Estimation of carbon dioxide emissions from the megafires of Australia in 2019–2020

Catastrophic fires occurred in Australia between 2019 and 2020. These fires burned vast areas and caused extensive damage to the environment and wildlife. In this study, we estimated the carbon dioxide (CO2) emissions from these fires using a bottom-up method involving the improved burnt area approach and up-to-date remote sensing datasets to create monthly time series distribution maps for Australia from January 2019 to February 2020. The highest monthly CO2 emissions in Australia since 2001 were recorded in December 2019. The estimated annual CO2 emissions from March 2019 to February 2020 in Australia were 806 ± 69.7 Tg CO2 year−1, equivalent to 1.5 times its total greenhouse gas emissions (CO2 equivalent) in 2017. New South Wales (NSW) emitted 181 ± 10.2 Tg CO2 month−1 in December 2019 alone, representing 64% of the average annual emissions of Australia from 2001–2018. The negative correlation observed between CO2 emissions and precipitation for 2001–2020 was 0.51 for Australia. Lower than average precipitation and fires in high biomass density areas caused significant CO2 emissions. This study helps to better assess the performance of climate models as a case study of one of the major events caused by climate.


Scientific Reports
| (2021) 11:8267 | https://doi.org/10.1038/s41598-021-87721-x www.nature.com/scientificreports/ Pickrell 4 and Nolan et al. 2 documented the extent of the land areas affected by these fires up to December 2019. The fires continued in NSW and Victoria until February 2020 3 , and CO 2 emissions for January and February 2020 have not been quantified. Therefore, we estimated the CO 2 emitted by the Australian fires until February 2020 and created monthly CO 2 emission maps from these fires to understand changes in the time series and distribution of the fires across the whole of Australia. This paper covers the following: (1) providing CO 2 emissions and its spatio-temporal distribution; (2) quantitation of the effect on the CO 2 emission estimation by input sources; (3) evaluation of the relationships between CO 2 emissions and precipitation and CO 2 emissions and temperature. It is important to evaluate many case studies of major events to understand the global environment. As an assessment of one of the major events, this study helps to better assess the performance of climate and fire models.

Results
CO 2 emissions from fires were estimated for six states and one territory (Fig. S1). The spatio-temporal distribution of estimated monthly CO 2 emissions between January 2019 and February 2020 are shown in Fig. 1. CO 2 emissions, which began in northern Western Australia (WA) in March 2019, passed through Northern Territory (NT) and reached Cape York Peninsula in Queensland in June 2019. While continuing to emit CO 2 in the northern area of the country, the CO 2 emissions increased in the eastern parts of both Queensland and NSW from April 2019. Afterwards, CO 2 emissions were estimated in Queensland and NSW from November to December 2019, and in Victoria in January 2020. Other CO 2 emissions were also noted in southwestern WA from April to May 2019. Although CO 2 emissions were estimated in Queensland in February 2020, emissions had disappeared in most of the regions by this time (Fig. 1).
Annual CO 2 emissions from fires in Australia for 2019 were 674 ± 57.6 Tg CO 2 year −1 , which was estimated to be 2.4 times the average annual CO 2 emissions in 2001-2018 (Table 1). Focusing on the monthly CO 2 emissions, the average emissions from April to October 2001-2018 were consistently large and the emissions in 2019-2020 season increased again after October. In particular, the emissions in November, December 2019, and January 2020 were 157 ± 23.1, 304 ± 16.9, and 173 ± 6.14 Tg CO 2 month −1 , respectively. These emissions correspond to 3.6, 9.7, and 6.4 times the average monthly CO 2 emissions in 2001-2018, respectively.
To comprehensively understand the CO 2 emissions from the fires, we focus on the emissions in each state. The largest estimated monthly CO 2 emissions were 181 ± 10.2 Tg CO 2 month −1 in NSW, in December 2019 (Table 2). CO 2 equating to 52% of the emissions from the fires in Australia was emitted from NSW across the 14 months evaluated in this study. The emissions from NSW in the latest year, March 2019 to February 2020, were 443 ± 40.4 Tg CO 2 year −1 , equivalent to 1.6 times average annual emissions through Australia in 2001-2018. In Victoria, 126 ± 8.50 Tg of CO 2 was emitted over just two months between December 2019 and January 2020, and the annual emissions in the latest year, March 2019 to February 2020, were 149 ± 14.3 Tg CO 2 year −1 , equivalent to 53% of Australia's average annual emissions for 2001-2018.
Although the highest monthly CO 2 emissions through Australia since 2001 occurred in December 2019, the largest burnt area did not occur in this month (Fig. 2). The burnt areas in September 2011 and October 2012 were both larger than that of December 2019, however CO 2 emissions were not as high during this month (46.8 ± 5.07 and 57.3 ± 7.58 Tg CO 2 month −1 , respectively). One of the reasons for this phenomenon is the aboveground biomass (AGB) density in the fire regions. We estimated the CO 2 emissions from fires by multiplying the burnt areas by the AGB densities and few coefficients (Eqs. (1) and (2)). The inland areas in NT and WA, which comprised most of the burnt area in September 2011 and October 2012, respectively, had relatively low AGB distributions. Eastern NSW, however, which exhibited the highest emissions in December 2019, had a higher AGB distribution than the above two regions (Fig. 3).
The annual precipitation in 2019 was only 53% of the average for 2001-2018, with the monthly precipitation values in each region in 2019 often below the average precipitation of 2001-2018 (Fig. 4). The annual precipitation in 2019 in NSW, Queensland, and Victoria, the areas that emitted large amounts of CO 2 , were 50%, 73%, and 73%, respectively, of the average precipitation of each region for 2001-2018. The precipitation in NSW, Queensland, and Victoria for the three months from October to December 2019 was 29%, 29%, and 57%, respectively, of the average precipitation for the same months for 2001-2018. This was one month ahead of the period with significantly increased CO 2 emissions (from November 2019 to January 2020). Focusing on the relationship between CO 2 emissions, which was conducted using base 10 log transformation, and precipitation, all regions excluding SA had a negative correlation coefficient including Australia with the negative correlation (0.51) as shown in Figure S2. These results indicate that lower than average precipitation was one of the causes inducing the significantly greater CO 2 emissions for the 2019-2020 seasons. Figure 5 shows the results of the monthly CO 2 emissions and the monthly mean temperatures 14 between the average values in 2001-2018 and the values in 2019-2020, respectively. The monthly mean temperatures for evaluation in each region (six states and one territory) were used in the capitals in each region as a representation. The temperatures in each region in 2019-2020 were on average 2-5% higher than those of 2001-2018. Although the CO 2 emissions in NSW from November 2019 to January 2020 and in Victoria from December 2019 to January 2020 were conspicuously higher than those of 2001-2018, the significant difference does not prevail when compared with the other regions and the other periods. There was a weak correlation between the CO 2 emissions and the temperatures in Tasmania, Victoria, and NT; however, no evident relationship was found in the other regions (Fig. S3). The higher-than-normal temperatures may have little effect on CO 2 emissions directly; however, they indirectly contribute to the expansion of burning areas and CO 2 emissions by causing the drying of fuel and soil.  Table S2 for every inventory). The highest CO 2 emission inventory was MWL by 69.8 ± 92.6 Tg CO 2 month −1 , the smallest inventory was GEN by 55.7 ± 76.1 Tg CO 2 month −1 , and the difference was 25% ( Fig. S4 and Table S2).
To understand the effect of AGB on the emissions, we compared the inventories with the same inputs of LC and FD (GWN and GEN, GWL and GEL, MWN and MEN, and MWL and MEL, respectively). The inventories using GEOCARBON had 11-14% more emissions than those of Globbiomass (Table S3). One of the reasons is       www.nature.com/scientificreports/ the difference in AGB density between AGB maps. The AGB density in GEOCARBON is 14% higher in Australia than those of Globbiomass.
To understand the effect of LC on the emissions, we compared the inventories with the same inputs of AGB and FD (GWN and MWN, GWL and MWL, GEN and MEN, and GEL and MEL, respectively). The inventories using MCD12Q1 had 4-6% more emissions than those of GLC2000. CO 2 emissions from forest areas with high AGB density were generally greater than the other LC areas. However, although the emissions of inventories using MCD12Q1 were greater than those of GLC2000, the forest areas in GLC2000 were 11% larger than MCD12Q1 (Table S4). The large difference region for the forest area is NT with low AGB density. Furthermore, NSW, Tasmania, and Victoria, where there is high AGB density, were evaluated to be 8-14% larger in forest area than GLC2000 on MCD12Q1. These results indicate the AGB density is more effective in CO 2 emission estimation than LC.
To understand the effect of FD on the emissions, we compared the inventories with the same inputs of AGB and LC (GWN and GWL, GEN and GEL, MWN and MWL, and MEN and MEL, respectively). The inventories using LC-M had 6% larger emissions than those of NC-M. The method of creation of FD maps caused the difference. The LC-M was created from the three confidence flags on MOD14Q1 and includes the whole burnt area of the NC-M created from the two flags. We consider input sources to be of influence on the CO 2 emission estimation, especially AGB density.
Comparison with previous studies. Previous studies calculated burnt areas to cover 3.0 million hectares in the eastern states of Queensland and NSW 4 , 3.8 and 0.5 million hectares in the temperate forest of NSW and Victoria, respectively, for the fire season until 29-12-2019 2 , and 5.8 million hectares of temperate broadleaf forest across NSW and Victoria between September 2019 and early January 2020 5 (Table S5). We measured the burnt areas from NC-M and LC-M to be approximately 4.3 and 4.5 million hectares, respectively, in NSW; 2.8 and 3.0 million hectares, respectively, in Queensland; and both 0.6 million hectares in Victoria between September and December 2019. Although the evaluated area and period do not completely match with the three previous studies, our results were 12-16% higher for NSW and 17% higher for Victoria than those of Nolan et al. 2 , 58-60% higher than those of Pickrell 4 , and 6-12% higher than those of Boer et al. 5 . The difference in the burnt areas between our results and the previous studies may be because we measured the burnt areas for the entire states, whereas there is the possibility that the previous studies concentrated on forest areas. MOD14A1 was mainly updated to decrease the omission errors in fire detection of all sizes and the obscuring fire detection by thick smoke. However, with MOD14A1, burnt areas were larger than actual owing to the low spatial resolution (1 km) because a burnt grid may have areas that are not burnt, smoldering areas, or flame areas. The factors that contributed to smaller evaluation for the burnt areas are the detection omission by burning periods outside satellite observation timing and the fire detection failure due to thick smoke or cloud cover as it is reported that the fire detection rate is 84% in Australia 15 . The CO 2 emission estimation using a newly burned area product and the development of an accurate burned area product will be studied in future work.
The combined CO 2 emissions of Australia and New Zealand (AUST region in Fig. S5) were determined to be comparable with previous studies (Fig. 6). Data from the Global Fire Emissions Database (GFED4.1 s) 9 and Global Fire Assimilation System (GFASv1.2) 7 were used to estimate average monthly CO 2 emissions from January 2003 to December 2019 to be 32.4 and 38.6 Tg CO 2 month −1 , respectively. Our estimated CO 2 emissions (an average of one standard deviation for the eight results for combined input sources) in the same period were 33.5 ± 7.59 Tg CO 2 month −1 . These were 3% larger and 15% smaller than that of GFED4.1s and GFASv1.2, respectively. However, the uncertainty (one standard deviation) of our estimated emissions by input sources is 23%, and the average emissions in both of GFED4.1s and GFASv1.2 were within the uncertainty. One of the reasons for these products showing relatively close values against the different estimation approaches is that the product used a common input dataset. GFED4.1s uses GEOCARBON for adjusting the AGB as one of the input sources of the Carnegie-Ames-Stanford Approach model, which is the basis for calculating the carbon pools 9 . Furthermore, GFASv1.2 sets the several scaling parameters for the estimation to fit the emissions of GFED 9 . www.nature.com/scientificreports/ Uncertainty. The uncertainty in the estimated CO 2 emissions was propagated from the remote sensing data, scaling coefficients, and features of this method itself. Regarding the uncertainty of remote sensing products, the overall fire detection rate of MOD14A1 has been calculated to be 84% for Australia 15 ; the overall accuracies of GLC2000 and MCD12Q1 for LC maps were 68.6% 16 and 73.6% 17 , respectively; and the root mean square error values of GEOCARBON 18 and Globbiomass 19 for the AGB maps were 87-98 Mg ha −1 and 52.8 Mg ha −1 , respectively. The combination of these remote sensing datasets, which were used for the CO 2 emissions estimation as inputs, introduced significant deviation into the estimation results. Our method used one-time fire instance for estimation and did not consider the burning term or the fire scale. Although the incinerated biomass density is considered in Eq. (2), biomass growth and recovery were not considered. These uncertainties influence each other and complicate evaluations of estimation results, which means that it is difficult to specify the uncertainty, similar to previous studies 9 .
Although we used BE and EF data sourced from Mieville et al. 20 and Shi et al. 21 as shown in Table S1, several authors have reported other values for Australia's regions and LC categories. Regarding BE and EF for temperate forest, Paton-Walsh et al. 22 reported the values of 0.88-0.91 and 1620 ± 160, respectively; furthermore, Guérette et al. 23 reported 0.89-0.91 and 1620 ± 160 in NSW, 0.91-0.93 and 1650 ± 170 in Victoria, and 0.88 and 1621 ± 160 in Tasmania, respectively. Similarly, for savanna, Smith et al. 24 reported the values of 0.90 ± 0.06 and 1674 ± 56, respectively, whereas Desservettaz et al. 25 reported 0.90 ± 0.06 and 1536 ± 154, respectively. In forest, BE contributes particularly high levels of CO 2 emissions and are 55-73% higher than the value we used. Therefore, to understand the impact of BE and EF on CO 2 emission estimation as an examination, we estimated CO 2 emissions using the BE and EF (g Kg −1 ) values of 0.895 and 1620 for forest, and 0.90 and 1613 for savanna, respectively, for January 2001 to February 2020 (Fig. S6). The estimated average monthly CO 2 emissions was 40.0 ± 6.19 Tg CO 2 month −1 , which is 1.8 times higher than our result shown in Sect. 3. Both BE and EF are known to change, depending on the season and precipitation levels; they further strongly influence CO 2 emission estimations from fires. Further verification of our results is required by comparing with the atmospheric concentration of CO 2 using a top-down method, because the emissions estimated in this experiment were high compared to previous studies.

Conclusions
This study presents the monthly changes in the time series and distribution of CO 2 emissions from Australian fires across 2019-2020. In our results, although the burnt area was not the largest to have occurred since 2001, the CO 2 emissions from this period were the highest, by 806 ± 69.7 Tg CO 2 year −1 from March 2019 to February 2020. The emissions in the latest year were equivalent to 2.9 times the average annual emissions in 2001-2018, and 1.5 times total GHG emissions without land use, land use change and forestry emissions of CO 2 equivalent for the whole of Australia in 2017. We found that lower than average precipitation and fires in high biomass density areas caused large CO 2 emissions, and there was a correlation between CO 2 emissions and precipitation for 2001-2020. The CO 2 emission inventories shown in this study will be opened to include all inventories by combining them into an input dataset. The scope for future research in this topic includes a reflection of the time series change of biomass density and the incorporation of the scale and duration of fires into the estimation method to reduce the uncertainty associated with estimated CO 2 emissions. Optimal BE and EF scenarios based on seasonal and precipitation changes, comparison of our estimated result with atmospheric concentrations, and the effect analysis of the emissions on regional/global carbon cycle need to be determined in future research. We expect that the CO 2 emissions estimation and its evaluations from the catastrophic fires in Australia help to better assess the performance of climate and fire models.

Materials and methods
Estimation method of fire CO 2 emissions. The remote sensing datasets were resampled at a 500 m spatial resolution, using the NEAREST function in ArcGIS version 10.5 to match the same spatial resolution. As MOD14A1 is a daily dataset, we created monthly burnt area datasets, including the number of fires occurring, to evaluate the burnt biomass in more detail. CO 2 emissions from fires (EMISSION, g CO 2 ) were generally calculated using Eq. (1) 20,21,26 . However, this equation cannot evaluate the number of fires occurring within a single region over a period. Therefore, here we represented the decrease in biomass density by fires over a year using Eq. (2) to determine the AGB density in Eq. (1), though this method does not consider annual changes in biomass density.
where m is the target month for calculating CO 2 emissions, p is the grid position on the map, c is the LC categories of the grid (p), i and I are the cumulative number of fire occurrences until the last month (m-1) and the target month (m), respectively, BA is the burnt area (m 2 ), BD is the total burnt biomass density (kg m −2 ), Agb is the biomass density (kg m −2 ) from the AGB map, BE is the burning efficiency (0 to 1), and EF is the emission factor of dry matter (g CO 2 kg −1 ). We assigned the BE and EF values sourced from Mieville et al. 20 and Shi et al. 21 to fit the categories of GLC2000 and MCD12Q1, respectively, as shown in Table S1. Finally, the eight types of estimated CO 2 emissions were combined into input datasets (2 3 ), namely two FD maps (NC-M and LC-M), two LC maps (GLC2000 and MCD12Q1), and two AGB maps (GEOCARBON and Globbiomass), were applied as an ensemble average to estimate optimal CO 2 emissions.
(1) EMISSION (m,p) = BA (m,p) · BD (m,p) · BE (c) · EF (c) , www.nature.com/scientificreports/ Remote sensing data. The remote sensing products of FD, LC, and AGB were used to estimate CO 2 emissions from fires. FD maps were used with the Thermal Anomalies and Fire MODIS data product version 6 (MOD14A1), which provides daily fire data with 1 km spatial resolution 27,28 . Every fire pixel is assigned as having either low (0-30%), nominal (30-80%), or high (80-100%) confidence levels 29 . We used two types of FD maps with data on the number of fire occurrences, dependent on confidence level: NC-M, with high and nominal confidences; and LC-M, with high, nominal, and low confidences. We counted the number of fire occurrences recorded on the maps, and an ongoing fire on the same grid position in MOD14A1 daily datasets was considered a single fire.
LC maps, the Global Land Cover 2000 Project (GLC2000) data product 30,31 and the MODIS Land Cover Type (MCD12Q1) Version 6 data product 17,32 , were used to obtain optimal scaling factors for each LC category. GLC2000 is a global LC map for the year 2000 and has 1 km spatial resolution. MCD12Q1 comprises a series of global LC maps from 2001 to 2018, with 500 m spatial resolution. The land use types used for the LC category were obtained from the Food and Agriculture Organization Land Cover Classification System (LCCS) for GLC2000 and from the International Geosphere-Biosphere Program for MCD12Q1. Note that MCD12Q1 of 2019 was applied to estimate the CO 2 emissions for 2020, because the 2020 datasets were not published at the time of study.
AGB maps, namely the GEOCARBON global forest biomass map 18,33 and the Globbiomass AGB map 34 , were used. The GEOCARBON map is a global AGB map with 1 km spatial resolution. Globbiomass is also a global AGB map with 25 m resolution; it is produced by the European Space Agency (ESA) 19 .
The Global Precipitation Measurement (GPM) level 3 product, with 0.1 degrees spatial resolution and monthly temporal resolution 35 , was used to evaluate the relationships between the estimated CO 2 emissions, burnt areas, and precipitation. The monthly mean temperatures from ClimatView system 14 from Japan Meteorological Agency were used to evaluate the relationship between CO 2 emissions and temperatures.