gmd 2019 116

Transcript

1 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © frican biomass burning emissions and the effect of Modelling A spatial resolution 1 1 Dave van Wees Guido R. van der Werf , 1 , , the Netherlands 1081 HV Amsterdam Department of Earth Sciences , , Vrije Universiteit d.van.wees @ vu.nl ) 5 Correspondence to : Dave van Wees ( - Abstract. Large scale fire emission estimates may be influenced by the spatial resolution of the model and input datasets used and cover, a coarse model resolution might lead to substantial errors n areas with relatively heterogeneous l Especially i . In this paper, w satellite e developed a model using in estimates. MODIS ) Moderate Resolution Imaging Spectroradiometer ( observations of burned area and vegetation characteristics to stud y the impact of spatial resolution on model l ed fire emission - estimates. We estimated fire emissions - meter spatial resolution (native MODIS burned area) 10 for sub Saharan Africa at 500 model ing framework , 2017 period, using a simplified version of the Global Fire Em - issions Database (GFED) for the 2002 l ° ). We estimated fire emissions of ° , 0.125 ° and compared this to model runs at a range of coarser resolutions (0.050 , 0.250 1 1 - - ° and 0.82 PgC yr meter resolution - at 0.25 resolution; a difference of 24%. At 0.25 ° resolution, our at 500 0.68 PgC yr relatively similar to GFED4 , which also runs at 0.25 ° resolution , whereas our 500 - meter estimates were model results were resolutions are 15 finer substantially lower. We found that lower emissions at mainly the result of reduced representation errors , led estimates of fuel load and consumption when comparing model to field measurements as part of the model calibration. - 1 difference in and lead to an at coarse resolution simulation Additional errors stem from the model additional 0.02 PgC yr - estimates . These errors exist due to the aggregation of quantitative and qualitative model input data; the average or aggregated values are propagated in the coarse resolution simulation and affect the mod majority - el parameterization and the 20 difference the - meter and s We identified at least three error mechanisms responsible for final result. in estimates between 500 resolution simulations , besides those stemming from representation errors in the calibration pr ocess , namely : 1. biome 0.25 ° errors due to the averaging of input data and the associated 2. misclassification leading to errors in parameterization, temporal mechanism related to the aggregation of burned area in particular . Even though reduction in variability, a and 3. each other and only modestly affect estimates at a continental scale, they lead to neutralized these mechanisms largely affect large - scale estimates differently for other 25 substantial error at regional scales with deviations up to a factor 4 , and may These findings could prove valuable in improving coarse resolution models and suggest the need for increased continents. spatial resolution in global fire emission models. 1

2 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © Introduction 1 se of trace gases and aerosols into the atmosphere (Andreae and Fires exert a key influence on the global climate by the relea Merlet, 2001; Ciais et al., 2013; Ward et al., 2012) . Furthermore, fires partly shape, and in the long - term sometimes ffecting the storage capacity of carbon . About 7 0% (Rabin et al., 2017) determine, the vegetation state of landscapes, thus a roughly half of the global fire carbon emissions of global burned area occurs in Africa (van 5 (Giglio et al., 2018) , leading to an der Werf et al., (Archibald et al., 2009; v The majority of fires in Africa occur in the savannas . der Werf et al., 2010) 2017) , an ecosystem that is dependent on fires and where trees have evolved to tolerate fire (Beerling and Osborne, 2006) . and agricultural demographic changes African savannas are currently undergoing major sh ifts in fire activity due to (Andela and van der Werf, 2014) fire occurrence decrease in . expansion, leading to a 10 the eighties (Seiler and Crutzen, 1980) . Early estimates were Efforts to estimate global fire emissions have been made since - specific parameterizations of fire return times and biomass c based on biome onsumption rates, extrapolated using vegetation maps. More recently, satellite products have become an important tool for improved estimates of fire emissions, mapping based approaches to estimate - fire events globally and giving insight in fire impacts and dynamics. Two main satellite emissions exist, based either on observed burned area in combination with a biogeochemical or fuel load model, or based on 15 fire radiative power (FRP), which is directly related to fire emissions after integration over time to obtain fire radiative al., 2012; Wooster, 2002) . Burned area is determined after a fire has occurred, signified by a change energy (FRE) (Kaiser et associated with the burn scar in surface reflectance (Giglio et al., 2018) , whereas FRP is based on the fire size and intensity, determined by detection of the thermal hot spot during a satellite overpass. 20 In fire emission models, aboveground biomass and resulting fuel load are key variables for estimating emissions. Biogeochemical models dynamically simulate biomass buildup and degradation, and come with different levels of process (Hély et al., 2003, 2007; Hoelzemann et al., 2004; Schultz et al., 2008; van der Werf et al., 2017) In regional complexity . models, parameterizations derived from field data can be used to accurately represent local relations between e.g. precipitation and , and resulting fuel load can be 25 plant productivity , and between soil moisture and combustion completeness Hély et al., 2007; Korontzi et al., 2004; Russell - Smith et al., 2009) . Some of (Alleaume et al., 2005; calibrated at local scale (Ito and Penner, 2004) . However, in global - scale models, simple these models are based on pre determined fuel load maps parameterizations are often inaccurate due to the large variety in e.g. vegetation dynamics and fire characteristics across l., 2014; Rogers et al., 2015) , these models often depend heavily on continents and biomes (Lehmann et a . As a result 30 - derived climate and weather data, and land and vegetation characteristics. However, global satellite data on fire - satellite specific processes is scarce (Pettinari and Chuvieco, 2016) . Therefore, field measurements are crucial in constraining modelled fuel load and consumption (Hély et al., 2003; van Leeuwen et al., 2014) . Modelled fuel load can be combined with timate based burned area maps to es - combustion completeness factors to estimate fuel consumption, and then with satellite 2

3 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © mission factors emissions into emissions of trace used to convert dry matter Finally, e dry matter emissions. are or carbon key inputs for atmospheric and Earth system models which are , gases and aerosols (Akagi et al., 2011; Meyer et al., 2012; Wooster et al., 2011; . Yokelson et al., 2013) is limited by the spatial resolution of the satellite detector, as burned patches smaller than the 5 The detection of burn scars satellite footprint are often not detected. When these relatively small fires are active during the satellite overpass, the t hermal anomaly and its FRP may be detectable. Recent burned area products combine both of these detection methods to . In a first study complement burned area based on burn scar detection with relatively small fires from active fire detection looking into this on a global scale, 2012) ( Randerson et al. found an increase in global burned area of approximately 35% - induced (prescribed, agricultural, 10 due to the addition of small fire burned area. These small fires are often hu man deforestation) and mainly occur in croplands, woody savannas and tropical forests. Consequently, by the inclusion of these s Database (GFED) increased small fires, global fire emission estimates based on burned area from the Global Fire Emission - - 1 1 in GFED4s (“s” for small fires) on average over 1997 2016 (van de r Werf et al., in GFED4 to 2.2 Pg C yr from 1.5 PgC yr - 1 - 1 - emissions increased from 0.8 Pg C yr in GFED4s. in GFED4 to 1.1 Pg C yr , Saharan Africa alone 2017) . For sub - 15 gs), Besides the error in burned area due to limitations of the satellite detector and undetected small fires (amongst other thin the accuracy of fire emission estimates may also be affected by the coarse spatial resolution of most fire emission models. Emission models based on burned area, such as GFED4, often perform at a spatial resolution significantly coarser than the data used to calculate emissions, especially input nativ e resolution of the burned area dataset. This is necessary because much coarser than satellite data . Because of this and the necessary tradeoff between model 20 meteorological data, is usually aggregated to coarser resolution prior to the model spatially complexity and computational resources, the burned area data is used in GFED4). However, there might be large heterogeneity of fu els and simulation (e.g. 0.25 spatial resolution ° . Whether aggregation (Alleaume et al., 2005; Hély et al., 2003) combustion characteristics within aggregated burned area , , - leads to significant errors in large and the associated loss in heterogeneity scale averaged model estimates such as GFED is aggregation for the accuracy of modelled fire 25 not known. Therefore, it is necessary to understand the implications of spatial emissions. resolution could lead to biases in the results coarse of remote sensing Previous studies have examined how relatively spatial ( 1998) analyzed biases in 30 m Landsat TM burned area fo r Central Africa after spatial studies . For example, Eva & Lambin ( studied the burned area classification error in 30 García Lázaro et al. aggregation to a resolution of 1 km. Similarly, 2013) ), as compared to the 30m Landsat Iberia for several satellite products that span a range of resolutions (250m, 1100m, 0.05 ° ( 2005) for Africa and Miettinen & Liew ( 2009) for product. Comparable studies were done at conti nental scale by Silva et al. udies found that at coarser resolution, small and fragmented burned area mentioned st previously Southeast Asia. All of the the finest resolution available data , whereas large tends to be underestimated and spatial homogeneity fires compared to 3

4 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © . with a tendency to overestimate 2009) specifically studied the impact of spatial ( Nelson et al. leads to better estimates ) ( based forest classification (forest or non - based aggregation of an inventory - aggregation, by comparing majority and average - - forest). For majority d area based aggregation, they reached conclusions analogous to the previously mentioned burne namely that at coarser resolution the forest proportion is underestimated for sparsely forested area, whereas it is studies, 5 overestimated for heavily forested area. For average - based aggregation however, the mean forest proportion remained pixels. Furthermore, image variability decreased for aggregate , as binary area is averaged to fractional area in the constant coarser resolutions, because the average - aggregated pixel values converge towards the mean value of the entire image (Bian, . 1997) resolution input datasets to coarse r resolution are propagated in the models 10 Errors introduced by spatially aggregating fine used in a nonlinear model, an additional error driven by these datasets (Crosetto et al., 2001) . When aggregated datasets are (Jensen, 1906) . In general, for arises due to the nonlinear propagation of averaged values, known as Jensen’s Inequality every nonlinear function there exists an inequality between taking the average of the function result afterwards, versus averaging the function input variables beforehand. We could, for example, consider a fire emission model as a single 15 nonli near function. When running this model at aggregated resolution, an inequality (i.e. error) exists compared to the native resolution model. The magnitude of the inequality is dependent on the variance of, and covariance between, the input variables, and th e amount of local curvature (second derivative) of the function, which is a measure of its non - linearity ., 1983; Duursma and (Cale et al (Denny, 2017) . Jensen’s inequality is mostly discussed in literature in relation to ecology and Robinson, 2003; Pierce and Running, 1995; Ruel and Ayres, 1999) , but also in relation to biology (Denny, 2017) 20 , in the context of spatial, temporal and class averaging (e.g. plant functional types, (Heuvelink and Pebesma, 1999) geology r fire emission estimates is not known. The resulting error in emission PFTs). However, the implications of this inequality fo (Randerson et al., 2012; estimates could be of particular importance, since fire processes are generally highly heterogeneous . Roy and Landmann, 2005) 25 fire In this context, the aim of this study is to better understand the impact of spatial resolution on the resulting biomass and from aforementioned modelling at aggregated resolutions result in significant errors emission estimates. Whether the s error scale averaged fire emission estimates such as GFED, and fire adapted Dynamic Global Vegetation Models in - large FireMIP) (Rabin et al., 2017; van der Werf et al., 2017) has until now not been investigated To , e.g. those used in DGVMs ( . meter spatial resolution , to produce a first emission this end , we developed a fire emission model capable of running at 500 - this resolution at - Saharan Africa. We then compared these emission estimates to three ad ditional 30 estimate for sub , 0.05 ° ), in order to study the impact of ° ° simulations using the same model for a range of aggregated resolutions (0.25 , 0.125 ork spatial resolution on model results. Besides a comparison of large - scale emission estimates, a substantial part of our w scale biases due to aggregation, and identifying the underlying error mechanisms. As part of this - was to understand local analysis, we also considered the role of modelled biomass, a key precursor for resulting emissions. Finally, we compare d our 4

5 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ° m (van der Werf et al., 2017) , and tr ied - 500 eter and 0.25 resolution model results to the emission estimates from GFED4(s) modelling at aggregated resolutions in respect to changes due to to e the changes in emission estimates due to contextualiz model validation improvements and the incorporation of small fires. The insights gained in this study could possibly form an resolution, and in implementing step forward satellite in the direction of global fire emission modelling at native important counter - measures for reducing errors when modelling at aggregated resolutions. 5 2. Methods 2017 period with a monthly time step. - - Africa for the 2002 We developed a model to estimate fire emissions for sub Saharan l We start with describing the model, which was derived from the GFED mode ling framework and adapted to run at a range input datasets (2.2). We then describe the model of spatial resolutions (2.1). This is followed by a description of the various optimization using satellite - based reference data and field measurements of fuel load (FL) and fuel consumption (FC) (2.3). 10 Finally, we describe the simulations performed (2.4) and the methods used to compar e different model resolutions (2.5). .1 Model description 2 Approach model used. GFED is rooted in the Carnegie - Ames - Stanford - simplified For this study a version of the GFED was cycle using satellite data to constrain (CASA) biosphere model, which was developed to simulate the terrestrial carbon V an der Werf et al. carbon uptake and other fluxes 2003) extended this model to 15 (Field et a l., 1995; Potter et al., 1993) . ( he (sub)tropics . Over time, further for t include fire processes, and provided spatially resolved estimates of fire emissions modifications were made to GFED, including improved burned area identification (Giglio et al., 2006, 2013) and distinction (van der Werf et al., 2006, 2010) . The most recent version, between different sources of fire emissions on a global scale (Randerson et es that remain undetected by most burned area algorithms GFED4s, also aims to account for relatively small fir 20 al., 2012; van . These small fires add about 15% burned area in our study area in Africa. Recent der Werf et al., 2017) . For this study we have simplified research suggests this increase in burned area may be conservative (Roteta et al., 2019) resolution of - the meter resolution on continental scale; as compared to the 0.25 ° GFED model so it can be run at 500 . Only the main GFED functionality relevant for aboveground dynamics in biomass, litter GFED4s running on a global scale . M maintained belowground dynamics, and fire emissions wa s ore refined mechanisms represented in GFED, such as Furthermore, no specific deforestation mechanisms re 25 re not implemented. we herbivory, grazing and fuelwood collection we not only ma d e req uired computational resources manageable, but also ma d e it easier to modelled. These simplifications s our key objective wa , which disentangle mechanisms that cause differences between the model runs at different resolutions . i based structure wherein ) ctivity ( s partitioned over various biomass pools, - The model has a pool Net Primary Produ NPP turnover and fire processes. Aboveground biomass (AGB) and belowground biomass 30 that are affected by losses due to above and below t he ground , and the total aboveground re considered as the live part of the total available carbon a (BGB) 5

6 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © live and dead carbon is referred to as aboveground biomass and litter (AGBL), all expressed in mass of carbon per unit area 2 - photosynthetically active ). (g C m wa NPP s calculated as the product of incoming solar radiation (SSR), the fraction of ): ! - specific light - radiation (fPAR) and a biome use efficiency (LUE; "#$%& ) = //0 ( * , , ) ∙ 2(30 ( * , , ) ∙ ! , '(( ( (1) * , , "#$%& is the time in months. NPP wa s distributed over tree and non - tree vegetation 5 is the grid location coordinate and where * , classes by multiplication with fractions of tree and non - tree vegetation cover, and further distributed in equal parts over the we third of tree allocated re represented as leaf, stem and root pools, all receiving one - corresponding biomass pools. Trees - tree allocated NPP. In this s represented as grass and root pools, both receiving half of non tree vegetation - NPP. Non wa simplified categorization other non - tree vegetation types, such as shrubs, are part of the g rass pool. For trees the root pool 10 s subdivided into separate fine and coarse root pools, with 20% of the stem NPP allocated to the coarse roots, whereas for wa of fine roots. We used biome - specific LUE values based on those reported by non ed - tree vegetation all root biomass consist orted for the savanna biome, we used the open shrubland value of 0.208 gC/MJ for was not rep Field et al. ( 1995) . Since LUE open savannas, and an empirically determined value of 0.280 gC/MJ for woody savannas (see also Table 1). The LUE value for woody savannas was chosen to be in - between values re ported for forest and grassland biomes. 15 When the model reaches its equilibrium state after the spin up phase, the carbon input from NPP is balanced by the carbon r rates and fire processes, specific turnove - output via fires and respiration because of decomposition. Depending on pool biomass decays into three litter pools fine litter, coarse woody debris (cwd), and soil organic matter. The pool - specific : specific values in turnover rates, loosely based on those used in GFED4 (van der Werf et al., 2017) , were optimized to biome - s a series of model validation steps (see 20 section 2.3, ‘Model optimization’). The vegetation exposed to fire i either combusted biomass and litter and emitted directly, killed and converted to litter, or unaffected by the fire. The amount of as carbon h pixel: s calculated by multiplication of the available flammable wa exposed to fire and the burned fraction for eac carbon ∑ ( ( ) ( ( ) ) ) ( ) A D E 6 * * [email protected] 3789 ∙ 44 , , ! 4 * , , , , = (2) ∙ 83 ∙ * , , < ∙ 2 , I :$$;H # BC :$$; =>&& :$$; mortality scalar < is a fire - induced tree , 44 where is the amount of carbon combusted and released to the atmosphere, 4 =>&& is the burned area, i.e. fraction of pixel burned , 2 fraction of carbon in fuel is the is the c , for ombustion completeness, 83 25 I The part of fire - exposed carbon that is combusted wa s determined by pool - 50%. which we use d specific combustion re scaled linearly between a predefined minimum and maximum value dependent on an we completeness values that empirically defined soil moisture scalar. This scalar wa s defined as: ) M , L ( JK R S . T N . PQ ( ) , , ! * VW, ℎ 0 . 1 < ! (3) < 1 . 0 , = BC BC S . U wa where 30 is the volumetric soil water content in units of volume fraction. The scalar /< s obtained by first standardizing the a range between 0 and 1 , and then dividing by 0.6 and capping at 1 to remo ve anomalously high values related /< values to 6

7 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © re capped to not be lower than 0.1, simulating a minimum soil moisture level Additionally, the scalar values we to wetlands. below which moisture dependent processes are not further affected. Dry conditions result in CC va lues closer to the whether trees exposed to maximum, and vice versa for wet conditions. A mortality scalar for woody vegetation simulate d . This scalar s expressed as , wa or left as litter t al., 2003) (van der Werf e fire are killed, and consequently directly combusted - induced mortality in open 5 the squared fraction of tree cover to total vegetation, to resemble the range from low fire dense tropical forests where trees are not adapted. When a high mortality in landscapes (where trees are adapted to fire) to unburned aboveground and belowground biomass is transferred to the litter pools. More specifically, tree is killed, all of leaves and grass become fine litter, dead stems are added to the cwd pool, and dead roots are added to the soil pool. 10 i wa s based on s litter decomposition of The dependent on temperature and moisture conditions. The rate of decomposition wa ) pool s defined as: - specific turnover rates, and scaled by an abiotic scalar. The abiotic scalar ( ! \ ] ] JK ^ = VW, ℎ 0 . 1 < ! (4) < 1 . 0 ! , \ \ . _ S is the temperature scalar: ! where ` ^ c PN dN a ! VW, ℎ ! > 1 . 0 = 1 . 0 , = (5) ` bS ` is the temperature in ° C, and a where is the temperature coefficient, for which we used a value of 1.5, and a capped 15 @ bS 2013) . A a ° value of 1.5 implies a 50% increase for every ( 10 ° C C, similar to van der Werf et al. maximum of 1.0 at 30 bS s standardized to a range from 0 to 1, and capped at a rise in temperature. Just like the moisture scalar, the abiotic scalar wa carbon minimum of 0.1. Part of the turnover - exposed is respired directly, based on a respiration fraction of 0.5. The h the cwd (only originating from trees), fine litter and soil pools, and finally consecutively throug remaining part degrade s 20 enters the slow decomposition stage. Every degradation step is again subject to direct respiration. The belowground organic matter algorithm wa s simplified compared to GFED, b ecause the belowground dynamics are not relevant for fire dynamics in our study area; fires do generally not occur in wetlands and peatlands in Africa. The LUE values and turnover rates used 1 . This table also gives the average effective Table for the biomass and litter pools for each biome are summarized in the abiotic scalar. turnover rates for the litter and cwd pools, after appl ication of Input datasets .2 2 25 Moderate resolution Imaging Spectroradiometer) Collection 6 satellite observation products with a MODIS ( The model use d - spatial resolution as input where available, and previous MODIS collections or coarser non - MODIS datasets - 500 meter we re based on 0.25 ° resolut ion ERA - Interim reanalysis data otherwise (see Table 2 ). The meteorological input parameters (Dee et al., 2011) from the European Centre for Medium Range Weather Forecasts (ECMWF). The datasets used cover the time period from 2002 to 30 2017, unless noted otherwise. The MODIS MCD15A2H product of fraction photosynthetically wa s used in combination with reanalysis SS R ( Dee et al., 2011) to calculate active radiation (fPAR; Myneni et al., 2015) wa s based on the MODIS MOD44B tree vegetation classes - NPP (see Eq. 1). The distribution of biomass over tree and non 7

8 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © product for the fractions of tree cover (FTC) and non ) Dimiceli et al., 2015 vegetation continuous fields (VCF; - tree (Giglio et al., s based on burned area (BA) from the MODIS MCD64A1 dataset wa vegetation cover (NTV). Fire extent . The decomposition of litter - wa s based on temperature and soil moisture scalars derived from ERA Interim reanalysis 2018) water layer 1) data. The classification of biomes meter temperature) and soil moisture (volumetric soil - air temperature (2 5 wa s based on the MODIS MCD12Q1 land cover type product, collection 5.1 (Friedl et al., 2010) . The last available year of L C over T ype 2 classification scheme data and for this product r subsequent years. For this study, the , 2013, was also used fo produced by the University of Maryland (UMD) was used. MODIS input datasets were spatially aggregated to 0.050 ° , 0.125 ° , and 0.250 The resolution using 500 - meter resolution ° - 10 average - based aggregation. As an exception, the qualitative land cover type data was aggregated using majority based resolution ° land cover class to the aggregate grid cell. The 0.25 aggregation, by assigning the most frequently occurring - ° neighbour meter resolution by nearest - reanalysis data was resampled to 500 interpolation, i.e. by using the reanalysis 0.25 resolution were averaged to grid cell value nearest to each MODIS pixel. All MODIS data with sub - monthly temporal month as weights. he number of days in monthly resolution, using t the 15 2.3 Model optimization We tuned our model to match satellite - (Avitabile et al., 2016) and field based data on aboveground woody biomass (AGBw) ., 2014) . Since NPP i s the driver for biomass growth, we first me asurements of fuel load and consumption (van Leeuwen et al d level NPP corresponde GFED4 (van der Werf et al. , 2017) . Then, the AGBw was optimized to agree - ensured that biome to ( 2016) ates per biome. As a first order Avitabile et al. based gridded estimates by - with observation , by tuning the turnover r 20 we tuned AGBw with the stem turnover rate, approximation since the stems of trees hold at least 95% of the total AGB for 2 - m - tree) biomass is typically below 250 gC and therefore within all forest biomes (Poorter et al., 2012) . Herbaceous (i.e. non . After optimization of the stem biomass, the turnover rates of 2016) ( Avitabile et al. the uncertainty range of the dataset by leaf biomass ratios (i.e. root the leaf, grass and root pools were adjusted to attain root shoot ratios) in line with the - stem - - Poorter et al. ( 2012) . The previously described subdivision of root biomass into separate biome - specific ratios as reported by shoot - 25 se and fine root pools was used to improve root coar ratios. The chosen turnover rates also influence the amount of litter produced. Even though the amount of tree biomass is not always relevant for fires, since most African fires are ground fires and def orestation mechanisms are not specifically part of the model, it does determine the amount of cwd and part of the fine litter produced. d values by van Leeuwen 30 In the final validation step, model l ed FL and FC were compared to the compilation of field measure . For the African continent the database contain 16 measurement records that reported FL and FC, of which 9 ed 2014) ( et al. are grouped into different fuel classes (e.g. grass, leaves, litter, cwd). Additional field studies compiled by Scholes et al. on FL in African savannas were included in the comparison, giving 73 measurements on total FL. For all field ( 2011) 8

9 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ter, cwd, and occasionally leaves. As a consequence, the total FL estimates records used, measured FL consists of grass, lit a variable number of model pools, dependent on what fuel classes involve reported in our comparison to the measurements d not include the stem fuel class (and thus is not the ion of FL d Our definit were measured in the corresponding field study. oes dominated was - same as AGBL) as this class not reported in any field record used, which all consider ground fires in grass biomes where trees are generally not affected. The field measurements were collocated with model results based on the field 5 led plot coordinates. The model FL was optimized for each biome separately by tuning turnover rates of the grass and litter root pools to match measured average FL and spread in measurements. The - shoot ratios were not significantly affected by this parameter tuning. Due to the lack of sufficient field data on FC (16 records, at only 6 unique locations), we validated FC using the average of all records over all biomes. only 10 ents in the database were taken in savanna type biomes, and all except one (in Burkina Faso) were taken All field measurem - ° , resulting in the sample set being less representative of other biomes ° in Southeast Africa (south of 12 S and east of 23 E) , the majority of records d not report separate measurements of specific fuel classes, and thermore Fur and regions in Africa. id As a result , the model validation wa s restricted to provided thus only an overall fuel load value for the combined fuel classes. s also complicated by the large spatial 15 wa nd FC. The validation of individual fuel classes a comparison of total FL a variability in biomass allocation to fuel classes for field plots with similar properties, and because field conditions that nown for most records (such as last fire occurrence, slash - and - burn or not, early o r determine the allocation ratios we re unk late season fires, etc.). we re given with a precision of two decimal degrees. This yields an 20 For field records, the field plot coordinates most which is larger than the model pixel size of 500 meter. Therefore, more accurate coordinates with uncertainty of about 1 km, picked based on the field site descriptions using Google Earth. Where possible, - four decimal degrees precision were hand homogeneously vegetated areas were p 500 meter scale. For - icked to remove the influence of other land cover types at sub some field records the coordinate of a settlement or city nearby the field plot was reported instead of the actual plot, in w hich . 25 case again a neighboring pixel was chosen, or the actual field plot was retraced in the vicinity of the reported coordinates our study Many of the reported measurements were conducted before period, in which case the first model year, 2002, was was known, the month in the middle of the regional fire season of the used. For studies where only the year of measurement indicative of influence of f there were recent burned area and related drops in biomass in a pixel Finally, i pixel was used. , recent fires, a neighboring unburned pixel was chosen. 30 2.4 Simulations 2017 period, monthly temporal resolution. We ran our model at 500 with a - meter native MCD64A1 resolution for the 2002 - up was done based on the 2002 - 2006 climatology, in order to stabilize the model pools and match total - year spin A 200 - and outflow. Additional simulations were performed for the three aggregated resolutions (0.050 ° , 0.125 ° , 0.250 ° ) carbon in - 9

10 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ed biomass, litter and emissions. We restrict l to study the effect of spatial resolution on model our analysis to the African ed n particular to the Northern Hemisphere Africa (NHAF) and Southern Hemisphere Africa (SHAF) regions as continent, i These two regions contain the African continent south of 23 (van der Werf et al., 2006) defined in GFED . ° N latitude and Saharan Africa. - will be referred to as sub 2.5 Resolution comparison 5 We calculated the differences that occur in modelled AGBL and fire emissions due to running the model at different spatial of differences : those that occur as part of the model calibration and those that ed two categories resolutions. We consider ndent on resolution because it includes parameter tuning occur as part of the model simulation. The model calibration is depe error exists due to the scale This to match model pixels with field measurements, which is subject to a representation error. mismatch in comparing field measurements to model grid cell averages 10 (Janji ć et al., 2018) . At coarser resolution, the error is larger and as a consequence the model calibration is more biased. This leads to different model results at different and 500 - meter resolutions, due to resolution dependent model settings. We will refer to the differences between aggregated - differences. Besides calibration related differences, resolution model results due to different model calibration as calib ration - coarse differences in the model simulation result from the spatial aggregation of input datasets and the subsequent aggregated - 15 computation of the model algorithm. We will refer to the related differences in and 500 model results between differences. We define simulation difference as the difference that occurs when running meter resolution as simulation identical models with the exact same calibration, but at different spatial resolutions. Calibration differences 2.5.1 - meter resolution with an additional calibration We studied differences by comparing the model calibrated at 500 calibration The 0.25 resolution calibrated model was also compared direc tly to GFED4 as 20 resolution (Table 1, in parentheses). at 0.25 ° ° justified. This also it is based on a similar coarse resolution calibration , to determine whether our model simplifications were re . allowed GFED4 to serve as an indirect reference to validate modelled emissions where FC field measurements we lacking For the comparison to GFED4, the database without small fires was used (GFED4 instead of GFED4s), in order to compare the models using the same amount of burned area. The discrepancy between the burned area from GFED4 (without small 25 ed on MCD64A1 Collection 5.1) and MCD64A1 Collection 6 was accounted for by raising GFED4 emissions fires but bas additional according to the fraction of burned area in MCD64A1 Collection 6 compared to Collection 5.1. . 2 Simulation differences 2 . 5 as a result of running the model at quantified simulation differences Besides studying calibr ation differences, we additionally spatial resolutions . For this analysis we used the parameters based on the calibration at 500 - meter resolution , to differen t comparison . Using the same calibration, absolute and relative differences in simulation were 30 data - have the best model 10

11 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © meter meter native resolution results. Beforehand, the 500 - calculated as the coarse resolution results minus those of the 500 - based aggregation. A positive simulation results were aggregated to the coarse - resolution to be compared with, using average difference indicates higher estimates at coarser resolution, and a negative difference indicates lower estimates at coarser 2017 annual average spatial fields, to focus on spatial resolution effects. we re limited to the - resolution. The analyses 2002 5 In order to understand the error mechanisms that lead to simulation differences and to quantify their contributions, we performed additional simulations with altered model algorithms and compared these to the base model simulation. For ould example, the contrib ution of fire in the overall simulation difference, and its contribution to error at coarser resolution, c be quantified by comparing an altered simulation without fire processes (i.e. BA = 0) to the base simulation with fires. This lculated as the relative simulation difference with fires minus the difference without fires (altered minus contribution wa s ca 10 induced tree mortality, to study the contribution of that - base). Similarly, we compared simulations with and without fire process in the overall sim ulation difference. Notably, this method c ould only be used to quantify relative differences and not absolute differences, because the altered ed simulation in different model results, making absolute differences incomparable. In order to enable un biased 15 s result the log relative difference as: subtraction of relative differences, we calculate d difference = ln ( s / s ) log (6) relative , >&u in our case , s is the coarse simulation result and s 1985) is the 500 - meter simulation result . Törnqvist et al. ( where , >&u s proposed this method as a replacement for the ordinary relative difference calculated as ) / s ( , because of its − s >&u >&u - additive, symmetric and 1 to infinity, normed properties. For positive values, the ordinary relative difference ranges from 20 , we e log difference results in a positive bias when performing addition or subtraction. Using th which is asymmetric and c quantify the isolated contri bution of a process, by subtracting the log relative simulation difference of the altered ould * ) approximates simulation from that of the base simulation, without introducing bias. The log relative difference (i.e. log − the ordinary relative difference (i.e. ) for small values, but deviates strongly for large values due to the non - linear w = 1 * For example, an ordinary relative difference of 0.5 is scale, which has to be considered when interpreting the results. 25 equivalent to a log relative difference of 0.41, and analogously 1.0 translates to 0.69 log relative difference. In Using the same method, we also isolated the error that originates from the use of biome - specific LUE’s and turnover rates. class he most commonly occurring land cover the base model simulation, t wa s used for the entire aggregate grid cell, and the 30 at grid cell re then applied to th This leads to misclassification of the we majority biome turnover rates and LUE of the . (Foody, 2002) , which we will refer to as biome misclassification . Furthermore, the biome - minority land cover classes we sidering the minority biomes in the grid cell, leading to re used without con majority biome specific parameters of the as the biome - specific parameter error . We c ould account for this error by running the model for each biome what we refer to 11

12 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © re correctly used for specific parameters s each individual biome. Then, the overall result - separately, so that the biome we wa computed by summing the individual biome results, weighted by their respective fractional cover in the grid cell. Again, the contribution of the error mechanism. altered run wa s compared to the base model simulation, in order to quantify the treated all average - based aggregated input data on a per - biome basis. In a second ‘per - biome’ approach, we additionally 5 result afterwards, as for Because our model algorithm is nonlinear, Jensen’s inequality exists between averaging the model meter model resolution, and averaging the input data (i.e. FTC, NTV, fPAR and BA) beforehand, as for the - the 500 input aggregated resolution simulations. We will refer to the error due to average - based aggregation of input datasets as t he - aggregation error . In order to account for this error, we aggregated all MODIS 500 meter resolution input datasets to coarser 10 resolution according to the individual biome fractions in each grid cell. In other words, we created average based aggreg ated - input datasets for each biome area separately, instead of one aggregate for the entire grid cell area. These aggregation products were then used in the corresponding simulation of the individual biome, and the individual results were again is only meaningful when specific as well, summed afte - the LUE and turnover rates are biome rwards. This altered simulation - and should thus be seen as an addition to the altered simulation that accounts for the biome specific parameter error, as described in the previo us paragraph. 15 van der Finally, a simulation was performed with the incorporation of a modified burned fraction (MBF), as described by ( 2017) and used in GFED4. They introduced the MBF to account for the underestimation of emissions in Werf et al. model resolutions. The uniform burning of a fraction of an aggregated grid cell leads to frequently burning areas at 0.25 ° ell is , because in this case fuel in the whole grid c when fires occur in the subsequent months underestimation of emissions 20 , also in areas that did not burn. In reality, the fuel is only lowered in the s burning in previous months lowered by the fire - extent grid cell area that did not burn yet. The fraction of the grid cell that actually burned, and subsequent fires burn the sub resolution model allow us to directly test meter - 500 of underestimation is mainly dependent on the fire return time. Our ed . We used a 4 the effectiveness of implementing an MBF month time period per burning season, at coarser resolution s - 2017) . 25 ( analogous to van der Werf et al. 3. Results lutions, based on model calibrations for either 500 - meter or 0.25 ° resolution, to We ran our m odel for a range of spatial reso differences. First, we discuss the results of the model calibration and and simulation better understand the calibration validation for AGBw, FL and FC at 500 n. Next, we discuss the resulting AGBL and emission estimates - meter resolutio 30 ° based on this model. Then we compare this to the results for the 0.25 resolution model calibration and relate this to GFED4. the differences that occur between simulations with The remainder of this chapter is dedicated to simulation differences , i.e. we used the model calibrated at 500 - meter resolution. In e study of simulation differences, different spatial resolutions. For th 12

13 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © mainly and 500 - meter resolutions, and finish with the results for section focus on the differences between the 0.25 this we ° . ° the intermediate resolutions of 0.125 ° and 0.050 3.1 Woody biomass and fuel load ( 2016) reference dataset. The modelled total AGBw for Africa wa s 40.2 Pg C, compared to 40.6 Pg C for the Avitabile et al. s wa . 1). This wa s able to capture most of the spread in values (Fig 5 , and the model ed Area averaged values correspond well wa s as expected , since the model was calibrated to match the reference dataset. On a biome level, AGBw for tropical forest s 6.9 Pg C and 1.1 wa 31.2 Pg C, which s 1.4 Pg C lower than the reference dataset. For woody and open savannas AGBw wa re slightly overestimated by 0.1 Pg C. The tropical forest, woody savanna and open we avannas Pg C, respectively. Open s the vast majority of tree biomass in Africa. Even though area averaged AGBw for other forest ed savanna biomes contain id d not contribute 10 only 0.6% of African land surface and d types wa constitute s significant, these biomes together significantly to total AGBw. The comparison between model . 2a. l ed FL and field measurements based on the 500 - meter model calibration is shown in Fig 2 to overestimate low FL and A robust agreement was found, with an r value of 0.78 (r = 0.60). The model tend ed well between model ed and 15 d underestimate high FL. Overall, the average, median and spread over all field sites agree l 2 b). On average, FL for woody savanna and grassland wa s overestimated by approximately 15% (46 measured values (Fig . - - 2 2 re underestimated by 11% (20 g C m g C m ), whereas for shrubland values ). Model estimates for and 28 we , respectively well with measuremen ts, even though the range of values s underestimated for this biome. The d pen savannas agree wa o we shrubland and grassland statistics re both based on only 4 or 5 field measurements, which explains the large differences in of averages and ranges. 20 quantiles, and restrict ed the analysis to a comparison resolution. As in Fig 2a and b, Figure 2c and d shows the same comparison to field measurements, but simulated at 0.25 ° . - ed in simulation resolu tion. s based on the 500 wa this comparison meter resolution calibration, and thus only differ l ed at 0.25 ° resolution ha d a much lower range and wa s Compared to the 500 - meter resolution simulation, FL model substantially underestimated in most biomes except grasslands. The flat regression slope indicates that the spread in 25 s partly because several FL measurements we re within one 0.25 s not captured at coarse resolution. This wa measured FL wa ° the same model value. The 0.25 ° simulation show ed large estimation errors, especially for high FL grid cell and thus yield ed wa s incorrectly classified as tropical forest, leading to a large measurements. One woody savanna measurement 2d, black triangles). overestimation of FL (Fig . 13

14 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © 3.2 Fire emissions - 1 2017 annually averaged total fire emissions for sub 0.68 Pg C yr The 2002 we based on the 500 m model, re - - Saharan Africa 2 - . s dictated by burned wa burned (Fig with an average FC of 249 g C m 3, solid blue). The spatial distribution of emissions in the subtropical savanna regions. About 90% of the fire emi ssions (0.61 occurred 4a). The majority of emissions . area (Fig - 1 5 from the woody and open savanna regions, where the majority (87%) of the annually averaged burned Pg C yr ) originate d - 2 area wa s found. The highest FC wa s found in the tropical forest, where the average wa s 998 g C m burned. However, the s relatively low in this biome, resulting in low emissions. wa ned area bur Saharan Africa These 500 re 24% lower than GFED4 (without small fires), which - meter emission estimates for sub - we 1 - 2 - , and an average FC of 331 g C m burned (Fig . 3). All biomes with substantial emissions 10 estimate d 0.90 Pg C yr re the . The lower emissions in our model we re the biggest contributor to this difference, but woody savannas contribute we d compared to GFED4, because the amount of BA of the two estimates wa s identical. T he direct result of differences in FC in our model wa - dominated biomes. The spatial distribution of FC s less variable for most biomes , except for the forest a similar way as GFED4. s represented in wa variability of FC across biomes in the model 3.3 Calibration differences due to spatial resolution 15 resolution model calibration differ ed from the 500 - meter calibration in terms of slower The parameters used for the 0.25 ° in roughly a 2.5 Pg C increase in ed turnover rates for the stem, grass and litter pools for some biomes (Table 1). This result AGBw and a 3.0 Pg C increase in AGBL. The majority of this increase ted for by the savanna biomes (both open s accoun wa more than woody AGB. Figure 5 shows the resulting and woody). Comparatively, non d - woody AGB and litter increase resolution model ° 5 - m et er 20 equivalent to Fig. 2, but for the 0.2 – modelled FL compared to field measurements instead of 500 ° calibration. For this coarse calibration, again simulations for both 500 - meter (panel a and b) and 0.25 resolution (panel c and d) are shown. ° resolution agree d be tter with measurements for all By calibrating the model at 0.25 ° resolution, the FL simulated at 0.25 25 . 5c, d). On the other hand, with this coarse calibration, the 500 - meter resolution 2c, d and Fig . biomes (compare Fig FL, and perform ed poorer than the 0.25 ° resolution simulation in terms of biome simulation significantly overestimate d 5a and b). An exception 0.25 ° model pixels respective . average and distribution (Fig wa s the shrubland biome, for which all biomass in a consistent underestimation of . This result ed - to the field sites we re strongly influenced by low areas in that pixel calibration, the 500 ed much better meter resolution simulation still show FL for both calibration resolutions. For the 0.25 ° - 2 2 r r = 0.07). Th e 0.25 ° calibration led to a 30 correlation with measurements ( = 0.57), compared to the coarser simulation ( wa s amplified relative to low FL. s steeper and closer to 1 for both resolutions, as high FL wa regression slope that 14

15 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © th resolutions ° The increase in biomass and litter when using the 0.25 calibration led to higher emissions; 0.82 Pg C for bo s much closer to the estimate of 0.90 Pg C from GFED4 (without small fires) wa (Fig 3, transparent bars). As expected, this . e - than for the 500 s in the open and woody savanna biomes. Th wa meter calibrated model. The largest change in emissions s resemblance to GFED4 emissions suggest i s able to roughly reproduce GFED4 when calibrated at that our simplified model enabling 500 - meter resolution modelling. the same resolution of 0.25 ° , while additionally 5 3.4 Simulation differences due to spatial re solution because of resolution differences during the actual simulations. We ied Besides calibration differences, model results var - meter ° meter and 0.25 - calculated simulation differences between 500 resolution runs, using the model calibrated for 500 resolution (Fig resolution simulation, AGBw and AGBL estimates we re 4.0 Pg C and 4.6 Pg C lower than . 6). For the 0.25 ° ig 7a). The main positive differences we re 10 the 500 - meter simulation, respectively, with contributions from all biomes (F . o found at the transitions from barren to vegetated landscapes (e.g. at the fringes of the Sahara and Kalahari) and from land t re also found in the southern part of water. Notably, smaller positive differences West Africa. The average AGBL s wa we . lower at coarser resolution for all biomes, with a larger difference for biomes with more biomass (Fig 7b). - 1 meter resolution sim we , which wa s 3% lower than the 500 - run ulation 15 ° Total fire emissions for the 0.25 re 0.66 Pg C yr 3, solid orange versus blue bars). Even though the total emission estimates at different resolutions we re relatively (Fig . factor of , with deviations up to a (up occurred similar, significant regional differences in emissions 1.5, and higher deviations we . re mostly the to a factor of 4) at the border of water bodies and deserts (Fig 6b). The lower emissions at 0.25 resolution result of lower emissions in savannas and other grass dominated biomes, whereas for tropical forests (i.e. Congo Basin) and - 7a and b). Note that the relative differences shown for emissions are the 20 re higher (see also Fig ther forests, emissions o . we same as for FC, because the burned area is equal for both resolutions. 3.5 Disentangling of mechanisms were modest at continental scale, but simulation differences at regional that The results in section 3.4 i ndicated substantial . The various mechanisms that explain part of these differences were identified and quantified by doing additional scales section 2.5.2 ). 25 simulations with altered model confi gurations, and comparing them to the base model (see AGBL 3.5.1 Simulation differences for . 6 ) . Figure 8 shows the contributions of several error mechanisms to the total simulation difference in AGBL ( shown in Fig Figure 8a and b specific parameter error, and the input aggregation error, respectively . depict the contribution of the biome - . More Figure 8c and d show the remaining simulation difference, after subtraction of these two error mechanisms . 8c shows the part rel ated to fire processes (by doing a fire - off simulation), and Fig . 8d shows the 30 specifically, Fig 15

16 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © s negligible and therefore not shown ( 10). This is wa unexplained remainder. The effect of the MBF on AGBL Fig but see . by ground fires that only burn grass and litter, whereas AGBL is because frequent fires are mainly found in areas dominated mostly determined by stem biomass and thus mainly affected by canopy fires, which are less frequent. simulation difference in AGBL - a very substantial role in the base model Errors due to biome ed specific parameters play 5 (Fig - . 8a). For most of the African continent the majority of difference c ould be explained by this mechanism. The biome - - d ominated dominated biomes and higher AGBL for forest for lower AGBL for grass ed specific parameter error account biomes at coarser resolution. This shows that this error mechanism does not explain the strong negative difference in the we re strongest at the transition borders of biomes, where the distribu tion tropics in the total difference (Fig . 6a). Differences of land cover types is generally more heterogeneous. Examples are the transition of open savanna to woody savanna towards 10 ° S latitude), the transition of woody savanna to tropical forest towards the equator (at 5 ° N and ° N and 15 the equator (at 10 S latitude), and the transition towards the Sahara Desert. 5 ° Figure 8b shows the simulation differences in AGBL due to the input aggregation error. Compared to the biome - specific the two error mechanisms partly neutralize d 15 parameter error, the input aggregation error wa s relevant in other areas, and . s 8a and b). Substantial negative differences we re found at the transitions to forest biomes each other (opposite sign in Fig found at the transition to deserts and re (e.g. Congo rainforest, eastern South Africa, Madagascar). Positive differences we that these large positive differences occur where the majority - aggregated biome water bodies. Further investigation show ed zero biomass. - is water or desert, leading to large relative difference due to near The remaind er of simulation difference in s mixed positive and negative, specific parameter error and input aggregation error, - AGBL, after subtraction of the biome 20 wa . difference in the tropics be attributed to fire processes (Fig ould of which all negative difference c 8c and d). This negative . 8c, compare to Fig . explain 5b). This leaves an ed a large part of the total simulation difference in that region (Fig unexplained simulation difference of solely positive values, analogous to an overestimation of AGBL in the 0.2 5 ° resolution simulation. 25 ed the Because the various error mechanisms influence each other, the order of isolation of different mechanisms affect for which the resulting relative difference. This wa s especially the case for the isolation of fire related error mechanism, vary by up to difference dependent on the order of isolation. This wa s the case because the % ould relative difference c 25 ical for a part of the negative fire related difference in the woody savannas and the trop ed input aggregation error account 8b and c), and the biome - specific parameter error account ed for a 30 forest edge (note overlapping negative pattern in Fig . a negligible impact. d positive part in the open savannas. For the other mechanism, MBF, the order of isolation ha 16

17 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © differences for fire emissions 3.5.2 Simulation fuel load available for The results shown above concerned simulation differences in AGBL, which directly dictates the ally less re gener we burning, and is thus a key precursor for fire emissions. The simulation differences in emission ed as a dipole of positive and negative . - 9). The biome specific parameter error show pronounced than for AGBL (Fig 5 9a), similar to the pattern seen for AGBL. The difference due to the input . difference around biome transitions (Fig aggregation err 9b), and partly neutralize d the biome - specific parameter error. or wa s mostly positive (Fig . ed for a substantial part of the total simulation s negative everywhere, and account wa The isolated error related to the MBF s that this measure i s indeed able to remove errors at aggregated difference in savanna regions (Fig . 9c). This show van der Werf et al. s 2017) . The traditional way of accounting for 10 resolution related to short fire return times, as reasoned by ( s aggregated fire in a model (unmodified burned fraction) causes an underestimation of emissions at resolution in frequently . in Fig. 9c landscapes, which translates to a negative simulation difference as shown burning The remainder of simulation difference in emissions, after subtraction of all identified error mechanisms, wa s predominantly positive, especially in the 9d). region of the Con go tropical rainforest (Fig . 15 3.5.3 Relative contribution of error mechanisms explain the majority of total simulation difference in AGBL and For all biomes, the error mechanisms that we identified - 8 average relative difference in AGBL 10). From the (blue dot in . emissions, except for tropical forest emissions (Fig 0.0 between the base model running at 0.25 ° versus 500 - meter resolution , 47 % c ould be attributed to biome - Fig. 10) for Africa aggregation errors. Furthermore, 16 % to input wa s related to fire, 5 % to the 2 specific parameter errors and an additional % 20 s unexplained. This analysis - 0.0 1 s also performed for emissions, showing that from the wa wa MBF, and the remaining 3 0 % between the base model running at 0.25 average relative difference versus 500 - meter resolution , (orange dot in Fig. 10) ° be attributed to biome - specific parameter errors and an additional 15% to input aggregation errors. The MBF ould 30% c for 2 2 %, and 33% remain s unexplained. account ed we specific parameter errors. For AGBL, this 25 On a biome level, most of the simula tion differences re explained by biome - - dominated biomes. For tree - dominated biomes, input mechanism explain ed the large majority of difference for the grass nt instead. For most biomes, various error mechanisms partly we aggregation errors and fire processes re more importa each other, resulting in a reduced overall difference. The MBF mostly affect ed emissions, and as expected the neutralize d quent fires. The unexplained remaining difference s largest for biomes with considerable burned area and fre wa contribution we re fully explained. was 30 positive for all biomes, and only the emission differences for the grassland and cropland biomes 17

18 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © 4 3.5. Simulation difference as a function of resolution , 0.050 ° - meter (which we assume to be 0.005 ° for this Across the four analyzed , 0.125 ° , 500 ° spatial resolutions (0.250 ed comparison)), the absolute differences in AGBL and emissions follow a gradual trend described by a natural logarithmic re the difference in cropland AGBL, and in woody savanna emissions. The absence of a logarithmic we function. Exceptions trend in these cases negative difference s within the biome. This suggests 5 wa s caused by a mixture of pixels with positive and for different that for these biomes the trend i s explained by a c ombination of concave upward and downward logarithms pixels, e variability within th as a result of biome. In case of emissions, the absence of a logarithmic trend for the whole of the pattern of woody savannas. We sub d the sensitivity of the simulation difference as the - Saharan Africa reflect ed estimate derivative of the fit function. The general form of the fit function is: ) ( log + { , (7) 10 * + z x x * z , and { are constants. Given the logarithm quotient rule: , is the spatial resolution in degrees, and where ~  " d ( ) ) ( (8) * * + z x = x log } + z log − x log Å , | b " ~  Ä , ≪ twofold increase or decrease in resolution results in a constant change in simulation difference. z and assuming each * - 2 ) ( log 1 2 roughly = 0 . . gC m 17 ∙ per twofold change in is issions For example, the sensitivity for open savanna em 81 resolution (increase or decrease) (see Fig , independent of the initial resolution, a twofold finer 15 . 11b). This means that (coarser) spatial resolution always leads to the same decrease (increase) in simulation difference. In other words, both at fine is not and coarse resolutions the model results are equally sensitive to reso lution changes. Importantly, in cases where z resolution (e.g. for tropical forest emissions) much smaller than x , the sensitivity decreases towards finer . 4. Discussion 1 - - Saharan Africa using a n emission model based on native MODIS 20 We estimated fire emissions of 0.68 Pg C yr for sub coarser resolution estimates were compared to , as - satellite spatial resolution of 500 meter. These relatively high resolutions . We (Rabin et al., 2017) used for most previous fire emission estimates such as from GFED and fire modules in DGVM’s ion occurring in both the calibration and simulation stage of our model. analyzed differences due to spatial resolut 25 With our simplified emission model, we were able to reproduce emissions from GFED4 on a continental and biome scale - 1 (Fig - meter resolution model were 0.22 Pg C yr . lower ( - 3 ). However, emissions for sub - Saharan Africa based on our 500 24%) than GFED4, with the largest difference for woody savannas. The difference with GFED4s emissions was larger not incl ude small fire burned area. The emission estimates for our model calibrated at 0.25 ° resolution because our model d id - 1 - 8%). Comparison of the 500 - were 0.14 Pg C yr meter version, and more in line with GFED4 ( - higher than for the 500 s illustrated that turnover rates governing biomass turnover and decomposition 30 resolution model calibration ° meter and 0.25 aggregated resolution under the used calibration approach , which led to higher fuel load and were required to be slower at 18

19 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © likely overestimates relatively coarse consequently higher emissions. This suggests that GFED due to its fuel consumption model resolution for similar reasons. - meter calibrated model version fuel consumption estimates and underlying faster turnover rates for the 500 The lower version, can partly be explained by a larger representation error when comparing model to 5 compared to the 0.25 ° resolution meter spatial resolution the representation error for model pixels compared to - field measurements. We showed that at 500 s greatly re duced, especially in heterogeneous landscapes with large spatial variation in wa individual field measurements additionally le d to a larger sample of usable measurements, because multiple field plots biomass. The improved resolution ould that would otherwise be located in one 0.25 be compared individually. However, also at 500 - meter resolution, ° pixel c he Furthermore, t part of the representation error remain s , mostly because of the uncertainty in field plot location and time. 10 demands inc reased model complexity, since small - scale heterogeneity comparison to field measurements at finer resolution is no longer averaged out and thus has to be represented in the model. ° we found that When comparing our 0.25 calibrated resolution model version to GFED4, which runs at the same resolution, 15 we the turnover rates governing biomass turnover and decomposition in our model re generally faster, despite the emission partly be explained by the simplifications made in our model when compared to estimates being relatively smaller. This c an as the absence of herbivory, grazing, fuelwood collection and explicit deforestation mechanisms; all processes GFED4, such i that remove additional biomass. Furthermore, because GFED4 s optimized globally and not only for Africa, turnover rates ( Lehmann et al. and Rogers et al. ( 2015) for example, discussed same biome across continents. 2014) can be different for the stics among continents. This is also influenced by the availability of field data the differences in vegetation and fire characteri 20 also be per continent, which is relatively poor for Africa. Finally, the faster turnover rates for the litter pools in particular can explained by the use of different soil m oisture data and subsequent parameterization of litter decomposition in our model. were slower and closer to GFED4 (Table Indeed, the effective turnover rates (i.e. after scaling by the abiotic scalar) for litter agree with the dataset developed by Avitabile et al. ( 2016) ). We have optimized our model (see section 2.3 ). 1 l ed AGBw to 2018) showed large negative biases in this dataset for savanna biomes, which suggests savanna stem 25 However, Bouvet et al. ( specific - ield data on biome turnover rates tend to be too fast in both our model and GFED4. This indicates that additional f biomass and turnover rates is required to better evaluate our model. - - 1 , the simulation difference in emissions of 0.02 Pg C yr Compared to the calibration difference in emissions of 0.14 Pg C yr 1 6 ). At 30 was much smaller. Regionally h owever, simulation differences in AGBL and emissions were substantial (Fig . ° N to 10 ° S belt but lower aggregated resolution, AGBL was lower almost everywhere and emissions were higher in the 10 in the surrounding latitudes. In order to explain these differences, we identified at least three error mechanisms that can specific parameter error, 2) input aggregation error, and 3) temporal effec ts due to - amplify or dampen each other: 1) biome 19

20 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © and fuel build grid fire return tim - - aggregation of burned area fractions specifically, as spatial averaging affects sub e (MBF) fire fuel recovery, explained below). - up rates after a fire (post - specific parameter error 4.1 Biome - ied from the use of biome med he simulation difference stem Overall, most of t specific parameters, especially where they var considerably between biomes. In our model, the turnover rates 5 we re particularly variable, and they var ied among biomes for specific values, which we expect - st, in GFED4 only the stem pool turnover rates are set to biome all biomass pools. In contra - to result in relatively small errors. The biome specific parameter error wa s largest in aggregate grid cells with a large sub - grid heterogeneity in land cover types and a large gradient in parameter values between neighboring pixels. Isolation of this s highest error mechanism showed the largest AGBL differences in biome transition regions, where the variation in biomes i ° resolution for grass - dominated regions is explained by the misclassification of grid cell 10 . 8a). Lower AGBL at 0.25 (Fig minority forest patches as grassland. At coarse resolution, average biomass is underestimated because the whole grid cell is simulated as a grassland (the majority biome), wher eas at finer resolution the presence of a forest is revealed, with ° accompanying different turnover rates and LUE. Conversely, higher AGBL at 0.25 dominated regions is resolution for tree - explained by the misclassification of grid cell minority grassland a s forest. This is comparable to previous studies that found 15 an underestimation of small fragmented burned or forest area and an overestimation of large homogeneous burned or forest to coarser resolution (Eva and Lambin, 1998; area, as a result of majority - based aggregation of binary classified pixels . Miettinen and Liew, 2009; Nelson et al., 2009; Silva et al., 2005) wa s mostly related to stem biomass, because AGBL mostly consist ed of stem The biome - specific parameter error in AGBL the largest range of turnover rates. 20 d biomass, and because the stem pool ha The stem turnover rates for the open and woody otably large biome ed savanna biomes differ - by a factor of 7, which explains the n specific parameter errors at transitions . ° 8a). The large negative differences on the Eastern S latitudes (Fig N and 15 ° between those biomes, such as around the 10 flank of the continent are also likely explained by a combination of the heterogenic mosaic of agriculture/savannas/forest biomes, be and thus a large variability in tree cover and related turnover rates in this region. However, the error can similarly ied with up to a factor of 4 (see Table 1). 25 significant for biomass pools other than the stem pool, since other turnover rates var This is especially important in biomes with little tree cover. cific parameter error in AGBL directly affect ed emissions by determining the amount of fuel, and The biome - spe relative we re differences in emissions consequently again AGBL via the removal of fuel. However, for most pixels the 9 ). Compared to 500 - meter resolution, running the model at 0.25 ° resolution generally increased 30 smaller than in AGBL (Fig . emissions in savannas. A small area of grassland burning in an area d emissions in tropical forests and decrease resolution s predominately covered with forest results in an overestimation , since the grid cell of emissions at aggregated 20

21 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © average fuel load is mostly determined by forest biomass. The resulting emissions resemble a misclassified forest fire, tch with high fuel load surrounded by a majority of instead of the actual grass fire. On the contrary, the burning of a small pa 9a shows positive . low fuel load leads to an underestimation of emissions at coarse resolution. For these reasons, Fig - negative dipole patterns around the biome transitions. Emissions we re overestimated o n the more forested side of each biome we - 5 transition, and underestimated on the grassier side of each transition. These patterns re the direct result of the biome re no biome - specific parameters for fire, so no additional error specific parameter error in AGBL (Fig s . 8a). There we wa s much smaller than for wa introduced in the calculation of emissions from AGBL. Notably, the relative error for emissions wa s mostly determined by stem biomass, whereas emissions we re AGBL. This can be explained by the fact that AGBL mostly determined by grass and litter (and leaf) biomass. 10 - We expect that the AGBL after a fire is only minorly influenced by the biome specific parameter error in emissions. Since most emissions originate from grass fires, there is a minor impact on stem biomass and thus AGBL. This is also indicated by is error in emissions is small where emissions are - . like pattern in Fig 9a, suggesting that th the absence of an emission - related significant. By performing a simulation without fire induced tree mortality, we established that virtually all fire ects AGBL by 15 8c) is caused by mortality, instead of by direct emissions. Fire mostly aff . resolution difference in AGBL (Fig killing trees, but this does not directly translate to emissions because there is a time lag in the combustion of dead stems (i.e. cwd) and the CC is relatively low. 4.2 Input aggregation error - The non linear behavior of the model algorithm le d to an additional input aggregation error because of Jensen’s inequality. s largest for input data with high spatial variability, and thus most apparent in grid cells with large wa This error 20 ng is strongest. Previously, we identified fire induced tree heterogeneity in land cover types, where the bias due to averagi . 8c). This pattern is clearly isolated resolution difference in AGBL (Fig - mortality to be the main reason for the fire resembled in Fig . 8b, which suggests that mortality is strongly affected by the input aggregation error. This can be explained by the quadratic factor in calculating mortality, which amplifies Jensen’s inequality. 25 By aggregating the input data for each biome separately, the input aggregation error wa s reduced by decreasing the sing this method, the spatial resolution wa s effectively increased va riability related to the heterogeneity in land cover types. U equal to the number of biomes. However, variability inside individual biomes remains roughly by a factor and is not accounted fo (Yuan et al., accounts for likely r using this approach, and the remaining unexplained simulation difference ith in those 30 2007) . Besides biomes, more or other aggregation classes can be chosen that ideally reduce the variability w classes as much as possible with as little classes as possible. This could for example be a division based on tree cover intervals. However, the input aggregation error is unavoidable when modelling at aggregated resolutions, unless an ate estimator for Jensen’s inequality can be derived to account for this error. Since the reanalysis climate datasets we appropri 21

22 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ° wa s d used ha a 0.25 resolution for both our coarse and fine resolution simulations, no additional input aggregation error , Jensen’s inequality will exist resolution climate data is in case y these input datasets. However, introduced b finer aggregated error will probably be less substantial, as climate data is generally smoother and more The . as well for these datasets homogenous spatially. 5 4.3 Burned area aggregation (temporal effects) owin g to temporal mechanisms. The aggregation of BA fractions to coarser resolution in particular led to additional errors , fire return time s, resulting in an Firstly, aggregated the aggregation of BA alter ed underestimation of emissions at van der Werf et al. 2017) accounted for this effect by resolution s , especially in frequently burning areas. In GFED4, ( resolution model, we were able to trate the effectiveness of the MBF. meter demons introducing the MBF. With our 500 - aggregated resolution , especially in 10 Indeed, the MBF (Fig. 9c) account ed for the majority of underestimation in emissions at frequently burning areas (mostly savannas). which occurs due to the temporal non linearity of the modelled biomass - We introduce a second mechanism related to this, 12 for a hypothetical . buildup. This is a case of Jensen’s inequality in the temporal dimension. The effect is illustrated in Fig 100% CC and a uniform fuel load. At 500 - meter resolution, the 15 case of stem biomass buildup. For this example, we assu me d burned fraction is binary; a pixel is either completely burned or unaffected. In case the pixel burns, all stem biomass is removed by th is hypothetical fire. In the next month, the biomass r egrows from the start of the regrowth curve, where the s in the case of slope is relatively steep, which leads to fast regrowth. On the contrary, aggregated burned fraction the same the total biomass. This results in slower regrowth fraction net amount of biomass is removed from the grid cell of , but only a his leads to an underestimation of fuel load at coarse resolution, due to slower T from a later point on the regrowth curve. 20 biomass regrowth on average after a fire. Even though in normal model scenarios with partial CC the effect would be less for extreme, it could still be of importance, especially in case of canopy fires. In general, the effect is stronger slow turnover rates and short fire return times. In case of a grass fire, most biomass has alr eady recovered in the first months after the the fire. Only in case of a short fire return time an effect could be noticeable. However, additional analysis is required to 25 quantify the contribution of this error mechanism. 4.4 Small fires - 1 0.02 = 0.12 Pg C yr We found that emission s (calibration minus simulation difference) we re in total lowered by 0.14 – to 500 ° when increasing model resolution from 0.25 - meter resolution. For these results, we have relied solely on MCD64A1 not acc id burned area dataset which d ount for small fire burned area. Therefore, this difference may be offset by an increase - 1 due to 30 in emissions due to the inclusion of small fires. In GFED4s, emissions increased by 0.36 Pg C yr in our study region ivalent increase in emissions due to small fires in our model would offset the small fires, as compared to GFED4. An equ - meter resolution model the inclusion of effects of spatial resolution threefold. However, our findings suggest that in a 500 22

23 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © D4s, small fire burned area is added to the MCD64A1 product burned small fires may affect emissions differently. In GFE area (Collection 5.1), followed by the calculation of emissions at 0.25 resolution as in GFED4. However, small fires mostly ° eforestation), where the land cover is particularly heterogeneous occur in croplands and at the border of tropical forests (i.e. d Consequently calculating small fire emissions 2012; van der Werf et al., 2017) (Randerson et al., . , the GFED4s approach for suggest that small fire Our results is prone to the error mechanisms that occur at coarse resolution as described in this work. 5 resolution or at least on a per - biome basis. in particular, should be model l ed at emission estimates, finer resolution burned area products. For example, finer This is especially relevant considering the ongoing development towards published a 30 - meter resolution burned area dataset for Northern America based on Landsat imagery. Hawbaker et al. ( 2017) ( Furthermore developed a dataset of 20 - meter resolution burned area for Africa derived from the 10 , Ro teta et al. 2019) d Sentinel that a very substantial amount of burned area is still - 2 MSI sensor. Their preliminary product assessment indicate whi ch includes small fires using a statistical approach . Our work illustrates how the development of missed , even in GFED4s resolution emission models or finer resolution burned area datasets should be accompanied by the development of finer s . Even in our 500 - meter resolution emission model, sub 500 - meter better parameterization , in order to red uce error . s heterogeneity in burned area and fuel load is not accounted for, and could introduce additional error 15 spatial resolution, the model s roughly equally sensitive to resolution changes as wa We have shown that for relatively fine d that a twofold increase in resolution leads to a linear for coarser resolution (Fig . 11). The natural log relation implie ing of fire emission could reveal - new sources of error related to small - reduction of error. However, sub 500 meter model l relation might no longer be applicable. Dependent on the model precision - scale heterogeneity. At these scales the log 20 demands, an optimal spatial resolution can be chosen for which the simulation difference becomes in significant. However, A study by Nelson et al. ( 2009) , who the calibration difference can still be substantial , dependent on the representation error. there is an optimal resolution of around 300 that looked at the effect of spatial resolution on forest inventories, concluded meters at which the pixel size is slightly smaller than the forest patch size and the essential heterogenic characteristics of the landscape are captured. In line with our findings, this suggests there is a similar optimal resolution for burned area, and o ther 25 spatial data us ed in fire emission modelling, at which finer resolution does no longer significantly improve captured and computational resources are optimized. variability in the data used, 5. Conclusions - Saharan Africa, using the native spatial We have developed a carbon cycle model to estimate fire emissions for sub meter resolution with 30 - data (500 meter). A key objective was to compare fire emission estimates at 500 resolution of MODIS resolution of GFED4. We estima ted total fire emissions for sub - more often used coarser resolutions such as the 0.25 ° 23

24 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © 1 - averaged over 2002 Saharan Africa of 0.68 PgC yr - 2017. This is 24% lower than the most recent estimates from GFED4 (without small fires). s mainly caused by reduced representation errors in model calibration at resolution, when tuning the wa finer The difference model to match field measurements of fuel load. In addition, estimates we re different dependent on the resolution of the 5 local scale, these simulation differences were substantial, with differences up to a factor 4 in model simulation. At a more regions with large landscape heterogeneity, such as biome transition zones. The error mechanisms we identified as main used, and the consequent nces are all the result of spatial aggregation of the datasets contributors to these simulation differe coarse resolution model simulation. Spatial aggregation leads to a reduction of data variability, both in case of majority - based aggregation of land cover types, and in 10 case of average - based aggregation of all other, continuous, input data. The most of the simulation difference, and the remaining unexplained difference is most identified error mechanisms explain ed likely caused by the variability inside individual biomes , which wa s not accounted for in our method. This variability can only be fully accounted for by running the model at native resolution. Furthermore, temporal effects, such as differences in T hese temporal effects should be further - fire fuel recovery, may also explain part of the remaining d post ifference. investigated. 15 As a next step, we plan to run our model for the globe to improve global emission estimates, with a focus on highly heterogeneous regions such as deforestation zones. Understanding the unde rlying mechanisms that create errors in coarse enables the development of error reduction measures. This knowledge can be used to improve the next resolution model s 20 t least for Africa. Whether version of GFED. Our results indicate that fuel consumption in GFED may be overestimated, a correcting for the resolution global dependent errors discussed in this work will lead to lower emissions in the next version - of GFED, depends on to what exten small fire burned area may offset the decline in emissions. t availability Code an d data GFED4s data is publicly available at https://www.globalfiredata.org/. and model results are available upon request. Code Author contribution 25 van Wees and G uido R. van der Werf , and performed by Dave van Wees under principal The research was designed by D ave manuscript Dave van Wees with contributions from was written by Guido R. van The , supervision by Guido R. van der Werf der Werf. 24

25 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © Competing interests they have no conflict of interest. The authors declare that Acknowledgements We would like to thank Ioannis This work was supported by the Netherlands organization for scientific research (NWO). Bistinas for useful discussions. 5 References Akagi, S. K., Yokelson, R. J., Wiedinmyer, C., Alvarad o, M. J., Reid, J. S., Karl, T., Crounse, J. D. and Wennberg, P. O.: Emission factors - 4039 - – 4072, doi:10.5194/acp for open and domestic biomass burning for use in atmospheric models, Atmos. Chem. Phys., 11(9), 4039 11 - 2011, 2011. Alleaume, S., Hély, C., Le Roux, J., Korontzi, S., Swap, R. J., Shugart, H. H. and Justice, C. O.: Using MODIS to evaluate heterogeneity 10 4237, of biomass burning in southern African savannahs: A case study in Etosha, Int. J. Remote Sens., 26(19), 4219 – 2005. doi:10.1080/01431160500113492, tion, Nat. Andela, N. and van der Werf, G. R.: Recent trends in African fires driven by cropland expansion and El Niño to la Niña transi 795, doi:10.1038/nclimate2313, 2014. Clim. Chang., 4(9), 791 – Andreae, M. O. and Merlet, P.: Emission of trace gase 15 966, – s and aerosols from biomass burning, Global Biogeochem. Cycles, 15(4), 955 2001. Archibald, S., Roy, D. P., van Wilgen, B. W. and Scholes, R. J.: What limits fire? An examination of drivers of burnt area in Southern 613 – 630, doi:10.1111/j.1365 - 2486.2008.01754.x, 2009. Africa, Glob. Chang. Biol., 15(3), n, L., Bayol, N., Avitabile, V., Herold, M., Heuvelink, G. B. M., Lewis, S. L., Phillips, O. L., Asner, G. P., Armston, J., Ashton, P. S., Bani Gonzalez, G., Lucas, R., - rardin, C. A. J., Kearsley, E., Lindsell, J. A., Lopez Berry, N. J., Boeckx, P., de Jong, B. H. J., Devries, B., Gi 20 Malhi, Y., Morel, A., Mitchard, E. T. A., Nagy, L., Qie, L., Quinones, M. J., Ryan, C. M., Ferry, S. J. W., Sunderland, T., L aurin, G. V., - , H., Wijaya, A. and Willcock, S.: An integrated pan tropical biomass map using multiple reference Gatti, R. C., Valentini, R., Verbeeck datasets, Glob. Chang. Biol., 22(4), 1406 – 1420, doi:10.1111/gcb.13139, 2016. - Beerling, D. J. and Osborne, C. P.: The origin of the savanna biome, Glob. Chang . Biol., 12(11), 2023 – 2031, doi:10.1111/j.1365 25 2486.2006.01239.x, 2006. Bian, L.: Multiscale Nature of Spatial Data in Scaling Up Environmental Models, in Scale in Remote Sensing and GIS, pp. 13 26., 1997. – ground biomass map of African Mathieu, R., Naidoo, L. and Asner, G. P.: An above Bouvet, A., Mermoz, S., Le Toan, T., Villard, L., - savannahs and woodlands at 25 m resolution derived from ALOS PALSAR, Remote Sens. Environ., 206, 156 – 173, doi:10.1016/j.rse.2017.12.030, 2018. ardner, R. H.: Aggregation error in nonlinear ecological models, J. Theor. Biol., 100(3), 539 Cale, W. G., O’Neill, R. V and G 30 – 550, - 5193(83)90446 - doi:https://doi.org/10.1016/0022 0, 1983. Ciais, P., Sabine, C. L., Bala, G., Bopp, L., Brovkin, V., Canadell, J. G., Chhabra, R., Galloway, J. , Heimann, M., Jones, C. D., Le Queré, C., Myneni, R. B., Piao, S. L. and Thornton, P.: Carbon and other biogeochemical cycles, in Climate Change 2013: The Physical Science overnmental Panel on Climate Change, in Working Basis. Contribution of Working Group I to the Fifth Assessment Report of the Interg 35 group I contribution to the IPCC fifth assessment report., 2013. 25

26 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © s. Crosetto, M., Moreno Ruiz, J. A. and Crippa, B.: Uncertainty propagation in models driven by remotely sensed data, Remote Sen – - 5, 2001. Environ., 76(3) 4257(01)00184 385, doi:10.1016/S0034 , 373 - , Bauer, P., Dee, D. P., Uppala, S. M., Simmons, A. J., Berrisford, P., Poli, P., Kobayashi, S., Andrae, U., Balmaseda, M. A., Balsamo, G. mann, N., Delsol, C., Dragani, R., Fuentes, M., Geer, A. J., Haimberger, Bechtold, P., Beljaars, A. C. M., van de Berg, L., Bidlot, J., Bor - Sanz, B. M., 5 L., Healy, S. B., Hersbach, H., Hólm, E. V., Isaksen, L., Kållberg, P., Köhler, M., Matricardi, M., Mcnally, A. P., Monge - snay, P., Tavolato, C., Thépaut, J. N. and Vitart, F.: The ERA Morcrette, J. J., Park, B. K., Peubey, C., de Ro Interim reanalysis: – 597, doi:10.1002/qj.828, 2011. Configuration and performance of the data assimilation system, Q. J. R. Meteorol. Soc., 137(656), 553 Denny, M.: The fallacy of the average: o n the ubiquity, utility and continuing novelty of Jensen’s inequality, J. Exp. Biol., 220(2), 139 – 146, doi:10.1242/jeb.140368, 2017. 10 nuous Dimiceli, C., Carroll, M., Sohlberg, R., Kim, D. H., Kelly, M. and Townshend, J. R. G.: MOD44B MODIS/Terra Vegetation Conti Fields Yearly L3 Global 250m SIN Grid V006 [Data set], NASA EOSDIS L. Process. DAAC, doi:10.5067/MODIS/MOD44B.006, 2015. 3), – Duursma, R. A. and Robinson, A. P.: Bias in the mean tree model as a consequence of Jensen’s inequality, For. Ecol. Manage., 186(1 1127(03)00307 373 4, 2003. – 380, doi:10.1016/S0378 - - Eva, H. and Lambin, E. F.: Remote Sensing of Biomass Burning in Tropical Regions: Sampling Issues and Multisensor Approach, , 15 – 315(July 1997), 292 315, 1998. Field, C. B., Randerson, J. T. and Malm ström, C. M.: Global net primary production: Combining ecology and remote sensing, Remote 88, doi:10.1016/0034 4257(94)00066 - V, 1995. – Sens. Environ., 51(1), 74 - - Foody, G. M.: Status of land cover classification accuracy assessment, Remote Sens. Environ., 80 (1), 185 – 201, doi:10.1016/S0034 4257(01)00295 - 4, 2002. Friedl, M. A., Sulla Menashe, D., Tan, B., Schneider, A., Ramankutty, N., Sibley, A. and Huang, X.: MODIS Collection 5 global land - 20 182, ote Sens. Environ., 114(1), 168 cover: Algorithm refinements and characterization of new datasets, Rem – doi:10.1016/j.rse.2009.08.016, 2010. García Lázaro, J. R., Ruiz, J. A. M. and Arbeló, M.: Effect of spatial resolution on the accuracy of satellite - based fire scar detection in the 4753, doi:10.1080/01431161.2013.781290, 2013. – t. J. Remote Sens., 34(13), 4736 northwest of the Iberian Peninsula, In 25 Giglio, L., van der Werf, G. R., Randerson, J. T., Collatz, G. J. and Kasibhatla, P.: Global estimation of burned area using MODIS active 6 - 957 - – fire observations, Atmos. Chem. Phys., 6(4), 957 974, doi:10.5194/acp - 2006, 2006. eration Giglio, L., Randerson, J. T. and van der Werf, G. R.: Analysis of Daily, Monthly, and Annual Burned Area Using the Fourth Gen Global Fire Emissions Database (GFED4), J. Geophys. Res. Biogeosciences, 1 18, 317 – 328, doi:10.1002/jgrg.20042, 2013. Giglio, L., Boschetti, L., Roy, D. P., Humber, M. L. and Justice, C. O.: The Collection 6 MODIS burned area mapping algorithm and 30 – 018. 85, doi:10.1016/j.rse.2018.08.005, 2 product, Remote Sens. Environ., 217(March), 72 ldwell, M. K., Hawbaker, T. J., Vanderhoof, M. K., Beal, Y., Takacs, J. D., Schmidt, G. L., Falgout, J. T., Williams, B., Fairaux, N. M., Ca - series of sing dense time Picotte, J. J., Howard, S. M., Stitt, S. and Dwyer, J. L.: Remote Sensing of Environment Mapping burned areas u Landsat data, Remote Sens. Environ., 198, 504 – 522, doi:10.1016/j.rse.2017.06.027, 2017. 2000 characterization of fuels, fire behavior, combustion - Hély, C., Alleaume, S., Swap, R. J., Shugart, H. H. and Justice, C. O.: SAFARI 35 compl – 394, eteness, and emissions from experimental burns in infertile grass savannas in western Zambia, J. Arid Environ., 54(2), 381 doi:10.1006/jare.2002.1097, 2003. O.: A temporally explicit production Hély, C., Caylor, K. K., Dowty, P., Alleaume, S., Swap, R. J., Shugart, H. H. and Justice, C. - 007 - 9082 - 3, 2007. efficiency model for fuel load allocation in southern Africa, Ecosystems, 10(7), 1116 – 1132, doi:10.1007/s10021 2), 47 65, doi:10.1016/S0016 - – – oderma, 89(1 Heuvelink, G. B. M. and Pebesma, E. J.: Spatial aggregation and soil process modelling, Ge 40 7061(98)00077 - 9, 1999. 26

27 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © Hoelzemann, J. J., Schultz, M. G., Brasseur, G. P., Granier, C. and Simon, M.: Global Wildland Fire Emission Model (GWEM): Res. D Atmos., 109(14), doi:10.1029/2003JD003666, 2004. Evaluating the use of global area burnt satellite data, J. Geophys. phys. Res. D Ito, A. and Penner, J. E.: Global estimates of biomass burning emissions based on satellite imagery for the year 2000, J. Geo 18, doi:10.1029/2003JD004423, 2004. – Atmos., 109(14), 1 ć 5 Janji , T., Bormann, N., Bocquet, M., Carton, J. A., Cohn, S. E., Dance, S. L., Losa, S. N., Nichols, N. K., Potthast, R., Waller, J. A. and 1278, doi:10.1002/qj – Weston, P.: On the representation error in data assimilation, Q. J. R. Meteorol. Soc., 144(713), 1257 .3130, 2018. – 193, Jensen, J. L. W. V: Sur les fonctions convexes et les inegalites entre les valeurs moyennes, Acta Math., 30, 175 doi:10.1007/BF02418571, 1906. J., Razinger, M., Schultz, M. G., Suttie, M. Kaiser, J. W., Heil, A., Andreae, M. O., Benedetti, A., Chubarova, N., Jones, L., Morcrette, J. 10 and van der Werf, G. R.: Biomass burning emissions estimated with a global fire assimilation system based on observed fire ra diative - 9 - 2012, 2012. - 554, doi:10.5194/bg – power, Biogeosciences, 9(1), 527 527 zi, S., Roy, D. P., Justice, C. O. and Ward, D. E.: Modeling and sensitivity analysis of fire emissions in southern Africa du ring Koront SAFARI 2000, Remote Sens. Environ., 92(3), 376 – 396, doi:10.1016/j.rse.2004.06.010, 2004. van Leeuwen, T. T., van der Werf, G. R., Hoffmann, A. A., Detmers, R. G., Rücker, G., French, N. H. F., Archibald, S., Carvalho, J. A., 15 do, E. C., Cook, G. D., De Groot, W. J., Hély, C., Kasischke, E. S., Kloster, S., McCarty, J. L., Pettinari, M. L., Savadogo, P., Alvara , S., Meyer, C. P., Siegert, F., Trollope, L. A. and Trollope, W. S. W.: Biomass burning fuel consumption rates: A Boschetti, L., Manuri - 2014, 2014. - field measurement database, Biogeosciences, 11(24), 7305 – 7305 11 - 7329, doi:10.5194/bg Lehmann, C. E. R., Anderson, T. M., Sankara n, M., Higgins, S. I., Archibald, S., Hoffmann, W. A., Hanan, N. P., Williams, R. J., - Smith, J., Ryan, C. M., Durigan, G., Fensham, R. J., Felfili, J., Hutley, L. B., Ratnam, J., San Jose, J., Montes, R., Franklin, D., Russell - Hiernaux, P., Haidar, R., Bow 20 man, D. M. J. S. and Bond, W. J.: Savanna Vegetation - Fire Climate Relationships Differ Among Continents, 552, doi:10.1126/science.1247355, 2014. . )., 343(6170), 548 - Science (80 – Meyer, C. P., Cook, G. D., Reisen, F., Smith, T. E. L., Tattaris, M., Russell - Smith, J., Maier, S. W., Yates, C. P. and Wooster, M. J.: Direct 14, – ), 1 measurements of the seasonality of emission factors from savanna fires in northern Australia, J. Geophys. Res. Atmos., 117(20 doi:10.1029/2012JD017671, 2012. area mapping in insular South - East Asia, Int. J. Wildl. 25 . C.: Burn - Miettinen, J. and Liew, S scar patterns and their effect on regional burnt - 847, doi:10.1071/WF08102, 2009. Fire, 18(7), 837 – Myneni, R., Knyazikhin, Y. and Park, T.: MCD15A2H MODIS/Terra+Aqua Leaf Area Index/FPAR 8 - d ay L4 Global 500m SIN Grid V006 [Data set], NASA EOSDIS L. Process. DAAC, doi:10.5067/MODIS/MCD15A2H.006, 2015. Nelson, M. D., McRoberts, R. E., Holden, G. R. and Bauer, M. E.: Effects of satellite image spatial aggregation and resolutio n on 30 orest land area, Int. J. Remote Sens., 30(8), 1913 estimates of f – 1940, doi:10.1080/01431160802545631, 2009. Biogeosciences, Pettinari, M. L. and Chuvieco, E.: Generation of a global fuel data set using the Fuel Characteristic Classification System, :10.5194/bg - 13 - 2061 - 2016, 2016. 13(7), 2061 – 2076, doi - grid land surface variation on large - Pierce, L. L. and Running, S. W.: The effects of aggregating sub scale estimates of net primary production, Landsc. Ecol., 10(4), 239 – 253, doi:10.1007/BF00129258, 1995. Poorter, H., Nikla - analysis s, K. J., Reich, P. B., Oleksyn, J., Poot, P. and Mommer, L.: Biomass allocation to leaves, stems and roots: meta 35 8137.2011.03952.x, 2012. - – of interspecific variation and environmental control, New Phytol., 193(1), 30 50, doi:10.1111/j.1469 em Potter, C. S., Randerson, J. T., Field, C. B., Matson, P. A., Vitousek, P. M., Mooney, H. A. and Klooster, S. A.: Terrestrial ecosyst 841, – production: a process model based on global satellite and surface data, Global Biogeochem. Cycles, 7(4), 811 doi:10.1029/9 3GB02725, 1993. Rabin, S. S., Melton, J. R., Lasslop, G., Bachelet, D., Forrest, M., Hantson, S., Kaplan, J. O., Li, F., Mangeon, S., Ward, D . S., Yue, C., 40 an, T., Voulgarakis, A., Kelley, D. I., Arora, V. K., Hickler, T., Kloster, S., Knorr, W., Nieradzik, L., Spessa, A., Folberth, G. A., Sheeh 27

28 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © erimental and Colin Prentice, I., Sitch, S., Harrison, S. and Arneth, A.: The Fire Modeling Intercomparison Project (FireMIP), phase 1: Exp – - 1175 - 2017, 2017. analytical protocols with detailed model descriptions, Geosci. Model Dev., 10 10 1197, doi:10.5194/gmd (3), 1175 - sions Randerson, J. T., Chen, Y., Van Der Werf, G. R., Rogers, B. M. and Morton, D. C.: Global burned area and biomass burning emis 2012JG002128, 2012. from small fires, J. Geophys. Res. G Biogeosciences, 117(4), doi:10.1029/ Rogers, B. M., Soja, A. J., Goulden, M. L. and Randerson, J. T.: Influence of tree species on continental differences in bore 5 al fires and 234, doi:10.1038/ngeo2352, 2015. – climate feedbacks, Nat. Geosci., 8(3), 228 , A., Padilla, M., Storm, T. and Chuvieco, E.: Development of a Sentinel - 2 burned area algorithm: Generation of a Roteta, E., Bastarrika 17, doi:10.1016/J.RSE.2018.12.011, 2019. small fire database for sub - Saharan Africa, Remote Sens. Environ., 222, 1 – Roy, D. P. and Landmann, temporal reflective wavelength data, Int. - T.: Characterizing the surface heterogeneity of fire effects using multi – 4218, doi:10.1080/01431160500112783, 2005. 10 J. Remote Sens., 26(19), 4197 Ruel, J. J. and Ayres, M. P.: Jensen’s inequality predicts effects o f environmental variation, Trends Ecol. Evol., 14(9), 361 – 366, X, 1999. doi:10.1016/S0169 - 5347(99)01664 - Smith, J., Murphy, B. P., Meyer, C. P., Cook, G. D., Maier, S., Edwards, A. C., Schatz, J. and Brocklehurst, P.: Improving - Russell burning emissions for greenhouse accounting in northern Australia: Limitations, challenges, applications, Int. J. estimates of savanna – 18, doi:10.1071/WF08009, 2009. Wildl. Fire, 18(1), 1 15 Saharan Afr Scholes, R. J., Archibald, S. and von Maltitz, G.: Emissions from Fire in Sub ica: the Magnitude of Sources, Their Variability - – and Uncertainty, Glob. Environ. Res., 15, 53 63, 2011. Schultz, M. G., Heil, A., Hoelzemann, J. J., Spessa, A., Thonicke, K., Goldammer, J. G., Held, A. C., Pereira, J. M. C. and v an Het 17, doi:10.1029/2007GB003031, Bolscher, M.: Global wildland fire emissions from 1960 to 2000, Global Biogeochem. Cycles, 22(2), 1 – 2008. 20 Seiler, W. and Crutzen, P. J.: Estimates of gross and net fluxes of carbon between the biosphere and the atmosphere from biom ass burning, im. Change, 2(3), 207 – Cl 247, doi:10.1007/BF00137988, 1980. Silva, J. M. N., Sá, A. C. L. and Pereira, J. M. C.: Comparison of burned area estimates derived from SPOT - VEGETATION and Landsat 201, – , Remote Sens. Environ., 96(2), 188 ETM+ data in Africa: Influence of spatial pattern and vegetation type doi:10.1016/j.rse.2005.02.004, 2005. 25 – Törnqvist, L., Vartia, P. and Vartia, Y. O.: How should relative changes be measured?, Am. Stat., 39(1), 43 46, doi:10.1080/00031305.1985.10479385, 1985. , Mahowald, N. M., Rogers, B. M., Randerson, J. T. and Hess, P. G.: The changing radiative forcing of fires: global Ward, D. S., Kloster, S. 2012, 2012. 10857 - 10886, doi:10.5194/acp – - model estimates for past, present and future, Atmos. Chem. Phys., 12(22), 10857 - 12 van der Werf, G. R., Randerson, J. T., Collatz, G. J. and Giglio, L.: Carbon emissions from fires in tropical and subtropical ecosystems, 30 2486.2003.00604.x, 2003. Glob. Chang. Biol., 9(4), 547 – 562, doi:10.1046/j.1365 - atz, G. J., Kasibhatla, P. S. and Arellano, Avelino F., J.: Interannual variability in van der Werf, G. R., Randerson, J. T., Giglio, L., Coll 2006, 2006. – 3175 - - global biomass burning emissions from 1997 to 2004, Atmos. Chem. Phys., 6(11), 3423 3441, doi:10.5194/acpd 6 - io, L., Collatz, G. J., Mu, M., Kasibhatla, P. S., Morton, D. C., Defries, R. S., Jin, Y. and van van der Werf, G. R., Randerson, J. T., Gigl 2009), - 997 Leeuwen, T. T.: Global fire emissions and the contribution of deforestation, savanna, forest, agricultural, and peat fires (1 35 - - - 11707 11735, doi:10.5194/acp 2010, 2010. 10 Atmos. Chem. Phys., 10 (23), 11707 – industrial biomass burning emissions to - van der Werf, G. R., Peters, W., van Leeuwen, T. T. and Giglio, L.: What could have caused pre doi:10.5194/cp - 9 - 289 - 2013, 2013. exceed current rates?, Clim. Past, 9(1), 289 – 306, Morton, D. van der Werf, G. R., Randerson, J. T., Giglio, L., van Leeuwen, T. T., Chen, Y., Rogers, B. M., Mu, M., Van Marle, M. J. E., C., Collatz, G. J., Yokelson, R. J. and Kasibhatla, P. S.: Global fire emissions estima 40 - tes during 1997 2016, Earth Syst. Sci. Data, 9(2), 28

29 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © 9 720, doi:10.5194/essd - 2017, 2017. - – 697 - 697 scale experimental testing of fire radiative energy for quantifying mass combusted in natural vegetation fires, - Wooster, M. J.: Small – 24, doi:10.1029/2002GL015487, 2002. Geophys. Res. Lett., 29(21) , 21 Wooster, M. J., Freeborn, P. H., Archibald, S., Oppenheimer, C., Roberts, G. J., Smith, T. E. L., Govender, N., Burton, M. an d Palumbo, path FTIR spectroscopy and fire radiative power I.: Field determination of biomass burning emission ratios and factors via open - 5 assessment: Headfire, backfire and residual smouldering combustion in African savannahs, Atmos. Chem. Phys., 11(22), 11591 11615, – 2011, 2011. - 11 doi:10.5194/acp - - 11591 Yokelson, R. J., Burling, I. R., Gilman, J. B., Warneke, C., Stockwell, C. E., De Gouw, J., Akagi, S. K., Urbanski, S. P., Veres, P., ung, H. and Roberts, J. M., Kuster, W. C., Reardon, J., Griffith, D. W. T., Johnson, T. J., Hosseini, S., Miller, J. W., Cocker, D. R., J eld and laboratory measurements to estimate the emission factors of identified and unidentified trace gases for Weise, D. R.: Coupling fi 10 13 2013, 2013. 89 - - prescribed fires, Atmos. Chem. Phys., 13(1), 89 - – 116, doi:10.5194/acp Yuan, W., Liu, S., Zhou, G., Zhou, G., Tieszen, L. L., Ba ldocchi, D., Bernhofer, C., Gholz, H., Goldstein, A. H., Goulden, M. L., dy covariance Hollinger, D. Y., Hu, Y., Law, B. E., Stoy, P. C., Vesala, T. and Wofsy, S. C.: Deriving a light use efficiency model from ed – 4), 189 – 207, flux data for predicting daily gross primary production across biomes, Agric. For. Meteorol., 143(3 15 doi:10.1016/J.AGRFORMET.2006.12.001, 2007. 29

30 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © - Biome - use efficiency ( LUE ; ! , unitless ) a nd turnover rates ( t , years ) for the stem, level m odel parameter values for light : 1 Table , T leaf, grass, litter and coarse woody debris (cwd) pools. w o additional columns give t he average effective turnover rates (t eff. ° or the 0.25 years ) for the litter and cwd pools, after scaling by the abiotic scalar. Turnover rates that were different resolution f s model calibration are given in parenthese . t t t eff. t Biome t t eff. t ! cwd cwd litt grass leaf stem litt leaf 0.284 60 2 0.5 0.5 - 4 - Everg reen needle leaf 1 0.354 60 broad 0.5 0.5 0.8 4 6.1 Everg reen needle 0.280 60 2 0.5 0.5 - 4 - uous Decid leaf Broad leaf 0.255 35 (60) 0.5 0.5 0.5 1.7 4 - Decid uous 0.283 1 0.5 0.5 0.8 4 6.3 Mixed forest 35 0.5 0.3 0.2 1.2 0.299 4 - Closed shrub land 30 0.3 0.208 0.5 0.3 0.1 30 1 2.6 land Open shrub 0.280 35 (40) 0.5 0.3 (0.5) 0.15 (0.2) 0.5 1 (2) 3.5 Woody savan na 0.4 0.208 0.5 0.3 (0.5) 0.2 5 (10) 1 2.0 Open savan na 18 0.5 0.2 0.1 0.2 1 2.4 Grassland 0.229 2.1 0.5 0.2 0.1 0.2 1 15 Cropland 0.242 5 30

31 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © : Overview of model input datasets. Table 2 patial res. T emporal Product res. Reference Variable S 500 m 8 - daily Myneni et al. ( 2015) fPAR MCD15A2H Annual Dimiceli et al. ( 2015) FTC, NTV MOD44B 250 m 2018) 500 m Monthly Giglio et al. ( MCD64A1 BA 500 m Annual Friedl et al. ( 2010) Land cover MCD12Q1 - ERA Shortwave radiation Monthly Dee et al. ( 2011) Interim ° 0.25 - Inter im ( Air temperature Monthly Dee et al. ERA 2011) 0.25 ° ( Interim Soil moisture ERA - Monthly Dee et al. 2011) ° 0.25 31

32 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © - Saharan Africa as a whole , - and for individual biomes with 1 Area : averaged woody aboveground biomass (AGBw) for sub Figure 500 - meter) and those derived from the reference AGBw dataset by Avitabile et l ed values significant tree cover area. Both model ( th th th th percentiles (boxes), and the 5 are shown. Boxplots show the mean (dots), median (horizontal line), 25 and 95 5 and 75 al. (2016) th th percentiles for Africa and the open savanna biome approach zero 5 due to an abundance of 2 percentiles (whiskers). The 5 and pixels where AGBw is zero . 32

33 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © meter resolution, shown as scatter plot ed FL and field measurements based on calibration for 500 - : 2 Comparison of model Figure l ° resolution simulations. Boxplots show the mean (colored and boxplots per biome for (a, b) 500 5 - meter resolution and (c, d) 0.25 th th and 75 percentiles (boxes), and the range of values (whiskers). The number of measurements dots), median (horizontal line), 25 involved is given below each boxplot. 33

34 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © Total fire emissions for sub Saharan Africa and individual biomes, as compared to GFED4 (without small fires) and : - 3 Figure model calibration. - meter resolution GFED4s (with small fires). Solid orange and blue bars show the estimates based on the 500 ° Transparent bars show the estimates based on the 0.25 the calibration difference is , highlighting that resolution model calibration difference . 5 larger than the simulation 34

35 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ° C (b) for the 500 meter resolution model, aggregated to 0.05 Figure resolution for display. - Fire emissions (a) and F : 4 35

36 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ° Figure ed FL at and field measurements based on calibration for 0.25 : resolution, shown as scatter plot 5 Comparison of model l ° 5 The Boxplot description is resolution simulations. and boxplots per biome for (a, b) 500 - meter resolution and (c, d) 0.25 equivalent to Figure 2. Black triangles depict whiskers outside the plot range. 36

37 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ° over 500 : Relative simulation - meter resolution results, for (a) AGBL and (b) 6 Figure difference as the natural logarithm of 0.25 ° - meter resolution model results are aggregated to 0.25 fire emissions (equivalent to FC). The 500 resolution beforehand. The ° relative difference in emissions is equivalent to the difference in FC. Positive values show 0.25 meter values higher than 500 - values, and vice versa for negative values. 5 37

38 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ° n difference as 0.25 : minus 500 simulatio meter resolution results for (a) total AGBL and fire emissions, and (b) Absolute 7 Figure - th th and 75 - average AGBL and FC (per area burned). Boxplots show the mean (colored dots), median (horizontal line), 25 area th th percentiles (whiskers). and 95 5 percentiles (boxes), and the 5 38

39 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © ° Relative differences in AGBL for various isolated mechanisms, as the natural logarithm of 0.25 model result : Figure simulation 8 specific parameter error, (b) the additional - - over 500 meter model result. (a) shows the isolated difference due to the biome difference due to the input aggregation error, (c) the remaining difference related to fire processes, and (d) the re maining ° , ( b ) and ( c ) . The 500 - meter resolution model results are aggregated to 0.25 unexplained difference after subtraction of 5 ( a ) resolution. 39

40 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © Relative differences in fire emissions (and FC) for various iso lated mechanisms, as the natural logarithm of : Figure 9 simulation ° meter model result. (a) shows the isolated difference due to the biome - specific parameter error, (b) the model result over 500 0.25 - additional difference due to the input aggregation error, (c) the difference accounted for by implementation of the modified ) , ( b ) and ( c ) . The 500 - meter 5 burned fraction (MBF), and (d) the remaining unexplained difference after subtraction of ( a ° resolution. resolution model results are aggregated to 0.25 40

41 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © Saharan Africa and for individual biomes, as the : difference in AGBL and fire emissions for sub - simulation Relative 10 Figure ° - meter model result. Stacked bars depict the contribution of var natural logarithm of 0.25 ious error model result over 500 specific - mechanisms to the overall resolution difference in AGBL and fire emissions, by successive isolation of the biome parameter error, the input aggregation error, the MBF and finally the remaining difference related to fire (Fire rest). 5 Fur thermore , the unexplained remainder after removal of all identified mechanisms is shown (Unexplained). 41

42 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © meter, for (a) AGBL and (b) fire emissions, ver 11 difference per biome compared to 500 - Average absolute sus : Figure simulation spatial resolution. Data points are fitted with a natural logarithmic function where applicable. Inserted formulas show the d a 5 corresponding fit functions for the two outermost lines. The plot axes are linear. For this plot, 500 - meter resolution is assume ° of 0.005 resolution in degrees . 42

43 Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-116 Manuscript under review for journal Geosci. Model Dev. Discussion started: 2 May 2019 c Author(s) 2019. CC BY 4.0 License. © Typical stem biomass growth curve. Figure 12 : 43

Related documents

CityNT2019TentRoll 1

CityNT2019TentRoll 1

STATE OF NEW YORK 2 0 1 9 T E N T A T I V E A S S E S S M E N T R O L L PAGE 1 VALUATION DATE-JUL 01, 2018 COUNTY - Niagara T A X A B L E SECTION OF THE ROLL - 1 CITY - North Tonawanda TAX MAP NUMBER ...

More info »
TPM Main Part 3 Commands v1.2 rev116 01032011

TPM Main Part 3 Commands v1.2 rev116 01032011

1 2 1 2 3 4 5 6 7 8 TPM Main Part 3 Commands 9 10 Specification Version 1.2 11 Level 2 Revision 116 12 1 March 2011 13 TCG Published 14 15 16 17 Contact: [email protected] 18 19 20 21 22...

More info »
ECI Historical Listing    Volume III   March 2019

ECI Historical Listing Volume III March 2019

Employment Cost Index National Compensation Survey – Volume III Historical Listing April www.bls.gov/ect 2019 (December 2005=100) 2019 -March Current Dollar, March 2001 The estimates from 2001 to 2005...

More info »
TPM Main Part 2 TPM Structures v1.2 rev116 01032011

TPM Main Part 2 TPM Structures v1.2 rev116 01032011

TPM Main Part 2 TPM Structures Specification version 1.2 Level 2 Revision 1 1 6 1 1 1 March 20 TCG Published Contact: [email protected] TCG Published Copyright © 2003 2011 Trusted Comput...

More info »
WEF GlobalInformationTechnology Report 2014

WEF GlobalInformationTechnology Report 2014

Insight Report The Global Information Technology Report 2014 Rewards and Risks of Big Data Beñat Bilbao-Osorio, Soumitra Dutta, and Bruno Lanvin, Editors

More info »
Fourth National Report on Human Exposure to Environmental Chemicals Update

Fourth National Report on Human Exposure to Environmental Chemicals Update

201 8 Fourth National Report on Human Exposure to Environmental Chemicals U pdated Tables, March 2018 , Volume One

More info »
JO 7400.11C   Airspace Designations and Reporting Points

JO 7400.11C Airspace Designations and Reporting Points

U.S. DEPARTMENT OF TRANSPORTATION ORDER FEDERAL AVIATION ADMINISTRATION 7400.11C JO Air Traffic Organization Policy August 13, 2018 SUBJ: Airspace Designations and Reporting Points . This O rder, publ...

More info »
SARA R4/N4 series

SARA R4/N4 series

R4/N4 series SARA - System Integration Manual System Integration Manual - SARA R4/N4 Abstract This document - R4/N4 series describes the features and the integration of the size - optimized SARA cellu...

More info »
COMM 750 Network S Directory   508

COMM 750 Network S Directory 508

SM Provider Directory – Blue Network S UPDATED SEPTEMBER 2016

More info »
FY 2019 PB Green Book

FY 2019 PB Green Book

NATIONAL BUDGET DEFENSE ESTIMATES FOR FY 201 9 OFFICE OF THE UNDER SECRE TA RY OF DEFENSE (COMPTROLLER) APRIL 201 8

More info »
FedEx Service Guide

FedEx Service Guide

SERVICE GUIDE fedex.com/serviceguide January 7, 2019 Updated April 29, 2019 SERVICES R AT ES TERMS INDEX

More info »
Second National Report on Biochemical Indicators of Diet and Nutrition in the U.S. Population

Second National Report on Biochemical Indicators of Diet and Nutrition in the U.S. Population

Second National Report on Biochemical Indicators of Diet and Nutrition in the U.S. Population Second National Report on Biochemical Indicators of Diet and Nutrition in the U.S. Population 2012 Nationa...

More info »
Annual Energy Review 2011   Released September 2012

Annual Energy Review 2011 Released September 2012

D / E I A - 0 3 8 4 ( 2 0 1 1 ) E | S e p t e m b e r 2 0 1 2 O 1 1 0 2 w e i v e R y g r A n n u a l E n e . r a / v o g e a i e . w w w

More info »
PPI Table 9. Commodity and service groupings and individual items

PPI Table 9. Commodity and service groupings and individual items

Table 9. Producer price indexes and percent changes for commodity and service groupings and individual items, not seasonally adjusted [1982=100 unless otherwise indicated] Percent change to Apr. Other...

More info »
C:\Users\cbenson\AppData\Roaming\SoftQuad\XMetaL\7.0\gen\c\h803 enr.xml

C:\Users\cbenson\AppData\Roaming\SoftQuad\XMetaL\7.0\gen\c\h803 enr.xml

H. R. 803 One Hundred Thirteenth Congress of the United States of America AT THE SECOND SESSION Begun and held at the City of Washington on Friday, the third day of January, two thousand and fourteen ...

More info »
cat51a metric

cat51a metric

Wheaton Facility Aabybro Facility Springeld Facility P.O. Box 7900 Mølhavevej 2 1801 Business Park Drive Wheaton, Illinois DK 9440 Aabybro Springeld, Illinois 60187-7901 USA Denmark 62703 USA www.te...

More info »
Monthly Energy Review – April 2019

Monthly Energy Review – April 2019

DOE/EIA ‐ 0035( 2019/4 ) April 2019 Monthly Energy Review www.eia.gov/mer

More info »
RIE Tenant List By Docket Number

RIE Tenant List By Docket Number

SCRIE TENANTS LIST ~ By Docket Number ~ Borough of Bronx SCRIE in the last year; it includes tenants that have a lease expiration date equal or who have received • This report displays information on ...

More info »