4.1. Annual and Seasonal Fire Dynamics
The interannual dynamics of the SEVIRI-derived FRE estimates for the northern and southern African hemispheres (shown in
Figure 1a) match quite closely those of the MODIS burned area measurements (
Figure 1b) and GFED total fuel consumption estimates (
Figure 1c), though in general somewhat greater interannual variability is seen in the FRE data. The annual variation in SEVIRI FRE indicates that, in the northern hemisphere, fire activity peaked in 2007 and is decreasing by 11 Tj/year. A decreasing trend in also evident in the annual burned area (−2.75 Mha/year) although it is less pronounced than that of the FRE. Using the same burned area product although a different time-series (2001–2012), reference [
16] found a decreasing trend in burned area in the northern hemisphere which was attributed to the expansion of cropland into savanna reducing the occurrence of fire. In the southern hemisphere, trends in these parameters are similar and are rather weak but slightly negative over the thirteen years examined and with local peaks around 2010/11. Over a different time frame, reference [
16] found an increasing burned area trend in southern hemisphere Africa which was attributed to El Nino Southern Oscillation (ENSO) climate anomalies leading to increased fire activity.
Figure 1b clearly demonstrates a fall in burned area since 2012, however. As evident from
Figure 1a,b, annual dry matter (DM) fuel consumption estimates from GFED (version 4; not yet including the unvalidated small fire correction of reference [
58];
Figure 1c), follow a similar trend to burned area and FRE, but do exhibit some different dynamics.
Using the conversion factor (0.37 kg DM MJ
−1) described in reference [
20] and whose magnitude has recently been confirmed as valid for satellite data retrievals by reference [
28], SEVIRI-derived FRE retrievals deliver annual fuel consumption estimates ranging from 127–214 Tg and 208–299 Tg in northern and southern hemisphere Africa respectively, far lower than the GFED fuel consumption estimates (by 77 to 85%). Cloud obscuration, tree cover obscuration of surface fire FRP, and the inability of SEVIRI to detect fires burning below ≈40 MW very likely contribute to these low totals, but the relatively consistent ratio between the FRE-derived and GFED estimates (0.14–0.23 between 2004–2016) shown in
Figure 1a,c suggests a relationship between these two approaches which apply quite different methods to estimate landscape fire fuel consumption. Despite the findings of reference [
28], other relatively recent research suggests that the relationship between FRE and fuel consumption might not be the same for larger fires, which contain propagating fire fronts and greater fuel moisture variability [
35,
59].
The annual dynamics of monthly FRE-derived fuel consumption in northern and southern hemisphere Africa (
Figure 2a,b) indicates that the months in which fuel consumption is greatest are broadly consistent on an annual basis, whilst fuel consumption outside of the fire season is more variable albeit much lower in magnitude. June, July, August and September are the peak months of fuel consumption in the southern hemisphere although annual variations of 25–50% are evident in the latter two months. The peak months of fuel consumption (June–September) are characterised by very weak trends over time (both positive and negative). Fire activity during the biomass burning season, which in the northern hemisphere occurs between November–March and May–October in the southern hemisphere, drive the interannual trends observed in
Figure 1a. The large increases in southern hemisphere fuel consumption in August and September in 2008 and 2010 coincide with the extensive fire activity that occurred in Botswana in these two years, where 11 and 13 Mha burned, respectively [
60]. The peak fire season in the northern hemisphere extends between November and March which, as evident in the southern hemisphere, has large annual variations in monthly fuel consumption of 20–30% and typically display negative trends. Note that some SEVIRI data was missing in 2004 (January) and 2013 (March, April, May). A noticeable feature in
Figure 2b is that several months outside of the biomass burning season (e.g., November–March) have strong positive trends in fuel consumption albeit at low magnitudes, which may have implications for emissions mitigation schemes employing early season burning as fires at these times tend to be low intensity due to higher fuel moisture content and relative humidity [
61]. Korontzi et al., [
62] found the modified combustion efficiency (MCE) to vary seasonally over southern African grassland and woodland, which can have a significant impact on the chemical composition of the smoke emissions although similar seasonality was not found by reference [
63] in Australian savanna fires. In the northern hemisphere, a weak positive trend in fuel consumption in months outside of the fire season is evident although the large variability makes the interpretation of these dynamics more uncertain than those in the southern hemisphere.
4.2. MOD17 Productivity Assessment
Prior to their use in fuel load estimation, an assessment of the MODIS productivity estimates (which include both growth and maintenance respiration terms) was carried out using the Copernicus dry matter productivity (DMP) product of reference [
64] which is driven by Proba-V data and available from the Copernicus Global Land Service at 10-day temporal resolution and 1 km spatial resolution. Using three years of data (2013–2016), the productivity estimates were integrated over the MODIS burned areas (e.g., ‘fire clusters’) which burned in successive years, as described in
Section 3. The results shown in
Figure 3 indicate a stronger relationship between both datasets in southern hemisphere Africa (
r = 0.96) than in the northern hemisphere (
r = 0.8). A difference in magnitude is also evident between the MODIS and Copernicus products, with the former lower (albeit the latter does not account for below ground biomass or the costs associated with growth and maintenance respiration in woody tissue).
To assess the impact of these respiration terms, the average ratio between annual MODIS NPP and annually accumulated MODIS PSN was calculated and reveals that, over Africa, 88% of pixels have a NPP/PSN ratio between 0.7–0.8 and that 40% of the pixels have a value of 0.79. The annual NPP/PSN ratio is more variable in northern hemisphere Africa where lower ratios occur in the tropical forests of central and west Africa (0.68–0.75) and the parts of Sahel and CAR (<0.65). The basis for comparing FRE-derived fuel consumption estimates to MODIS PSN estimates, which includes maintenance respiration and growth respiration of woody material, rather than the Copernicus DMP estimates is that these data only cover 2013–present whilst a longer time-series is available from MODIS.
4.3. Comparison between FRE-Derived Fuel Consumption and MODIS Accumulated PSN and Copernicus Dry Matter Productivity (DMP)
Here we provide the first comprehensive assessment of the relationship between FRE-derived fuel consumption (Tg DM) and temporally integrated MODIS PSN productivity (Tg) and Copernicus DMP (Tg DM) estimates. MODIS PSN estimates (2004–2016) are used as a surrogate for NPP since the latter is only available on an annual basis, which is inconsistent with the seasonal dynamics of landscape fires. It is acknowledged that FRE-derived fuel consumption is not directly comparable to PSN, since the latter contains contributions from woody material maintenance, growth respiration and below ground biomass. Therefore, irrespective of variations in the combustion completeness, the temporally integrated PSN value should be greater than the FRE-derived fuel consumption estimates.
The focus was on areas whose burning encompassed twenty or more MCD64A1 MODIS 500 m burned area pixels, thus omitting the smaller fires which SEVIRI finds harder to identify [
41,
65]. The accumulated productivity estimates between the two years of burning were remapped into SEVIRI’s grid and the total available fuel per burned area ‘fire cluster’, or contiguous cluster of pixels, calculated using the burned area (m
2) measurements (assuming 47% fuel carbon content; reference [
66]). The matching FRE estimates were calculated via temporal integration of the SEVIRI FRP data between the first and last day the pixel was detected as being burned in the current year, with a 10-day buffer added to account for any offset between the burned area and active fire observations (e.g., due to cloud cover or differences in fire detectability using the burned area and active fire approaches; references [
26,
67,
68]). FRE-derived fuel consumption for each fire cluster was then estimated using the conversion coefficient of 0.37 kg MJ
−1 found by reference [
20], which provided good relations between fuel availability and FRE-derived fuel consumption metrics in reference [
39]. A timeseries of SEVIRI FRP retrievals and the mean MODIS PSN and GPP estimates from a single 57.7 km
2 fire which burned in an area of woody savanna are shown in
Figure 4 (a and b respectively). This fire affected an area of 18 discrete SEVIRI pixels over its lifetime (with a mean of 22 FRP observations per pixel), and the accumulated PSN and total fuel consumed over the 57.7 km
2 burned area were calculated as 96,950 and 13,700 tonnes respectively. The FRE-derived fuel consumption per unit area is 0.23 kg DM m
−2 which is somewhat lower than field measurements made in African savannas which ranges between 0.29–0.45 kg DM m
−2 [
36].
Active fire detection and FRP retrievals are influenced by a number of factors including sensor characteristics, obscuration due to cloud/smoke or an upper canopy, fire size, fuel moisture content and flame emissivity [
59,
69,
70,
71]. One approach to mitigate for certain of these issues is to focus analyses on “well-observed” active fire clusters. Using data between 2005 and 2016,
Figure 5 illustrates the relationship between FRE-derived fuel consumption (Tg DM) and accumulated MODIS PSN (Tg) for fire clusters in the northern (
Figure 5a) and southern (
Figure 5b) African hemispheres as a function of the mean number of per-pixel SEVIRI FRP retrievals in each cluster. Clusters with, on average, fewer than 10 FRP retrievals per-pixel exhibit greater variability whilst fires observed with an average of >10 observations display a much stronger relationship. This is clear in the northern hemisphere which displays a weaker relationship and greater variability than found in the southern hemisphere. For the remainder of the analyses presented, only fire clusters where pixels are observed on average ≥10 times are retained, omitting 46% and 48% of the fire clusters in the northern and southern hemispheres respectively.
Figure 6 shows scatterplots between FRE-derived fuel consumption and integrated PSN in the northern and southern African hemispheres for the remaining fire clusters for each year (2005–2015). In the northern and southern hemispheres, the average annual number of fire clusters per year is 2040 and 2588 respectively. Overall, there is little interannual variation in terms of changes in the correlation and slope of the relation between FRE-derived fuel consumption and accumulated PSN. The relationship in the northern hemisphere is weak (mean
r = 0.76) and the slope indicates that the FRE-derived fuel consumption estimates are between 11 and 18% of the accumulated PSN. Of 24,249 fire clusters that cover the full time-series, 99% of the clusters have a lower FRE-derived fuel consumption estimates, with an average relative difference of −73% (mode of 92%) and the average scatter (0.13 Tg), bias (−0.03 Tg) and RMSD (0.004 Tg) are also high. In contrast, a strong relationship is evident in the southern hemisphere (mean
r = 0.96) with the fuel consumption being 7–8% of the accumulated PSN. Of 30,948 fire clusters, 99% have a lower fuel consumption estimate, with an average percentage difference of −88% (mode of −96%) and the average annual scatter (0.57 Tg), bias (−0.12 Tg) and RMSD (0.07 Tg) are also relatively high.
The FRE-derived fuel consumption and Copernicus DMP in northern and southern hemisphere Africa (
Figure 7) also display a strong relationship although the fuel consumption is 37% (northern hemisphere) and 31% (southern hemisphere) of the accumulated DMP. Accounting for the difference in the ratio between MODIS NPP and MODIS PSN discussed in
Section 4.2, which indicates PSN estimates were ≈21% greater than NPP over southern hemisphere Africa, suggests FRE-derived fuel consumption is between 29–31% of the primary productivity estimate. Roberts et al., [
39] compared SEVIRI fuel consumption estimates with integrated SPOT VGT NPP estimates over 18 fires and found closer agreement (i.e., higher combustion completeness) than found here, albeit across a far smaller number of samples (and using NPP data derived from SPOT VGT using a process similar to that used for the Copernicus DMP product). This difference in primary production data may help to explain the difference in slope between the analysis conducted here and that of reference [
39].
Unlike the comparison between accumulated MODIS PSN and SEVIRI FRE-derived fuel consumption (
Figure 6a,b), where the correlation between the datasets differed in northern and southern hemisphere Africa, the relationship shown in
Figure 7 is similar in both hemispheres. The comparison between the MODIS PSN and Copernicus DMP (
Figure 3), also shows good agreement in the southern hemisphere but a weaker relationship in the northern hemisphere. Studies have found MOD17 productivity estimates to be underestimated in the Sahel region (e.g., references [
52,
72]) which may explain the closer agreement in magnitude between FRE-derived fuel consumption and the accumulated MODIS productivity in the northern hemisphere. Fensolt et al., [
53] suggests the underestimation of MODIS productivity estimates may result from uncertainty in the parameterisation of the Biome specific lookup-table (BLUT). Sjostrom
et al., [
52] highlights the importance of meteorological input data (e.g., vapour pressure deficit and photosynthetic active radiation) on GPP estimation and found that replacing the National Center for Environmental Prediction Department of Energy (NCEP-DOE) reanalysis data with tower measured meteorological data improved the correlation between eddy covariance and MODIS GPP at sites in northern and southern hemisphere Africa. Northern hemisphere Africa has been subject to increased land cover change since 2001 [
16] which may influence MODIS productivity estimates if this change is not adequately captured in the MOD17 algorithm. The similar ratio between FRE-derived and GFED v4 fuel consumption estimates in both hemispheres suggests there is little bias in the SEVIRI FRE data. Further assessment of MODIS productivity dataset over Africa is needed to elucidate the apparent differences observed between the northern and southern hemispheres.
Fuel Consumption Per Unit Area (m−2)
Despite the consistency in the relationship between the FRE-derived fuel consumption and accumulated PSN, it is clear that SEVIRI FRE-derived fuel consumption is underestimated. Here we assess the agreement between fuel consumption and productivity as a function of land cover type and fuel consumption per unit area (kg DM m−2). The land cover type for a given fire cluster is assigned on the basis of the dominant land cover within the fire cluster.
Figure 8 illustrates the relationship between FRE-derived fuel consumption and accumulated PSN for fire clusters in northern and southern hemisphere Africa for land cover types with >50 fire clusters. Similar to that shown in
Figure 6, the correlation between fuel consumption and integrated PSN is strong in the southern hemisphere (the northern hemisphere delivers a weaker relationship), although there is also limited variation in the proportion of fuel consumed as a function of land cover type. Savanna, cropland and grassland are the dominant cover types in both hemispheres, and demonstrate similar slopes between FRE-derived fuel consumption (Tg DM) and MODIS PSN (Tg). This may result from grass and litter comprising the largest fraction of combusted fuel in African savanna fires, with relatively little of the woody material being consumed by comparison [
73,
74]. The fuel moisture content of these fine fuels is also typically low during the peak fire season, which leads to high (a mean of ≈0.93) combustion completeness [
36,
73,
74,
75].
The fire clusters in this analysis were selected on the basis that they burned in consecutive years, and this frequent fire occurrence may help explain the similarity in the relationship between fuel consumption and integrated PSN between land cover types. Savanna fuel loads, principally grasses, tend to be greater in areas where fires occur on a biennial or longer timescale due to fuel build up (e.g., in moribund and unpalatable grass), which is influenced by the fire return interval, precipitation patterns, productivity and grazing pressure [
61]. Large areas of Africa burn frequently (
Figure 9) and the majority (81%) of the pixels in the fire clusters used in this analysis burned every three years or fewer between 2005 and 2016. Differences in fire occurrence may lead to variations in fuel build up within each fire cluster, since the accumulated PSN only accounts for the contribution of productive vegetation through fPAR measurements. Variations in the build-up of senescent fuel are not represented, and could increase the variability between fuel consumption and accumulated PSN measures.
The similarity in the proportion of fuel consumed relative to the accumulated productivity as a function of land cover type is reflected in
Figure 10, which presents the FRE-derived fuel consumption per unit area (kg DM m
-2) and the integrated PSN (kg m
-2) for northern (
Figure 10a,b) and southern (
Figure 10c,d) hemisphere Africa. The upper and lower bounds of the bars depict the 25th and 75th percentiles and the black line is the 50th percentile. As with many natural phenomena, the fuel consumption estimates are best described by a power law where the 25th percentile is close to the highest frequency.
The average fuel consumption estimates are approximately 0.1 and 0.2 kg DM m
−2 at 25th and 50th percentiles respectively. Savanna, grassland and croplands show similar fuel consumption estimates, which lends support to the notion that herbaceous fuel is the primary material burned. In the southern hemisphere, the MODIS PSN estimates for forest, savanna and croplands are higher than the fuel consumption estimates at all percentile ranges, whilst shrublands and grasslands are lower in most cases. The FRE-derived fuel consumption estimates should be lower than the accumulated PSN since combustion completeness is rarely 100% and the PSN estimates contain contributions from woody respiration and maintenance and below ground biomass. The average accumulated PSN estimates are 1.32 and 1.65 (kg m
−2) at 25th and 50th percentiles respectively and are greater than those of the northern hemisphere. The FRE-derived fuel consumption estimates for savanna, which is the dominant land cover type, is 89% and 70% lower at the 25th and 75th percentiles, respectively, which is close to the slope shown in
Figure 8.
Fuel consumption per unit area estimates for dambo grassland (0.22–0.29 kg DM m
−2), savanna (0.35 kg DM m
−2) and miombo woodland (0.42–0.45 kg DM m
−2) contained in references [
27,
74,
76,
77] are typically around twice those of the FRE-derived estimates. Despite constraining the fire clusters to those which are best observed (i.e., ≥10 observations on average per-pixel), the FRE-derived fuel consumption estimates appear underestimated. Roberts
et al., [
26] compared MODIS and SEVIRI FRE-derived fuel consumption estimates (g DM m
−2) and found those from SEVIRI were around three times lower than those from spatially coincident ground measurements. Applying this adjustment to fire cluster fuel consumption estimates for savanna, grassland and cropland in southern hemisphere Africa increases it to 0.33 (kg DM m
−2) at the 25th percentile. This is closer to the range of fuel consumption estimates found in African grassland savanna (0.21–0.65 kg DM m
−2) and woody savanna (0.29–0.73 kg DM m
−2; reference [
36]), though the slope characterising the relationship between FRE-derived fuel consumption and accumulated PSN for fire clusters in southern hemisphere Africa in savanna, grassland and cropland remains far below unity (at 0.23, 0.23 and 0.15 respectively).
A number of factors contribute to the underestimation of FRE-derived fuel consumption including FRP retrieval, fire detection bias and fuel characteristics. SEVIRI FRP has been compared to spatially and temporally coincident MODIS FRP for fire clusters >50 MW where over half (57%) were within 20% of the corresponding MODIS FRP [
42]. However, whilst good agreement is found on sufficiently intense fires, SEVIRI underestimates fire activity over a wider area due to the omission of small and/or low intensity fire which fall below its detection limit [
41,
42], or are omitted due to cloud mask over efficiency for example [
69]. Sensor imaging characteristics can also impact the retrieved FRP. For example, the point spread function (PSF) distributes fire emitted radiance into neighbouring pixels, and may result in these not being detected or falling outside of the MODIS burned area [
78,
79]. Vegetation structure can also reduce the retrieved FRP through preventing the detection of actively burning pixels [
65,
69] and intercepting the fire emitted radiance from fires beneath an upper canopy [
34]. Fuel moisture content also has an impact since energy is lost to vapourisation which reduces the potential for combustion and, should combustion occur, the retrieved FRP [
59,
80]. In this study, the majority of the fires were detected in the dry season, and account for 91% and 88% of all fire clusters in the northern and southern hemisphere respectively. The fuel moisture content of herbaceous fuels is typically low in the dry season but does vary over time and with the proportion of green grass. These factors are typically at a minimum at the peak of the dry season [
73,
81]. Recent studies have also found different FRE to fuel consumption coefficient values to that found by reference [
20] which could be caused by differences in sensor characteristics, fuel moisture, fuel type and fire size [
35,
79,
82,
83,
84].