Interactive comment on “ Pyrogenic carbon from tropical savanna burning : production and stable isotope composition ”

Abstract. Widespread burning of mixed tree–grass ecosystems represents the major natural locus of pyrogenic carbon (PyC) production. PyC is a significant, pervasive and yet poorly understood "slow-cycling" form of carbon present in the atmosphere, hydrosphere, soils and sediments. We conducted 16 experimental burns on a rainfall transect through northern Australian savannas with C4 grasses ranging from 35 to 99% of total biomass. Residues from each fire were partitioned into PyC and further into recalcitrant (HyPyC) components, with each of these fluxes also partitioned into proximal components (>125 μm), likely to remain close to the site of burning, and distal components (


Introduction
Pyrogenic carbon (PyC) describes carbon (C) in a continuum of thermally altered materials produced by incomplete biomass combustion and ranging from partly charred organic matter to condensed polyaromatic compounds (Hammes et al., 2007;Masiello, 2004).Components of the PyC continuum have been referred to by a variety of other terms, including soot, char and black carbon (Bird and Gröcke, 1997;Seiler and Crutzen, 1980).This range of names reflects the compositional complexity of PyC and the wide array of analytical methods employed for its quantification (Hammes et al., 2007;Ascough et al., 2009).While there is wide recognition of the large relevance of PyC affecting environmental processes (Bird et al., 1999;Cope and Chaloner, 1980;Lehmann et al., 2008;Zimmermann et al., 2012), it remains a poorly understood component of the global carbon cycle (Forbs et al., 2006).Estimates of global PyC production rates have varied considerably since first reported (Seiler and Crutzen, 1980).The difficulty in establishing a reliable estimate of PyC production rates arises from the range of definitions used for PyC, usually operationally defined by the analytical technique used, and also by widely divergent experimentally derived estimates of PyC production (Forbes et al., 2006).Calculations of the size of a global atmospheric C sink to PyC are also complicated because some components of PyC appear to be susceptible to degradation on comparatively short timescales (Bird et al., 1999;Zimmermann et al., 2012), while some are resistant to degradation, remain-G.Saiz et al.: Pyrogenic carbon from tropical savanna burning ing in soils and sediments for thousands to millions of years (Cope and Chaloner, 1980;Lehmann et al., 2008;Masiello and Druffel, 1998).While the use of different techniques does indeed lead to different results, it has become quite apparent over the years that the PyC continuum is also associated with a PyC degradation continuum (Bird et al., 2015).
Of the total PyC produced by biomass burning, it has been estimated that > 90 % remains (initially) close to the site of production and < 10 % is emitted into the atmosphere as aerosols (Kuhlbusch et al., 1996), with much of the PyC emitted into the atmosphere being ultimately deposited in the oceans (Bird et al., 2015).PyC that remains on the ground will potentially (i) be re-combusted in subsequent fire events (Saiz et al., 2014;Santín et al., 2013) (ii) be re-mineralised by biotic/abiotic processes (Saiz et al., 2014;Santín et al., 2013;Zimmerman, 2010;Zimmermann et al., 2012), (iii) be remobilised by bioturbation, wind or water in either particulate (Major et al., 2010;Rumpel et al., 2006) or dissolved form (Dittmar, 2008;Dittmar et al., 2012) and/or (iv) accumulate in the soil organic carbon (SOC) pool (Lehmann et al., 2008).
With fire return intervals of 5 years or less (Furley et al., 2008), tropical savannas are major loci of global annual PyC production.Forbes et al. (2006) estimated the annual PyC production rate from savanna fires to be 4-40 Tg yr −1 (of a global total of 50-270 Tg yr −1 ), assuming 1-2 % of carbon exposed to fires was converted to PyC.Van der Werf et al. (2010) estimated that 39 % of a total of 2.1 Tg yr −1 PyC emissions into the atmosphere derived from the burning of tropical savannas and grasslands (including extra-tropical grasslands; http://www.falw.vu/),whereas Bond et al. (2004) estimated savanna burning at 51 % of total "open-burning" emissions.
In tropical savannas, woody vegetation uses the C 3 photosynthetic pathway (δ 13 C < 24 ‰), whereas grasses primarily use the C 4 photosynthetic pathway (δ 13 C values ≥ 15 ‰) (O 'Leary, 1988).Randerson et al. (2005) estimated that 31 % of global fire emissions have a C 4 origin and that 20 % of total C 4 biomass returns to the atmosphere each year as a result of fire.The production of PyC from C 3 biomass generally has little impact on the δ 13 C of the PyC relative to the precursor biomass (< ∼ 1 ‰; Krull et al., 2003;Das et al., 2010).However, Krull et al. (2003) found that δ 13 C values of PyC from C 4 grasses were 1.5-5 ‰ lower than precursor biomass.Das et al. (2010) also reported PyC products with δ 13 C values lower than original biomass, but noted a difference between "smoke" (δ 13 C value lower by 0.5-7.2‰, species dependent) and "ash" (generally lower by up to 3.5 ‰).Therefore, while δ 13 C values provide a useful tracer of the fate of treeand grass-derived PyC both during and after formation, the robust interpretation of δ 13 C values requires a more nuanced understanding of the processes controlling fractionation effects.
At the global scale, models of terrestrial 13 C discrimination suggest about one-quarter of the gross primary produc-tivity (GPP) of the terrestrial biosphere is attributable to the C 4 photosynthetic pathway (Lloyd and Farquhar, 1994;Still et al., 2003).In contrast, recent work has suggested the fraction of C 4 -derived biomass in SOC in savanna systems is much lower than GPP estimates would imply, partly due to C 4 -derived C decomposing faster than C 3 -derived C in the soil (Wynn and Bird, 2007;Wynn et al., 2006).However, it is also plausible that during savanna burning (i) C 4 carbon is preferentially combusted relative to C 3 woody biomass and/or (ii) C 4 biomass produces finer PyC particles than C 3 woody biomass, and thus C 4 -derived PyC is more likely to be exported by wind, leading to a relative accumulation of C 3derived PyC at the site of burning (and ultimately in the soil), with C 4 -derived PyC preferentially accumulating in sites remote to burning and particularly in lacustrine sediments and the ocean (Bird and Cali, 1998).Differences between the δ 13 C values of C released into the atmosphere vs. δ 13 C values of the C as it was fixed, or vs. the δ 13 C value of the SOC pool, are referred to as "isotopic disequilibria" in mass balance models of atmospheric CO 2 (Alden et al., 2010;Enting et al., 1995).In savannas, large disequilibria are largely due to differences in the residence time of C 3 and C 4 biomass (Randerson et al., 2005;Ciais et al., 1999Ciais et al., , 2005;;Buchmann and Ehleringer, 1998).
Quantification of the contribution of savanna fires to terrestrial isotopic disequilibrium due to potential differences in the residence times of C 3 -and C 4 -derived C in vegetation and soil is important for (i) correctly interpreting soil and palaeosol carbon isotope data (Cerling et al., 2011), (ii) informing modelling studies that use variations in the CO 2 δ 13 C record to apportion sources and sinks of CO 2 (Randerson et al., 2005;Ciais et al., 1999Ciais et al., , 2005) ) and (iii) enabling an assessment of physical redistribution biases that might complicate the interpretation of the δ 13 C record of PyC in terrestrial, lacustrine and marine sedimentary records (Bird and Ascough, 2012).
In this paper, we use a series of controlled field burning experiments in four savanna environments of northeastern Australia, where the strong climatic gradient exerts a major influence on vegetation structure existing across this large region (Table 1, Figs. 1-2).The experiments were designed to quantify the production and fate of PyC during savanna burning.Specifically, we seek to determine (i) the proportion of savanna biomass that is converted to PyC during burning, (ii) the distribution of this PyC between different size classes and (iii) the proportion of PyC fraction remaining on the ground against that exported as fine particulates during burning.
Our study takes advantage of controlled conditions, and a purpose-built field apparatus designed to determine the stocks and fluxes of PyC from fires occurring in a variety of savanna types (Fig. 3, see Sect.2.2).In addition, we used mass and carbon isotope balances to test the hypothesis that there is a "savanna isotope disequilibrium effect" (SIDE) during burning, which leads to (i) overall more negative δ 13 C values for PyC produced in savanna fires due to the more   proved understanding of PyC dynamics, and for better interpreting records of biomass burning, climate and vegetation change in the past (Bird et al., 2015;Wurster et al., 2012).Here, we use the term PyC to describe all post-fire carbon, which in our experiments might also include non-thermally altered material.

Study sites
The study sites were located across four contrasting savannagrassland ecosystems in Queensland (Australia).The sites extended from the drier inner region to the more humid environments located near the northeast coast (Fig. 1).The different climatic conditions prevailing at the various locations strongly influenced both the species and structural composition of these ecosystems.While there are many definitions of "savanna" (Torello-Raventos et al., 2013;Domingues et al., 2010), the functional definition used for the purpose of this study is that of a biome consisting of a continu-   Total organic carbon (TOC) and HyPyC are determined from the mass of the material collected in the 1 m 2 burned area for each of the fractions, and their C content measured before and after the HyPyC treatment.Non-HyPyC is further distinguished between TOC and PyC for organic carbon (OC) measured before and after the fire, respectively.Burning locations are indicated by (G) and (T) for grassand tree-dominated localities, respectively (see Materials and methods).Particle size fractions of each sample were summed into (A) pre-fire ground, i.e. material that was collected from the surface before the fire, (B) post-fire proximal, i.e. material that was collected from the soil surface after the fire, combined with the > 125 µm fraction of material collected at the filter collected at the chimney outlet.Notes: 1 mass balance in this experimental burn was not possible as the extraction pipe accidentally detached from the structure during the fire. 2 Sample was contaminated during manipulation and storage.
ous graminoid component coexisting with woody plants at varying densities.The regional classification of vegetation was done according to Torello-Raventos et al. (2013).We chose a wide climatic range in order to provide the experiment with a broad spectrum of woody vs. grass biomass proportions (woody is equivalent to C 3 and grass to C 4 ).Indeed, there was a noticeable trend both in total biomass and in the relative contribution of woody (C 3 ) vegetation to the total biomass across the climatic gradient (Tables 1  and 2).The transect ranged from dry Mitchell grasslands (site: MIT; MAP = 435 mm), characterised by > 97 % C 4 grasses, to more humid tall savanna woodlands (site: DCR; MAP = 2050 mm).The tall savanna woodlands exhibit a larger net contribution of 50-60 % C 3 vegetation to the total biomass.While C 4 biomass was relatively constant across all sites, woody biomass increased with MAP (Fig. 2, Tables 1  and 2).Tree canopy cover was determined by means of sitespecific allometric equations and visual estimates (Torello-Raventos et al., 2013;Domingues et al., 2010), and ranged from < 5 % for a site established in a heavily dominated grassland ecosystem (Mitchell grassland -MIT) to 55 % for a tall savanna woodland located at about 20 km from the coast (Davies Creek National Park -DCR) (Fig. 1).
The biomass present at all fire experiment areas was mainly composed of a grass layer dominating with sparse Eucalyptus and Acacia tree seedlings with variable amounts of coarse woody debris and leaf litter.The most abundant grass species were Themeda australis, Imperata cylindrica and Heteropogon contortus.

Plot set-up and pre-burn sampling
The experimental set-up was designed to provide an estimate of initial total organic C, PyC abundance and δ 13 C values, serving thus as reference to compare against post-burn samples.Burning experiments were carried out on a small scale (m 2 ), and were conceived to ensure the capture of all particulates to enable a full isotope mass balance.As such, the experiments were not explicitly designed to simulate all characteristics of a "natural" burn, but given the methodological issues surrounding all such techniques, the estimates of PyC production derived from this work are unlikely to be more uncertain than those obtained with other systems (see Forbes et al., 2006, for a review).Finally, we did not consider or include burnt coarse woody debris, making the reasonable assumption that most PyC production in savanna fires occurs close to the ground from grass, litter and shrubs (Randerson et al., 2005).
A total of four burning experiments were conducted at each of the four study areas, with specific assessments of both initial biomass and PyC inventory carried out before each burn.At a local scale, the abundance and δ 13 C values of litter and standing biomass are quite heterogeneous in tropical savannas.This heterogeneity is largely controlled by the distribution of trees, with lower average δ 13 C values for litter and soil C around trees compared to grass-dominated areas away from the influence of trees (Wynn et al., 2006;Saiz et al., 2012).Therefore, duplicate burns were conducted at locations at half crown distance from trees (Tree; T locations), while two additional burns were also conducted at locations remote from trees (Grass; G locations).Biomass was quantified in the near vicinity of each planned burn location by means of destructively collecting all aboveground plant material from two 1 m 2 census quadrats.In order to estimate PyC lying on the soil surface, each quadrat was subsequently vacuumed using a DC 23 Motorhead vacuum cleaner (Dyson Appliances Ltd., NSW, Australia).This task was systematically performed by the same user and over the same length of time across all burning experiments to allow for inter-comparison of results.Both the biomass and the vacuumed material were stored in separate labelled plastic bags.This sampling procedure was carried out in duplicate at each burning location, which resulted in a total of 8 × 1 m 2 census quadrats being sampled per studied site (Fig. 3).

Experimental burns
Four small-scale (1 m 2 ) burning experiments were conducted at each of the four study areas on sites with no recorded fires for 2 years during the late dry season of 2011.This period represents maximum fuel load available throughout the year as a result of the presence of senesced grasses that have already reached their full growth potential, and the fact that a number of tree species preferentially shed their leaves at this time.All the burns took place on days with no prior recorded rainfall for at least 2 weeks and shortly after midday, as this time generally corresponds to maximum daily temperature (T ).This combination of factors results in optimum condi-tions for the occurrence and spread of fire in these ecosystems.At the time of the burns, wind speeds were lower than 10 km h −1 in all cases, air T ranged from 24.2 to 40.0 • C, and relative humidity varied from 45 to 17 %.
A rectangular area of approximately 3 × 2 m was consistently left undisturbed at the centre of each planned fire.At this location, a purpose-built stainless steel structure was deployed to allow for both the assessment and containment of the experimental burning (Fig. 3).The final placement of the structure was chosen considering both the prevailing wind and topography with a view to strategically allocate an untouched area of vegetation (> 1 m 2 ) over which the fire could be ignited promoting a natural progression of the flames towards the unit.Once the structure was positioned on the ground, a 10 m radius buffer zone was cleared to the ground (Fig. 3).
The structure consisted of a series of leaning panels made up of lightweight stainless steel, which was several times the volume occupied by the enclosed vegetation.The leaning panels provided extra volume for an optimal collection of smoke and airborne particles, with their lower sides positioned at 15 cm above ground-covering metal sheets to ensure a combustion with no artificial oxygen limitation.The front panel of the unit was vertical and was open in its lower section to allow the entrance of the fire.The structure was covered by an enclosed pyramid-like chamber culminating in an outlet to a 15 cm diameter steel tube connected ultimately to a Romac 2042K particle extraction unit (0.755 m 3 s −1 , 2 hp motor; Ron Mack, Perth, WA, Australia).The extractor unit was in operation during the entire course of the burn and was only disconnected after the cessation of any noticeable smouldering.A 125 µm stainless steel sieve was installed in the flow path at the outlet of chamber.This mesh size was chosen to conform to standard dimensions used to differentiate between micro-and macro-charcoal in palynological studies (e.g.Blackford, 2000, andHaberle, 2005).Smaller particles contained in the airflow downstream of the first mesh were then subsampled and semi-quantitatively estimated using a particle collector and a constant flow sampling pump (Quick Take 30, SKC Inc., Eighty Four, PA, USA).
Continuous air flow monitoring was performed on a vertical pipe positioned after the extractor unit by means of a calibrated 160S series "S" type Pitot tube (Dwyer Instruments, Inc., Michigan City, IN, USA) connected to a Testo 435-4 multi-function instrument (Testo AG, Lenzkirch, Germany).Fire T was recorded by means of K-type wire thermocouples connected to a data logger (Simple Logger II L642, AEMC Instruments, Dover, NH, USA) logging at 5 s intervals.The thermocouples were placed at the centre of the burned area at 0.02 m above the ground, and at the outlet of the chamber.Fire residence time was defined as the length of time that air T at the chimney's outlet exceeded 100 • C.This temperature threshold accounts for most of the particulate emissions derived from a flaming front, and it is also likely to encompass the majority of smouldering emissions in grass-dominated ecosystems.
The fire was allowed to burn the biomass contained within the unit without any external intervention.After the flames self-extinguished, the leaning side panels of the structure were lowered to the ground to minimise any lateral export of the burned material.Once the unit cooled down, any remaining stubble within the enclosed burn was cut at ground level, and the same vacuum procedure employed on the two adjacent biomass quadrats was used again to determine total organic carbon (TOC) and PyC remaining on the soil surface.In order to quantify and analyse the fine soot produced during each burn, all the panels making up the structure were dismantled and thoroughly cleaned by manually brushing with a 1 : 1 water / methanol solution, with the resultant mixture stored in a glass flask.Likewise, particulates collected on the steel mesh and in the particle collector were then retrieved and carefully stored to subsequently weighed and analysed as described in the laboratory methods below.
In summary, each of the four experimental burns at each of four sites along the climate transect resulted in the field collection of the following samples: -two biomass (vegetation quadrats) -three ground-vacuumed (two vegetation quadrats and one burnt quadrat) -three airborne particulate samples from the burnt quadrat (one from the steel sieve > 125 µm, one from the particle collector and one from the soot adhered to the metal panels).

Initial preparations
Biomass collected from each unburned reference quadrat was weighed, dried at 60 • C for 5 days and re-weighed to determine fuel moisture.This material was then sorted according to either grass (C 4 ) or woody biomass (i.e.tree/shrub leaves, twigs; C 3 ).The vacuumed material from each quadrat was also sorted, and any biomass fragments found were cleaned and added to the corresponding vegetation category (i.e.grass or woody).Thereafter these were weighed, recombined (keeping some aliquots separate for further testing) and milled.The samples vacuumed from the quadrats (both preburn and post-burn) were sieved to 2 mm and weighed.Separate aliquots were then size fractionated by wet sieving at 125 and 10 µm to conform with the definition of microcharcoal (10-125 µm; Haberle 2005).This procedure enables separate analysis of very fine < 10 µm PyC that is likely to be a major component of aerosol PyC (Andreae and Merlet, 2001).
In the case of airborne particles captured in the system, the following preparation procedures were conducted: (i) the coarse airborne particles collected at the 125 µm steel sieve positioned at the chamber outlet were flushed with a water / methanol solution (1 : 1) and underwent the same wet fractionation procedure as described above; (ii) the particles that passed through the steel mesh (< 125 µm) were wetsieved to 10 µm and (iii) the soot retrieved after cleaning the structure with the water / methanol solution was placed in a rotary evaporator to remove the liquid phase, but no attempt was made to size separate this soot, as intense physical brushing may have significantly altered the "natural" particle size distribution.This fraction was subsequently added to the distal (< 125 µm) pool.In all cases presented above, the resultant fractions were subsequently freeze-dried, weighed and finely milled in preparation for further analyses.

Hydrogen pyrolysis
Hydrogen pyrolysis has been described in detail in a number of publications (e.g.Ascough et al., 2009;Meredith et al., 2012;Wurster et al., 2012Wurster et al., , 2013)).The technique separates PyC in aromatic clusters with a ring size > 7 (HyPyC = PyC that is likely to be resistant to environmental degradation) from other organic C and has been shown to perform well in characterising PyC abundance in a range of environmental matrices (Meredith et al., 2012).Briefly, the solid samples were loaded with a molybdenum catalyst (∼ 10 % of dry weight) using an aqueous / methanol solution of ammonium dioxydithiomolybdate [(NH 4 ) 2 MoO 2 S 2 ].Dried, catalyst-loaded samples were placed in a reactor and pressurised with 15 MPa H 2 under a sweep gas flow of 5 L min −1 , then heated at 300 • C min −1 to 250 • C, then stepped at 8 • C min −1 to a final hold T of 550 • C for 2 min.
Because the catalyst that is loaded undergoes weight loss during hydrogen pyrolysis, the abundance of carbon in the sample after hydrogen pyrolysis is determined relative to TOC (the mass of carbon after treatment/the mass of carbon loaded) and reported as %HyPyC / TOC or %HyPyC / sample.

Carbon abundance and isotope composition
Carbon abundance and isotope composition of samples were determined using a Costech elemental analyser (EA) fitted with a zero-blank auto-sampler coupled via a ConFloIV to a ThermoFinnigan DeltaV PLUS mass spectrometer using continuous-flow isotope ratio mass spectrometry (EA-IRMS) at James Cook University's Cairns Analytical Unit.Stable isotope results are reported as per mil (‰) deviations from the Vienna Pee Dee Belemnite (VPDB) reference standard scale for δ 13 C values.Precisions (SD) on internal standards were better than ±0.2 ‰.

Calculations and modelling
Measured quantities of TOC, PyC and HyPyC in each fraction were summed into two fluxes from each of the fires: "distal" and "proximal".Proximal (likely to remain close to the site of burning) includes that collected from the surface after the fire in addition to the > 125 µm fraction collected from the sieve.Distal (likely to be transported from the site of burning) includes the < 125 µm fraction collected from the stainless steel sieve, as well as that collected in the particle collector attached to the pump, and the soot cleaned from the apparatus.
The percent grass biomass at each site was estimated from the mass of separated grass and woody biomass collected from two plots adjacent to the burning apparatus.An estimate of the initial isotopic composition of the biomass (δ 13 C biomass ) was determined by two methods.First, a bulk sample of the biomass from the two vegetation plots was homogenised and its δ 13 C value measured.Second, the separately summed masses of grass and woody vegetation were used to estimate the bulk isotopic composition, assuming specific values for C 3 woody biomass (−27.2 ‰) and for C 4 grass biomass (−13.1 ‰; values from Cerling et al., 1997).The estimated percent grass biomass values produced by these two methods agreed (linear regression r 2 = 0.88).The average of the two estimates is used as the central tendency, while the range of values is used as an error in these estimates.The carbon isotopic composition of CO 2 and other trace gases produced by combustion for each of the fires was calculated by mass balance using the amount and isotopic composition of initial biomass and of the residual products of combustion (note that this mass difference calculation of CO 2 likely includes other carbonaceous trace gases such as CO and CH 4 ).
The value of SIDE can be expressed as an "isodisequilibrium forcing coefficient" following the terminology of Alden et al. (2010), which is the difference between the δ 13 C values of forward and reverse fluxes between two reservoirs.The values of SIDE were calculated for each of the component fluxes from the fires as the difference between the δ 13 C value of the flux (δ 13 C flux , for example, the distal HyPyC component, δ 13 C flux = δ 13 C distal HyPyC ) and the initial δ 13 C value of the biomass (δ 13 C biomass ).The values of SIDE from each of the four fluxes were regressed with respect to the estimate of percent grass biomass using the Matlab curve-fitting toolbox with a power function.The significance of the power law curve fit was calculated as a p value, which describes the probability, with an F test, that the relationship is a better estimate than simply the mean value of SIDE.Confidence intervals for the curve fit equations were calculated at p = 0.9.The distribution of residuals was tested for normality with a Jarque-Bera test.
The savanna isotopic disequilibrium flux (D SIDE_CO 2 = F bur (SIDE CO 2 ), where F bur is the flux of C from annual biomass burning) was calculated for global savanna environments using an estimate of the fraction of C 4 photosynthesis from Still et al. (2003).The calculated F bur value for savannas was averaged over the period from 1997 to 2011 using the Global Fire Emissions Database (GFED; http://globalfiredata.org;van der Werf et al., 2010), clipped (b) Relationship between TCE and the proportion of pyrogenic carbon produced (HyPyC) relative to TCE.The regression (solid black line) takes the following form: HyPyC / TCE (%) = 48.6-30.9(TCE) 0.066 ; r 2 0.61; p < 0.05; BIC 6.27; n = 14.Different symbols group the individual burning experiments into the different ecosystems shown in Fig. 2; individual data are presented in Table 2.
to the area where C 4 grasses > 1 % (Still et al., 2003).The global distribution of the isotopic disequilibrium flux was calculated and plotted using Generic Mapping Tools (GMT; Wessel et al., 2013).

Production of PyC during savanna burning
The total carbon exposed (TCE) to combustion showed a discernable pattern across the precipitation gradient, with higher potentially combustible C in standing biomass and surface litter at the more humid sites (Table 2, Fig. 2).The median (range) production of PyC and HyPyC was 16.0 (11.5) % and 2.5 (4.9) % of TCE across all experimental burns (Table 2).Similarly, the amount of total carbon combusted (TCC) defined as the difference between TCE and post-fire carbon was used to calculate combustion completeness (described as TCC relative to TCE), and whose median (range) for all fires was 84.0 (11.5) %.The median (range) amount of HyPyC in the proximal component (> 125 µm) of the flux from the fires was 96.9 (14.1) % of the total HyPyC flux across all experimental burns.Fire residence time, defined as time over which air T at the chimney's outlet exceeded 100 • C, increased with TCE (Fig. 4), and thus with the proportion of woody biomass (Table 2).The proportion of TCE converted to HyPyC significantly decreased with initial TCE, and thus decreased with the proportion of woody biomass (Fig. 4; Table 2).

Isotopic disequilibria of proximal and distal fluxes of PyC and HyPyC
The values of SIDE were computed from δ 13 C values shown in Table 3.When considering the PyC pool, the median (range) SIDE value of proximal PyC (SIDE proximal -PyC; −3.3 (4.2) ‰; where SIDE = δ 13 C flux -δ 13 C biomass , with superscripts proximal or distal and PyC or HyPyC relevant to the flux) was not significantly different from the median (range) SIDE value of distal PyC (SIDE distal -PyC; −4.7 (6.3) ‰; p = 0.737, Mann-Whitney U test; Fig. 5).However, the median (range) SIDE value of proximal HyPyC (SIDE proximal -HyPyC −1.1 (5.3) ‰) was significantly different from the median (range) SIDE value of distal HyPyC (SIDE distal -HyPyC; −2.9 (8.0) ‰; p = 0.024; Fig. 5).Comparing the two proximal fluxes, the median value of SIDE proximal -PyC was significantly different from the median value of SIDE proximal -HyPyC (p < 0.001 Fig. 5).The median value of SIDE distal -PyC was not significantly different from the median value of SIDE distal -HyPyC (p = 0.245;  et al. (1997).b TCE: total carbon exposed, mass-weighted isotopic composition of pre-fire biomass and ground TOC.c Mass balance in this experimental burn was not possible as the extraction pipe accidentally detached from the structure during the fire.d Sample was contaminated during manipulation and storage.

Production of PyC across contrasting savanna environments
The proportion of biomass converted to PyC and HyPyC during the experimental burns varied across the range of savannas studied (Table 2, Fig. 4).The median PyC / TCE conversion rate of 16.0 % is in the upper range of values reported in previous studies, although those could have underestimated actual PyC conversion rates due to a combination of factors, which include the use of analytical techniques focusing only on recalcitrant PyC components, the non-inclusion of all PyC materials being produced, and the choice of experimental fires not being representative of typical wildfire conditions (Santín et al., 2015).However, the observation that the relative amount of HyPyC produced in grass-dominated savannas was larger than in woody-biomass-dominated savannas runs counter to our initial assumption that anticipated that more woody biomass would promote the production of proportionally more HyPyC.Longer fire residence times were positively correlated with greater TCE, which resulted in less HyPyC being formed due to the opportunity for more complete combustion (Fig. 4).Indeed, incomplete combustion of biomass can mainly be ascribed to the combined effects of low combustion T , short fire residence times and an overall lack of available oxygen (Loo and Koppejan, 2002).Typically, low fire residence time is reported for fastmoving fires characteristic of landscapes with abundant fine fuel, such as savannas (Hartford and Frandsen, 1992;Wright and Bailey, 1982).While fire residence time was shorter for the grass-dominated sites, maximum T was quite comparable across all the studied sites (Table 2), and agrees well with results reported for savanna fires elsewhere (Miranda et al., 1993;Trollope, 1984).Therefore, the short-lived fires occurring in grass-dominated ecosystems may result in the production of proportionally more HyPyC (Table 2, Fig. 4).The median (range) production of HyPyC as a proportion of both TCE and TCC were 2.5 (4.9) and 2.9 (5.9) %, respec- tively.These values compare well with estimates of "black carbon" (charcoal) production for savanna and grassland fires reviewed by Forbes et al. (2006).

Allocation of PyC produced during savanna fires
In order to better constrain the global PyC cycle, there is an obvious need for a deeper understanding of the factors controlling formation, translocation and mineralisation of PyC and its recalcitrant compounds, represented here by HyPyC (Bird et al., 2015;Conedera et al., 2009;Zimmermann et al., 2012).In this context, the partitioning of PyC between the proximal and distal fluxes may have a strong influence on preservation potential (Thevenon et al., 2010).Blackford (2000) reports an order of magnitude decrease in the percentage of charred particles > 125 µm observed at just 7 m from a recently burned area.In contrast, smaller-sized particles can be transported by wind up to thousands of kilometres, with this range largely depending on the particle size, height of the convection column and environmental conditions (Clark, 1988).The total amount of HyPyC contained in both fluxes varied across the burning experiments, but with most of the HyPyC in the proximal flux in all cases., 2010), clipped to the area where C 4 grasses > 1 % (Still et al., 2003), with an estimated fraction of combusted CO 2 of 0.83 (estimated from Table 2).
The proportion of HyPyC in the proximal flux was very large in grass-dominated savannas (> 96 % of total HyPyC at MIT), compared to the proportion of HyPyC in the proximal flux in woody-biomass-dominated savannas (86-91 % of total HyPyC at DCR).This also suggests that longer fire residence times result in not only a more complete combustion of biomass; they may also promote more effective comminution and volatilisation of fuel into fine particles and gases.Such an effect was also observed by Kuhlbusch et al. (1996), who found burning experiments with the highest fuel loads showed the highest degree of C volatilisation.

Savanna isotope disequilibrium effects (SIDE)
In general, the calculated SIDE values are consistent with relative 13 C-depletion of PyC produced during savanna fires (Krull et al., 2003), with some significant distinctions between proximal and distal fluxes, as well as between the PyC and the HyPyC fluxes (Fig. 6).A relative 13 C-depletion is more consistently observed for the proximal PyC, with δ 13 C values always lower than the original biomass.The δ 13 C values of the proximal flux of HyPyC were also generally lower, with the exception of two fires (UND 3-4; Table 3, Fig. 5).
The distal flux of PyC also had lower δ 13 C values than the precursor biomass.The observation that PyC is relatively 13 C-depleted with respect to the original biomass across all sites supports the hypothesis that, where woody (C 3 ) and grass (C 4 ) biomass are both present, the (C 4 ) grass biomass is preferentially combusted relative to (C 3 ) woody material, leaving PyC from woody biomass to contribute dispropor-  5).This differential fractionation between the products of combustion and original biomass may represent a significant contribution to the global carbon isotope disequilibrium terms used in 13 C double deconvolution models of atmospheric CO 2 (Alden et al., 2010;Ciais et al., 2005;Enting et al., 1995).The isodisequilibrium forcing coefficient for biomass burning, here referred to as I bur , is defined as the difference in the 13 C value of CO 2 produced by biomass burning and the 13 C value of CO 2 into biomass regrown after burning: while the disequilibrium associated with this term (D bur of Ciais et al. (2005) is where F bur is the atmospheric source flux of CO 2 due to biomass burning.However, as used in double deconvolution mass balance models, I bur assumes no isotopic fractionation associated with combustion, i.e. that the δ 13 C value of CO 2 of combustion (δ 13 C CO2_bur ) is equal to that of biomass burnt (δ 13 C biomass-bur ).Because our results show that CO 2 of combustion is 13 C-enriched with respect to biomass burnt (by a SIDE CO 2 value of about 0.6 ‰), we can explicitly account for this during the calculation of I bur : An asterisk is used to indicate that the term accounts for fractionation during combustion.The total burning disequilibrium becomes The first term is the disequilibrium as calculated by Ciais et al. (2005), and assumes no fractionation associated with combustion.The second term explicitly accounts for SIDE CO 2 associated with differential combustion of C 3 and C 4 biomass.The global value of D SIDE_CO 2 is estimated in Fig. 7 to have a global sum of about 0.75 Gt C yr −1 ‰.This value may be an upper bound for D SIDE because some of the 13 C-depleted PyC may be mineralised after the fire, as was observed by Zimmermann et al. (2012).
Figure 5 shows that in both the proximal and distal fluxes, the significance of SIDE during burning is more pronounced when considering the PyC produced by the fires than when considering the HyPyC component alone.The observed patterns highlight the importance of SIDE for the more labile or semi-labile components of the PyC flux, into which C 4derived PyC may be preferentially partitioned.However, the significance of SIDE is less pronounced in the more stable pool of PyC flux, into which C 3 -derived PyC may be preferentially partitioned.
We used particle size separations of PyC fluxes to test the hypothesis that PyC remaining near the site of production shows a more pronounced SIDE due to preferential export of C 4 -derived PyC in the finer particle size fractions.While no significant trends are observed for the proximal fluxes of PyC (either the total or HyPyC components), the relationship is significant for both distal PyC and for distal HyPyC (Fig. 6).These observations indicate that in grassdominated environments, the preferential export of fine particle size fractions is likely to be dominated by low δ 13 C components of PyC.Thus, the SIDE is more pronounced in grassdominated fires either because (a) the small amount of C 3derived PyC flux in these systems is preferentially exported in the fine (soot) fraction of the distal component or (b) compounds with low δ 13 C values are preferentially exported to the fine fraction (Das et al., 2010;Krull et al., 2003).Krull et al. (2003) attributed this effect to the occlusion of protected 13 C-depleted compounds in phytoliths of C 4 grasses, and our results further suggest that 13 C-depleted compounds (e.g.lipids, lignins and phytoliths) may be preserved in PyC (O'Malley et al., 1997), as they are preferentially exported as highly condensed aromatic structures in the fine soot components of natural fires.Another significant contributor to SIDE may also be the plant methoxyl pool, which is extremely depleted in 13 C (∼ −29 ‰ in leaves relative to bulk biomass) and comprises about 2.5 % of plant biomass (Keppler et al., 2004).Moreover, the same authors report that the fractionation associated with the methoxyl groups may be further enhanced during their conversion to volatile compounds (i.e. through burning), which could also contribute to the low δ 13 C values we observed in the soot fraction.Nonetheless, we cannot rule out the possibility that a significant proportion of this soot component may derive from the small amount of C 3 biomass present, even in C 4 -dominated ecosystems.
The small-scale experimental set-up used here did not consider coarse woody debris (CWD), which is common in most savanna environments and may burn along with other biomass on the ground.As this material is exclusively composed of C 3 biomass, the inclusion of PyC from CWD would likely increase the contribution of proximal PyC to the total flux, further enhancing the SIDE on SOC δ 13 C (Bird and Pousai, 1997;Wynn and Bird, 2008).

Implications of SIDE
The calculation of the global SIDE disequilibrium flux (Fig. 7) reinforces the conclusion that the effects of SIDE during savanna burning are most pronounced in grassdominated savannas, where C 4 plants are the most significant fraction of biomass.Scaling these trends up to the global scale indicates a significant SIDE forcing for the savanna biome, which may contribute to a 13 C disequilibrium flux on the order of 0.75 Gt C yr −1 ‰, a value which is of significant magnitude in comparison to other 13 C disequilibrium fluxes used in mass balance models (Ciais et al., 2005).This 13 C disequilibrium flux (D SIDE_CO 2 ) accounts for isotope fractionation associated with differential combustion of C 3 and C 4 biomass, which is otherwise not explicitly accounted for, and would contribute to the total biomass burning isotopic disequilibrium (D bur ) caused by regrowth of ecosystems after fires (1.66 Gt C yr −1 ‰; Ciais et al.;2005).Because grass-dominated savannas show high SIDE for PyC, the implication is that SOC in savannas with frequent grass-dominated fires will become increasingly 13 Cdepleted, as the stable components of PyC accumulate in the soil over time.This observation is consistent with studies of SOC at large spatial scales, for example, where grassdominated Mitchell grasslands show lower δ 13 C values of surface SOC (ca.−16 ‰) than is typical of C 4 biomass (ca. −12 ‰) (Bird and Pousai, 1997;Wynn and Bird, 2008), and agrees well with findings by Dümig et al. (2013) which indicate both the presence of charred grass residues and accumulation of alkyl C in soil fractions as the most likely contributors to the observed decreasing δ 13 C values from grass biomass to C 4 -derived surface SOC.Such a conclusion has implications for interpretation of palaeo-records derived from ancient SOC or its by-products in sedimentary records, which may be biased towards 13 C-depleted values by this SIDE during savanna burning.

Conclusions
The production of PyC and HyPyC was quantified in 16 experimental fires conducted along a transect of sites in northern Australian savannas.After burning, the residues were partitioned into proximal and distal fluxes, each measured for their PyC and HyPyC contents and the carbon isotope composition of each component.The production of PyC across all experimental burns ranges from 11.6 to 23.1 % of TCE.Total carbon exposed is positively correlated with fire residence time, and increasing fire residence time reduces the proportion of HyPyC produced due to the opportunity for more complete combustion.Thus, the short-lived fires occurring in grass-dominated ecosystems result in the production of proportionally more HyPyC compared to more woody-dominated savannas.The median (range) production of HyPyC as a proportion of both TCE and TCC across all experimental burns were 2.5 (4.9) and 2.9 (5.9) %, respectively.These values are in good agreement with other estimates of black carbon (charcoal) production for savanna and grassland fires (Forbes et al., 2006).The relative amount of HyPyC in the proximal flux is significantly larger in grassdominated savannas (up to > 99 %) compared to that observed in woodier savannas (> 86 %).This is a significant observation as these two fluxes are likely to have different preservation potentials.The savanna isotope disequilibrium effect (SIDE, difference in carbon isotopic composition between grass biomass and pyrogenic carbon components) was highest at sites with the greatest proportion of grass biomass although this trend was only observed for the distal flux of fine-grained material (< 125 µm).Scaling the SIDE values observed here to the global savanna biome suggests that SIDE contributes a significant 13 C disequilibrium flux to the global mass balance of 13 C exchange between the atmosphere and other reservoirs (Ciais et al., 2005).The observation that SIDE increases with increasing grass biomass also suggests that frequently burned savannas are subject to depletion of 13 C in the pyrogenic products.As the PyC produced by savanna fires accumulates, either in situ or distally, SIDE produces soil or sedimentary organic carbon with δ 13 C values which are relatively 13 C-depleted with respect to the original biomass.This effect may have significant implications for the interpretation of stable carbon isotopic composition of palaeo-records which are based wholly or in part on PyC.

Figure 1 .
Figure 1.Geographical locations of the studied sites in northeast Australia.Source: "North East Australia" Image Landsat (2014) Google Earth.

Figure 2 .
Figure 2. Characteristics and regional classification of the vegetation; the reader is referred to Torello-Raventos et al. (2013).

Figure 3 .
Figure 3. Schematic diagram depicting an experimental burn, with pre-and post-fire C pools.Initial biomass and PyC inventory were quantified in the near vicinity of each planned burn location by means of destructively collecting all aboveground plant material and vacuuming the soil surface from two 1 m 2 census quadrats, as shown in left bottom picture.

Figure 5 .
Figure 5. Box and whisker plot of isotopic disequilibrium values (SIDE = δ 13 C flux -δ 13 C pool ) for proximal and distal fluxes of PyC and HyPyC, and combustion CO 2 calculated by mass balance (n = 16 for proximal components; n = 15 for distal components, n = 14 for combustion CO 2 ).Line at centre of box represents the sample median, box represents 25th and 75th percentiles and whiskers represent range.

Figure 6 .
Figure 6.Relationship of isotopic disequilibrium values (SIDE = δ 13 C flux -δ 13 C pool ) to the initial proportion of grass biomass (percentage by mass).Error bars of percentage grass biomass represent the range of values measured from two 1 m 2 quadrats.Error bars of δ 13 C represent values calculated using the difference from the isotopic composition of standing biomass (δ 13 C pool ) as determined by one of two methods: (1) measurement of the bulk C isotopic composition of biomass collected from the two quadrats and (2) estimated carbon isotopic composition based on the mass of woody and grass biomass, assuming specific values for C 3 woody biomass (−27.2 ‰) and for C 4 grass biomass (−13.1 ‰; values from Cerling et al., 1997).Solid line shows best curve fit of the power-law equation, with dashed lines representing 90 % confidence intervals.n = 16 for proximal components; n = 15 for distal components.Residuals of the power law fits to the distal HyPyC and distal PyC data showed a normal distribution (Jarque-Bera), with mean residual SIDE values of −0.03 and −0.02 ‰, respectively.

Figure 7 .
Figure 7. Global map of savanna isotope "disequilibrium flux" (following terminology of Alden et al. (2010) D SIDE_CO 2 = SIDE CO 2 × F bur ) of the flux of CO 2 of combustion.The value of SIDE CO 2 used is 0.61 ‰.F bur -savanna is the mean annual flux of CO 2 from savanna fires, averaged over the period from 1997 to 2011 using the Global Fire Emissions Database (GFED; http://globalfiredata.org;van der Werf et al., 2010), clipped to the area where C 4 grasses > 1 %(Still et al., 2003), with an estimated fraction of combusted CO 2 of 0.83 (estimated from Table2).

Table 1 .
Characteristics of the studied sites.
Domingues et al. (2010)ipitation and MAT is mean annual temperature.Climate data derived from the Australian Bureau of Meteorology.For the calculation of woody canopy cover (CC), the reader is referred toDomingues et al. (2010).

Table 2 .
Mass balance of C fluxes in each of the fires.

Table 3 .
Stable isotopic composition (δ 13 C values) of C fluxes from each of the fires; standard errors of the means are shown in parentheses.Explanations for the different fractions are as in Table2.
Notes: a estimated carbon isotopic composition based on the mass of woody and grass biomass, assuming specific values for C 3 woody biomass (−27.2 ‰) and for C 4 grass biomass (−13.1 ‰); values from Cerling

Biogeosciences, 12, 1849-1863, 2015 www.biogeosciences.net/12/1849/2015/ tionately
to the PyC flux (i.e.SIDE PyC values are negative).Also, because the δ 13 C values of PyC are relatively 13 Cdepleted with respect to original biomass, mass balance dictates that the δ 13 C values of CO 2 of combustion are relatively 13 C-enriched with respect to original biomass (i.e.SIDE CO 2 values are positive; Fig.