Gas flaring activity and black carbon emissions in 2017 derived from Sentinel-3A SLSTR

Gas flares are a regionally and globally significant source of atmospheric pollutants. They can be detected by satellite remote sensing. We calculate the global flared gas volume and black carbon emissions in 2017 by applying (1) a previously developed hot spot detection and characterisation algorithm to all observations of the Sea and Land Surface Temperature Radiometer (SLSTR) instrument on-board the Copernicus satellite Sentinel-3A and (2) newly developed filters for identifying gas 5 flares and corrections for calculating both flared gas volumes (Billion Cubic Meters, BCM) and black (BC) carbon emissions (g). The filter to discriminate gas flares from other hot spots uses the observed hot spot characteristics in terms of temperature and persistence. The variability of the detection opportunity is corrected with an approximate method. 6232 flaring sites are identified worldwide. The best estimates of the annual flared gas volume and the BC emissions are 129 BCM with a confidence interval of [35 BCM, 419 BCM] and 73 Gg with a confidence interval of [20 Gg, 239 Gg], respectively. Comparison of 10 our activity results with those of the Visible Infrared Imaging Radiometer Suite (VIIRS) Nightfire data set and SWIR-based calculations show general agreement but distinct difference in several details. The calculation of black carbon emissions using our gas flaring data set with dynamic emission factors lie in the range of recently published black carbon inventories, albeit towards the lower end. The data presented here can therefore be used e.g. in atmospheric dispersion simulations. The advantage of using our algorithm with Sentinel-3 data lies in the previously demonstrated ability to detect and quantify small flares, 15 the long-term data availability from the Copernicus programme and the increased detection opportunity of global gas flare monitoring when used in conjunction with the VIIRS instruments. The flaring activity and related black carbon emissions are available as "GFlaringS3" on the Emissions of atmospheric Compounds and Compilation of Ancillary Data (ECCAD) web site (https://eccad3.sedoo.fr/#GFlaringS3, DOI 10.25326/19).

Satellite remote sensing has been utilized for regional and global identification and characterization of GF (Casadio et al., 2012b, a;Anejionu et al., 2014;Faruolo et al., 2014;Chowdhury et al., 2014;Anejionu et al., 2015;Faruolo et al., 2018). The advantages and limitations of satellite remote sensing for the observation and monitoring of GF have recently been reviewed by Obinna C. D. Anejionu (2019). The most prominent system is NOAA's VIIRS (National Ocean and Atmospheric Administration's Visible Infrared Imaging Radiometer Suite) Nightfire (VNF) dataset (see https://ngdc.noaa.gov/eog/viirs/download_ 5 viirs_fire.html), developed by Elvidge et al. (2013Elvidge et al. ( , 2016 and still under active development (Elvidge et al., 2019). It provides a consistent global survey of GF volumes from 2012 onwards (https://ngdc.noaa.gov/eog/viirs/download_global_flare.html). Caseiro et al. (2018) describe an adaptation and extension of the VNF algorithm, with which observations of the Sea and Land Surface Temperature Radiometer instrument (SLSTR) on-board the Copernicus Sentinel-3 satellites can be analysed, too. The algorithm detects and quantifies hot sources, including gas flares, using the night-time readings of the shortwave 10 infrared (SWIR), mid-infrared (MIR) and thermal infrared (TIR) channels. SLSTR observations (night-time overpasses at 10:00 PM) complement those of the VIIRS instruments (0:40 AM on JPSS-1/NOAA-20 and 1:30 AM on Suomi-NPP) by filling observation gaps in the time series. Both instruments are planned to allow long term data availability.
Here, we present a new dataset of global GF volumes (Billion Cubic Meters, BCM) and BC emissions (g), which is based on the observations by Sentinel-3A in 2017. Section 2 describes the newly developed methods for identifying gas flares among 15 the observed hot sources, correcting for intermittent observations opportunities, and dynamically determining appropriate BC emission factors from the observations. The results are presented in Section 3 and the conclusions are summarised in Section 4.

Hot spot detection and characterization
The steps involved in the study presented here are outlined in Figure 1. We base our work on the persistent hot spots detection 20 and characterisation algorithm described in Caseiro et al. (2018) to process a full year (2017) of Sentinel-3A's Sea and Land Surface Temperature Radiometer (SLSTR) data. A hot event at temperatures typical of a gas flare will produce a local maximum in the night-time readings of the shortwave-and mid-infrared (SWIR and MIR) channels of SLSTR. The SWIR band centered at 1.61 µm (S5) is closest to the expected spectral radiance maximum and serves as the primary detection band. Hot SWIR and MIR pixels are searched for using a contextual approach and their radiances are extracted. The radiometrically calibrated 25 radiances of the different bands are used to fit the sum of two Planck curves: one representing the hot source and the other the cool background. Both curves are fitted to the night-time spectral infrared (short-wave, mid-wave and thermal infrared: SWIR, MIR and TIR) observations to characterize temperature (T) and area (A) of the observed gas flares. Detections are divided into spurious signals (if the temperature retrieval is outside the 500-5000 K range) and valid hot spots. Hot spots are further divided into high accuracy (detection in at least another band besides S5, i.e. S6 or S7 or F1, meaning that with the LWIR there are at 30 east 4 points for the Planck curve fit, and at least 3 cloud-free pixels in the background to compute the background radiances) and low-accuracy. Subsequently, the radiative power (RP) of the flare is calculated with the Stefan-Boltzmann equation. Figure 1. Flowchart of the methodology used in the present work. Starting point is the hot spot algorithm detection described in Caseiro et al. (2018) which, after applying a filter on the retrieved temperature (500 K < T < 5000 K), produces a database of valid hot spots detections.
The valid hot spot detections are further filtered: detection in at least another band besides S5 (S6 or S7 or F1, meaning that with the LWIR there are at east 4 points for the Planck curve fit, PCF) and at least 3 cloud-free pixels in the background to compute the background radiances.
This step produces a database of high-accuracy hot spots detections. For each of these high accuracy hot spot detections, the activity (Billion Cubic Meters, BCM) and Black Carbon (BC) emissions are computed. All hot spot detections are gridded into a global 0.025 • × 0.025 • grid.
For each grid cell, the following quantities are computed: the number of valid detections (n Obs ), of high-accuracy detections (n Obs HA ) and of days of operation (nOps) (see below, Section 2.2.2), temperature, area and radiative power (minimum, average, maximum and standard deviation), the date of the first and the last detection and the sum of the activity (BCM) and BC emissions. A volcano filter is applied to the global grid in order to mask out volcanic activity (see Section 2.2.1). The last step is the discrimination of grid cells for flaring activity (see Section 2.2.2). While in principle the methodology used is based on VNF developed for Visible Infrared Imaging Radiometer Suite (VIIRS) (Elvidge et al., 2013(Elvidge et al., , 2016, it differs by (1) analysing the radiances of clusters of contiguous hot pixels instead of treating spatial maxima as individual pixels and (2) using the TIR channels when fitting the sum of the two Planck curves. By considering all the hot pixels and therefore using all the information on the spatial scale we expect our representation to be more realistic for the characterisation of large gas flare arrays. The use of the TIR channels implies a more stringent constraining of the 5 background temperature. These options may lead to the retrieval of a lower flaring temperatures than those retrieved using the VNF approach (Caseiro et al., 2018). The lack of ground truth hinders the evaluation of the accuracy of the different options. Caseiro et al. (2018) already tested the method within a limited time-span using oil and/or gas producing regions and compared the results to the VNF "flares only" product. The results showed good agreement of the hot source detection when investigating persistent hot spots with the advantage of the Sentinel-3A's SLSTR algorithm in detecting and quantifying smaller 10 flares, due to the night-time availability of a second SWIR channel. The characterisation in terms of temperature, area and radiative power reached similar values. However, temperatures were slightly lower, while areas and radiative power were slightly larger, in the SLSTR-based results. Note that VIIRS Nightfire also uses two SWIR channels since January 2018 and that the detections are now conducted by two VIIRS instruments (at 0:40 and 1:30 AM).

Volcano filter
In order to filter out volcanic activity from, we use the data available from the Global Volcanic Program of the Smithsonian institution (Venzke, 2013). A mask at 0.025 • resolution was calculated from all eruptions in 2017. It also contains buffering zones stretching to 0.25 • distance from the volcanic center coordinates because these coordinates may not correspond to the location of the thermal radiation emission: Many volcanoes do not consist of a single edifice and many individual eruptive 20 fissures through which lava erupts may be present in a volcanic field (Siebert et al., 2010).

Discrimination of gas flares from other industrial hot sources
In order to discriminate gas flares from other night-time hot spots we investigate persistence and retrieval temperature filtering.
We project all the hot spot detections onto a 0.025 × 0.025 degrees global grid. This is a spatial resolution that reflects the reality on the ground, where several flares, separated by tens to hundreds of meters, may be operated within a single oil and gas 25 facility, while facilities are tens of kilometers apart. This resolution is also adequate to the pixel footprint of medium-resolution imagers (Facchinelli et al., 2019). The following quantities are then computed for each grid cell: the number of hot spot detections (n Obs ), the number of high-accuracy (Caseiro et al., 2018) hot spot detections (n ObsHA ), the minimum, maximum, average and standard deviation of the high-accuracy hot spot retrieved temperature (T ), area (A) and radiative power (RP ) as well as the earliest and the latest hot spot observation date. 30 The temperature value used in the selection process of the discrimination strategy is based on Elvidge et al. (2016) and on the recent work by Liu et al. (2018), who derived gas flaring temperatures of 1000 K to 2600 K from the VIIRS Nightfire database, depending on the type of operation (shale oil or gas, offshore, onshore or refinery). Most of the gas flares display temperatures between 1650 K and 1850 K. However, temperatures can occasionally be as low as 1300 K. We therefore test for the minimum and/or for the maximum temperature for all the high-accuracy detections within a grid cell (T min and T max , respectively). The temperature range reported by Elvidge et al. (2016) and Liu et al. (2018) overlaps with particularly hot detections from the coal chemical industry and steel plants. Therefore, additional criteria are needed for identifying gas flares in the hot source dataset.

5
In order to select the discriminating strategy we test several subsets of the gridded high-accuracy hot spot database. For each of the 8 subsets described in Table 1, a sample of 100 random onshore grid cells complying to the defined thresholds have been tested by examining high-resolution imagery (Google Earth) and the locations are classified into four categories: The user's accuracy is the accuracy from the point of view of the user. This metric represents the frequency with which a class on the map corresponds to the ground truth. In our case, UA is lies between at least the verified flaring locations and at most the sum of the verified and the likely flaring locations. The commission error is calculated by reviewing the classified sites for incorrect classifications. In the present work, the commission error is lies between at least the locations without any 20 industry or infrastructure and at most the sum of the locations without any industry or infrastructure and the unlikely flaring locations.
In some industrial areas, facilities that use gas flares may be close to other hot spots, such as iron smelters or steel mills. In those cases, a commission error occurs when the flare is off and our methodology samples the other hot spot. The omission error can be divided into two categories: flares that the hot spot detection and characterization algorithm failed to detect and 25 flares that were detected as hot spots but which the discrimination strategy left out. The former will be the same for any discrimination procedure considered.
Detections located in grid cells where the yearly maximum temperature retrieval is above 1500 K and a persistency and quality criterion (n ObsHA > 5) is met are considered as originating from gas flares and called flaring locations hereafter. This Table 1. User's accuracy (UA, %) and commission error (C, %) of the hot spot discrimination strategies considered. n Obs is the number of hot spot detections within a grid cell, n ObsHA is the number of high-accuracy hot spot detections within a grid cell, Tmin is the minimum temperature retrieved among all the hot spots detected within a grid cell, Tmax is the maximum temperature retrieved among all the hot spots detected within a grid cell, n cells is the number of grid cells that comply with the thresholds. In order to discriminate gas flares from other hot spots we discriminate hot spots based on their persistency (n Obs and n ObsHA ) and on their temperature time series (Tmin and Tmax).
We have tried 8 combinations (discrimination strategies) of thresholds on those variables. Each column represents a tested discrimination strategy. For each of the 8 combinations, we examine high-resolution imagery for 100 random onshore locations (800 in total) in order to verify the presence of a gas flare. The goal is to maximize user's accuracy (UA) and minimize commission error (C) while minimizing the omission error (here, the variation in n cells is used as a proxy). The discrimination strategy #5 was selected as the most suitable.
results in a relatively large number of grid cells with detections, i.e. low omission error, combined with a high user's accuracy and a low commission error of 85±11% and 3±1%, according to Equations 1 and 2, respectively. Inserting a limitation on the minimum temperature or increasing the maximum temperature or the number of high-accuracy observations would not significantly increase the user's accuracy or decrease the commission error, but would reduce the number of grid cells complying with the discrimination criteria by several hundreds, which we interpret as an increase in the omission error. Although our 5 discrimination strategy is satisfactory at the planetary level, global aggregating approaches may miss individual occurrences, in our case especially where the oil exploitation frontier is under expansion (Facchinelli et al., 2019).

Determination of flared volumes and black carbon emissions
In order to compute the activity and emissions, we estimated the number of days of operation per site by expressing the maximum number of observations as a function of latitude (in 10 • bins, see Figure 2). The function computes the maximum 10 number of observations n Obsmax per grid cell for a given latitude which we assume expresses a continuous hot spot (365 days a year). The number of days of operation n Ops is then estimated by scaling following Equation 3. By this scaling, we approximately correct for the limitations of gas flaring detection from space (cloud cover and overpass frequency).
n Ops = n Obs × 365 n Obsmax (3) For the estimation of the flared volume, we applied the calibration derived by Elvidge et al. (2016) to each single detection within a flaring location grid cell. The calibration relationship uses a modified formulation of the radiative power to output the yearly flared volumes (in Billion Cubic Meters, BCM), which is then scaled to a daily activity in m 3 .
The lower bound BCM min for the activity estimate assumes that the flare is only active on the n Obs days with hot spot detections. The upper bound for the activity estimate assumes that the flare is constantly active: The best activity estimate corresponds to assuming that the flare is active for n Ops days: The emissions of black carbon (BC) from gas flares are estimated using reported emissions factors (EFs). Klimont et al. are associated with higher BC emissions. Between the two extremes, the BC emission is scaled linearly as a function of the 10 flaring temperature (see Figure 3). To the best of our knowledge, this is the first time that operating practices are taken into consideration when assigning the EF. The country-level BCM and BC estimates are computed by summing the individual flares estimates within the borders of each country and its exclusive economic zone.
With this methodology we estimate a wide range of possible activity (BCM) and BC emissions (g), where our best estimate falls between the 'flaring only when there is a detection' (BCM min and BC min ) and 'constant flaring' (BCM max and BC max ). We consider the range between BCM min and BCM max to be the confidence interval for the possible activity of a 5 flaring site. Similarly, [BC min , BC max ] is the confidence interval for the possible emission.

Flaring locations
All the hot spots detected globally for 2017 using the algorithm as described in Figure 1 are shown in Figure 4. Their classification is shown in Figure 5 10  Noise, such as detections in the open ocean and the South Atlantic anomaly, is filtered out when considering only the high-accuracy hot spots as defined by Caseiro et al. (2018). There are 427202 high-accuracy hot spots in the dataset. Their distribution pattern suggests that they encompass a large range of hot sources, such as wildfires, volcanoes and industrial sources.
After gridding, filtering out active volcanoes (71 volcanic eruptions in 2017) and retaining the grid cells passing the GF 5 criteria, the number of globally detected flaring sites in 2017 is 6232. This nubmer is lower than the estimate of 7467 by Elvidge et al. (2016). Although the SLSTR-based methodology is able to detect smaller gas flares due to the availability of the second SWIR channel, the VIIRS Nightfire algorithm is expected to detect a larger number of hot spots because it considers local maxima, whereas our methodology clusters adjacent hot pixels, and the instrument has a wider swath, i.e. more detection opportunities.

Flaring characteristics
Statistics on retrieved temperature and radiative power are reported in Figures 7 to 9. 20 The average temperature at the flaring locations approximately ranges from 950 K to 2250 K. This is slightly smaller than the range 750-2500 reported by Liu et al. (2018), who used VIIRS Nightfire data. This conforms to the findings of (Caseiro et al., 2018). The distribution of average retrieved temperatures confirms the bi-modal distribution with modes around 1750 K and 1200 K that has also been observed by VIIRS.
The geographical distribution of the average retrieved flaring temperature and radiative power are shown in Figures 8 and 10, 25 respectively.
The average radiative power at flaring locations range from a few tenths of MW up to approximately 1 GW (Figure 9), spanning the 5 orders of magnitude reported by Elvidge et al. (2016).
There were 197610 detections at flaring locations in the SLSTR dataset, most were clusters of up to 20 S5 pixels, as shown in Figure 11.

Flared volumes
The activity for all the individual flaring locations is shown in Figure 12. The distribution is very close to that reported by Elvidge et al. (2016), who additionally find a few extremes in the range 0.8-1.1. The most active flare burned 0.623 BCM in 2017 and is located south of Punta de Mata in Venezuela. This is also where Elvidge et al. (2016) found the most active flare for 2012, though more active (1.13 BCM). The total activity of the region is shown in Figure 13, and the cumulative activity 5 throughout the year in Figure 14.

BC emissions
The black carbon (BC) emissions for all the individual flaring locations are shown in Figure 16. The most active flare emitted 0.35 Gg BC in 2017 (south of Punta de Mata). The cumulative BC emissions throughout the year for the whole Persian gulf region is shown in Figure 17. Our best estimate for the global BC emissions from gas flaring is 73 Gg (lower estimate: 20 Gg, upper estimate: 239 Gg).

5
As for the flared volume, 4 countries account for more than half of that value: Iraq (11 Gg), Iran (11 Gg), Russia (10 Gg) and Algeria ( Figure 19 shows all Nightfire flaring detections in 2017 projected onto the same 0.025 • × 0.025 • global grid as used for our results. The original VNF database lists a total of 10825 observations.

5
After the projection, these produced 10185 flaring locations. This is an indication that the gridding spatial resolution used is adequate.  Both methodologies retrieve temperatures roughly in the same range and the correlation is reasonably good with R 2 = 0.34.
The SLSTR-based method retrieved temperature are more spread out over the values' range than in the VNF product though.
The BCM retrievals are well correlated with R 2 = 0.74. This shows that the differences, and presumably errors, in the temper-5 ature and flare area retrievals partly compensate each other. The regerssion line shows that the BCM in the VNF dataset is on average 16% larger than the one retrieved by the SLSTR-based method. We interpret the large difference in the number of detections as the effect of several factors: (1) small geolocation errors,

10
(2) clustering versus local maxima detections, (3) difference in the overpass time between platforms. In Table 2 we attempt to quantify the importance of these factors.
Distances up to one grid cells (adjacent cells) may arise from small geolocation inaccuracies. This accounts for 2562 SLSTRbased locations and for 1507 VNF locations.  The Nightfire algorithm considers pixels which are local maxima and our methodology aggregates contiguous hot pixels within a unique cluster. These clusters may be very large, contain more than one local maximum and spread over grid cells, but only one cell is finally considered as its location. This was already indicated in Caseiro et al. (2018), where data retrieved from a natural gas and condensates production site in the Yamal peninsula (northern Siberia) was compared between VIIRS, SLSTR and HSRS (Hot Spot Recognition System). In that particular case, the methodologies relying on clustering of hot pixels 5 (SLSTR and HSRS) reported fewer gas flaring sites than VIIRS Nightfire. In the current database, about half of the 197611 hot spots at flaring locations were clusters of up to 20 S5 pixels. At nadir, a S5 pixel has a ground footprint of 500×500 square meters. Therefore, 20 S5 pixels span a distance of up to 10 km, which corresponds to, at most, about 4 grid cells. Distance between VNF flaring locations which are not adjacent and less than 4 grid cells away from the closest SLSTR-based location may therefore arise from the clustering methodology. This is the case for 73 SLSTR-based locations and 1651 VNF locations.

10
The remaining 633 SLSTR-based locations and 4063 VNF locations were further apart. 2222 of the 4063 distant VNF grid cells had at least 1 SLSTR-based detection, 1964 of which had the maximum retrieved temperature above 1500 K, see Figure 22. Regarding the other criteria we use to classify a grid cell as a flaring location, only 159 grid cells (out of the 2222 distant VNF locations that had at least one SLSTR-based detection) had at least 5 high-accuracy SLSTR-based detections, see Figure 23. 15 We thus conclude that at VNF distant locations there are not enough SLSTR detections to be considered as flaring by our procedure. The difference in the overpass time between the sensors may play some role: the Suomi-NPP platform observes later during the night, thus possibly having more opportunities to detect gas flares at night. However, the effect of the different swaths of the instruments is likely more important. VIIRS has a much larger swath (3040 Km against 1420 Km for SLSTR) and has therefore more opportunities of detection. This is more important if the flare exhibits little activity and, indeed, the  The very large majority of the VNF-based activity was detected at locations coincident with SLSTR-based locations (87% at coincident locations, 92% at coincident or adjacent locations). 98% of the SLSTR-based activity was detected at coincident or adjacent locations. This shows that both methodologies detect similar global levels of activity (151 BCM computed by VNF and 129 BCM for the present work) and this at the same locations, despite the large discrepancy in the number of flaring locations.
5 Table 2 shows that the difference comes mostly from coincident and adjacent flaring locations, which indicates that the computation of the activity levels is more likely the source of the discrepancy (22 BCM globally) rather than the detection itself.
One possibility is the scaling we use to obtain the annual activity from the individual detections (see Section 2, Equation 3 and Figure 2). Another source of discrepancy could be the lower temperature retrieved by our method, which considers the TIR for the Planck curve fitting.
10 Figure 24 shows the global distributions of the average retrieved flaring temperatures. The generally lower temperature in the present work can be traced back to the clustering methodology we take and was discussed in our previous paper (Caseiro et al., 2018). The use of the TIR channels in the dual Planck curve fitting may also lower the retrieved hot source temperature.  3.6 Comparison with the SWIR-radiance method Fisher and Wooster (2018) developed a methodology to derive the radiative power from flares using a single SWIR channel, analogous to a previous methodology which used the MWIR radiance to derive the fire radiative power in the case of landscape fires (Wooster et al., 2005). The methodology is based on the near-constant ratio between the SWIR spectral radiance and the total emitted radiance at typical flaring temperatures (i.e. 1600-2200 K). The radiative power can be computed, using a sensor-5 specific optimized relationship and without the input of the emitter's temperature, with a theoretical precision of 13.6 %.
Fisher and Wooster (2019) used the SWIR-radiance methodology to derive a global time series of flaring activity using (A)ATSR and SLSTR data. For SLSTR, the data used spanned from May 2017 to May 2018. The identification of nighttime thermal hotspots was conducted using a simple static threshold: pixels having a radiance above the instrumental noise level were considered as hot spots. The thermal hotspots are subsequently gridded onto a 1-arcminute global geographic grid 10 (approximately 2 km). For SLSTR, a grid cell is considered as a flaring location if there were at least two different months with hotspots observations. The authors report 10428 locations, with a global distribution very similar to what we report here.
We apply the SWIR-radiance approach developed for SLSTR's S5 SWIR channel by Fisher and Wooster (2019) to our results (197611 valid detections at flaring locations). We limit the comparison to detections for which the retrieved temperature falls in the range 1600-2200 and which were limited to one S5 pixel in order to comply with the pre-requisites of the SWIR-radiance 15 approach. Figure 25 shows the comparison between the RP derived from the SWIR-radiance method and the Planck-curve fitting followed by the application of the Stefan-Boltzmann equation. There is a good correlation, although the SWIR-radiance methodology outputs radiative power values about 25 % lower than those from this work. The trend is more recognizable at lower temperatures, which might indicate that the relationship between the SWIR spectral radiance and the total emitted radiance 20 deviates more from linearity at lower temperatures. Figure 25. Comparison between the SWIR-radiance method (Fisher and Wooster, 2018) and the SLSTR-based Planck curve fitting. The method derived by Fisher and Wooster (2018) and applied to SLSTR by Fisher and Wooster (2019) was applied to the hotspots detected by our method at flaring locations. Only detections with one pixel and whose retrieved temperature was in the range considered valid for the SWIR-radiance method (between 1600 and 2200 K) were considered. The continuous line represents the 1:1 relationship. The dashed line is the regression line between the methods.
In this work we apply a previously described hot spot detection and characterisation procedure (Caseiro et al., 2018) to the 2017 observations of Copernicus Sentinel-3A's Sea and Land Surface Temperature Radiometer (SLSTR) instrument. We adapt the procedure most commonly used to discriminate gas flares (based on two characteristics of gas flares: persistence and temperature). This allows us to refine the characterisation procedure in order to separate gas flares from other sources of 5 heat emissions. Validation through referencing with high-resolution images yields a detection accuracy of 85±11%, with a commission error of 3±1%.
Our methodology detects 6232 flaring sites worldwide in 2017. Over half of these are located in five countries only: Russia, the United States, Iran, China and Algeria.
Additionally, we calculate the volume of flared gas based on the previously observed relationship between the flared volume  30 We also compare our activity estimates to the SWIR-radiance method. Both methodologies show a good agreement for high-temperature detections, but a constant under-estimation by the SWIR-radiance method for lower temperature detections.
We further quantify the black carbon (BC) emissions due to gas flaring. We assume that temperature is an indication of the completeness of combustion and use the retrieved flame temperature to determine the emission factor for each single detection, bounded by the range of previously published emission factors. We assume that a temperature close to the adiabatic flame temperature corresponds to the lower bound of the emission factor range considered, while the lowest temperature corresponds to the upper bound. The minimum, maximum and best estimates for the BC emissions are computed in the same way as the flared volume. Our resulting best estimate for the emission of BC to the atmosphere by gas flaring in 2017 is 73 Gg with a confidence interval of [20,239] Gg, with Iraq, Iran, Russia and Algeria being responsible for roughly one half of those 5 emissions. The most active flaring location, which is also the one with the highest emissions, is estimated to have yearly emission of 0.35 Gg of BC in 2017.
This study shows that the SLSTR instruments on-board the Copernicus Sentinel-3 satellites are well suited for the quantitative characterization of the gas flares in terms of flame temperature, size and radiative power, as well as BC emissions.
This will allow to increase the detection opportunity for gas flares in an observation system that comprises SLSTR and VIIRS 10 instruments.

Data availability
The flaring activity and the related black carbon emissions product are available with 0.