We estimate the emissions of carbon monoxide (CO) and black carbon (BC) from open vegetation fires in the Southern Hemisphere Africa from 1998 to 2005 using satellite information in conjunction with a biogeochemical model. Monthly burned areas at a 0.5-degree resolution are estimated from the Visible InfraRed Scanner (VIRS) fire count product and the MODerate resolution Imaging Spectroradiometer (MODIS) burned area data set associated with the MODIS tree cover imagery in grasslands and woodlands. The monthly fuel load distributions are derived from a 0.5-degree terrestrial carbon cycle model in conjunction with satellite data. The monthly maps of combustion factors and emission factors are estimated using empirical models that predict the effects of fuel conditions on these factors in grasslands and woodlands. Our annually averaged effective CO and BC emissions per area burned are 27 g CO m−2 and 0.17 g BC m−2 which are consistent with the products of fuel consumption and emission factors typically measured in southern Africa. The CO and BC emissions from open vegetation burning in southern Africa range from 45 Tg CO yr−1 and 0.26 Tg BC yr−1 for 2002 to 75 Tg CO yr−1 and 0.42 Tg BC yr−1 for 1998. The monthly averaged burned areas from VIRS fire counts peak earlier than modeled CO emissions. This characteristic delay between burned areas and emissions is mainly explained by significant changes in combustion factors for woodlands in our model. Consequently, the peaks in CO and BC emissions from our bottom-up approach are identical to those from previous top-down estimates using the Measurement Of the Pollution In The Troposphere (MOPITT) and the Total Ozone Mapping Spectrometer (TOMS) Aerosol Index (AI) data.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Biomass burning releases many trace gases and aerosols, such as carbon monoxide, oxygenated volatile organic compounds and carbonaceous aerosols, into the atmosphere, which have a significant influence on ambient air quality and climate [Crutzen et al., 1979; Andreae and Crutzen, 1997]. Most of the biomass burning takes place in the tropical belt [Crutzen and Andreae, 1990]. However, fire statistics are rather insufficient in many areas of the tropical regions, even though the savanna and grassland fires are mainly associated with human practices. Thus many researchers have utilized satellite information to quantify the emissions from the biomass burning [Chuvieco and Kasischke, 2007].
 The MODerate resolution Imaging Spectroradiometer (MODIS) burned area data product is available at a 500 m resolution only for 2 months for the year of 2000 [Roy et al., 2002]. Using the MODIS burned area data, Ito and Penner [2005b] showed that the estimated ranges from regional modeling approaches for carbon monoxide (CO) emissions from open biomass burning were within the range of the estimates constrained by chemical transport models and measurements in 2000 [Arellano et al., 2004; Pétron et al., 2004]. Thus the MODIS burned area product may serve as a useful data to calibrate the active fire counts to the burned areas in the Southern Hemisphere Africa until long-term burned area data sets become available.
 In southern Africa, Swap et al.  showed the apparent inconsistency of the results between the commonly held understanding maximum burning seasons of August and September and the satellite measurements of burned areas, which indicated that the largest burned areas were observed during June and July [Silva et al., 2003]. The updated version of this burned area data set [Tansey et al., 2004] was used to estimate the seasonal variations of CO emissions, which also showed the peak in July [Ito and Penner, 2004; Sinha et al., 2004; Korontzi, 2005; Jain et al., 2006]. Here the peak in monthly data is defined as the calendar month during which the maximum value is estimated for a year. The 1- to 2-month offset was also reported between the peak in the MODIS fire counts [Justice et al., 2002] and the peaks in the Measurement Of the Pollution In The Troposphere (MOPITT) [Edwards et al., 2006] and the SCanning Imaging Absorption spectroMeter for Atmospheric CHartographY (SCIAMACHY) [Gloudemans et al., 2006] CO retrievals in southern Africa.
 In addition to the fire affected areas, seasonal variations in the CO emissions are also influenced by the available biomass for combustion (i.e., fuel load) and burning conditions (i.e., flaming and smoldering fires). Litter fall occurs as the dry season progress, so that the amount of leaf litter increases seasonally [Hoffa et al., 1999]. van der Werf et al.  used a biogeochemical model to represent the seasonality in the delivery of leaves to the litter pools.
 The combustion factor (CF) and the emission factor (EF) are mainly characterized by composition, availability, and moisture content of biomass. To exploit the spectral signature of green vegetation for a measure of the fuel moisture content, the strong absorption of chlorophyll in the red and the high reflectance of vegetation in the near-infrared portion provides a useful information [Tucker, 1979]. The Normalized Difference Vegetation Index (NDVI) is calculated as the difference in reflectance between the near-infrared and visible red wavelengths divided by their sum. Korontzi  used the monthly NDVI data composites at a spatial resolution of approximately 64 km2 to represent the temporal and spatial variations of the regional fuel moisture condition, and related them to CF and EF. Verbesselt et al.  used the measurements of time series greenness from VEGETATION instrument on board the Systèm Pour l'Observation de la Terre (SPOT) at finer spatial resolution of about 1 km2 to examine them for a fire risk assessment.
 The purpose of this paper is to improve the mismatch of 1–2 months, by incorporating the seasonal variations of burned areas with fuel loads, combustion factors, and emission factors at a fine resolution. For this purpose, satellite information and a biogeochemical model are used to estimate CO and black carbon (BC) emissions from open biomass burning in southern Africa during the 8 year period from 1998 to 2005. Section 2 describes data sets used for estimating the CO and BC emissions from open vegetation fires using satellite data and a biogeochemical model. Section 3 shows multiyear estimates of CO and BC emissions and comparisons with previous estimates. We also present an analysis of the sensitivity of the CO emissions to the fire products and effective CO emissions per area burned to explore the relative importance of temporal and spatial variations in different treatments of burned areas and seasonality in fuel conditions. Section 4 presents a summary of our findings.
2. Materials and Methods
 In a modeling approach, the emissions of gases and particles per month ([Q]lit g species month−1) within a land cover type (l) at location (i) for month (t) from open vegetation fires can be described by the following equation [Seiler and Crutzen, 1980]:
where X is chemical species, BA is the burned area per month (m2 month−1), FL is the fuel load (kg m−2) expressed on a dry weight (DM) basis within each grid, CF is the combustion factor, and EF is the emission factor in grams of species per kilogram of dry matter burned.
2.1. Area Burned
 Monthly burned areas at a 0.5-degree resolution are estimated from the Visible InfraRed Scanner (VIRS) fire count product [Giglio et al., 2003] and the MODIS burned area data set [Roy et al., 2002] associated with the MODIS tree cover (TC) imagery [Hansen et al., 2003] in grasslands and woodlands. Monthly 0.5-degree VIRS fire product is available for the Tropical Rainfall Measuring Mission (TRMM) latitudes from 38°N to 38°S, which covers the entire African continent [Giglio et al., 2003]. The MODIS burned area data set was obtained on the approximate day of burning at a 0.5 km resolution for 2 months in 2000 [Roy et al., 2002]. We used the pixels with burns occurring in July and September and on the last day of the previous months [Roy, 2003]. The MODIS tree and herbaceous cover data sets were collected at a resolution of 0.5 km × 0.5 km from November 2000 to December 2001 [Hansen et al., 2002]. Comparing the validation database in Zambia with a subset of the global MODIS TC map in a validation exercise showed that the general tree cover distribution was well represented [Hansen et al., 2006].
 In southern Africa, two classifications of land cover types (i.e., grasslands and woodlands) have been used to estimate the emissions from open vegetation fires [Korontzi et al., 2003; Sinha et al., 2004; Ito and Penner, 2005b]. Following these studies, land cover types were classified into two classes, based on the MODIS percentage of tree cover. Grasslands were defined as those areas with less than or equal to 15% tree cover, and woodlands were defined as those regions with greater than 15% tree cover.
van der Werf et al.  related burned area at a spatial resolution of 1° to the VIRS active fire counts and fractional tree cover using a linear regression model. Giglio et al. [2006a] calibrated the MODIS active fire observations to the burned areas at 1° spatial resolution as a function of tree and herbaceous vegetation cover, and the mean size of monthly cumulative fire (i.e., pixel clusters). On the basis of these studies, we developed equations between the ratio of fire counts to burned areas and tree cover in grasslands and woodlands, separately. The 0.5-km resolution MODIS burned areas in grasslands and woodlands were aggregated to 0.5° resolution. The ratio between burned areas and fire counts was binned in each one percentage of tree cover, and then the average of the ratio was related to the percentage tree cover, which was the average value for the corresponding 0.5° grid cells.
 For grasslands (Figure 1a), we find the following linearly decreasing function of the mean tree cover (%) for the ratio of fire count (FC) to burned area (km2),
The negative slope in the equation (2) is explained by one possible factor that increasing tree cover reduces the fire spread rate and increases the detection probability of hot spots and thus decreases the effective burned area per fire count [van der Werf et al., 2003; Giglio et al., 2006a].
 For woodlands (Figure 1b), we find the following two linear functions of tree cover,
In woodlands, intensity of fires is likely increased by the amount of fine fuel load [Hoffa et al., 1999], which may increase as tree cover increases because more fine litters fallen from trees are available for open fires. Thus the positive relationship in the equation (3) and negative in (4) may be the combined results of the two opposite factors of fuel injections for fire intensity and of obstacles against fire spread. We note that the uncertainties associated with any calibration methods are likely to be comparatively large. The MODIS burned area of Roy et al.  (0.81 × 106 km2) for the sum of July and September was smaller (30%) than our estimate (1.16 × 106 km2). This amount of difference was also found in the study between the different calibration techniques of Giglio et al. [2006a] (0.83 × 106 km2 yr−1) and van der Werf et al.  (1.16 × 106 km2 yr−1).
 In addition to the burned area observations, the errors in the estimated burned areas also depend on the accuracy of the fire count data. Thus we examine the sensitivity study to investigate the effect of different seasonal variability of fire counts on modeled CO emissions. In southern Africa, VIRS fire counts [Giglio et al., 2003] peak about two months earlier than the Along Track Scanning Radiometer (ATSR) [Arino and Plummer, 2001] and the MODIS [Justice et al., 2002]. We used the Collection 4, version 3 Terra MODIS monthly Climate Modeling Grid (CMG) fire product at a 0.5° spatial resolution from January 2001 through December 2005 [Giglio et al., 2006b]. The method used to produce the seasonal variations is based on the overlap period between the VIRS and the MODIS fire data sets. The monthly 0.5-degree burned areas and the MODIS fire counts were summed to annual data in southern Africa, and then the monthly variations of burned areas were scaled and distributed by the MODIS fire counts at 0.5-degree resolution.
2.2. Fuel Consumption and Emission Factors
 Our fuel consumption model considers herbaceous vegetation in grasslands and that plus the fine litter of leaves fallen from trees in woodlands. The primary biomass density was derived from a 0.5-degree terrestrial carbon cycle model from 1998 to 2003. The ecosystem model used in this study is a modified version of Simulation model of Carbon cYCle in Land Ecosystems (Sim-CYCLE) [Ito and Oikawa, 2002; Ito et al., 2006], which is composed of 18 carbon pools representing the canopy tree layer, floor C3 grass layer, floor C4 grass layer, soil surface litter, and mineral soil. Carbon flows such as photosynthesis, respiration, and decomposition are calculated using physiological relationships with moderate simplicity for the global-scale applications. The model parameters were calibrated using field data from a number of sites, so that the model validation shows a good agreement with observations of carbon storage and productivity [Ito and Oikawa, 2002]. The carbon pools for herbaceous living biomass and tree fine litter are taken from the calculations of the process-based model [Ito et al., 2006]. The majority of fires in southern Africa are surface fires and the woody debris is generally collected for use as fuels [Shea et al., 1996; Desanker et al., 1997; Hoffa et al., 1999]. Thus the coarse fuels (i.e., stems and coarse woody debris) are not treated in this work and implementation of this assumption is discussed in section 3.
 Monthly meteorological data for the carbon cycle model were obtained from the United States National Centers for Environmental Prediction and the United States National Center for Atmospheric Research (NCEP/NCAR) [Kistler et al., 2001] for the period from 1948 to 2003. Since the meteorological data were available to 2003, we used the monthly averages of the output at each grid for 1998–2003 for the years of 2004 and 2005. Trenberth and Guillemot  evaluated the NCEP/NCAR reanalysis precipitation data set. They found a reasonable agreement in annual mean precipitation over southern Africa, but excessive precipitation, for example, in June 1995. In future work, correction using observation-based precipitation data may help to improve the carbon cycle model [Sheffield et al., 2006], which is currently postprocessed by incorporating the satellite data of tree cover [Hansen et al., 2003] and the Global Precipitation Climatology Project (GPCP) product [Adler et al., 2003], as described below.
 The herbaceous biomass density from the prognostic model generally showed good agreement during the wet season but overestimated during the dry season in southern Africa, mainly because the herbs continued to grow even during the dry season owing to the excessive precipitation from the reanalysis data. Thus we used the living biomass density at the end of growing season from the model for the fuel load of the herbs during the burning season.
 When the herbaceous biomass density from the carbon cycle model was calculated to be zero owing to the light limitation for the under-story vegetation but the MODIS herbaceous cover was plus, the biomass density based on the relationship between the annual rainfall (mm) and in situ biomass density (kg DM m−2) measurements at about 150 sites in West African savannas from Menaut et al.  (slope: 4.9 × 10−4, intercept: −0.058, correlation coefficient: 0.7) was used to fill the gap in each grid. Thus we applied this relationship to the annual rainfall data from 1998 to 2005 at a 2.5-degree resolution using monthly rainfall data from Adler et al.  [Ito and Penner, 2004].
 Fractional tree cover from the process-based model, which simulates potential vegetation, is generally higher than that derived from the satellite observation. Thus we scaled the biomass density for tree fine litter from the model by the ratio of tree cover between the model and the MODIS. We also scaled the biomass density for herbaceous vegetation from the model by the ratio of herbaceous cover between the model and the MODIS.
 The NDVI and cloud mask data set were derived from the sensor VEGETATION on board the SPOT satellite platforms by the Flemish Institute for Technological Research (VITO). The images had a spatial resolution of approximately 1 km2 after geometric, radiometric, and atmospheric corrections [Maisongrande et al., 2004]. Clouds obscure the land surface, reducing NDVI considerably. The cloud screening for VEGETATION products is based mainly on a threshold approach using the contrast between clouds and Earth surface in the blue band [Kempeneers et al., 2000]. We used the cloud mask data to identify cloud-free pixels. We also used evergreen forests and desert masks determined by NDVI to remove pixels where the NDVI-based method to assess the PGREEN was unreliable [Korontzi, 2005]. The cloud-free NDVI data in grasslands and woodlands are calculated from the 1998–2005 time series of NDVI and cloud mask data. Then the relative greenness indices, PGREEN, are calculated using the following equation [Kogan, 1997; Burgan et al., 1998],
where [NDVI]limin is the minimum of all NDVI values at the pixel over the preceding growing season, [NDVI]limax is the maximum of all NDVI values at the pixel over the preceding growing season. If fine-scale data are averaged over coarse-grid data from different land cover types, the results are often biased. Thus we calculated the averages from the cloud-free NDVI data over a 0.5-degree resolution in grasslands and woodlands, separately. Since the NDVI data were available from April 1998, we used the monthly averages of the PGREEN from 1999 to 2005 for the period from January to March in 1998.
 For the fine fuels in grasslands and woodlands, we adopted combustion factors that depend on the fuel moisture. Thus we used the relationship between the combustion factors and the PGREEN [Hoffa et al., 1999]. Ito and Penner [2005b] recalculated the relationships between CF and PGREEN in grasslands (slope: −2.0, intercept: 1.4, correlation coefficient: 0.7) and woodlands (slope: −2.1, intercept: 0.87, correlation coefficient: 0.6), based on the measurements at 8 sites presented by Shea et al. , Ward et al. , Hoffa et al. , and Korontzi et al. . The relationships include measurements obtained in both the early and late burning season.
 Fire extent and intensity in southern Africa are strongly influenced by biomass availability for fires, and hence rainfall [Lindesay et al., 1996; Swap et al., 2003]. Fires in woodlands tend to be more intense in areas of low tree cover and high mean annual rainfall, where grass production is high but where grass quality and therefore grazing pressure is low [Desanker et al., 1997]. The grasses dry faster and are consumed during the flaming combustion which increases the MCE, while the litter fuels dry slower and tend to involve more smoldering combustion, which can decrease the MCE [Ward et al., 1996; Bertschi et al., 2003]. The dependence of MCE on fuel types has been determined for southern Africa in the late dry season by Ward et al.  and in both the early and late dry season by Korontzi . A reanalysis of the measurements in woodlands presented by Shea et al. , Ward et al. , Hoffa et al. , and Korontzi et al.  provide the following relationship for MCE as a function of fuel types,
where P denotes the proportion of each fuel type of green grass (gg), dead grass (dg) and fine litter (fl) in the fuel mixture and the three proportions sum to 1. These measurements took place at 9 sites in Zambia from June to September. The fuel loads of green and dead grasses were calculated from the NDVI-based PGREEN values and the fuel-load modeled estimates of grass [Korontzi, 2005].
 We recalculated the regression equation between MCE and EF for BC (g-BC kg-DM−1) from 10 fires measured in southern Africa by Sinha et al. [2003, 2004].
 To quantify the relative importance of different aspects of our model in explaining the seasonal delay of the CO emissions peak relative to burned area, we set two different scenarios. For case 1, the monthly averaged fuel loads, combustion factors, and emission factors were used to estimate the monthly CO emissions. The annually averaged effective CO emissions per area burned were calculated in grasslands (18 g CO m−2) and woodlands (30 g CO m−2) in southern Africa from the calculations of case 1. For case 2, these averages for each land cover type are multiplied by the monthly burned areas to estimate the monthly CO emissions.
3. Results and Discussion
3.1. Comparison of Relative Greenness Indices, Fuel Consumption, and Emission Factors With Observations
 We compare our PGREEN, FC and EF model results with the field measurements as a function of time and land cover type. The comparisons of monthly variations of PGREEN in the area affected by open vegetation fires are shown in Figure 2. Monthly averaged observations were taken from Shea et al.  and Hoffa et al.  in Zambia. The median and range of observations are shown because of the limited number of the observations for different fires. Overall, our estimates of PGREEN were in good agreement with observations. The monthly averages of PGREEN in grasslands (Figure 2a) decreased from June to July. From July to September, the dead grasses in grasslands were already dry and the monthly averages of PGREEN were almost constant in time. The monthly averages of PGREEN in woodlands (Figure 2b) decreased from the early to late burning season.
 The comparisons of monthly variations of FC from open vegetation fires in southern Africa are shown in Figure 3. Monthly averaged observations were taken from Shea et al. , Hoffa et al. , and Hély et al. . The ecology of miombo ecosystems is closely related to that of the savanna-like cerrado in South America [Desanker et al., 1997]. Thus the two measurements of fuel consumption in the cerrado [Ward et al., 1992] were also used for the averaged FC in woodlands. Our fuel consumption in grasslands (Figure 3a) was generally higher than the measurement. These overestimates in grasslands do not cause a significant error in the total emissions of CO because woodland fires (80%) are a much larger source of CO than are grassland fires (20%) in southern Africa. Our fuel consumption in woodlands (Figure 3b) increased from 0.24 (kg DM m−2) in June to 0.68 in September, while the measurement increased from 0.01–0.25 to 0.45–0.84. The seasonal increase in FC is mainly caused by the variation in fuel moisture during the burning season [Hoffa et al., 1999]. At the beginning of the burning season, the fuel is wetter and burns less completely. As it dries out, it burns more completely.
 The comparisons of monthly variations of EF for CO from open vegetation fires in southern Africa are shown in Figure 4. Monthly averaged observations were taken from Ward et al. [1992, 1996], Korontzi et al. , and Yokelson et al. . Our emission factor in grasslands (Figure 4a) decreased from 74 (g-CO kg-DM−1) in June to 48 in July, while that of measurements decreased from 52–101 to 33–54. The seasonal decrease in EF for grasslands is associated with the burning of moister vegetation at the beginning of the dry season [Korontzi et al., 2003]. The burning of moist vegetation produces a higher fraction of the products of incomplete combustion in the early season, which increases the production of CO emissions. Our emission factor in woodlands (Figure 4b) did not significantly change with month, which was consistent with the measurements in southern Africa. Our averaged EF in woodlands (69) was within the range of the observations for grassland fires in June and significantly smaller than the average for tropical forest fires (104 ± 20) [Andreae and Merlet, 2001]. This indicates that the woodland fires involve smoldering smokes as much as in early grassland fires but more flaming than the forest fires.
 Our averaged BC emission factors in grasslands and woodlands were 0.31 and 0.45 (g-BC kg-DM−1) and in good agreement with those of measurements (0.24 and 0.47) in southern Africa [Sinha et al., 2003, 2004] and with the global average (0.48 ± 0.18) [Andreae and Merlet, 2001]. We note that the measurement data is not available for comparisons of monthly variations of EF for BC.
3.2. Seasonal Variability
 The monthly variations of burned areas and CO emissions are shown in Figure 5. The peaks in BA often occurred in early burning season of June and July, while those in CO emissions were found in late burning season of August and September.
 The monthly averages of burned areas and CO emissions in different scenarios are shown in Figure 6. Two different fire counts of (Figure 6a) VIRS and (Figure 6b) MODIS were used to estimate the burned areas and CO emissions. The monthly averaged burned area from VIRS fire counts peaked earlier than MODIS, as well as monthly mean CO emissions when the CO emission were calculated from annually averaged effective CO emissions per area burned (case 2). In contrast, the peaks in monthly averaged CO emissions were identical between two different fire counts of VIRS and MODIS when the CO emissions were derived from time-dependent CO emissions per area burned (case 1). Even though more fire affected areas were observed in June and July by VIRS, the effective CO emissions per burned area were substantially low in the early burning season when the fires might cease owing to the wet fuels. These results suggest that the VIRS is sensitive to subpixel small fires as well as large ones, but large-scale fires within one pixel in late burning season are not differentiated owing to the coarser VIRS spatial resolution (4.8 km2 before 7 August 2001 and 5.8 km2 after 24 August 2001) [Kummerow et al., 1998] than MODIS (1 km2).
 The fire season in southern Africa typically starts in May in the northwest, moves southeast, and ends in October near the east coast of South Africa [Cahoon et al., 1992]. The peaks in burned areas, combustion factors and CO emissions for 2000 are shown in Figure 7. The peaks in burned area occurred on the west side of the continent around between 5°S and 10°S in June and July, and on the east side in July–September. In contrast, the peaks in combustion factors and CO emissions were found on the west side of the continent in July–September, and the east side below 10°S in September. This characteristic delay between burned areas and CO emissions is mainly caused by significant changes in CF. In woodlands, the averaged combustion factor increased from 0.25 in June to 0.71 in September by a factor of 2.80, while the burned area (106 km2 yr−1) decreased from 0.36 to 0.24. The fuel loads and the emission factors did not change significantly. As a result, CO emission (Tg CO yr−1) increased from 4.9 to 11.9.
3.3. Interannual Variability
 The interannual variations of BA, CO and BC emissions are presented in Table 1. The annual estimates of BA from open vegetation burning in southern Africa ranged from 1.8 × 106 km2 for 2002 to 3.0 × 106 km2 for 1998, while those of CO and BC emissions ranged from 45 Tg CO and 0.26 Tg BC for 2002 to 75 Tg CO and 0.42 Tg BC for 1998. The annual burned area was generally coupled with the fire emissions. This is consistent with the result of van der Werf et al. , but our averaged BA from 1998 to 2004 (2.3 × 106 km2) was significantly larger than their estimate (0.81 × 106 km2) from Giglio et al. [2006a]. Since both of BA were based on the calibration methods, the major difference in the total amounts might be associated with the burned area data sets of Roy et al.  and Giglio et al. [2006a], both of which used the MODIS observations but different techniques. Appropriate validation is needed to assess the uncertainties associated with satellite-data-based products [Roy et al., 2005a].
Table 1. Annual Amounts of Burned Areas, CO, and BC Emissions From Open Vegetation Fires in Southern Africa
Barbosa et al.  showed significantly smaller biomass burned after an El Niño in 1986/1987. Their possible explanation was that the drought might have lead to a lower biomass production with a consequent reduction in the number of fires, burned areas and biomass burned. During the 1997/1998 episode, the typical drought that occurs during the ENSO warm events was not as severe and extensive in southern Africa [Anyamba et al., 2002]. Therefore the biomass production did not significantly change during the growing season for 1997/1998 in southern Africa. Our estimate of CO emission for 1998 was larger (24%) than the average, which was consistent with the result of van der Werf et al.  (25%).
3.4. Carbon Monoxide Emissions
 The annual amounts of total area burned, the effective CO emissions per area burned and CO emissions for southern Africa are compared among various modeling studies in Table 2. Our CO emission was in good agreement with those of van der Werf et al. , although our annually averaged burned area was significantly larger by a factor of 2.8. Our annually averaged CO emission per area burned was 27 (g CO m−2), which was in good agreement with the regional estimates of Sinha et al.  (34) and Ito and Penner [2005b] (35). That of van der Werf et al.  (104) was in better agreement with that of global emission model of Ito and Penner  (130). The main reason is that in the global models, woody vegetation contributes substantially to emissions in savanna grid cells, while the fine fuel is a dominant component in the regional models. Thus the average fuel loads in the global models are higher than the values in regional models, which are based on the regional measurements [Shea et al., 1996; Hoffa et al., 1999]. The fuel wood collection is accounted for the global models in the coarse grid cells, but this human practice must be treated as sub-grid-scale phenomena, as Shea et al.  and Hoffa et al.  reported that the majority of coarse woody debris (CWD) had been previously removed for use as fuels prior to the fires by local farmers. Satellite measurements of vegetation changes together with the burned areas may provide better space-and-time-resolved data in the future.
Table 2. Comparison of Annual Amounts of Areas Burned, CO Emissions per Area Burned and CO Emissions From Open Vegetation Fires in Southern Africa Among Various Modeling Studies
 In order to examine differences in bottom-up estimates of CO emissions from different data sets with those from top-down estimates, we compare the estimates of CO to the ranges deduced from the inverse modeling studies [Arellano et al., 2006] (also G. Pétron et al., Multiyear inversion of CO surface sources using the MOPITT satellite data, manuscript in preparation, 2007) (hereinafter referred to as Pétron et al., manuscript in preparation, 2007). In the Arellano et al.  and Pétron et al. (manuscript in preparation, 2007) studies, the monthly amounts of the prescribed CO emissions were optimized, so that the predicted CO mixing ratios from a chemical transport model fit the measured monthly average CO mixing ratios derived from the MOPITT instrument. Biofuel emissions were separated from the open vegetation fire category in these studies and thus were also not included here.
 The comparisons of monthly variations of CO emissions are shown in Figure 8. The seasonal variations of our CO emissions, which were derived from the burned areas using VIRS fire counts and time-dependent effective CO emissions per area burned, were better agreement with those from previous top-down estimates using the MOPITT data for 2000 and 2001. Although our annual emissions were smaller than the top-down estimates, the smaller peaks derived from the observations from January to March were likely influenced by biomass burning in the Northern Hemisphere Africa. Thus sum of monthly emissions from May to September was calculated for the comparison. Better agreement was obtained among the four estimates of Arellano et al. , Pétron et al. (manuscript in preparation, 2007), van der Werf et al. , and this work (74 ± 14 Tg CO for 2000). In later years, however, the agreement was worse (75 ± 38 Tg CO for 2003). Our calibration methods for the burned area estimates are based on the data sets solely for 2000. Thus the extrapolation to different years causes additional uncertainties in later years. van der Werf et al.  used the Giglio et al. [2006a] data set, which was extrapolated from the selected areas and months of the prototype MODIS burned areas between January 2001 and December 2004 to fourteen geographic regions during the period. Gloudemans et al.  compared the SCIAMACHY CO column with the model CO estimates using the van der Werf et al.  data set and found significant underestimates of the model results in southern Africa for 2003 and 2004. Accurate measurements of the burned areas are required to improve the emission estimates in southern Africa.
3.5. Black Carbon Emissions
 We compare our estimates of BC emissions from open biomass burning for 2000 to those deduced from the inverse modeling study [Penner et al., 2004] in regions of southern Africa defined by Zhang et al. . We note that the values shown in Figure 9a is smaller (about 10%) than those presented in section 3.1 (Table 1) because the land area in the domain defined by Zhang et al.  is slightly smaller than the area in the region defined in this paper as the Southern Hemisphere Africa. In the Penner et al.  study, the monthly amounts of the prescribed BC emissions from open vegetation fires were optimized using the method developed by Zhang et al. , based on the Total Ozone Mapping Spectrometer (TOMS) Aerosol Index (AI), a 3-D chemical transport model (IMPACT), and the Bayesian inversion approach [Tarantola and Valette, 1982a, 1982b]. The TOMS AI is a measure of the change of spectral contrast in the near ultraviolet due to the radiative transfer effects of aerosols in a Rayleigh scattering atmosphere [Herman et al., 1997]. The IMPACT model was developed at the Lawrence Livermore National Laboratories [Rotman et al., 2004] and extended to treat aerosols and detailed chemical reactions for a wider set of volatile organic compounds at the University of Michigan [Liu et al., 2005; Feng and Penner, 2007; Ito et al., 2007].
 Here we show the monthly averaged TOMS AI data in Figure 9b, corresponding to absorbing aerosols (0.2 < AI) [Torres et al., 2002]. For each month, we used the total number of days of available data to calculate the weighted monthly averages. We also show the monthly averaged MODIS Aerosol Optical Depth (AOD) data [Remer et al., 2005]. The peak in our BC emissions during open biomass burning season for 2000 was identical to those from previous top-down estimate and the TOMS AI data.
 Our estimated amount of peak emission of BC was in good agreement with the inverse modeling estimate, but we underestimated CO (about 30%). Although this difference is close to the uncertain range of inverse estimates [Arellano et al., 2006; Pétron et al., manuscript in preparation, 2007], this systematic bias may be associated with the assumption of no coarse fuels burning in this work. Generally, the coarse fuels are consumed during the smoldering process in savanna [Ward et al., 1996; Bertschi et al., 2003]. Because the emissions from the prolonged, smoldering fires associated with coarse woody debris are injected near the surface, the transport of these emissions from the surface to the altitude where the satellite instrument has enough sensitivity to measure the mixing ratios associated with these emissions is particularly important [Ito and Penner, 2005a; Arellano and Hess, 2006]. Especially, the TOMS AI may not detect smoke plumes from low-intensity fires [Duncan et al., 2003].
 We estimated the emissions of CO and BC from open vegetation fires in southern Africa over the 1998–2005 period, using satellite derived information and the biogeochemical model. Overall, our estimates of relative greenness indices, fuel consumption and emission factors from June to September were in good agreement with those of the measurements.
 This study implies that process-based carbon cycle models may reasonably retrieve the distributions of fuels in natural land ecosystems. However, these models were not sufficiently validated for the spatial and temporal variability in fuel load, which is actually affected by fire history and human activities. Incorporating the impacts of disturbance is a challenging issue for carbon cycle model study in conjunction with observational study; especially, wild fire regime is interactive with the carbon cycle through fuel accumulation and consumption.
 The monthly averaged burned area from VIRS fire counts peaked earlier in June and July, than those in CO emissions in August and September. This characteristic delay between burned areas and emissions was mainly explained by significant changes in combustion factors for woodlands. Even though more fire affected areas were observed in the early burning season by VIRS, the averaged combustion factor was substantially low owing to the wet fuels.
 This is the first bottom-up emission model which shows that the peaks in CO and BC emissions during open biomass burning season for 2000 were identical to those from previous top-down estimates using the MOPITT and the TOMS AI data [Penner et al., 2004; Arellano et al., 2006; Pétron et al., manuscript in preparation, 2007]. Sum of monthly emissions during burning season from May to September in 2000 was in good agreement between our bottom-up estimates and previous top-down estimates. In later years, however, the bottom-up estimates were lower than the top-down estimates. Clearly, further work is required to reduce the differences between these two approaches.
 The SPOT VEGETATION data sets were distributed by Spot Image and produced by VITO, Belgium. We wish to thank the Oak Ridge National Laboratory Distributed Active Archive Center at Goddard Space Flight Center, for producing the SAFARI 2000 data in its present format and distributing them.