Supplement to : Tropospheric BrO column densities in the Arctic derived from satellite : retrieval and comparison to ground-based measurements

This document contains two sets of maps corresponding to two field campaigns conducted in the course of the International Polar Year 2007–2008. The maps show gridded tropospheric column densities of bromine monoxide (BrO) measured by the GOME-2 instrument in the vicinity of two different locations where ground-based measurements of BrO and many other tracegas species were performed. In spring 2008, ship-based measurements were performed aboard the research icebreaker Amundsen south of Banks Island in the Amundsen Gulf. In spring 2009, ground-based measurements were performed at Barrow, Alaska. Both field initiatives were conducted in the scope of the Ocean – Atmosphere – Sea Ice – Snowpack (OASIS) project. The BrO maps are averaged tropospheric column densities retrieved from measurements within the denoted day. On some days, however, there is no valid data present. Furthermore, only those measurements are averaged, which feature a solar zenith angle (SZA) smaller than 80◦, and a tropospheric air-mass factor (AMF) larger than 1. The location of the ground-based measurements – the Amundsen research vessel, or Barrow, respectively – are marked with a white circle in the center of the map.

1 Introduction Barrie et al. (1988) discovered bromine activation as the phenomenon behind polar ozone depletion events (ODEs) in the Arctic troposphere.Since then, considerable progress in understanding the phenomenon of ODEs has been made.However, even after two decades, key questions remain open: what are the sources of reactive halogens, what triggers their release, and what is the impact on the global tropospheric ozone budget?For a review of the current understanding of the halogen chemistry in the polar troposphere, see also Simpson et al. (2007) and references therein.
Bromine monoxide (BrO) is a radical that catalytically destroys ozone.Its first observation from space was enabled by the Global Ozone Monitoring Experiment (GOME) instrument aboard the ERS-2 satellite (Wagner and Platt, 1998;Richter et al., 1998;Chance, 1998).Areas up to 2000 km across (covering several million km 2 ) with elevated columns of BrO were detected to appear from one day to another implying local production of BrO.BrO is remotely probed from space by the technique of differential optical absorption spectroscopy (DOAS) which uses characteristic narrow absorption bands of molecules (Platt and Stutz, 2008).
Compared to ground-based measurement techniques like long-path DOAS (LP-DOAS) (Tuckermann et al., 1997;Hausmann and Platt, 1994;Hönninger et al., 2004;Pöhler et al., 2010;Liao et al., 2011), multi-axis DOAS (MAX-DOAS) (Hönninger and Platt, 2002;Hönninger et al., 2004;Wagner et al., 2007;Frieß et al., 2011), or chemical ionization mass spectrometry (CIMS) (Liao et al., 2011), observations from space offer a much better spatial coverage (at polar latitudes, full coverage is reached once per day) while temporal resolution and information about the vertical distribution is comparatively sparse.Another advantage of satellite instruments is the relatively long time of operation of several years.Satellite data sets are particularly suitable to answer open questions or test hypotheses on a more general basis (Wagner et al., 2001;Richter et al., 2002;Hollwedel et al., 2004;Kaleschke et al., 2004).
Before an existing data set on BrO column densities can be analyzed for tropospheric BrO activation, systematic errors need to be minimized.One of the largest uncertainties comes from the variability of the stratospheric BrO column (Wagner and Platt, 1998;Wagner, 1999;Theys et al., 2009;Salawitch et al., 2010;Choi et al., 2011) which needs to be assessed in order to study BrO in the troposphere from satellite observations.For extremely low tropopause heights, the stratospheric partial column may become comparable to that of a tropospheric event.When the tropopause lowers, the stratospheric air is adiabatically compressed and hence the stratospheric column of BrO increases additionally to an increase of the overall thickness of the stratosphere.Spatial structures mimicking tropospheric bromine events may thus appear in maps of the total BrO column.The challenge is to separate possible tropospheric events from stratospheric disturbances.
Several retrievals of tropospheric BrO columns use the output of stratospheric chemistry models for stratospheric BrO correction (Theys et al., 2009(Theys et al., , 2011;;Begoin et al., 2010;Salawitch et al., 2010;Toyota et al., 2011;Choi et al., 2011).These algorithms either use simulated stratospheric columns of BrO directly or derive a parameterisation of the stratospheric BrO column based on model results first and then apply a climatology from which the stratospheric BrO column is calculated using measured O 3 and NO 2 column data.However, chemistry models are potentially biased because the chemical mechanisms may be incomplete, and necessary parameterisations may result in systematic errors.Model results also depend on the choice of initial values which are usually difficult to obtain.
The retrieval proposed in this work overcomes these drawbacks by retrieving the parameters to estimate the stratospheric BrO column using only the measurements themselves.In brief, our algorithm uses the simultaneously retrieved O 3 column density to account for dynamic effects and the retrieved NO 2 column density to account for chemical effects.The algorithm contains the following four steps (details are given in Sect.2.2): first, the column measurements of BrO and O 3 are binned according to the respective NO 2 column, the solar zenith angle, and the line of sight angle.In a second step, the measurements without a significantly enhanced BrO/O 3 column ratio are considered to calculate the mean stratospheric BrO/O 3 column ratio within each bin.Then, the stratospheric BrO column for each pixel is calculated using the measured O 3 and NO 2 column densities and the solar zenith angle.Finally, the difference between measured total and calculated stratospheric column yields a residual BrO column.This approach is completely independent from models.
In a second step, an algorithm assuring the sensitivity of the satellite measurement towards BrO located in the boundary layer (BL) is developed and also included in the retrieval.Hence, it is possible to study surface processes involved in bromine activation on a per-pixel basis.Parameters affecting this sensitivity are the surface albedo as well as the thickness and height of overlying clouds.In polar regions, the detection of clouds from satellites is particularly difficult for instruments measuring in the UV and visible spectral range due to ambiguities between cloud particles and the ice-or snow-covered underlying surface.Various studies to measure optical properties of clouds over ice in polar regions from space (e.g.Vasilkov et al., 2010;O'Byrne et al., 2010) are based on data from a multitude of sensors and satellites.The ice-mode of FRESCO+ (Fast Retrieval Scheme for Clouds from the Oxygen A band, Koelemeijer et al., 2001;Wang et al., 2008) derives the surface height of a Lambertian reflector with monthly averaged climatological albedo value using O 2 absorption measurements.In this work, we chose a slightly different approach: individual reflectances are combined with the corresponding differential absorption of the (O 2 ) 2 collision complex (denoted O 4 in this work) in order to assure the sensitivity above a given threshold.The scale height of O 4 is approx.4 km (Greenblatt et al., 1990;Acarreta et al., 2004) thus providing a better sensitivity to nearsurface concentrations compared to O 2 .
This paper is organized as follows: in Sect. 2 we describe our new algorithm to retrieve tropospheric columns of BrO during periods of halogen activation in Arctic spring.Different parts of the algorithm are either compared to simulated data or measurements from other satellite instruments in Sect.3. Retrieved tropospheric BrO columns are compared with ground-based measurements of BrO obtained during two field campaigns to the Arctic in 2008 and2009 (Sect. 4).Conclusions are drawn in Sect. 5.

Spectral evaluation, column separation and sensitivity filter
The GOME-2 instrument (second Global Ozone Monitoring Experiment) is a high-resolution nadir scanning spectrometer aboard the MetOp-A satellite (e.g.Callies et al., 2000;Munro et al., 2006).MetOp-A, launched in 2006, is the first of a series of three polar-orbiting satellites of identical design.
The satellite is flying in a sun-synchronous orbit crossing the Equator at 09:30 LT.It is a platform for a set of instruments primarily designed for meteorological applications.Data from two of these instruments -the AVHRR (Advanced Very High Resolution Radiometer) and GOME-2 -are used in this study.The retrieval of tropospheric columns of BrO from GOME-2 is the main focus, while AVHRR data will later be applied to evaluate the sensitivity to near-surface trace-gas concentrations (Sect.3.3).GOME-2 has four main spectral channels in the UV/vis spectral range between 240 and 790 nm.The instrument scans in a whisk-broom scheme with a swath-width of 1920 km, which allows an almost global coverage each day.Polar regions, however, are sampled several times a day.The nominal ground pixel size is approx.80 × 40 km 2 with an integration time of 187.5 ms per spectrum.
Three fundamental steps are needed in order to retrieve a residual tropospheric vertical column density of BrO (VCD trop ) from GOME-2 spectra: i. retrieval of total slant column densities (SCDs) of BrO, O 3 , NO 2 , and O 4 from Earth radiance spectra using the DOAS method (Sect.2.1).
iii. calculation of VCD trop using a tropospheric air-mass factor (AMF) retrieved from O 4 SCD and reflectance measurements (Sect.2.3).The algorithm, which is also capable of filtering measurements with a low sensitivity to near-surface concentrations of BrO, is described in Sect.2.3.1, and its implementation in Sect.2.3.2.
The results from the retrieval as well as its advantages and disadvantages are discussed in Sect.2.4.

Evaluation of GOME-2 spectra
Differential optical absorption spectroscopy (DOAS) is a common technique to derive slant column densities of numerous trace-gases in the atmosphere (Platt and Stutz, 2008).In this work, DOAS is applied in three different wavelength ranges to derive SCDs of BrO, O 3 , O 4 and NO 2 from calibrated GOME-2 spectra.Table 1 summarizes the parameters and molecular absorption cross-sections applied in the DOAS evaluation.
For the retrieval of BrO SCDs, several modifications compared to previously published retrievals are applied to the settings of the DOAS fit.The wavelength range between 336 and 360 nm combines the standard wavelength ranges used for the first GOME (Wagner and Platt, 1998;Aliwell et al., 2002) and the Scanning Imaging Absorption Spectrometer for Atmospheric Chartography (SCIAMACHY) (Afe et al., 2004;De Smedt et al., 2004) instruments and encompasses four absorption bands of the BrO molecule.Furthermore, formaldehyde (HCHO) was excluded from the spectral evaluation in order to reduce the noise level of the BrO retrieval.This approach is appropriate when only polar regions are taken into account where HCHO abundances are generally low.In addition to molecular absorption cross-sections, two spectra are included in the evaluation procedure to account for the wavelength-dependent Ring effect (Grainger and Ring, 1962) following the suggestions of Wagner et al. (2009a); both are calculated and normalised using the DOA-SIS software version 3.2 (Kraus, 2004).Furthermore, a reciprocal intensity spectrum is included in the fit in order to account for possible stray light within the instrument.A fourth-order polynomial is finally included to account for broad-band effects like surface reflection as well as Mie and Rayleigh scattering.
The fit result provides total SCDs of BrO S, which need to be subsequently normalised for several reasons: (i) the SCDs of weak absorbers potentially contain an unknown offset due to spectral structures varying over time as discussed by Richter et al. (2002); (ii) the GOME-2 instrument suffers from sensor degradation leading to increased statistical and, more problematic, systematic errors of the BrO SCDs as revealed by Dikty et al. (2011); (iii) the proposed retrieval algorithm for tropospheric BrO VCDs is intended to be applicable also on satellite sensors other than the GOME-2 on MetOp-A.The normalisation step introduces the possibility to homogenise the BrO data gained from the measurements of different satellite instruments.
Measured BrO SCDs are normalised to a VCD of V norm = 3.5 × 10 13 molec cm −2 within a reference sector over the Pacific Ocean as suggested by Richter et al. (2002).This normalisation is performed for each pixel number of one scan separately (GOME-2: 32 pixels per scan; pixel numbers correspond to discrete VZA angles).The boundaries of the reference sector are ±10 • latitude and 150 • E to 100 • W longitude.Pixels with a footprint significantly different from the nominal ≈ 80 × 40 km 2 (narrow-mode and backscan pixels) are excluded from counting as reference measurements.The normalised SCDs S are calculated by subtracting the median difference between SCDs in the reference sector and the normalised SCD S norm = V norm •A geom from the measured SCDs applying the geometrical AMF A geom .While the AMF is defined as the ratio of SCD and VCD in general, A geom displays an adequate approximation for stratospheric absorbers for SZA < 80 • .A geom is defined as where ϑ denotes the SZA, and ψ denotes the viewing zenith angle (VZA).The asterisk ( * ) denotes that the reference in the column to the left is applied.
Total VCDs V of BrO can be approximated from S using again applying A geom .As an example, total VCDs of BrO measured on 25 March 2009 over the Arctic are plotted in Fig. 1a.
Owing to the strong differential structure of ozone, the SCD of O 3 may also be derived from the same DOAS evaluation as BrO at much higher signal-to-noise ratio.The O 3 SCD is calculated as the sum of the fit results of both O 3 references corresponding to different temperatures.This approach of a retrieval for O 3 potentially leads to SCDs with a systematic error which, however, cancels out later during the parameterisation of the stratospheric BrO-column (Sect.2.2).O 3 VCDs computed from O 3 SCDs using A geom are shown in Fig. 1b for 25 March 2009.
Using radiances in another wavelength interval, the SCD of the oxygen collision complex (O 4 ) is retrieved in the range between 355 and 390 nm using the setting compiled in Table 1.This spectral range includes two absorption bands of O 4 at 360 nm and 380 nm.
Finally, the SCD of NO 2 is retrieved from radiances measured in band 4 of the GOME-2 instrument in the range between 431 nm and 453 nm.In contrast to the previous two settings, a single Ring spectrum (also calculated using the DOASIS software) was found to be sufficient due to a weaker Raman signal at longer wavelengths in connection with the rather narrow fit range.

Separation of tropospheric and stratospheric BrO slant-columns
This section describes how the measured total SCD of BrO is separated into background stratospheric and residual tropospheric column density S strat and S trop , respectively.
Furthermore, the standard deviation of the measurement σ 0 of BrO is estimated.This allows to evaluate the significance of a possible tropospheric signal.For the sake of clarity, SCDs S and VCDs V without subscripted chemical formula denote BrO column densities throughout the paper.

Concept of the BrO column separation
The main task of the BrO column separation is to compute the SCD of BrO contained in the stratosphere, S strat .Two substances, O 3 and NO 2 , are used to parametrise S strat similar to the approach initially proposed by Theys et al. (2009) but without utilising any model output.O 3 is chosen as a parameter for tropopause dynamics, whereas NO 2 is used as a parameter for variations in the stratospheric chemistry.
The ratio z 0 of the stratospheric BrO SCD to the O 3 SCD, S strat,O 3 , is defined as where S strat,O 3 is expressed in molec cm −2 using the definition of the Dobson unit (1 DU = 2.69 × 10 16 molec cm −2 ).The PV at the 475K isentrope (f) may be used to identify regions within the polar vortex (see Sect. 2.2.2).All VCDs are calculated using a geometric AMF; gray areas contain no data.
Knowing z 0 and S strat,O 3 would allow us to compute the SCD of BrO directly, This approach implicitly relies on similar vertical profiles of BrO and O 3 , which is further discussed in Sect.3.2.However, measurement data obtained by GOME-2 furnish us only with a set of values of the ratio z between the total BrO SCD S and of the total O 3 SCD S O 3 in the stratosphere and troposphere combined: Since almost the entire O 3 column is located in the stratosphere, z becomes where z is defined as the ratio between S trop and S strat,O 3 .In addition, measurement errors are included in z, which allows us to write where z0 is the mean of z 0 and ζ σ 0 is Gaussian distributed with zero mean and σ 0 standard deviation.The quantity z can be interpreted as an error contribution due to elevated concentrations of BrO in the troposphere.The distribution of z is unknown a priori.However, it leads to an overestimation of S strat , if the simple mean z is used as an estimator for z0 .
The ratio z 0 in the stratosphere, Eq. ( 4), depends mainly on the stratospheric NO 2 chemistry which is parameterised by the VCD of NO 2 and the SZA.The stratospheric chemistry leads to significant deviations between S strat and S strat,O 3 .The partitioning of inorganic bromine species Br y = {BrO+BrONO 2 +Br 2 +HOBr+ HBr + . . .} is not constant (Dorf et al., 2006;Theys et al., 2009;Salawitch et al., 2010).It turns out that the BrO/Br y concentration ratio, which is typically of the order of 0.6 during daytime, depends primarily on the stratospheric NO 2 concentration.This is due to the fact that NO 2 acts as a sink for stratospheric BrO producing bromine nitrate (BrONO 2 ): which decreases the BrO concentration while leaving the concentration of Br y unchanged.BrONO 2 is the second most abundant Br y -species during daylight (e.g.Sinnhuber et al., 2002;Atkinson et al., 2007;Theys et al., 2009).The main loss mechanism of BrONO 2 , however, is photolysis leading to a quasi-stationary state between BrO and BrONO 2 depending on the NO 2 concentration and the actinic flux.As a result, the ratio z 0 in Eq. ( 4) decreases with increasing concentration of NO 2 also depending on the SZA ϑ determining the actinic flux.The concentration of NO 2 is not accessible from nadir measurements alone, and therefore the NO 2 vertical column density V N is used in the column separation process instead.Furthermore, our algorithm also accounts for a slight dependence of z 0 on the VZA ψ.The stratospheric BrO SCD S strat is therefore mainly a function of ϑ, V N , and ψ: Unfortunately, however, the assumptions made so far are not applicable to the chemistry inside the polar vortex and during ozone-hole conditions.Extremely cold temperatures alter the chemistry of the stratosphere rendering reaction Eq. ( 9) insufficient to describe the chemistry affecting BrO.Moreover, there can be massive chemical loss of stratospheric O 3 so that S strat,O 3 can no longer be used to account for dynamical effects.Therefore, our algorithm in its present form is inapplicable for an estimation of the stratospheric BrO within the polar vortex occurring in springtime Antarctica in general and in some areas of the Northern Hemisphere during winters with low stratospheric temperatures as depicted in Fig. 2. Finally, in order to compute S strat as a function of ϑ, V N , and ψ, we need to compute z0 (ϑ, V N , ψ) from The precise procedure of how this is done is explained in the following subsection.

Implementation of the column separation algorithm
This section describes the implementation of the algorithm to calculate the tropospheric SCD of BrO.The algorithm is divided into four steps: (i) selection of reference measurements for one day and partitioning of reference measurements in the (ϑ, V N )-plane for five different ψ-ranges; (ii) calculation of z0 and σ 0 in each partition after filtering significantly enhanced z ; and (iii) mapping of z0 (ϑ, V N , ψ) on all observations and calculation of S strat according to Eq. ( 10).
The statistical analysis to retrieve z0 requires a sufficiently large base population of measurements T 0 .The analysis is performed separately for each day D. In order to increase the size of T 0 , all measurements within a 7-day period P = [D − 3, D + 3] are considered.This approach improves the statistical significance and reduces noise.It is similar to a running average filter and relies on a stratospheric chemistry changing only slightly within one week.i.A subset T ⊂ T 0 of measurements is selected to avoid interferences with anthropogenic NO 2 emissions and to increase the accuracy of the stratospheric information in the nadir observations.T , from which the stratospheric correction is computed, contains only those observations with an SZA smaller than 80 • , latitudes above 30 • N, a fit-error for BrO smaller than 5 × 10 13 molec cm −2 , an O 4 SCD larger than 6.5 × 10 42 molec 2 cm −5 , and a non-negative NO 2 VCD smaller than 8 × 10 15 molec cm −2 below 60 • N. Narrow-mode and backscan pixels are excluded from T as well as potential measurements within the polar vortex, as for example depicted in Fig. 2f.Areas inside the polar vortex are identified using information about the potential vorticity derived from meteorological model data (ECMWF operational analysis, regular 1 • × 1 • grid with 91 hybrid pressure levels, 6 h time resolution).Columns exceeding a potential vorticity of 35 PVU at the 475 K isentrope surface or 75 PVU at 550 K are discarded from the further analysis.Furthermore, T does not contain any measurements with a ground elevation above 1000 m and no measurements over land masses at latitudes below 73 • N. The latter selection rule accounts for areas with a strong anthropogenic NO 2 signal like Prudhoe Bay or Norilsk which would interfere with the algorithm.After applying these filters, the final subset T contains N α ≈ 10 5 reference observations for each day from which the stratospheric BrO column is estimated.
ii.A mean stratospheric BrO/O 3 background ratio z0 (ϑ, V N , ψ) is calculated from T .Obtaining an estimate of z0 , Eq. ( 11), by measured values of z α ∈ T with coordinates (ϑ α , V N α , ψ α ), α = 1, . . ., N α , containing an arbitrary error with a Gaussian and a positive unknown contribution, requires a technique of approximating a function on an unstructured set of points where the data to be approximated contain uncertainties.Traditionally, least-squares approximations (Quarteroni et al., 2002) are used to approximate scattered data.More elaborate methods use radial basis functions or kriging (Press et al., 2007) in order to treat scattered data.Common to these methods is that some knowledge about the distribution, such as the variance, is necessary in order to compute an approximant.In addition, they are relatively costly, given that the number of measured values N α is large (≈ 10 5 ), making it necessary to have an efficient method to process a large number of these data sets.
For the method proposed in this paper, we take advantage of the fact that the function z0 (ϑ, V N , ψ) depends only weakly on ϑ, V N , and ψ.In addition, we use trilinear interpolation in order to avoid spurious oscillations which can occur when using polynomials of higher degree.Since z0 depends only weakly on ϑ, V N , and ψ, we can regroup the measured z in subsets for which z0 is almost constant.For a domain this boils down to finding a partition β , β = 1, . . ., N β of , such that β contains enough points to allow for statistics on z α for which (ϑ α , V N α , ψ α ) ∈ β .On the other hand, β should be small enough, such that z0 does not vary too much with respect to ϑ, V N , and ψ in β .It is clear that such a partition, as shown in Fig. 3, is not unique and that the shape of the subsets β might influence the accuracy of the present method.In Appendix A, we present in more detail how is partitioned in quadrilaterals, allowing a trilinear reconstruction of z0 .For each partition β , a filter algorithm is applied as presented in the following.The filter algorithm is based on the assumption that an ensemble of z is normally distributed around z0 .Significant outliers, if any, are mostly due to enhancements of the tropospheric BrO column and, to a lesser degree, due to a partially depleted O 3 column.Both effects lead to an increase of a particular z by z in Eq. ( 11), which in turn leads to an increasing asymmetry of the otherwise symmetric normal distributed z (Fig. 4).The asymmetry a β of the distribution of z = z( β ) in partition β is defined as where z denotes the mean, z the median, and σ the standard deviation of z.If a β is larger than a threshold, i.e. the distribution is skewed towards higher BrO/O 3 SCD ratios, a subset of z accounting for the stratosphere needs to be calculated before z and σ can be used as estimators for z0 and the standard deviation σ 0 of z 0 , respectively.
A filter algorithm is designed to find a subset of z with a symmetric distribution identified as the stratospheric mode (Fig. 4).The asymmetry of the distribution of z is iteratively minimized by cropping values with an offset z = |z − z| larger than a given threshold δz.In step k of the iteration, the asymmetry a k of the distribution of is calculated with zk−1 denoting the mean of the distribution in the previous step.Starting with δz 0 = max(z)− z, the threshold δz k is iteratively decreased until a k ≤ 0.001 or a maximum of k = 20 steps is reached (see green bars in Fig. 4).The minimal asymmetry calculated from this algorithm is limited by numerical accuracy, and the termination condition of 0.001 was found to provide still a reasonably small residual asymmetry of the output.The result is a filtered mean zβ = zk .The standard deviation σ β , however, is not calculated based on the cropped distribution of z k .This approach would lead to an underestimation of the true standard deviation, because the cropped distribution (green bars in Fig. 4) has a larger kurtosis than the normal distribution.Therefore, it is computed only using measurements with z m < zβ and because this selection is assumed not to include any measurements with a significant tropospheric signal.
iii.The above computed values zβ and σ β are mapped to the center of gravity of the points (ϑ α , V N α , ψ α ) in β and used for trilinear interpolation, which furnishes us two functions: z0 (ϑ, V N , ψ) and σ 0 (ϑ, V N , ψ).Now, the SCD of BrO in the stratosphere S strat and its standard deviation, σ strat , can be computed by Table 2. Summary of all modelled geometries for which the threshold parametrization is performed.The solar zenith angle (SZA), relative azimuth angle (RAA), and viewing zenith angle (VZA) are defined in the satellite system, respectively.
The detailed steps of the partitioning algorithm are presented in Appendix A.

Sensitivity filter and air-mass factor
In order to finally retrieve the desired residual tropospheric VCD of BrO from the tropospheric SCD using we need to calculate the tropospheric air-mass factor A trop .This section describes how A trop can be retrieved from radiance measurements and O 4 SCDs and that each measurement can be classified into sensitive to the boundary layer (BL) and possibly obscured.The concentration of O 4 is proportional to the square of the O 2 concentration, and therefore its scale height is approximately 4 km.Hence, its absorption is a good indicator for the photons having penetrated the lower part of the atmosphere (e.g.Wagner and Platt, 1998).

Concept of the sensitivity filter
Ground-based measurements showed that most of the enhanced tropospheric BrO column is located within the BL and often close to the surface (e.g.Hönninger et al., 2004;Wagner et al., 2007;Pöhler et al., 2010;Prados-Roman et al., 2010;Frieß et al., 2011).We therefore assume, as an approximation, that the residual tropospheric column of BrO is entirely located between 0 and 500 m above the ground with a constant concentration (box profile).It is noted that the exact value of the BrO mixed layer-height may differ in reality.Radiative transfer simulations, however, showed that its choice is not critical for the presented considerations because the sensitivity of nadir measurements only slightly depends on altitude above surfaces with high albedo which are typical for polar regions.Therefore, instead of a real AMF trop , the AMF for the lowest 500 m (AMF 500 , denoted A 500 ) is retrieved and used in this work.
For nadir satellite observations, the sensitivity to the ground mostly depends on the surface albedo and whether clouds with a large cloud optical density (COD) are present.Under clear-sky conditions, the absorption signal from tracegases located close to the ground is reduced over dark surfaces due to little reflection by the ground compared to Rayleigh and Mie scattering in the atmosphere.But over bright surfaces, a substantial fraction of the observed photons penetrates near-surface layers.To a large extent, this is still true even for cloudy scenes.Thick clouds, however, effectively shield the absorption signal from these layers.
The distinction between sea ice, snow, thick aerosol layers, water clouds, and ice clouds by satellite remote sensing is not unambiguously possible (Vasilkov et al., 2010), and therefore surface albedo and COD cannot be readily derived from our measurements.
Instead, we chose an approach relying on proxies to parametrise A 500 .The two proxies used in the proposed algorithm are the reflectance R and the O 4 AMF, A O .On the one hand, R is a well suited measure to discriminate either clouds/ice (bright) and ocean/land (dark).A O , on the other hand, helps to discriminate between ice and clouds and furthermore provides information about the height and optical thickness of potential clouds.R is calculated as where L and E are the Earth radiance and solar irradiance measured by GOME-2 at 372 nm, respectively.The wavelength of 372 nm for R was chosen in order to minimize where V O 4 = 1.33 × 10 43 molec 2 cm −5 is the O 4 VCD integrated from sea level to the top of the atmosphere.Equation ( 19) furthermore applies an empirical correction factor of 0.8 which has already been suggested by Wagner et al. (2009b) and Clémer et al. (2010) and was confirmed by sensitivity studies conducted for this work.The same definition is used for the computation of A O , and, hence, the reduction of the real O 4 VCD over an elevated surface cancels out in the comparison between measurement and model.However, the illustration in Fig. 6b depicts A O measured on 25 March 2009 depending on the surface elevation.
Results from a computational radiative transfer model are used to study the interrelation between modelled values for R, A O and A 500 .For this purpose, triples of (R, A O , A 500 ) were modelled for a comprehensive set of surface albedos and aerosol/cloud scenarios.The main objective in the next step is to identify the range (or area in the (R, A O )-plane) where A 500 exceeds a certain sensitivity threshold AMF min 500 .The range limits are geometrically approximated, parameterised, and saved in lookup tables (LUTs) for discrete viewing geometries.When finally analysing the measurements, the LUT parameters are interpolated depending on the viewing geometry.Whether a measurement fulfils the AMF min 500criterion or not is then decided based on the measured R and A O .
Finally, the AMF for the boundary layer A meas 500 depends on R and A O .The parameters a 0 , a x and a y of the surface are derived from a least-squares fit of a selection of modelled (R, A O , A 500 )-triples with A 500 > AMF min 500 .In analogy to the surface sensitivity algorithm, a 0 , a x and a y are also stored in LUTs.

Implementation of the sensitivity filter and AMF calculation
This section describes the implementation of the surface sensitivity filter algorithm.For each viewing geometry, the algorithm consists of five steps: (i) modelling of (R, A O , A 500 )triplets for a fixed set of aerosol scenarios and surface albedos, (ii) interpolation of additional (R, A O , A 500 )-triplets accounting for partial cloud cover and different surface scenarios, (iii) parameterisation of the range of the (R, A O )-plane where A 500 exceeds a given threshold AMF min 500 , (iv) derivation of the a-parameters in Eq. ( 20), and (v) allocation of derived parameters in lookup tables.The LUTs are finally needed to interpolate the stored parameters for each GOME-2 pixel depending on its viewing geometry.The interpolated parameters are needed to decide whether a pixel is sensitive to the boundary layer and to calculate A 500 using Eq. ( 20).
There are four parameters defining the satellite viewing geometry: the SZA ϑ, the solar relative azimuth angle (SRAA), the viewing zenith angle (VZA) ψ and the ground elevation.These parameters span the four dimensional LUTs whose discretisation nodes are summarized in Table 2.Each LUT has a total of 6720 entries corresponding to 6720 different viewing geometries.i. R, A O and A 500 are modelled for different surface albedos and aerosol scenarios using the McArtim software package (Deutschmann et al., 2011).Two different wavelengths are used in the radiative transfer calculations: R is derived from radiative transfer simulations at 372 nm, whereas A O , and A 500 are simulated at 360 nm.For each LUT entry, R, A O , and A 500 are calculated for the albedos 0.03, 0.09, 0.24, 0.39, 0.54, 0.66, 0.78, 0.90, 0.96 for a pure Rayleigh atmosphere (clear-sky) and the aerosol/cloud scenarios summarized in Table 3.For the calculation of A 500 , a tropospheric box profile between 0 and 500 m is assumed.
Before the entire LUTs have been calculated, the scenarios summarised in Table 3 were found to be largely representative for the presented sensitivity filter through extensive radiative transfer simulations.However, two scenarios (0-1 km, OD 20 and 3-4 km, OD 50) were added at a later stage in order to further improve the accuracy of the algorithm.It is noted that future studies may benefit from using even more selected scenarios yet increasing the computational cost of the algorithm.
ii.Further (R, A O , A 500 )-triplets are interpolated from the Monte Carlo model results for two reasons: firstly, interpolation increases the number of values populating the (R, A O )-plane and hence increasing the accuracy of the subsequent parameterisation, and, secondly, it may be accounted for real gradients of surface albedo and partial cloud cover through interpolation.Large albedo gradients are typical for ice edges over oceans or areas of freshly fallen snow over land.Therefore, the surface albedo is parameterised by two properties: the albedo at a wavelength of 372 nm and the high albedo fraction of the surface η s .

H. Sihler et al.: Tropospheric BrO in the Arctic derived from satellite
η s is the geometric fraction of the ground pixel assumed to have a very high albedo R high = 0.96.The reflectance R defined by Eq. ( 18) then depends on η s and the modelled reflectances R high and R low over surfaces with an albedo of 0.96 and below 0.96, respectively.
The number of photons crossing the boundary between both parts is assumed to be negligible (independent pixel approximation).Accordingly, the modelled AMF depends on η s following where the modelled A low and A high are weighted by the modelled radiances (Martin et al., 2002).
Furthermore, scattering media in the atmosphere, i.e. clouds and/or aerosol layers, are modelled as a single layer with a geometric thickness of 1 km containing particles with a single scattering albedo of 1.00 and a Henyey-Greenstein asymmetry parameter of g = 0.85 (King, 1987).The parametrization of scattering media in our model atmosphere has three dimensions: the cloud fraction (η c ), the cloud height (CH) and the cloud optical density (COD).η c is defined as the fraction of a scenery which is covered by clouds.In analogy to the definition of η s , photons are assumed to travel either through cloud-free (cf) or cloud-covered (cc) sceneries.The radiances and AMFs depending on the cloudfraction may then be interpolated using and respectively.
Summing up the interpolation steps for the radiance and both AMFs for fractional η s and COD, (1) η c is varied from 0.2 to 0.8 for every constant albedo using Eqs.( 23) and ( 24), respectively.(2) For the clear-sky case, η s ranges from 0.05 to 0.95 with steps of 0.05 using Eqs.( 21) and ( 22), respectively.(3) With clouds, η s and η c were varied from 0 to 1 and from 0.2 to 1, respectively, both with steps of 0.2.This scheme results in 938 modelled and interpolated (R, A O , A 500 )-triplets.As an example, all triplets are shown in Fig. 7a for a nadir looking geometry and SZA = 76 • .R is plotted along the abscissa-axis, and A O is plotted along the ordinate-axis.
A 500 values are colour-coded.The comparison between modelled and measured (R,A O )-pairs in Fig. 7b shows that the range of modelled values (for a specific viewing geometry) includes almost all corresponding measurements.Obviously, the numerical radiative transfer model McArtim is capable of reproducing the range of real measurements for the considered cloud scenarios.
Figure 7 furthermore illustrates the advantages of using the two parameters R and A O instead of using just a single A O threshold.There is a significant number of measurements located in the sensitive range featuring an O 4 AMF below point A but also at a lower radiance.These measurements would be lost if only one threshold criterion based on S O 4 was applied.Furthermore, the measurements gained from using the two-parameter approach are particularly precious for the investigation of bromine activation in the Arctic.These measurements are more likely located at the sea-ice edge, because, at a given radiance R, A O is maximal for clear-sky scenarios over pixels partially covered by sea ice.
iii.The limits of range P in the (R, A O )-plane containing A 500 -values smaller than AMF min 500 are parameterised.P is the inverse of the range where an A 500 exceeding AMF min 500 can be assured.The limits of P are geometrically approximated in order to obtain a suitable parameterisation.Therefore, a convex hull H containing all A 500 smaller than AMF min 500 is constructed.As depicted by the shaded area in Fig. 7a, the characteristic shape of H enables us to approximate its upper edge with a parabola g: Before g is approximated to the upper edge of H , we introduce an intensity threshold h.Using the reflectances of the upper right corner A and left corner B of H , R A and R B , respectively, h is given by the mean Finalizing the parameterisation of the edge, g is derived from a least-squares fit using the points of the upper edge of H with R ≥ h.
iv.A least-squares surface fit of all modelled and interpolated triples in the upper right section (greater than g, and h) is performed using the model function in Eq. ( 20). Figure 8 compares A 500 resulting from the bilinear model to the modelled and interpolated input values of the fit for one example geometry (SZA = 76 • , same as in Fig. 7).This plot reveals that a single value (mean or median) would add a significant systematic error to the retrieved A 500 compared to the real A 500 .It is therefore concluded that using the two proxies (R and A O ) for the determination of A 500 offers the opportunity to even quantify A 500 to some degree instead of e.g. using a constant value.v.For a given AMF min 500 , h, g 0 , g 1 , g 2 and the surface fit parameters a 0 , a x and a y are stored in seven separate LUTs, which are then used to interpolate the thresholds and AMF meas 500 for any observation geometry for any measured R and A O .
Finally, an observation is flagged as sensitive if A O and R are larger than g and h, respectively, or otherwise as possibly obscured.If the measurement is sensitive, A 500 is derived using interpolated values for a 0 , a x and a y .

Results and discussion of the retrieval algorithm
Tropospheric VCDs of BrO resulting from the column separation algorithm are displayed for 25 March 2009 in Fig. 1e and for 1 April 2007 in Fig. 2e.Both figures illustrate the capability of the algorithm to separate the residual tropospheric column from the measured total column and to reduce the correlation to the tropopause height on a large scale.Fine-structured areas of elevated BrO remain in the retrieved tropospheric columns.For 1 April 2007, the pixels in the east sector fall into areas where the O 3 VCDs are reduced due to ozone hole conditions.These are removed from the retrieval.
Figure 6 shows the tropospheric BrO columns for different choices of the AMF min 500 threshold.The algorithm successfully removes measurements over areas outside the Arctic with a relatively low surface albedo.The comparison between Fig. 6c through 6f shows that the maps depend only slightly on the choice of AMF min 500 .
It is important to note that the presented algorithm -compared to previously published algorithms -depends neither on results from stratospheric chemistry models, gridded measurements from other satellite instruments nor surface albedo climatologies avoiding the disadvantages of using a potentially biased model description and possible short-term deviations from climatological values.Apart from the potential vorticity data provided by ECMWF to identify areas potentially disturbed by ozone hole chemistry, only data measured by the GOME-2 instrument are required.
Another distinct advantage of the column separation algorithm is that measurement errors are derived based on observations and not based on the mathematical fit error of the SCD retrieval.As pointed out by Stutz and Platt (1996), the fit error may underestimate the true error in the presence of erroneous reference cross-section alignment and systematically structured residual spectra.However, these malicious influences are difficult to quantify.In this work, empirically derived measurement errors are derived in order to provide a realistic error estimation which also includes the error of the column separation.Hence, it is possible to decide whether a measured BrO column density significantly exceeds the stratospheric background in the SCD space.This can be particularly advantageous when calculating the correlation to independent data sources by avoiding a systematic bias from potentially flawed assumptions of the vertical distribution and the state of the atmosphere, which are necessary to solve the radiative transfer.

Validation
The last section described the methods of a new satellite retrieval for tropospheric BrO column densities.Several parameters of the implementation to separate the tropospheric from the total column and the sensitivity filter algorithm for the boundary layer were determined by numerical inspection.The algorithm proved to be stable, and varying the different parameters within reasonable limits resulted in minor variations of the result.Due to its complexity, however, it is especially important to validate the presented algorithm in order to unravel potential flaws.
A validation requires independent measurements.Unfortunately, for tropospheric BrO columns, there is no independent satellite data to compare with and therefore the different steps of the algorithm are validated separately using either simulated data or measurements from instruments other than GOME-2: (1) the decomposition algorithm of the total BrO column is tested on simulated measurements (Sect.3.1) and ( 2

Proof of concept of column separation algorithm using simulated measurements
In Sect.2.2 we presented an algorithm to retrieve the ratio of stratospheric SCDs of BrO and O 3 .The algorithm mainly consists of a two-dimensional partitioning of the measurements (Appendix A) and an asymmetry filter.Here, the capability of the algorithm to retrieve the true z0 (ϑ, V N ) is benchmarked.
As a matter of fact, the true z0 is not known for the satellite nadir geometry.Therefore, the whole numerical algorithm is benchmarked using simulated measurements instead.The simulations are based on mathematical distributions without any a priori chemistry or radiative transfer.The retrieved z0 may then be compared to the known model function z m used as an input for the measurement simulation.
After the generation of measurements, the algorithm to derive BrO/O 3 -fractions is applied as described in Sect.2.2.The results are compiled in Fig. 9 with axes similar to Figs. 3, 4, and 5, respectively.The differences between z m and the retrieved surface z0 and the relative error σ 0 are illustrated in Fig. 9e and f, respectively.Both plots show that the algorithm succeeds in reproducing the model function within the sampled area.Residual linear structures of the difference are artefacts caused by the bilinear interpolation between the nodes of retrieved surface.The relative error almost never (<1 %) exceeds 2 %, and the relative mean error is 0.5 %.
In conclusion, the presented algorithm is capable of reproducing a given model surface for the stratospheric BrO/O 3 SCD ratio within the sampled area.The described simulator was used to test different combinations of parameters for the algorithm (number of nodes, partitioning scheme, interpolation method, convergence thresholds).The final implementation of parameters was found to provide a reasonable trade-off between resolution and sampling error.

Proof of concept of column separation algorithm using profiles simulated by EMAC
In addition to simulated measurements, it is also possible to benchmark the proposed column separation algorithm applying concentration profiles of BrO, O 3 , and NO 2 simulated by a chemistry climate model (CCM).SCDs of BrO and O 3 as well as VCDs of NO 2 are computed from an ensemble of profile triplets provided by the CCM and using radiative transfer calculations.Then, the algorithm presented in Sect.2.2 is applied on the computed SCDs and VCDs in order to retrieve again the stratospheric BrO SCDs.These BrO SCDs are compared to the original BrO SCDs and, hence, benchmarked.This approach is presented here and has two distinct advantages compared to the study in Sect.3.1: it incorporates radiative transfer effects which may lead to deviations due to differences in the concentration profiles, and the ensemble of computed values should be more realistic.The data basis for this study is concentration profiles of BrO, O 3 and NO 2 which were computed by the ECHAM5/MESSy Atmospheric Chemistry (EMAC) model described by Jöckel et al. (2010).This model, of which the results of a "nudged" (towards ECMWF analysis data) simulation in T42L90MA resolution are used, incorporates the Modular Earth Submodel System (MESSy) in the ECHAM5 global climate model (GCM) developed by the MPI for Meteorology, Hamburg.One distinct feature of the EMAC output is provided by the SORBIT submodel, which saves the result at the overpass times and locations of sun-synchronous satellite instruments like GOME-2 (Jöckel et al., 2010).Therefore, compared to the application of typical model output (global snapshots), a higher correlation between model and satellite measurement can be expected.It is noted that the output of EMAC used here features only a resolution of 128 times 64 grid cells in longitudinal and latitudinal direction, respectively.Therefore, model data of seven consecutive days between 22 and 28 March 2007 are used in order to increase the total number of different concentration profiles.The model profiles are filtered applying the same selection criteria as to the measurements (Sect.2.2.2).
An ensemble of n = 20 000 simulated satellite measurements of BrO, O 3 and NO 2 is generated from the EMAC profiles.n is similar to the typical number of measurements in one ψ-range.Hence, the choice of n is reasonable because only nadir measurements are considered here for the sake of simplicity.A random concentration between 10 and 40 ppt is added to the lowest 500 m of 50 % of the randomized BrO profiles in order to simulate events of enhanced near-surface BrO.From these profiles, the total SCDs of BrO and O 3 are computed using again the McArtim model applying a pure Rayleigh atmosphere without any aerosols and clouds, a random surface albedo between 3 % and 96 %, and the respective SZA of the profile.The computation of the NO 2 VCD is trivial.Finally, the column separation algorithm is applied on the simulated column measurements in order to retrieve a stratospheric BrO SCD S strat and its standard deviation σ strat according to Eqs. ( 15) and ( 16), respectively.
Figure 10a correlates the retrieved S strat to the "true" stratospheric BrO SCD S * strat without the random tropospheric BrO enhancement.An almost perfect correlation (r 2 = 0.99) is found between S strat and S * strat .The deviation of the slope (not shown) from the 1 to 1 line is of the order of the numerical error.Hence, it can be concluded that the proposed algorithm succeeded in retrieving the correct stratospheric BrO SCD with negligible systematic bias.This finding is particularly important because it indicates that the requirement of the column separation algorithm for sufficiently similar vertical profiles of BrO and O 3 is probably also fulfilled in reality.In reality, however, additional interferences due to clouds and more complex structures of the surface albedo may arise potentially decreasing the correlation.
Finally in this study, the differences between the retrieved and original BrO SCD strat = S * strat − S strat are compared to the σ strat as provided by the retrieval.Figure 10b shows the distribution of the strat divided by the retrieved σ strat .The red line is the normal probability density function with a standard deviation of unity.The agreement between the retrieved distribution and the model assumptions for normally distributed data is remarkable.Despite the small asymmetry, this figure demonstrates that the error computed by the proposed retrieval is a realistic estimate for the real measurement error of the separated stratospheric BrO SCD.

Comparison to AVHRR image data
As described in Sect.2.3, the sensitivity of GOME-2 measurements to surface near trace-gas concentrations is difficult to quantify over sea-ice and snow-covered land due to ambiguities in the optical properties of the surface (ice/snow) and clouds (water/ice).However, by comparing the results of the presented sensitivity algorithm to AVHRR reflectance measurements, it is possible to test the general response of the algorithm towards the shielding effect of (A) thin clouds over a dark lead, (B) thick clouds over ice, and (C) over only partially snow-covered land (Fig. 11).
The AVHRR/3 instrument is also borne by the MetOp-A satellite and measures reflectances at five spectral bands between the visible red and the thermal infra-red spectral range at a spatial resolution of 1.1 km.The black-and-white image in Fig. 11 shows AVHRR reflectance measurements at 630 nm (channel 1) of Northern Alaska and the Arctic Ocean from 4 April 2009 at 22:43:42 UTC.The scenery is dominated by large bright areas of sea ice in the Bering Strait, the Arctic Ocean, as well as snow between the northern coast of Alaska and the Brooks Range in the south.The colour-coded outlines of individual satellite pixels represent pixels assured to be sensitive to the surface with AMF min 500 = 3.While the center of the satellite swath features the highest A 500 (dark red pixels in the upper part), the algorithm manages to detect regions with a reduced sensitivity to the surface.Clearly, the sensitivity to the surface is reduced over dark surfaces like (A) the Barrow lead at the north-west coast of Alaska and (C) over the darker slopes of the Brooks Range to the bottom of the figure.A little bit more subtle (B) is the shielding effect of clouds in the east.The linear cracklike features in the sea ice are almost completely blurred by clouds which can be identified by their shadows towards the north-west.

Comparison to CALIPSO cloud data
In order to validate the selectivity and response of the presented sensitivity filter (Sect.2.3) towards clouds over bright surfaces, filter results are compared to measurements of the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) instrument.CALIOP is an active instrument measuring the time-resolved backscatter signal of a pulsed laser beam from which, among other parameters, the height and optical density of clouds may be derived independently from the surface albedo (NASA, 2006;Winker et al., 2007).CALIOP is the primary instrument carried by the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) satellite.
The specification of the satellites MetOp-A and CALIPSO and the measuring principle of the respective GOME-2 and CALIOP instruments differ fundamentally (see Table 4).The most dominant difference is the footprint of each instrument.CALIOP samples a single 70-m-wide cross-section of the atmosphere while GOME-2 averages over 3200 km 2 .CALIOP thus only probes 0.2 % of the atmospheric area within one GOME-2 pixel at most.A one-to-one comparison of GOME-2 and CALIOP measurements is therefore problematic, but it is still possible to compare averages assuming the cloud properties CALIOP measures are to some extent representative for the whole GOME-2 pixel.Furthermore, CALIPSO flies on different orbit than MetOp-A.CALIPSO crosses the Equator around 13:30 LT in ascending node while MetOp-A has an Equator crossing time of 09:30 LT in descending node.In polar regions, however, the orbits of both satellites partly overlap.The time difference between a CALIPSO and a MetOp-A overpass varies periodically and there are chances for almost simultaneous measurements.
In this study, four years of provisional CALIPSO Lidar Level 2 5 km cloud layer data are compared to the classification of GOME-2 pixels regarding the sensitivity to the surface (Sect.2.3).The comparison focuses on the ability of the algorithm to detect clouds over bright surfaces possibly reducing the sensitivity to trace-gases at the surface.Therefore, only pixels featuring a high sea-ice concentration of 95 % are considered here.Sea-ice concentration maps derived from microwave-radar measurements were provided by the Integrated Climate Data Center (ICDC); see Kaleschke et al. (2001) andSpreen et al. (2008) for a detailed description of the product.Additionally, only measurements in the Northern Hemisphere below 83 • N latitude are compared.The time difference between both measurements is limited to 30 min, and every GOME-2 pixel taken into account must contain at least 70 km of the CALIPSO ground-track.Finally, 15 374 collocated measurements meet these selection criteria in the months February to June of the years 2007 to 2010.
Depending on AMF min 500 , two properties of the CALIPSO data set, the cloud optical thickness (COT) and layer top altitude (LTA) of the uppermost layer, are selected, averaged and classified following the sensitivity algorithm applied on the specific GOME-2 pixel they are collocated with.Figure 12 shows the comparison between the surface sensitivity filter, retrieved A 500 and collocated CALIPSO measurements.The comparison to COT (Fig. 12, left column) is discussed first and followed by the comparison to LTA (Fig. 12, right column).
The histogram in Fig. 12a shows the distribution of all collocated measurements compared to measurements classified as possibly obscured by the sensitivity filter for different thresholds AMF min 500 = 0.5, . . ., 3.5.There are two accumulation points: one for COT<1 and another between 3 and 3.5 COT.The first accumulation point is due to essentially cloudfree pixels, and the second one is probably caused by clouds that are optically thicker than can be resolved by CALIOP leading to a systematic underestimation for these clouds.For increasing AMF min 500 , however, an increasing percentage of measurements are flagged as possibly obscured which is also shown in Fig. 12c. Figure 12c furthermore illustrates that the percentage of flagged measurements increases with increasing COT and the choice of AMF min 500 as expected.Hence, it may be concluded that the proposed surface sensitivity filter is COT selective over sea ice and able to classify the majority of pixels with high COT as possibly obscured.The dependence of A 500 on COT plotted in Fig. 12e confirms that a larger COT on average leads to a smaller surface sensitivity.
The right column in Fig. 12 shows the respective plots for the LTA revealing a similar but weaker dependence of the sensitivity filter on LTA than COT.This is not surprising since there is presumably some cross-correlation between COT and LTA because clouds with a larger top altitude are potentially optically thicker.The histogram in Fig. 12a shows one dominating accumulation point for LTA < 1 km caused by cloud-free CALIOP measurements which are set to LTA = 0. Therefore, the dependence of the number of measurements classified as possibly obscured (Fig. 12d) shows the strongest gradient between 0 and 2 km LTA.The dependence on LTA vanishes between 2 and 8 km but increases again for high clouds (LTA>8 km).Finally, Fig. 12f shows the dependence of A 500 on LTA.The linear fit to all data shows a relatively slowly decreasing slope (black line).The slope becomes steeper, however, if only measurements below 2 km LTA are taken into account (red line).Hence, A 500 depends more strongly on the presence of low clouds which results from the concentration profile of O 4 , whose slope decreases with altitude.
Within the limitations of the CALIOP data set (relatively low maximum cloud optical depth, which can be measured) and of a comparison of different data sets, it can be concluded that the algorithm is capable of identifying the shielding effect of clouds over sea ice.GOME-2 pixels with a higher average COT and LTA are more likely classified as possibly obscured.A higher threshold AMF min 500 increases the sensitivity of the filter towards filtering thinner and higher clouds.The LTA, however, was expected not to play such an important role because A 500 is almost constant for clouds higher than 500 m.
The dependence on LTA illustrates the limits of the presented filter approach based on the utilisation of O 4 as a tracer for near-surface air.O 4 is also abundant above 500 m altitude implying the following limitations.Firstly, the shielding effect of very low and optically thick clouds may be underestimated because, in this case, S O 4 is almost not affected.Secondly, a pixel may also be filtered although the measurement is sensitive to the BrO present in that pixel.Therefore, filtered measurements are flagged as only possibly obscured.Example scenarios that appear as obscured but are in fact sensitive could be either a layer of BrO over a relatively dark surface elevated high enough to be detected anyway or near-surface BrO residing below high, optically thin clouds over a rather bright surface which may reduce S O 4 more strongly than the real A 500 .The strength of the presented filter algorithm, however, is that measurements flagged as sensitive are very likely to be actually sensitive to near-surface BrO as the first limitation can be assumed to be less frequent in reality than the second.

Comparison to ground-based measurements
In this section, tropospheric BrO VCDs from GOME-2 are compared to both LP-DOAS and MAX-DOAS measurements of BrO obtained during two Arctic field campaigns, respectively.
During March andApril 2008, Pöhler et al. (2010) measured BrO in the tropospheric boundary layer directly over the sea ice of the Amundsen Gulf from aboard the Amundsen research icebreaker.The data set from the Amundsen includes LP-DOAS and yet unpublished MAX-DOAS measurements.The LP-DOAS measured concentrations of BrO averaged between 1 and 19 m height above the sea ice.From MAX-DOAS measurements, BrO VCDs were calculated using the differential SCD between 10.6 • and 90 • divided by the differential AMF assuming clear-sky conditions and a surface albedo of 0.99 (Grenfell et al., 1994).The other data set was collected between March and April 2009 during the Ocean-Atmosphere Sea-Ice Snowpack (OA-SIS, http://oasishome.net/)field initiative at Barrow, Alaska.The LP-DOAS measurements were presented in Liao et al. (2011).Vertical profiles of BrO and aerosols were retrieved from MAX-DOAS data using optimal estimation as described by Frieß et al. (2011).Tropospheric BrO VCDs were determined by integrating the retrieved MAX-DOAS profiles.
Both ground-based and satellite measurements offer their particular advantages and disadvantages to study the same phenomenon.Resolution and coverage differ between both approaches spatially and temporally.Ground-based instruments usually offer a higher spatial and temporal resolution, whereas satellite measurements observe the same property over a vast area at moderate spatial and temporal resolution.In conclusion, a high correlation may only be expected if both techniques sample the same volume of air, and spatial as well as temporal variabilities are small.

Collocating satellite data and ground-based measurements
In order to assure spatial and temporal coherence of groundbased and satellite measurements, coincident measurements have to be selected.Averages of collocated subsets are calculated for each overpass of the satellite.Note that a polar orbiting satellite may pass the same site several times a day during daylight depending on latitude, season, and type of swath.The swath of GOME-2 covers the studied sites up to three times a day at an SZA below 80 • .First of all, only those satellite measurements are selected that are sensitive to the ground (Sect.2.3) and whose pixel footprint includes the location of the ground-based measurement.In a second step and for each satellite pixel, the time interval is calculated in which ground-based measurements are averaged.This calculation combines surface wind-speed and direction with the relative location of the measurement site within the GOME-2 pixel.If the measurement site is located close to the pixel edges, the averaging time-interval includes only those measurements corresponding to the air probed by both the satellite and ground-based instrumentation.This approach leads to a mean duration of approx.three hours which is limited to two hours prior to and after the time of the satellite overpass.Correlation coefficients as well as slope a and y-intercept b of a linear bivariate model (Cantrell, 2008) are calculated based on these overpass averages.Additionally, daily means including the overpass averages of each day are computed.

Results from comparing satellite with ground-based measurements
This section summarizes the results from both field campaigns, which are then discussed in Sect.4.3.
For the Amundsen measurements, the time series of MAX-DOAS and LP-DOAS are shown together with the GOME-2 overpass data using a sensitivity threshold of AMF min 500 = 1 in Figs. 13 and 14 and the corresponding correlation plots in Fig. 18a and 18b, respectively.The comparison to the MAX-DOAS measurements encompasses more than a month beginning on 9 March 2008.Both instruments, MAX-DOAS and GOME-2, captured several events of elevated tropospheric BrO VCDs including one major event around 14 March.Figure 15 shows a map of this particular event.The GOME-2 data furthermore reveal another particularly strong enhancement on 16 April 2008.The amplitudes of the collocated MAX-DOAS VCD time series are almost identical (slope close to unity), but there is a significant bias.The tropospheric column densities retrieved from GOME-2 measurements are systematically smaller by approx.3 × 10 13 molec cm −2 than measured by MAX-DOAS (Fig. 18a).The LP-DOAS measurements of the BrO mixing ratio, however, encompass only 16 days with collocated GOME-2 measurements interrupted by 5 days of cruise.The LP-DOAS measured up to 42 pmol mol −1 of BrO on 15 March when major enhancements were also observed by GOME-2 (see Supplement).The slope in Fig. 18b is approx.500 m, which represents an estimate for the BrO layer thickness.
The time series of both MAX-DOAS and LP-DOAS measurements at Barrow are shown in Figs. 16 and 17,respectively.More than a month of collocated measurements with GOME-2 from mid-March to mid-April 2009 are available.Compared to the measurements from aboard the Amundsen, the amplitudes of BrO enhancements (VCDs and mixing ratios) were generally smaller at Barrow.The correlations between GOME-2 and MAX-DOAS (Fig. 18c) as well as GOME-2 VCD and LP-DOAS mixing ratio (Fig. 18d) are weaker but statistically significant (n = 77, r 2 = 0.1, p = 0.005).The slope in Fig. 18c reveals that the BrO VCDs retrieved from MAX-DOAS measurements are approx.twice as high as the collocated GOME-2 measurements.However, the offset between both data sets is significantly smaller than for the Amundsen measurements.
It is important to note that the temporal variability of LP-DOAS mixing ratios at Barrow occasionally deviates from GOME-2 tropospheric VCDs (Fig. 17).There are several days where the LP-DOAS measured above 10 pmol mol −1 while the GOME-2 VCDs are close to zero (16,23,29 March,and 11 April).On 14 March, however, both GOME-2 and MAX-DOAS show significantly elevated BrO columns while the LP-DOAS measured comparatively moderate 7 pmol mol −1 .These differences are discussed in Sect.4.3.
Finally, the dependence on the sensitivity threshold AMF min 500 is studied.The correlation coefficient r 2 is calculated for all four ground-based data sets and collocated GOME-2 measurements for different AMF min 500 between 0.5 and 4 (Fig. 19).Furthermore, the respective number of remaining collocated measurements are shown.As already mentioned, the correlation between satellite and groundbased measurements is larger for the Amundsen than for the Barrow data.The increase in r 2 with AMF min 500 is consistent for both MAX-DOAS comparisons indicating that the proposed sensitivity filter in fact identifies measurements with ambiguous sensitivity (Fig. 19a and c).For the LP-DOAS measurements, however, the trend for r 2 is less clear.For the Amundsen data, the correlation with LP-DOAS mixing ratios decreases significantly for AMF min 500 > 2. At Barrow, r 2 increases only for rather high thresholds AMF min 500 ≥ 3 when more than half of the collocated measurements are filtered.For AMF min 500 = 3.2 the correlation between LP-DOAS and GOME-2 is r 2 = 0.19 (see discussion bellow).Gridded maps of daily satellite measurements corresponding to both time series can be found in the Supplement to this paper.

Discussion of comparison with ground-based measurements
The comparisons between ground-based and satellite measurements of BrO show a good agreement, demonstrating the capability of the presented method to retrieve realistic  (Fig. 18b) in agreement with the height of the simultaneously measured O 3 -depleted layer reported by Seabrook et al. (2011).The map of tropospheric BrO column densities (Fig. 15) shows a particularly large area affected by bromine activation.It is hence concluded that this particular "BrO cloud" is located at the surface but not in elevated layers, at least not at the location of the Amundsen vessel.At Barrow, however, the correlation between LP-DOAS and GOME-2 measurements was much weaker than seen on the Amundsen vessel.There are several potential explanations for this behaviour.
i.The Barrow time series is much longer and the correlation comprises several strong events of elevated BrO levels.The height of the chemically perturbed boundary layer varied strongly between these events (Frieß et al., 2011;Helmig et al., 2012).Therefore, the assumption of a linear correlation may not be appropriate.
ii. Located onshore, the local meteorology as well as surface processes related to bromine activation at Barrow may differ fundamentally from that over the sea ice.Furthermore, ground-based O 3 measurements at Barrow revealed large horizontal heterogeneities (Liao et al., 2011;Helmig et al., 2012) which may bias the LP-DOAS measurements with respect to the satellite.
iii.The uncertainties of the sensitivity of the satellite measurement to the surface due to the nearby opening in the sea ice are large (cf.Sect.3.3).In fact, increasing AMF min 500 at Barrow resulted in an increased r 2 supporting this hypothesis.However, the comparison between MAX-DOAS and GOME-2 measurements at Barrow shows a good agreement despite the offset.Furthermore, the correlation of surface concentrations as measured by LP-DOAS and retrieved from MAX-DOAS measurements is significant (Frieß et al., 2011, Fig. 3).This indicates that the VCDs measured by GOME-2 are realistic, and, hence, the variations of the BrO profile at Barrow were probably larger than during the Amundsen campaign where only one major event was captured by LP-DOAS.
Occasionally enhanced near-surface BrO concentrations correspond to a shallower average mixing height of BrO as reproduced in Fig. 18d.In the following, two example days are selected in order to illustrate the vertical variability of the BrO profile.As mentioned above, the LP-DOAS data in Fig. 17 show on several days considerably higher enhancements than GOME-2.This discrepancy can be explained by a very  (Frieß et al., 2011).
shallow layer of BrO (Fig. 20b).All shown MAX-DOAS profiles decrease to background values at altitude above 250 m.Another interesting event was captured at Barrow on 14 March 2009.GOME-2 observed a significantly elevated BrO column while the LP-DOAS shows only moderate levels of bromine activation.However, the comparison with the integrated tropospheric VCD from MAX-DOAS again shows good agreement, and the apparent discrepancy between LP-DOAS and GOME-2 therefore indicates the presence of an elevated layer of enhanced BrO concentrations as suggested by e.g.Hönninger and Platt (2002), Wagner et al. (2007), and Frieß et al. (2011).In this context, the term elevated denotes a layer of enhanced BrO at a few hundred metres which are still within the boundary layer.On this day, O 3 levels were below 1 nmol mol −1 suggesting a limited production of BrO at the surface (Helmig et al., 2012).As sampled by an ozonesonde launched at Barrow (Fig. 20a), the O 3 mixing ratio increases with altitude allowing for a more efficient BrO production at these altitudes.Furthermore, the potential temperature gradient profile suggests a highly stratified boundary layer as well as a strong temperature inversion at the surface (∼ 50 m) hampering O 3 mixing from aloft.This conclusion is supported by BrO profiles retrieved from MAX-DOAS measurements from the same day as depicted in Fig. 20a (Frieß et al., 2011).Despite one outlier retrieved from measurements around 11:30 AST, all retrieved profiles feature a positive BrO gradient close to the surface.However, most BrO is still located within the boundary layer.
From the data presented in this study, we conclude that events of enhanced BrO are well captured by satellite measurements, and that ground-based observation and tropospheric VCDs retrieved from GOME-2 data are significantly correlated.The near-surface concentrations measured by LP-DOAS furthermore indicate that the satellite observations in turn are linked to surface processes as observed from aboard the Amundsen.At Barrow, however, deviations from this general dependence could be explained by local meteorological perturbations and variations of the surface BrO column height.Occasionally, satellite measurements may underestimate the influence by bromine activation when the surface BrO layer is extremely shallow and horizontal gradients of both the chemistry and the surface sensitivity are large.The general applicability of this observation, however, needs to be tested further because the presented study comprises only a relatively small number of ground-based observations over the sea ice.

Conclusions
We present a new algorithm to retrieve residual tropospheric columns of Arctic BrO solely based on data from a single satellite instrument.Two important properties of our algorithm are that it identifies measurements (a) with significantly enhanced tropospheric BrO amounts which cannot be explained by stratospheric processes and (b) are sensitive for near-surface layers.Unlike earlier attempts to solve this task, the presented approach does not depend on extensive chemistry models and climatological data.Only potential vorticity fields are supplied from external sources allowing to identify the polar vortex, where the retrieval algorithm is not applicable.This procedure is necessary to provide a consistent data set without artefacts caused by a disturbed stratospheric chemistry.Based on this work, possible surface processes involved in Arctic bromine activation can be studied.
Both the decomposition of the total column of BrO into stratospheric and residual tropospheric contribution as well as the surface sensitivity filter algorithm were validated through real measurements and simulated data.The resulting tropospheric BrO columns were compared to four independent BrO ground-based data sets and significant correlation was found.The comparison to ground-based data from two field campaigns taking place on-and off-shore confirmed that near-surface processes are the source of activated bromine compounds.While the correlation with MAX-DOAS VCDs was generally significant (r 2 > 0.6), especially the comparison with LP-DOAS measurements at Barrow is less straightforward.However, the major deviations between LP-DOAS and GOME-2 could be explained by the stratification of the surface layer and the profiles of both BrO and O 3 .Furthermore, it is occasionally possible that shallow surface layers and horizontal heterogeneities may obfuscate active bromine chemistry from satellite measurements.
Though only GOME-2 measurements were analysed, the presented algorithm is applicable to measurements performed by similar satellite instruments like GOME, SCIA-MACHY, and the Ozone Monitoring Instrument (OMI) as well.Additionally, the algorithm to determine the sensitivity to trace-gas concentrations close to the surface is neither limited to retrievals of BrO nor to the Arctic.Any DOAS retrieval from satellites intended to study surface concentrations over bright surfaces may, in principle, apply the presented surface sensitivity filter, e.g.retrievals of iodine monoxide (IO) in the Antarctic troposphere.

Computing of a partition
A partition β of = [ϑ a , ϑ b ] × [V N a , V N b ] × [ψ a , ψ b ] is computed as follows.First of all, all previously selected reference measurements T are divided in the ψ-direction into N ψ bins defined by the limits ψ k with k = 1, . . ., N ψ − 1 in order to separate the weak ψ-dependence from the following algorithm.Also, the viewing zenith angles are almost equally distributed over the entire instrument swath.In the ϑ and V Ndirection, however, a partitioning algorithm accounting for non-uniform distributions is required.Partitions all containing a similar number of observations are desirable in order to achieve homogeneous statistics of the asymmetry filter applied on every partition separately.
The two-dimensional partitioning algorithm applied on each ψ-bin is based on a partition of which can be indexed by two indices i, j for the ϑ and V N -direction, respectively: (A1) The necessary steps are explained by means of a concrete example.
In order to sample the z-surface in the two dimensions of the (ϑ, V N )-plane, T is subdivided into N ϑ times N V partitions i,j ⊂ with i = 1, . . ., N ϑ and j = 1, . . ., N V , respectively.The partitioning is performed on the twodimensional domain : (1) T is divided into N ϑ preliminary partitions along the ϑ-axis each containing an equal number of observations except the last two columns (i = N ϑ − 1, N ϑ ), whose sizes are weighted by two in order to allow for a higher density of bins at this edge of the domain.The borders between the partitions are denoted h i (i = 1, . . ., N ϑ − 1) in Fig. A1.(2) Subsequently, the preliminary partitioning nodes defining the boundaries of each partition in V N -direction are calculated.For each i = 1, . . ., N ϑ − 1, we construct a set union T i : which contains all observations in the column left and right of the respective h i .Each T i is then divided into N V partitions in V N -direction defining the (ϑ, V N )-coordinates of the preliminary partitioning nodes.This time, however, the first and last partition (top and bottom row) contain only half as many observations compared to the six partitions in between.
The modification to the basic scheme of partitions containing an equal number of measurements prevents the outer partitions from becoming too large at the cost of retrieval noise.

1Fig. 1 .
Fig. 1.Illustration of the decomposition of the total BrO VCD measured by GOME-2 into stratospheric and tropospheric contribution for 25 March 2009.The top row shows VCDs of BrO (a) and O 3 (b) assuming a geometric AMF.Coinciding spatial structures of enhanced VCDs are visible e.g. over Eastern Europe and Northern Siberia, which are attributed to stratospheric dynamics and variations of the tropopause height (d).The BrO SCD strat (c) is retrieved from measurements of BrO, O 3 , and NO 2 alone.The BrO VCD trop (e) is the difference between (a) and (c).The PV at the 475K isentrope (f) may be used to identify regions within the polar vortex (see Sect. 2.2.2).All VCDs are calculated using a geometric AMF; gray areas contain no data.

1Fig. 2 .
Fig. 2. Same as in Fig. 1, but for 1 April 2007.The white contour in (f) marks the 75 PVU-isoline at the 550 K isentrope.The decomposition into stratospheric and tropospheric column fails within the polar vortex, because there is no clear correlation between O 3 VCD and the tropopause height (d) any more.

1Fig. 3 .
Fig. 3. Partitioning of GOME-2 reference measurements (colourcoded density) in the (ϑ, V N )-plane for the near-nadir direction (|ψ| ≤ 14 • ) for 25 March 2009.Each partition contains an almost equal number of measurements from which BrO/O 3 SCD ratios are retrieved.The example partition 6,3 is depicted in Fig. 4. The mean of measurements within each partition (blue crosses) is used as nodes for interpolating the results (Fig. 5).

1Fig. 4 .
Fig. 4. Frequency distribution of measured BrO SCD to O 3 SCD ratios (blue) from the example partition 6,3 in Fig. 3.The algorithm retrieves the limits of a subdivision with minimal asymmetry (green) containing mostly measurements of the stratospheric background.Significantly enhanced measurements (high BrO, low O 3 ) appear in the right tail of the distribution (red).See text for details.

Fig. 5 .
Fig. 5. Interpolation (a) of the BrO/O 3 SCD ratio surface and (b) its standard deviation σ depending on SZA and NO 2 VCD.The nodes of the bilinear surface interpolation (squares) are the mean of the partitions displayed in Fig. 3.The distributions of some partition with a negligible asymmetry are not filtered before the interpolation and indicated by white squares.

1Fig. 6 .
Fig. 6.Illustration of the sensitivity filter and tropospheric AMF applied on GOME-2 measurements for 25 March 2009 (same as Fig. 1).(a) The retrieved tropospheric BrO VCDs are filtered according to the respective minimum sensitivity to trace gas concentrations close to the surface using (b) measured AMFs of O 4 .Panels (c) through (f) show tropospheric BrO VCDs for different sensitivity thresholds AMF min 500 = 0.5, 1, 2, 3, respectively.Note that the sensitivity to the choice of AMF min 500 is low.(a) is calculated using A geom ; (c)-(f) are calculated using A 500 .Areas without any sensitive measurements are left gray.

1Fig. 7 .Fig. 8 .
Fig. 7. (a) Modelled and interpolated (R, A O , A 500 )-triplets for a nadir geometry at SZA = 76 • .The convex hull H (shaded area) including all A 500 < 1 = AMF min 500 is parameterised in order to provide a threshold for the surface sensitivity filter.(b) Classification of all GOME-2 nadir observations of 2008 at the same SZA based on measured R and A O with a threshold of AMF min 500 = 1.The described filter distinguishes between measurements sensitive to the lowest 500 m of the atmosphere (black dots) and those that are possibly obscured by clouds and/or too low albedo (grey area, bright dots).The convex hull (magenta) of modelled values contains approx.88 % of the measurements.
) using concentration profiles of BrO, O 3 and NO 2 provided by atmospheric chemistry model simulations (Sect.3.2).(3) Results of the surface sensitivity filter are validated through a case study of imaging satellite data in the red spectral region (Sect.3.3) and (4) compared to optical properties of clouds measured by the CALIOP instrument (Sect.3.4).(5) The comparison of retrieved tropospheric VCDs to ground-based measurements of BrO is described in Sect. 4. It is noted that cross-validations to airborne DOAS measurements have already been published in Prados-Roman et al. (2010) and Heue et al. (2011).

1Fig. 9 .
Fig. 9. Benchmark of the separation algorithm (Sect.2.2 and Appendix A) using simulated measurements modelling a (a) known surface.(b) Partitioning of measurements, (c) interpolated surface nodes, (d) application of asymmetry filter.(e) The difference between true and retrieved surface function shows only small deviations.(f) The relative error almost never exceeds 2 %.

1Fig. 10 .
Fig. 10.Benchmark results of the column separation algorithm using an ensemble (n = 20000) of concentration profiles of BrO, O 3 , and NO 2 simulated by the EMAC model.(a) Retrieved stratospheric BrO SCDs against "true" input BrO SCD.(b) Distribution of the difference between retrieved and input BrO SCD normalised by the BrO standard deviation σ strat as provided by the retrieval (see text).

1Fig. 11 .
Fig. 11.Overlying image of derived A 500 (colour-coded) and AVHRR reflectance measurements (monochrome background, channel 1, 630 nm) over Northern Alaska and the Arctic Ocean.Blue quadrangles indicate GOME-2 satellite pixels with an assured sensitivity to near-surface absorbers (A 500 ≥ 3).Pixels above leads at the north-west coast of Alaska (A), above clouds (B), and over the dark Brooks Range (C) are labelled possibly obscured (A 500 < 3) and not plotted.

1Fig. 12 .
Fig. 12.Comparison between results of the surface sensitivity filter and collocated CALIPSO measurements over sea ice: CALIPSO cloud optical thickness (COT, left column) and CALIPSO layer top altitude (LTA, right column).(a) and (b) histograms of the unfiltered measurements compared to the histograms of measurements identified as possibly obscured at different sensitivity thresholds AMF min 500 ; (c) and (d) ratio of filtered measurements depending on COT and LTA, respectively; (e) and (f) respective dependence of A 500 for AMF min 500 = 0.5.

Fig. 18 .Fig. 19 .
Fig. 18.Correlation and orthogonal fit of the different ground-based data sets with VCD trop measured by GOME-2 (AMF min 500 = 1).Note that the slopes in (b) and (d) are given in unit metre comparable to the height of the mixing-layer (assuming ρ air = 2.9 × 10 19 molec cm −3 ).Error bars are omitted for the sake of clarity.
Fig. 20.(a) Profiles of potential temperature, O 3 , and BrO mixing ratio at Barrow on 14 March 2009.(b) BrO profile on 29 March 2009.The marks at the bottom of the BrO profiles denote the values measured by LP-DOAS.Potential temperature and O 3 are measured by ozonesonde (courtesy of NOAA Earth System Research Laboratory).BrO profiles are retrieved from MAX-DOAS measurements(Frieß et al., 2011).

Table 1 .
Compilation of fit ranges, reference cross-sections and parameters of the three DOAS evaluations of calibrated radiance spectra measured by GOME-2.The slant column densities (SCDs) of BrO, O 3 , O 4 and NO 2 are retrieved.The synthetic Ring spectra account for (wavelength-dependent) inelastic Raman scattering, and the reciprocal intensity spectrum accounts for instrumental stray light (see text).The fit-polynomial models broadband absorption.

Table 3 .
Modelled layers of scattering media (aerosols and/or clouds) defined by their lower and upper edge over ground and the optical density (OD).