Current good practice guidelines for national greenhouse gas inventories requires that seasonal variation in emission factors from savanna fires be considered when compiling national accounts. African studies concluded that the emission factor for methane decreases during the dry season principally due to curing of the fuels. However, available data from Australian tropical savannas shows no effect of seasonality on emission factors, consistent with observations that the fine fuels appear to cure fully soon after the start of the fire season. To test whether the seasonality in greenhouse gas emission factors reported for Africa also occurs in Australia, methane and nitrous oxide emission factors were measured in early and in late dry season fires in Western Arnhem Land, a region typical of much of the northern Australia savanna zone. We found no significant seasonality in methane emission factors, but there was substantial variation in emission factors associated with inter-fire differences in vegetation and fuel. This variation could be explained almost completely by combustion efficiency. Nitrous oxide emission factors were not related to combustion efficiency but showed some variation across vegetation and fuel size class. Both methane and nitrous oxide emission factors were consistent with previous work in northern Australia and with some published values from Africa. The absence of a significant seasonal trend in emission factors indicates that savanna fire emissions in northern Australia can be managed by strategic prescribed burning.
 Burning of savannas and grasslands consumes more than one third of the total annual biomass burnt globally and much effort has been put into refining the emissions estimates from these fires [Andreae and Merlet, 2001]. In Australia, on average approximately 400 000 km2 of tropical savannas and 150 000 km2 of arid savannas burn annually [Russell-Smith et al., 2007]. The emissions of nitrous oxide and methane from savanna burning comprise about 2%–4% of the annual accountable greenhouse gas emissions from the Australian continent [Department of Climate Change and Energy Efficiency (DCCEE), 2010]. The direct CO2 emissions from those fires, which are not included in the national inventory, are of a similar magnitude to all national emissions of CO2 from fossil fuel combustion.
 Accounting of the emissions from savanna fires in Australia is done according to the principles outlined in Good Practice Guidance and Uncertainty Management in National Greenhouse Gas Inventories [Intergovernmental Panel on Climate Change (IPCC), 2000], which advises:
Since the emission factor for CH4 can decrease by 50–75% as the burning season progresses, it is strongly suggested that each inventory agency collect seasonal data on the fraction of savanna area burned, the aboveground biomass density, and the fraction of aboveground biomass burned in each savanna ecosystem from the early dry season to the late dry season.
(chap. 4, §A.1.1.3, p. 4.87)
This recommendation follows from a premise stated earlier in the same chapter:
It is desirable to develop the seasonal-dependent activity data and the emission factors of CH4 and N2O from savanna burning in various savanna ecosystems in each country if data are available. Fewer savanna areas and a smaller percentage of aboveground biomass are burned in the early dry season than in the late dry season. Therefore, as the dry season progresses in different savanna ecosystems, it is critical to monitor (i) the fraction of burned savanna area; (ii) the aboveground biomass density; (iii) the percentage of the aboveground biomass burned; and (iv) combustion efficiency.
(chap. 4, §A.1.1.1, p. 4.85)
 The IPCC's assertion that methane emission factors decrease during the burning season appears to be supported by a range of studies of fires in southern Africa [Hoffa et al., 1999; Korontzi, 2005; Korontzi et al., 2003a, 2003b, 2004]. However, this contrasts with conclusions from the main Australian study on emissions from savanna burning, that of Hurst et al. [1994a], which found no seasonal variation in methane emission factors. Based on the Australian finding, it was recommended that strategic early dry season burning be used to reduce greenhouse gas emissions from fires in northern Australia [Cook, 1994; Cook et al., 1995]. The methodology is elaborated in Cook and Meyer  and in Russell-Smith et al. . If the assertion of the IPCC is applicable to northern Australia then it could invalidate this entire approach to reducing greenhouse gas emissions from Australian fires since preferential use of early dry season fires may increase rather than decrease total methane emissions. Land managers would then be forced into a trade off between the requirements to minimize late dry season fires due to their threat to biodiversity, life and property and requirements to minimize early dry season fires due to their greater methane emissions.
 Following the recommendations in Good Practice Guidance [IPCC, 2000], Russell-Smith et al. defined a Tier 2 (i.e., regionally stratified country-specific) accounting methodology potentially applicable to northern Australia. They argued that use of single emission factors for methane and nitrous oxide applied to all fuel types was inadequate, but was based on the best available data. It was recommended that further research be undertaken to address variation in emissions factors across seasons and fuel types, a position that was also strongly agued at a recent international fire experts' workshop [Hyer et al., 2012]. Accordingly, this paper aims to quantify the effects of fire season on emission factors and how these vary across fuel types (grass, leaf litter from trees, coarse woody debris and shrubs).
2.1. Site Descriptions
 The experiment was conducted over two field campaigns, (1) an early dry season (EDS) campaign between 2 July 2009 and 7 July 2009 in which six experimental fires were sampled and (2) a late dry season (LDS) campaign between 30 September 2009 and 4 October 2009 in which five experimental fires were sampled.
 The sampling sites were within the Western Arnhem Land Fire Abatement Project area near Kulnguki (12°38′S 133°55′E; Figure 1) on two landscape types. The first was on the gently undulating sand plains of the Queue landsystem [Lynch and Wilson, 1998]. Here, the vegetation is dominated by Eucalyptus tetrodonta, E. miniata and Corymbia ferruginea trees with Acacia mimulain the mid-storey. The understorey was either annual sorghum (Sorghum intrans) or perennial sorghum (Sorghum plumosum) with a variety of shrubs to about 1 m high. This system will be referred to herein as tussock grass open woodland (TGOW). The second landscape type was the rugged dissected quartz sandstone plateaux of the Buldiva land system. Here, the dominant trees are E. miniata and C. arnhemensis with Acaciaspp. in the mid-storey. The understorey is dominated by Spinifex hummock grass (Triodia spp.) and a variety of shrubs including Hibbertia spp., Grevillea spp., Jacksonia spp. and Calytrixspp. This system will be referred to as hummock grass open woodland (HGOW). While the tussock grasses are similar to those of many tropical savannas world-wide, the hummock grasses are an endemic Australian group of arid-adapted grasses with a shrub-like growth form.
 The soils of the Queue land system are sands to a depth of more than 1 m classified as Orthic Tenosols, while those of the Buldiva landsystem are either exposed rock or shallow sands classified as Leptic Rudosols [Isbell, 1996; Lynch and Wilson, 1998].
 The climate is tropical monsoonal with about 95% of the mean annual rainfall occurring between November and April. The mean annual rainfall of Oenpelli (12.33°S 133.06°E), 100 km to the west-northwest, is 1439 mm. Mean maximum temperatures at Oenpelli are about 32°C in July and 33°C in January.
2.2. Fire Descriptions
 The sampling was conducted in three vegetation types: (1) TGOW; (2) HGOW; and (3) pure swards of annual sorghum or isolated hummocks of Spinifex surrounded by rocky outcrops. These isolated hummocks could be ignited individually without risk of establishing a fire front. These isolated fires facilitated sampling smoke from a single fuel class (grasses) in the absence of tree leaf and twig litter which is ubiquitous in all other locations.
 In total, eleven experimental fires were ignited: six in the EDS campaign and five in the LDS campaign. The six TGOW fires were conducted in a series of adjacent plots (designated Blocks A–F) established along an access road prior to the first campaign. Each plot was approximately 500 m × 200 m in size, bounded on three sides by mineral earth barriers and on the fourth side by the road. The HGOW fires were ignited in isolated country accessible only by helicopter approximately 10 km from the tussock grass open woodland plots. Two areas of approximately 30 ha each surrounded by natural fire breaks were identified prior to the campaign.
 Each experimental plot was surveyed prior to ignition for plant species composition, fuel composition, load and moisture content following the protocols described by Russell-Smith et al. . The characteristics of all plots, fuels and fires are given in Table 1.
Table 1. The Characteristics of the Vegetation, Fuels and Fires Sampled at Kulgnuki
TGOW: tussock grass open woodland; HGOW: hummock grass open woodland; Spinifex: pure sward of Spinifex hummocks; Sorghum: pure sward of Sorghum tussocks; NM: not measured.
 The fires in the tussock grass open woodlands were ignited with a fire line on the down-wind boundary, which back-burns to produce a fire break. The upwind boundary was then ignited and the majority of the block burned with a heading fire driven by the prevailing wind. In the hummock grass open woodland, fires were lit from a point ignition.
 Within the fires in TGOW and HGOW treatments measurements were made of combustion of fine (6 mm or less diameter) fuels, and logs, as described in section 2.4. Additionally, a series of tests was conducted in which smoke was sampled from individual fuel components (grass, fine tree leaf litter, coarse fuels, heavy fuels and green leaves for several species). These fuels were fed onto the hot coal bed of a campfire and burned to completion during which a series of bag samples were collected. These tests extended the range of MCE beyond those encountered in the experimental plot fires.
2.3. Calculation of Emissions and of Emission Factors
 The methodology used in this study for estimating trace gas emissions from fire is a variant of the IPCC algorithm. It is described in full by Russell-Smith et al. . In brief, the emission of a trace species i (Ei) is determined by the mass of fuel consumed during combustion (FP) and the emission factor for each trace species (EFi).
 The mass of combusted fuel is the product of the area exposed to fire (A), the fuel load (FL) and the burning efficiency (BEF), according to the commonly used Seiler and Crutzen  approach. Fuel is typically stratified into a range of size classes (j) specifically tree leaves, grass, coarse fuels and heavy fuels. The area exposed to fire is the area of the fire scar A′ corrected for the patchiness of the fire (P), i.e.
 The emission factors (EFs) can be defined either relative to combusted fuel mass [Andreae and Merlet, 2001; IPCC, 2006] or relative to the fuel elemental content [Hurst et al., 1994a]. We use the latter definition. For carbon species, CH4, CO and volatile organic compounds (VOC), EFs are expressed relative to fuel carbon, and the nitrogen species N2O and NOx are expressed relative to fuel nitrogen. Fuel carbon mass is determined from fuel mass by the fuel carbon content (CCj) while fuel nitrogen is derived from the fuel mass by the product of CCj and the fuel nitrogen to carbon ratio NCj. The parameter Mi converts from elemental mass of species i to molecular mass. Combining these equations, for CH4, CO and VOC,
For N2O and NOX
 Emission factors are measured following the approach of Hurst et al. [1994a]. The emission ratio for trace species i in a smoke sample is defined as
 Where ΔCi = [Ci]smoke − [Ci]amb, i.e., the difference between the molar concentrations (or mole fraction) of trace species i in the smoke sample and its concentration in ambient air upwind of the combustion source, ΔCO2 = [CO2]smoke − [CO2]amb is the difference between the molar concentrations (or mole fraction) of CO2/C in the smoke sample and in the upwind air, etc.
 Emission factors are commonly defined relative to fuel mass [as in Andreae and Merlet, 2001]; however, in the definition proposed by Hurst et al. [1994a], which is used in this study, emission factors and emission ratios are closely related; for the carbon species, e.g., CH4,
For the nitrogen trace species, e.g., N2O
Hao and Ward  found that the emission factors of many organic species, are strongly correlated to the combustion efficiency (CE), which is defined as
i.e., emission factor for CO2 (EF CO2).
 This is commonly approximated by the modified combustion efficiency (MCE) defined as
because usually more than 95% of the carbon is emitted as CO and CO2.
2.4. Gas Sampling Methodology
 The experimental fires were, necessarily, for reasons of environmental and personal safety, small in size (∼10 ha) compared to the much larger fire events that can occur in the area [Edwards and Russell-Smith, 2009]. The small size of these fires resulted in small plumes, which precluded smoke sampling from aircraft. Fortunately, low–intensity savanna fires can be reliably sampled at ground level. Two sampling approaches were used in this study; via direct sampling close to the emission source (i.e., within 1 m) and via open path infrared spectroscopy adjacent to or within the fire boundary.
2.4.1. Source Sampling
 We collected smoke from flaming combustion, from targeted sampling of smoldering-phase combustion behind the fire front and from burning heavy fuels and logs. We refer to it subsequently as the “bag method.”
 For the bag method, we designed a portable backpack smoke collector suitable for sampling close to the emission sources of savanna fires in remote areas often accessible only by helicopter. The unit comprises a 2.5 m × 12 mm diameter stainless steel sampling probe, in-line Teflon filters and an air pump that delivers the filtered smoke sample to a 10 L Tedlar bag mounted on a backpack. During sampling the tip of the probe was positioned approximately 500 mm above the flame within the entrainment zone. In this region, combustion has ceased due to cooling and dilution by entrained air, but the smoke concentration remains high. The air sampling rate was set at approximately 1 LPM and therefore each sample bag contains smoke sample collected over approximately 10 min. The backpacks are equipped with three additional gas lines; two for continuous measurement of PM2.5, CO2 and CO, and a third for collection of total suspended particulate matter (TSP) on filters. Concentrations of CO2and CO were measured continuously with a Q-Trak (model 7565, TSI, Shoreview, Minn., USA) and particulates were measured continuously with a DustTrak (model 7451, TSI, Shoreview, Minn., USA. The Q-Traks were calibrated daily with gas standards (2500 ppm CO2, 100 ppm CO and 500 ppm CO, CalGaz, Air Liquide America Corp., Cambridge, Md., USA), while the DustTraks were calibrated against the gravimetric PM mass measurements. The analyses of the particulate emissions will be presented in a subsequent paper.
 Tedlar bags were selected for the storage container because they are free from the well known artifact that occurs with some stainless steel canisters, The artifact is caused by reactions catalyzed at the stainless steel surface involving SO2 [Muzio and Kramlich, 1988] The issue was explicitly raised in Good Practice Guidance [IPCC, 2000], which states: “Since N2O is not stable during storage of smoke samples, the molar emission ratio of N2O to CO2 has been derived from laboratory experiments in which different types of vegetation were burned…” (chap. 4, §A.1.1.2, p. 4.86). It does not occur with glass flasks (which were used by Hurst et al. [1994a]) or with Tedlar bags.
 The daily operating protocol for the bag method comprised: (1) calibrating and zeroing the Q-Traks; (2) setting and calibrating the sample flow rates; (3) flushing and evacuating Tedlar gas sample bags; (4) preparing two field blanks (Tedlar bags filled with high purity nitrogen); and (5) filling two Tedlar bags with ambient air for background correction of the smoke concentrations.
 Typically, five to eight 10-L smoke samples were collected by each of two sampling units at each experimental fire. All samples were analyzed for CO2 and CO concentration within 12 h of collection. The bags were then returned to Darwin (approximately 300 km to the west), where they were analyzed for CH4, N2O, CO2 and CO concentration within 3–5 days of collection. CH4 concentration was measured by a Flame Ionization Detector using a TEI Model 55C total hydrocarbon analyzer (Thermo Scientific, Franklin, Mass., USA). N2O was measured by GC-ECD with an HP 5890 Series II Gas Chromatograph on Poropak QS column with 5% CH4 / Argon carrier gas. The TEI55C was calibrated against a BOC Beta standard (4.2 ppm CH4 / 0.9 ppm propane in air). N2O was calibrated against an ambient clean air standard collected under baseline conditions at Cape Schanck Victoria and calibrated by CSIRO on the AGAGE global calibration scale against standards prepared by the Scripps Institute of Oceanography. CO2was measured by NDIR using both a Q-Trak and a Li-Cor Model 6262 CO2/H2O analyzer (Li-Cor, Lincoln, Neb., USA). CO concentration was measured using a Q-Trak. Analytical precision was approximately 0.5 ppb N2O, 50 ppb CH4, 1 ppm CO2 and approximately 3 ppm CO.
2.4.2. Corrections for Sample Diffusion From the Tedlar Bag During Storage
 It is well known [e.g., Fan et al., 2001] that trace gas species stored in Tedlar bags slowly exchange with the surrounding atmosphere and therefore bag manufacturers and most standard sampling protocols and standard operating procedures recommend that storage time should be minimized, ideally to 2 days or less depending on the gas species under investigation. In this study the field site was one day's transport time from the analytical laboratory and analysis time was typically one day therefore storage times ranged from 3 to 5 days before the first laboratory analysis. However, with stable gases such as CH4 and N2O, the rate of exchange for stable gases is predictable and therefore correction can be made for these losses.
 Losses of these gases occur mostly by diffusion through the Tedlar membrane and leakage around the valve seals. The process can be described as
 Where C, and Ca are the concentration of the trace species inside the bag and in the ambient air, respectively, A and V are the surface area and volume of the bag, and τ is the porosity of the bag. Integrating equation (10) from t = 0 (the time of filling) to t = T (the time of analysis,
 Field blanks were prepared on each day of sampling and analyzed for CH4 and N2O at least twice after filling. Substituting the values of Ca, C0, CT and T from these analyses into equation (11) yielded estimates for average Aτ/V for CH4 and N2O of 0.07 ± 0.01 d−1 and 0.08 ± 0.01 d−1, respectively. A previous unpublished study by C. P. Meyer in which these bags were used indicated that Aτ/V for N2O and CO2 were also similar. Therefore, if we assume that the permeability of Tedlar to CO is similar to the other gas species and that the bags are stored in clean ambient air, then from equations (5) and (11) it is evident that the emission ratios will be independent of storage time if all gases from a single bag are analyzed at approximately the same time. Where this is not the case, equation (12) can be applied to correct measured concentrations to the time of sampling, T0.
 To confirm that these assumptions held in practice, the samples from the fire of 2 July, which were initially analyzed on 6 July were reanalyzed on 9 July. Emission factors for CH4 were, on average, 17% higher in the second analysis while EF N2O were, on average, 9% higher (Figure 2a). Laboratory concentrations of CH4 and CO2 were observed to be well above ambient concentrations, particularly CH4 which was persistently about 75% higher than ambient due to venting of the Argon/CH4 carrier gases into the laboratory. Because this slows the rate of CH4 loss from the Tedlar bags relative to the other gases during the periods that the bags were connected to the sample manifold in the laboratory, it is not surprising that the EFs calculated from the second analysis tended to be higher than those from the first analysis. In the second campaign, the gases were vented outside the laboratory and, therefore, the differences between EFs calculated from the first and second analyses in October (Figure 2b) were less than in July; within 0.3% for EF CH4 and within 8% for EF N2O. Comparison between the initial MCE estimate determined from the measurements of CO and CO2 made in the field and the first and second laboratory analyses are shown in Figure 3. The relative change between the initial estimate and the first laboratory analysis averaged 0.1%, while the average difference between the initial and the second laboratory estimate was 2%. Therefore, the repeatability of EF estimates was good for all the gases of interest, and thus the assumptions stated above appear to be valid. However, to minimize the risk of errors caused by elevated concentrations of CH4 (and potentially also of CO2 and N2O) in the analytical laboratory, all reported values of MCE were calculated from the initial CO and CO2 bag concentrations, and all reported EF CH4 and EF N2O were calculated from the first laboratory analyses.
2.4.3. Open Path Infrared Spectrometry
 The second method used to quantify the relative abundance of the different traces gases within the smoke, and thus to derive their emissions ratios and emission factors, was extended open path Fourier Transform Infrared (FTIR) spectroscopy. The procedure followed the approach detailed in Wooster et al. , building on that of Griffith et al. . An FTIR spectrometer (MIDAC Corporation, Irvine Calif., USA), fitted with a 76 mm Newtonian telescope, was used to view a collimated IR source positioned downwind of the experimental plots at a distance of 30–70 m. IR spectra were recorded at 0.5 cm−1resolution and co-added to improve the signal-to-noise ratio. IR-absorbing (i.e., GHG) molecules within the smoke cause the appearance of species-specific absorption features within the spectra, allowing them to be identified. The path length averaged mixing ratios of the gases CO2, CO and CH4 can then be retrieved via quantitative analysis of the IR spectra [e.g., Smith et al., 2011; Wooster et al., 2011]. For this we used the MultiAtmospheric Layer Transmission (MALT) model of Griffith combined with a nonlinear least-squares-fitting procedure [described byGriffith et al., 2003; Smith et al., 2011]. Mixing ratios of CO2, CO and CH4 were retrieved from the spectra using this approach and the optimum spectral windows identified by Smith et al. . That study indicated that this analysis procedure can retrieve path length averaged mixing ratios to a precision of <5% over the range of total column molecular abundances found here for CO2, CO and CH4. Emission ratios were derived from bivariate plots of CO and CH4 against CO2, and emissions factors calculated using the Carbon Mass Balance Method of Ward and Radke . See Wooster et al.  for details of these ER and EF calculations using a similar set of IR spectra recorded during a series of savanna fires in southern Africa.
 In terms of measurement geometry, in TGOWs the IR spectrometer was located adjacent to the downwind boundary of the plot [as in Wooster et al., 2011], and in HGOW, the FTIR was downwind of the point of ignition. While our extended open path approach to these spectrometric measurements did not allow the sampling of combustion from individual fuel types (as did the “campfire” measurements made using the bag method) and provides a measurement accuracy somewhat lower than those of laboratory based methods, it does offer the advantage of being able to rapidly characterize emissions chemistry based on long transects through the emitted smoke. This removes potential impacts of unrepresentative point-based sample collection, and allows measurements to be made with minimal (i.e., few tens of seconds) difference between the smoke generation and the measurement time, significantly limiting the effects of any post-combustion chemical conversion processes. It therefore offers a very complementary approach to the bag method described insection 2.4.1.
 Values of EF CH4 for flaming combustion in TGOW varied from 0.13% to 0.62% of emitted carbon, while EF CH4 values measured in the HGOW fires were substantially lower, ranging from 0.06% to 0.3% of emitted carbon. The means and variances of the data are presented in Table 2. The direct measurements of pure grass swards, either native Sorghum or Spinifex were similar to EF CH4 values in the HGOW and ranged from 0.05% to 0.23% emitted carbon. These EF CH4 values corresponded to MCEs of 0.88–0.97, 0.94–0.97 and 0.94–0.98, respectively, for the TGOW, the HGOW and pure grasses (both Sorghum and Spinifex). The EF CH4 for heavy fuels was three times larger than EF CH4 for fine fuels (Table 3).
Table 2. The Mean Values of MCE, EF CH4 (% Emitted Carbon) and EF N2O (% Emitted Nitrogen) for Flaming Combustion of Each Plot Sampled in July and September 2010 at Kulnguki
TGOW: tussock grass open woodland; HGOW: hummock grass open woodland.
 The regression equations relating EF CH4 (% emitted carbon) for July, October and the combined data set for flaming combustion data are, respectively:
 The July and October regression lines are not significantly different, so in subsequent calculations we use equation (15) (Figure 4).
 A typical time course for path length averaged mixing ratios of CO, CO2 and CH4 as measured by the FTIR (fire 3 in the TGOW) is shown in Figure 5. The data shows a series of diminishing peaks from smoke eddies from the fire front adjacent to and upwind of the FTIR. The smoke in these eddies is a mixture of flaming combustion and smoldering combustion of fine and coarse fuels in the zone immediately behind the front. This post fire zone was at, at most, 5–10 m deep, and, in comparison to the flaming zone, a relative low intensity emission source. As the fire front moved away from the instrument, the likelihood of smoke eddies from the front intersecting the FTIR path rapidly diminished and the smoke detected by the FTIR was mostly from the increasing area of smoldering combustion. In the smoldering phase, the trace gas concentrations were close to the background concentration.
Figure 6 shows the relationship between EF CH4 and MCE calculated from the complete set of FTIR measurements. These data align extremely closely with the bag data. This is independent verification that the EF CH4 estimates for flaming combustion are reliable.
 In the tussock grass open woodland, the EF N2O values varied from 0.466% to 1.04% emitted nitrogen, while for the hummock grass open woodland, the values varied from 0.085% to 1.19% emitted nitrogen. The range in values for native Sorghum and Spinifex, respectively, were 0.48%–0.77% and 0.36%–0.93% emitted nitrogen. The EF N2O was not correlated to the MCE (Figure 7). Therefore, in contrast to EF CH4, EF N2O was independent of vegetation class or combustion efficiency. Much of the variability occurred within individual fires and is therefore statistically unexplained.
 An analysis of variance of the emission factor data shows that there is no seasonal change in EF CH4 but that there are substantial and significant differences in EF CH4 across vegetation types (Table 4). Most of variance in EF CH4 was explained by MCE. There is complete agreement between the bag method and the FTIR method of measuring EF CH4. The issues that need to be addressed are (1) the consistency of the data collected here with previously reported data for Australian savanna and forest fires, savanna and forest fires from other regions and other classes of biomass combustion and (2) the implications of the findings for the national greenhouse gas inventory.
Table 4. Analysis of Variance of MCE, EF CH4 and EF N2O for Emissions From Flaming Combustion With Vegetation Class and Season
 In Figure 8a, the relationship between EF CH4 and MCE from our measurements is compared with previous Australian measurements [Hurst et al., 1994a, 1996]. The linear regression between EF CH4 and MCE (15) appears to apply for a wide range of combustion states down to MCE values of about 0.88. Below this value, the relationship becomes nonlinear. This relationship is also consistent with a wider range of published data from African savannas [Cofer et al., 1996; Korontzi et al., 2003a; Ward et al., 1996], from laboratory experiments in combustion hoods [Kuhlbusch et al., 1991] and from domestic woodheaters [Gras et al., 2002]. At MCE values greater than 0.88 most EF CH4 versus MCE data fall close to our linear regression line (Figure 8b), and therefore within this domain the linear relationship described by equation (15) appears to be a robust description for most EF CH4 data.
 At MCE below 0.88 published data (Figure 8b) are consistent with the emissions we see from coarse and heavy fuels (Table 3, Figure 8a), however a nonlinear function is required to best explain the EF CH4 versus MCE relationship. Over the full domain of MCE, an exponential relationship provides a good empirical description of both our measurements and previously published data (Figure 8b, equation (16)). It explains 74% of the variance (16) and asymptotes to equation (15) at higher MCE values.
 One consequence of a nonlinear relationship between MCE and EF CH4 is that linear approximations calculated over MCE values obtained from both smoldering and flaming combustion will tend to overestimate the slope we observed for flaming combustion alone (i.e., MCE greater than 0.88). This situation occurs in measurements in mixed plumes and in studies where emission factors measured in smoke collected near the source were weighted by the relative contributions of smoldering and flaming combustion [e.g., Ward et al., 1996]. This could account for the twofold higher values for the slope of EF CH4 vs MCE of 8.6% emitted C (unit MCE)−1 reported by Korontzi et al. [ 2003b] compared to the slope of 3.9% emitted C (unit MCE)−1 of equation (15). Outliers to (15) reported by Kuhlbusch et al. , Cofer et al.  and Ward et al.  might also be explained by this phenomenon.
 The cause of the large variation in MCE is clearly complex. Korontzi et al. [2003b] ascribe it mostly variation in fuel moisture content, extending from the study by Hoffa et al. , which found such a relationship in both Miombo and Dambo vegetation communities in Zambia. Much, probably most, of the variation in moisture content reported in these studies was associated with grass curing. In the savanna woodlands of Northern Australia, this is less likely to be a significant factor. Australian savannas experience a very protracted dry season between April and October, when little or no rainfall occurs, and Australian native grasses senesce very quickly. At Oenpelli, the nearest meteorological observatory to Kulnguki, no rainfall occurred in 2009 between 17 May and 29 October 2009. By early June the fuels were fully cured and remained at equilibrium fuel moisture content (approximately 8% dry weight) throughout the dry season. It is therefore unlikely that fuel moisture content is the source of the MCE variation that we recorded in the two campaigns. This is in accord with Hurst et al. [1994a], who also found no seasonal difference in EF CH4 in elevated smoke plumes from fires in Kakadu National Park, NT, which they sampled from light aircraft. Fuel structure is the more likely candidate. Fine fuels, comprised of leaf litter and small twigs in a loose bed on the soil surface, were poorly aerated during combustion and were inclined to smolder. In contrast, grass tussocks burned quickly and intensely during both the EDS and LDS campaigns.
 We found no seasonal change in EF N2O but we observed substantial variability among replicates and among fires from different vegetation classes. This variability is probably because N2O is an intermediate reaction product. It is produced by reaction of NO with either NH or NCO and destroyed by collisions with a third body (i.e., by thermal decomposition) or by reaction with hydrogen or hydroxyl radicals to produce N2 [De Soete, 1990; Ogawa and Yoshida, 2005]. Using a laboratory furnace, Winter et al.  confirmed that N2O emission occurs at the end of the devolatilisation phase following flame extinction, when the destruction reactions are quenched faster than the formation reactions. The duration of this state is less than half of the devolatilisation phase and less than 5% of the time required for complete combustion. Because the net production is the balance between competing processes, the N2O emission rate will be highly dependent on both the combustion conditions and the dispersion rate of the combustion gases.
 Our measurement of EF N2O (0.7% N; see Table 3) was at the lower end of the range of values reported in the published literature. The measurement by Hurst et al. [1994b] from northern Australian savannas was 0.8% N and very similar to ours. Good Practice Guidance [IPCC, 2000] summarizes EF N2O measurements from savannas in Africa and South America; EF N2O averages 0.068 g N2O (kg fuel)−1 or 1.15% N. Andreae and Merlet's  summary reports EF N2O ranging from 0.21 g N2O (kg fuel)−1 for savannas, 0.26 g N2O (kg fuel)−1 for forests and 0.07 g N2O (kg fuel)−1 for agricultural residues. These translate to EF N2O values of 2.5%, 3.5% and 0.8% of fuel N respectively.
 Recently, Sahai et al.  measured even higher values in smoke emissions from wheat stubble burning in India, ranging from 0.34 to 0.74 g N2O (kg fuel)−1, or approximately 4%–9% N. However, these values are questionable for two reasons. First, we found that EF N2O is independent of MCE. In contrast, Sahai et al.  report that EF N2O calculated from ground-level observations of plumes strikes, declines with CE from 1.6 g kg fuel−1 to 0.3 g kg fuel −1. They also reported declining EF NOx with increasing CE. It is chemically implausible that production of a terminal oxidation product declines with increasing combustion efficiency and it is contrary to the response of NOx to MCE observed in other well based studies [e.g., Gras et al., 2002]. Second, Sahai et al.  reported that N2O comprises a large proportion of the oxidized nitrogen; increasing from 22% at 0.93 combustion efficiency to 48% at 0.98 combustion efficiency. This contrasts to most previous studies in which NOX emissions exceeded N2O emissions by more than an order of magnitude [Andreae and Merlet, 2001]. Given these discrepancies, the conclusion of Sahai et al.  should be treated with caution.
Winter et al.  give some insights into possible sources of the EF N2O variability. Similar to our findings, they reported wide variation in EF N2O from wood fuels ranging from 0.1% to 5% N, and these values were insensitive to oxygen concentrations above 10% but very sensitive to combustion temperature. Maximum EF N2O occurred at 800°C. At 700°C and 900°C, the N2O emission was approximately threefold lower; consistent with a process in which net emission of N2O is a balance between formation and destruction mechanisms. Ogawa and Yoshida  specifically addressed the extent of destruction processes in open combustion for rice straw combustion and concluded thermal decomposition and reduction removed 15% of the N2O produced by the formation reactions during combustion. Across all forms of biomass burning, the magnitude of the proportionate destruction of N2O is likely to depend on the balance of production and destruction processes which are affected by many factors such as combustion temperature. Combustion temperature is often reported as flame temperature, a parameter which is influenced to a degree by the known method of measurement. Due to thermal capacitance, the apparent flame temperature measured by thermocouples is influenced by thermocouple wire diameter and flame duration at the thermocouple junction, while, due to infrared absorbance by the combustion gases, the flame temperature measured by infrared thermometers is weighted to the outer envelope of the combustion gases. Keeping these sources of variation in mind, apparent flame temperature appears to be highly variable, ranging from below 400°C to above 800°C within and between fires of differing intensity. Meyer et al.  for example report flame temperatures averaging 610°C (480°C to 710°C) in 18 test fires of straw, native Sorghum, sugarcane and forest litter conducted in a test tunnel. Moore et al.  measured peak flame temperatures of 680°C at Kapalga, NT, while Freeborn et al.  report flame temperatures near 800°C measured with infrared thermometers in laboratory combustion tests. A comprehensive study in temperate Eucalypt forest in Western Australia [Gould et al., 2008] reported apparent flame temperatures ranging from about 200°C to 1000°C depending on flame length. Applying the Gould et al.  relationship to Australian savanna fires, in which flame lengths typically range from 0.5 m to 2 m, apparent flame temperatures will range from 400°C to 700°C. The balance between N2O formation and destruction in the combustion zone will vary widely across this temperature range, and it is therefore feasible that much of the variability we observed in EF N2O arises from differences in local fire intensity and hence from apparent flame temperature along the fire front. This phenomenon may also go some way toward explaining the wide variation in EF N2O reported in the literature. The savanna fires in Northern Australia are at the cooler end of the spectrum of wildfire intensities, which may explain the low EF N2O values reported by Hurst et al. [1996, 1994b] and confirmed by our study, compared to higher values of EF N2O reported for forest fires by Andreae and Merlet . The wide variation in reported EF N2O emissions factors clearly indicates that further studies are required to characterize the determinants.
 We found no evidence that seasonality significantly affects emission factors, but our finding that emission factors vary strongly with vegetation type and the fuel components makeup has substantial implications for the calculation of the national emissions inventory. Here we show how the changes affect the Australian national greenhouse gas inventory for savanna burning, but note that the principles should apply generally to savannas across the world. The current Australian methodology [DCCEE, 2010] follows the 1996 revised IPCC guidelines for greenhouse gas inventories [IPCC, 1996] with country-specific emission parameters that include fuel loads, burning efficiencies and emission factors.Russell-Smith et al. present an elaboration of this principle for the savanna woodlands which explicitly describes the effect of vegetation class, fuel type and seasonality that affects fire patchiness and burning efficiency but not emission factors. This was required for consistency between greenhouse gas accounts from carbon-offset management projects and national greenhouse gas accounts. Both methodologies applied a single emission factor for methane and for nitrous oxide across strata. However, the data presented here indicate that the use of a single emission factor for methane in particular is no longer appropriate. A sensitivity analysis indicates that, from 2003 to 2009, taking account of the innovations ofRussell-Smith et al.  reduces the annual average emissions by 5%. Introducing a separate emission factor for smoldering logs increases emissions by 15%, while the new emission factors for fine fuels reduce emissions by 21% (Table 5). These changes lead to a combined reduction in emissions estimates of 4%. While this combined impact is relatively small for Australia, it may be larger for some other savanna systems around the world, and for smaller regions within Australia. For example, Wooster et al.  recently demonstrated a twofold variation in methane emissions factors for late dry season fires conducted in Kruger National Park (South Africa) in different experimental burn plots containing different proportions of the various savanna fuel components.
Table 5. Sensitivity of Mean Annual Emissions of CH4 and N2O in 2003–2009 to Methodology Revisions for Fire Season, Vegetation Class and Fuel Size Class, Emission Factors and N:C Ratio
Change From Base
Change From Current Method
Revised EFs for Fine and Coarse Fuels
Fully Updated EFs and NC
 Thus our findings indicate that variation in emission factors across fuel types and vegetation types need to be considered across the world's savannas, which show considerable structural variation.
5. Conclusions From Direct Measurements of Emissions
 Because the Australian national greenhouse gas inventory algorithm for emission from savanna burning is a multiplicative combination of factors, robust regional averages of emission factors, fuel loads and burning efficiencies provide for an accurate national account. However, emissions offset programs that reduce annual emissions through fire management may require accounting methodologies which do not have this property and, therefore, cannot be simplified to the Australian national algorithm without loss of accuracy. Meyer  and Russell-Smith et al.  provided a partial solution that reconciled the methodologies required for emission offset projects with the Australian national methodology but, in light of the findings by Korontzi et al. [2003b] about seasonality of methane emission factors in southern Africa, both papers acknowledged that further work was required to improve the understanding of seasonal variation in emission factors. The issue has been addressed more completely here and we have shown that there was no change in either methane or nitrous oxide emission factors at our study sites between the EDS and the LDS and, therefore, that reduction in greenhouse gas emissions from savanna fires in northern Australia through judicious fire management is a viable strategy. Nevertheless there was substantial variation in emission factors associated with vegetation and fuel type differences and these need to be explicitly included in inventory methodologies if accounting is to be accurate. The study sites were selected because they were representative of the majority of the tropical savanna of northern Australia, and therefore our detailed findings should be directly applicable to the Australian national accounts.
 The extent to which these findings apply globally, particularly the absence of seasonality in MCE and EF CH4, requires further investigation. There are substantial differences between the findings presented here and those presented by the IPCC  and substantial differences between the various studies of African savanna fires [Hoffa et al., 1999; Korontzi, 2005; Korontzi et al., 2003a, 2003b, 2004]. Whether these reflect differences between regions or result from differences in measurement and interpretation are yet to be resolved.
 We thank the North Australian Indigenous Land Sea Management Alliance (NAILSMA), Indigenous Rangers of Warrdeken Land Management Limited, the Mok Clan as traditional owners of the field sites and, in particular, Wamud Namok (deceased). This was a collaborative project supported by the Australian Government Caring For Our Country Program through NAILSMA. We are also grateful for the technical support provided by Bushfires NT and CSIRO.