INFERNO : a fire and emissions scheme for the UK Met Office ’ s Unified Model

Warm and dry climatological conditions favour the occurrence of forest fires. These fires then become a significant emission source to the atmosphere. Despite this global importance, fires are a local phenomenon and are difficult to represent in large-scale Earth system models (ESMs). To address this, the INteractive Fire and Emission algoRithm for Natural envirOnments (INFERNO) was developed. INFERNO follows a reduced complexity approach and is intended for decadalto centennial-scale climate simulations and assessment models for policy making. Fuel flammability is simulated using temperature, relative humidity (RH) and fuel load as well as precipitation and soil moisture. Combining flammability with ignitions and vegetation, the burnt area is diagnosed. Emissions of carbon and key species are estimated using the carbon scheme in the Joint UK Land Environment Simulator (JULES) land surface model. JULES also possesses fire index diagnostics, which we document and compare with our fire scheme. We found INFERNO captured global burnt area variability better than individual indices, and these performed best for their native regions. Two meteorology data sets and three ignition modes are used to validate the model. INFERNO is shown to effectively diagnose global fire occurrence (R = 0.66) and emissions (R = 0.59) through an approach appropriate to the complexity of an ESM, although regional biases remain.


Introduction
Fire is a key interaction between the atmosphere and the land surface (Bowman et al., 2009).Its impacts are wide-ranging: it influences forest succession (Bond and Keeley, 2005), is a tool for deforestation (van der Werf et al., 2009) and an important natural carbon source (Bowman et al., 2013), while it also provides a major natural hazard to humans through property and infrastructure destruction and air quality degradation (Johnston et al., 2012;Marlier et al., 2013).Biomass burning emissions are not only substantial in magnitude (Lamarque et al., 2010), but also drive the variability of atmospheric composition (Spracklen et al., 2007;Voulgarakis et al., 2010Voulgarakis et al., , 2015) ) and impact short-term climate forcing (Tosca et al., 2013).
There are feedbacks between fire and climate: lowhumidity conditions cause droughts, which enhance fire activity (Field et al., 2009), which, in turn, emits aerosols and trace gases (Akagi et al., 2011), influencing the abundances of radiatively active atmospheric constituents, cloud formation and lifetime, and in turn precipitation, and surface albedo (Voulgarakis and Field, 2015).Bistinas et al. (2014) showed global fire frequency is correlated with land-use, vegetation type and meteorological factors (dry days, soil moisture and maximum temperature) and that human presence tends to noticeably reduce fire activity (landmanagement, landscape fragmentation and urbanisation).
Examining and quantifying such impacts and feedbacks is paramount to Earth system models (ESMs), yet to integrate vegetation fires presents many challenges as it intricately links multiple disciplines from ecology to atmospheric chemistry, physics and climate science.
Here, we present and evaluate the INteractive Fire and Emission algoRithm for Natural envirOnments (INFERNO) and its implementation.INFERNO is a necessarily simple parameterization that focuses on the large-scale occurrence of fires and is suitable for ESM application.The model uses a few key driving variables while retaining a broadly accurate parameterization for fire emissions.INFERNO's performance against observations and well-established and operationally relevant fire indices is presented.
2 Model description

INFERNO
INFERNO was constructed upon the simplified parameterization for fire counts proposed and evaluated for the presentday by Pechony and Shindell (2009), which was subsequently shown to provide a good estimate for large-scale fire variability over climatological timescales (Pechony and Shindell, 2010).In short, that parameterization uses monthly mean temperature, relative humidity (RH) and precipitation to simulate fuel flammability.It also uses human population density and lightning to represent ignitions.To incorporate this parameterization within the Joint UK Land Environment Simulator (JULES; Best et al., 2011;Clark et al., 2011), several changes were applied.Upper layer soil moisture is used to represent precipitation memory while precipitation acts as a rapid fire deterrent.Vegetation density was replaced by fuel load index, dependent on leaf carbon and decomposable plant material (DPM), i.e. litter.Such a relationship with fine fuel and moisture was used in Thonicke et al. (2001).Furthermore, we developed a parameterization to obtain burnt area (BA), emitted carbon (EC) and fire emissions of different species (E X ), and our fire diagnostics are made for each of the nine plant functional types (PFTs) in the current version of JULES (Harper et al., 2016).
Figure 1  1.67 ignitions km −2 month −1 .This corresponds to 1.5 ignitions km −2 month −1 due to humans (I A ), heuristically determined, and 0.17 ignitions km −2 month −1 natural ignitions due to lightning (I N ), derived from the multi-year annual mean of 2.7 strikes km −2 year −1 (Huntrieser et al., 2008) assuming that 75 % of strikes are cloud-to-ground (Prentice and Mackerras, 1977).This mode inherently suppresses the variability in fires due to any anthropogenic or natural ignition changes (Pechony andShindell, 2009, 2010).Second, human ignitions and suppressions can be assumed to remain constant at the global mean value mentioned above (I A = 1.5 ignitions km −2 month −1 ); however cloudto-ground lightning strikes may vary and, in addition, each strike is assumed to start a fire.This mode accounts for natural variability in fire ignitions, which can be simulated within an ESM, or prescribed from observations.Third, varying human ignitions and suppressions and varying natural ignitions (cloud-to-ground lightning strikes, as in mode 2).This was the original ignition approach in Pechony and Shindell (2009), which was left unchanged and is detailed below.In this ignition mode, anthropogenic ignition and suppression depends on population density (PD), as proposed by Venevsky et al. (2002).
PD is in units of people km −2 , and k (PD) = 6.8 × PD −0.6 is a function that represents the varying anthropogenic influence on ignitions in rural vs. urban environments.The parameter α = 0.03 represents the number of potential ignition sources per person per month per km 2 .Both natural and anthropogenic ignitions have the potential to be suppressed by humans, such that the fraction of fires not suppressed is f NS = 7.7 (0.05 + 0.9 × e −0.05PD ).

Flammability (F )
We adapt the Pechony and Shindell (2009) scheme for flammability to function interactively within an ESM (see Eq. 6).Starting from the saturation vapour pressure (e * , Eq. 4; Goff and Gratch, 1946) and its temperature dependence, we introduce a fuel load index (FL PFT , Eq. 5) as well as RH, precipitation and soil moisture in order to obtain flammability (Eq.6).The land surface model (JULES) determines soil moisture content (θ ) and fuel load (DPM C and Leaf C,PFT ).
(5) Equation ( 5) shows FL PFT is taken as the PFT-specific leaf carbon (Leaf C,PFT , aboveground) plus the carbon within DPM (DPM C ). DPM is a soil carbon pool of which we assume 70 % is available to fires, i.e. near surface (DPM is shared across all PFTs).FL scales linearly between 0 (at a threshold of Fuel low = 0.02 kgC m −2 ) and 1 (at a threshold of Fuel high = 0.2 kgC m −2 ).Similar approaches to represent fuel availability within fire parameterizations have commonly been adopted (Arora and Boer, 2005;Li et al., 2012;Thonicke et al., 2010).
for RH low ≤ RH ≤ RH high 0 for RH low > RH (6) RH is the relative humidity (%) and R is the precipitation rate (mm day −1 ).The influence of RH scales between (and is bound by) 0 (at a threshold of RH low = 10 %) and 1 (at a threshold of RH up = 90 %).We then adapt the formula by replacing a vegetation index dependent on leaf area index (LAI) with the fuel load index (FL).Finally, flammability (F PFT ) is dependent on upper-level (down to 0.1 m) soil moisture: θ is the unfrozen soil moisture as a fraction of saturation.The individual importance of these variables to our model is illustrated in Fig. A1.

Burnt area (BA)
Our approach is to associate an average burnt area per fire to each PFT, effectively decoupling the fire-spread stage from local meteorology and topography, which is typically not resolved in the relatively coarse grid of an ESM.An average burnt area (BA PFT ) was heuristically determined for each PFT: 0.6, 1.4 and 1.2 km 2 for trees, grass and shrubs, respectively, such that grass and shrubs will fuel larger fires than trees.Sub-categories of trees, grass and shrubs are not differentiated.Observational evidence supports that the land cover type is an efficient way to characterise fires, which tend to be larger in grasslands than in forests (Chuvieco et al., 2008;Giglio et al., 2013).The BA is then calculated following Eq.( 7): Here, BA PFT is the burnt area (fraction of PFT cover burnt per second) for each PFT; meanwhile the number of ignitions times the flammability (I T F PFT ) represents the number of fires.
Inferring burnt area from number of fires in this manner stands out from other fire models that utilise wind speed (Arora and Boer, 2005;Thonicke et al., 2010;Li et al., 2012), effectively modelling the fire rate of spread.Wind is key to the modelling of individual fires; yet implementing wind effectively within fire models designed for the relatively coarse grid of ESMs was found to be problematic (Lasslop et al., 2014(Lasslop et al., , 2015)).Conversely, Hantson et al. (2014) found global fire size was mostly influenced by precipitation, aridity and human activity (population density and croplands).

Emitted carbon (EC)
To account for the wetness of fuel in INFERNO, combustion completeness (the fraction of biomass exposed to a fire that was volatised) scales linearly with soil moisture (as a fraction Equation ( 8) shows how the PFT-specific EC (in kgC m −2 s −1 ) is computed.BA is the burnt area (fraction s −1 ), CC min and CC max are the minimum and maximum combustion completeness for both leaves (CC min = 0.8 and CC max = 1.0) and stems (CC min = 0.0 and CC max = 0.4) and C i is the carbon stored in each PFT's leaves or stems (kgC m −2 ).The parameters used for combustion completeness (CC min and CC max ) are similar to the Global Fire Emission Database (GFED; van der Werf et al., 2010), albeit with lower minimum combustion of stems (0.0 as opposed to 0.2).Nevertheless, GFED uses a more complex representation of moisture across multiple fuel types and only accounts for fires that were observed.In comparison, our scheme only relies on soil moisture and was much more sensitive to minimum combustion, such that the contribution from moist forested areas (e.g.rainforests) needed to be reduced by increasing the impact of soil moisture (reducing stems' CC min ).

Emitted species (E X )
There has been a significant amount of work on estimating emission factors (EFs) across fire biomes (such as savannahs, boreal forest, etc.).This was synthesised in Akagi et al. (2011) as well as Andreae and Merlet (2001) and its updates.Updated EFs for Akagi et al. (2011) were not used in this version of INFERNO.To convert biome-specific EFs to PFT-specific EFs, each PFT was linked to a fire biome (see Table A1).INFERNO uses these to estimate emissions (Eq.9).
Here, E X is the amount of species X emitted by fires (in kg m −2 s −1 ), EC is the emitted carbon (in kgC m −2 s −1 ), EF X is the PFT-specific emission factor (see Table 1) (in kg of species emitted per kg of biomass burnt) and [C] is the dry biomass carbon content, which we assume as 50 % (a common simplification; Lamlom and Savidge, 2003).INFERNO currently provides emissions for basic trace gases: CO 2 , CO, CH 4 , NO x , SO 2 and aerosols: organic carbon (OC) and black carbon (BC).

Implementation within JULES
INFERNO is currently implemented within the JULES (Best et al., 2011;Clark et al., 2011) its carbon fluxes and vegetation dynamics.The results shown here used JULES v4.3.1 and INFERNO will be included in JULES from version 4.5 onwards.INFERNO utilises soil moisture (see Eqs. 6, 8), which JULES calculates as the balance between precipitation (following the scheme for rainfall interception in Dolman and Gregory, 1992) and extraction by evapotranspiration and runoff (Cox et al., 1999;Best et al., 2011).JULES has four soil layers, and INFERNO uses the top layer unfrozen soil moisture (0 to 0.1 m depth).Note that in its current state, JULES does not associate carbon pools with depths; hence, it is not possible, for example, to access only the top-most DPM.The vegetation dynamics and litter carbon used obey the TRIFFID DGVM (Cox, 2001).Fractional coverage of PFTs in any grid cell is based on competition for resources (light and water), governed by Lotka-Volterra competition equations and based on a tree-shrub-grass dominance hierarchy (Cox, 2001).
In JULES, vegetation carbon content is determined by the balance between photosynthesis, respiration and litterfall.Within JULES, TRIFFID (the Top-down Representation of Foliage and Flora Including Dynamics; Cox, 2001) predicts changes in biomass and the fractional coverage of nine PFTs (Table A1) based on accumulated carbon fluxes and heightbased competition, where the tallest trees have first access to space (Harper et al., 2016).Vegetation can grow in height, and the carbon in leaves, roots, and wood is related allometrically to the "balanced LAI", L b (Cox, 2001).L b is the seasonal maximum LAI and a function of plant height.Within INFERNO, leaf carbon (Leaf C , used for calculating FD and emissions) is Meanwhile, wood carbon (Wood C , which affects emissions), is calculated as PFT dependent parameters (σ l , the specific leaf density, a wl , the allometric coefficient and b wl , the allometric exponent) are given in Table A1.When using JULES in its standalone version, IN-FERNO can use inputs of population density (in people km −2 ) and cloud-to-ground lightning flash rates (in flashes km −2 month −1 ) from ancillary data sets.Interestingly, lightning can be interactively simulated in atmospheric models (not population), although this will not be explored in this paper.Similarly, meteorology needs to be prescribed and is then interpolated from its native temporal resolution to the model's time step.Although designed to be integrated within an ESM, the capability to run INFERNO only with JULES is particularly useful for present-day comparison with observations, and to dissociate causes of biases in results.In its current early state, INFERNO provides a diagnostic tool, it does not remove carbon from vegetation nor does it lead to tree mortality.Akagi et al. (2011), such that each PFT was attributed a fire biome (see Table A1).This method of attributing emission factors to PFTs is similar to that presented in Thonicke et al. (2010), and can be extended to include all species of trace gases and aerosols compiled in Akagi et al. (2011).

Fire weather indices
Three other well-established daily fire indices are also available within JULES.These indices have been used for several decades to help plan operational response to wildfires on numerical weather predictions (NWP) timescales.Although unit-less and ill-defined risk-based quantities, comparison to INFERNO is still useful for understanding the results in the context of practically established metrics.The Canadian Fire Weather Index (Forestry Canada, 1992;Van Wagner and Pickett, 1985) consists of six components, calculated from basic meteorological parameters.Three are fuel moisture codes designed to represent the drying of different fuel types, their characteristics are displayed in Table A2.Two intermediate quantities, the Initial Spread Index and the build-up index are calculated from these, and are in turn used to yield the final Fire Weather Index (FWI): FWI = e 2.72(0.434ln B) 0.647 , B > 1 where B = 0.1 ISI FD with ISI (Initial Spread Index) and FD the fuel density.We refer to the original publications for detailed equations for the complex Canadian FWI and each of its components.The McArthur Forest Fire Danger Index (FFDI; Noble et al., 1980;Sirakoff, 1985) was developed for use in Australia.Simpler in its formulation than the Canadian index, it consists of a drought component modified by the local temperature, humidity and wind speed.The calculation of the drought component depends on the soil moisture deficit (the amount of water needed to restore the soil moisture content of the top 800 mm of soil to 200 mm), which is related to the JULES soil moisture.
The FFDI (F McArthur ) is given by where T is the daily maximum temperature, H the daily minimum RH and W the daily mean wind speed.D is the drought factor, given by where N is the number of days since the last rain, R the total rain in the most recent day with rain and I the amount of rain needed to restore the soil moisture content to 200 mm in the top 800 mm of soil.Finally, the Nesterov index (Nesterov, 1949) is the simplest fire index implemented in JULES.It uses only the daily mean temperature, mean daily dew point (or suitable substitute), daily total precipitation and the previous day's index.The index is incremented daily, unless daily precipitation exceeds 3 mm, in which case it is reset: where T is the mean daily temperature, D the mean daily dew point, P the daily total precipitation and N 0 the previous day's index.The Nesterov index is a key component for other fire models (Venevsky et al., 2002;Thonicke et al., 2010).

Model configuration
Monthly lightning data were obtained from LIS-OTD (Lightning Imaging Sensor -Optical Transient Detector) observations for 2013 (Christian et al., 2003) and was recycled for every year in the simulation.These detections were converted to cloud-to-ground strikes using the relationship presented in Prentice and Mackerras (1977).Land use and population density were obtained from the HYDE data set (Hurtt et al., 2011) and then linearly interpolated to create inter-annually varying data.Finally, annual CO 2 concentrations, which affect vegetation dynamics, were prescribed as a global aver- age following the data set prepared for the global carbon budget (Le Quéré et al., 2015).
To test the sensitivity to the meteorological input, JULES simulations were driven by meteorology from both CRU-NCEP (Climate Research Unit and National Center for Environmental Prediction) v5 (https://crudata.uea.ac.uk/cru/data/ ncep/), and WFDEI (Weedon et al., 2014) with precipitation from the GPCC (Schneider et al., 2013).Both data sets were used on a 6-hourly basis.
Outside of these driving variables, JULES was configured according to the TRENDY project (Sitch et al., 2015;Peng et al., 2015); 100-year spin-up was performed repeating the 1990-2000 conditions 10-fold.Four configurations were used to create simulations covering 1990-2013, although to validate INFERNO only the 1997-2010 period was analysed.The first three use CRU-NCEP meteorology with each of our three ignitions modes (see Sect. 2.1.1);constant ignitions (mode 1), prescribed lightning and constant anthropogenic ignitions (mode 2), and both natural and anthropogenic ignitions varying with prescribed lightning and population density (mode 3).The fourth simulation assumes mode 1 (constant ignitions), while meteorology is prescribed from WFDEI and precipitation from GPCC.
Evaluation was performed against the published data for GFEDv3, FINNv1, Global Fire Assimilation System (GFAS) and GFEDv4.We also used the data from GFEDv4s (http: //globalfiredata.org,manuscript in preparation) and GFEDv4 (Giglio et al., 2013) to calculate grid-specific emissions and burnt area.The GFED passes satellite observation of burnt area through the Carnegie-Ames-Stanford Approach (CASA) biogeochemical model in order to obtain emissions from open burning.GFEDv4 (Giglio et al., 2013) innovates on GFEDv3 (Giglio et al., 2010) mainly through an updated algorithm to retrieve burnt area from MODIS satellite products and an increased spatial and temporal resolution, to 0.25 • and daily resolution (this resolution was assessed in Mangeon et al., 2015).Meanwhile GFEDv4s also includes the contribution from small fires (Randerson et al., 2012).The Fire Inventory from NCAR version 1.0 (FINNv1; Wiedinmyer et al., 2011) provides high-resolution (both temporal and spatial) global emissions of trace gas and particle emissions from open burning of biomass.It focuses on rapid availability and assimilation in real-time forecast and follows a similar process to GFED to estimate emission, but its burnt area is obtained directly from fire pixel using land cover (Wiedinmyer et al., 2011).The GFAS (Kaiser et al., 2012), unlike the aforementioned products, directly assess emissions from satellite-observed fire radiative power more apt at detecting small fires and avoiding the uncertainty of biogeochemical models.

Results
Maps of the burnt area and emitted carbon are displayed in Fig. 2, their resolution is 192 longitude by 145 latitude grid cells (1.875 • × 1.24 • ).The results from INFERNO used a configuration with CRU-NCEP meteorology and the third ignition mode: interactive lighting and anthropogenic ignitions.We compare our results with downscaled means from GFED.INFERNO accurately diagnoses total fire occurrence and emissions over the 1997-2010 period: we found a spatial correlation of R = 0.66 for burnt area and R = 0.59 for emitted carbon, both passing the t test with 95 % significance.In addition, regional mean yearly budgets are compared with GFED in Table B1.Compared to GFEDv4, we notice INFERNO estimates higher burnt area in all regions apart from Australia and New Zealand, and southern hemisphere Africa.Meanwhile emitted carbon is underestimated in boreal regions and equatorial Asia, but overestimated in most other regions (significantly in Southern Hemisphere America).Over the studied period, C4 grass were the main contributors to burnt area in INFERNO (a mean 2.34 million km 2 year −1 ), meanwhile broadleaf evergreen trees (tropical) led to the most emitted carbon (a mean 1.48 Pg year −1 ).GFEDv4 projects the grid-box with maximum burnt area within the Central African Republic (87 % of grid fraction burnt per year), while INFERNO finds a maximum burnt area of 57 %, slightly to the north (south-east of Lake Tchad).The discrepancy is much larger for emissions, with a maximum emitted carbon of 1.47 kg m −2 year −1 in Indonesia predicted by GFEDv4s, against 0.4 kg m −2 year −1 for INFERNO, in Angola.These results could be expected, as INFERNO focuses on capturing global biomass burning, it will not represent such extremes of burning; furthermore, the immense emitted carbon observed in Indonesia follows from undiagnosed peat fires.INFERNO's approach to burnt area only considers trees, grass and shrub cover and was determined heuristically; meanwhile, Hantson et al. (2014) found global fire size was mostly influenced by precipitation, aridity and human activity (population density and croplands).Further parameterizations for fire size exist (e.g.Hantson et al., 2015Hantson et al., , 2016)), which could improve INFERNO burnt area estimates while maintaining simplicity and traceability.
Figure 3 shows the modelled global annual average biomass burning emissions and burnt area from 1997 to 2010.The three ignition methods are evaluated: fully interactive ignitions (red) predict the highest carbon emissions while interactive lightning with constant human ignitions (blue) the lowest.WFDEI was observed to lead to more biomass burning emissions in tropical forest areas (and in particular the borders of rainforests), while CRU-NCEP favoured burning in near-desert areas (the Sahel, India and south American grasslands).We expect this result to be significantly influenced by differences in precipitation (GPCC for WFDEI runs and CRU for CRU-NCEP; Schneider et al., 2013).
Comparisons to FINNv1, GFEDv4, GFASv1 and GFEDv3 were restricted to their budgets published in Kaiser et www.geosci-model-dev.net/9/2685/2016/Geosci.Model Dev., 9, 2685-2700, 2016 Biomass burning emissions and burnt area simulated by the model follow similar trends to GFEDv3, although with a smaller inter-annual variability in the model.Carbon emissions from all simulations fall within 1 standard deviation of GFEDv3, apart from 3 years: 1997, 1998 and 2001.Note that for these years, emissions in GFED were obtained from the lower-resolution AVHRR rather than MODIS.1997 and 1998 were strong El Niño years during which droughts in equatorial Asia led to extreme emissions from land-clearing fires, a recurrent problem in the region (Field et al., 2009).Indeed in 1997, in the region contained between 20 • S-20 • N and 90-160 • E (or equatorial Asia), GFEDv3 estimate emissions of 1.07 PgC, while INFERNO (with CRU-NCEP and fully interactive ignitions) estimates 0.15 PgC.Unfortunately, peat is not modelled in JULES and thus neither is peat present in our fire scheme.It was estimated tropical peat fires alone produced an average of 0.1 PgC per year from 1997 to 2009, and 0.7 PgC in 1997 in particular (van der Werf et al., 2010).Furthermore, 2002 and 2006 also saw important peat burning, with GFEDv3 estimating peat emissions of 0.16 and 0.21 PgC respectively.In both of these years, the trend in INFERNO differs from GFEDv3's (stagnation in 2002 anddecrease in 2006).Peat-lands can be significant in equatorial Asia but also boreal regions where their combustion leads to the release of long-stored carbon (Turetsky et al., 2015).In 1998 and 2001, the difference in emissions could not be attributed to a particular location.While fire emissions from Equatorial Asia were underestimated, GFEDv3 observed lower emissions over Africa compared to INFERNO, which seems to be the key driver of our discrepancies.
Table 2 shows the budgets for four latitudinal bands across the various simulations performed.The second ignition mode (constant anthropogenic and interactive lightning ignitions at any time and place) appears to consistently predict lower emissions and burnt area (with the exception of low latitudes).Furthermore, the main impact of using an ignition model that varies with both natural and anthropogenic ignitions is a reduction of fires at low (tropical and sub-tropical) latitudes, and an increase in equatorial regions.Indeed, when compared to constant ignitions (mode 1), interactive ignitions (mode 3) predict more emissions in forest encroachment regions (noticeably surrounding the Congo and Amazon rainforests), and less in heavily populated areas (Nigeria, India).Meanwhile, we observed interactive lightning ignitions (mode 2) significantly reduced burning in grasslandsavannah environments.We link this to the predominance of cloud-to-ground lightning strikes in a wet environment within the LIS-OTD data set (e.g. the Congo rainforest; Christian et al., 2003) and fewer strikes (and ignitions) in the more flammable grasslands and savannahs.These issues are visible in Fig. B1, which shows difference maps of the four model configurations, for 1997-2010 mean yearly totals.Equatorial and boreal regions include peat that leads to large fuel consumption, which is unaccounted for in JULES, suggesting that our model will inherently underestimate emissions from these regions.
Species-specific average emissions produced by the IN-FERNO scheme are shown in Table 3 in Tg per year for the 1997-2010 period.CO and CH 4 appear to be produced in noticeably larger quantities than in observation-based emission estimates.This hints at an overrepresentation of smouldering-type combustion.In INFERNO this might be due to the emission factors used, or the type of vegetation burnt.
In order to examine whether our flammability can represent fire occurrence, three other fire indices were diagnosed, Table 3.Average annual emission (Tg year −1 ) for INFERNO with the interactive ignition mode and CRU-NCEP reanalysis (3 -CRU-NCEP) and the constant ignition mode and WFDEI reanalysis (1 -WFDEI), comparison to GFASv1 (Kaiser et al., 2012), GFEDv3 (van der Werf et al., 2010) and FINNv1 (Wiedinmyer et al., 2011) (1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) shown for four latitudinal bands.R coefficients were obtained between either of the three simulated fire indices or INFERNO's burnt area (ubiquitous ignitions -ignition mode 1, using CRU-NCEP meteorology) and burnt area from GFEDv4 (Giglio et al., 2013).Italics mean the correlation was not significant (p value above 0.05).We restrict our analysis to grid-boxes in which GFEDv4 observed burning.Latitudes were bound to: beyond 50 These indices were obtained seamlessly during the model runs, therefore utilizing the same meteorological and hydrological driving variables, and the same vegetation conditions.Their predictions were regressed with GFEDv4 1997-2010 annual burnt area (Giglio et al., 2013).This analysis relies on the assumption that fire indices can be used as a proxy for the variability of fire occurrence and spread, and eventually of burnt area (not the magnitude).Only areas that had been observed to burn sometime between 1997 and 2010 were sampled; to avoid accounting for high fire indices in non-vegetated areas such as the Sahara.
Table 4 shows the result of our analysis.Ignitions followed mode 1: in this mode ignitions are constant; therefore, the only variability in burnt area (and performance) is due to INFERNO's flammability scheme.The McArthur index performs poorly at high latitudes (it was made for Australia), but outperforms the other indices in low latitude regions.The Canadian and Nesterov indices correlate best with observed burnt area in high latitude regions (for which they were developed).Altogether, INFERNO's burnt area appears to follow observed burnt area better than the sole usage of a fire index.

Conclusion
Through a minimalistic approach we propose a parameterization for fire occurrence of appropriate complexity for application at large spatial scales within an ESM context: the INteractive Fire and Emission algoRithm for Natural envi-rOnments (INFERNO).It directly only varies according to precipitation (and resulting soil moisture), temperature and humidity, and indirectly it utilises vegetation.It is also capable of explicitly simulating ignitions using lightning and anthropogenic information.While our scheme manages to represent fire occurrence on large scales (both spatial and temporal), it performs best at low latitudes.INFERNO's burnt area scheme appears superior to the use of fire indices alone (Nesterov, McArthur and basic Canadian) for capturing annual burnt area variations, and thus one form of fire impact.However, due to the nature of our analysis (fire danger and burnt area remain different quantities) this does not imply INFERNO should supersede fire weather indices for operational purposes, neither has our algorithm been built for numerical weather prediction or seasonal fire danger forecasting.
Nonetheless, our current simulations suggest the variability in emissions is underestimated by INFERNO, in particular the impact of the 1997-1998 El Niño and the subsequent La Niña, which may be attributable to the lack of representation of peat in the model, critical to biomass burn-ing in equatorial Asia and boreal areas.The use of different present-day meteorological data sets has an important impact on the magnitude and variability of our diagnostics.Using WFDEI-GPCC rather than CRU-NCEP led to more burnt area but lower fuel consumption and eventually less emitted carbon (this follows from grasslands burning rather than forests).Vegetation zone interfaces were key to this difference.Similarly, lightning appears to more frequently ignite fires in wet environments (rainforests) while flammable environments (savannah, grasslands) with rarer lightning are sensitive to the presence of an anthropogenic ignition source.Including a scheme to parameterise human impacts appears to significantly reduce fires in heavily populated areas, while favouring their encroachment of rainforests (the vicinity of which are an anthropogenic ignition "sweet spot" in our parameterization).Nevertheless, there is much uncertainty attributed to human induced emissions and effects on fire regime (Marlon et al., 2008;Thonicke et al., 2010).Accordingly, we include different modes to examine the impact of ignitions (human or natural) in INFERNO.
The implementation of INFERNO within the Met Office's Unified Model and its significance for present-day atmospheric composition and climate will be investigated in a separate paper.To close the vegetation-fire feedback, INFERNO will eventually need to remove carbon from vegetation and to include tree mortality.While a strength of the model is its minimalistic approach, the scheme holds potential for im-provements.For instance, litter influences flammability but only live vegetation leads to emissions while in reality litter significantly contributes to observed fuel consumption (van Leeuwen et al., 2014).Similarly, we predict that the inclusion of peat within JULES would improve its fire diagnostics, especially for locations with large fuel consumptions (e.g.equatorial Asia and boreal climates; van der Werf et al., 2010).Given the predictability of emissions from peat fires in relation to precipitation (van der Werf et al., 2008), this would be a promising area of exploration.The value of this model being its simplicity and linearity, any improvements to INFERNO should follow this vision; complex parameterizations are better suited for process-based fire schemes (e.g.Lasslop et al., 2014;Li et al., 2013, p. 1).

Code availability
Information on the JULES land surface model can be found at http://jules-lsm.github.io/.INFERNO is included in JULES v4.5 and is included in this documentation.The JULES source code can be accessed via the Met Office's science repository (requires registration): https://code.metoffice.gov.uk/trac/jules.In particular, the version of the code used to produce the outputs included in this study can be accessed at https://code.metoffice.gov.uk/trac/jules/browser/main/branches/dev/stephanemangeon/vn4.3.1_inferno.Table A1.The key JULES PFT-specific parameters for allometry and vegetation carbon used in our simulations (Clark et al., 2011).This appendix contains additional results illustrating the dependence of INFERNO with ignitions and its performance on a regional basis.

Figure 1 .
Figure 1.Schematic summarising the INteractive Fire and Emission algoRithm for Natural envirOnments (INFERNO) and its key components and behaviour.Ignitions can be accounted for in a variety of ways (see Sect. 2.1.1),meteorology influences flammability (see Sect. 2.1.2),while plant coverage influences burnt area (see Sect. 2.1.3),finally emissions are calculated according to leaf and stem carbon for each PFT (see Sect. 2.1.4).

Figure 2 .
Figure 2. 1997-2010 mean yearly burnt fraction (above) and emitted carbon (below, in kg m −2 ).Shown for INFERNO on the left (with CRU-NCEP meteorology and interactive ignitions; ignition mode 3) and for GFED on the right.

Figure 3 .
Figure 3. 1997-2010 biomass burning emissions and burnt area predicted by INFERNO.Two driving data sets were used, CRU-NCEP (solid lines) and WFDEI (green dotted line).Observations are shown in black (MODIS-based estimates).The grey shading represents 1 standard deviation within GFEDv3's estimates.

Figure A1 .
Figure A1.The mathematical functions used for individual dependencies of INFERNO on key driving variables for flammability (a, b, c,d, e) and ignitions (f), within the range of reasonable Earth observations.Note the population density only influences the model output if ignition mode 3 is selected (interactive lightning and human ignition).

Figure B1 .
Figure B1.Emitted carbon difference maps between the four runs performed to analyse the sensitivity of INFERNO to ignitions (our three ignition modes; see Sect.2.1.1)and meteorology (CRU-NCEP and WFDEI-GPCC).

Table 2 .
Mean yearly emission budgets in Peta-grams of emitted carbon and mean yearly burnt area budgets in million km 2 for the 1997-2010 period.Latitudes were bound to beyond 50 • (high latitudes), 35 to 50 • (mid-latitudes), 15 to 35 • (low latitudes) and below 15 • (equatorial).Four configurations of INFERNO are presented, with CRU-NCEP and WFDEI driving meteorology coupled with three ignition modes: mode 1 indicates constant anthropogenic and lightning ignitions, mode 2 is for constant anthropogenic with interactive lightning ignitions, and mode 3 for interactive lightning and anthropogenic ignitions.

Table A2 .
The characteristics of the Canadian's Fire Weather Index's three fuel moisture codes.