Overwintering fires rising in eastern Siberia

Overwintering fires are a historically rare phenomenon but may become more prevalent in the warming boreal region. Overwintering fires have been studied to a limited extent in boreal North America; however, their role and contribution to fire regimes in Siberia are still largely unknown. Here, for the first time, we quantified the proportion of overwintering fires and their burned areas in Yakutia, eastern Siberia, using fire, lightning, and infrastructure data. Our results demonstrate that overwintering fires contributed to 3.2 ± 0.6% of the total burned area during 2012–2020 over Yakutia, compared to 31.4 ± 6.8% from lightning ignitions and 51.0 ± 6.9% from anthropogenic ignitions (14.4% of the burned area had unknown cause), but they accounted for 7.5 ± 0.7% of the burned area in the extreme fire season of 2020. In addition, overwintering fires have different spatiotemporal characteristics than lightning and anthropogenic fires, suggesting that overwintering fires need to be incorporated into fire models as a separate fire category when modelling future boreal fire regimes.


Introduction
Arctic-boreal ecosystems have been a long-term carbon sink [1], storing approximately double the amount of carbon as currently in the atmosphere [2]. Fire is the primary disturbance in boreal regions where the climate is warming faster than in other parts of the world [3]. Recent increases in boreal fires threaten the large boreal soil carbon reservoirs by direct carbon emissions from combustion [4] and indirect carbon emissions that may result from fireinduced permafrost thaw [5,6]. The burned area within a fire season usually stems from a combination of anthropogenic and lightning ignitions in the southern boreal forest, shifting towards a lightningdominated fire regime in the remote northern boreal forest [7]. Recently, researchers discovered another unexpected source of burned area, overwintering ('zombie') fires, which, in part, may be responsible for the early start of recent boreal fire seasons [8].
Overwintering fires are seemingly extinguished on the surface at the end of the boreal fire season; however, they locally smoulder belowground in carbon-rich organic soils throughout the winter. In the subsequent fire season, smouldering fires re-emerge as flaming forest fires when weather conditions favourable for fire spread arrive in spring after snowmelt (figure 1) [8][9][10]. Smouldering overwintering fires burn deep into the organic soils and may as such emit soil carbon that had been preserved over several fire cycles [9,11]. The first evidence of the widespread occurrence of overwintering fires came from Scholten et al [9] who combined records from fire managers with satellite imagery and other geospatial data over Alaska, USA, and the Northwest Territories, Canada. They found that between 2002 and 2018, overwintering fires were responsible for less than 1% of the burned area in these regions. However, after large fire years, overwintering fires accounted for more than 5% of the annual burned area in some years and more than 30% of the annual burned area in one individual year. The majority of burned area in the circumpolar boreal forest is located in Eurasia [12]. The prevalent re-emergence of overwintering fires after hot and dry summers with extreme fire activity suggests that overwintering fires may be more widespread, including in Siberia, with the ongoing intensification of boreal fire regimes [11,13].
Eastern Siberia experienced anomalously high fire activities in 2019 and 2020 [14], and the 2020 fire season spurred the first scientific discussion on the proliferation of overwintering fires in eastern Siberian fire regimes [8,10]. Still, the role of overwintering fires and their contribution to the fire regimes in eastern Siberia has remained unexplored. We here created daily burned area maps by merging the 500 m Moderate Resolution Imaging Spectroradiometer (MODIS) Collection 6 burned area product and the 375 m Visible Infrared Imaging Radiometer Suite (VIIRS) I-band active fire product between 2012 and 2020. Subsequently, we adapted and optimized an overwintering fire detection algorithm [9] to operate over Yakutia, jointly using fire, lightning, and infrastructure data (figure S1 available online at stacks.iop.org/ERL/17/045005/mmedia). For the first time, we assessed the contribution of overwintering fires to the annual burned area in Yakutia, eastern Siberia, between 2012 and 2020. In addition, we compared the spatio-temporal characteristics of overwintering fires with those from anthropogenic and lightning fires.

Study area
Yakutia, also known as Sakha, is a republic of Russia, located in eastern Siberia, and occupies most of the Northeastern part of the Eurasian continent, covering over 3 million km 2 . The whole territory of Yakutia lies in the zone of perennially frozen soils and over 40% of its territory lies within the Arctic Circle. Yakutia consists of Arctic and boreal vegetation communities. Arctic tundra vegetation covers the ecosystems near the northern coastline and in higher elevation subarctic regions. Forests dominate the central and southern parts of Yakutia: boreal forests cover the majority of Sakha and these forests are dominated by larch tree species [15]. Cold continental climate prevails in Yakutia with most of the area covered with snow for between 225 and 250 days yr −1 [15]. Yakutsk, the capital of Yakutia, has an average January temperature of −40.9 • C and is considered one of the coldest cities in the world [16]. Climate change has resulted in higher temperatures over much of Yakutia in recent decades [17,18]. For example, a new record temperature of 38 • C within the Arctic Circle was recorded in Verkhoyansk, Yakutia, in June 2020 during a prolonged heatwave that increased droughts and stimulated wildfires [14]. Forest fires in Yakutia include low-intensity surface fires as well as high-intensity canopy fires [19,20]. The majority of carbon emissions from these fires stem from burning in organic soils [12], which may facilitate the survival of overwintering fires [9].

Daily burned area maps
We derived daily burned area maps for Yakutia between 2012 and 2020 at 500 m spatial resolution by combining the MODIS Collection burned area product (MCD64A1) [21] and the VIIRS daily active fire product (VNP14IMG) [22]. The MCD64A1 burned area product has relatively large omission errors due to its coarse resolution [23,24], which affected the characterization of burned areas in Yakutia (figure 2(a)). Therefore, we complemented the burned area estimates from the MCD64A1 product with the VIIRS active fire product as it proved to fill some of the omission errors from the MCD64A1 products thanks to its high detection rate (figure 2(b)) [25]. The VIIRS active fire data also has omission errors when used to indicate burned area due to cloud and smoke cover, or when fires are spreading fastly between subsequent acquisitions [25]. As such, the MCD64A1 and VNP14IMG were used in a complementary fashion in order to minimize omission errors. We classified 500 m by 500 m pixels that included the centre of a VIIRS active fire detection but were not classified as burned in the MCD64A1 product as burned pixels, complementing the burned area pixels from the MCD64A1 product. Given the inclusion of other thermal anomalies than fire in the VNP14IMG product, we only complemented the MCD64A1 burned area with burned area from the VNP14IMG product within a 5 km buffer of MCD64A1 product. The resulting burned area in Yakutia between 2012 and 2020 was 44% larger than the burned area from the MCD64A1 burned area product alone (table S1). To separate burned area in individual fire perimeters, we created a 500 m buffer around all burned area pixels, and we attributed fragmentary burned pixels that were not attached to the core burned area but within the buffer to the fire perimeter of the core burned area (figure 2(c)). These fragmentary burned area pixels at the edges of large fire scars result from omission errors in the remotely sensed burned area product. When not accounted for, these fragmentary burned pixels can erroneously be seen as separate ignitions, and this results in overestimations of the number of fire ignitions in boreal regions in existing global fire segmentation products [26]. Due to the uncertainties in the temporal reporting accuracy of the MCD64A1 product [27], all burned pixels of our combined burned area product were assigned a day of burning from the nearest VIIRS active fire observation (figure 2(d)). When no VIIRS pixels were present within 1 km of burned pixels from the MCD64A1 product, we retained the day of burning from the MCD64A1 product. As a result, we created annual burned area progression (day of burning) maps. Large boreal fires often coalesce from multiple fire starts. We thus retrieved fire starts by searching for local minima within the day of burning maps within each fire perimeter by using a search radius of 10 km. This approach allows retrieval of multiple fire starts for larger burn complexes that originated from multiple fire starts (figure 2(d)). When the local minimum contained multiple pixels with the same day of burning, we estimated the locations of the first day of burning (hereafter referred to as fire start locations) as the centroid of these neighbouring pixels.

Ignition attributions
We derived fire start locations based on the daily burned area maps and annual fire perimeters. We used spatial and temporal constraints based on a snow cover product (MCD10A1), the Global Lightning Detection Network (GLD360) lightning data, and OpenStreetMap (OSM) infrastructure data, to attribute fire starts to lightning, anthropogenic, and overwintering fires (figure 3). We buffered all lightning strike locations in Yakutia using a 3 km buffer based on the positional accuracy of the lightning detection network [28]. Lightning fires often smoulder for several days after the strike in boreal organic soils before being detected by satellites [9]. We accounted for this holdover time when attributing fire starts to lightning by allowing a time lag of up to six days between a buffered lightning strike and a fire start [9]. We used vector data on roads and powerlines to attribute fire starts to anthropogenic fires. We classified fire starts that occurred within 5 km of roads and powerlines as anthropogenically ignited as suggested from prior work on relationships between distance from roads and anthropogenic fire starts in parts of Siberia [29,30]. Overwintering fires re-emerge early after spring snowmelt and in close proximity of the burn scar of the year before [9]. We computed maps of the annual first snow-free day (when fractional snow cover drops below 15%) in spring from the MCD10A1 snow cover product, and we calculated the local snowmelt onset day within 10 km buffers of each fire start location. Fire starts within 1 km of the burned area from the year before that occurred within 60 days after the local snowmelt onset were attributed to overwintering fires [9]. Fire starts that could not be attributed to lightning, anthropogenic and overwintering fires, remained unclassified in the unknown class.

Burned area attribution
Spatially adjacent burned areas from large boreal fires may stem from multiple different fire starts that grew together. For these fires, we estimated the contribution of each fire cause category, lightning, anthropogenic, overwintering, and unknown causes, to the total burned area by proportionally dividing the burned area in each fire perimeter among its constituting fire starts from different causes. In our initial attribution of fire causes, some fire starts were assigned to multiple fire causes. To account for the attribution uncertainty for these fire starts with multiple possible fire causes, we developed six scenarios that included all possible attribution priorities between the different fire causes (figure 3). These scenarios are only applicable to fire starts that have at least two possible fire causes. The burned area from lightning, anthropogenic, and overwintering fires was also estimated using the same scenarios We calculated the mean burned area in each fire category from the six different fire cause attribution scenarios.

Fire size distributions
Overwintering fires are known to often remain small [9], yet fire sizes from overwintering fires have not yet been compared with fire sizes from lightning and anthropogenic fires. To do so, we calculated frequency-fire size probability distributions. Frequency-fire size probability distributions follow a power-law (heavy-tailed) relationship [31,32]. The power-law takes the form of: where the number of fires per fire size bins of 1 km 2 was estimated, with β and α as regression coefficients. We estimated the distributions of small versus large fires for lightning, anthropogenic, and overwintering fires separately using fire size probability distributions. Because cumulative frequencies may obscure underlying trends in finite data sets, we followed Malamud et al [33] using the frequency densities defined as where δN is the number of fires within a bin of width δfire size. The frequency densities f (fire size) are then the number of fires per bin. We increased the bin width for fire size to approximate equal bin widths after logarithmic transformation. For each fire category, the frequency densities were plotted as a function of fire size, and we calculated the linear regressions in the log-log space: log (frequency densities) = −β · log (fire size) in which β is the regression slope, and logα is the regression intercept. The power-law exponent β quantifies the ratio of the number of large to small fires in a given fire regime, with β value of zero indicating the small and larger fires are equally represented in the fire size distribution, whereas larger β values denote a dominance of small fires in the fire size distribution [33].

Results
Yakutian landscapes include extensive larch forests and carbon-rich peatlands in the continuous permafrost zone, which feature frequent fires with a burned fraction of more than 5% yr −1 of the total area between 2012 and 2020 ( figure 4(a)). The estimated burned area in Yakutia of 7.2 Mha in 2020 far exceeded the annual burned area estimates between 2012 and 2019, and was, in part, driven by an unprecedented heatwave resulting in early snowmelt and rapid drying of fuels ( figure S2) [14,19,34]. In 2020, the mean snowmelt at locations of overwintering fires was earlier (mean day of the year = 134.0, standard deviation = 6.5 days) than between 2012 and 2019 (mean day of the year = 137.6, standard deviation = 11.3 days). Overwintering fires reemerged earlier in 2020, too, (mean day of the year = 165.9, standard deviation = 20.0 days) compared to the period between 2012 and 2019 (mean day We assessed all fire start attributions based on the spatiotemporal constraints and quantified the burned areas accordingly using the six different fire cause attribution scenarios (figure S4, tables S2 and S3). Overwintering fires in Yakutia accounted for 3.2 ± 0.6% of the total burned area between 2012 and 2020, compared to 31.4 ± 6.8% from lightning ignitions, 51.0 ± 6.9% from anthropogenic ignitions, and 14.4% of the burned area had unknown cause ( figure 4(b)). Overwintering fires contributed remarkably to the fire activity in 2020 and accounted for 7.5 ± 0.7% of the annual burned area, which was about seven times as much as the burned area from overwintering fires between 2012 and 2019. Lightning, anthropogenic, and overwintering fires occupy different parts of the landscape. Much of the burned area is concentrated between 60 • N and the Arctic Circle, where a human-dominated fire regime prevails because of human accessibility. Farther North, within the Arctic Circle, lightning fires dominated most of the landscape ( figure 4(c)). Overwintering fires were widespread and scattered along the latitudinal gradient in Yakutia. Notably, overwintering fires contributed to 9.2 ± 0.9% of the burned area within the Arctic Circle in Yakutia in 2020 ( figure 5).
Re-emerging overwintering fires demonstrated different spatiotemporal characteristics than anthropogenic and lightning fires. The overwintering fires started in early spring, around the same time as anthropogenic fires, but well before the occurrence of lightning ignitions ( figure 6). They, however, have a shorter temporal niche than anthropogenic fires, which occur almost throughout the year, but a longer temporal niche than lightning fires, which are  concentrated in summer in co-occurrence with a seasonal nadir in fuel moisture. The fire size distribution of overwintering fires was also markedly different from lightning and anthropogenic fires as denoted by the power-law exponent β, which quantifies the ratio of large to small fires [30]. A larger β coefficient denotes a larger proportion of small fires in the fire size distribution. The majority of overwintering  fires remained small (β = 1.67), while proportionally more large fires were ignited by humans (β = 1.22) and lightning (β = 0.89) (figure 7). For example, 59.2% of overwintering fires remained smaller than 2500 ha, compared to 36.4% for anthropogenic fires, and 16.1% for lightning fires. Overwintering fires remain comparatively small, possibly because of fuel limitations that constrained fire growth in the areas burned the year before. We found that 28.6% of overwintering fire starts detected by the VIIRS active fire data were unrecorded in the MODIS burned area data. Notably, overwintering fires in 2020 were clearly larger than in other years with a median fire size of 13.5 km 2 , compared to median sizes of overwintering fires of 1.5 km 2 between 2012 and 2019 (table 1).

Discussion
Boreal fires are increasing in frequency, extent, and severity as fire regimes are changing with climate warming [4]. Our work provides additional evidence that overwintering fires are an emerging property of an intensifying boreal fire regime in Siberia. Overwintering fires are mainly driven by summer temperature extremes, deep burning, and large annual fire extent influenced directly by climate warming [9]. While the contribution from overwintering fires to the total burned area is less than that of lightning and anthropogenic fires, overwintering fires extend the legacy of a large fire season into the subsequent fire season. Given their re-emergence in early spring, they can contribute to increasingly earlier starts of boreal fire seasons. This may pose additional challenges for fire management agencies that need to be operational in early spring to suppress these fires. Overwintering fires are generally smaller than lightning and anthropogenic fires. The increased sensitivity of the VIIRS active fire algorithm to small fires allowed detection of some of these small overwintering fires that are omitted by burned area products that are primarily derived from fireinduced reflectance changes at moderate resolution (e.g. 500 m) [9]. The inclusion of VIIRS active fire data thus allowed improvements in characterizing spatiotemporal characteristics of overwintering fires, and a combined virtual constellation of higher resolution polar-orbiting satellites (e.g. Sentinel-2 and Landsat 8/9) may help detect more overwintering fires with fainter thermal anomalies [35,36], and possibly earlier in the season, in future work.
McCarty et al [8] and Irannezhad et al [10] suggested that overwintering fires may have contributed to an early and unprecedented fire season in the Arctic in 2020. We found evidence that overwintering fires in 2020 notably deviated from overwintering fires between 2012 and 2019 because of their earlier reemergence, larger size and as a consequence larger burned area. This suggests that the extended spring heatwave and drought over Yakutia (figure S2) with consequent early snow melt (figure S3) may have facilitated the early re-emergence of overwintering fires. The large fire sizes of overwintering fires in 2020 compared to overwintering fires between 2012 and 2019 also suggests that the drought conditions in early 2020 may have enabled these overwintering fires to grow larger, despite being partly surrounded by areas that had burned in 2019. Future work could focus on how the growth of these overwintering fires has developed in relation to the availability and moisture conditions of fuels, and fire weather conditions such as vapour pressure deficit, and wind speed and direction. Such an analysis will necessitate a combination of high-resolution datasets on topography, fuels, and fire weather.
Overwintering fires result from deep-burning and smouldering in organic soils, and may thereby influence soil functioning and post-fire recovery trajectories [37]. In addition, overwintering fires may emit significant amounts of carbon [9], but the extended smouldering phase also increases the likelihood of these fires emitting relatively more methane compared to flaming fires [38], further exacerbating climate feedbacks. So far, the phenomenon of overwintering fires has mainly been observed and investigated from satellite imagery. A better understanding of the occurrence and carbon cycle impacts of overwintering fires will need to include detailed in situ measurements at locations where overwintering fires hibernated and re-emerged. Such measurements could include assessing local topographic drainage conditions, soil bulk density, carbon content, and emission measurements when overwintering fires are in a smouldering phase. Given the remote occurrence of overwintering fires, these measurements can be constrained by logistical challenges, which may overcome by close collaborations with local communities near areas where overwintering fires have re-emerged.

Conclusions
Although overwintering fires were responsible for a smaller proportion of burned area compared to anthropogenic and lighting fires in Yakutia, eastern Siberia between 2012 and 2020, our work highlighted the importance of overwintering fires and their contribution to the elevated burned area in eastern Siberia in 2020. Given the differences in timing, location of affected landscapes, and fire size distributions of overwintering fires compared to anthropogenic and lightning fires, we call to include overwintering fires as a separate fire cause category in fire models. This study provides the first quantification of the role of overwintering fires in eastern Siberia. Such information is important to further optimize predictions of the fate of overwintering fires in the rapidly changing boreal biome.

Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: https://doi.pangaea.de/10.1594/PANGAEA.938118. Science Center. We thank the following institutions/individuals for their public data set. MODIS burned area product (MCD64A1) are available from United States Geological Survey (USGS) (https://lpdaac. usgs.gov/products/mcd64a1v006/). The VIIRS active fires product (VNP14IMG) is available from Fire Information for Resource Management System (FIRMS) (https://earthdata.nasa.gov/firms). Lightning data were retrieved from the global lightning detection network GLD360 developed by Vaisala (www.vaisala.com/en/products/systems/ lightning/gld360) for which permissions can be obtained by Vaisala. Infrastructure data are available for Yakutia from the OpenStreetMap Data Extracts (https://download.geofabrik.de/index.html). The snow cover product (MCD10A1) and derived snow melt dates are available through Google Earth Engine (GEE) and processed based on a script adapted from the GitHub account provided by Dr Koen Hufkens (https://github.com/khufkens/ MCD10A1/tree/v1.1).