Fire as carbon sink? The global biome-dependent wildfire carbon balance

19 20 Wildfires generally result in biospheric recovery approximating the pre-disturbance 21 state. However legacy carbon(C) gains and losses that have until now been overlooked in 22 global-scale theory and modelling indicate that post-fire C gains through pyrogenic 23 carbon (PyC) production, and losses via fire regime shifts, post-fire mortality, topsoil loss 24 and inland water export, may be central to whether 20 th century fires have imposed a net 25 terrestrial C source or sink. Here, we integrate PyC production and soil accumulation 26 into a global terrestrial model (ORCHIDEE-MICT) and estimate wildfire C-gains and 27 losses over 1901-2010, quantifying the fire-C balance at global, regional and vegetation 28 scales. Excluding the effect of PyC mineralisation, fires provide a land storage of +177 29 TgC yr -1 (63% PyC production), dominated by grasslands. The global balance is 30 nuanced,

global-scale theory and modelling indicate that post-fire C gains through pyrogenic 23 carbon (PyC) production, and losses via fire regime shifts, post-fire mortality, topsoil loss 24 and inland water export, may be central to whether 20 th century fires have imposed a net 25 terrestrial C source or sink. Here, we integrate PyC production and soil accumulation 26 into a global terrestrial model (ORCHIDEE-MICT) and estimate wildfire C-gains and 27 losses over 1901-2010, quantifying the fire-C balance at global, regional and vegetation 28 scales. Excluding the effect of PyC mineralisation, fires provide a land storage of +177 29 TgC yr -1 (63% PyC production), dominated by grasslands. The global balance is 30 nuanced, with forest fires resulting in strong terrestrial net C loss:gain ratios (>-2:1) that 31 are greatest in tropical regions (>-3:1). Frequent tropical grassland fires are responsible 32 for the bulk of the land PyC sink and its environmental persistence, whose theoretical 33 minimum mean residence time we quantify at 2760yrs. We highlight the dependency of 34 the global fire-C balance on vegetation coverage and the potential role of preserving 35 grasslands, particularly those in the tropics, in that regard. 36 37 Wildfires are a key driver of disturbance-recovery cycles in many regions of the world. While 38 fires emit large quantities of C to the atmosphere (~2 PgC yr -1 ) 1 , subsequent vegetation 39 recovery re-captures the emitted C on decadal timescales and results in an uncertain but likely 40 small net impact on atmospheric C in the long run. Natural shifts in fire regimes and vegetation 41 occur infrequently and are largely driven by climatic and human perturbations, such that 42 biomes tend towards quasi-steady state outside of these. It is thus assumed that on decadal to 43 centennial time-, and biome to global spatial-scales: 44 45 (1) 46 47 Where ()* " is fire CO2 emissions due to vegetation combustion, and +)* " is uptake of 48 atmospheric CO2 by post-fire vegetation recovery, poles referring to flux direction with respect 49 to C stocks in the terrestrial biosphere.However, recent research on both sides of the flux 50 complicates this perspective. A range of long-term 'legacy' C fluxes traced back to source fire 51 events lead to either C accumulation or loss by land ecosystems, however their balance is yet 52 to be determined. On the 'legacy sink' side, the charring of biomass by fire creates a by-53 product known as pyrogenic C (PyC) (~10-20% annual fire CO2 emissions) 2,3 which is 54 significantly more resistant to biochemical oxidation than bulk soil organic carbon (SOC) 4-6 . 55 Most studies find that PyC degrades with a 'mean residence time' (MRT) of 100s to 1000s of 56 years 7-10 (1-2 orders of magnitude higher than non-PyC SOC), suggesting a sequestration flux 57 from the atmosphere which exceeds the temporal boundaries of the fire-recovery cycle in most 58 fire regimes, driving long-term terrestrial PyC accumulation ( ,-) ' ). Of the PyC produced, 59 we assume a fraction additional to ,-) ' consists of a lightweight 'labile' component that is 60 likewise readily mobilised, hereafter denoted The magnitude of the global PyC sink is dependent on its production rate and MRT with respect 63 to degradation processes, which is a positive function of maximum flame intensitiesy 4,9 related 64 to biomass loading and moisture content 11 . 65 66 Thus, frequently-burned biomes like grasslands can be expected to host low-intensity fires that 67 produce relatively labile PyC compared to that of forests which host less frequent and more 68 intense fires due to lower and generally wetter fuel stocks. A trade-off between PyC input and 69 MRT determines PyC storage across biomes: Increasing grass cover leads to more frequent 70 PyC production, however this PyC tends to be relatively labile (Fig. 1). 71 72 On the 'legacy source' side, fires impose long term C deficits on the terrestrial biosphere via 73 several mechanisms: First, the return of biomes to their pre-fire biomass state (Fig.1a) requires 74 a stable fire regime, in which the biomass recovery interval (BRI; the time period of complete 75 vegetation recovery) is shorter than the fire return interval (FRI; the time between fires). 76 Violation of this condition (BRI < FRI) entails a change in the fire regime and an overall C-77 deficit, representing a step-wise decrease in biomass C ( ∆(1234 ) and land coverage (Fig. 1a). 78 Second, tropical rainforests exposed to pre-fire disturbances such as drought 12 are vulnerable 79 to episodes of aboveground vegetation-C (VC) mortality ( 5678 ) in the decades following 80 fires ( 5678 can exceed 25% of VC) 13,14 . Third, in areas where the existing fire regime results 81 in higher fire frequencies than the average for that vegetation type, large fractions of SOC can 82 be lost through combustion, erosion and microbial mineralisation ( 9:;6<< ). Topsoil SOC 83 losses through this mechanism have been observed to exceed 20% on average in grasslands 84 and broadleaf forests 15 . Fourth, PyC is liable to export from land to oceans via rivers 85 ( ,-)=>? % ) in particulate or dissolved form (Py-POC; Py-DOC, respectively), totalling >40 86 TgPyC-C yr -1 7,16 . A fraction of this exported PyC is later deposited to the ocean floor up to 87 1E+04 yrs 7 after its initial production, with some proportion of photo-oxidative degradation 88 occurring en route 17 . Although considered as a loss term in the terrestrial C budget equation 89 below,      ' is an absolute 158 quantity, meaning simply that the more fire there is the more PyC is injected into the global 159 soil mass (Fig.1). Due to high fire frequency and recovery rates, grassland biomes are both the 160 main source of PyC globally (~250 TgCyr -1 ), and, compared to other vegetation types, pull the 161 relative sink and source terms of Eq. 2 towards the former. 162 163

PyC MRT ; PyC production
In grasslands, simulation-average annual PyC production fell globally by 8% relative to the 164 simulation mean pre-and post-1930, when conversion to agriculture or plantation forest during 165 the 'Dust Bowl' era resulted in a reduction in grassland cover in the Great Plains by up to -166 96% 26,37 , with subsequent distributional changes in fire and PyC production (Fig. S4). The 167

(a) (b)
- 5  -4  -3  -2  -1  0  1  2  3  4  5  6  7  8  9   80  respectively. Our vegetation maps show -21% (-1. 16 Mkm 2 ) and -12% (-3.6 Mkm 2 ) net 170 declines in C3 and (tropical) C4 grasslands between the first and last decades of simulation 171 (Fig. S5), leading to global, correlated decreases PyC production (Fig. 3i, Fig S6), in spite of 172 global forest PyC production doubling over the same period (Fig. S3). This is consistent with 173 a generalised shortening of fire return intervals over the 20thC (Fig. S7) and fire duration post-1940 (Fig. S9). The net effect of these dynamics has been to decrease 186 the partial fire C sink by up to half over the 20 th Century (Fig. S10). 187 188 Emerging Constraints on PyC Storage 189 190 We approximate the vegetation-specific potential in situ MRTs of PyC by considering mean 191 and maximum flame temperatures for each vegetation type (Table S2; see Methods). 192 Surprisingly, we find that Tropical (C4) grasslands and savannah regions host some of the 193 highest maximal and mean flame intensities, which confers correspondingly high PyC MRTs 194 ( Fig. 1), in contrast with C3 grassland fires whose intensities are lower than most vegetation 195 types. The data also indicate that fires tree-dominated vegetation types are generally more 196 intense than in grasslands despite the absence of crown fire representation, which can release 197 very large amounts of energy 11,42,43 , in ORCHIDEE 24 . This implies that the grass-forest 198 dichotomy for PyC MRT proposed here may only hold for temperate grasslands (Fig.1, Table  199 S2, Text S6), with tropical grasslands producing not only the highest quantity and but also 200 some of the most recalcitrant PyC. Note that MRT is a global average, aggregating the broad 201 variation in spatial (lateral/vertical) residence time distributions globally, as per the literature. 202 203 We is dotted to highlight that it is a terrestrial 218 export flux, not an atmospheric flux. Bottom panel shows aggregations of (h) fire C losses (sum of (càf)) and (i) 219 the fire C balance net of PyC mineralisation, i.e. the sum of (aàb) and (càf). Maximum mineralisation (g) is thus 220 equal to the residual, (i).

Regional Distribution of Fire-induced C Sinks and Sources 222 223
While absolute fire C gains and losses are highest in the tropics (Fig. 4a), the highest ratio of 224 gains to losses (efficiency of the fire C sink) lies in the temperate north (30-60N), with a factor 225 of ~3, and lowest (factor 1.6) in the tropical south (0-30S). The positive partial balance is 226 strongest in the south, consistent with the quantity of PyC production, with some negative 227 balance areas occurring as a result of combined FRI decrease and topsoil loss (Figs. 4c, S4-228 S7). However, the partial balance over all vegetation types masks the fact that when 229 considering grasslands and forests separately, the latter on aggregate exhibit strongly negative 230 partial balance estimates, with the greatest negative factorials in the southern hemisphere at 0-231 30 and 30-60S (factors of -3 and -8, respectively, Fig. 4b), providing powerful suggestive 232 evidence that grasses and forest play contrasting functional roles with respect to fire in the 233 terrestrial C balance. 234 235 Overall, over the period of our simulation, our results indicate that when aggregated over space 236 and vegetation types, fires may provide a long-term C sink made possible by the buffers of 237 PyC production 2 and biospheric uptake, provided that the fire regime remains in state that does 238 not lead to ecosystem degradation from failed recovery (FRI≥BRI

Annual C fluxes relative to mean PyC Production
? ? ?

Annual C fluxes relative to mean PyC Production
? ? ?

Annual C fluxes relative to mean PyC Production
? ? ?

253
The balance is calculated as the net sum of Eq. 2, excluding the PyC mineralisation term ( 012 , ), which is unknown.

254
Re-thinking the Role of Fire in the Carbon Cycle 255 256 Our work indicates that net of legacy fluxes fires in may be a force for Earth System C equilibrium, 257 provided that they ignite over biomes (e.g. grasslands) that carry evolutionary adaptations to cope 258 with them. This is somewhat counterintuitive given the dramatic nature of wildfires and their 259 large initial emissions, which feed a perception of destabilisation 48 . Forest fire-induced C 260 losses are compensated globally by the dynamics of grassland ignition, despite considerable 261 post-fire grassland SOC losses (~52vs.5 TgC yr -1 for grass, forest respectively). 262 263 We estimate that fire regime change-induced C loss ( ∆(1234 ) in grasslands is less than that 264 associated with forests (~20 vs 26 TgC yr -1 ). This result is afforded by grasses' capacity for 265 fire recovery 49-53 , providing large stocks of aboveground fuel (potential PyC), thereby 266 mitigating losses and maximising soil PyC gains. Thus, PyC production is less correlated to 267 flame intensity/temperature for grassland than forests (Fig. S15). In the absence of grasslands, 268 fire phenomena would impose a net terrestrial C source. This aligns with the proposition that 269 the co-evolution of grassland fire and herbivory led to the formation of PyC-rich Mollisols that 270 may have been central to climatic cooling since the Mid-Cenozoic (~40-0Ma) 35 . Conversely, 271 the fire-as-destabilisation perspective is justified in biome-specific cases, generalisable to non-272 grass (forest) fires (Fig. 4b). 273 274 In forests, fire C losses can overwhelm PyC gains even without considering PyC 275 mineralisation. This is not surprising, particularly in the humid tropics, where tree species are 276 ill-adapted to catastrophic disturbance 54,55 and massive post-fire mortality is commonplace 12-277 14,54 (Fig. S14). Similarly, subtropical and semi-arid regions may be more prone to fire events 278 of higher intensity 11 , resulting in topsoil loss and an incapacity for biomass recovery, in 279 addition to aquatic outflows of C (Fig 4a, S13). 280 281 The dependency of the partial fire C balance on fire-affected vegetation composition has 282 important implications for a world in which the frequency and intensity of droughts, heatwaves 283 and wind extremes 56-64 , are forecast to increase 65-70 , potentially increasing all terms in Eq.2. 284 The preservation and restoration of native grasslands may be seen as an important vector for 285 increasing C stocks/decreasing C losses from future fire activity, and would apply to both 286 temperate and tropical systems, given the efficiency of the former as a C-sink (Fig.4b) and the 287 contribution of the latter to PyC production (Fig. S3) and bulk PyC MRT (Table S2) widely used sub-branch of ORCHIDEE that is global in scope but includes some soil, hydrological and thermal 507 processes specific to boreal regions 1-3 , whose use here will facilitate future assessments of PyC stocks in deep 508 permafrost soils. At the core of the model is terrestrial biomass fixed by photosynthetic C uptake, performed by 509 13 plant functional types (PFTs) with distinct primary production, senescence and carbon dynamics 4 . Biomass is 510 allocated to foliage, fruit, roots, above/below -ground sap, heart wood and carbon reserves which are transfered 511 to two reactivity-differentiated litter pools. ORCHIDEE-MICT is integrated with a model-specific version (see 512 ( 5,6 ))of the SPITFIRE fire module 5,7,8 , which takes the aboveground portion of these biomass components and 513 allocates them to potential fire fuel classes differentiated by their potential time to combustion/oxidation.

514
ORCHIDEE-SPITFIRE has been involved in multiple phases of FireMIP 9 and its predictions found to be within

536
C mass balance is maintained by removing PyC produced from other C pools. PyC produced is first subtracted 537 from the fraction of biomass going to litter pools in that SPITFIRE timestep (1 day). If PyC produced > biomass 538 going to litter in that timestep, then the remaining quantity is taken from CO2 emissions, whose reduction 539 recursively reduces total PyC production. PyC is then introduced to the biosphere-pedosphere interface by its

555
The simulations used for this study were forced with imposed historical 13-PFT vegetation (ESA-LUH2 v1.

565
PFT-specific FRI is defined as the interval between consecutive fires affecting a consistent area, which is not a 566 standard output of ORCHIDEE and so was determined probabilistically. To do so, first we find the annual 567 fractional fire contribution of each PFT ( 0&5 ) to total CO2 emissions:

571
From this the probabilistic fire incidence per PFT, pixel and year can be estimated:

575
Where ( ) 0&5 is the annual probabilistic fire incidence per PFT and pixel, and 0'; the fraction of each 576 PFT occupied by vegetation from a given PFT (

612
Loss of biospheric C due to fire regime change ( ∆ ) This loss term is calculated for each PFT and includes net C losses from areas where the biospheric disturbance 615 steady state condition is not satisfied (BRI<FRI) as a result of a change in fire regime. We treat areas that 616 experienced decreases in FRI of >10% between the first and last three decades of the simulation (Fig. S7), as 617 having exhibited a fire regime shift. Then, we estimate the system biomass loss per fire event for these areas as

645
To capture only those areas that may have experienced drought and hence drought induced fire mortality, we 646 assume that drought occurs in a pixel if annual precipitation for that pixel is at or below the 25 th percentile of 647 precipitation for that pixel over the entire simulation period and mask out pixels in the dataset TE 0&5 which 648 do not satisfy this condition. We then assume that total post-fire mortality loss is approximated from the mean 649 literature value of -24.8% (±6.9%) 19 and define this fraction as the total C-loss. However, since this biomass loss 681 estimated that boreal, tropical (<30° N/S) and temperate regions export 3.8 (± 0.6), 12.4 (± 4.9) and 1.8 (± 0.8)

685
To integrate ∑`' $ with model output we estimate the contribution of each PFT to global PyC slow and PyC

697
We assume that Py-POC export occurs proportionally to Py-DOC export based on their literature-reported global 698 export rates, such that total Py-POC+DOC export occurs at a rate 2.39 ( = (18+25)/25) times that of Py-DOC. The

699
total Py-SOC that is hydrologically mobilized from each soil pool ( .