Retrieval of near-surface sulfur dioxide (SO2) concentrations at a global scale using IASI satellite observations

Sulfur dioxide (SO2) is an atmospheric trace gas with both natural and anthropogenic sources. In the troposphere, SO2 released by industrial activities mainly stays close to the ground level. The IASI/MetOp infrared remote sensor has shown over the years good performances for tracking SO2 plumes in the free troposphere. Probing anthropogenic SO2 pollution on the other hand is a challenge due to the generally low sensitivity of infrared measurements to the near-surface atmosphere, itself caused by the weak thermal contrasts between the ground and the air above it. Recent studies, which have focused on local sources (the industrial area of Norilsk and of the North China Plain), have however demonstrated that IASI was able to retrieve SO2 near-surface concentrations in favorable meteorological situations, and in particular in case of large temperature inversions. Expanding on these findings, this work presents new observations of near-surface SO2 at global scale from IASI observations. The method, which includes the determination of the SO2 plume altitude and SO2 boundary layer column, will be briefly presented. Global distributions of anthropogenic pollution will be shown, focusing on the identification of the principal hotspots and of exceptional pollution events. A first assessment of the retrieved columns with correlative measurements will be provided for some local sources. IASI measurements and new OMI SO2 retrievals will be compared. This will highlight the complementarity of these current TIR and UV sounders for measuring SO2 pollution, which could be exploited in the future with IASI-NG and Sentinel-5 instruments.

In contrast to volcanoes, SO 2 pollution from anthropogenic activities is difficult to monitor because it is often confined horizontally and vertically. In the UV spectral range, different methods have successfully been developed to retrieve surface SO 2 . These are included in different global products such as the operational planetary boundary layer (PBL) OMI SO 2 product (Krotkov et al., 2006(Krotkov et al., , 2008 or the recent OMI algorithm based on a multi-windows Differential Optical Absorption Spectroscopy (DOAS) scheme developed by Theys et al. (2015). The latter will be used for comparison purposes later in this paper. The availability of the satellite-derived columns from the UV nadir sounders have allowed SO 2 anthropogenic emissions to be inferred (e.g. Carn et al., 2007;Fioletov et al., 2011Fioletov et al., , 2013Fioletov et al., , 2015McLinden et al., 2012McLinden et al., , 2014). This has not yet been possible using TIR instruments, which suffer from lower sensitivity to the near-surface atmosphere due to generally low temperature differences between the surface and the PBL atmosphere (hereafter called thermal contrast). Recently Bauduin et al. (2014) and Boynard et al. (2014) have nevertheless demonstrated the capability of IASI to measure near-surface SO 2 locally. Both studies revealed that the presence of large thermal inversions (associated to high negative thermal contrasts) and low humidity (preventing opacity in the ν 3 band) has allowed near-surface SO 2 to be retrieved. However, the detection of SO 2 by IASI could theoretically be achieved in other situations, particularly in the case of large positive thermal contrasts, which correspond to a surface much hotter than the atmosphere. This is demonstrated and discussed in Sect. 3, which provides the first global distributions of near-surface SO 2 from IASI over the period 2008-2014. The method used to retrieve SO 2 columns, described thoroughly in Sect. 2, relies on the calculation of a hyperspectral range index (HRI), similarly to the work of Van Damme et al. (2014) for ammonia (NH 3 ). As the aim of this work is to retrieve nearsurface SO 2 , the determination of the altitude of SO 2 defines the first step of the method and relies on the work of Clarisse et al. (2014). It is used to remove all plumes located above a height of 4 km, which likely correspond to volcanic eruptions. The retrieval of 0-4 km SO 2 columns is performed in a second step, where calculated HRIs are converted into columns using a look-up table (LUT) approach. This calculation is performed using the ν 3 band (1300-1410 cm −1 ). This spectral band has the advantage of being more intense than the ν 1 band (a factor of 7.8 larger for the strongest lines, Bauduin et al., 2014). It is however more affected by the absorption of H 2 O, which can cause opacity in this spectral region in the near-surface atmosphere and therefore can reduce the IASI sensitivity to SO 2 down to the surface. This parameter, along with thermal contrast, has been taken into account in the whole retrieval procedure (described in Sect. 2) and the results are analysed with respect to these two parameters.

IASI and methodology
The IASI instrument is a Michelson interferometer onboard MetOp platforms (A and B) circling the Earth on a Sunsynchronous polar orbit. The IASI effective field of view is composed of 2 × 2 footprints, each with a diameter of 12 km at nadir. IASI has global coverage twice a day thanks to a swath of 2200 km and two overpasses per day (morning at 09:30 LT and evening at 21:30 LT at the equator). The instrument covers the thermal infrared spectral range from 645 to 2760 cm −1 with a spectral resolution of 0.5 cm −1 after apodization. Each measurement consists of 8461 radiance channels (0.25 cm −1 sampling) and is characterized by a noise of 0.2 K on average. Details about the instrument can be found elsewhere Hilton et al., 2012). Only measurements of the IASI-A instrument have been considered here.
The methodology used in this work to retrieve near-surface SO 2 amount relies on the sensitive detection method of trace gases introduced by Walker et al. (2011), and used later by Van Damme et al. (2014) for retrieving NH 3 columns at a global scale from IASI. Retrieval approaches based on spectral fitting generally consist of simultaneous iterative adjustments of the atmospheric parameters of interest (here the 0-4 km SO 2 column) and spectrally interfering unknown variables. The idea of the proposed method is to consider the interfering variables as permanent unknowns and to incorporate them in a generalized noise covariance matrix S. This matrix should include all variability coming from the parameters affecting the IASI spectrum in the spectral range under consideration (here 1300-1410 cm −1 , corresponding to the ν 3 band) but not SO 2 . In this way, instead of iteratively adjusting the SO 2 columns and the interfering parameters, a normalized range index is calculated according to Atmos. Meas. Tech., 9, 721-740, 2016 www.atmos-meas-tech.net/9/721/2016/ where K is a Jacobian, a derivative of the IASI spectrum y with respect to SO 2 ; y is the mean background spectrum with no detectable SO 2 associated with the matrix S. This matrix S acts as a weight in the projection of the observed spectrum onto the SO 2 signature, giving more importance to IASI channels less influenced by interfering parameters. Practically, y and S are built from a sufficiently large sample of SO 2 -free IASI spectra to include the global atmospheric variability in the absence of SO 2 (see Sects. 2.2 and 2.3). The HRI, which is unitless, can be seen as an index of detection, whose value represents the strength of the SO 2 signal in the IASI radiance spectrum, which is related to the amount of SO 2 in the atmosphere. The larger its value, the more likely the enhancement of the gas. An ensemble of SO 2free spectra has a mean HRI of 0 and a standard deviation of 1, and a HRI of 3 (which corresponds to 3σ ) can reasonably be considered as the limit of detection. Because the HRI does not correspond to the real column of SO 2 , it is therefore needed to convert it in a subsequent step. This can be done by using look-up tables built from forward model simulations, which link the simulated HRI values to known SO 2 columns. Prior to this, as this work focuses on near-surface pollution, we use the algorithm of Clarisse et al. (2014) to select the spectra with detectable low-altitude SO 2 enhancements. Only plumes located below 4 km are kept. These successive steps of the retrieval scheme are detailed in the next sections.

Retrieval of the altitude of the plume
The altitude of SO 2 is retrieved using the algorithm presented in Clarisse et al. (2014). In short, a HRI is computed for different altitudes following where the K h is the Jacobian for SO 2 located at an altitude h. If there is a detectable amount of SO 2 , the function HRI(h) will peak at the altitude of the plume. Indeed, the overlap between the IASI spectrum and the SO 2 spectral signature is maximal at this altitude. The height determination therefore consists in calculating the function HRI(h) at predefined altitudes and finding the altitude of its maximum. To this end, K h Jacobians have been precalculated using the finite difference method for the spectral range 1300-1410 cm −1 using monthly averaged H 2 O and temperature profiles in 10 • × 20 • boxes. These averages were calculated from the meteorological fields from the EU-METSAT L2 Product Processing Facility (PPF) (Schlüssel et al., 2005;August et al., 2012) using the 15th of each month of 2009, 2011 and 2013. One set of 30 vectors K h has been generated for each box and for each month, considering a 1 km thick SO 2 layer of 5 DU (1 Dobson unit = 2.69 × 10 16 molecules cm −2 ), located every 1 km from 1 to 30 km. For each IASI observation, local Jacobians are then calculated using a bilinear interpolation of the four closest grid boxes, to better take the variation of the atmospheric conditions into account when observations move away from the centre of the boxes. The mean background spectrum y and the associated covariance matrix S needed to calculate the HRI (Eq. 2) have been built using a sample of 1 million randomly chosen IASI spectra. Those with detectable SO 2 have been filtered out using an iterative approach: first, spectra with observable SO 2 signatures were rejected using a brightness temperature difference method (see Clarisse et al. (2008) for details) and a first estimate of the matrix S was made. The second step uses this initial matrix to exclude spectra with measurable HRI from the remaining set of measurements (a similar method is used in Van Damme et al., 2014 andClarisse et al., 2013). Similarly to the Jacobians, y and S are calculated over the spectral range 1300-1410 cm −1 . Note that an altitude was retrieved when an HRI larger than 2 was found, even though in practice we expect the altitude retrieval to be accurate only for HRI values above 4 or 5.
As explained in Clarisse et al. (2014), because of the use of averaged Jacobians, the retrieved altitude can be biased, especially close to the surface. The best accuracy is achieved between 5 and 15 km, and the altitude estimate is provided within 1-2 km. Below 5 km, the retrieved altitude is more uncertain. However, this is not an important issue here as the altitude is used only to filter out SO 2 plumes emitted by volcanoes directly in the free troposphere. In the following, only plumes located between the surface and 4 km above ground are selected. The retrieved SO 2 corresponds therefore to a 4 km thick layer (hereafter called 0-4 km column).
From the calculated K h , we can estimate where in the retrieved 0-4 km layer IASI is the most sensitive to SO 2 . We found that, for favourable conditions of thermal contrast and humidity, IASI is sensitive down to the surface but has its maximum sensitivity in the upper part of the 0-4 km layer. In the case of low thermal contrast (TC) and/or large column of H 2 O, IASI becomes insensitive to the lower part of the 0-4 km layer.

Retrieval of near-surface SO 2 concentrationslook-up tables
For the low SO 2 plumes, the next step consists in computing a HRI (according to Eq. 1) for each IASI measurement and converting it to a SO 2 column. Different Jacobian y and S have been built for this second step (see section below). Because a constant Jacobian is used in the calculation of the HRI, there are several parameters that impact its value in addition to the SO 2 abundance itself and they need to be ac-counted for. We have considered the impact of viewing angle (by building angle-dependent matrices for the HRI calculation -see Sect. 2.3.1) and the impact of humidity and thermal contrast, which are separate entries in the look-up tables (Sect. 2.3.2).

Angular dependency
The dependence of the signal strength on the viewing angle has to be taken into account in the conversion of HRI values. As reported by Van Damme et al. (2014), the application of a cosine factor to account for the increased path length tends to overcorrect the HRI and leads to a bias for larger angles. As they suggested, angle-dependent K, y and S have been used. Specifically, between 0 and 55 • , 5 • angle bins have been defined and a last one of 4 • is considered for 55-59 • (IASI zenith angle ranges between 0 and 58.8 • ).
For the median angle of each bin, a Jacobian has been generated for a standard atmosphere (Anderson et al., 1986), with a scaling factor applied to the methane profile according to Bauduin et al. (2014). A thermal contrast of 10 K has been considered. All K values have been calculated with the finite difference method for 200 ppb SO 2 well mixed between 4 and 5 km, and over the 1300-1410 cm −1 range. For y and S, almost the totality of cloud-free (i.e. cloud fraction below 20 % and available EUMETSAT L2 surface temperature, atmospheric temperature and H 2 O profiles) observations of the 15th of each month of 2009 and 2011 have been used, sorted by angle bins. Measurements with detectable SO 2 have been filtered as above. In this way, for each angle bin, y and S have been calculated from about 750 000 spectra.

Look-up tables
The conversion of the HRI into an SO 2 column is done using look-up tables, which, as for y and S, have been separated per angle bin. The LUTs include four dimensions linking thermal contrast, total column of water, HRI and SO 2 column. To build the LUT, forward simulations of IASI spectra have been performed for a series of situations, summarized in Table 1, using the line-by-line Atmosphit software (Coheur et al., 2005). More specifically, the following parameters were varied to provide a representative set of atmospheric conditions.
-SO 2 columns: to obtain a reference SO 2 vertical profile for anthropogenic emissions, we relied on the global chemistry transport MOZART model (Emmons et al., 2010) outputs of January, April, July and October 2009 and 2010. An average profile was calculated from all modelled profiles above the eastern United States, Europe and eastern China, with the SO 2 concentration above 4 km set to zero. The resulting reference profile is shown in Fig. 1 (blue). The set of atmospheric SO 2 columns included in the LUT was then obtained by scaling this reference profile by the 16 factors listed in Table 1, leading to a range of 0-4 km SO 2 columns (ground to 4 km above it) going from 0 to 415 DU.
-H 2 O: in a similar way, the water vapour profile from the US Standard model ( Fig. 1 in red) has also been varied using 16 scaling factors (Table 1), covering a range of H 2 O total column from 9.5 × 10 19 to 2.3 × 10 23 molecules cm −2 .
-Temperature: a single temperature profile has been used (US standard, Fig. 1, right panel). To include a range of thermal contrast values, which are defined as the difference between the temperature of the ground and the temperature of the air at 500 m (see Fig. 1), we have varied the surface temperature to provide 25 different situations, listed in Table 1. These include extreme cases of thermal contrasts, from −30 to +40 K, but also a range of low and moderate values. Note that a thermal contrast of 0 corresponds to an isothermal 0-1 km layer and implies that the outgoing radiance of this layer is that of a black body . Note also that a constant emissivity of 0.98 has been used in the forward simulations; for most cases, differences between using a spectrally varying emissivity and a constant emissivity are on the order of the noise of the instrument.
The LUTs constructed as described above have been interpolated on a finer grid (TC, H 2 O and SO 2 dimensions). An example of resulting LUT is shown in Fig. 2a (for constant water vapour) and Fig. 2b (for constant thermal contrast). The retrieval scheme consists in determining, for each IASI measurement, the satellite zenith angle, the HRI, the thermal contrast, the total column of water vapour and, using the LUT, the 0-4 km column of SO 2 . From Fig. 2, it can be seen that the HRI has the same sign as the thermal contrast. In the case of positive thermal contrast, this is explained by the fact that SO 2 spectral lines are in absorption in IASI measurements, resulting in a negative difference (y − y). Given the fact that the Jacobians are also negative (see definition in Sect. 2.3.1), the calculated HRI is positive. As a rule, for constant zenith angle, H 2 O column and SO 2 column, the value of the HRI increases with the thermal contrast. This increase in spectral signal corresponds to an increase of IASI sensitivity to near-surface SO 2 . However (see Fig. 2b for a constant TC of 15 K), for increasing water vapour, which renders the atmosphere opaque in the low layers, the IASI sensitivity decreases along with the HRI for constant SO 2 columns, thermal contrast and viewing angle. In the case of negative thermal contrast, SO 2 lines are in emission and the calculated HRI is negative too. For decreasing negative thermal contrast, the HRI value usually decreases. But from Fig. 2a it can be seen that some HRI values are positive for negative thermal contrasts. We explored this seemingly odd behaviour with the help of   2.4 × 10 20 molecules cm −2 . From 0 to 66.33 DU, HRI decreases for increasing SO 2 . Above 66.33 DU, the HRI starts to increase with increasing SO 2 . From about 250 DU, the HRI becomes positive. This behaviour can be explained by the competition between emission (mainly in the 0-1 km layer) and absorption (above 1 km). Figure 3 (right panel) presents the contributions (in absolute value) of the emission in the 0-1 km layer and the absorption in the 1-4 km layer to the total spectral signal as a function of the 0-4 km SO 2 column. They have been evaluated at 1355 cm −1 using similar techniques as in Clarisse et al. (2010). In Fig. 3, for concentrations ranging from 0 to 66.33 DU, emission in the 0-1 km layer increases more rapidly than absorption in the 1-4 km layer. This results in decreasing HRI (more and more negative). From 66.33 DU, emission comes closer to saturation; its increase is slower than the one of absorption, whose saturation occurs for larger SO 2 columns, and the HRI begins to increase. From around 250 DU, absorption totally counterbalances emission and HRI values become positive. This competition between emission in the lowest layers and absorption higher up depends on the value of the temperature inversion, as the latter determines the strength of the emission. Note that this competition also depends on the altitude of the thermal inversion, but which is constant here (just above the ground). The consequence is that, for negative thermal contrast, a negative HRI can be converted into two SO 2 columns (see Fig. 3, left panel): a small one (emission combined with lower absorption above 1 km) and a large one (larger emission partly counterbalanced by a more rapid increase of absorption above 1 km). In that case, because the very large columns included in the LUTs are not expected above anthropogenic sources, only the smallest one is considered. Note that the large columns for which the HRI is positive for negative thermal contrast have been kept. From the LUT, we can estimate the detection limit of IASI to near-surface SO 2 . In Fig. 4, the lowest detectable 0-4 km column of SO 2 is presented as a function of thermal contrast and the total column of H 2 O. These columns have been calculated using the LUT assuming a detection threshold of 3 on the value of HRI (see Sect. 2.1). As expected, this limit of detection largely depends on thermal contrast and humidity. Indeed, when the former is close to 0, IASI stays insensitive, even to large SO 2 columns. For large thermal contrasts, 0-4 km columns lower than 1 DU can be measured. This also depends on the humidity. Below 2 × 10 22 molecules cm −2 , the limit of detection stays below 2 DU for both high positive and high negative thermal contrasts. For larger H 2 O amount,  Left panel: evolution of the HRI as a function of the 0-4 km column of SO 2 , in the case of negative thermal contrast (−10 K). In this simulation, the total H 2 O column is 2.4 × 10 20 molecules cm −2 and the angle bin is 15-20 • . Right panel: contributions of the emission in the 0-1 km layer (blue) and of absorption in the 1-4 km layer (red) to the IASI spectrum at 1355 cm −1 , expressed in brightness temperature difference (absolute value). Details are given in the text. this limit rapidly increases for negative thermal contrast but stays relatively low for large positive thermal contrasts. From above 4 × 10 22 molecules cm −2 of H 2 O, the detection threshold starts to increase for positive thermal contrasts.

Error characterization
For each LUT, an associated table of errors has been generated by propagating the uncertainties of the different LUT parameters: where σ SO 2 is the absolute error of the SO 2 column, σ TC and σ H 2 O are the errors on thermal contrast and total column of water vapour, which are respectively taken equal to √ 2 K and 10 %, relying on early validation of the IASI level 2 meteorological fields from the PPF (Pougatchev et al., 2009); σ HRI is the standard deviation of the HRI and is equal to 1.
An example of error tables is given in Fig. 5 (Fig. 5a relative errors and Fig. 5b absolute error) for the angle bin 0-5 • and for a total column of water vapour of 2 × 10 20 molecules cm −2 . As expected, the errors are directly linked to the IASI sensitivity to near-surface SO 2 , with large errors (above 100 % and 10 DU respectively for relative and absolute errors) occurring in the case of small thermal contrasts. The errors decrease with increasing thermal contrasts and drop to 20 % (2 DU) or less in the most favourable situations. As discussed above, for large total columns of H 2 O, IASI is also less sensitive to near-surface SO 2 and errors in- crease accordingly. The errors are used in the following to filter out the data (see Sect. 3). Another source of errors, which is not taken into account in the error calculation, is the assumed SO 2 vertical profile. A given column amount of SO 2 located at different altitudes corresponds to different HRI values. For instance, we have estimated the error on the SO 2 column to be of the order of 30 % when SO 2 is confined to the 0-1 km layer only (for a TC of 10 K and a total column of H 2 O of 9.5 × 10 21 molecules cm −2 ). Note also that the assumed temperature profile can be a source of error (see Sect. 3.3).

Results
In this section, global distributions, time series and a first product evaluation are presented. For this, only SO 2 columns with less than 25 % relative error and less than 10 DU in absolute error have been used. The second criterion was necessary to remove spurious data over the cold Antarctic region with unrealistic high columns. The first condition has been chosen to reject measurements for which IASI sensitivity to near-surface SO 2 is limited, and thus for which associated retrieved SO 2 columns have large uncertainties. However, this procedure tends to favour large SO 2 columns. As a consequence, the presented averages are expected to be biased high. It is important to stress though that individual measurements that pass the filter are not a priori biased but have random uncertainties related to errors on the different input parameters (errors on TC, choice of temperature and SO 2 profile). It is also important to note that by using the HRI > 2 criterion early in the retrieval procedure, we have made the choice not to treat observations with small or undetectable amounts of SO 2 . This potentially throws away one category of useful observations: those where a low HRI is found together with favourable atmospheric conditions. In this case, a low HRI can be an indication of the absence of large SO 2 concentrations at the surface. Future versions of the retrieval algorithm could be expanded to include those, and this would potentially decrease the high bias when calculating averages.

Global distributions
The SO 2 retrievals have been performed on 7 years of IASI observations (1 January 2008-30 September 2014). In Fig. 6, an average global distribution of the near-surface column of SO 2 for this period is presented, separately for day (top panel) and night (bottom panel) observations. Only measurements with less than 20 % cloud fraction in the IASI field of view and with available surface temperature, profiles of temperature and H 2 O from the EUMETSAT IASI level 2 PPF have been used. A selection based on the error, described in the beginning of Sect. 3, has also been applied. The columns that pass these posterior filters have been averaged on a 0.5 • × 0.5 • grid for cells including more than five IASI measurements. The bottom-right inset in the daytime map presents the total number of successful measurements (those which pass the error filtering) in each grid boxes. The bottom-right inset in the night-time map presents the global anthropogenic emissions (in kg s −1 m −2 ) of SO 2 provided by the EDGAR v4.2 inventory (downloaded from the ETHER/ECCAD database) (EDGAR-Emission Database for Global Atmospheric Research, 2011). Figure 6 reveals several anthropogenic and volcanic hotspots, numbered from 1 to 13. Most of them are observed during the morning overpass, when the thermal contrast is large. They are detailed here.
1. China: China is one of the world's largest emission sources of SO 2 , mainly due to energy supply through coal combustion (Lu et al., 2010;Smith et al., 2011;Lin et al., 2012). A large region of enhanced SO 2 columns, from 1 to 8 DU on the 7-year average, is seen over the industrial area surrounding Beijing. The largest columns are found close to Beijing, where emissions are the largest according to the EDGAR database, and they then decrease westwards.
2. Norilsk: located above the polar circle, Norilsk is an industrial area where heavy metals are extracted from sulfide ores. It is also well known for its extremely high levels of pollution (Blacksmith Institutes, 2007), and more particularly for its emissions of SO 2 (AMAP, 1998(AMAP, , 2006. The Norilsk smelters are also observed with IASI in Fig. 6, with averaged SO 2 columns varying between 1 and 9 DU. A comparison between measurements obtained in this work and those retrieved using an optimal estimation method  is given in Sect. 3.4.1. 4. Iran: several SO 2 sources are observed above Iran. Columns of 1 to 4 DU are measured above the smelters of Sar Cheshmeh copper complex (Rastmanesh et al., 2010(Rastmanesh et al., , 2011. Emissions of oil industries located on the Khark Island (Ardestani and Shafie-Pour, 2009;Fioletov et al., 2013) are also observed, with columns around 1 to 2 DU.

5.
Balkhash: in Fig. 6, we can see that IASI is able to measure SO 2 above the region of copper smelters located in Balkhash, Kazakhstan (Nadirov et al., 2013;Fioletov et al., 2013). Columns around 1 to 2 DU are retrieved.
6. Mexico and Popocatepetl: columns reaching more than 10 DU are measured in the region of Mexico City and are regularly detected. These can be attributed to lowaltitude plumes released by the Popocatepetl volcano (Varley and Taran, 2003;Grutter et al., 2008) and/or to SO 2 emissions of the Tula industrial complex, located north of Mexico City (De Foy et al., 2009). The proximity of these two sources does not however allow us to separate their individual contribution by using IASI observations only.
7. Kamchatka volcanoes: SO 2 columns of about 2-3 DU are observed above the Kamchatka region. These probably correspond to the activity of different volcanoes located in this region (e.g. Kearney et al. (2008); see also the archive of the Global Volcanism Program: http: //volcano.si.edu/). 8. Nyiragongo: above the Democratic Republic of Congo, a plume with columns larger than 10 DU is detected. It corresponds to SO 2 volcanic degassing from Nyiragongo (Carn et al., 2013) and also probably to emissions of its neighbour, the Nyamuragira (Campion, 2014). Fig. 6, Mount Etna is covered by SO 2 , with 0-4 km columns around 6 DU. This volcano is known for its periodic degassing activity and lava-fountaining events (Tamburello et al., 2013;Ganci et al., 2012).

Etna: in
10. Andes: a large SO 2 plume, with columns around 2-3 DU, is observed in the region of the Andes and can have several origins, which are difficult to distinguish. In the south of Peru, some volcanoes have shown activity in recent years (e.g. Ubinas or Sabancaya; see the archive of the Global Volcanism Program: http: //volcano.si.edu/). Copper smelters are located in Ilo (Carn et al., 2007), close to the coast, but are south of the observed plume. SO 2 measured above Bolivia and Chile can originate from active volcanoes of the central Andean volcanic zone (Tassi et al., 2011;g. the Putana volcano, Stebel et al., 2015). Smelters are also located in this area (Huneeus et al., 2006) and anthropogenic emissions are reported by the EDGAR database. The presence of an artefact in this region, due to the difficulty of representing the emissivity, can however not be totally rejected. Finally, SO 2 measured above Argentina, Ecuador and Colombia is mainly emitted by local volcanoes (Global volcanism Program, http://volcano.si. edu/). 11. Bulgaria: a narrow plume, with SO 2 columns of about 2 DU, is observed in Bulgaria. This corresponds to the Maritsa Iztok Complex of thermal power plants located close to Galabovo and Radnevo (Eisinger and Burrows, 1998;Prodanova et al., 2008). One unexpected pattern in the daytime distribution is the SO 2 plume at the extreme western part of China, corresponding to the Taklamakan Desert (number 13). In this region, the EDGAR inventory only documents a few small sources, but no strong ones are known. While this could be explained by an artefact of the calculated HRI due to sand emissivity, which strongly affects the thermal infrared measurements, it is noteworthy that the issue is not similarly observed above other deserts. For instance, in Fig. 7 we compare the distribution of measured HRI and the total column of H 2 O for three desert regions: the Sahara, the centre of Australia and the Taklamakan. We observe that the HRI values are for almost 90 % of the cases below the detection limit of 3 above the Sahara and Australian deserts, whereas 30 % of the measurements over the Taklamakan are associated to a HRI between 3 and 5, and in a few cases even above. It is therefore likely that the measured columns are real, with SO 2 being transported from the source regions in east China over the desert or being emitted there by developing gas and oil industries (Lin et al., 2013, http://www.cnpc.com.cn/en/ Taklamakan/Taklamakan.shtml). The very high thermal con- trast (up to 20 K) and very low humidity conditions found jointly in that region make it indeed possible to measure such weak columns. However, further investigations are still required to properly assess the source of this detected plume and to exclude possible false attribution due to surface emissivity effects. Finally, the low-altitude parts of the plume released by the Nabro eruption, which followed complex transport patterns , are also seen during day above Ethiopia. Note that SO 2 is observed above Iceland and corresponds to the Bardarbunga eruption that started in September 2014 (Schmidt et al., 2015). The different conditions and filters applied on IASI measurements (see the beginning of this section) are responsible for the small area covered by SO 2 .

Atmos
It is worth emphasizing that some of the measured points in the 7-year average are only representative of 1 year. For continuous/permanent sources, this indeed depends on the inter-annual variation of thermal contrast and water vapour that limit IASI sensitivity. Moreover, some particular events are typical of some years, like volcanic eruptions. Finally, as mentioned in the beginning of Sect. 3, because of the error filtering, the presented global averages are biased high. It is therefore an average of measurements which are sensitive to near-surface SO 2 .
Comparison with the EDGAR database has allowed us to identify the observed SO 2 plumes. It also points out the sources missed by IASI. Almost year-round low thermal contrasts (January-March and September-December) combined with high humidity in summer (May to September) lead to the absence of eastern US and eastern European sources in Fig. 6. Sources in India and in southeastern Asia are also not observed by IASI. This is likely because of large H 2 O amount in the atmosphere in the tropical region, which renders the near-surface atmosphere opaque in the ν 3 band. Note that the joint use of the ν 1 band, less impacted by H 2 O absorption, could allow these sources to be detected. The problem of these missing sources is not limited to IASI. OMI SO 2 distributions (Figs. 6 and 7 in Theys et al., 2015, andFigs. 1 and6 in Krotkov et al., 2015) shows the ability of the sounder to measure small sources above India, USA and Europe which are not detected by IASI. These distributions also reveal the absence of some of them, compared to those reported by the EDGAR database: South Eastern Asia (e.g. Thailand), Northern Europe and part of India. These Atmos. Meas. Tech., 9, 721-740, 2016 www.atmos-meas-tech.net/9/721/2016/ absences in OMI measurements are possibly caused by unfavourable geophysical conditions (presence of clouds, . . . ), but this has to be investigated more deeply. However, qualitatively, OMI and IASI global distributions are in good agreement. Both instruments are able to measure large sources such as Northeast China as well as smaller ones, like power plants in Turkey or Bulgaria. The two sounders are also complementary: regions characterized by high humidity and/or low thermal contrasts, undetectable by IASI, can be measured by OMI, whereas IASI better monitors SO 2 at high latitudes, especially during the winter, and is not limited to daytime. When examining Fig. 6, the differences between the SO 2 distributions retrieved from IASI measurements during morning (top panel) and evening (bottom panel) overpasses are also striking. In the evening distribution, the plumes are more spatially confined and the columns at the centre of the plumes are generally larger by about a factor of 3. These differences will be discussed in more detail in Sect. 3.3. after the description of the time series below, which provides additional clues about difference.

Time series
In Fig. 8, the 7-year time series (1 January 2008-30 September 2014) above Beijing and the smelters region of Sar Cheshmeh (Iran) are presented as examples. For both areas, daily averages of near-surface SO 2 columns, thermal contrast and H 2 O total column are shown, separately for the morning (blue) and evening (red) overpasses of IASI. The averages have been calculated within a radius of 125 and 75 km around Beijing and Sar Cheshmeh respectively. As before, only observations with less than 20 % cloud fraction and with available meteorological level 2 have been taken into account and only those satisfying the error filtering are considered.
For Beijing and Sar Cheshmeh, the daily-averaged SO 2 columns from the morning overpass vary around 3 DU, with maxima that can reach 15 and 25 DU respectively. The time series is incomplete for Beijing, with successful SO 2 retrievals from December to May associated with fairly high thermal contrast (10 K on average but up to 20 K - Fig. 8, middle panels) and low humidity (below 5 × 10 22 molecules cm −2 - Fig. 8, bottom panels). The favourable thermal contrast conditions persist mostly yearround in Beijing but the humidity is too high during the other months to allow us to probe the surface using this IASI scheme. For Sar Cheshmeh, the time series of SO 2 columns from the IASI morning overpass is more extensive and this is due to the dryness of the site as compared to Beijing (a factor 2), combined also with persisting high thermal contrast conditions, from 10 K in the colder months to more than 30 K in summer.
It is inferred from Fig. 8 that IASI is mostly not sensitive to surface SO 2 above the two sites in the evening due to the drop of thermal contrast close to 0. As already noted in the previous section, the retrieved SO 2 columns are larger by at least a factor of 3 in the evening compared to the morning (red vs. blue symbols). This is further investigated in the next section.

Morning-evening differences
To examine the differences in the SO 2 distributions from morning and evening overpasses, we focus hereafter on a large area (30-40 • N/105-117 • E) above China. For this region and for each month in the period 1 January 2008-30 September 2014, we first calculate for morning and evening the fraction of successful SO 2 retrievals, i.e. those that pass the prior and posterior filters described in the previous sections and for which the HRI has a correspondence in the LUTs, relative to all the retrievals performed in the considered area. Regarding the last condition, it is important to point out that we found that a number of IASI measurements, mainly associated with negative thermal contrast, were not covered by the LUTs (i.e. their HRI values have no correspondence in the LUTs). The fraction of these measurements is shown in Fig. 9 (left, second panel from top), along with the fraction of successful SO 2 retrievals (top panel), as time series. They are compared (as in Fig. 8) to the time evolution of thermal contrast (third panel from top) and water content (bottom panel).
From the top panel we see that the number of successful retrievals during the evening orbit is significantly smaller than during the morning orbit of IASI. In the morning the seasonality is marked, with successful retrievals varying from close to zero in the humid summer months to 20-60 % from January to May. For the evening measurements, the number of successful retrievals stays low year-round and is above 5 % only for 1 or 2 months in spring. The prime rejection criterion for the evening measurements is surprisingly the absence of correspondence, for given angles, thermal contrasts and humidity, between measured and simulated HRI in the LUTs, as obvious from the second panel. This is especially the case in winter (60-80 % of rejected measurements), when thermal contrast is negative and humidity is low.
The fact that such situations are not included in the LUTs comes very likely from a misrepresentation of night-time atmospheric temperature profiles and in particular, temperature inversions with the conditions used to build the LUTs (Table 1). To illustrate this, Fig. 9 (right panel) shows a comparison between the standard temperature profile used and a typical profile retrieved above China (35.81 • N/117.81 • E) on 29 December 2013. This is a situation for which the thermal contrast is −5 K and the water column is 2.42 × 10 22 molecules cm −2 , and for which the measured HRI value of −3.9 has no correspondence in the LUTs. The simulation of a IASI spectrum with these two temperature profiles, assuming a SO 2 column of 4.35 DU, results in totally different values of the HRI, +1.2 for the US Stan- dard temperature profile and −4.1 for the retrieved temperature profile. These results pinpoint a limitation of the current LUTs for slightly negative thermal contrast (it is not observed for large temperature inversions), which is a range where the competition between absorption and emission contributions to the HRI vary drastically. More work will be needed to avoid this shortcoming of the method in the future, either by including more temperature profiles in the calculation of the LUTs or by using alternative approaches to better account for the variety of real situations encountered. Note that errors on the thermal contrast and on the assumed SO 2 vertical profile also affect the HRI and, as a consequence, Atmos. Meas. Tech., 9, 721-740, 2016 www.atmos-meas-tech.net/9/721/2016/ the retrieved SO 2 column. They could be partly responsible for the observed non-correspondence between measured HRI and the LUTs. The small number of successful retrievals in the evening measurements, combined with the generally lower sensitivity of IASI in this period of the day, is likely responsible in part for the factor 2-3 difference observed in the SO 2 columns between morning and evening measurements. Indeed, as only the retrieved SO 2 columns with small errors are kept, and as in the evening these are mainly those with large columns, the averages are biased high. The effect also exists for the morning measurements but is less pronounced because of the better sensitivity to smaller columns with the atmospheric conditions -particularly thermal contrast -encountered. Another possible cause for the larger concentrations is photochemistry. During the day, the photochemistry is more active and the concentrations of oxidants such as OH, H 2 O 2 and O 3 are high, creating an important sink for SO 2 , which disappears at night, favourizing higher concentrations. Such a diurnal cycle of SO 2 has been observed previously in China (Wang, 2002;Wang et al., 2014) and in other regions of the world (Khemani et al., 1987;Psiloglou et al., 2013), but in others, noontime SO 2 peaks have also been observed (e.g. Lin et al., 2012;Xu et al., 2014), possibly as a result of other meteorological/dynamical effects (Xu et al., 2014).

Comparisons with an optimal estimation retrieval scheme
As already mentioned in Sect. 3.1, Norilsk is an industrial area located in Siberia that emits large quantities of SO 2 . Recently, taking the advantages of the large temperature inversions that develop in winter in the region, we have monitored SO 2 in this area for the first time with the IASI sounder over several years , by using a method relying on the iterative optimal estimation (Rodgers, 2000) and exploiting a generalized spectral noise covariance matrix (see also Carboni et al., 2012). As a first assessment of the low-altitude SO 2 column product developed in this work, we compare the resulting 0-4 km SO 2 columns with the 0-5 km columns retrieved by Bauduin et al. (2014). The results are presented in Fig. 10. For the comparison, we consider observations located within a 150 km radius around the city of Norilsk. Only measurements of Bauduin et al. (2014) with less Figure 10. Comparison between SO 2 columns retrieved above the industrial area of Norilsk using the retrieval method developed in this work and the 0-5 km SO 2 columns retrieved in Bauduin et al. (2014). Measurements located within a radius of 150 km centred in Norilsk with less than 25 % cloud coverage and retrieved with relative errors smaller than 25 % and absolute errors smaller than 10 DU have been taken into account. Furthermore, only measurements with a H 2 O amount at 350 m lower than 4 g kg −1 and with thermal contrast larger than 5 K in absolute value have been considered (see detailed explanations in the text). The pink line corresponds to the linear regression (reduced major axis) calculated between the two sets of data. The colour bar represents the humidity at 350 m expressed in g kg −1 .
than 25 % cloud fraction, with a thermal contrast larger than 5 K in absolute value and with a humidity below 4 g kg −1 at 350 m above ground have been selected (this altitude corresponds to the average height of the temperature inversions). These last conditions ensure that near-surface SO 2 is indeed probed, as explained in Bauduin et al. (2014) and also well visualized in their Fig. 3. Finally, we only consider the SO 2 columns retrieved with the method described in this work satisfying the error criteria. The entire period 2008-2013 is analysed, resulting in a total of 1233 pairs of columns to compare. The comparison between the two sets of SO 2 columns is shown in Fig. 10. A linear regression is also shown between the two coincident sets of data using the reduced major axis method (Smith et al., 2009) to account for the fact that both data sets come with errors. The agreement between the columns is very good, characterized by a correlation coefficient of 0.94. The intercept, which is close to zero, and the slope of 0.80 indicate that the SO 2 columns retrieved using the LUT tend to be 20 % smaller than those retrieved with the iterative method of Bauduin et al. (2014). This difference is partly due to the difference in columns (0-4 km with the newly developed method against 0-5 km in Bauduin et al., 2014) and to the difference in the profile used to build the LUTs with the a priori profile used in the iterative method. The use of the constant temperature profile for the LUTs can also cause this difference.
Finally, it is worth emphasizing in Fig. 10 the measurements for which the LUTs provide a column above 2.5-3 DU and the optimal estimation retrieval of a column close to 0 corresponding to the a priori column. As obvious from the colour scale, these retrievals are all associated with a relatively high humidity of 3 g kg −1 . These measurements have a significant HRI around 5, indicating small signal strength, which is probably the reason of the difference observed between the two methods.

Comparisons with OMI-derived SO 2 columns
OMI is an imaging spectrograph that operates in a nadirviewing mode in the ultraviolet-visible spectral range 270-500 nm, and was launched in 2004 on the EOS-Aura NASA platform (details are given in Levelt et al., 2006). We perform a comparison of the retrieved 0-4 km SO 2 column from IASI and those retrieved from OMI using the algorithm of Theys et al. (2015) for anthropogenic SO 2 . For the IASI column, the same filters on cloud fraction and errors as described in Sects. 3.1 and 3.2 are applied. For the OMI columns, only those retrieved in the spectral range 312-326 nm from measurements not affected too much by the row anomaly, with solar zenith angles smaller than 65 • and less than 30 % cloud fraction are used (see Theys et al. (2015) for details). The comparison is performed on monthly averages for the period 2010-2013 and for an area corresponding to a radius of 125 km around Beijing. Figure 11 shows the comparison in terms of (top panel) a time series of the monthly averaged columns of IASI (blue) and OMI (red squares for the standard retrieval and green triangles for the retrieval using a different air mass factor -AMF), and in terms of regional maps (2010-2013) above China (bottom panels). From Fig. 11, it can be seen that the IASI columns are on average a factor of 2.5 larger. The mean relative difference between the monthly averages of the two instruments is −135 % (OMI-IASI/OMI). This difference has several origins. Firstly, monthly means calculated from IASI are probably overestimated by the fact that only the columns with low errors are kept, which favours the higher values of the columns. Secondly, it is likely that OMI SO 2 columns are underestimated. This has already been observed by Theys et al. (2015) above Xianghe (China), in the comparison with MAX-DOAS measurements (Wang et al., 2014) and explained by the inappropriate AMF used to convert the OMIderived SO 2 slant column densities in vertical column densities. In their study, Theys et al. showed that the use of better AMF significantly improves the agreement between MAX-DOAS and OMI observations. For the sake of illustration, we also show in Fig. 11 the SO 2 vertical column densities from OMI retrieved using the method of Theys et al. (2015), but with a constant AMF of 0.4, which is the one used in the operational OMI PBL SO 2 product (Krotkov et al., 2008). With this correction, the agreement between the two instruments is improved; the mean relative difference becomes −65 %. Dis-Atmos. Meas. Tech., 9, 721-740 crepancies are within the range of what we can expect given the difference in the overpass times of the two satellites and given the high bias introduced by averaging only the IASI observations with a low relative uncertainty. Note also that the difference between the vertical sensitivity profiles of the two instruments can also contribute to observed differences. Finally, we have calculated the correlation coefficient between IASI and OMI measurements. It is very low for both OMI data sets, −0.03 and 0.15 respectively for the varying AMF and AMF = 0.4. These low values are mainly caused by one outlier on May 2012, where IASI has a large SO 2 column of 5.7 DU, which corresponds, however, to very few measurements (three). These coefficients significantly improve when this outlier is removed; they become respectively 0.23 and 0.54.

Conclusions
In this work, we have presented a method for retrieving SO 2 in the low troposphere from IASI at a global scale. The method follows two steps, both relying on the calculation of hyperspectral range indexes (HRIs), which represent the strength of SO 2 spectral signal in IASI measurements. In the first step, the altitude of SO 2 plumes is retrieved and all plumes with altitude above 4 km (from ground), likely from volcanic origin, are rejected. For the remaining low-altitude plumes, HRI values are converted into 0-4 km (ground to Tables of errors have been associated with each LUT, allowing the error characterization of each retrieved SO 2 column and the posterior selection of the retrieved SO 2 columns for which IASI is sensitive enough. We have estimated for day and night, respectively, that 10 and 7 % of the total IASI observations are taken in the favourable conditions required to measure near-surface SO 2 . Also, from calculated Jacobians, we have shown that IASI's sensitivity to SO 2 increases with altitude, but in the case of favourable conditions of thermal contrast and water vapour, the instrument is sensitive to SO 2 down to the surface. The method has been applied to IASI data from 1 January 2008 to 30 September 2014 to provide global distributions and time series for the 0-4 km column. The average global distribution reveals the large known anthropogenic SO 2 sources, such as the Norilsk and Sar Cheshmeh smelters, the power plants in South Africa and the large industrial region in Northeast China. Smaller sources, e.g. power plants in Bulgaria, are also measured. In addition to this, lowaltitude plumes from degassing volcanoes are also detected. Non-negligible SO 2 columns have been retrieved above the Taklamakan Desert and this was explained by enhanced sensitivity of IASI in this region characterized by extremely low humidity and high thermal contrast; the source of the SO 2 remains to be assessed. Similarly, the generally favourable conditions occurring in Sar Cheshmeh (Iran) have allowed us to acquire the daily time evolution of the SO 2 column almost completely over the entire 7 years. This was not the case for the Beijing area that we selected as another example region, where we show that IASI sensitivity has a strong seasonal cycle such that the SO 2 columns can only be retrieved with small errors for the period December-May, corresponding to the driest months. The retrieved 0-4 km SO 2 columns from IASI have been compared to those of OMI on a monthly averaged basis. A high bias of 135 % has been revealed, decreasing to 65 %, depending on the choice of the AMF used in the OMI retrievals. The high bias is likely linked to an overestimation of IASI averages due to the error filtering applied on the data. More comparisons and validation work are needed to investigate the observed differences between the two instruments more deeply. The two instruments are nevertheless complementary; regions characterized by high humidity and/or low thermal contrasts can be measured by OMI, whereas IASI better monitors SO 2 at high latitudes, especially during the winter, and is not limited to daytime. Another assessment of the retrieval method was provided by comparing the retrieved SO 2 columns from IASI with the LUT-based approach to those retrieved from the same measurements with an iterative optimal estimation scheme; a good correspondence was found between the two column data sets (correlation coefficient of 0.94, with the LUT lower by 20 %) considering the different assumptions and input profiles used in the two methods. This excellent agreement shows how well the new method is able to retrieve near-surface SO 2 . It has the advantage of being very fast; it-erations and the retrieval of interfering parameters are not needed. In contrast to the optimal estimation method, it is, however, not able to provide a full retrieval characterization of the retrieved columns/profiles, notably in terms of vertical sensitivity.
Finally, striking differences between morning and evening SO 2 distributions retrieved from IASI were shown, with the SO 2 columns retrieved in the evening being more spatially confined and larger than those in the morning by a factor of 2-3. While changes in photochemistry and a larger highbias in the night averages could explain part of this effect, we have shown that it is most probably due to a shortcoming of the LUT, which relies on a single temperature profile and is not able to deal well with temperature inversions, which develop in the evening and in winter. Further developments will be needed to correct for this and to allow a better representativeness of the variety of temperature, SO 2 and H 2 O profiles occurring globally. The use of both ν 3 and ν 1 bands can also be envisaged to reduce the impact of H 2 O absorption and to increase the IASI sensitivity to surface SO 2 . Despite this, given the preliminary comparisons, the results obtained with this new method are however very encouraging, especially for daytime, and constitute the first successful attempt to retrieve near-surface SO 2 globally with the IASI thermal infrared sensor. The continuation of this program is ensured by the upcoming launch of MetOp-C (2018) and in the longer term, by the IASI-NG mission onboard MetOp-SG (Crevoisier et al., 2014).