A new global burned area product for climate assessment of fire impacts

Aim This paper presents a new global burned area (BA) product developed within the framework of the European Space Agency’s Climate Change Initiative (CCI) programme, along with a first assessment of its potentials for atmospheric and carbon cycle modelling. Innovation Methods are presented for generating a new global BA product, along with a comparison with existing BA products, in terms of BA extension, fire size and shapes and emissions derived from biomass burnings. Main conclusions Three years of the global BA product were produced, accounting for a total BA of between 360 and 380 Mha year 2 1 . General omission and commission errors for BA were 0.76 and 0.64, but they decreased to 0.51 and 0.52, respectively, for sites with more than 10% BA. Intercomparison with other existing BA datasets found similar spatial and temporal trends, mainly with the BA included in the Global Fire Emissions Database (GFED4), although regional differences were found (particularly in the 2006 fires of eastern Europe). The simulated carbon emissions from biomass burning averaged 2.1 Pg C year 2 1 .


INTRODUCTION
Biomass burning is one of the key components of atmospheric and terrestrial systems, as it influences emissions of various gases and aerosols, global carbon budgets and ecosystem dynamics (Krawchuk et al., 2009). Naturally occurring wildfires have maintained less tree cover on some lands than there would otherwise have been and therefore they have reduced the capacity of terrestrial ecosystems to sequester carbon (Staver et al., 2011;Yue et al., 2015). Fires also have biophysical effects by altering the vegetation type and structure, thus changing the surface albedo and energy balance (Beck et al., 2011;Rocha & Shaver, 2011). For these reasons, fire disturbance has been named by the Global Climate Observing System (GCOS) programme (GCOS 2011) as an essential climate variable (ECV). In 2010 the European Space Agency (ESA) launched a climate change initiative (CCI) programme to generate satellite-derived products that follow the specifications of GCOS (Hollmann et al., 2013). The Fire_cci project is part of the CCI programme and aims to develop global information on burned area (BA) following the requirements of atmospheric and vegetation modellers .
Global BA products have been developed in the past using data from various satellites/sensors . The most reliable (Padilla et al., 2015) are those based on the NASA's Moderate Resolution Imaging Spectrometer (MODIS) sensor, the MCD45 (Roy et al., 2008) and the MCD64 (Giglio et al., 2009). This latter BA product is one of the inputs of the Global Fire Emission Database (GFED) (v.3, Giglio et al., 2010;v.4, Giglio et al., 2013). This paper presents a new global BA product, which was developed within the Fire_cci project to meet the particular requirements of the climate modelling community. The main requirements relate to spatial and temporal resolution, file formats, information on fire patch distribution and precise reporting on the uncertainty and accuracy of the final product.
Our BA dataset is the first global one based on Envisat-MERIS images. As this sensor has higher spatial resolution than MODIS BA products (300 m versus 500 m at nadir), it was hypothesized that it would have greater potential for the detection of small fires (<50 ha), which make an important contribution to global fire effects (Randerson et al., 2012). This paper briefly presents the methods developed to generate the BA product, and compares this product with existing global BA datasets, both in terms of spatial and temporal trends and with regard to its potential for the estimation of fire emissions.

Product generation
The BA algorithm combined temporal changes in nearinfrared (NIR) MERIS-corrected reflectances with active fire detection from the standard MODIS thermal anomalies product, following a two-phase algorithm . Global processing of a 3-year time series (2006-08) was used to prototype the product.
Burned pixels detected by the MERIS BA algorithm were transformed to two BA products, following the recommendations of a user requirement questionnaire conducted with climate modellers. 1. The pixel product is delivered in monthly GeoTIFF files at the spatial resolution of the MERIS images (300 m at nadir, which was translated to a pixel size of 0.0029761905 geographical degrees). Files are arranged by continental tiles that make up the globe. The product includes the date when the burned pixel was first detected, a confidence level value reflecting the probability that the pixel was burned (taking into account number of detections, pre-processing assumptions and BA classification likelihood) and the land cover burned from the Glob-cover2005 land-cover map (Arino et al., 2007). Metadata information is provided as a separate XML file. 2. The BA grid product is released in NetCDF-CF format files at 0.58 resolution covering the whole globe every 15 days. This BA product includes the total BA within each grid cell, the standard error of the BA estimation, the fraction of the observed area within each grid cell over the reporting period and the number of distinct burn patches. Furthermore, the grid product also reports BA in each grid cell for each of the 18 vegetated land-cover classes in the GLOBCover 2005 product (Arino et al., 2007). The metadata are built into the NetCDF format. Further product information and downloads are available at http://www.esa-fire-cci.org/.
Validation of the final product was derived from multitemporal pairs of Landsat images, following CEOS Cal-Val guidelines (http://lpvs.gsfc.nasa.gov/fire_home.html, last accessed November 2015). These reference data were generated for the year 2008, selecting the 105 sites from a stratified random sample (see Padilla et al., 2014Padilla et al., , 2015.

Cross-analysis with existing BA products
Spatial and temporal patterns derived from our time series  were compared with the GFED4 BA product. GFED4 is the updated version of GFED3, which is the most widely used inventory in global biogeochemical and atmospheric modelling studies. The actual change in global BA from GFED3 to GFED4 during the period 2006-08 is only 20.1%. Some analyses were also done with GFED4s (released in July 2015). GFED4s consists of GFED4 complemented by the small fire database, following methods proposed by Randerson et al. (2012) based upon scaling active fire counts to BA.
To estimate emissions derived from different BA products, the ORCHIDEE dynamic global vegetation model (DGVM) (Krinner et al., 2005) was used. This model was recently coupled with the prognostic fire model SPITFIRE . Fire_cci and GFED4 BA data were originally provided with different land-cover maps (MOD12Q1 for GFED4 and Globcover2005 for Fire_cci). As DGVMs are based on the concept of plant functional types (PFTs; ORCHIDEE has 13 PFTs), BAs from observational data needed to be harmonized on a few common vegetation types. Five common vegetation types were identified: tropical forest, temperate forest, boreal forest, natural grassland and croplands. In this phase of the comparative analysis, agricultural fires were not simulated, as is common in most fire models embedded within large-scale vegetation models (except Li et al., 2013). The transformation of MOD12Q1 and Globcover2005 land-cover types to the five vegetation types followed the approach of Poulter et al. (2011) and involved three steps: (1) assumptions were made on different PFT fractions with respect to land-cover types; (2) PFTs with the same climate types (tropical, temperate and boreal) were pooled together within the same grid cell; (3) the 13 PFTs of ORCHIDEE were further regrouped into five vegetation types.
We compared the simulated global carbon emissions and CO emissions using the ORCHIDEE model with BA being prognostically simulated, and being forced by the Fire_cci data and GFED4 data. For the prognostic simulation, the fuel combustion completeness (CC) was dynamically simulated in the model (for details see Yue et al., 2014). For the simulations where observational BA data sets were used as input, the regional combustion completeness values were used (see Table 4 in , as constant CC (as those from GFED4 emission data set were not available). The prognostic simulation was conducted on a daily time step and the forced-BA simulations were conducted on monthly time step.
We choose fire C and CO emissions for comparison, because fire C emissions contribute to the CO 2 build-up in the atmosphere and are always used as the basis for deriving emissions for other trace gases and because CO is frequently used as a tracer of fire emissions (Leeuwen et al., 2013;Poulter et al., 2015). The CO emissions were derived in ORCHIDEE by using the emission factors (in units of g CO per kg dry matter consumed in fire) and C emissions (C 5 0.45 3 the dry mass of consumed fuel). Besides using the mean BA as reported in the Fire_cci data set, the high and low end of Fire_cci BA estimations (mean 1 SE and mean -SE, respectively) were also used to force the model, as a simple approach to construct the emissions uncertainty with respect to the BA uncertainty.
In addition to comparing results with existing products, new analyses focused on burn patch size and shape analysis were carried out with our Fire_cci product. Global BA analysis at the patch level has recently emerged for fire size distribution analysis based on pixel aggregation methods (Archibald et al., 2010;Hantson et al., 2015), helping to bridge the gap between landscape-scale and global fire regime characterization . Most fire behaviour models embedded in DGVMs simulate elliptical burn patches. The elongation of those ellipses (the ratio between their X-and Y-axis), area and orientation are relevant for a better understanding of regional fire processes (Barros et al., 2013;Mansuy et al., 2014). We tested for the usability of the pixel-level product to identify individual patches within each global grid cell and generate global-scale patch metrics. We used the spatio-temporal flooding algorithm proposed by Archibald & Roy (2009) and further used in Hantson et al. (2015) within each 28 3 28 resolution grid cell, a grid cell size in accordance with the maximum fire patch size observed world-wide  and previously used for the global patch size distribution analysis by Hantson et al. (2015). We kept the largest patch computed within each grid cell and its surrounding 0.58 buffer zone, and for which the central ellipse coordinates were located within the 28 3 28 grid cell. In doing so, large fires that potentially overlapped neighbouring grid cells were still considered and were referenced in the grid cell where their central coordinates belong. We cropped the continental-scale pixel product for each 28 3 28 grid cell tile (plus its 0.58 buffer zone) and we re-projected it into an equal-area resolution map. Each fire patch was then characterized by several metrics: the number of cells in each patch as an indicator of fire size, the perimeter to area ratio (P/A) as an index of patch complexity, the X/Y ratio (ratio between the longest and the shortest ellipse axis defining the patch elongation) and the orientation azimuthal angle theta (angle of the fitted ellipse enveloping the fire patch: azimuthal deviation from the north of the longest axis of the ellipse). We computed these metrics for the July-August season of 2006. For comparison, a similar patch analysis was performed with fire polygons delivered by forest services at the national level for the USA (http://wfdss.usgs.gov/wfdss/), Canada (http://cwfis.cfs. nrcan.gc.ca/ha.nfdb) and Australia (Bradstock et al., 2014). All analyses were performed with the R package 'raster'.

Product generation and validation
Figure 1 (top) shows the spatial distribution of the Fire_cci average BA for the period 2006-08. The most extensive burnings occurred in the tropical regions, particularly in the African continent, followed by the northern regions of Australia, central Brazil, Venezuelan and the Colombian Llanos and Southeast Asia. A second belt of burned regions is noticeable in the temperate grasslands and croplands of central Asia and the south-east USA. The boreal forests of Russia and Canada also have a substantial role in global biomass burnings. Total BA obtained with the MERIS algorithm ranged from 360 to 380 Mha year 21 for the 3-year period.
The accuracy measures revealed a very high overall accuracy (99.6%), with higher errors for the burned than the unburned pixels, which is common in BA products as burned patches are much less extended globally than unburned ones. The accuracy of burned pixels was observed to be higher for areas with greater fire occurrence (omission errors decreased from 0.76 to 0.51 and commission errors from 0.64 to 0.52 in sampling sites with more than 10% BA: Alonso-Canas & . Comparison with existing BA products (Padilla et al., 2015) showed that MERIS BA results have a similar overall accuracy to the two MODIS BA products (MCD45 and MCD64), but with higher commission and omission errors than MCD64 and higher commission error than MCD45. In terms of error balance, an overall trend towards underestimation was found (35%).

Comparison with BA products
In general, the Fire_cci product was found to be in close agreement with the GFED4 product in terms of temporal and spatial distribution ( Figure 1). Mean annual BA in the Fire_cci product was 369 Mha year 21 , 6.6% higher than the estimated 346 Mha year 21 of GFED4 (Table 1).
Absolute differences in the proportion of BA were within 5%. Larger discrepancies were confined to tropical Africa and Australia. The spatial patterns of relative differences reflected that Fire_cci tended to be notably higher in regions where small agricultural fires predominate. For example, across Europe, except for the Kazakhstan steppe region, Fire_cci estimates were two to five times higher than those from GFED4. Across agricultural India, Fire_cci estimates were typically more than four times higher than in GFED4.
In terms of land-cover classes (Fig. 2), Fire_cci exhibited the strongest contrast with GFED in woody savanna regions, where Fire_cci annual BA was higher than GFED4 by 30 Mha (or 26%). Open shrublands ranked second in terms of absolute differences; here, Fire_cci annual BA was 10 Mha (or 44%) lower than in GFED4, probably due to the confusion with savannas. Croplands ranked third in terms of absolute changes, with an increase in global BA of 30 Mha (or 35%). The higher spatial resolution of the MERIS sensor (300 m 3 300 m compared with 500 m 3 500 m in MODIS) could partially explain the higher performance of the Fire_cci product for areas with agricultural fires. Around 5.4% of all Fire_cci burned grid cells had BA sizes smaller than the lowest monthly BA in GFED4, which was 21 ha.
The global temporal patterns of Fire_cci and GFED4 BA were strongly correlated (r 2 5 0.86). Temporal coherence was also high on a regional scale (r 2 > 0.5 using the standard GFED regions, except for the boreal areas: r 2 5 0.18 for boreal Asia and 0.37 for boreal North America).
While Fire_cci and GFED4 showed large similarities when the products were integrated over large regions and/or temporal scales, pronounced differences emerged for specific fire events. Figure 3 shows, as an example, the differences in estimated fire-affected area during spring 2006 in eastern Europe and Central Asia. This event caused record levels of air pollution in the European Arctic, as documented by Stohl et al. (2007). The predominantly agricultural fires implied a strong increase in radiant energy between April and May, measured by the GFAS product which is based on detecting radiation coming from active fires using middle and thermal spectral bands (Kaiser et al., 2012). While the locations of major fires were well reflected in the Fire_cci product, they were largely missed by GFED4. Yet Fire_cci shows a much lower spatial extent of the fires when compared with the recently released GFED4s database, which includes BA estimates from small fires based on active fire detections (Randerson et al., 2012). This example nicely illustrates that each satellite fire product has different commission and omission characteristics dependent upon the region and the period selected.

Fire patch analysis
We first used the spatial distribution of fire patch indices (for the largest fire within each 28 3 28 grid cell) across the globe to assess the coherence of pixel aggregation into patches in addition to the BA at the grid cell level for the different products. Figure 4 illustrates the global pattern of patch direction (theta, Fig. 4a), patch elongation X/Y (Fig.  4b) and patch perimeter to area ratio (P/A, Fig. 4c), and their corresponding frequency distribution according to latitudinal thresholds (458 S to 08, 08-458 N, 458-908 N).
Patch orientation was equally distributed according to latitude and spatially heterogeneous with no major regional pattern, indicating no major bias in the processing chain. The patch elongation index (X/Y) is normally distributed across all latitudes around a value of 1.5, with lower values for the 458 S to 08 zone indicating less elongated patches in the most fire-prone regions of tropical savannas. The regional pattern remains globally heterogeneous with contrasted values for  neighbouring grid cells and a low spatial aggregation of similar fire types. The global distribution of the patch complexity index (P/A) is more related to the latitude, with a wide range of values in the northern latitudes progressively narrowing toward low values when going southward, indicating lower patch complexity when BA (and patch area) is higher.
To better assess the ability of the global pixel product to capture the actual patch structure, we compared the results obtained from fire patches derived from the BA pixel product with a similar analysis derived from fire polygons produced by national forest services in North America and Australia. Grid cell to grid cell cross-tabulations of continental maps at 28 3 28 resolution are presented in Fig. 5, for patch area thresholds of 10, 50, 100 and 200 pixels (corresponding, respectively, to 90, 450, 900 and 1800 ha), with their corresponding correlation coefficient r 2 and regression slope being given in Table 2. We observed a linear relation for patch area with r 2 > 0.968 for all patch size thresholds and a regression slope around 0.94 indicating a 6% underestimation of BA from MERIS compared with forest service estimates. The correlation on patch elongation (X/Y) was very low when considering all fires (r 2 5 0.024), but increased to 0.382 when considering fires larger than 900 ha. A similar result was observed for the patch orientation (theta), with r 2 5 0.49 for patch sizes >900 ha and r 2 5 0.75 for patch sizes >1800 ha. This Figure 4 Global pattern of largest fire shape indices for the summer (July-August 2006) fire season (28 resolution). X/Y, patch elongation as the ratio between the longest and shortest ellipse axes (log 10 scale); P/A, perimeter to area ratio; theta, azimuthal angle of the longest ellipse axis (degrees deviation from north). analysis illustrates the potential use of the MERIS pixellevel product for large fire shape assessment in DGVM simulations of patch elongation and direction driven by wind conditions, and as yet unexplored due to the lack of available data in global fire products. However, small fire shapes seem to remain uncertain due to the sensitivity of patch indices to burned pixel uncertainty, and the difference in fire patch resolution between global remote sensing products and local fire polygons obtained from highresolution remote sensing or field observation. We finally obtained a low correlation between forest service polygons and MERIS P/A index for large fires (r 2 5 0.027 for patch sizes >900 ha), indicating a high uncertainty in patch boundaries. The correlation including all fires is much higher (r 2 5 0.697) as a consequence of the intrinsic strong correlation between P/A and patch size observed in the global pattern analysis.

Comparison of atmospheric emissions
BA and estimated CO and C emissions are presented in Table  1. They were based on ORCHIDEE with the prognostic simulation, and as forced by GFED4 and Fire_cci BA data, as well as on the CASA model with GFED4 BA data. As most existing fire models (Lasslop et al., 2014) including ORCHIDEE are calibrated against GFED3.1 data, BA and Figure 5 Relationships between patch metrics computed from the Fire_cci burned area polygons (Y-axis) and those derived from the forest services (X-axis): area (ha), elongation X/Y (log 10 ), perimeter/area ratio (P/A, m m 22 ), and orientation (theta, degrees deviation from north). The size of the circles indicates the surface of the considered fire patch, reclassified into size classes >90, >450, >900 and >1800 ha. The correlation coefficients (r 2 ) and slopes of the regressions are presented in Table 2. Table 2 Correlation coefficients r 2 and regression slopes (in brackets) of the regressions between forest services fire patches and MERIS Fire_cci pixel-based fire patches. Analysis on fire size (no. of cells), patch elongation (X/Y ratio), patch complexity (perimeter/area ratio, P/A) and patch ellipse orientation (azimuthal angle theta, degrees) are presented for fire size classes >10 ha (at 18 and 28 resolution), >450 ha, >900 ha and >1800 ha. (see Fig. 5  carbon emissions from GFED3.1 data are also provided in Table 1. Besides those values, data from the recently released GFED4s have also been included. The annual BA estimated by the Fire_cci product was found to be similar to those obtained with the GFED3.1 and GFED4 products, but lower than the most recent estimations from GFED4s, which should be related to the consideration of small fires (Randerson et al., 2012). The ORCHIDEE prognostic simulation reported a low estimate of BA (232 Mha year 21 ), because calibrating the model gross primary productivity according to observational data (Jung et al., 2011) has generally reduced the fuel load and consequently BA (see Yue et al., 2015, for more details). ORCHIDEE simulations being forced by GFED4 and Fire_cci BA data, and the prognostic simulations, yielded annual fire carbon emissions of 2.0-2.1 Pg C year 21 , being close to those from GFED3.1 data (1.9 Pg C year 21 ) and GFED4s data (2.1 Pg C year 21 ). These estimations were all higher than the GFED4 data set (1.6 Pg C year 21 ). GFED4 and GFED4s data have generally lower fire fuel consumption than GFED3.1 data, and this partly explains the higher emissions in model runs because GFED4 and GFED4s emissions have not yet been publicly released (http://www.globalfiredata.org/data.html) and all fire models are calibrated against GFED3.1 data.
The two uncertainty simulations (i.e. ORCHIDEE-Fire_cci HIGH and ORCHIDEE-Fire_cci LOW) provide the lower and higher bounds of emissions (1.3-2.9 Pg C year 21 ), but it should be noted that a rather simple approach was used here as the heavy computation requirements of ORCHIDEE preclude a more complex uncertainty estimation approach such as Monte Carlo simulation.
The monthly carbon emission time series of fire carbon emissions are presented in Fig. 6. The correlation coefficient between the monthly carbon emission time series for the ORCHIDEE prognostic simulation and GFED4 data is 0.30, indicating that ORCHIDEE could only fairly capture the seasonality of emissions as revealed by the GFED4 data. The correlation with the GFED4 data increased to 0.65 when the model was forced by the Fire_cci BA data, and further to 0.75 when it was forced with GFED4 BA data. This showed that errors in prognostically simulating carbon emission seasonality could to a large extent be attributed to the quality of input BA data. The replacement of prognostic BA with observations improved the carbon emission simulation. The correlation between the simulations forced with Fire_cci and GFED4 was 0.92, higher than their respective correlation with the GFED4 emission data. This again confirms that Fire_cci and GFED BA data yield comparable results in fire emissions simulations.
As in the case of C emissions, prognostic and forced simulations were compared to analyse CO emissions. CO emissions forced with Fire_cci and GFED BA data gave similar estimates (329-351 Tg CO year 21 ) to those provided by GFED3.1 and GFED4s data (334-339 Tg CO year 21 ). The monthly time series of CO emissions resembled those of C emissions. As CO is considered as an effective tracer of fire emissions (Leeuwen et al., 2013;Poulter et al., 2015), we fed the fire CO emissions as reported by GFED4 and various model configurations into the atmospheric chemistry transport model LMDz-INCA and compared the agreement of resulting simulated CO mixing ratios with ground observations for 14 sites, where the CO mixing ratios are highly affected by fires (see Figure S1 in Appendix S1 in the Supporting Information for site distribution). The CO emissions from other sources and a simplified atmospheric chemistry process were included in the transport process (see Yin et al., 2015, for more details). The root mean square deviation (RMSD) between de-trended monthly simulated and measured CO mixing ratio time series for 2006-08 was calculated for each site. Of the 14 sites, minimum RMSDs were reached by using ORCHIDEE-Fire_cci CO emissions data for five sites [mean RMSD by ORCHIDEE-Fire_cci is 48.7 parts per billion (p.p.b.) for these five sites], by ORCHIDEE prognostic CO emissions for seven sites (mean RMSD by ORCHIDEE is 102.2 p.p.b. for these seven sites) and by GFED4 CO emissions data for two sites (mean RMSD by GFED is 50.0 p.p.b. for these two sites) (Table S1 in Appendix S1). Despite this approach being prone to errors, in terms of atmospheric transport and chemistry processes the results show comparable performances of Fire_cci BA to GFED4 in terms of CO emissions quantification.

DISCUSSION
The Fire_cci BA product was developed as a contribution to climate modelling efforts to improve our understanding of fire dynamics in the Earth system, and the general role of fires in the climate. Being part of the GCOS programme, the requirements of climate users were recognised as a priority for the Fire_cci project. An analysis of GCOS-stated requirements (GCOS 2011) for this ECV, along with the actual demands of the climate modelling community, identified the following characteristics for an ideal BA product. 1. Long-term time series (>30 years) with a high temporal consistency (6 5%). 2. Daily temporal resolution at the original spatial resolution of the sensor (weekly, 10-day or monthly basis) on a global scale. 3. Detection of BA patches smaller than 25 ha. 4. Maximum omission and commission errors should be below 15%, with demonstrated systematic and adequate validation using internationally agreed validation protocols. 5. Error traceability and uncertainty characterization should be included in both the pixel and grid products in a manner that can be easily understood and utilized by users in various applications. 6. Easy access to data, including auxiliary information and metadata on burned cover, burn patch distribution and burn severity/efficiency. Summarizing the main contribution of the Fire_cci BA product, we should point out that our product has made significant advances in several areas. First we were able to improve the detection of small fires by increasing the spatial resolution with respect to existing BA products (from 500 m to 300 m), which was demonstrated in the characterization of agricultural fires in eastern Europe and Central Asia. We performed the most extensive accuracy assessment done so far with global BA products (more than 300 Landsat images were used to generate the reference perimeters), using a statistically designed sample to obtain global and regional validation metrics, following standard CEOS protocols. In addition, the error and uncertainty characterization was provided at both pixel and grid level. This information was employed in a simple modelling scenario to estimate lower and upper carbon emission bounds, but additional exercises will be undertaken when longer time series are available. Finally, the Fire_cci BA product offers daily information at pixel size (date of detection) and biweekly BA totals at grid size, meeting the requirements of climate users, it includes the information layers demanded by modellers and it is easily accessible, including standard file formats and metadata.
In terms of product limitations in the framework of the GCOS requirements, the prototype Fire_cci product currently only offers a short time series (3 years), which will be extended to 15 years in the second phase of the Fire_cci project, with MERIS data complemented with other sensors. Omission and commission errors of the BA category were found to be much higher than those required by GCOS (<15%), although it should be emphasized that none of the existing global BA products meets those requirements (Padilla et al., 2015). Our global BA estimations have higher errors than those of MODIS BA products, but were found to be better balanced, with less underestimation than the MODIS products (but still close to 35%). A second phase of the project will process data from higher-resolution sensors, with the target of estimating the contribution of small fires, particularly in the more fire-prone regions such as Africa.
In terms of new data analysis not previously reported at a global scale, such as the fire shape indices, the Fire_cci pixel product has been found to be very useful for describing and better understanding fire size and fire shape distributions at regional and global scales. The impacts of these spatial distributions on fire regime characterization are clear, particularly when models are available to analyse fire behaviour and fire vulnerability globally . The implications of this shape analysis on improving fire emission estimations in DGVMs has to be further evaluated, but at least it has great potential given the fact that fire size and duration in current DGVMs are poorly evaluated .
Grid-level products derived from pixel-level analysis remain the main source of validation for coupled DGVM/fire models, but increasing the need for intermediate steps in the validation process has been pointed out by Yue et al. (2014). This includes fire patch size distribution and fire duration, assessed from regional datasets or regional pixel-level patch analysis (Archibald et al., 2010). Fire patch information would in turn enhance model validation and improve simulations, for instance to better estimate the completeness of fuel combustion in DGVMs as larger fires are probably associated with more complete burning. On the regional level, landscape fire succession models would also benefit from fine-scale datasets as provided in the 300-m resolution Fire_cci pixel product.
From the emissions simulation analysis it can be concluded that the Fire_cci BA product can be a good complement to existing BA estimations based on MODIS data. We have demonstrated that spatial and temporal BA trends of our product are consistent with those included in GFED, but regional variations provide new insights into potential problems or strengths of existing BA databases. The extreme fires in spring 2006 in eastern Europe are an example, as the Fire_cci product appears to capture BA more realistically than GFED. Air pollution model studies including particle transport models (Evangeliou et al., 2015) focusing on this event could therefore achieve better predictability with respect to measurements when using the Fire_cci product as input. Cross-comparison of multiple products is a means to validate and possibly improve individual fire products. It also helps user communities to select which product (or combination thereof) should be used for a given research question.
Finally, the climate scope of the CCI programme emphasizes the need to foster collaboration between ECV teams, which will greatly benefit both fire analysis as well as that of other ECVs. For instance, potential synergies between soil moisture, ozone, greenhouse gases or even sea temperature trends and fire occurrence and impacts have been previously suggested (Archibald et al., 2013;Hantson et al., 2015). The investigation of various relations among these variables, with the advantage that all the products are generated under the CCI umbrella using a consistent approach, will be undertaken within the framework of the Climate Modelling User Group (CMUG), which is also part of the CCI project.