (Tachiiri and Shinoda 2012; Vrieling et al., 2016). However, they still lack abrupt and stochastic changes in the projections of socioeconomic processes or infectious disease potentially affecting livestock density, as shown in Africa by Holdo et al. (2009) after rinderpest virus curation.
Paleo-fire reconstructions have already captured grazing as a keystone process through dung fungi spore residues in sediments (Cordova et al. 2019). Current modeling develop-ments cover disease projections (Perry et al. 2013), and modeling efforts are increasingly devoted to the more complex human aspect, including human consumption behavior, inequality, and subsequent potential issues (Motesharrei et al. 2016) or social conflicts (Neumann et al. 2011). Our study demonstrates that grazing can be highly variable as a fast response to or abrupt change in agricultural policies or political regime. These abrupt changes can have a significant impact on fire activity. Better integration of human process on grazing activities in DGVMs, even as stochastic events, would capture this important process to account for probable political collapse/agricultural policies, societal decisions or widespread animal diseases. These improbable factors could affect the future global carbon budget.

Conclusions
The spatial and temporal extent of the area burned was examined in different ecosystems in northern Eurasia from 2002 to 2016. We conclude that the burned area in grasslands declined 74 %, from ∼ 282 000 km 2 in 2002 to ∼ 73 000 km 2 in 2016, or at a rate of 1.3 × 10 4 km 2 yr −1 . The area burned in forest did not show a trend. Grassland fires in Kazakhstan accounted for 47 % of the total area burned but comprised 84 % of the decline in the total area burned in northern Eurasia during the 15 years. A wetter climate and the increase in grazing livestock in Kazakhstan are the major factors contributing to the decline in the area burned in northern Eurasia. The population of livestock increased in most of Kazakhstan from 2002 to 2016, decreasing the burned area due to fuel removal from grazing. The major factors affecting the availability of the fuels resulting in the decline in burned area are climate, the proportion of the grassland cover, the aboveground net primary production and the livestock density. These factors interact to reduce the area burned in Kazakhstan, especially in grasslands.
Data availability. All data and materials are available in the paper or Supplement. The original geospatial dataset of the burned area is large and will be made available upon reasonable request. However, a derived dataset has been used to estimate black carbon emissions from fires in the same region. This dataset has been archived on the Forest Service Data Archive website (https://doi.org/10.2737/RDS-2016-0036; Hao et al., 2016b).
Author contributions. WMH led both the project and the writing of the paper. MCR simulated aboveground biomass ANPP and offered advice on the statistical analysis. LSB was responsible for the statistical analysis. YB, PC and FM suggested the use of the PDSI and livestock population to explain the declining burned areas. BLN analyzed the data and contributed certain figures. AP mapped the burned areas. REC conducted GIS analysis. SPU offered advice on the execution of the project. CY advised on the trend in the burned areas. All authors contributed to writing the paper.
Competing interests. The authors declare that they have no conflict of interest.
Acknowledgements. Wei Min Hao received funding from the US Department of State, the US Forest Service Research and Development, and the NASA Terrestrial Ecology Program. Yves Balkanski and Philippe Ciais received funding from the European Union's Horizon 2020 Research and Innovation program (grant no. 641816; CRESCENDO). Florent Mouillot received funding from the European Space Agency FireCCI program.
Financial support. This research has been supported by the European Union's Horizon 2020 Research and Innovation program (grant no. 641816; CRESCENDO), the European Space Agency (FireCCI program), the Department of State, the US Forest Service Research and Development, and the NASA Terrestrial Ecology Program.
Review statement. This paper was edited by Nicolas Brüggemann and reviewed by two anonymous referees.