Fresh snow samples were collected along a north–south transect between Tomsk and the Yamal coast in February–March 2014 (23 samples, Table S1 in the Supplement), while in March 2015 sample collection took place on the Kindo peninsula and near the port of Arkhangelsk in the White Sea (11 samples, Table S1). Finally, in February–May 2016 samples were collected on the Kindo peninsula, in Arkhangelsk, and between Tomsk and Yamal (20 samples, Table S1). These areas have been reported to receive pollution from both urban and gas flaring sources (Stohl et al., 2013). For example, the gas flaring sources located in Yamal and Khanty-Mansiysk (Russia) are in the main pathway along which sub-Arctic air masses travel to the Arctic (Stohl et al., 2006). All sampling points were located more than 500 m away from roads to minimize the direct influence from local traffic emissions. Information about sample collection such as the location of sampling, the amount of snow collected and the depth at which snow was sampled is reported in Table S1, and the sample locations are plotted in Fig. 1.

Sampling was performed using a metal-free technique using pre-cleaned plastic shovels and single-use vinyl gloves. Samples were stored in polyethylene bags that had been thoroughly washed with 1 M HCl and rinsed with abundant deionized ultrapure water in the laboratory prior to their use. After returning the samples to the laboratory, the snow was allowed to melt at ambient temperature (18–20 ∘C) and immediately filtered through quartz 47 mm fibre filters (2500QAT-UP Pall for samples collected in 2014 and QM-A Whatman for samples collected in 2015 and 2016). The filters were dried at 60–70 ∘C, wrapped in aluminum foil and stored in a refrigerator. Quartz fibre filter collection efficiency of BC in liquid samples can be less than 100 % (Hadley et al., 2010; Ogren et al., 1983). To what extent this has affected the levels reported in the present study is unknown. Thus, the results presented should be regarded as conservative estimates based on the assumption that some BC might have been lost during filtration.

EC content of the filters was measured at NILU's laboratories with TOA, using the Sunset laboratory OC–EC instrument operated according to the EUSAAR-2 protocol (Cavalli et al., 2010). A 1.5 cm2 punch was cut from the filtered snow samples for the analysis. Transmission was used for OC charring correction. Performance of the OC–EC instruments is regularly inter-compared as part of the joint European Monitoring and Evaluation Programme (EMEP) Aerosols, Clouds, and Trace gases Research InfraStructure Network (ACTRIS) quality assurance and quality control effort (Cavalli et al., 2016).

The content of carbonate (CO32-)–carbon on the filters was measured with TOA, following thermal-oxidative pretreatment based on the approach described by Jankowski et al. (2008). A punch of 1.5 cm2 from each filter was heated at 450 ∘C for 2 h in ambient air to remove OC and EC but not CO32-–carbon. The filter punch was subjected to TOA immediately (30 s) after thermal-oxidative pretreatment. The split time (between OC and EC) obtained for each filter punch used to determine the filter samples' content of EC (Sect. 2.2) was also used to apportion CO32-–carbon to OC and/or EC. The influence of CO32-–carbon evolving as EC was accounted for using the following equation: ECCO32-corr=EC-ECCO32-, where ECCO32-corr is EC corrected for CO32-–carbon that evolved as EC during TOA, EC is elemental carbon and ECCO32- is CO32-–carbon that evolved as EC during TOA. Applying this correction, EC values were 5–22 % lower (see Supplement).

The concentrations of BC in snow were simulated with the LPDM FLEXPART version 10 (Stohl et al., 1998, 2005). The model was driven with operational meteorological wind fields retrieved from the European Centre for Medium-Range Weather Forecasts (ECMWF) of 3 h (for the years 2014 and 2015) and 1 h (for the year 2016) temporal resolution. The ECMWF data have 137 vertical levels and a horizontal resolution of 1∘ × 1∘ for the 2014 and 2015 simulations and 0.5∘ × 0.5∘ for the 2016 simulation.

The simulations were conducted in backwards time (retroplume) mode, using a new feature of FLEXPART to reconstruct wet and dry deposition with backward simulations (Eckhardt et al., 2017). This new feature is an extension of the traditional possibility of simulating atmospheric concentrations backward in time (Seibert and Frank, 2004; Stohl et al., 2003). It is computationally efficient because it requires only two single-tracer transport simulations (one for wet deposition, one for dry deposition) for each measurement sample. To reconstruct wet deposition amounts of BC, computational particles were released at altitudes of 0 to 20 km at the locations where snow samples were taken, whereas to reconstruct dry deposition, particles were released between the surface and 30 m at these locations. All released particles represent a unity deposition amount, which was converted immediately (i.e. upon release of a particle) to atmospheric concentrations using the deposition intensity as characterized by either dry deposition velocity or scavenging rate (for further details, see Eckhardt et al., 2017). The concentrations were subsequently treated as in normal “concentration mode” backward tracking (Seibert and Frank, 2004) to establish source–receptor relationships between the emissions and deposition amounts. The termination time of the particle release was the time at which the snow sample was collected, whereas the beginning time was set as the time when the ECMWF precipitation at the sampling site, accumulated backward in time, was equal to the water equivalent of the snow sample, up to the specified sampling depth.

The model output consists of a spatially gridded sensitivity of the BC deposition at the sampling location (receptor) to the BC emissions, equivalent to the backwards time mode output for concentrations (Seibert and Frank, 2004; Stohl et al., 2003). BC deposition at the snow sampling point can be computed (in mass per unit area) by multiplying the emission sensitivity in the lowest model layer (the footprint emission sensitivity) with gridded emissions from a BC emission inventory and integrating over the grid. The deposited BC can be easily converted to BC snow concentration by taking into account the water equivalent depth of the snow from ECMWF (in millimetres). In the present study, the ECLIPSE (Evaluating the Climate and Air Quality Impacts of Short-Lived Pollutants) version 5 emission inventory (Klimont et al., 2017; Stohl et al., 2015) was used (http://www.iiasa.ac.at/web/home/research/researchPrograms/air/Global_emissions.html). The total emissions of BC from ECLIPSE in the areas of study are shown in Fig. 1a.

BC was assumed to have a density of 2 gm-3 in our simulations and a logarithmic size distribution with an aerodynamic mean diameter of 0.25 µm and a logarithmic standard deviation of 0.3. Each computational particle released in FLEXPART represents an aerosol population with a lognormal size distribution (see Stohl et al., 2005). Assumed aerodynamic mean diameter and logarithmic standard deviation are used by FLEXPART's dry deposition scheme, which is based on the resistance analogy (Slinn, 1982), and they are consistent with those used in other transport models (see Evangeliou et al., 2016; Shiraiwa et al., 2008). Below-cloud scavenging was determined based on the precipitation rate taken from ECMWF. The in-cloud scavenging was based on cloud liquid water and ice content, precipitation rate, and cloud depth from ECMWF (Grythe et al., 2017). The FLEXPART user manual (available from http://www.flexpart.eu) provides more information. All modelling results for this sampling campaign can be viewed interactively at the URL http://niflheim.nilu.no/NikolaosPY/SnowBC_141516.py.

The spatial distribution of EC measured in snow samples from northwestern European Russia and western Siberia is shown in Fig. 1c for each of the campaigns (2014, 2015 and 2016) and is also summarized in Table S2. There was large spatial variability in the distribution of EC in snow in 2014 ranging from 3 to 219 ngg-1, with a median (± interquartile range) of 23 ± 49 ngg-1. The highest EC concentrations in 2014 were observed in western Siberia near Tomsk (147 to 219 ngg-1). FLEXPART emission sensitivities for these samples showed that the air was coming from the north and the east (see in http://niflheim.nilu.no/NikolaosPY/SnowBC_141516.py). This explains the high concentrations of EC, as most of the anthropogenic BC sources are located in these regions. In the rest of the snow samples for 2014, EC concentrations between 4 and 170 ngg-1 were observed. High concentrations were observed near the Ob River coinciding with air masses arriving mainly from Europe. During the 2015 field campaign, EC concentrations were the highest near Arkhangelsk (175 ngg-1), for which FLEXPART showed that the air was coming from nearby areas (http://niflheim.nilu.no/NikolaosPY/SnowBC_141516.py). Therefore, it is likely that the samples were affected by direct emissions from the city or the port of Arkhangelsk. During the same campaign, snow samples collected on the Kindo peninsula (on the White Sea coast) showed high variability in EC concentrations (range: 46–152 ngg-1, median = 70 ± 34 ngg-1). According to FLEXPART emission sensitivities, air masses were transported to Kindo peninsula from central and southern Europe driven by an anticyclone over Scandinavia (http://niflheim.nilu.no/NikolaosPY/SnowBC_141516.py). Finally, for the snow samples collected outside Arkhangelsk, on the Kindo peninsula, and close to the Yamal Peninsula in western Siberia in 2016, EC concentrations ranged between 7 and 161 ngg-1 (median = 40 ± 47 ngg-1). Outside Arkhangelsk, EC concentrations varied widely from 31 to 161 ngg-1 with a median concentration in this region of 61 ± 43 ngg-1. This is far below the 175 ngg-1 observed in 2015, although there was only one sample collected in that year. On the Kindo peninsula, EC was relatively constant in 2016, ranging between 25 and 35 ngg-1 (median = 28 ± 4 ngg-1), which is more than 60 % lower compared with the 2015 values (median = 70 ± 34 ngg-1). Finally, between Tomsk and Yamal, EC concentration was highly variable (7–119 ngg-1) due to the different EC sources affecting snow (median = 50 ± 34 ngg-1). For instance, it is expected that gas flaring affects snow close to Yamal, while snow collected in the south (Tomsk) is likely influenced by sources in Europe or local urban emissions. Nevertheless, the highest concentrations (>100 ngg-1) were observed north of 68∘ N, on the Yamal Peninsula.

We compared the measured EC concentrations in the snow samples with those calculated by FLEXPART. For this, the emission sensitivities were multiplied with the total emission fluxes from ECLIPSE (Sect. 2.4). A scatter plot of modelled and measured snow concentrations is presented in Fig. 1b. The results show a good correlation between modelled BC and measured EC concentrations for the 2015 and 2016 campaigns (R2015=0.83 and R2016=0.68, p value < 0.05), but weaker correlation for 2014 (R2014=0.53, p value < 0.05). The FB for individual samples is shown in Fig. S1. The MFB of the model for the 2014 snow measurements was -42 %, which shows that the model underestimated observations. In total, the model underestimated concentrations by 30–168 % for 17 out of 23 samples, whereas for the rest (six samples) FB values ranged between 20 and 148 % (median MFB: -56 ± 72 %) (Fig. S1). In 2015, the model underestimated observations by 48 % (median MFB: -56 ± 29 %) for 11 out of 12 samples (FB between -101 and -7 %, while one value was found to be 12 %). For 2016, FB values of the simulated concentrations of BC in snow showed another set of underestimation (median: -13 ± 60 %) between 0.3 and 198 % for 12 out of 19 samples. For the remaining seven samples, the model predicted higher concentrations compared with observations (10 to 75 %) (Fig. S1). RMSE values were estimated to be quite high, between 37 and 49 ngg-1, due to the large variation in the observed EC concentrations.

The levels of EC in snow presented here are relatively high compared to previously reported concentrations in the Arctic. Apart from Aamaas et al. (2011), who measured maximum EC concentration in snow close to the airport of Svalbard of more than 1000 ngg-1, most of the reported levels of EC in the relevant literature are close to our findings. For instance, Ruppel et al. (2014) found that EC concentrations have been increasing up to 103 ngg-1 since 1970 in Svalbard. McConnell et al. (2007) reported that the BC concentrations measured at the D4 ice-core site in Greenland were 10 ngg-1, at maximum, which most likely originated from BB in the conifer-rich boreal forest of the eastern and northern United States and Canada. Forsström et al. (2013) reported concentrations as high as 88 ngg-1 in Scandinavia and lower ones at higher latitudes (11–14 ngg-1 in Svalbard, 7–42 ngg-1 in the Fram Strait and 9 ngg-1 in Barrow). Svensson et al. (2013) collected snow samples from Tyresta National Park and Pallas-Yllästunturi National Park in Sweden. Tyresta is a relatively polluted site located circa 25 km from the city centre of Stockholm with a population of about 2 million people. Pallas-Yllästunturi National Park is located in Arctic Finland and a clean site with no major city influencing the local and regional air. The concentration of EC in Pallas-Yllästunturi was between 0 and 140 ngg-1, while in Tyresta the BC concentrations were up to more than 7 times higher (53–810 ngg-1). Furthermore, Doherty et al. (2010) in the most complete dataset for the Arctic snow and ice BC reported highly variable concentrations (up to 800 ngg-1) for 5 consecutive years (2005–2009). Finally, in the most recent dataset for snow BC, Macdonald et al. (2017) reported BC concentrations ranging from 0.3 to 15 ngg-1 for the samples collected near the Alert observatory (see Sect. 4.1).

We further analysed the model output in order to calculate relevant contributions from various BC source types to BC concentrations in snow (for method description, see Sect. 2.4). ECLIPSE emissions include waste burning (WST), industrial combustion and processing (IND), surface transportation (TRA), power plants, energy conversion, and extraction (ENE), residential and commercial combustion (DOM), and gas flaring (FLR) while BB emissions were adopted from the Global Fire Emissions Database, version 4 (GFEDv4.1) (Giglio et al., 2013). The results are depicted in Fig. 2 for the sampling campaigns of 2014, 2015 and 2016 in western Siberia and northwestern European Russia, sorted from the northernmost to the southernmost sampling location.

In 2014, TRA contributed about 18 %, on average, to the simulated BC in snow, DOM 28 % and FLR 44 %, whereas ENE and IND were less significant. Maxima of TRA, DOM and FLR contributions were observed at a latitude of about 65∘ N, where measured EC and modelled BC were similar. An example of the contribution from the aforementioned dominant sources to snow BC concentrations for the highest measured EC concentration in snow is shown in Fig. 3. The transport sector includes emissions from all land-based transport of goods, animals and persons. It is more significant in southern Russia and close to the borders with Kazakhstan and Mongolia, where a large number of major Russian cities (e.g. Moscow, Kazan, Samara, Yekaterinburg, Tomsk, Novosibirsk, Krasnoyarsk) are located and connected with each other by federal highways. Residential and commercial combustion includes emissions from combustion in households and public and commercial buildings. Therefore, it is expected to be high for areas that consist of large population centres (Fig. 3). FLR emissions were found to contribute the most in this example, with a total concentration from this sector of 19.7 ngg-1 (compared with 12.6 and 16.5 ngg-1 in TRA and DOM, respectively) (Fig. 3).

On the Kindo peninsula and in Arkhangelsk, where snow sampling took place in 2015, the main contributions to snow BC were from DOM (47 %), TRA (30 %), BB (7 %) and FLR (6 %) (see Fig. 2). Similar to EC measurements in snow, simulated BC was also higher than in 2014, as the sampling sites were located closer to strong sources in Europe (Kindo) and close to a populated area (Arkhangelsk) with a strong regional impact. The highest concentration of EC was observed on the Kindo peninsula (33.13∘ E–66.53∘ N). Figure 4 shows the spatial distribution of emissions that contributed to simulated snow BC at the sampling point where the highest BC concentration was observed. In this case, TRA and DOM emissions from Europe mostly affected snow on the Kindo peninsula whereas FLR emissions were very low due to the long distance from the sampling point. Emissions from an unusual late winter–early spring episode of BB at the borders of Belarus, Ukraine and Russia also affected BC concentrations in snow in northwestern European Russia (Fig. 4). The importance of episodic BB releases in Russia, the miscalculation of satellite-retrieved BB emissions and their impact on Arctic concentrations in early spring has been explained by Evangeliou et al. (2016) and Hao et al. (2016). BB emissions, originating mostly from eastern Europe, contributed about 19.4 ngg-1 to the snow concentration at the receptor point (Fig. 4). TRA and DOM emissions were the dominant sources for this sampling point, contributing 33.6 and 47.2 ngg-1, respectively (Fig. 4).

Finally, in 2016, when samples were collected at the Kindo peninsula, in Arkhangelsk and in Yamal, DOM, FLR and TRA contributed, on average, 31, 29 and 27 %, respectively (see Fig. 2c). Similar to the measured EC concentrations in snow, simulated concentrations of BC in 2016 were lower than those in 2015, on average. The highest measured EC concentration was observed in the Khanty-Mansiysk region (72.94∘ E–65.36∘ N), which mirrors the simulated BC concentration at the same point very well. The much higher contribution from TRA at this sampling point (38.6 ngg-1) (Fig. 5b) is attributed to emissions from southern Russia (e.g. Tomsk), where all the main cities in Russia are located. Another large fraction of TRA emissions comes from central and eastern Europe (see also in http://niflheim.nilu.no/NikolaosPY/SnowBC_141516.py). Similar to TRA, emissions from DOM were mostly transported to Khanty-Mansiysk from central and eastern Europe, as well as from Turkey, contributing 36.6 ngg-1 (Fig. 5). As previously mentioned, the sampling point where the highest EC concentration was measured is located inside the largest gas flaring region of Russia. In addition, the corresponding emission sensitivity maps showed that the air was coming from the south passing directly through this high-emission region, making FLR emissions the highest contributing source (88.8 ngg-1) (Fig. 5).

In this section, we present an effort to further validate our model calculations of BC concentrations in snow. For this purpose, BC concentrations in snow that were adopted from Doherty et al. (2010) were compared with modelled BC concentrations in snow that were simulated with FLEXPART as described in Sect. 2.4. Samples were collected in Alaska, Canada, Greenland, Svalbard, Norway, Russia and the Arctic Ocean during 2005–2009, on tundra, glaciers, ice caps, sea ice, frozen lakes and in boreal forests. Snow was collected mostly in spring, when the combination of snow cover and exposure to sunlight was at a maximum and before the snow had started to melt. Samples of melting snow collected in the summer of 2008 from Greenland and from Tromsø, Norway, were removed from the study, as we have no knowledge about the depth of the melt layer and effects of the percolation of meltwater through the snowpack. All samples were collected away from local sources of pollution. In many locations (Canadian Arctic, Russia, Greenland, Tromsø and Ny-Ålesund) samples were gathered at different depths throughout the snowpack, giving information on the seasonal evolution of BC concentrations as the snow accumulated (and/or sublimated) throughout the winter. In these cases only the surface BC was taken into account. The snow was melted and filtered, and the filters were analysed in a specially designed spectrophotometer system to infer the concentration of BC (for more information see Doherty et al., 2010). In contrast to our findings for the origin of snow BC in the Russian Arctic, a source apportionment analysis performed on the 2008 and 2009 measurements (Hegg et al., 2010) from this dataset showed that the dominant source of BC in the Arctic snow pack was BB. Specifically in eastern Siberia biomass burning of crops and grasslands contributed more snow BC in high latitudes than boreal forest fires, in contrast to the Canadian Arctic.

A comparison of modelled (FLEXPART) and measured BC concentrations (Doherty et al., 2010) in snow is depicted in Fig. S2. The model captures snow BC concentrations relatively well in most of the Arctic regions except for the Canadian Arctic, where the modelled concentrations of snow in 2007 were significantly higher. Samples from the same region in other years showed moderate agreement with modelled values. Similar to our finding for the new Russian measurements, the model underestimated deposition by 51 %. The RMSE was estimated to be 52 ngg-1, which is acceptable considering that the variation in snow concentrations in the dataset ranged from 0.3 to 783 ngg-1. The highest measured concentrations of snow BC were observed in Russia, where the model showed a good spatial agreement. For instance, the highest values were obtained in western Siberia, close to the gas flaring regions of the Nenets and Komi oblasts, as well as in southeastern and northeastern Russia, where air masses were arriving from high-emitting sources in southeastern Asia. Lower biases in modelled BC concentrations were observed in northern Siberia with the exception of a few samples at the coasts of the Kara Sea and northeastern Siberia. Furthermore, biased BC concentrations were also observed in Greenland and northern Canada. In western Siberia, BC in snow presented in Doherty et al. (2010) between 2005 and 2009 was 80 ± 63 ngg-1 on average, which is very close to the average value of measured EC obtained from the sampling 2014–2016 campaigns (50 ± 46 ngg-1).

From the total number of samples presented in Doherty et al. (2010) that were used here for validation, only six were collected on the Yamal Peninsula similar to some of the data presented in the current paper. The rest were collected in the Nenets and Komi regions and in eastern Russia and cannot be directly compared with snow EC measurements from the 2014–2016 campaigns. BC concentrations on the Yamal Peninsula in 2007 ranged from 4.1 to 17.6 ngg-1 (median ± interquartile: 10.3 ± 4.9 ngg-1). In the same region, we report EC concentrations to be more than double, varying between 6.6 and 55 ngg-1 (median ± interquartile: 27.8 ± 25.5 ngg-1), whereas there were two samples that showed EC concentrations of more than 100 ngg-1. As mentioned in Sect. 2.1, the sampling of snow for the EC analysis took place more than 500 m away from roads to minimize influence from traffic emissions, while a similar statement is also found in the Doherty et al. (2010) data. It is not clear whether the observed discrepancy arises as a measurement artefact (even though every effort has been taken in both papers to follow a robust protocol) or from real spatio-temporal variation.

Modelled BC concentrations simulated with FLEXPART were also compared with snow BC concentrations from samples collected at the Global Atmosphere Watch observatory at Alert, Nunavut, from 14 September 2014 to 1 June 2015 and they are available in Macdonald et al. (2017). Alert is a remote outpost in the Canadian high Arctic, at the northern coast of Ellesmere Island (82∘27′ N, 62∘30′ W), with a small transient population of research and military personnel. Sampling details and analytical methodologies used for the analysis of BC can be found in Macdonald et al. (2017). BC concentrations in FLEXPART were simulated as in all previous analyses described in this paper (see Sect. 2.4). Time series of simulated and measured BC are depicted in Fig. S3 for the whole sampling period. As before, a correlation coefficient (R) of 0.63 indicates that our model captures the temporal variation in the measured BC in snow. The RMSE was estimated to be almost 63 ngg-1, a relatively high value. The MFB of 47 % indicates a strong overestimation of snow concentrations, although in many samples the opposite was also observed (Fig. S3). This is in contrast to the previous datasets discussed, for which the model underestimated measurements.

Further analysis was carried out to adequately understand the origin of the aforementioned overestimations in the Canadian Arctic in both datasets (Doherty et al., 2010; Macdonald et al., 2017), as they are shown to be rather systematic. For this reason, we have calculated the average footprint emission sensitivities and the average BC contribution from the major sources in ECLIPSE for the 2007 snow samples in the Canadian Arctic and for Alert samples. We have chosen these samples because they were 3 or more times higher than the observations and in this way we can locate the observed overestimations predicted with FLEXPART (Fig. 6).

Regarding the model overestimation for the 2007 samples, the average footprint emission sensitivity showed that the air was coming from continental regions of Canada with a smaller contribution from Scandinavia (Fig. 6). The highest emission sources for these samples were TRA and DOM, which contributed almost 80 % to the snow concentrations, whereas forest fires were less important at the time of sampling. Two hot spots were identified, one along the border of Canada with the US and another, of smaller intensity, in southeastern Asia. A similar emission sensitivity was obtained for the same area of the Canadian Arctic in 2009, only slightly shifted to the north; simulated concentrations were in very good agreement with observations (Fig. S2). This shows that the model overestimation for the 2007 samples is likely attributed to an overestimation of TRA and DOM sources in North America in ECLIPSE for 2007. For the Alert samples, for which the model strongly overestimated BC, the major sources were TRA and FLR, which contributed 55 %, and BB, which contributed about 7 ngg-1 (22 %) on average (Fig. 6). Anthropogenic BC arriving from Europe and Russia has been previously shown to be important for Alert air pollutant concentrations (Sharma et al., 2013). The model overestimation of BC in snow samples at Alert needs further investigation. It is likely that it originates from anthropogenic emissions in northwestern America or in Europe because forest fires in Canada and Russia, although important for Alert (e.g. Qi et al., 2017), were not significant in the present comparison.

It has been shown that measured concentrations of EC in snow in northwestern European Russia and western Siberia were underestimated in FLEXPART (Fig. 2). This was confirmed by the calculated fractional bias (see Sect. 3.2), the spatial distribution of which is shown in Fig. S1. To examine whether this underestimation was due to missing emission sources or errors in modelled transport and deposition, we have calculated the average footprint emission sensitivity for those sampling points, for which FLEXPART strongly (FB < -100 %) and slightly (-100 % < FB < 0 %) underestimated the observed values. The average footprint emission sensitivities are shown in Fig. 7 together with the locations of active fires in the last 2 months before the sample collection. The fire data were adopted from MODIS (Moderate Resolution Imaging Spectroradiometer) (Giglio et al., 2003) and the gas flaring facilities from the Global Gas Flaring Reduction Partnership (GGFR) (http://www.worldbank.org/en/programs/gasflaringreduction).

When the model strongly underestimated the measured EC (FB < -100 %), the average footprint emission sensitivity showed the highest values over the Yamal Peninsula and the agglomeration of many gas flares in Khanty-Mansiysk (Fig. 7b). This might confirm the findings of Huang et al. (2014) that gas flaring emissions in the ECLIPSE inventory, while very high, are still underestimated. According to a related study by Huang and Fu (2016), Russia contributes 57 % to the global BC emissions from gas flaring. Underestimation of modelled atmospheric concentrations compared to observations from the Barents and Kara seas was recently also reported by Popovicheva et al. (2017), although the underestimation was relatively small.

When FLEXPART showed a moderate underestimation of EC concentrations in snow (-100 % < FB < 0 %), the emission sensitivity was high near Arkhangelsk and over Scandinavia (Fig. 7). BC emissions in Scandinavia are considered relatively low in most inventories and contribute no more than 6.5 % to the global emissions in ACCMIP (Aerosol Chemistry Climate Model Intercomparison Project) (Lamarque et al., 2013), 6.2 % in EDGARv4.2 (Emission Database for Global Atmospheric Research) (Olivier et al., 2005), 2.1 % in MACCity (Monitoring Atmospheric Composition and Climate and megacity Zoom for the Environment) (Hollingsworth et al., 2008; Stein et al., 2012) and 3.3 % in ECLIPSE (Klimont et al., 2017). The highest emission sensitivity was found over northwestern Russia (Fig. 7), a region that includes Murmansk. Pollution levels in Murmansk could be high due to emissions from local industry, mining, heating and transport (Law and Stohl, 2007). Another potential source region was the Nenets–Komi area and western Kazakhstan, where a few other flaring facilities are located (Fig. 7).

Figure 7 shows that the underestimation of observed EC concentrations in snow strongly depends on the region where samples are collected. In western Siberia, the underestimation was larger than in northwestern European Russia. For this reason, we have computed the average region-specific emission sensitivities and the average region-specific contribution from the major polluting sources identified in the ECLIPSE dataset. We distinguish between three regions: northwestern European Russia, western Siberia (north of 62∘ N) and western Siberia (south of 62∘ N) (Figs. S4–S6). For the samples collected in northwestern European Russia (Fig. S4), an average contribution of 21.6 ngg-1 from all sources was estimated to have originated mainly from TRA (7.7 ngg-1) and DOM (10.4 ngg-1) sources in Finland. The contribution from BB and FLR emissions was insignificant (8 and 6 %, respectively), whereas the rest of the ECLIPSE sources were negligible (IND, ENE, WST). For the samples collected at high latitudes in western Siberia, the average contribution from all sources was more than 4 times higher (86 ngg-1) than those observed in northwestern European Russia (Fig. S5). FLR emissions accounted for 40 % of the total contribution, which reflects the proximity of the sampling site to the main flaring facilities of Russia. The average contribution from TRA activities in Europe and southeastern Russia to the northern part of western Siberia was 24 %. DOM emissions in eastern Europe also contributed another 28 %. Finally, for the samples that were collected in the southern part of western Siberia an average contribution of 47.4 ngg-1 was estimated from all sources included in ECLIPSE (Fig. S6). The highest contributing categories were TRA and DOM, whereas FLR appeared to contribute less, although the sampling site is close to the Khanty-Mansiysk flaring region. This is attributed to the prevailing winds that forced flaring emissions to a northernmost direction opposite to the location of the sampling stations (see Fig. S6).

Overall, the region-specific analysis of the sources contributing to modelled BC in snow showed that the DOM, FLR and/or TRA sources might explain the model underestimation in the high Arctic. However, in the most recent assessments of BC of the higher Arctic (Popovicheva et al., 2017; Winiger et al., 2017), it was shown that ECLIPSE captures levels of BC quite well, whereas FLR emissions might have a smaller impact in the central Siberian Arctic (Tiksi) than previously estimated. Surprisingly, the average contribution from BB in lower latitudes was extremely low in all western Siberia (Figs. S5 and S6), despite the fact that sampling took place in springtime, when BB becomes important. Evangeliou et al. (2016) reported that using a different dataset, which is based on the same approach as GFED but includes updated emission factors for Eurasia, surface concentrations of BC in the Arctic stations can be substantially higher. This shows the need for further investigation of BC sources in the Russian Arctic.