Reliability of biomass burning estimates from savanna fires: Biomass burning in northern Australia during the 1999 Biomass Burning and Lightning Experiment B field campaign

Authors

  • Jeremy Russell-Smith,

    1. Tropical Savannas Cooperative Research Centre and Bushfires Council of the Northern Territory, Darwin, Northern Territory, Australia
    Search for more papers by this author
  • Andrew C. Edwards,

    1. Tropical Savannas Cooperative Research Centre and Bushfires Council of the Northern Territory, Darwin, Northern Territory, Australia
    Search for more papers by this author
  • Garry D. Cook

    1. Tropical Savannas Cooperative Research Centre and CSIRO Tropical Ecosystems Research Centre, Darwin, Northern Territory, Australia
    Search for more papers by this author

Abstract

[1] This paper estimates the two-daily extent of savanna burning and consumption of fine (grass and litter) fuels from an extensive 230,000 km2 region of northern Australia during August-September 1999 encompassing the Australian continental component of the Biomass Burning and Lightning Experiment B (BIBLE B) campaign [Kondo et al., 2002]. The extent of burning for the study region was derived from fire scar mapping of imagery from the advanced very high resolution radiometer (AVHRR) on board the National Oceanic and Atmospheric Administration (NOAA) satellite. The mapping was calibrated and verified with reference to one Landsat scene and associated aerial transect validation data. Fine fuel loads were estimated using published fuel accumulation relationships for major regional fuel types. It is estimated that more than 43,000 km2 was burnt during the 25 day study period, with about 19 Mt of fine (grass and litter) fuels. This paper examines assumptions and errors associated with these estimates. It is estimated from uncalibrated fire mapping derived from AVHRR imagery that 417,500 km2 of the northern Australian savanna was burnt in 1999, of which 136,405 km2, or 30%, occurred in the Northern Territory study region. Using generalized fuel accumulation equations, such biomass burning consumed an estimated 212.3 Mt of fine fuels, but no data are available for consumption of coarse fuels. This figure exceeds a recent estimate, based on fine fuels only, for the combined Australian savanna and temperate grassland biomass burning over the period 1990–1999 but is lower than past estimates derived from classification approaches. We conclude that (1) fire maps derived from coarse-resolution optical imagery can be applied relatively reliably to estimate the extent of savanna fires, generally with 70–80% confidence using the approach adopted here, over the major burning period in northern Australia and (2) substantial further field assessment and associated modeling of fuel accumulation, especially of coarse fuels, is required.

1. Introduction

[2] It is widely recognized that biomass burning of tropical savanna biomes is a globally significant driver of CO2 cycling and a source of CO2 and other chemically reactive, atmospheric trace gases [Cahoon et al., 1992]. However, accurate quantification of such emissions is problematic, being reliant on reliable estimation of various parameters including, the spatial and temporal distribution of burning, and appropriate fuel load estimates for different fuel fractions and types. The application of different assumptions and approaches, particularly estimation of the extent of biomass burning, has resulted in various global or regional estimates of savanna biomass burning and derived emissions [e.g., Seiler and Crutzen, 1980; Crutzen and Andreae, 1990; Hao and Liu, 1994; Scholes et al., 1996a, 1996b]. The Australian contribution of biomass burning from savannas is ranked third after Africa and the Americas [Hao and Liu, 1994].

[3] Recent estimates of the extent of biomass burning in Australia, derived from various sources including interpretation of NOAA AVHHR (National Oceanic and Atmospheric Administration's advanced very high resolution radiometer) active fires and fire scar mapping, indicate that an average of 387,867 km2 of savanna and temperate grassland was burnt annually over the period 1990–1999 [National Greenhouse Gas Inventory Committee (NGGI), 2000]. Available estimates for northern Australian savannas for the period 1997–1999, derived from interpretation of NOAA AVHRR fire scar mapping, indicate that an average of 301,046 km2 was burnt annually [Russell-Smith et al., 2002], or 75% of the mean annual national estimate of 401,477 km2 for savanna and grassland burning over the same three year period [NGGI, 2000].

[4] The amount of biomass burnt derived from combined Australian grassland and savanna data is estimated as 155 Mt DM/yr for the 10-year period 1990–1999 [NGGI, 2000], or ∼116 Mt DM/yr for tropical savannas only, for the period 1997–1999, assuming 75% of the national total as above. Such estimates of the contribution of Australian savanna biomass burning are significantly lower than those derived from “classification” (in the sense of Scholes et al. [1996a]) approaches (e.g., 420 Mt DM/yr from savanna fires in Oceania [Hao et al., 1990]; 290 Mt DM/yr from all Australian forest and savanna fires [Hao and Liu, 1994]).

[5] In this paper, we consider the extent of savanna burning, and biomass consumption of fine (grass and litter) fuels, from an extensive region of northern Australia, for the period 21 August to 14 September 1999 encompassing the Australian continental component of the Biomass Burning and Lightning Experiment-B (BIBLE B) campaign. We restrict our detailed assessment of biomass burning to a ∼230,000 km2 region for which we have sufficient reliable data for various analyses presented. The assessments comprise (1) estimation of the extent of burning derived from interpretation of NOAA AVHRR imagery, qualified and calibrated by ground verification studies and mapping of fire scars from fine-resolution Landsat Thematic Mapper (TM) imagery, and (2) estimation of the quantity of fine fuels burnt using published regional fuel accumulation relationships and associated detailed vegetation mapping. The broader extent of biomass burning in Australian savannas for 1999 is considered in discussion.

2. Methods

2.1. Northern Territory Study Region

[6] The extent of burning, and derived estimates of biomass burning and trace-gas and particulate emissions, is assessed here primarily for a region encompassing 230,700 km2 of the Top End of the Northern Territory, Australia, north of 15°S (Figure 1). Despite the recent availability of fire maps for all northern Australia savannas derived from interpretation of NOAA AVHRR, this smaller region was selected given availability also of digital vegetation mapping at 1:1,000,000 scale, and associated plot data [Wilson et al., 1990], fuel accumulation relationships for major ground fuel types [Cook et al., 1995; Russell-Smith et al.,1998], and detailed fire mapping ground truth and calibration data for the period encompassing the BIBLE B field campaign.

Figure 1.

Location of study area and distribution of major ground fuel types, following Wilson et al. [1990]. The diagonal rectangle represents the area of the LANDSAT scene used for verification and calibration of fire scar mapping derived from NOAA AVHRR.

[7] The vegetation cover of the study region, as for much of northern Australia generally, ranges from open forest or woodland savanna, dominated by Eucalyptus over a range of highly flammable annual and perennial grasses, to hummock grasslands occupying rugged sandstone formations [Wilson et al., 1990]. With the exception of coastal and riverine floodplains, soils, where present, are typically deeply weathered and infertile. An implication of such poor soils is that grazing by introduced cattle (mostly Bos indicus) is restricted mostly to more fertile, run-on sites, especially floodplains.

[8] The major fire period occurs over the long dry season, typically between Apr/May–Oct/Nov. Rainfall, which occurs over most of the Australian savanna region between Oct–Mar under the influence of the Asian monsoon, declines rapidly from over 2000 mm pa in the north–west, to around 800 mm pa in the south. Although the amount of rainfall received in any one area is highly variable from year to year, the wet season is a highly reliable event [Taylor and Tulloch, 1985]. Such climatic conditions are conducive for the production of grassy fuels sufficient for carrying ground fires on an annual basis in higher rainfall areas, to once every few years under lower rainfall conditions [Walker, 1981; Williams et al., 2002]. Most fires are lit by people for a range of land management purposes, fires ignited by lightning at the start of the annual wet season are relatively few.

[9] Based on regional mapping of fires from satellite imagery (mostly NOAA AVHRR from the early 1990s, and Landsat since 1980), Russell-Smith et al. [2000] identified two broad patterns concerning the application of fire in northern Australia. In north–western and northern Australia, and possibly also on parts of western Cape York Peninsula in the northeast, vast tracts are burnt annually, typically by intense wildfires late in the dry season. Ecological studies indicate that such fire regimes are having catastrophic impact on native fire-sensitive species, communities, and habitats. Conversely, elsewhere across northern Australia but especially on more productive pastoral lands, the restricted application/absence of burning is leading in some situations to native and exotic woody species thickening/invasion, likewise with profound ecologic and economic consequences.

2.2. Extent of Burning

[10] The extent of burning in the study region was assessed for a 25-day period, encompassing the regional component of the BIBLE B experiment, from 21 August to 14 September 1999. For northern Australia, fire scar mapping is undertaken on a ∼9 day basis, coinciding with afternoon nadir overpasses of the NOAA 14 satellite carrying the AVHRR sensor (pixels ∼1.1 km2 at orbital nadir). Mapping of fire scars is currently undertaken on a regional basis with reference to bands 2, 3 and 5, utilizing a change analysis and visual interpretation technique [Craig et al., 2000; Langaas et al., 1999] by the Department of Land Administration (DOLA) in Western Australia. Active fire detection is also undertaken by DOLA using an automatic contextual algorithm for fire hot spot detection [Lee and Tag, 1990; Craig et al., 2000], utilizing nighttime overpasses only. For analyses presented here, fire scar mapping and active fire detection was undertaken on dates as described in Table 1. Mapping of fire extent between nine-day intervals was first vectorized, then interpolated with reference to the locations of daily fire hot spots where these were available in the absence of cloud cover (Table 1).

Table 1. Dates of Satellite Data Acquisition and Validation Exercisea
AVHRR Fire Scar MappingAVHRR Hot Spot Data AcquisitionLANDSAT Scenes for Fire Scar MappingAerial Validation
  • a

    Refer to section 2 for details. Read 18/08/99 as 18 August 1999.

  17/05/99 
   6/08/99
18/08/99   
 21/08/9921/08/99 
 23/08/99  
 25/08/99  
27/08/9927/08/99  
 29/08/99  
 31/08/99  
 2/09/99  
 4/09/99  
5/09/99   
 6/09/99  
 8/09/99  
 10/09/99  
 12/09/99  
13/09/99   
 14/09/99  
  30/09/99 

[11] As well as this broader regional mapping of fires from AVHRR, a further more refined assessment was undertaken for a subregion defined by a single Landsat TM image (Figure 1). For this assessment, AVHRR subregion data for each nadir overpass for the period April to 13 September 1999 (Table 1) were first registered to available 1:250 000 topographic mapping. Starting with images for the first two dates, a difference image was created using all available bands, followed by unsupervised Isoclass classification to determine burnt and unburnt areas. After prior masking of identified burnt areas at each iteration, the procedure outlined above was repeated for all consecutive nadir images resulting in a final fire scar map for the subregion.

[12] The above fire mapping derived from AVHRR was verified and calibrated with reference to independent mapping of fire scars for the sample subregion (Figure 1), utilizing relatively fine resolution Landsat TM imagery (pixels 25 × 25 m). Mapping of fire scars in the sample subregion from Landsat TM imagery was undertaken digitally by first masking out large areas where no fire activity was evident by applying a threshold to the thermal band (band 6), performing an unsupervised Isoclass classification using the remaining 6 bands, and then selecting burnt classes on-screen. Mapping was undertaken using imagery sampled early in the dry season, in May, and then for the ensuing period up until the start of BIBLE B assessment period (Table 1); no cloud-free Landsat imagery was then available until the end of September.

[13] Ground truth validation of the Landsat-derived fire map was undertaken using aerial transect data collected in early August, 16 days prior to the second sampled Landsat scene use for mapping (Table 1), following the methodology as outlined by Russell-Smith et al. [1997]. This aerial assessment was flown on the date of the previous Landsat overpass but the scene ultimately proved to be too cloudy. To use these aerial validation data, therefore, all areas affected by fire, as identified from AVHRR daily hot spots and associated fire mapping over the 16 day period, were first masked from the Landsat -derived fire map and associated ground truth data. The validation data were then overlaid over the fire map, and the error matrix calculated.

[14] Calibration of the actual extent of burning derived from both AVHRR mapping procedures (see above) was undertaken with reference to the more detailed Landsat mapping in the sample subregion as follows. First, fire scar maps derived from interpretation of both sensors for the sample subregion were intersected, and error matrices calculated. On the basis of this analysis, a corrected estimate of the extent of burning determined from AVHRR was derived for the whole study region. Second, fire map interpretations derived from both sensors were compared for 5 individual fire events, covering a range of fire sizes. And third, a least squares regression of the extent of burning derived from Landsat, against extent of burning derived from AVHRR, was undertaken for 10 × 10 km cells describing an internal grid within the sample subregion, following the general approach as outlined by Eva and Lambin [1998a].

2.3. Biomass Burning

[15] Estimates of the amount of biomass burnt throughout the study period were calculated as the product of area burnt, and fine (<6 mm diameter) grass and litter fuel loads (t.ha−1) expressed on a dry weight (DM) basis. Regional data for combustion of larger fuels (e.g., logs) are not yet available.

[16] The mapped distribution of major regional fuel types (i.e., Sorghum-dominated, Triodia-dominated, and floodplain; following Wilson et al. [1990]) is given in Figure 1. Given, however, that relatively little burning of floodplains was undertaken over the study period, for practical purposes fuel accumulation of floodplain fuels was estimated as for Sorghum-dominated fuels.

[17] Fuel loads were calculated for regional Sorghum and Triodia fuel types, using fuel accumulation relationships as given by Cook et al. [1995] and Russell-Smith et al. [1998], respectively (Figure 2). Given that the calculation of grass and leaf litter contributions to Sorghum-dominated fuels is dependent on tree stem basal area [Cook et al., 1995], mean Sorghum fuel load values for the study region were determined respectively for each of 44 vegetation types described by 760 individual polygons, using data available for 985 plots associated with the 1:1,000,000 scale vegetation map of the Top End of the Northern Territory [Wilson et al. 1990].

Figure 2.

Fuel accumulation curves for (a) sorghum: fuel load = (L/K) * (1-e(−K.TSB)), where K = 0.79 and L = annual fuel production = tree litter fall + grass production = (0.35 * TBA) + (3.5 − (0.125 * TBA), TSB is time since burnt, and TBA is total basal area = 6.84 m2/ha (mean value for vegetation units containing Sorghum north of 15°S). Expression from Cook et al. [1995]. Fuel accumulation curves for (b) triodia: fuel load = 4.04 log(TSB) + 2.82. Expression from Russell-Smith et al. [1998].

[18] Fuel accumulation over time (i.e., time since last burnt) was calculated with reference to GIS intersection of the mapped regional distributions of the two fuel types, with regional digital fire history map data derived from interpretation of AVHRR imagery for the period 1993–1998. Where time since last burnt was greater than the 6 years of available fire history, an arbitrary value of 10 years was assigned.

[19] The fuel accumulation data derived are an estimate of the maximum possible fuel combusted, giving a combustion efficiency of 1, at this stage we do not have data to describe the combustion efficiency under different fire regimes.

[20] In sum, biomass burning over the 25-day study period was calculated at least every two days as the product of the calibrated area of fire scars mapped from AVHRR, with detailed estimates of fine fuel load derived from available fuel accumulation relationships.

3. Results

[21] Verification and calibration data associated with finer-resolution fire scar mapping derived from Landsat TM imagery for the sampled subregion over the 1999 dry season, prior to the commencement of the BIBLE B campaign, are given in Tables 2a and 2b. Intersection of 1362 aerial transect validation points (Figure 3a) with the Landsat -derived fire map (Figure 3b) indicate an overall level of agreement of 84%; in general there was better agreement of mapped burnt areas than unburnt areas (Table 2a). These validation data thus inspire a generally high degree of confidence in the use of the Landsat -derived fire mapping as a basis for calibrating fire maps derived from NOAA AVHRR.

Figure 3.

Western Arnhem Land subregion: (a) aerial validation transects overlaid on fuels map; (b) LANDSAT TM-derived fire scar mapping for 1999, including vectors delineating individual fire events as per assessment in Table 2b; (c) NOAA AVHRR-derived fire scar mapping for 1999 by Department of Land Administration (DOLA), Western Australia; and (d) NOAA AVHRR-derived fire scar mapping for 1999 from this study.

Table 2a. Verification of LANDSAT TM-Derived Fire Scar Mapping From Aerial Transect Dataa
Mapped DataAerial Transect Data
BurntUnburntTotal
  • a

    Numbers in parentheses are percent agreement.

Burnt54268610 (88.9)
Unburnt149603752 (80.2)
Total691 (78.4)671 (89.9)1362 (84.1)
Table 2b. Calibration of Fire Maps Derived from NOAA VHRR Against LANDSAT TM-Derived Fire Mappinga
LANDSATAVHRR Mapping From DOLAbAVHRR Mapping From This Study
Burnt Area, km2Intersection, %Omission, %Commission, %Burnt Area, km2Intersection, %Omission, %Commission, %
  • a

    Numbers in parentheses are the ratio of burnt area mapped from NOAA AVHRR against LANDSAT.

  • b

    Department of Land Administration, Western Australia.

  • c

    Refer to Figure 3b for locations of individual fires.

Over entire LANDSAT scene: 4396 km23,873 (0.88)6535224,866 (1.11)792131
Five individual firesc
   1293 km21,365 (1.06)8911171390 (1.08)881219
   341 km2232 (0.68)524816394 (1.16)742642
   235 km2239 (1.02)772325318 (1.35)901045
   202 km2147 (0.73)554517179 (0.89)80209
   32 km252 (1.61)90107148 (1.48)831765
Mean (n = 5) 72.627.429.2 831736

[22] Calibration data for fire mapping from NOAA AVHRR undertaken by the Department of Land Administration (DOLA), Western Australia (Figure 3c), and for the sample subregion (this study; Figure 3d), against above fire mapping derived from Landsat TM, are provided in Table 2b. Overall, an area of 4396 km2 was mapped as burnt based on interpretation of Landsat. For the same subregion, fires mapped from AVHRR by DOLA described 3873 km2, and our own mapping described 4866 km2, or 0.88 and 1.11 the area estimate mapped from Landsat, respectively.

[23] Using a complementary regression approach for mapped Landsat and AVHRR fire extent data (km2) per 10 × 10 km grid cells (n = 219), highly significant regressions were obtained for predicting the “true” extent of burning as estimated from Landsat, against DOLA and our own fire mapping from AVHRR, as follows: Landsat fire extent = 0.87 × AVHRR fire extent + 3.5, R2 = 0.82 (DOLA); Landsat fire extent = 0.85 × AVHRR fire extent + 1.2, R2 = 0.91 (this study). In contrast at least to the underestimate of “true” fire extent using DOLA's AVHRR fire mapping for the whole Landsat subregion (Table 2b), the slopes of both regression expressions indicate that the actual extent of burning is slightly overestimated by mapping from AVHRR in the sample of grid cells used.

[24] The general correspondence between AVHRR and Landsat fire scar mapping described in the two statistical approaches described above, however, masks considerable variability in the mapping accuracy attained. Thus, overall intersection between AVHRR fire scar mapping produced by DOLA, and ourselves with Landsat, was 65% and 79%, with associated errors of commission of 22% and 31%, respectively (Table 2b). Such variable mapping accuracy is further illustrated with reference to calibration data presented for five individual fire events (refer to Figure 3b for locations). Accuracy ratios of mapped burnt extent (AVHRR versus Landsat) for individual fire events ranged from 0.68–1.61 (DOLA), and 0.89–1.48 (this study). Intersections of individual fires mapped from AVHRR with Landsat ranged from 52–90% (DOLA), and 74–90 % (this study), with corresponding errors of commission of between 16–71%, and 9–65%, respectively (Table 2b). While intuitively one might assume smaller departures from unity for accuracy ratios as fire events increase in area, no consistent trends are apparent in these data (Table 2b).

[25] Assuming that DOLA's fire mapping over the entire 230,000 km2 study region for the study period underestimates the “true” mapped fire extent similarly as apparent for the Landsat study subregion (Table 2b), the calibrated actual extent of burning was derived as the product of the area mapped from AVHRR by DOLA (37,560 km2) with 1.14 (i.e., 1/0.88 from Table 2b). Thus, over the 25-day study period an estimated total of 42,818 km2 of the study region was burnt. The distribution of the extent of burning over the study period was nevertheless highly variable (Figure 4a).

Figure 4.

(a) Extent of burning and (b) fine fuel biomass burnt on a two-day basis from 21 August to 14 September 1999 in the 230,000 km2 study region.

[26] Using published fuel accumulation relationships in conjunction with available time-since-last-fire data for two major fuel types (see methods), an estimated 18.75 Mt DM was burnt over the study period and, as with the distribution of the extent of burning, the amount of biomass burnt over respective two-day periods was highly variable (Figure 4b).

4. Discussion

4.1. Data Reliability

[27] Levine [1996, p. xxxvi] suggests that “perhaps the greatest single challenge to the scientific community studying biomass burning is to accurately assess the spatial and temporal distribution of burning over a given period of time, that is, weeks, months, or a year”. In order to overcome limitations associated with the estimation of area burnt based on regional assumptions of savanna fire frequency (e.g., the “classification” approach of Seiler and Crutzen [1980], Crutzen and Andreae [1990], and Hao and Liu [1994]), over the past decade various studies have attempted to apply potentially more accurate remote sensing estimation approaches (e.g., savannas [Eva and Lambin, 1998a, 1998b], tropical forests [Siegert and Hoffmann, 1999], and boreal forests [French et al., 1996]).

[28] In savannas, such approaches have included the calibration of active fire counts derived from AVHRR data to burnt extent estimates derived from interpretation of Landsat imagery [Setzer and Pereira, 1991; Pereira and Setzer, 1996; Scholes et al., 1996a]. Few studies, however, have actually used coarse resolution imagery to map savanna fires. Among these, Beringer et al. [1995] mapped fires in the Northern Territory, Australia, in 1992 using AVHRR; no validation of these data was undertaken. Eva and Lambin [1998a, 1998b] calibrated 1 km resolution ATSR-1 data over central Africa against Landsat and airborne video data. Regardless of the approach, significant errors are associated with the estimation of burnt extent given well documented limitations with applied optical remote sensing methods and technologies [e.g., Robinson, 1991; Eva and Lambin, 1998a].

[29] Using a multisensor approach, we have first verified the finer-resolution fire mapping derived from Landsat against aerial transect data. The resultant verified level of overall agreement (84%; Table 2a) conforms to similar verification exercises undertaken for regional mapping of fire scars with Landsat, for example, 83–86% [Russell-Smith et al., 1997] and 82–91% [Edwards et al., 2001]. Those authors also demonstrate that differential mapping errors are associated with different vegetation and landscape types. This has particular ramifications for the development and application of regional fire histories based on multiyear data [Russell-Smith et al., 1997]. It is notable that few biomass burning studies which adopt a multisensor approach (but see Eva and Lambin [1998a, 1998b]) qualify the reliability of their finer-resolution mapping.

[30] Second, we have calibrated the regional estimate of burnt extent derived from NOAA AVHRR based on the above Landsat assessment. This calibration was admittedly referenced against only one Landsat scene, but one which encompassed both major regional fuel types (Figure 1) and one for which we possessed independent verification data. Further, we compared available regional AVHRR fire map products provided by DOLA, with our own best estimate, and examined errors associated with the mapping of fires of different sizes (Table 2b).

[31] Collectively, these assessments indicate that (1) overall, available regional AVHRR fire map products for northern Australia compared satisfactorily with more detailed, localized assessments; (2) there was, however, significant variability with respect to the accuracy of mapping individual fires; (3) the extent of burning derived from relatively coarse resolution AVHRR imagery provided highly predictable estimates of “true” burnt extent, when regressed against finer resolution Landsat imagery as proposed by Eva and Lambin [1998a]; (4) nevertheless, depending on the statistic used, for the period May–September regional fire mapping provided by DOLA either slightly understated or overstated the actual area burnt; and (5) using active daily (albeit night time overpass), “hot spot” AVHRR fire data in association with fire mapping from 9-day nadir overpasses, we successfully interpolated the extent of burning on a two-daily basis.

[32] Given suitable qualification of fire map data, therefore, we contend that fire maps derived from coarse-resolution optical imagery can be reliably applied to estimate the extent of savanna fires over the major burning period in northern Australia, under relatively cloud-free dry season conditions. While inherent scaling difficulties are involved in estimating error associated with such mapping, from assessments presented here it would appear that the confidence with which mapping fire extent at regional scales (as opposed to mapping smaller individual fires) that can be achieved realistically lies generally between 69% (i.e., 84% accuracy of Landsat mapping, from Table 2, ×82% explained by regression of Landsat versus AVHRR mapping undertaken by DOLA, from results) and 82% (i.e., 90% accuracy of Landsat mapping potentially obtainable, see discussion above, ×91% explained by regression of Landsat versus AVHRR mapping undertaken in this study, from results).

[33] Fuel loads for the study region north of 15°S were derived from relationships established for widespread Eucalyptus savannas with a Sorghum-dominated understorey, and for Triodia- (or spinifex-) dominated hummock grasslands occupying rugged sandstone terrain [Russell-Smith et al., 1998]. These formations occupy 62% and 16% of the region, respectively [Wilson et al. 1990]. Fuel accumulation for Sorghum-dominated understorey fuels was derived from an expression [Cook et al., 1995] which incorporates grass and litter production over time as a function of woody biomass (expressed as basal area, or m2/ha), and taking into account annual decomposition. While this relationship may have general application to Sorghum-dominated communities, its relevance to other nonspinifex fuel types requires testing. From those sites whence the fuel accumulation relationship was derived (i.e., Kapalga Research Station in present-day Kakadu National Park), the coefficient of variation (CV) of total Sorghum-dominated fuel loads was 22%.

[34] Likewise, for spinifex-dominated fuels, the applied fuel accumulation relationship was derived from an assessment of fuel accumulation on plots (n = 33) occupying a limited geographic range. However, given that spinifex grasses typically are confined to rugged, seasonally xeric substrates, the actual length of the growing season (typically 4–6 months) in any one wet season probably does not vary significantly across this relatively high rainfall region. While spinifex typically comprises the bulk of the ground fuels, shrub and leaf litter components comprise perhaps a further third to one half the total (Russell-Smith et al., unpublished data, 2002); these latter fuel components are not considered in the spinifex fuel accumulation relationship published by Russell-Smith et al. [1998].

[35] Further, coarse fuels such as tree stems have not been taken into account in fuel load estimates. Typically, savanna forests dominated by Eucalyptus miniata and E. tetrodonta in the study region contain up to about 100 t/ha of live woody biomass, with savanna woodlands commonly carrying half this amount (R.J. Williams and G.D. Cook, unpublished data, 2000). In one study, five years of annual burning reduced live woody biomass by about 7 t/ha if fires occurred early in the dry season, but by nearly 28 t/ha under late dry season fires [Williams et al., 1999, R.J. Williams, unpublished data, 2000]. Most of these dead standing stems would be burnt by subsequent fires and, in the case of late dry season fire regimes, coarse fuel biomass may be equivalent to that of the fine fuel burnt over this five-year period (∼25 t/ha [Williams et al., 1998]). Clearly, further work is required to better estimate the biomass and long-term dynamics of regional fuels.

[36] Such work requires substantial field assessment of the dynamics of fine and coarse fuels across the major vegetation types of the region, including: improved knowledge of the production and decomposition of litter from both trees and grasses; and better understanding of fire behavior including fire return intervals, and burning efficiencies in early and late dry season fires. These results would then need to be extrapolated using various modeling approaches. With appropriate calibration to north Australia, the approach of Scholes et al. [1996a] in southern Africa would suffice for modeling fine fuel loads, but it takes no account of coarse fuels arising from the death of trees under high intensity fires. The recently developed FLAMES model (G.D. Cook and A.C. Liedloff, personal communication, 2001), which simulates the dynamics of both coarse and fine fuels under various fire management regimes, would be more satisfactory but requires further development.

4.2. Biomass Burning from Northern Australian Savannas in 1999

[37] In order to place the BIBLE-B campaign in a broader regional perspective, but cognizant of the above qualifications, we present below available data for biomass burning over the entire Australian savanna region for 1999. As presented by Russell-Smith et al. [2002], the extent of biomass burning from northern Australian savannas in 1999 was estimated as ∼417,500 km2, derived from interpretation of AVHRR imagery (Figure 5a). Of this total, 30% (136,405 km2) was burnt in the 230,000 km2 study region examined in this paper. The extent of biomass burning in northern Australian savannas in 1999 exceeds the average 387,867 km2 of combined Australian savanna and temperate grassland burnt annually over the period 1990–1999 [NGGI, 2000].

Figure 5.

(a) Two-monthly extent of fires in the Australian tropical savannas in 1999 from fire mapping undertaken by Department of Land Administration, Western Australia. (b) Estimate of fine fuels consumed per ha, for respective biogeographic regions [Thackway and Cresswell, 1995] of the Australian tropical savannas.

[38] For estimation of the quantity of fine fuels consumed from Australian savanna fires in 1999, a different approach to that used above for estimation of fine fuels in the Northern Territory study region is required given the absence of habitat-specific fuel accumulation models for the entire region. Generally, however, the amount of fuel consumed by savanna fires can be estimated using the following equation, which is used in the Australian greenhouse gas inventory [NGGIC, 1994]:

display math

where M is mass of fuel burnt in fires (tons), A is estimated area of fires (hectares), FL is fuel load (tons/hectare), and j is burning efficiency of fires.

[39] The area burnt is typically assessed using remote sensing tools such as NOAA AVHRR imagery. However, as considered above, not all the burnt areas defined by these techniques are actually subjected to fires, and this is taken into account in the burning efficiency factor. This factor combines two effects: (1) the ratio of net area burnt to gross area within the fire perimeter identified using tools such as remote sensing; and (2) within areas over which flames have passed, the ratio of fuel pyrolized to fuel load. Further work is required to provide reliable estimates for the subcomponents of the burning efficiency parameter, because they are likely to vary considerably across different land types and fire seasons. At present a value of 0.72 has been used in the Australian National Greenhouse Gas Inventory for all savanna fires.

[40] There are a range of estimates of fuel loads across the tropical savanna biome [e.g., Walker, 1981; NGGIC, 1994, 1996, 1997]. For the monsoon tall-grass and mid-grass systems which dominate much of the savanna belt of north Australia, Walker [1981] estimated typical fuel loads to be 4 t ha−1 and 1 t ha−1, respectively. Estimates used in the Australian National Greenhouse Gas Inventory are given on a State by State basis and have been revised on several occasions. For example, the estimates for the Northern Territory have varied from 5 t ha−1 (1994), down to 3 t ha−1 (1996), and up to 4.1 t ha−1 (1997), while similar ecosystems in Western Australia and Queensland were estimated to carry 8.3 t ha−1 and 3 t ha−1 (1997). The large magnitude of the variation in these estimates indicates the difficulty in providing an adequate synthesis of fuel dynamics across the range of land types, rainfall regimes, and land management regimes of northern Australia.

[41] A modeling approach applied at finer scales is arguably a better way to capture the variation in fuel loads. The GRASP model [Littleboy and McKeon, 1997] estimates total standing biomass of grass and grass litter, but because fine fuel in tropical savannas comprises both tree and grass litter, this does not provide a good estimate of total fine fuel. Modeling tools that explicitly incorporate the dynamics of trees and tree litter are under development (e.g., the FLAMES model of Cook and Liedloff [2001]), estimation of biomasss burning from tropical savannas rests with fuels modeling as opposed to assessing the extent of north Australia. We have thus applied the following approach for estimating fine fuel loads for northern Australian savannas.

[42] Fuel load, F (t ha−1), depends on the time since the last fire, t (years), and the balance between the rates of litter production, L (t ha−1 years−1), and litter decomposition, K, such that [Olson, 1963]

display math

[43] Here we use available records of fire scars based on NOAA AVHRR imagery for three years (1997–1999) to estimate the average fire return interval, derived as the mean of the reciprocal of the proportion of respective bioregions [Thackway and Cresswell, 1995] (Figure 5b) burnt in each year across the north Australian savanna lands. Annual rates of litter production and decomposition are derived from estimates given by Walker [1981], and average fuel loads when a fire occurs are estimated using equation (2). Applying equation (1), the total amount of fuel consumed by fires across this region in 1999 was then estimated as 212.3 Mt, assuming a combustion efficiency of 0.72 (Table 3). This is significantly greater than the combined estimated Australian grassland and savanna average of 155 Mt DM/yr for the 10-year period 1990–1999 [AGO-NGGI, 2000], or ∼116 Mt DM/yr for tropical savannas only, for the period 1997–9 using burning extent data given by Russell-Smith et al. [2002]. As noted previously by Beringer et al. [1995], such estimates of biomass burning are substantially lower than those reported previously by Hao et al. [1990] and Hao and Liu [1994], based on “classification” approaches; however, the contribution of coarse woody fuel components is unaccounted for in our estimate.

Table 3. Estimates of Extent Of Area Burnt, and Fine Fuels Consumed, in Northern Australian Savannas in 1999a
Biogeographic RegionArea, km2Area Burnt, km2Mean Fuel Load, t/haFuel Burnt (Including Burning Efficiency of 0.72), MtFuel Burnt per Hectare of Biogeographic Region, Mt/ha
Brigalow Belt North112,7583,1378.21.90.2
Cape York Peninsula115,48433,7667.718.81.6
Central Arnhem36,90323,8996.611.43.1
Central Kimberley76,89729,0907.615.92.1
Central Mackay Coast11,5082228.20.10.1
Daly Basin20,93210,9405.24.12.0
Dampierland89,60833,4387.117.11.9
Desert Uplands68,8109418.20.60.1
Einasleigh Uplands128,06110,3468.26.10.5
Gulf Coastal27,80310,4837.05.31.9
Gulf Fall and Uplands118,98342,3326.820.81.7
Gulf Plains211,58534,6798.220.51.0
Mitchell Grass Downs319,7551,0148.20.60.0
Mount Isa Inlier66,5773,1988.21.90.3
Northern Kimberley87,03134,4206.917.22.0
Ord-Victoria Plains125,17216,7318.19.80.8
Pine-Creek Arnhem51,58129,7075.511.72.3
Sturt Plateau99,71740,6677.822.82.3
Top End Coastal68,62729,9456.113.11.9
Victoria Bonaparte72,92928,5696.212.81.8
Total1,910,721417,524 212.35 

[44] The geographical distribution of estimated fuel consumption by fires in 1999 for Australian savannas is given for respective bioregions, expressed as the estimated fuel consumed divided by the total area of each bioregion (Figure 5b). Figure 5b thus takes into account not only an estimate of typical fine fuel loads, but also an estimate of the area burnt. The greatest rates of fuel consumption were in central Arnhem Land.

5. Conclusion

[45] The above assessment of the extent of biomass burning in northern Australia illustrates two general methodological issues. First fire maps derived from coarse-resolution optical imagery can be applied relatively reliably to estimate the extent of savanna fires, generally with 70–80% confidence using the approach adopted here, over the major burning period in northern Australia. And second, it is evident that substantial further field assessment and associated modeling of fuel accumulation, especially of coarse fuels, is required. Such fuels modeling is complicated by a number of factors, including the effects of herbivory (especially by cattle in certain Australian savanna regions), and interactions with fire intensity on death of woody stems. In contrast to Levine [1996], therefore, our assessment indicates that the greater challenge for reliable estimation of biomass burning from tropical savannas rests with fuels modeling, as opposed to assessing the extent of burning.

Acknowledgments

[46] Fire mapping derived from NOAA AVHRR was provided by the Department of Land Administration (DOLA), Western Australia, under contract to the Tropical Savannas Cooperative Research Centre, Darwin, and the Bushfires Council of the Northern Territory, Darwin. Lisa Roeger undertook the interpolation of daily fire mapping for the period 21 August to 14 September 1999, and Cameron Yates undertook the GIS assessment of fire extent for the northern Australian savanna region, 1999. Bob Yokelson and two anonymous reviewers are thanked for their critical comments on earlier drafts.

Ancillary