Effects of long-range aerosol transport on the microphysical properties of low-level liquid clouds in the Arctic

The properties of low-level liquid clouds in the Arctic can be altered by long-range pollution transport to the region. Satellite, tracer transport model, and meteorological data sets are used here to determine a net aerosol-cloud interaction (ACI) parameter that expresses the ratio of relative changes in cloud microphysical properties to relative variations in pollution concentrations while accounting for dry or wet scavenging of aerosols en route to the Arctic. For a period between 2008 and 2010, ACI is calculated as a function of the cloud liquid water path, temperature, altitude, specific humidity, and lower 5 tropospheric stability. For all data, ACI averages 0.12± 0.02 for cloud droplet effective radius and 0.16± 0.02 for cloud optical depth. It increases with specific humidity and lower tropospheric stability and is highest when pollution concentrations are low. Carefully controlling for meteorological conditions we find that the liquid water path of arctic clouds does not respond strongly to aerosols within pollution plumes. Or, not stratifying the data according to meteorological state can lead to artificially exaggerated calculations of the magnitude of the impacts of pollution on arctic clouds. 10


Introduction
Due to growing concentrations of greenhouse gases and complex feedback processes, the Arctic region has warmed approximately two times faster than the global average (Serreze and Francis, 2006;Serreze et al., 2009;Richter-Menge and Jeffries, 2011), a trend that is anticipated to continue through this century (Yoshimori et al., 2013;Overland and Wang, 2013). Further, the Arctic is not pristine, even if it is remote from industrialized areas and major aerosol sources. Mid-latitude aerosols can 15 be transported to northern latitudes in relatively high concentrations when precipitation rates are low and there are strong temperature inversions that inhibit vertical mixing (Sirois and Barrie, 1999;Law and Stohl, 2007;Quinn et al., 2007;Law et al., 2014). The origins of arctic haze tend to be pollution from Eurasia (Shaw, 1995;Stohl, 2006;Shindell et al., 2008;Ancellet et al., 2014), and boreal forest fires in North America, Eastern Europe and Siberia (Stohl, 2006;Stohl et al., 2007). Between spring and summer, the atmosphere becomes cleaner due to an increase in wet-scavenging (Garrett et al., 2010). 20 Such aerosols have the potential to alter cloud properties in the Arctic (Garrett and Zhao, 2006;Lance et al., 2011). On one hand, thin low-level clouds with more numerous smaller droplets can radiate more long wave radiation thereby warming the surface (Garrett et al., 2002(Garrett et al., , 2004Garrett and Zhao, 2006). On the other, polluted clouds can reflect more sunlight, leading to a cooling effect (Lubin and Vogelmann, 2007). Zhao and Garrett (2015) found that seasonal changes in surface radiation associated with haze pollution range from +12.2 W m −2 in February to −11.8 W m −2 in August. Annually averaged, the longwave warming and shortwave cooling nearly compensate, although the seasonal timing of the forcings may have implications for rates of sea ice melt (Belchansky et al., 2004;Markus et al., 2009).
The influence of aerosols on cloud microphysical properties is often quantified using an indirect-effect (IE) or aerosolcloud interactions (ACI) parameter that expresses the ratio of relative changes in cloud microphysical properties to variations in pollution concentrations, most typically aerosol index, the aerosol optical depth, the aerosol concentration, or the cloud condensation nucleus (CCN) concentration (Feingold et al., 2001;Feingold;2003b). Where the parameter is expected to decrease with increasing aerosols or cloud condensation nuclei (e.g. the effective radius), the ratio is multiplied by negative one so that the IE or ACI is positive. Garrett et al. (2004) used ground-based retrievals of cloud-droplet effective radius and surface measurements of dried aerosol 10 light scattering from Barrow (Alaska) to obtain a value for the dloud-droplet effective radius ACI that lies between 0.13 and 0.19. Satellite measurements show that ACI values for cloud-droplet effective radius range from 0.02 to 0.20 for mid-latitude continental clouds (Nakajima et al., 2001;Feingold, 2003a;Lohmann and Feichter, 2005;Myhre et al., 2007) and from 0.03 to 0.15 for mid-latitude oceanic clouds (Bréon et al., 2002;Sekiguchi, 2003;Kaufman et al., 2005;Myhre et al., 2007;Costantino and Bréon, 2013;Wang et al., 2014). Satellite instruments have the advantage of providing data over large spatial scales, 15 however satellite retrievals of aerosol concentrations are normally obtained from air columns close to the analyzed cloud. The assumption is made that plumes are horizontally homogeneous both within and without the cloud, and that they are vertically co-located with cloud top (Nakajima et al., 2001;Feingold et al., 2001;Sekiguchi, 2003). For large-scale cloud studies, this method potentially introduces bias since it is not obvious that pollution should be uniform for different meteorological regimes.
Co-locating satellite cloud retrievals with pollution tracer output from a chemical transport model offers an alternative 20 approach for assessing the effect of pollution on clouds (Berg et al., 2011;Lance et al., 2011;Tietze et al., 2011). Cloud microphysical properties and pollution concentrations can be estimated at the same time, location, and meteorological regime (Schwartz et al., 2002;Kawamoto et al., 2006;Avey et al., 2007). Active tracers experience both sources and sinks through wet scavenging, dry deposition, and chemical reactions that can be difficult to accurately model. Passive pollution tracers, on the other hand, are determined only by source emission strength and subsequent dilution. An example of a passive tracer 25 is carbon monoxide (CO), which is a combustion by-product that correlates with the anthropogenic CCN close to pollution sources (Longley et al., 2005). Since both are found in pollution plumes, CO can serve as a passive proxy for CCN that is relatively straight-forward to model. A difference is that, unlike CCN, CO is unaffected by wet and dry scavenging. In the absence of scavenging, a linear relation exists between CCN and CO and it should be possible to see changes in clouds when CO concentrations are high. But if CCN have been scavenged from pollution plumes, then the observed sensitivity of clouds 30 to the pollution plumes should be expected to be low. Thus passive tracers serve as an indicator of the net effect of pollution plumes on clouds that accounts for the effect of scavenging during transport en route to the clouds themselves.
More generally, the primary control on cloud microphysical properties is not aerosols but rather meteorological conditions during cloud formation (Chang and Coakley, 2007;Brenguier and Wood, 2009;Kim et al., 2008;Painemal et al., 2014;Andersen and Cermak, 2015). For example, a reduced stability of the environmental temperature profile can allow for enhanced 35 cloud-droplet growth through increasing convection (Klein and Hartmann, 1993). This would be expected to lead to greater mixing of the aerosols with the cloudy air and greater aerosol impacts on cloud microphysical properties (Chen et al., 2014;Andersen and Cermak, 2015). Also, in the Arctic during the winter, pollution plumes from Asia are often associated with higher values of potential temperature than pollution plumes from Europe (Stohl, 2006). Thus, the observed impact of pollution plumes on clouds may be correlated with a particular meteorological regime.

5
Using the approach of co-locating a passive tracer from a tracer transport model with satellite observations, Tietze et al. (2011) presented an analysis of pollution-cloud interactions over the Arctic from March to July 2008. Anthropogenic and biomass burning aerosol pollution was represented using a CO passive tracer in the FLEXPART (FLEXible PARTicle dispersion model) tracer model (Stohl et al., 2005) as a proxy. CO was co-located with POLDER-3 (Polarization and Directionality of the Earth's Reflectance) and MODIS (Moderate Resolution Imaging Spectroradiometer) observations. Tietze et al. (2011) 10 showed that the sensitivity of liquid cloud effective radius (r e ) and optical depth (τ ) to pollution has a maximum around the freezing point, and that the sensitivity decreases for both higher and lower temperatures. The optical depth was generally up to four times more sensitive than the effective radius. Their results also suggested that biomass burning pollution has a smaller yet significant impact on liquid cloud microphysical properties than anthropogenic pollution, and that the ACI parameters depend on altitude, LWP, and temperature. 15 Our study extends the Tietze et al. (2011) research by adding two years of data, 2009 and 2010 and, in addition to temperature, stratifying the results by lower tropospheric stability (LTS) and atmospheric specific humidity (Matsui et al., 2006;Mauger and Norris, 2007). Our results highlight the importance of considering meteorological conditions when assessing the aerosol impact on cloud microphysical properties to show that r e and τ have similar sensitivities to pollution.

20
The analyses in this study are based on a co-location of satellite retrievals of cloud properties, tracer transport model simulations of pollution locations and concentrations, and reanalysis data sets for meteorological fields.

Satellite cloud property retrievals
In this study we used two instruments, both part of the A-train mission (Stephens et al., 2002). The MODIS instrument on board the Aqua satellite measures radiation in 36 different spectral bands with central wavelength from 400 nm to 14 400 nm 25 in wavelength. For the effective radius, optical depth, and cloud top temperature we use Collection 5 Level-2 products (Platnick et al., 2003;King and Platnick, 2006). Regarding the technique applied for computation of the MODIS Level-2 products cloud top temperature is derived from the 11 µm infrared band. Cloud-droplet effective radius (r e ) and optical depth (τ ) are retrieved from simultaneous cloud-reflectance measurements in three water absorbing bands (1.6, 2.1, 3.7 µm) and three nonabsorbing bands (0.65, 0.86, 1.2 µm) (Platnick et al., 2003). The pixel resolution of the retrievals at nadir is 1 km×1 km for 30 cloud microphysics and 5 km×5 km for cloud top temperature.
The POLDER-3 camera on the PARASOL satellite platform (Polarization & Anisotropy of Reflectances for Atmospheric Sciences coupled with Observations from a Lidar) captures a wide field of view through spectral, directional, and polarized measurements of reflected sunlight (Fougnie et al., 2007). Multidirectional observations allow for a pixel to be observed from up to sixteen different view angles. The instrument measures radiance in 9 spectral channels between 443 and 1020 nm, including three polarized channels at 490, 670 and 865 nm. POLDER-3 cloud microphysical properties retrievals have a 36 km×36 km 5 spatial resolution. Cloud top pressure is derived from the cloud oxygen pressure (Bréon and Colzy, 1999). Cloud top pressure retrieved by POLDER-3 appears to be a better proxy for low-level cloud height than MODIS cloud top pressure derived using a thermal signature assuming a given temperature profile (Buriez et al., 1997;Weisz et al., 2007;Tietze et al., 2011;Desmons et al., 2013).
To determine the cloud thermodynamic phase we use a combination of MODIS and POLDER-3 measurements. The al-10 gorithm takes advantage of multi-angle polarization data, shortwave, thermal infrared, and visible measurements to retrieve a thermodynamic phase index Φ between 0 for liquid clouds and 200 for ice clouds with varying degrees of confidence (Riedi et al., 2010). Figure 1 shows the distribution of the thermodynamic phase index for clouds between 200 and 1000 m and between 1000 and 2000 m altitude from 2008 to 2010 over a region with latitudes greater than 65 • . We observe different modes in the phase index corresponding to liquid clouds with Φ lower than 70, clouds with undetermined phase, mixed phase or multiple 15 cloud layers, for which Φ lies between 70 and 140, and ice clouds with Φ greater than 140.

Anthropogenic pollution tracer fields
For determining anthropogenic pollution tracer fields, we use the Lagrangian particle dispersion model FLEXPART (Stohl et al., 1998(Stohl et al., , 2005. The model is driven with 3 hourly operational analysis wind fields from the European Centre for Medium-Range Weather Forecasts (ECMWF) with 91 model levels and a horizontal resolution of 1 • ×1 • . We use the same simulations as 20 described by Stohl et al. (2013), which consider a black-carbon tracer undergoing removal processes, two fixed-lifetime black carbon tracers, and a carbon monoxide (CO) tracer. The CO tracer used for this study is considered as passive in the atmosphere but is removed from the simulation 31 days after emission, thus focussing the simulation on "fresh" pollution. For the CO emissions, ECLIPSE (Evaluating the CLimate and Air Quality ImPacts of Short-livED Pollutants) version 4.0 emission data (Klimont et al., 2015;Stohl et al., 2015) are used. For the anthropogenic emissions considered here, the ECLIPSE emissions are 25 based on the GAINS (Greenhouse gas -Air pollution Interactions and Synergies) model (Amann et al., 2011). The emissions are determined separately for every year of this study and, notably, include gas flaring emissions, which have been shown to be important for black carbon in the Arctic (Stohl et al., 2013). Emissions from the residential sector are temporally disaggregated using a heating degree day approach (Stohl et al., 2013).
Studies that have used FLEXPART CO concentration fields (χ CO ) have found satisfactory agreement between model output 30 and measurements (Stohl, 2006;Paris et al., 2009;Hirdman et al., 2010;Sodemann et al., 2011;Stohl et al., 2013Stohl et al., , 2015Eckhardt et al., 2015). In the Alaskan Arctic for the day of 18 April 2008, Warneke et al. (2009) described a slope of 0.9 for a linear fit between FLEXPART model output of χ CO and airborne measurements of CO with a least-squares correlation coefficient of 0.63.
The FLEXPART model outputs used here have a temporal resolution of 3 h and a spatial resolution of 1 • ×2 • (in latitude and longitude) divided into 9 different vertical levels. FLEXPART CO concentration (χ CO ) output is provided in units of mg m −3 but converted to units of ppbv (parts per billion by volume) to remove the atmospheric pressure dependence. Since the focus here is on the effect of anthropogenic pollution on clouds, only FLEXPART spatial bins where anthropogenic sources comprise more than 80 % of total CO concentrations are considered for comparison with cloud properties. 5 2.3 Meteorological data ERA-Interim (ERA-I) reanalysis data from ECMWF (Berrisford et al., 2009) extends from 1989 to the present with an improved version released in 2011 (Dee et al., 2011). The temporal resolution is 6 h at 60 pressure levels. Reanalysis data from ERA-I shows good agreement with satellite retrievals and aircraft data for cloud fraction and cloud radiative forcing in the Arctic (Zygmuntowska et al., 2012). Wesslén et al. (2014) analyzed ERA-I data with the Arctic Cloud-Ocean Study (ASCOS) 10 campaign measurements in 2008 and calculated biases of about 1.3 • C, 1 % and −1.5 hPa respectively for temperature, relative humidity and surface pressure, root mean square errors of about 1.9 • C, 3.7 % and 8.7 hPa respectively, and correlation coefficients of approximately 0.85 for temperature and surface pressure and 0.31 for the relative humidity.
The goal of this study is to use satellite, tracer transport model, and meteorological data sets to determine the effects of longrange aerosol transport on cloud microphysics due only to the pollution itself and not to the meteorological state. The focus is 15 on temperature, specific humidity, and LTS since these have been identified as a basic meteorological quantities that correlate with cloud microphysical properties (Matsui et al., 2006;Mauger and Norris, 2007). Defining the potential temperature (θ) as where T and P are the air temperature and pressure, P 0 equals 1000 hPa and R, and c p are respectively the gas constant for air, and the isobaric heat capacity, the LTS is defined as the potential temperature difference between 700 and 1000 hPa (Klein 20 and Hartmann, 1993).
We also consider clouds with values of LWP greater than 40 g m −2 separately from clouds with LWP values less than 40 g m −2 . This approach separates clouds according to their thermal radiative properties since a cloud with low LWP will tend to act as a graybody and potentially be radiatively susceptible to pollution at thermal wavelengths (Garrett and Zhao, 2006;25 Lubin and Vogelmann, 2006;Mauritsen et al., 2011). Thick clouds act as blackbodies, and their longwave radiative properties are determined by temperature only (Garrett and Zhao, 2006;Garrett et al., 2009).
The different data sets used in this article are summarized in Table 1.

Methodology
This study examines data between 2008 and 2010 over the ocean at latitudes greater than 65 • . Passive satellite sensors measure interactions of solar radiation with the atmosphere so as to retrieve cloud microphysical parameters of interest from visible wavelength measurements so analyses are restricted to the period between 1 March and 30 September.
3.1 Co-location of satellite retrieval and model pollution tracer fields 5 CO tracer concentrations from a FLEXPART grid cell are defined as the average between two temporal points, averaged over a spatial box. For example, model CO concentrations at 03:00 UTC and at the latitude-longitude co-ordinates of (70 • , 80 • ), represent an average over a box between the latitudes of 70 to 71 • and longitudes of 80 to 82 • and between 00:00 UTC and 03:00 UTC.
For an A-train satellite overpass time of 00:45 UTC, we match space-based retrievals to FLEXPART concentration output 10 at 03:00 UTC representing the average concentration between 00:00 UTC to 03:00 UTC and then linearly interpolate ECMWF meteorological fields to the LTS and specific humidity values for 01:30 UTC.
Here, as with many prior studies looking at aerosol-cloud interactions in the Arctic, we consider only low-level clouds (Garrett et al., 2004;Garrett and Zhao, 2006;Lubin and Vogelmann, 2006;Mauritsen et al., 2011), with POLDER cloud top altitudes between 200 and 1000 m, and between 1000 and 2000 m. The cloud top pressure is translating to cloud top altitude 15 by a pressure profile specific to the Arctic region. These two layers correspond to the FLEXPART vertical bin resolution.
We average POLDER and MODIS data that falls within the height bins so that they are co-located with the corresponding FLEXPART CO concentrations.
Regarding horizontal co-location, Fig. 2 shows how the datasets are combined. We project data from satellite, model, and reanalysis data sets onto an equal-area sinusoidal grid such that the grid-cell resolution is 0.5 • ×0.5 • at the equator correspond- 20 ing to an area of 54 km × 54 km. The sinusoidal projection conserves the grid-cell area independently of longitude and latitude.
One grid-cell can include up to 81 POLDER-MODIS pixels. Satellite and tracer transport model data are averaged over each grid-cell.
One consequence of the averaging is that each grid-cell can include up to 81 different pixel-level values of Φ. We place a limit on the SD of the averaged phase index within each grid-cell such that σ Φ < 10 in order to satisfy an a priori requirement 25 of there being a nearly homogeneous phase within each grid cell.

The net aerosol-cloud interactions parameter
Assuming a constant LWC and a mono-disperse size distribution of cloud droplets, the droplet effective radius (r e ) decreases as the CCN number concentration N c increases following the relation (Feingold et al., 2001): Here, we take a different approach which is to examine how cloud properties change in response to changes in a CO tracer under the presumption that the CO tracer serves as a proxy for the potential of long-range pollution transport, of which CCN may be a part. Of course, N c and CO tracer concentrations (χ CO ) do not represent the same quantity. However, cloud condensation nuclei and CO are both by-products of combustion. The two quantities are expected to be highly correlated close to pollution sources where relative changes in one can serve as a proxy for relative changes in the other (Avey et al., 2007;5 Tietze et al., 2011). The reason for using CO is twofold. First, CO is passive and therefor easier to represent in a dynamic model. Second the analysis here is less focussed on the local physics of aerosol-cloud interactions and more focused on the actual impact of anthropogenic activities on clouds far from combustion sources. These are similar but not identical questions.
The aerosol-cloud interactions (or the ACI) parameter addresses the precise physics of the extent to which aerosols can modify clouds. However, interactions are a two-way street: where aerosols have been scavenged en route to distant clouds, then the 10 potential is for a pollution plume to be present but its impact on cloud properties weak. To account for scavenging, we employ the term ACI net or the net aerosol-cloud interaction parameter. ACI net is the same as the ACI while additionally accounting for any reduction of the ACI due to dry or wet scavenging of aerosols during transport. ACI net can be interpreted as a measure of the sensitivity of a cloud at any given location to pollution plumes from distant sources. It allows for the passive components of a plume (e.g. CO) to remain while aerosols have been removed: For example, in the absence of scavenging then N c , from Eq. (3) is linearly related with χ CO and Eq. (5) will be bound by a theoretical maximum value of ACI net re = 1/3 (Twomey, 1977;Feingold et al., 2001). Further from source regions, the 20 correlation of CO concentration and aerosols is invariant to dilution but it may be affected by wet and dry scavenging (Garrett et al., 2010(Garrett et al., , 2011. If scavenging rates are low, CO and CCN tend to covary, but when precipitation is high along transport pathways then aerosols are removed and values of ACI net will tend to be lower. Garrett et al. (2010) found that at Barrow, Alaska, when the temperature exceeds 4 • C at the surface, wet scavenging efficiently removes CCN from the atmosphere. In this event, cloud microphysical properties will not affected by pollution plumes. 25 Since r e and the optical depth (τ ) are linked through τ = 3 2 LW P ρwre for homogeneous clouds it follows that or, ACI net τ = ACI net re + ACI net LWP (7) Figure 3 shows an of how ACI net is calculated for temperatures between −12 and 6 • C and altitudes between 1000 and 30 2000 m, for all LWP values. We first calculate ACI net re as the linear fit of the natural logarithm of the effective radius to the natural logarithm of CO concentrations. The fit used in this study is based on the robust linear method (RLM) (Huber, 1973(Huber, , 1981Venables and Ripley, 2013). RLM uses an iterative least squares algorithm: every measurement has initially the same weight; The weights of each point are updated giving a lower weight to points that appear as outliers with respect to the entire dataset. The process iterates several times and stops when the convergence tolerance of the estimated fitting coefficients lies below 10 −8 . The slope is therefore less sensitive to outlier points. In Fig. 3, points indicated by the red line are weighted 5 similarly to those indicated by black and blue lines. The slope retrieved by the linear fit, in Fig. 3, is -0.13±0.016. Referring to Eq. (5), ACI net re equals +0.13 ± 0.016.
3.3 Stratifying the data for specific humidity and lower tropospheric stability  Table 2 describes the method used here to stratify the dataset according to meteorological conditions. We identify a range in LTS and specific humidity that occupies 15 % of the total space of observed values but that is centered at the mode of the respective distributions. The total LTS range is 2.1 to 37 K, so the interval size is 5.3 K. The specific humidity is distributed over several orders of magnitude. To better represent the distribution, we use a logarithmic scale for this parameter. The logarithm 15 base 10 of specific humidity has an interval of 0.28. The most common values of meteorological state, defined here as the maximum number of measurements, is delimited by the red rectangle in Fig. 4. The rectangle corresponds to a range between 2.0 g kg −1 to 4.0 g kg −1 for specific humidity, and a range of 16.5 K to 21.8 K for LTS. It is these ranges that are focussed upon for calculation of the ACI net parameter. We assume these intervals are sufficiently narrow that the variability within the interval has limited impact on cloud microphysics. stratified. The number of grid-cells used to calculate each ACI net parameter per bin ranges from 100 to 3300. The ACI net parameter is almost always positive but sometimes close to zero. ACI net re ranges from 0 for graybody clouds between 1000 to 2000 m altitude with a cloud top temperature between −6 and −4 • C, to 0.34 for blackbody clouds between 1000 and 2000 m altitude with a cloud top temperature between 4 and 6 • C. ACI net τ ranges from −0.10 for all clouds between 200 and 1000 m altitude with a cloud top temperature of −11 • C, to 0.35 at 3 • C for blackbody clouds between 1000 and 2000 m altitude. In general, ACI net τ and ACI net re are of the same order of magnitude and the maximum values of ACI net are found for clouds with temperatures above the freezing temperature.
We define the uncertainty in ACI net as the 95 % confidence limit in the calculation of the slope of the linear fit. The uncertainty in the calculated values of ACI net re is generally less than 0.1, except for clouds with temperatures between 4 and 6 • C and between −12 and −10 • C, where the uncertainty bar is approximately 0.2. For the optical depth, the uncertainty is 5 typically approximately 0.1, although larger values are observed for high and low cloud top temperatures.
For blackbody clouds between 1000 and 2000 m altitude, the average values of ACI net τ and ACI net re equal 0.20 and 0.14 respectively. For cloud tops between 200 and 1000 m altitude, ACI net τ and ACI net re equal 0.14. For graybody clouds between 1000 and 2000 m, ACI net τ and ACI net re equal 0.12 and 0.08 respectively. For cloud tops between 200 and 1000 m altitude, ACI net τ and ACI net re equals 0.14 and 0.12 respectively. Thus, the value of ACI net appears to be fairly robust to altitude and 10 cloud thickness and to whether r e or τ is considered.

Dependence of ACI net on pollution concentration, specific humidity, and lower tropospheric stability
In what follows, we examine the influence of LTS, specific humidity and pollution concentration on ACI net . Table 4 shows values of ACI net re and ACI net τ for graybody and blackbody clouds, and for χ CO < 5.5 ppbv and χ CO > 10.0 ppbv, corresponding 15 respectively to the lower and upper quartile of CO tracer concentration, and for a range in LTS and specific humidity. For graybody and blackbody clouds, ACI net τ and ACI net re are highest for low values of χ CO . The difference in ACI net values between low and high polluted environments is slightly greater for ACI net re than for ACI net τ . Table 4 suggests that cloud effective radius and cloud optical depth are most sensitive to pollution when pollution concentrations are low. Previous studies have hypothesized that the effect of CCN on cloud microphysical properties saturates when cloud-droplet concentrations are high 20 (Bréon et al., 2002;Andersen and Cermak, 2015). This effect does not explain the differences presented in Table 4 because Eq. (4) and (5) already take into account the potential for linear saturation by considering the logarithms of χ CO and cloud parameters.
We now present the sensitivity of the ACI net parameter to 5 different ranges of meteorological parameters delimited by the percentiles values presented in Table 5. We do not stratify the data according to LWP. Figures 6 and 7 show the influence of 25 pollution loading on the cloud-droplet effective radius and cloud optical depth for each of the different specific humidity and LTS regimes. Figure 6 presents the ACI net parameter with respect to the cloud optical depth and cloud-droplet effective radius as a function of the specific humidity, stratifying the data by LTS according to the method described in Sect. 3.3. Figure 7 is the same as Fig. 6 except that it shows ACI net as a function of LTS for a range of specific humidities. Figure 6 shows that ACI net re and ACI net τ tend to increase with the specific humidity independent of LTS. The ACI net param- 30 eter is close to zero, or negative, for low values of specific humidity. It increases rapidly with specific humidity, saturating at a maximum value of about 2.5 g kg −1 . We note that cloud top temperature and specific humidity are weakly correlated. The correlation coefficient (r 2 ) of the linear regression of the two parameters is 0.20. The same applies for the specific humidity and the LWP. The correlation coefficient of the two parameters is 0.05.
ACI net re increases with LTS from 0.02 for values of LTS ranging between 2.1 and 14 K to 0.09 for values of LTS between 23 and 38 K. The ACI net τ dependence on LTS is larger: ACI net τ equals 0.10 for LTS values between 2.1 and 14 K and it equals 0.32 for LTS values between 21 and 38 K.

Discussion
The results presented here show values of the ACI net parameter with respect to the cloud-droplet effective radius and optical 5 depth, for clouds over oceans north of 65 • lying between 200 and 2000 m, and for the years between 2008 to 2010. We find ACI net values that range from 0.00 to 0.34 for the cloud-droplet effective radius, and from −0.10 to 0.35 for the optical depth.
Prior studies examining the Arctic region have retrieved ACI values ranging from −0.10 to 0.40 (Garrett et al., 2004;Lihavainen et al., 2009;Zhao et al., 2012, Sporre et al., 2012. Tietze et al. (2011) calculated ACI net values ranging from 0.00 to 0.17 using a similar satellite-FLEXPART co-location method. What differs in this study is that we are examine solely 10 anthropogenic pollution, and that we extend the dataset from one to three years, stratifying the dataset according to specific humidity and LTS. The larger ACI net values we find in this study suggest a higher sensitivity of cloud microphysical properties to pollution plumes from distant sources than was found by Tietze et al. (2011).
However, Tietze et al. (2011) also found values of ACI net τ that were greater than ACI net re , by put to a factor of four, and they attributed this difference to unknown dynamic or precipitation feedbacks that make ACI net LWP greater than zero (Eq. 7).

15
In contrast, our results show that the ACI net re and ACI net τ parameters are more similar, suggesting no such feedback. Table 6 compares the differences between ACI net τ and ACI net re that are presented in Table 3, along with their corresponding values when no control is made for specific humidity and LTS. The difference between ACI net re and ACI net τ is largest when meteorological parameters are not controlled for. For all clouds considered, the maximum difference increases from 0.04 when the data are considered within narrow meteorological bands to 0.12 when the data are not. This is result is important since it suggests that 20 the hypothesized feedbacks discussed by Tietze et al. (2011) may have in fact been due to the natural sensitivity of clouds to local meteorology. Not controlling sufficiently for meteorology may lead to artifacts that exaggerate the magnitude of the aerosol indirect effect.
In contrast to most prior efforts, satellite-retrieved cloud properties are not compared to CCN or aerosol concentrations but rather to pollution concentrations -specifically CO simulated from a tracer transport model. CO serves as a proxy for CCN 25 close to pollution sources although far from sources CCN and CO can become decoupled due to scavenging en route to distant clouds. The comparison we present through the ACI net parameter implicitly accounts for this possibility. For temperatures below −6 • C, low values of ACI net are observed. Tietze et al. (2011) hypothesized that at such temperatures, cloud supersaturations may be too small to activate aerosols as CCN or that clouds with lower temperatures have followed longer transport pathways nearer the surface (Stohl, 2006) and therefore had greater exposure to dry deposition. Table 4 suggests that ACI net values are lowest when pollution concentrations are high. Figure 8 presents the normalized distribution of potential temperature for polluted and pristine clouds, defined as the upper and lower quartile, for graybody clouds.
We present results for graybody clouds because the ACI net differences between polluted and clean cases are largest; Results for blackbody and all clouds are not shown here, but have similar results regarding the potential temperature distribution.
Highly polluted air parcels are associated with potential temperatures around 280 K whereas pristine air parcels have a lower potential temperature -around 272 K. We hypothesize that higher values of potential temperature suggest pollution sources from further south, so wet scavenging is more likely to occur during transport and this decreases the correlation between a CO 5 tracer and CCN, therefore lowering the ACI net parameter. Also, polluted air parcel and aerosols do not necessarily have the same physical and chemical properties at lower and higher latitudes, and this difference may impact the influence of aerosols on cloud microphysics and aerosols (Bilde and Svenningsson, 2004;Dusek et al., 2006;Ervens et al., 2007;Andreae and Rosenfeld, 2008).
In general, we observe that when moisture increases, the cloud sensitivity to pollution increases. From model simulations 10 of stratocumulus, Ackerman et al. (2004) found that when the relative humidity (RH) above cloud top is high, cloud LWP increases with N c consistent with theoretical arguments (Albrecht, 1989;Pincus and Baker, 1994), but that when the RH is low, the LWP decreases when N c increases, as supported by some observations (Coakley and Walsh, 2002). The difference was attributed to the consequence of dry air into a cloud layer. Humidity inversions are common above low-level cloud tops in the Arctic (Nygärd et al., 2014), so similar phenomena may be playing a role.

15
Studies of the indirect effect at mid-latitudes suggest that values of ACI are highest under unstable conditions (Chen et al., 2014;Andersen and Cermak, 2015). Our results from the Arctic show the reverse, that conditions of high LTS are associated with higher values of ACI net . Klein and Hartmann (1993) showed that, in general, higher values of LTS lead to greater stratiform cloudiness, except in the Arctic where radiative cooling prevails over convection as the driving mechanism for cloud formation. This result is similar to results found by Kim et al. (2008) who found that aerosol-cloud interactions are 20 strongest in clouds with adiabatic liquid water content profiles. Such clouds might be expected more frequently when LTS is high and there is reduced vertical mixing.
Finally, we find ACI net τ is more sensitive to changes in LTS than ACI net re . A consequence is that for values of LTS greater than 23 K, ACI net τ and ACI net re differ by about 0.20. In a stable atmosphere with high LTS it appears that ACI net LWP increases more strongly in response to aerosols than in unstable environments (Klein and Hartmann, 1993;Qiu et al., 2015). 25

Conclusions
Satellite, numerical model, and meteorological reanalysis data sets from 2008 to 2010 are used here to calculate the sensitivity of cloud-droplet effective radius and optical depth in the Arctic to anthropogenic pollution transported from mid-latitudes. We focused on latitudes north of 65 • for a period between March 2008 and October 2010. Using ECMWF reanalysis data, we stratified the dataset according to temperature, LTS, specific humidity, altitude, and LWP. We find that the sensitivity of cloud 30 properties to pollution, as quantified by values of ACI net , lies close to a theoretical maximum value of 1/3, assuming that a simulated CO tracer correlates well with CCN. Further, ACI net re and ACI net τ seem to increase with specific humidity and LTS, highlighting that meteorological parameters have an important impact on aerosol-cloud interactions.
Globally, Klimont et al. (2013) have estimated that there was a drop of about 9281 Gg in anthropogenic sulphur dioxide emissions between 2005 and 2010 due to a reduction in European and American emissions and a flue gas desulfurization program on power plants in China. This reduction in emissions has led to a decrease of sulfate concentrations at Arctic surface station (Hirdman et al., 2010). In the Arctic, the effect of a decrease in mid-latitude pollution emissions may some day be offset by greater levels of Arctic industrialization (Lindholt and Glomsrød, 2012) and shipping (Pizzolato et al., 2014;Miller 5 and Ruiz, 2014) that introduce new local aerosol sources. Further, an increase in the extent of open-ocean due to sea-ice retreat may be expected to lead to an increase in the atmospheric humidity (Boisvert and Stroeve, 2015) and from the results presented here, a higher sensitivity of clouds to aerosols. However, this study also suggests that any associated decrease in LTS could partially counter-act this effect.
Climate warming is thought to stimulate boreal forest fires (Westerling et al., 2006). The impact of pollution from biomass 10 burning has not been included in the present research. Given biomass burning aerosol can act as efficient ice nuclei (Markus et al., 2009), the analyses presented here might be extended to explore aerosol-induced changes in cloud thermodynamic phase.
Acknowledgements. This paper is dedicated posthumously to Kyle Tietze, who developed many of the techniques used in this study while a graduate student at the University of Utah. The authors thank ICARE, NASA and CNES for the data used in this research. We acknowledge financial support from The University of Lille1. This material is based upon work supported by the National Science Foundation under the     Table 6. Difference between ACI net τ and ACI net re (i.e. ACI net LWP ) for graybody, blackbody, and all clouds when lower tropospheric stability and specific humidity are stratified and when they are not stratified. The averaged ACI net values are shown in Table 3.
All LWP Graybody Blackbody     Blue numbers indicate the number of grid-cells, in hundreds, that are used to calculate each ACI net value. In each figure the ACI net value averaged over the temperature and weighted according the inverse of the uncertainty is indicated.   greater than 10 ppbv and less than 5 ppbv. The values of ACI net re and ACI net τ associated with each histogram are presented also in Table 3.