Operational implementation of the burned area component of the Copernicus Climate Change Service : from MODIS 250 m to OLCI 300 m data

This paper presents a new global, operational burned area (BA) product at 300 m, called C3SBA10, generated 10 from Sentinel-3 Ocean and Land Colour Instrument (OLCI) near-infrared (NIR) reflectance and Moderate Resolution Imaging Spectroradiometer (MODIS) thermal anomaly data. This product was generated within the Copernicus Climate Change Service (C3S). Since C3S is a European service, it aims to use extensively the European Copernicus satellite missions, named Sentinels. Therefore, one of the components of the service is adapting previous developed algorithms to the Sentinel sensors. In the case of BA datasets, the precursor BA dataset (FireCCI51), which was developed within the 15 European Space Agency’s (ESA) Climate Change Initiative (CCI), was based on the 250m-resolution NIR band of the MODIS sensor, and the effort has been focused on adapting this BA algorithm to the characteristics of the Sentinel-3 OLCI sensor, which provides similar spatial and temporal resolution to MODIS. As the precursor BA algorithm, the OLCI’s one combines thermal anomalies and spectral information in a two-phase approach, where first thermal anomalies with a high probability of being burned are selected, reducing commission errors, and then a contextual growing is applied to fully detect 20 the BA patch, reducing omission errors. The new BA product includes the full time-series of S3 OLCI data (2017-present). Following the specifications of the FireCCI project, the final datasets are provided in two different formats: monthly fullresolution continental tiles, and monthly global files with aggregated data at 0.25-degree resolution. To facilitate the use by global vegetation dynamics and atmospheric emission models several auxiliary layers were included, such as land cover and cloud-free observations. The C3SBA10 product detected 3.77 Mkm, 3.59 Mkm, and 3.63 Mkm of annual BA from 2017 25 to 2019, respectively. The quality and consistency assessment of C3SBA10 and the precursor FireCCI51 was done for the common period (2017-2019). The global spatial validation was performed using reference data derived from Landsat-8 images, following a stratified random sampling design. The C3SBA10 showed commission errors between 14-22% and omission errors from 50 to 53%, similar to those presented by the FireCCI51 product. The temporal reporting accuracy was also validated using 4.7 million active fires. 88% of the detections were made within 10 days after the fire by both products. 30 The spatial and temporal consistency assessment performed between C3SBA10 and FireCCI51 using four different grid https://doi.org/10.5194/essd-2020-399 O pe n A cc es s Earth System Science Data D icu ssio n s Preprint. Discussion started: 14 January 2021 c © Author(s) 2021. CC BY 4.0 License.


Introduction
Interactions between climate and fire are bidirectional. On the one hand, climate impacts fire regimes (Marlon et al., 2013;Turco et al., 2019b), mostly by modifying temperature and precipitation patterns, which in turn impact fire ignition and behaviour through changes in soil and fuel moisture, vegetation productivity, and fuel availability (Andela and van der Werf, 2014; Daniau et al., 2013;Enright et al., 2015;Moritz et al., 2012;Bowman et al., 2020). On the other hand, biomass burning 40 is a critical source of aerosols and greenhouse gases that directly affect atmospheric chemistry (van der Werf et al., 2017;Ward et al., 2012) and carbon budgets Poulter et al., 2015) and carbon stocks (Friedlingstein et al., 2020). Emissions from fires increase aerosol optical depth, modifying the radiation budget and, thus, warming the lower atmosphere, which affects regional temperature, clouds, and precipitation patterns (Tosca et al., 2010;Tosca et al., 2014).
Fires have important ecological implications as well. They are a critical component supporting biodiversity (Kelly and 45 Brotons, 2017), since they alter vegetation composition, structure, and succession (Bowman et al., 2009). Furthermore, wildfires have significant societal impacts, particularly when extreme events occur as a result of climate anomalies. Large damages on people's lives, health and infrastructures have been observed in recent catastrophic fire seasons of Southern Europe, Australia and Western USA (Bowman et al., 2020;Turco et al., 2019b).
The Global Climate Observing System (GCOS) programme promotes international efforts to generate a wide range of 50 emissions for atmospheric chemistry modelling (Seiler and Crutzen, 1980;van der Werf et al., 2017) and the parameterization of fire-enabled Dynamic Global Vegetation Models (DGVMs) (Forkel et al., 2017;Forkel et al., 65 2019; Hantson et al., 2016;Hantson et al., 2020). These applications strongly benefit from an extended temporal coverage of BA datasets, while assuring their temporal and spatial consistency (Heil, 2019).
The FireCCI project has developed three BA algorithms to generate BA data from moderate-resolution sensors. The three resulting BA products were named FireCCI41, based on Envisat satellite's Medium Resolution Imaging Spectrometer (MERIS) data (Alonso-Canas and Chuvieco, 2015), covering the period 2005-2011 at 300m spatial resolution; the 70 FireCCI50 product, based on 250m resolution Moderate Resolution Imaging Spectroradiometer's (MODIS) bands (Chuvieco et al., 2018), and an improved version of this product, named FireCCI51, which extends from 2001 to 2019 (Lizundia-Loiola et al., 2020b). These three algorithms are based on a hybrid approach, combining Near-Infrared (NIR) and Red reflectance with active fires detections obtained from thermal anomalies. The three algorithms also followed a two-phase approach (seed detection + region growing), although the newest one overcame the previous version's weaknesses by improving BA 75 detection using active fire cluster-based thresholds (Lizundia-Loiola et al., 2020b).
A few years after the ESA CCI programme started, the European Commission launched the Copernicus programme, which included different services mostly based on Earth observation data. One of them is the Copernicus Climate Change Service (C3S, https://climate.copernicus.eu/, last accessed December 2020). C3S has been developed as an operational service that routinely produces datasets for more than 20 ECVs (Thépaut et al., 2018). It builds upon the research and development 80 carried out by the worldwide scientific community, especially the European one. In this regard, the C3S programme is now in charge of the operational processing of several ECVs, many of which were generated within the CCI programme (Thépaut et al., 2018). Since C3S is a European service, it aims to use extensively the European Copernicus satellite missions, named Sentinels. Therefore, one of the components of the service is adapting previous developed algorithms to the Sentinel sensors. consistent as possible with its predecessor FireCCI51 (2001FireCCI51 ( -2019 to ensure that multi-decadal analyses can benefit from both datasets uninterruptedly. Hence, quality assessment was performed in two stages. First, the period 2017-2019 was spatially validated using reference data obtained from multitemporal Landsat-8 images and temporally using active fire 100 information (Section 2.6). Then, an inter-comparison was performed between FireCCI51 and C3SBA10 datasets to provide users with insights about the consistency between the two products, pointing out similarities and discrepancies (Section 2.7).
The results of the BA detections for the common time series (2017-2019) as well as the quality assessment of the product are presented in Section 3, followed by the discussion and conclusion (Section 4 and Section 5, respectively).

Original burned area algorithm
The original FireCCI51 algorithm was initially developed for the Terra MODIS NIR band at 250m spatial resolution and, therefore, some minor changes were introduced to adapt it to the new sensor (Section 2.3). A detailed description of the original algorithm can be found on Lizundia-Loiola et al. (2020b), although a simplified description is provided in the following paragraphs and in Figure 1. 110 The FireCCI51 algorithm follows a hybrid approach that takes advantage of combining active fires and spectral information.
It first creates monthly composites of NIR reflectance by searching for minimum NIR reflectance while maximising the proximity to the fire date determined by the nearest active fire. Then, a second variable is computed from the relative NIR drop of consecutive composites. The third variable is generated from neighbour active fires that are close enough in space (<1875m) and time (<4 days) to be aggregated into spatio-temporal clusters (STC) (Lizundia-Loiola et al., 2020b). 115

OLCI Surface directional reflectances
The S3 satellites were designed to provide operational ocean and land observation services (https://sentinel.esa.int/web/sentinel/missions/sentinel-3, last accessed December 2020). They represent the continuity of the European Remote Sensing satellites (ERS-1 and ERS-2), the Envisat and the Satellite Pour l'Observation de la Terre (SPOT). S3 have a low (814.5 km), high inclination (98.65º), sun-synchronous Earth-orbit. With a period of ⁓101 minutes, 135 S3 need 27 days to complete a full cycle (385 orbits). The first satellite, called S3A, was launched in February 2016 and the second one, called S3B, was launched two years later. The orbit of S3B is identical to S3A but flies ±140º out of phase to improve revisit time. The design life of S3 is 7.5 years (consumables for 12 years), but the mission is expected to last longer, as two additional satellites will be launched during the first years of the current decade.
The OLCI sensor, one of S3's main payloads, is a push-broom imaging spectrometer composed of five cameras that are 140 12.6º tiled in western direction to mitigate potential sun-glint effects (https://sentinel.esa.int/web/sentinel/technicalguides/sentinel-3-olci, last accessed December 2020). It has a swath width of 1270 km (Field-of-View = 68.6º) and it offers global coverage at 300m every ⁓2 or ⁓1 day depending on the number of satellites available (S3A or S3A+S3B, respectively) and application (ocean or land). The OLCI instrument includes the same 15 spectral bands of MERIS (400-1020 nm), and six extra channels that were included to improve the atmospheric and aerosol corrections (Table 1). All this 145 information is included in the OL_1_EFR Level 1B product, which provides full resolution, calibrated, ortho-geolocated and spatially re-sampled Top-of-Atmosphere (TOA) radiances (https://sentinel.esa.int/web/sentinel/user-guides/sentinel-3olci/product-types/level-1b, last accessed December 2020).
This product was converted to surface directional reflectance (SDR) using an automated pre-processing chain that generates OLCI Level 3 SDR. The first module of the chain converted the TOA spectral radiance to the apparent SDR. Then, a pixel 150 identification module calculates a set of pixel classification attributes, like clear, cloud, snow/ice, cloud shadows, etc. The third module is an atmospheric correction algorithm, which included the correction for the absorbing and scattering effects of atmospheric gases, in particular ozone, oxygen and water vapour, of the scattering of air molecules (Rayleigh scattering) and the correction of absorption and scattering due to aerosol particles. The atmospheric correction including aerosol retrieval was first developed in the GlobAlbedo project (GlobAlbedo, 2013). The last processing module of the pre-processing chain 155 (compositing and mosaicking) was applied to retrieve the final Level 3 products. These 1-day SDR composites were derived from an input set of single satellite observations (i.e., SDR and pixel classification data described previously). Thus, those single observations were (i) reprojected onto a Plate Carrée grid, (ii) temporally aggregated for given binning cells (tiles), and (iii) mosaicked based on the binning cells onto a Level 3 product of 10x10 degrees tiles ( Figure 2

Auxiliary data 160
Information on active fires was used to guide both the seed and the growing phase of the BA algorithm. The input product was the MODIS MCD14ML Collection 6, which provides global monthly information on the location of thermal anomalies using both Terra and Aqua MODIS bands (Giglio et al., 2018b) at 1 km spatial resolution. The MCD14ML product is provided in ASCII format and one of its attributes specifies the presumed origin of the thermal anomaly: 0 for presumed vegetation fire, 1 for active volcano, 2 for other static land sources, and 3 for offshore thermal anomalies. Among these 165 categories, only those anomalies labelled as 0 were considered. This product was selected as it was the one used by the original FireCCI51 algorithm (Lizundia-Loiola et al., 2020b).
Global, annual land cover (LC) maps, which were provided by the Land Cover C3S, were used as well (https://cds.climate.copernicus.eu/cdsapp#!/dataset/satellite-land-cover?tab=overview, last accessed December 2020). LC production was transferred from CCI to C3S at the same time as the BA datasets, which meant that the same working group 170 was the responsible of adapting the algorithms. So, although the FireCCI51 BA algorithm used the latest LC v2.0.7 of CCI project (ESA, 2017), the LC version v2.1.1 from C3S was considered fully consistent with its predecessor (Defourny et al., 2020). In fact, the annual land cover maps from 1992 to 2015 developed within the CCI were brokered to C3S and extended for the years 2016-2018. Data at 333m spatial resolution from PROBA-V was used to generate the global, annual LC maps of 2016, 2017, and 2018. The use of the LC products had two main objectives. On the one hand, they were used to mask 175 unburnable areas, i.e. bare areas, urban areas, water bodies, and permanent snow and ice classes, to decrease the amount of data to be processed. On the other hand, they were used to report which land cover class was burned. Since the idea was to characterise the situation before the fire, the LC map that was applied in each year was the one prior to the BA analysis, e.g.
LC of 2016 was used for the BA processing of 2017. This product was selected to 1) ensure consistency with the original algorithm, 2) ensure consistency within the C3S, and 3) use a product that offered LC information at a similar spatial 180 resolution of the input OLCI SDR.

Adapted burned area algorithm
The adaptation was based on the analysis of the same study sites that were used for the FireCCI algorithm downloaded from the Canadian National Fire Database (CNFDB, http://cwfis.cfs.nrcan.gc.ca/ha/nfdb, last accessed December 2020). The remaining seven tiles were used to visually check that no major problems arise when applying the adapted algorithm to the new sensor, e.g. border effects in Central Africa, problems on the thresholds due to contrasted fire regimes within the same tile in Angola or region-growing problems due to high BA/active fire ratios found near Kazakhstan. 195 The adaptation of the algorithm to OLCI and its integration in the C3S operational system were carried out during 2018 and, hence, the training dataset's temporal coverage was limited to 2017, when only S3A was available.
Most of the parameters that control the FireCCI51 BA algorithm were independent from the base sensor used to detect BA, such as the time-gap used to temporally aggregate active fires or the absolute thresholds that were fixed for global detection (Lizundia-Loiola et al., 2020b). However, parameters linked to a distance might be affected by the spatial resolution of both 200 the surface reflectance product and the active fire product. The fixed distance used to spatially aggregate active fires into a same cluster (Section 2.1), for example, might depend on the spatial resolution of the input active fires product. Higher spatial resolution thermal bands can lead to higher densities of active fire pixels within a same burned patch (Oliva and Schroeder, 2015;Schroeder et al., 2014) and, therefore, the distance that is needed to spatially aggregate them could be smaller. In this case, there was no difference between the input active fire product used in the development of the original 205 algorithm and the adapted one (Section 2.2.2). Similarly, the distances used to define the unburned region around active fire clusters (Section 2.1) depend on both the spatial resolution of the input surface reflectance and the density of active fires per burned patch. The spatial resolution of the base sensor directly affects the number of surface reflectance pixels that fall into the unburned strip and, hence, could be used to estimate the thresholds. The same distance could imply the selection of much more pixels when the spatial resolution of the input sensor was high while in coarser resolutions could lead to a number of 210 pixels that may be insufficient. The difference of the spatial resolution of the input sensors used in the original algorithm (250m) and the adapted one (300m) were considered sufficiently similar to keep the parameters that defined the unburned strip.
Since the original FireCCI51 BA algorithm and, hence, the adapted one, needed only one input NIR band, the most suitable OLCI band (Table 1) had to be selected. In the case of MODIS it was not necessary to make such selection because it offers 215 only one NIR band (Lizundia-Loiola et al., 2020b). Based on a sensitivity analysis of the MERIS bands carried out by Oliva et al. (2011) during the development of FireCCI41, two NIR bands were considered candidates to be the input of the adapted BA algorithm: Oa12 centred at 753.75nm and Oa17 centred at 865nm (Table 1). These spectral bands were the most sensitive bands for BA detection, although no significant differences were found between them (Oliva et al., 2011).
The adapted BA algorithm was run two times with the training dataset, i.e. thirteen study sites and year 2017, using in each 220 process one of the mentioned bands. The assessment against the services' fire perimeters showed slightly more accurate results of the 865nm band over the 753.75nm one (same commission error and ⁓1% lower omission error). None of the bands showed unexpected anomalies in the rest of the tiles. Taking into account these results and the similarity between the OLCI 865 ± 10 nm band and the MODIS 858.5 ± 17.5 nm band used in the original FireCCI51 BA algorithm, which implied https://doi.org/10.5194/essd-2020-399 a greater consistency between the algorithms, the OLCI's 865nm NIR band was selected as main input for the adapted 225 algorithm.

Product generation
The adapted algorithm was processed in a monthly basis for each 10ºx10º tile and, hence, the raw outputs were provided per month and tile. Pre-processing and BA processing took advantage of the independence of tiles and were produced in parallel for multiple tiles. BA processing needed a time series of two prior composites to detect the NIR drops, and it required active 230 fire data as seeds for the growing step, which introduced constraints for sequential processing of the time series. These raw outputs were formatted, following mainly the specifications of the FireCCI project (Chuvieco et al., 2017), to obtain the pixel and grid products that were delivered to the final users in NetCDF4 format. The former provided the original BA classification at the original spatial resolution (300m) distributed in six continental tiles ( Figure 3). The latter computed a global aggregated summary of the raw results at 0.25º spatial resolution. Both the pixel and grid products provided auxiliary 235 layers suggested by the project end-users, including the land cover that was burned, and the uncertainty, fraction of observed area and burnable area (Table 2).
A total of 273 tiles that encompass the main burnable areas of the Earth, except from Antarctica, some small islands and regions northern than 80º latitude, were processed. The C3SBA10 extended the spatial coverage provided by the FireCCI, whose eastern border was extended from 50º W to 26º W in C3SBA10 ( Figure 3).
The operational dataset was processed since January 2017, ensuring up to three years of overlap (2017-2019) between the FireCCI51 and C3SBA10 products. The delivery frequency of the operational product ranged from three to six months 245 depending on the availability of the MODIS active fire MCD14ML product.

Uncertainty characterisation
Uncertainty characterisation is increasingly demanded by the end-users of BA products since it greatly helps the parameterization of climate models (Heil, 2019). Although important advances have been made in the last years regarding uncertainty characterisation of ECVs (Merchant et al., 2017;Mittaz et al., 2019;Sayer et al., 2019), methods to standardise 250 the generation of this variable are still needed (Chuvieco et al., 2019). Therefore, the CS3BA10 algorithm follows the same approach used in the FireCCI51 product. Four input variables related to different phases of the original algorithm were used to compute the estimation of BA detection uncertainty: monthly NIR composite (NIR), monthly relative NIR drop (RelΔNIR), number of valid observations in the first 10 post-fire days (obs) and the distance to the nearest seed-pixel (dist).
The uncertainty was computed using a logistic regression. Based on those four variables and the results of 2008 of the study 255 sites that were used to develop the original algorithm, the final coefficients of the regression were obtained. In the case of the adapted C3SBA10 algorithm, the coefficients of the logistic regression were updated (Eq. 1) using the four input variables and results of the training dataset described in Section 2.3: (1)

Spatial assessment
The spatial accuracy assessment was conducted for the period 2017-2019 for both the FireCCI51 and C3SBA10 products.
An independent reference BA dataset derived from Landsat-8 OLI imagery was produced in compliance with the CEOS LPVS stage 3 validation requirements (Morisette et al., 2006), which implies a rigorous statistical selection of reference sites representing diverse global conditions. Hence, for each calendar year 100 sample units were randomly selected using a 265 stratified random sampling design. The sampling units were spatially defined by the Thiessen scene areas (TSAs) constructed by Cohen et al. (2010) and Kennedy et al. (2010) from the Landsat-frames World Reference System 2 (WRS-2).
TSAs enable the partition of the earth's surface into non-overlapping spatial units, allowing the computation of unbiased estimators (Gallego, 2005).
To ensure that sampled units were distributed across the main biogeographic regions, a first level of stratification was 270 applied based on the biomes defined by the ecoregions2017 map (Dinerstein et al., 2017). The original 14 biomes were aggregated into 8 major biomes, i.e. boreal forest, deserts and xeric shrublands, Mediterranean, temperate forest, temperate savanna, tropical forest, tropical savanna, and tundra ( Figure 4). A second level of stratification was applied based on the FireCCI51 BA extent for each calendar year. Thus, each TSA within each biome and year was assigned to a high or low BA stratum following the approach of Padilla et al. (2014). Then, the sample's allocation within the 16 resulting strata (8 biomes 275 x 2 high-low strata) was established proportionally to ℎ √ ̅̅̅̅ ℎ , where ℎ is the number of units in stratum h and ̅̅̅̅ ℎ is the mean mapped BA for stratum h (Padilla et al., 2017).
The reference data for the selected sites were obtained from pairs of Landsat-8 images using a semi-automatic classification algorithm (Bastarrika et al., 2014), followed by a visual inspection to confirm the correct identification of burned patches.
Consecutive pairs of images were used to obtain long temporal reference data, covering several months, the entire fire season 280 or the whole year, depending on the image availability. These long temporal reference data allow for an extended temporal overlapping between reference and BA datasets, which makes it possible to improve the spatial accuracy assessment of the BA products, minimizing the impact of the product's temporal reporting accuracy (Section 2.6.2) in the spatial accuracy estimates (Franquesa et al., 2020b).
Accuracy metrics were computed based on the error matrix, which is widely used by the scientific community for thematic 285 accuracy assessment (Congalton, 1991). This matrix was derived from the cross tabulation of reference data and the BA products. The metrics used to assess the product accuracy were the Dice coefficient (DC) (Dice, 1945), the relative bias (relB) and the commission (Ce) and omission (Oe) errors. DC measures the similarity between the two sets of data; the higher its value, the more the accuracy of the burned category. The relB informs about the bias of the BA product relative to the reference BA, negative relB means BA product underestimation and positive relB, overestimation. Ce refers to the 290 burned area mapped by the BA product but not classified as burned in the reference data, and Oe to the burned area in the reference data not mapped in the BA product. All the accuracy metrics and their associated standard errors were estimated for the whole population applying the formulas described in Padilla et al. (2014) and .

Temporal reporting accuracy assessment
A validation of the temporal reporting accuracy of the products was performed as well. Unlike the assessment of the spatial 295 accuracy, which is already addressed in the previous section, the assessment of the temporal accuracy aims to show how accurately the day-of-burn (JD in Table 2) is determined by the BA product. Then, following the approach of Boschetti et al. (2010), the day-of-detection reported by the active fires (Section 2.2.2) was used as ground truth and compared with the dayof-detection reported by the corresponding pixel of the global BA product. These differences where then analysed per biome as well. 300

Consistency assessment
The estimates of accuracy metrics provided by the validation exercise provide an objective, valuable information regarding the spatial and temporal accuracy of the validated products. However, along with the spatial and temporal accuracy, consistency between products is also of primary interest for multi-decadal analyses, e.g. climate modelling (Chuvieco et al., 2019;Heil, 2019). From this perspective, the C3SBA10 operational product cannot be understood as a unique, independent 305 dataset, but as a continuation of its predecessor FireCCI51 product, ensuring uninterrupted global BA estimations from 2001 to present. To account for the similarities and differences that can arise between both datasets spatial and temporal trends of BA were compared.
A fundamental step when comparing global BA datasets is to define both spatially and temporally the comparison grid that will be used to aggregate the information. Although different spatial and temporal comparison units can be found in the 310 literature (Boschetti et al., 2004;Humber et al., 2018;Turco et al., 2019a), in this case the requirements of some of the most relevant applications of global BA datasets were used to determine the comparison framework (Chuvieco et al., 2019;Heil, 2019). However, in most cases, different applications use different spatial and temporal analysis units. The "fire model intercomparison project" (FireMIP), for example, which systematically compares different fire-enabled DGVMs, uses global, monthly BA data aggregated at 0.5º spatial resolution for model benchmarking (Hantson et al., 2016). Statistical 315 analyses that target at identifying factors that control fire patterns at global and regional scales commonly require monthly BA data aggregated at 0.25º or coarser spatial resolutions (Andela and van  temporal scales at 0.25º resolution. However, several global fire emission databases are already available at 0.1º (e.g. Kaiser 320 et al. (2012) updated) and even 0.05º (e.g. Kuenen et al. (2014)) following the increasing spatial resolution of some global atmospheric chemistry models (Aleksankina et al., 2018;Hu et al., 2018). Thus, it was decided to aggregate BA information into four different geographic grid cell sizes, i.e. 0.05º, 0.1º, 0.25º, and 0.5º, in a monthly and annual basis.
The results were aggregated spatially into eight biomes as well, following the first stratification of the validation exercise.
Four comparison cases were identified: biome-specific monthly BA, biome-specific annual BA, global monthly BA and 325 global annual BA. A total of 1404 scatterplots (3 years x (12 monthly +1 annual) x (eight biomes + 1 global) x 4 spatial resolutions) were generated based on these comparison cases and the four comparison grids defined above for the analysed years (2017-2019). A linear regression was calculated for each one (Humber et al., 2018;Turco et al., 2019a). Three metrics were used to determine the agreement of the products for a given case. The slope of the regression line was used to analyse the bias, the correlation coefficient (r) was used to determine the agreement between products and the root mean square error 330 (RMSE) was used to account for the dispersion of the estimations.

Trends on global burned area
More than three years (2017present) of S3 OLCI data were processed to generate the C3SBA10 global BA product. A total of 273 10ºx10º tiles were needed to cover almost all the burnable areas on the Earth. The algorithm results were 335 summarised and published in pixel (300 m resolution) and grid (0.25-degree resolution) products, using in both the NetCDF4 format. Figure 5 shows the annual accumulated grid BA product for the year 2019 while Figure 6 shows the latitudinal and seasonal contribution of BA for the same year. The amount of BA detect per biome for the three complete natural years can be found in Table 3.
The C3SBA10 product detected 3.77, 3.68, and 3.59 Mkm2 of BA for the years 2017, 2018 and 2019, respectively. The 340 main land cover contributor to global BA was tropical savanna, which had burned 2.8 Mkm 2 per year (Table 3) The influence of the dry season in these tropical biomes is quite evident, with a marked fire period from November to March in the Northern Hemisphere, and from June to October in the Southern hemisphere ( Figure 6). This trend slightly differs in South East Asia where the fire season continues until May.
Another significant belt of fire activity can be found in the temperate forest and grasslands of the Northern hemisphere ( Figure 5), with relative maximum around 50º N ( Figure 6). The C3SBA10 detected that on average the accumulated 355 contribution to the global fire activity of these biomes was 0.27 Mkm2 (7.5%). Most of this BA comes from the grasslands on the Asian steppe, which is dominated by large fires, although there is also an important agricultural activity (Hall et al., 2016;Zhu et al., 2017). The central plain of United States also exhibits fire activity linked to croplands as well as the forest areas of the north-western territories. In the southern hemisphere, the BA derived from the extreme wildfires of 2019-2020 in the temperate forest of south-eastern Australia (Bowman et al., 2020) is noticeable in Figure 5. The monthly variability of 360 BA in temperate areas is marked by both human agricultural practices and dry summer conditions, leading to two separated fire activity periods: one in March-April and the other in July-August.
The deserts and xeric shrublands' amount of BA detected by the C3SBA10 product showed a similar contribution to that of temperate areas, although with much more inter annual variability. The BA on 2017 was 0.26 MKm2 while on 2019 this number decreased by a half (0.11 Mkm2) (Table 3). Two areas on the Earth are responsible of most of this BA: central and 365 southern Kazakhstan are affected by large fires in summer (July-August) while the xeric shrublands and grasslands of central and western Australia show fire activity from October to January.
In the highest latitudes (>60ºN), the boreal forest showed a clear increasing trend in BA for the three-year period when the C3SBA10 product was computed, from 61,173 km2 in 2017 to 90,503 km2 in 2019 (50% increase). In the Tundra region, BA showed even higher increases, as in 2019 the product estimated 6 times more BA than the previous year (13,044 km2 370 versus 1,697 km2, respectively) ( Table 3). In fact, in 2019 the contribution of these two biomes was almost as much as that of temperate forest. Figure 6 shows how the fire season in the northern latitudes has a clear seasonal pattern, with highest activity in the Summer months (July-August).    (2018) commission errors (Ce) and higher omission errors (Oe) than FireCCI51 through the three-year period. Both products showed negative relative bias for all the years, indicating a systematic underestimation of BA, which is common to other 385 global BA datasets (Boschetti et al., 2019).

Spatial assessment
Although three years are not enough to extract significant conclusions about the temporal trends of accuracy, a similar behaviour can be observed in both products. Thus, the year that presented the highest accuracy metrics in one product also it did in the other product.

Temporal reporting accuracy assessment 390
About 4.7 million of active fires were used to estimate the temporal reporting accuracy of the three-year period of both products (Figure 7). On average (2017-2019) the results of the C3SBA10 showed that for deserts and xeric shrubland, both tropical and temperate savanna and Mediterranean biomes more than 90% of burned pixels were detected within the first 10 days after the fire (Table A1). This proportion was slightly lower in temperate and tropical forest where 80% and 86% of the pixels were labelled within 10 days of detection, respectively. The less accurate biomes were the boreal ones, where only the 395 70% (forest) and 69% (tundra) of the burned pixels were detected within the 10-day period. Although this period could be reasonable for some applications (e.g. dynamic vegetation modelling) there are other applications where it is important to know the precise date of burn. The analysis showed that, on average, only in both savannas and desert and xeric shrubland biomes C3SBA10 was able to detect within 0-or 1-day difference more than 20% of the cases. This accuracy significantly increased if the difference was set to a maximum of 3-days difference, where all biomes double the number of cases that 400 were detected within this accuracy.
The comparison between products showed that the temporal reporting accuracy is quite similar, having little or no impact the availability of more S3 images in 2019. In fact, for all forest biomes and tundra C3SBA10 provides a slightly more accurate reporting of the date of burn (Figure 7).

Consistency assessment 405
The relative differences on the amount of BA detected by each product are presented in Table 3. Table 5 shows the consistency metrics (correlation coefficient (r), slope and RMSE) at global and annual scale for the three-years period and the four comparison grid sizes. Figure 8 and Figure 9 show the spatial agreements between FireCCI51 and C3SBA10 for the  biomes where fire activity is frequent. Thus, the difference between both products was reduced more than a half in tropical 415 forest, temperate savanna, desert and xeric shrubland, and temperate forest and almost a half in tropical savanna (Table 3). In 2019, if tropical savanna was not considered, only 30,000 km2 of BA were underestimated by the C3SBA10 versus those of the FireCCI51. This underestimation was partially compensated by the detection of more BA (5,305 km2) in boreal forest and tundra.
Even though the annual BA detected by the global BA products differs through the study period, the correlation coefficient 420 (r) showed a significant agreement between them at all comparison grid sizes (Table 5). For instance, the r value at the finest comparison resolution (0.05º) was found to be 0.937 in the year with less agreement (2017)  There is a clear increase on the agreement between products in 2019 as well. The correlation metric is the highest in that year (from 0.95 at 0.05º to 0.99 at 0.50º grid resolutions), as well as the slope (from 0.88 at 0.05º to 0.91 at 0.50º grid resolutions), which means an increase in the amount of BA detected by the C3SBA10. The RMSE is also the lowest for that year, indicating that the availability of more images from both Sentinels makes the differences smaller. That is why a more 430 detailed analysis of the results of 2019 was conducted to better understand the spatial and temporal trends of the discrepancies between the products.
Through the eight biomes the lowest correlations at annual scale were those of deserts and xeric shrublands with an r=0.86 at 0.05º grid resolution (Figure 8e). Tropical savanna showed the highest annual agreement with correlations from 0.96 (at 0.05º) to 0.99 (at 0.50º). However, it had the highest RMSE (2.28 km 2 at 0.05º) due to the large amount of BA per grid cell. 435 Boreal forest and tundra show a very high agreement as well, with correlation coefficients between 0.94 and 0.98 for all grid sizes and the lowest RMSE, i.e. 0.27 km 2 for boreal forest and 0.14 km 2 for tundra at 0.05º. The rest of the biomes presented r values between 0.87-0.89 at 0.05º spatial resolution and above 0.90 from 0.10º upwards.
The two biomes that showed to be unbiased or with a little bias towards C3SBA10 were those located at highest latitudes correlations above 0.92 at 0.05º grid size, being higher in the rest of the comparison resolutions. In fact, at 0.50º the lowest correlation, which was found in April, showed a value higher than 0.95. A systematic underestimation of the C3SBA10 450 below 25% of the annual BA was noticeable through all the grid cells that belong to tropical savanna (Figure 9), which was traduced in slopes between 0.83 (April) and 0.90 (October) at 0.05º spatial resolution. The influence of tropical savanna is clear in the global metrics, which can in some cases obscure subjacent trends in other biomes.
In most cases, the correlation increases with the proportional contribution to the annual BA. This was the case of boreal forest ( Figure 8g) and tundra ( Figure 8h) whose peak months clearly matches the summer months, i.e. July and August, 455 which are responsible of the 70% and 77% annual BA in those biomes, respectively, and, at the same time, showed the highest agreement (r > 0.91 and slope > 0.92). However, not all biomes have a marked unimodal fire season since they encompass different continents and hemispheres and, therefore, fire activity is affected by diverse climatic and human conditions. In the case of deserts and xeric shrubland biome (Figure 8e) two separate peaks can be found, one in January The agreement shown between C3SBA10 and FireCCI51 is not as stable in the case of temperate and Mediterranean biomes.
In the former case there is a clear discrepancy in the monthly contribution of April to the annual BA. In the case of temperate 465 forest (Figure 8c), the monthly contribution corresponds to 21.3% for C3SBA10 while FireCCI51 estimated it at 16.8%. The same occurs in the temperate savanna (Figure 8d) where the contribution increases from 20.4% (FireCCI51) to 26.8% (C3SBA10). However, this discrepancy affected mainly the RMSE, which is much higher for this month than for the rest of the months in both biomes (RMSE=0.44 km 2 in savanna and RMSE=0.27 km 2 in forest at 0.05º grid size). In the Mediterranean biome (Figure 8f), from July to November there is a significant increase in the bias towards FireCCI51, losing 470 the stability shown in the first half of the year, and with slope values as low as 0.51, e.g. in the case of October.

Discussion
After adapting the original FireCCI51 algorithm to the characteristics of the C3SBA10, we found that the spatial and temporal trends in global BA of the two products showed consistent temporal and spatial trends. Most differences were observed in the first two years of the time series (2017 and 2018) when only one S3 satellite was operating, and therefore the 475 temporal resolution of OLCI (2-3 days) was much lower than the MODIS sensor (1 day). BA estimations between the two products were very similar in 2019, although C3SBA10 had 0.28 Mkm 2 less BA (global BA: 3.63 Mkm 2 for C3SBA10 and 3.91 Mkm 2 for FireCCI51). Most of this difference (92%) was found in tropical savanna, where a systematic underestimation of the C3SBA10 product was noticeable (Figure 9). This tendency towards underestimation was observed in other global BA products (generally based on coarse spatial resolution sensors) when comparing with regional products based on medium 480 https://doi.org/10.5194/essd-2020-399 resolution sensors. For instance, a continental BA product at 20 m derived from Sentinel-2 (S2) for the year 2016 and Sub-Saharan Africa found that global BA products significantly underestimate total BA, as they included 80% less area than the S2 BA product (Roteta et al., 2019). This was mainly caused by the poor detection of small fires (< 250 ha) in global products. C3SBA10 is likely missing more small fires than FireCCI51 due to its coarser spatial resolution (300 m vs 250 m).
This tendency towards underestimation was confirmed by the validation exercise carried out through the overlapping three-485 year period (2017-2019), which showed that both global BA products presented higher omission than commission errors. This trend agrees with the validation exercises of other global BA products that can be found in the literature (Boschetti et al., 2019;Chuvieco et al., 2018; as well as with the abovementioned inability of moderate resolution BA products to detect small fire patches. The omission errors presented in this paper were lower than those found by previous authors, e.g. 72.6% for the NASA's standard product in Boschetti et al. (2019) or 81% and 71% for two consecutive versions 490 of FireCCI products (FireCCI41 and FireCCI50) in Chuvieco et al. (2018). However, it must be highlighted that these numbers are not fully comparable to those presented in this paper since all these studies used the so called short units (reference data generated from a maximum period of 16 days) instead of long units (reference data covering several months, as done in this study) to estimate the spatial accuracy of the products. The short units' approach is more affected by the temporal reporting accuracy (Section 3.2.2) of the global BA products than when using long units. In any case, both the validation exercise and the consistency assessment showed a high agreement between the predecessor global BA product FireCCI51 and the operational C3SBA10. Although this consistency is still high in the years where only S3A is available (2017-2018) it significantly increased in 2019 due to the availability of the second S3B satellite. In fact, in 500 2019 the C3SBA10 product showed a similar accuracy (DC=0.59±0.04) to that estimated for the FireCCI51 (DC=0.61±0.03). The global and annual comparison metrics for that year were r = 0.952, slope = 0.88 and RMSE = 1.07 km 2 with a 0.05º spatial resolution, the finest spatial resolution used in the comparison (Table 5). The agreement between the products improved as the comparison resolution decreased (Table 5, and Figure 8).
A deeper analysis of the year 2019 showed different trends in the consistency between products among the biomes. It was 505 found that the changes in the correlation between products were linked to the BA detected for each month and biome.
Although this relationship did not seem to be linear, i.e. a specific increase in BA did not mean a parallel increase in the correlation, it was clear that the larger the BA detected the higher the correlation between products. The tropical savanna was an obvious example, were most of global BA was found. The boreal regions were another example where the correlation is the highest in the months with the highest fire activity, while it significantly decreases for the rest of the months. The 510 extreme wildfire event of Eastern Australia (Bowman et al., 2020) clearly showed this phenomenon since it significantly increased BA and agreement between products in December in the Mediterranean biome and in November and December in temperate forests. There were some unusual cases as well, where higher BA did not necessarily mean higher correlation. This was the case of the temperate areas where the correlation decreased in April. Temperate savanna and forest biomes encompass almost all the 515 Northern Hemisphere mid-latitude croplands located in the Russian Federation and Kazakhstan (Hall et al., 2016). Cropland fires are characterised to be small and transient (McCarty et al., 2009;Randerson et al., 2012), which made moderate resolution BA products (250-500 m) unable to properly characterise the extent of cropland BA patches (Hall et al., 2016;Zhu et al., 2017). It is reasonable to assume that the same issue was affecting both the FireCCI51 and C3SBA10 products, which clearly increased the uncertainty in those regions and, hence, decreased the agreement. In fact, it was very likely that the rest 520 of the biomes were affected by a similar issue (Vadrevu et al., 2019), showing a decrease in the agreement of products from March to May (Figure 8) when agricultural activity is high globally .
The impact of agricultural fire activity in the performance of moderate BA products is only an example of the importance of the human factor in global fire regimes (Andela et al., 2017). Benali et al. (2017) assessed the existence of bimodal fire seasons, i.e. areas with two separate peak months, that were directly related to human activities. They found that about 25% 525 of the areas with relevant fire activity showed this behaviour. In those cases, at least one of the two fire seasons was found to take place in sub-optimal weather conditions. Commonly, those periods have less valid images available due to clouds and cloud shadows, hindering BA detection, increasing uncertainty, and reducing agreement between products.
The sensor characteristics are also affecting the correlation between products as well. Different sensors imply different observation geometries, different angular effects and different times of overpass, directly influencing what is observed by 530 each sensor. This led to the detection of different burned patches in one product compared to the other, e.g. in cropland areas, or to the detection of different regions of the same burned patch as in some xeric shrublands, were fires can be large, but also irregular and with high spread rates (Laurent et al., 2018). Conversely, the temporal reporting accuracy (Section 3.2.2) was not significantly affected by this issue since a similar accuracy was shown by both products through the studied three-year period ( Figure 7). This was due to the use of monthly composites, which tend to select the most suitable observation from 535 the images available and did not require statistics of several pre-and post-fire images as in the case of multitemporal approaches.
The results presented in this paper represent a first step in the transition from MODIS-based to S3-based BA products. In that sense, OLCI sensor have shown a good capacity to detect BA globally, although these estimations still depend on MODIS active fires. Future work should be done to assess the impact of changing the source of the active fires to other 540 sensors such as the Visible Infrared Imaging Radiometer Suite (VIIRS) on board Suomi-NPP and NOAA-20 or the Sea and Land Surface Temperature Radiometer (SLSTR) on board S3. In the former case, VIIRS active fires can considerably increase the detection of small fires due to its improved spatial resolution of 375m (Oliva and Schroeder, 2015;Schroeder et al., 2014), although the impact on the identification of burned patches should be studied, due to the limitations of the moderate resolution reflectance data to detect small burned areas. In the latter case, a first study of S3 SLSTR active fire 545 product has shown a good performance of this product over small fires (Xu et al., 2020), which provides a unique opportunity to develop a hybrid BA algorithm entirely based on S3. Therefore, global BA estimations may be improved in the future if the operational algorithm used to detect BA within the C3S is updated with new and/or improved versions provided through research projects, such as FireCCI.

Data availability 550
The C3SBA10 product presented in this paper includes two types of BA files: monthly full resolution continental tiles and monthly global grid files at an aggregated resolution of 0.25 degrees, as described in Section 2.4 and Additionally, the reference files generated to spatially validate the data presented in this publication has been released and are available at https://doi.org/10.21950/BBQQU7 (Franquesa et al., 2020a).

Conclusions
This paper presents a new global BA product, called C3SBA10, that was developed and produced within the Copernicus Climate Change Service. The last version of a series of global BA algorithms developed by the ESA's FireCCI project, 560 known as FireCCI51 (Lizundia-Loiola et al., 2020b), was adapted to detect burned area for this Service. The algorithm follows a two-phase hybrid approach, which is based on S3 OLCI surface reflectance and MODIS active fires. Initially, the algorithm detects pixels with a high probability of being burned based on active fire information and NIR reflectance. Then, thresholds adapted to each fire and fuel conditions are computed using spatio-temporal active-fire clusters and their surroundings. Finally, a contextual growing is applied to fully detect the burned patch. Thus, the first step aims to reduce 565 commission errors while the second reduces omission ones. This algorithm is being operationally applied to produce the C3SBA10 product, which is the official BA product of the C3S service. At the time of writing this paper more than three years (January 2017present) of global BA are publicly available at the Climate Data Store in pixel (300 m) and grid            Tables   Table 1. OLCI bands characteristics. The band indicated in grey was the one used as input for the algorithm.

Band
Band centre (