Hemispheric contrasts in ice formation in stratiform mixed-phase clouds: Disentangling the role of aerosol and dynamics with ground-based remote sensing

Multi-year ground-based remote-sensing datasets were acquired with the Leipzig Aerosol and Cloud Remote Observations System (LACROS) at three sites. A highly polluted central European site (Leipzig, Germany), a polluted and strongly dustinfluenced eastern Mediterranean site (Limassol, Cyprus), and a clean marine site in the southern mid-latitudes (Punta Arenas, Chile) are used to contrast ice formation in shallow stratiform liquid clouds. These unique, long-term datasets in key regions 5 of aerosol-cloud interaction provide a deeper insight into cloud microphysics. The influence of temperature, aerosol load, boundary-layer coupling, and gravity wave motion on ice formation is investigated. With respect to previous studies of regional contrasts in the properties of mixed-phase clouds our study contributes the following new aspects: (1) Sampling aerosol optical parameters as a function of temperature, the average backscatter coefficient at supercooled conditions is within a factor of 3 at all three sites. (2) Ice formation was found to be more frequent for cloud layers with cloud top temperatures above −15◦C 10 than indicated by prior lidar-only studies at all sites. A virtual lidar-detection threshold of IWC needs to be considered in order to bring radar-lidar-based studies in agreement with lidar-only studies. (3) At similar temperatures, cloud layers which are coupled to the aerosol-laden boundary layer show more intense ice formation than de-coupled clouds. (4) Liquid layers formed by gravity waves were found to bias the phase occurrence statistics below −15◦C. By applying a novel gravity wave detection approach using vertical velocity observations within the liquid-dominated cloud top, wave clouds can be classified 15 and excluded from the statistics. After considering boundary layer and gravity-wave influences, Punta Arenas shows lower fractions of ice containing clouds by 0.1 to 0.4 absolute difference at temperatures between −24 and −8◦C. These differences are potentially caused by the contrast in the INP reservoir between the different sites. Copyright statement. TEXT

from the year-long lidar/radar dataset at Macquarie Island (54.6°S 158.9°E, Australia) the cloud observations focused on austral summer.
Using a shiporne dataset, Mace and Protat (2018) also found frequent liquid-dominated clouds with low radar reflectivities 90 and one-third of the liquid layers only observed with lidar. Comparing the observations with a Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) dataset from Hu et al. (2010), they found an overestimation of supercooled liquid in the satellite dataset, especially strong at temperatures above −15 • C. In a follow-up study, Mace et al. (2020) refined the CALIPSO classification scheme, leading to more frequent detections of the mixed phase, especially during wintertime and in the lower latitudes of the Southern Ocean. However, no CTT-resolved phase occurrence statistics is presented. Liquid layers 95 in deeper clouds, observed during another shipborne campaign (McFarquhar et al., 2021;Alexander et al., 2021), could only be reproduced in regional model simulations, when INP parametrization was tuned to lower concentrations (Vignon et al., 2021). Zaremba et al. (2020) investigated airborne active remote-sensing observations of Southern Ocean clouds south of Tasmania. They also found widespread liquid cloud tops at temperatures down to −30 • C. By investigating the ground-based remote-sensing dataset assembled at McMurdo (77.8°S 166.7°E, Antarctica), Silber et al. (2018) found frequent long-lived 100 liquid-topped clouds, also below −30 • C.
Yet, a statistical analysis of the relationship between both, aerosol conditions, cloud vertical dynamics, and the phase partitioning in stratiform cloud layers of the southern-hemisphere mid-latitudes based on long-term observations was not established. One reason is that, despite increased activity in the recent past, ground-based remote-sensing observations of clouds and aerosol are still sparsely distributed in the Southern Ocean and at the coast of Antarctica.
ice production is assessed in Sec. 3.2.4. The study concludes with a discussion of the contrast identified in aerosol load and stratiform cloud properties (Sec. 4) followed by a summary and outlook (Sec. 5).

Data and Methods
This section introduces the datasets and methods used in the remainder of this study. Starting with the campaigns and in-125 strumentation (Sec. 2.1), followed by a short description of the retrievals used (Sec. 2.2 and 2.3) and finally the methods for automated selection of shallow stratiform clouds (Sec. 2.4) and characterization of vertical motion (Sec. 2.5).

Datasets from Leipzig, Limassol, and Punta Arenas
Basis of the observational datasets is the Leipzig Aerosol and Cloud Remote Observations System (LACROS), the mobile ground-based remote-sensing supersite of the Leibniz Institute for Tropospheric Research (TROPOS), Leipzig, Germany. The 130 instrumentation used for the synergistic approaches applied in this study comprises a MIRA-35 35 GHz scanning cloud radar, a Polly XT multi-wavelength Raman polarization lidar, a Streamline XR 1.5 µm scanning Doppler lidar, a HATPRO 14-channel microwave radiometer, a 1064 nm ceilometer, an optical disdrometer, and radiation sensors. Main properties of the sensors are summarized in Table 1. of instrumentation of LACROS has been kept unchanged since the year 2014. Hence, this comparative study uses data from the 6 year period since then. During that period, LACROS was, besides several 4-8 weeks short-term deployments at other sites, stationed at Leipzig, Limassol, and Punta Arenas. First results of the CyCARE campaign are described by Ansmann et al. (2019). For the second long term campaign, LACROS  As a proxy for the aerosol conditions at the three measurement sites, we used vertically resolved observations from the Polly XT lidar system (Althausen et al., 2009;Engelmann et al., 2016). Quantities of interest are the aerosol backscatter coefficient, the extinction coefficient, the particle linear depolarization ratio, and the cloud-relevant concentration of INP. Their retrieval is explained below.
Basis for the statistical analysis are profiles of particle backscatter coefficient β p at 532 nm wavelength. The profiles are 160 computed with the Klett method (Fernald, 1984) by the PollyNET retrieval whenever atmospheric conditions are suitable (Baars et al., 2016(Baars et al., , 2017Yin and Baars, 2021). The PollyNET retrieval chain also ensures a homogenized analysis of the data from the three different Polly XT instruments, which were utilized in the frame of this study (see Fig. 1 and Engelmann et al., 2016;Baars et al., 2016).
Profiles of the particle linear depolarization ratio (hereafter referred to as particle depolarization ratio) are only calculated 165 when the ratio of molecular backscatter coefficient to β p is below a value of 18. Additionally, any particle depolarization ratios larger than 0.7 are masked, as they are indications for noise artifacts in the cross-polarized signal component. All profiles are then filtered with the co-located Cloudnet target classification (see Section 2.3) to exclude clouds, especially optically thin ice clouds that are only clearly classified by the cloud radar. Finally, a manual screening excluded fragments of thin liquid clouds, which would otherwise artificially increase β p . For the averages, the optical data of each retrieved profile is binned to vertical 170 intervals of 200 m or 3 K.
The derived average optical properties can be used to estimate aerosol microphysical properties, such as concentrations of ice nucleating particles (INP). This is an important step in order to evaluate the datasets of the three sites with respect to contrasts in the potential contribution of aerosol effects on heterogeneous ice-formation efficiency. Conversion from optical properties as observed by lidar to microphysical aerosol properties is based on the parametrizations described by Mamouri 175 and Ansmann (2016). By means of this approach, the lidar-measured aerosol extinction coefficient is converted to the number and surface concentration N 500 and S 500 of aerosol particles larger than 500 nm in diameter. These quantities are applied in available in-situ-based parametrizations for the retrieval of INP concentrations. Prerequisite for the retrieval is a correct aerosol typing, as different types of particles differ by orders of magnitude in their ice forming efficiency. In order to do so, the average backscatter profile is separated into the categories marine, continental, and mineral dust, based on air mass source 180 (see Appendix B and Radenz et al., 2021) and particle depolarization ratio (one-step POLIPHON; Mamouri and Ansmann, 2017). The average extinction is calculated from the profiles of β p by assuming a typical lidar ratio of 20 sr for marine, 50 sr for continental, and 45 sr for dust aerosol (Müller et al., 2007;Baars et al., 2017;Bohlmann et al., 2018). In a next step, the extinction coefficient is converted to N 500 and S 500 , using sun-photometer-based conversion factors (Mamouri and Ansmann, 2016

Cloudnet processing
Synergies between lidar, cloud radar, microwave radiometer, and meteorological data are utilized for determination of cloud macro-and microphysical properties and as basis for the stratiform cloud identification scheme. State-of-the-art routines for achieving this synergy are comprised in the Cloudnet retrieval (Illingworth et al., 2007). Cloudnet re-grids the observations to  and Löhnert, 2007). The statistical retrieval is based on long-term radiosonde observations (Leipzig and Limassol) and highresolution reanalysis data (Punta Arenas). Attenuated backscatter of the ceilometer is regularly cross-calibrated with Polly XT using the calibrated attenuated backscatter of PollyNET. Usually, the ceilometer data is used in the synergistic retrieval, as 200 the dataset is more robust and less prone to interruptions. Rare gaps in the observations are filled with the Polly XT attenuated backscatter at 1064 nm.

Automated cloud selection and characterization
While the Cloudnet algorithm provides a pixel-by-pixel classification of cloud phase and microphysical properties, information on temporal coherence and cloud evolution is not readily available. Hence, an automatic reproducible filtering algorithm for 205 the selection of targeted stratiform, supercooled cloud systems is necessary. Based on the data cube LARDA 3 (Bühl et al., 2021), the approach of Bühl et al. (2016) is implemented into an automated selection algorithm.
An example of the Cloudnet processing of measurement data and the application of the cloud selection scheme is shown in Fig. 2. Starting with a profile of the Cloudnet target classification mask (Hogan and O'Connor, 2004), consecutive pixels classified as cloud pixels (liquid droplets, ice, ice and supercooled droplets) are grouped together and defined as features. In 210 case similar types of hydrometeors were observed in matching heights, single features in neighboring timesteps are connected to coherent cloud cases. For the analysis, the cloud cases are filtered for shallow stratiform clouds, which are liquid topped and either have an ice virga or not (rectangles in Fig. 2d). An overview of the microphysical parameters sampled for each cloud case is provided in Table 3. It is assumed that if ice is formed in an liquid layer it also sediments out of the cloud. This is required, as the signal in the top layer is dominated by return from liquid droplets and Cloudnet provides no reliable 215 mixed-phase classification there.
To pinpoint potential effects of aerosol load, thermodynamic, and dynamic drivers of ice formation have to be constrained.
This is especially important, as the sites are situated in different climate zones. We presume for our study that thin stratiform clouds serve as a natural laboratory with only a limited number of microphysical processes being possible.  The cloud cases are filtered for a length of more than 20 minutes and smooth cloud top heights (standard deviation < 150 m) 220 to exclude convective clouds. Laboratory studies of Fukuta and Takahashi (1999)  within the virga. The ice water content (IWC) is derived from the radar reflectivity and the temperature using the Hogan et al.

Gravity wave detection
In order to pinpoint effects of aerosol and dynamicas on the phase partitioning in the stratiform cloud dataset, an approach is 235 required to assign vertical motion regimes to each cloud case. Here we focus on the temporal structure of vertical velocity, to constrain the dynamics forcing on a cloud. Usually, shallow clouds are characterized by a fully developed turbulence in the liquid-dominated cloud top (Bühl et al., 2019), where the vertical motion is driven by cloud top cooling (e.g., Shao et al., 1997;Fang et al., 2014;Simmel et al., 2015). In the turbulent layer at cloud top, up-and downdrafts alternate at horizontal scales in the order of 100 m or less.

240
However, orographic gravity waves can trigger vertical motion and associated cloud formation, as well. Microphysical processes in these wave clouds are governed by large-scale dynamics, where vigorous up-and downdrafts may appear stationary.
Due to this dynamics, the mixed-phase and the ice phase are horizontally separated, with the liquid drops predominantly in the ascending branch and the ice particles in the descending branch (Heymsfield and Miloshevich, 1993;Baker and Lawson, 2006). The properties of the horizontal wind field determines the regions of the up-and downdraft in such orographic clouds.

245
Observing these clouds by stationary ground-based remote sensing might thus not sample the full horizontal extent of the cloud, which causes a misclassification of the cloud case in terms of liquid-only and ice-producing. An illustrative example of such a wave cloud is shown in appendix A Also, the flow is highly laminar, opposed to the confined, fully developed turbulence found in layered mixed-phase clouds.
Cloud microphysics in wave conditions cannot directly be compared to layered clouds. For clouds in the heterogeneous freezing 250 regime, Cotton and Field (2002) found that only rapid evaporation freezing once the downdraft commences could explain their observations, where evaporation freezing can be better characterized as inside-out contact freezing of shrinking particles (Durant, 2005). A more recent study by Field et al. (2012) found that condensation and immersion freezing are needed together with deposition and evaporation freezing to explain their aircraft observations of ice formation in wave clouds.
On the other hand, the frequent occurrence of atmospheric gravity waves in a specific region might increase the frequency 255 of thermodynamic conditions that favor the presence of a sustained liquid phase. As Korolev (2007)  The autocorrelation function Ψ (τ ) for a time series of vertical velocities v t is defined as with the temporal shift τ and the vertical velocity v at time t. To compare different cloud cases, the autocorrelation function 270 Ψ (τ ) is normalized with Ψ (0). The temporal shift τ from the observations is converted into a horizontal shift or autocorrelation length l with the Cloudnet model-based horizontal wind velocity v hor : Similarly the vertical-velocity spectral power density is calculated by a Fast Fourier Transform of the vertical-velocity time series. High autocorrelation coefficients for large shifts and low power density are indications for wave-driven, low-275 turbulent flow. The inset in Fig. 2c shows the autocorrelation function for each of the identified cloud cases. All of them are weakly affected by gravity waves, but small-scale turbulence dominates. In contrast, the wave cloud in Fig. A1d shows high autocorrelation coefficients for longer shifts. As a characteristic value of the autocorrelation function, the shift at which the coefficient drops below 0.8 was chosen after visually inspecting the whole dataset. When the ice-formation frequency statistics is investigated for an influence of gravity waves (Sec. 3.2.3), the shift threshold is reduced step by step. As the autocorrelation 280 function is generally decreasing, a lower characteristic value than 0.8 would result in larger shift thresholds. For shifts larger than about 500 m, random fluctuations appear for clouds with a rapid drop in autocorrelation coefficients. Thus, 0.8 is a robust choice for the characteristic value.

Results
In this section, the thermodynamic phase partitioning and quantitative ice mass production in ice forming shallow cloud layers 285 are presented and compared between all measurement sites under study. First, the average profiles of aerosol optical and microphysical properties are presented. Then, instrumental detection thresholds, boundary layer effects and gravity wave activity are all analyzed as potential influencing factors on the retrieved ice-formation characteristics.

Aerosol conditions at Leipzig, Limassol, and Punta Arenas
To provide a general insight into the aerosol conditions at the three sites, the average optical and microphysical aerosol prop-290 erties as derived from the lidar observations (Sec. 2.2) are shown in Fig. 3 and Fig. 4. The impact of aerosol on clouds is controlled more strongly by temperature than geometrical height, hence the averages are also calculated with temperature as a vertical coordinate.

Optical properties
The average aerosol backscatter coefficient β p at 532 nm and the particle depolarization ratio derived from the Polly XT obser- When only the periods with co-located Polly XT observations (used in this study) are considered, the AOT is 0.216.
Mean β p at 532nm drops below 0.2 Mm −1 sr −1 only above 4 km height, which corresponds to and extinction coefficient of 1.0 Mm −1 assuming continental aerosol conditions and a corresponding lidar ratio of 50 sr.
Limassol is characterized by a distinct dry season with no precipitation and very few clouds during the summer. Generally,

305
Limassol is frequently affected by aerosol transport from Africa, the Middle East and Europe with aerosol characteristics including dust (mineral and soil), marine (organics and sea salt) and anthropogenic pollution as well as mixtures of these (Nisantzi et al., 2015). Mean AOT at 500 nm is 0.176 during the whole observational period and 0.165 during the 'cloudy season' from October to May. In the following, the non-cloud season from June to September is excluded from the statistics.
The profile of mean backscatter is similar to the one at Leipzig within a factor of 1.5, whereas the median generally is higher 310 at Leipzig.
The aerosol load at Punta Arenas can be separated into two distinct layers with an aerosol-rich boundary layer and pristine conditions aloft. The free troposphere is dominated by marine aerosol from the Southern Ocean and was reported to show no changes compared to the pre-industrial conditions (Hamilton et al., 2014). Nevertheless, events of aerosol long-range transport also occur occasionally (Foth et al., 2019;Floutsi et al., 2021). The boundary layer is laden with a mixture of marine and 315 continental aerosol, as Punta Arenas is located 230 km inland from the Pacific coast. Mean AOT at 500 nm is 0.055 during the whole campaign, but dropping to 0.047, when excluding the period of long-range wildfire smoke transport in early 2020 (Ohneiser et al., 2020). Average boundary layer height is around 1.5 km (Foth et al., 2019) with negligible β p above 2.0 km barren ground from the Sahara is the strongest terrestrial source with contributions around 15%. POLIPHON (see Sec. 2.2) 355 shows a peak at −12 • C with mean dust fractions of 0.3 (90% percentile 0.87). Hence, the backscatter is divided into dust and non-dust according to the dust fraction. The non-dust portion is then split up into continental and marine, with the 40% contribution of the continent above −12 • C and 20% below. The aerosol mixture at Leipzig is dominated by continental aerosol, with an average dust fraction of 0.1. First, the temperature-resolved fraction of occurrence of ice-forming clouds is depicted in Fig. 6. For the phase occurrence frequency, a cloud is classified as ice-producing, if an ice virga was observed at least during 5% of the duration of a cloud case (Sec. 2.4). Comparing the frequency of ice-containing clouds at different locations provides insights into differences of 395 primary ice formation (Choi et al., 2010;Kanitz et al., 2011;Seifert et al., 2015;Tan et al., 2014;Zhang et al., 2018). Generally, clouds contain ice more frequently for decreasing temperature (Fig. 6, solid lines). While at Leipzig and Limassol nearly all clouds with CTTs below −16 • C contained ice, at Punta Arenas a fraction of 0.4 ± 0.1 of shallow stratiform clouds at these temperatures were classified as liquid only. This behavior is discussed further in Sec. 3.2.3.
Compared to prior lidar-based studies (e.g., Kanitz et al., 2011;Choi et al., 2010;Seifert et al., 2010), the fraction of 400 ice-containing clouds in the synergystic dataset is higher at temperatures above −10 • C. At these temperatures, the amounts of ice produced and hence, the radar reflectivity and optical extinction are usually very low and stays undetected for lidar (Bühl et al., 2013a) and space-borne radars (Bühl et al., 2016). To quantify a lidar detection threshold in terms of optical extinction, the reflectivity-to-IWC and the reflectivity-to-extinction relationships by Hogan et al. (2006) are used. The only additional information needed for both parametrizations is the ambient temperature. Using these relationships, the response of 405 the occurrence statistics to arbitrary extinction detection thresholds can be tested. As shown in Fig. 6 (dashed lines), for the lidar data used in the Cloudnet classification, an extinction threshold of 12 Mm −1 had to be applied in order to best match the lidar-only statistics from Punta Arenas and Leipzig presented by Kanitz et al. (2011).

Effect of boundary layer aerosol load on phase occurence
As discussed in Sec. 3.1, the aerosol load at Punta Arenas is confined to the lowermost 2 km and the aerosol load at temper-410 atures above −10 • C is similar to Limassol. To check for possible impact of this boundary layer aerosol on the ice-formation efficiency, the basic temperature-resolved phase occurrence frequency (Fig. 6) is split into two subsets, one containing cloud cases with bases below 2 km height and one with cloud cases having bases above that threshold, in the following denoted as coupled and uncoupled clouds, respectively. The resulting ice-formation frequency is shown in Fig. 7. At any height, temperatures vary by more than 11 • C (10% to 90% percentile), providing ample coverage for the height threshold. Generally, coupled Total duration of the clouds in each bin is given by the numbers on top in hours.
Considering only uncoupled clouds (cloud base above 2 km height, Fig. 7b), stronger contrasts become evident. The wave clouds discussed in the following section are still included and impact the frequency at temperatures below −18 • C at Punta Arenas. The frequencies of ice-forming clouds are similar at Leipzig and Limassol, but 0.15 to 0.5 at −10 • C lower at Punta Arenas. Comparing the coupled and uncoupled state at Leipzig and Limassol, ice formation is more frequent in the coupled case by 0.15 between −5 and −12 • C.

Gravity wave influence on phase occurence at low temperatures
The lower frequency of ice-containing cloud layers at temperatures below −18 • C over Punta Arenas is a prominent feature of both, the lidar-radar and the lidar-only-equivalent datasets shown in Fig. 6. The associated high abundance of supercooled liquid water is frequently reported as a general phenomenon of stratiform clouds in the southern hemisphere mid and higher latitudes. Within this subsection, the reasons for the found behavior over Punta Arenas will be elaborated in more detail.

430
Frequently, the stratiform liquid-only cloud layers observed over Punta Arenas at temperatures below −18 • C are embedded in orographic gravity waves. Following the wave detection methodology introduced in Sec. 2.5, Fig. 8 shows the autocorrelation and power spectra of the Doppler lidar vertical velocity for clouds classified as liquid only over Leipzig, Limassol, and Punta Arenas, respectively. These supercooled liquid-only clouds at Punta Arenas show high autocorrelation coefficients at long shifts ( Fig. 8c1), whereas liquid only clouds with similar characteristics are absent at Limassol (Fig. 8b1) and Leipzig (Fig. 8a1). In 435 terms of spectral power density (Fig. 8c2), the strongly supercooled clouds at Punta Arenas show only low turbulence. In a next step, the autocorrelation is used to filter the dataset for clouds affected by gravity waves. As described in Sec. 2.5, the length at which the autocorrelation coefficient drops below 0.8 is used as a characteristic value. Decreasing threshold value for this characteristic length will remove gravity-wave influenced clouds from the dataset. For large values of the length threshold, e.g., larger 1000 m, only clouds that are strongly forced by gravity waves will be removed, whereas going to shorter 440 thresholds (< 500 m) will also remove weakly gravity-wave-influenced clouds. Fig. 9 shows an increase in the fraction of ice-containing clouds below −12 • C with decreasing correlation-length thresholds from 30000 to 300 m. At Punta Arenas, the fraction of ice-containing clouds increases from 0.5 to 0.85. In the temperature interval between −15 and −12 • C, the fraction is 0.05 to 0.1 lower at Punta Arenas compared to Leipzig and Limassol. Hence, clouds with fully developed turbulence show similar ice-formation frequencies, independent of the location, with indications for a still slightly reduced ice-formation 445 efficiency over Punta Arenas.

Comparison of radar reflectivity factor of the ice virga
Prior studies of Zhang et al. (2018) identified a strong contrast in radar reflectivity factor between the different 30-deg latitude bands of the globe, with the southern hemisphere mid-latitudes (30-60°S) showing the lowest mean reflectivity of all regions.
They concluded that this difference in reflectivity factor is associated to a respective difference in ice crystal mass and number 450 concentration. In the following we provide a similar representation of regional contrasts of ice-virga reflectivity from groundbased perspective that is based on a single radar instrument. As described in Sec. 2.4, the amount of ice formed in the mixed phase layer is measured at six height bins (180 m) below the base of the liquid dominated cloud top and hence at the top of the virga (Bühl et al., 2016). Fig. 10a shows the cloud top temperature-resolved statistics of reflectivity for the three stations, which is based on the full cloud dataset, including the wave-influenced clouds (as these are also included in the study by Zhang  We further investigated the properties of the ice-forming liquid-dominated cloud top layers and found that cloud thickness 460 agrees within 40 m above −30 • C. The ice-to-liquid content ratio (Fig. 10b) is smaller at Punta Arenas, than at the northern hemispheric locations, especially (factor 3) between −24 and −20 • C, but also above −10 • C. Hence, in these temperature regimes, the liquid phase is less efficiently converted into ice in stratiform clouds above Punta Arenas.

Discussion of observed contrasts in properties of shallow clouds
In the previous section, a comprehensive overview of the aerosol conditions and stratiform cloud properties at the strongly 465 contrasting sites of Leipzig, Limassol, and Punta Arenas was presented. Several aerosol-, temperature-, surface-coupling-, as well as dynamics-related differences have been identified and will be discussed in the following.
The analysis of the profiles of aerosol optical properties obtained from the Polly XT lidar observations reveals almost similar aerosol load at Punta Arenas and Limassol between 0 and −10 • C. Leipzig shows a higher aerosol load in this temperature interval. For lower temperatures the lowest average extinction was observed at Punta Arenas. Low values of particle depolar-

470
ization ratio indicate that non-spherical particles, such as mineral dust are completely absent at Punta Arenas below −10 • C.
The absence of continental aerosol species in the free troposphere is also consistent with the Southern Ocean south of Aus- Figure 11. Fraction of ice containing clouds over temperature for Leipzig, Limassol, and Punta Arenas when considering only fully turbulent clouds with an autocorrelation coefficient smaller than 0.8 for a horizontal shift of 300 m (see Fig. 7) and cloud bases in the free troposphere above 2 km height (see Fig. 9).
load. Fig. 11 also depicts the low frequency of ice formation in free-tropospheric fully-turbulent clouds at temperatures above −15 • C. With less coupling to the near-surface aerosol reservoir, ice formation is strongly suppressed compared to Leipzig and Limassol.

510
The ice mass formed by stratiform liquid layers is lowest at Punta Arenas, as was found by investigating the radar reflectivity factor in the ice virga as a proxy. Especially between −28 • C and −16 • C, the observed radar reflectivity factor is up to 7 dB lower, compared to both sites in the northern hemisphere. This result is consistent with estimates from space-borne sensors covering the full Southern Ocean (Zhang et al., 2018). Contradicting, Arctic mixed-phase clouds were found to respond with lower IWC to increased loads of anthropogenic pollution (Norgren et al., 2018). As for the lower frequency of ice formation 515 discussed above, the difference in ice mass coincides with a lack of dust INPs at these temperatures. Slight differences in IWC or Z, respectively, might be explained by a slower mass growth rate, caused by a smaller vapor diffusion coefficient at higher ambient pressure (Hall and Pruppacher, 1976). When temperature and particle size are considered similar, the stratiform clouds subject to this study, experience 10% to 20% larger growth rates at Leipzig and Limassol than at Punta Arenas. Hence, the difference of a factor 3-6 larger ice mass cannot be explained solely by this effect alone.

520
Frequent occurrences of ice forming clouds above −10 • C were found, which were not covered by studies based on spaceborne active remote sensing. Similar to their colder counterparts, they show a lower ice amount in the virga above Punta Arenas, but with smaller difference compared to the sites in the northern hemisphere. With an average reflectivity of −36 dBZ, they are usually well below the detection limit of the CloudSat satellite in the A-Train constellation (Bühl et al., 2013a(Bühl et al., , 2016. Occurrence of such clouds in other parts of the Southern Ocean cannot be ruled out. A misclassification of supercooled drizzle 525 clouds as ice containing is unlikely, as they exceed −30 dBZ neither at cloud top, nor in the virga. From previous studies it is known that the onset of drizzle formation is usually associated to higher reflectivities, either above approximately −20 dBZ (Liu et al., 2008;Acquistapace et al., 2019) or at least above −30 dBZ (Wu et al., 2020).

Summary and Outlook
This study investigated contrasts in aerosol-cloud interactions in shallow supercooled stratiform clouds observed with the 530 ground-based remote sensing supersite LACROS at Leipzig, Limassol, and Punta Arenas.
Sampling the profiles of optical properties with temperature as a vertical coordinate revealed aerosol load at temperatures between −15 and 0 • C being within a factor of 2 at Punta Arenas and Limassol. This finding is related to the cold and (compared to the free troposphere) aerosol laden boundary layer at Punta Arenas. At lower temperatures, the lowest β p and extinction was observed over Punta Arenas, the highest over Leipzig. The very low particle depolarization ratio at Punta Arenas between −25 535 and −10 • C, suggests the absence of mineral dust in a temperature regime, where dust is known to be an efficient INP. An estimate of INP concentrations at the respective temperatures based on the optical properties reveals differences of 1-4 orders of magnitude between Punta Arenas and the two northern hemispheric sites. In absence of abundant INPs from marine sources, the contribution of terrestrial sources causes strong variability.
The phase occurrence frequencies showed a higher fraction of ice containing clouds at weakly-supercooling temperatures 540 of above −10 • C compared to prior lidar-only studies. A cloud radar with a sensitivity better than −40 dBZ is needed to sufficiently characterize low ice water contents in the virga formed by shallow stratiform clouds in this temperature regime.
Coupling to the boundary layer increases the frequency of ice formation at slightly supercooling temperatures at all sites. The strongest contrasts in the ice-formation frequency between free-tropospheric and surface-coupled conditions were found for Punta Arenas. This finding is in compliance to the found contrasts in the INP profiles at the three sites and further indicates 545 that the free-tropospheric INP reservoir over the Southern Ocean is limited.
Frequent liquid only layers below −20 • C at Punta Arenas were found to be associated with orographic gravity waves, causing two implications: (1) potential phase misclassification by stationary observers due to horizontal separation of ice and liquid phase; (2) sustained liquid water in updrafts, because the associated vertical velocities allow supersaturation over water; The newly developed Doppler lidar autocorrelation approach helps to address "[. Appendix A: Wave-cloud example The automated time-height-resolved air mass source attribution described by  is used to characterize air mass origin for the LACROS observations. As in the original publication, 10-day HYSPLIT ensemble backward trajectories are calculated in intervals of 3 hours and 500 m throughout the period of the deployment. For the calculation of the residence times, a reception height of 2 km is used. Additionally to the MODIS based surface cover classification and the the named geography (individually for Leipzig, Limassol, and Punta Arenas), 30 • -wide bands of latitude are used to characterize merid-585 ional transport. The assignment of the surface cover classed and named geography are depicted in Fig. 2 and 3 of . The average residence time for each set of surface types is shown in Fig. B1. However, as for the aerosol optical