A Global Bottom‐Up Approach to Estimate Fuel Consumed by Fires Using Above Ground Biomass Observations

Real‐time estimates of the fuel consumed during a fire (dry‐matter) relies on indirect estimate from remotely sensed released energy combined with biome dependent conversion coefficients. The uncertainties in the conversions lead to the use of inflation factors to avoid large underestimations in the prediction of aerosol load during fires. In this study, adopting two different fire inventories for burned areas, we apply above ground biomass (AGB) observations [from Soil Moisture and Ocean Salinity (SMOS) L‐band vegetation optical depth] as proxy for fuel load in deriving estimates of dry‐matter. These new estimates are then converted into biomass burning aerosols and validated against independent aerosol optical depth observations from the AERONET in situ global network. Results showed that use of AGB as a proxy observation of fuel load improves fire emission estimates and substituting the need for indirect dry‐matter estimates from remotely sensed fire activity or for the use of inflation factors.


Introduction
Despite fuel availability playing an active role in fire occurrences, real-time global monitoring of landscape changes are neither employed in fire early warning systems for fire prevention and suppression (Bedia et al., 2018;Di Giuseppe et al., 2020;Mölders, 2010;Roads et al., 2005) nor in systems aiming at quantifying biomass burning emissions and their impact on air quality (e.g., Flemming et al., 2009).
Indeed, one of the most used models for fire forecasting worldwide, the fire weather index (FWI) (de Groot et al., 2007;Di Giuseppe et al., 2016;Noble et al., 1980), is specifically calibrated to describe the fire behavior in a Jack pine stand typical of the Canadian forests and only relies on weather parameters. The FWI is however applied globally (de Groot et al., 2007;Di Giuseppe et al., 2020;San-Miguel-Ayanz et al., 2002) on the assumption of a spatially and temporally constant fuel load, thus disregarding information on the available biomass. One of the consequences of this assumption is desert areas are classified at extremely high fire danger all year around despite no fuel to burn and the possibility of a fire occurrence is unfeasible.
Estimation of fuel load is the step introducing the greatest uncertainty in fire emissions inventories (Ellicott et al., 2009;Soja et al., 2004;Van der Werf et al., 2003). Most fire emission models adopt a 'bottom-up' approach, in which estimates of burned biomass are generated from remote observation of burned area (BA), active fire (AF) counts and/or fire radiative power (FRP). These burned biomass estimates are multiplied by biome-specific emission factors to convert each kilogram of burned dry matter to the amount of a trace gas or aerosol released into the atmosphere. Both the Global Fire Assimilation System (GFAS) DI GIUSEPPE ET AL. 10.1029/2021GL095452 2 of 11 (Kaiser et al., 2012;Di Giuseppe et al., 2017 and the Global Fire Emissions Database (GFED) (Field et al., 1998;Giglio et al., 2013;van der Werf et al., 2006) use bottom-up approaches, but with some important differences. GFED relies on post-event BAs detection and a fuel load estimated through a vegetation growth model, and is not available in real time. GFAS uses real-time observations of FRP converted into combusted matter through static coefficients, and is used in real time to provide the boundary conditions of fire emissions to the vastly used atmospheric composition configuration of the Integrated Forecast System (IFS-COMPO) (Flemming et al., 2009;Rémy et al., 2019). The two systems are not independent from each other with GFAS conversion coefficients calibrated so total annual emissions match the output of GFED, which is considered the benchmark. Kaiser et al. (2012) found that despite the ad-hoc calibration performed, the global budget of fuel burnt (often referred to as dry-matter, DM) estimated from the GFAS system was substantially divergent to what is independently observed from MODIS (Moderate Resolution Imaging Spectroradiometer) when converted into aerosol optical depth (AOD). As AOD measurements are operationally assimilated by IFS-COMPO, some of GFAS fire emission components such as black carbon or organic matters are inflated with factors that vary from 6.12 to 3.4 to avoid rejection of valid observations in the assimilation cycle (Benedetti et al., 2009;Rémy et al., 2019). The large errors introduced by the inaccuracies in the dry-matter estimation are known and the use of inflation factors is broadly accepted in other operational systems (Kolusu et al., 2015). It is also recognized that in regions of small undetected fires, the required multiplicative factors could be much larger (Petrenko et al., 2017;Ramo et al., 2021).
Recently, to overcome these limitations, top-down methods have been proposed to eliminate the need for the explicit knowledge of fuel loads to derive fire emissions (Ichoku & Ellison, 2014;Mota & Wooster, 2018). Mota and Wooster (2018) showed that satellite-derived fire radiative energy (FRE) can be used to directly quantify AOD, therefore short-cutting the need for dry-matter estimations. The benefit of the top-down approach is that it provides a framework to use FRE observations directly into assimilation systems of composition models. However, top-down methods hinder the potential benefit of an explicit representation of vegetation processes in Earth system models. Information about the fuel available for burning is crucial in fire forecasting and prevention systems (Di Giuseppe et al., 2016). Indeed fuel, weather and local topography, are the only controlling factors of fire intensity and spread (Williams et al., 2003). Thus, methods to properly estimate fuel load and the amount of dry-matter released during burning are not only important for fire emissions (Benedetti et al., 2019) but also in fire behavior models (Burgan, 1984), fire management practices (Finney, 2001) and early warning systems (Di Giuseppe et al., 2020).
One of the main problems is that daily fuel monitoring at a spatial scale useful for operational forecasts (order of 10-50 km) is still a challenge globally. While harvesting techniques provide the most accurate method for fuel estimation, it is not always practical and certainly not available at global level (Kumar et al., 2015;Kumar & Mutanga, 2017). Global applications must rely on remote sensing observations either by image classification and photo interpretation (Foody et al., 2003;Rahman et al., 2005) or by indirect mapping which derives fuel information from vegetation optical signals from active or passive microwave measurements and lidar sensors (Gleason & Im, 2012;Zolkos et al., 2013). In the last 10 years, two L-band microwave sensors have been performing systematic observations: the Soil Moisture and Ocean Salinity (SMOS) satellite (Kerr et al., 2010), launched by ESA in November 2009, and the Soil Moisture Active Passive (SMAP) satellite (Entekhabi et al., 2010), launched by NASA in January 2015. In particular, the full-polarization and multi-angular capabilities of SMOS allow the simultaneous retrieval of the soil moisture content and L-vegetation optical depth (L-VOD). Recently Rodríguez-Fernández et al. (2018) found a quasi-linear relationship between L-VOD and benchmark estimations of above ground biomass (AGB) derived through complex data fusion of in-situ inventory plots, lidar observations and optical and microwave imagery (Avitabile et al., 2016;Baccini et al., 2012;Saatchi et al., 2011). The logistic relationship found opens a way to obtain real-time AGB estimations from L-VOD with future missions with L-band capabilities (e.g., L-ROSE, SMAR) also providing potential avenues for real-time estimations.
Here we explore the benefits of directly employing AGB estimations as a proxy for fuel load by deriving estimates of dry-matter released into the atmosphere and comparing with those derived through FRP conversion. Fuel load is transformed into combusted dry-matter by employing two different datasets of BAs and an assumption on combustion completeness. To validate the quality of the dry-matter estimates, these are then 3 of 11 converted into biomass burning aerosols and compared with independent AOD measurements. The direct use of AGB, while maintaining a bottom-up approach, could substitute the modelling method used for vegetation growth proposed by GFED and could provide an estimation of dry matter which is independent of the fire emitted energy as in the GFAS system. By exploiting independent observations to derive dry-matter estimations this will limit the need for the application of inflation factors to the final AOD estimations.

Using Above Ground Biomass
The schematic in Figure 1a explains how DM estimation is calculated from AGB, providing a global database on a 25 km grid for the years 2010-2017. Following Seiler and Crutzen (1980) the portion of vegetation that is consumed during biomass burning is expressed as: where BA is the burned area; FL is the fuel load, the amount of biomass or organic matter an ecosystem contains per unit area; is the combustion completeness or burning efficiency, which is the fraction of fuel actually consumed during the fire. Each static map provide three L-VOD to AGB conversion curves fitting the 5th, 50th and 95th percentiles of the data using a logistic regression (see Figure S4 in Supporting Information S1) (Fernandez-Moran et al., 2017;Rodríguez-Fernández et al., 2018). In the following, the three different databases are named AGB-BA (Baccini), AGB-SA (Saatchi) and AGB-AV (Avitabile) and are based on the median AGB curve. Monthly total of BAs are available from several sources. Humber et al. (2019) performed an extensive inter-comparison of four global BA products and demonstrated no consensus on burn locations or timing, with commission and omission errors being found depending on geographic location, timing of the burning, season, and total amount of burning. Given these inaccuracies and necessity to estimate the impact of a choice of BA database, we select two BA databases that were found to perform quite differently (Humber et al., 2019). The first one, called hereinafter BA-GFED, is available through GEFD4.1s (van der Werf et al., 2017) and is based on MODIS MCD64A1 product. It also employs MODIS AF to add small fires that were undetected by MCD64A1 (Randerson et al., 2012). It has a 500 m pixel resolution and is also available on a regular grid of 0.25 deg. The second BA product, named hereinafter BA-CCI, is a multi-sensors product provided by the European Space Agency Climate Change Initiative (ESA-CCI) (Padilla et al., 2015). We use version FireCCI5.1 which is calculated using a two-phase algorithm, where MODIS AF locations are used to identify seed pixels corresponding to high confidence BAs. These areas are then grown using Medium Resolution Imaging Spectrometer (MERIS) vegetation input data, which are distributed in 10deg by 10deg tiles (Alonso-Canas & Chuvieco, 2015;Chuvieco et al., 2018).
Combustion completeness, , is assumed constant inside a biome ( Figure S1 in Supporting Information S1) and taken from the averaged values used in GFED4 [ Table 4 in van der Werf et al. (2006)]. has been measured for various biomes and fuel types (van Leeuwen et al., 2014), and varies over the course of the fire season with more complete combustion at the end when fuels is more cured (Hoffa et al., 1999). In general, fine and dry fuels burn more completely than coarse and wet fuels and can range from 98% for standing dry grass to less than 10% for dead logs (Liousse et al., 2004). By using a time invariant and grossly neglecting the sub-pixel variability of the combustion efficiency we are likely to introduce large inaccuracies in the final estimations of the fuel available for burning. However, here the focus is on the impact of using observed fuel-load from AGB in isolation, maintaining the other assumptions as close as possible to the GFED and GFAS approaches for comparisons purposes.
By selectively combining any AGB estimations with the two available BA datasets, a total of 6 products are generated for a test year, 2016. For convenience the data-sets are numbered 1 to 6

Using Fire Radiative Power
For comparison a dry-matter reference database from FRP observations is created following the conversion method proposed by (Kaiser et al., 2012) and implemented in GFASv1.2. Weighted mean FRP values from MODIS MOD/MYD14 products  are used to derive daily FRE estimates. Only two platforms, TERRA and ACQUA acquires MODIS FRP data so a limited number of daily over-passing is available. The FRP to FRE conversion fails to capture a full diurnal cycle of a fire (Freeborn et al., 2009) and a flat FRP emissions profile for the fire diurnal cycle is assumed. Daily FRE are then multiplied by biome-specific conversion factors (in kg MJ − 1 ) relating the FRE to total dry-matter . The conversion factors defined in Kaiser et al. (2012) are used as per the operational GFASv1.2. As these were calibrated over the totals of GFEDv3.1 (van der Werf et al., 2006;van der Werf et al., 2010) their accuracy depends on the assumption of combustion completeness and the limitation for BA detection in GFED.

Estimation of Aerosol Optical Depth From Dry-Matter
The conversion of DM into particulate optical depth is an important step as it allows to asses the quality of the bottom-up approach proposed using an independent validation. It follows the procedure explained in Kaiser et al. (2012) and implemented in GFASv1.2. Forty four constituents, , are diagnosed from to-tal DM using the formulation = ⋅ , where are the 44 conversion coefficients from Heil et al. (2010). Each DM data set produces a different set which are subsequently used to initialize one month of daily simulations for September 2016. The underlying aerosol model is the IFS-COMPO which includes several aerosol species (organic matter, black carbon, sulfates, nitrates, ammonium, desert dust and sea salt). In addition to fire emissions which are specified in the above methods, emissions for anthropogenic aerosol species are taken from established inventories (Lamarque et al., 2010) while other natural aerosol emissions are modeled on surface winds and other model parameters. More information on the aerosol model can be found in Rémy et al. (2019) and in the Supporting Information S1.
In total eight sets of AOD simulations were performed; one for each of the new dry-matter datasets (1-6) and two control runs using FRP prediction of dry-matter with and without inflation factors. While AOD from biomass-burning is available as model outputs, validation in isolation is difficult. Ground truth measurements for fire emissions are only available from controlled ignition experiments aiming at studying the combustion process in detail [e.g., the TROFEE campaign described in (Karl et al., 2007;Yokelson et al., 2008)]. Total AOD is instead provided by sensors on board of polar orbiting/geostationary satellites and in-situ stations from global networks. In this case, not only biomass combustion but all sources of emission (e.g., anthropocentric fuel burning, chemical reactions and sea salt and dust) contribute to these measurements. Still, signals in the total AOD should be expected, especially in regions where the contribution from biomass burning is important (e.g., African savanna, boreal forests). To validate our simulations we use the AOD observations from the in-situ AERONET network (Holben et al., 1998). Model outputs are bi-linearly interpolated to the AERONET site locations and averaged over 24 hr periods (from T + 3 to T + 24). AERONET data are similarly averaged, with each data value receiving a weight proportional to the time difference between the data values before and after it, up to a maximum of 3 hr.

Dry-Matter Comparison
A comparison of the newly created databases with the equivalent estimations from GFASv1.2 and GFEDv4.1s for the whole of 2016 is provided in Figure 1b with detailed analysis as a function of the biomes in Figure 1c. GFAS dry-matter estimations were calibrated to match GFEDv3.1 global monthly means and explains the similarities between these two datasets globally. Dry-matter estimations derived by AGB observations are however much larger than both GFAS and GFED with an increasing factor ranging globally between 2.7 and 6.1 depending on the AGB retrieval algorithm and the BA database adopted (see Table 1). Interestingly, most of the variability in the multiplicative factor can be ascribed to the choice of the AGB retrieval algorithm, the BA playing a minor role on the global average. It is important to stress that this variability is driven by the L-VOD to AGB conversions and the benchmark databases chosen. As shown in Figure S1 in Supporting Information S1 a measurement of L-VOD = 1 would produce estimates of AGBs equal to 270Mgh −1 , 284Mgh −1 and 353Mgh −1 depending on the benchmark chosen which correspond to a 30% spread around a mean value. Figure S4 in Supporting Information S1 also highlights that L-band microwaves are not sensitive to all vegetative structures and forests with L-VOD>1 saturate in the L-band at roughly 300-400 Mgh −1 . In addition accounting for the sub-grid scale variability of the combustion completeness is likely to increase even further this variability.
A more detailed depiction can be drawn from the analysis of the differences at biome level (Figure 1c). A substantial increase in the estimated DM is observed in Savanna with the AGB derived DM factors stretching to 10 compared to GFASv1.2 and GFEDv4.1s in agreement with other datasets (Nguyen & Wooster, 2021). Savanna is both the largest biome covering greater than 50% of all land and also experiences the largest fire  Total 2.7-5.8 2.9-6.1 2.6-5.5 2.8-5.8 Note. The variation is expressed as multiplicative factor and the interval represents the range min-max between datasets 1 to 3 and 4 to 6 as defined above.  Figure S2 in Supporting Information S1) making this underestimation very relevant for the total biomass burning aerosol budget. Interestingly, AGB-BA datasets project the largest variations in DM estimates in agriculture areas (AG + AGOS) possibly due to the interference in the L-VOD signal induced by irrigation practices where the vegetation moisture content largely contributes to the L-VOD retrieval (Rodríguez-Fernández et al., 2018). This is a relevant aspect as agricultural lands are important contributors to total fire activity with 22% of all fires occurring here, a biome that only covers 7% of the global land (see Figure S2 in Supporting Information S1). Fire occurrence only partially explains the contribution of a biome to the global annual fire emissions with fuel load having a substantial impact in the final budget. Still the coherent emerging picture is that in the more fire prone biomes, employing direct observations of AGB implies a substantial increase in dry-matter consumed during burning.

Verification With AOD
The conversion of monthly DM estimations based on BAs into daily fire emissions needed to initialize an air quality model is not straightforward, especially if we try to replicate the operation of a real-time system which relies on FRP for fire detection. Although BA and AF based emission products have been used widely in global and regional scientific studies, there are substantial differences between the two products (Roy et al., 2008) and much remains unknown about the exact relationship between them. Humber et al. (2019) found many AFs are observed for pixels without BA observation (cell B in Figure 1a) and vice versa, area averaged FRP can be zero in pixels that are classified as burned (cell A in 1a). On the one hand, this is the result of the sensitivity of the products; AF products ability to identify small fires while BA product lacks a visible scar from small fires. On the other hand, FRP products might miss detection due to poor satellite sampling, cloud masking and high vegetation interference (Freeborn et al., 2009;Kaufman et al., 2003). Figure 2a estimates the frequency of these cases. Data are grouped into two cases on the x-axis: given only the pixels where BA is above zero, the proportion between matching and non-matching FRP records is 61% (true, yellow) and 39% (false, purple) respectively.

Figure 2. (a)
The left column shows the percentage of occurrences of detected active fires from GFASv1.2 where burned area observations from BA-CCI dataset is above zero. Conversely, the right column shows the percentage of occurrence of burned areas where fire radiative power is recorded above zero. The matching records are shown in yellow, while the non-matching ones are in purple. (b) 2D density plot (percentage of occurrence, as a factor of the fire duration on the x-axis and the latitude on the y-axis). The first 10 outliers in terms of fire length with the lowest percentage of occurrence are marked with black circles.
• | 0 : given only the pixels where FRP is above zero, the proportion between matching and non-matching FRP records is 91% (true, yellow) and 9% (false, purple) respectively.
The most relevant finding is that there are almost 40% of cases in which a recorded FRP is not matched by a BA observations. Spurious signals in the FRP database can contribute to part of this mismatch but it also confirms the inability of MCD64A1 datasets to map small fires (Roteta et al., 2019;Tsela et al., 2014) even with the boosting introduced in recent versions. These discrepancies would pose great problems to the use of dry-matters derived from BAs in real-time applications that rely on FRP detection to initiate a burning process. On one hand, BAs not matching any FRP will give rise to an unaccounted DM budget. On the other hand there will be undefined DM values where a FRP is observed but no BAs recorded. Here, any unmatched FRP/BA sample is discarded. In the future these inconsistencies could be removed through the continuous assimilation of an available real-time burned scar (e.g., Nolde et al., 2020). From the prospective of the air quality prediction, blacklisting copious FRP observations has a detrimental affect on the quality of the AOD prediction. However, the aim is to assess the potential for an AGB-BA based dry-matter derivation, experiments therefore will be benchmarked against control runs performed with the same subset of FRP observations, which are degraded in comparison to an operational GFAS run.
Another challenge is the conversion of the monthly AGB-BA derived DM datasets into daily estimates as fire length is of only few days in most part of the planet. Thus, the available DM is consumed in an interval shorter than a month. DM estimations from BAs provide a bulk evaluation and might be more accurate then FRP estimations, in terms of average quantities. However, they are not able to capture the emissions' daily variability required to initialize operational composition forecasts. In this diagnostic study, to solve this problem, an average fire duration is estimated for every pixel using observed time series of FRP from GFASv1.2 for a given month. Figure 2b shows the percentage of fire occurrence in September 2016, as a factor of the fire duration on the x-axis and the latitude on the y-axis. Most of the burning occurs on the sub-equatorial bell between 0 and -30 degree-north due to the intensive bush cleaning practices in the African savanna (Trollope, 1984) and the agricultural fires in the Amazon forest. A second smaller peak is visible at around +40°-north associated to boreal forest fires. Outliers can reach 20 days of continuous burning but the average fire duration globally is 1.96 days.
Once monthly DM from datasets 1-6 are reduced to daily estimates on coinciding FRP detections they are converted into fire emissions and then in total aerosol optical depth. The global verification against the 300 AERONET stations (Figure 3) highlights extremely encouraging results. The performances are summarized on a Taylor diagram (Taylor, 2001). For each model, three statistics are plotted: the Pearson correlation coefficient, the normalized standard deviation ratio of the simulated versus observed values and the mean bias. It is evident from this summary diagram that all six data-sets on average outperform the FRP approach in terms of standard deviation and bias even when employing the inflation factors to boost emissions to match observed AOD adopted by GFASv1.2. In agreement with the findings highlighted in Figures 1b and 1c the choice of BA database does not project the same impact in the final results as the choice of the benchmark employed in the AGB conversions. Importantly not all databases perform homogeneously over space. Thus, while AGB retrieved from Baccini et al. (2012) provides the best overall results, AGB from Avitabile et al. (2016) allows for a zero bias and perfect RMS amplitude in Africa (Figure 3 top right) while producing standard deviation amplitude that is clearly much larger than that of the observed field in South America (Figure 3 bottom right) and smaller in North America (not shown).
A spatial overview of the impact of the new observations on the total AOD using several metrics (the modified normalized mean bias (MNMB) is provided as an example in Figure S3 in Supporting Information S1) highlights that using FRP conversions without inflation factors performs poorly in all the areas in which biomass burning is relevant, i.e., in the Amazon and central and South Africa in September 2016. While an improvement in skill is obtained by the introduction of the ad hoc inflation factor, greater improvements can be obtained by employing observed dry-matter. Geographical performances are different with AGB-BA providing the best performance in South America and AGB-AV the best in savanna Africa.

Conclusions
This paper shows the potential for AGB estimations to be directly employed to quantify fuel burned during a fire and estimate the particulate load released into the atmosphere. Results demonstrate that despite the uncertainties still present in AGB databases and in burnt scar detection, a bottom-up approach directly employing observations can improve fire emissions estimations.
The immediate consequence is that direct observations of biomass with the use of real time BAs could substitute indirect estimates of fuel load based on energy release. Interestingly, even if AFs have been employed to boost BAs detection of small fires, FRP is still a very different product, in which almost half of the cases is not encompassed in a detected burned scar. While spurious signals can play a role in the number of FRP detected, AF detection is very sensitive to small fires covering as little as 10-14 of the pixel area . On the contrary, 20% of the pixel area must be fire affected for a burn scar to be reliably detectable . This could mean a reduction in the reported benefit of the use of AGB in combination with BAs when compared to the absolute skill achievable by an FRP based method, especially if multiple sensors are employed to improve the FRP-FRE conversion.
Yet, as the documented improvements in AOD prediction highlights, there is a real opportunity arising from real-time L-VOD observations to characterize the biomass available in earth system models. Moreover these results suggest that fuel load could be initialized through L-VOD observations; the first step toward the inclusions of fire processes into land surface scheme of real time weather forecasting systems. Given the impact that biomass burning aerosols have on the energy budget of the planet (?, ?) this could in turn lead to improved weather forecasts and consequently better skill in downstream applications that rely on it. The six experiments are represented by different symbols on the diagram and numbered 1-6. The two controls are named FRP and FRP + fact. The size of the symbols represents the statistical weight. The point represented with the asterisk is the observation. For each experiment, three statistics are plotted: the Pearson correlation coefficient (blue lines); the RMS error (green contours centered around the asterisk); and the standard deviation ratio (red contours). The shade of the symbol represents the bias, according to the color-bar on the right.

Data Availability Statement
Dry matter databases derived by crossing burned areas databases with AGB estimations are available through Zenodo at https://doi.org/10.5281/zenodo.5582576. The FRP database used in this study and the corresponding dry matter estimated from FRP are available as experiments in the ECMWF MARS archive with codes defined in Table S1 in Supporting Information S1. Data from GFASv1.2 used for Figure 2 are available through the CAMS website at https://apps.ecmwf.int/datasets/data/cams-gfas/. GFEDv4.1s dry matter data used to create Figure 2 are available through GFED website at https://globalfiredata.org/pages/ data/.