Ship-Based Observations and Climate Model Simulations of Cloud Phase Over the Southern Ocean

Abstract

The Southern Ocean (SO) clouds exert a significant influence on the Earth's radiation budget. We analyzed ship-based remote sensing observations of SO clouds over a 5-month long Department of Energy (DOE) ARM Measurements of Aerosols, Radiation and Clouds over the Southern Oceans (MARCUS) field campaign to better understand cloud phase variability. We developed a method to classify eight categories of hydrometeors (ice, liquid, mixed phase, rain, drizzle, snow, aerosols, and clear sky) based on measurements of lidar, radar and radiosondes. Cloud thermodynamic phases (liquid, ice and mixed phase) at coarser scales were further derived to compare with the DOE Energy Exascale Earth System Model version 1 (E3SMv1) simulation and the Earth Model Column Collaboratory (EMC²) instrument simulator. For a scale-aware comparison with climate model simulations, we found that spatially averaging the raw remote sensing data (e.g., backscatter and reflectivity) results in increased cloud cover and cloud liquid, whereas, directly averaging cloud phase from higher to lower resolution maintains clear air regions and is thus recommended for in-cloud frequency comparisons. For cloud thermodynamic phases in stratiform clouds, the E3SMv1 underestimates cloud ice and overestimates cloud liquid at temperatures between −40°C and 0°C. When latitudes increase, both observations and simulations show a transition of dominant phase from liquid to ice for cloud tops as well as for the entire cloud columns, but the model underestimation of ice phase is more severe at higher latitudes. Such model bias is unlikely caused by spatial scale differences or lack of heterogeneity in cloud vertical structure in the simulation.

Key Points

• Cloud phase with respect to temperature and latitude over the Southern Ocean was identified using 5-month ship-based measurements

• Compared with observed cloud phase, the E3SMv1 model underestimates ice and overestimates liquid phase between −40°C and 0°C

• Observations and simulations show similar transition from liquid to ice-dominated cloud columns and cloud top layers from 43°S to 70°S

Plain Language Summary

The Southern Ocean region of Earth consists of vast open waters with extensive cloud cover throughout the year. This cloud cover prevents the ocean water below it from absorbing solar heat and maintains global ocean and atmospheric circulation patterns. Thus, it is important to improve our understanding and our ability to model this cloud cover to better predict future changes in the Earth's climate. Here, we analyzed atmospheric data obtained by a ship traversing the open waters between Australia and Antarctica for nearly 5 months. We examined observed cloud properties such as cloud ice and liquid proportions and compared them with global climate model results. We found that the model simulates more liquid and less ice at lower temperatures when compared with observations. Such model bias becomes larger at higher latitudes, indicating that Earth's surface may absorb more solar radiation than the model estimates, especially around Antarctica. Both the observations and the model show that ice clouds are common at higher latitudes whereas liquid clouds are common at lower latitudes.

1 Introduction

The Southern Ocean (SO) region is characterized by vast open waters and extremely high (80%) annual low-level cloud cover (Hu et al., 2010). These clouds prevent the darker ocean waters below them from considerable warming by reflecting most of the shortwave solar radiation back to space. The SO clouds thereby maintain a delicate surface energy balance which has a cascading effect on global ocean circulations as well as global atmospheric circulation patterns (Hwang & Frierson, 2013). Thus, a number of recent studies have focused their attention on the comparisons between observations and global climate model (GCM) simulations in terms of cloud properties over the SO. Recent modeling efforts showed that SO clouds in GCMs suffered from large uncertainties and biases and improved representations of cloud processes are needed (Bodas-Salcedo et al., 2014, 2016; Kay, Bourdages, et al., 2016; Trenberth & Fasullo, 2010). Tan et al. (2019) also showed that variations in thermodynamic cloud phase can significantly affect the cloud optical depth values over the SO using satellite measurements. However, compared with the Arctic where recent efforts have improved long-term observational statistics (Ireland, 2022), lack of land sites over the SO has hampered data collection efforts. While a number of studies have investigated long-term remote-sensing data sets over McMurdo Station, Antarctica (Lubin et al., 2020; Scott & Lubin, 2016; Yip et al., 2021), more observations are needed over the open waters. Recently, some airborne as well as ship-based campaigns were undertaken over the SO to evaluate cloud properties such as cloud phase, liquid water content (LWC) and ice water content (IWC), and the effect of anthropogenic and biological aerosols on the clouds (e.g., D’Alessandro et al., 2019; McFarquhar et al., 2021; Yang et al., 2021), providing opportunities for model evaluation in this remote region.

Mixed-phase clouds, which consist of both liquid and ice particles, can significantly alter cloud albedo depending on vertical phase partitioning. They can also affect cloud lifetimes through glaciation, resulting in changes in cloud cover and consequently modifying the radiative balance (Matus & L’Ecuyer, 2017). Evaluation of GCMs in the Coupled Model Intercomparison Project Phase 5 (CMIP5) often shows biases in high-latitudinal mixed-phase clouds, which lead to rapid glaciation and significant shortwave biases over SO (McCoy et al., 2016). This was largely due to an underestimation of cloud top supercooled liquid water (SLW). Tuning the model parameterization to allow more SLW has been found to reduce the shortwave bias (Kay, Bourdages, et al., 2016; Kay, Wall, et al., 2016). Observations have confirmed that cloud top SLW layers are common over the SO and Antarctica (Yip et al., 2021; D. Zhang et al., 2010; D. Zhang et al., 2019; Y. Zhang et al., 2019). Analysis of satellite data further shows that cloud top phase varies with latitudes (Mace, Protat, & Benson, 2021). However, since satellite measurements are often affected by cloud top beam attenuation and show large uncertainties in cloud vertical structure (Chepfer et al., 2013), other types of measurements such as ground-based observations are needed to further investigate the variation of cloud vertical structure and cloud phase vertical stratification with respect to latitudes.

In addition to previously reported biases of liquid phase in climate model simulations, several studies have shown that ice phase is likely underestimated in climate model simulations, potentially due to the underestimation of secondary ice production (SIP) processes in model parameterizations (Zhao, Liu, Phillips, et al., 2021; Zhao & Liu, 2021). Other studies using aircraft and satellite observations also indicated that SIP processes likely increase the amount of cloud hydrometeors in ice phase over the SO (Mace, Protat, & Benson, 2021; Zaremba et al., 2021). Several studies evaluated newer versions of climate models with improved representations of SLW in mixed-phase clouds compared with the CMIP5 models and found remaining biases in ice phase. For example, a previous study used aircraft observations over the SO to evaluate the cloud microphysical properties in several climate models and found lower IWC in model simulations (Yang et al., 2021). Another study used ground-based observations over McMurdo station, Antarctica to evaluate a climate model simulation and found underestimation of simulated ice phase frequency when cloud fraction is higher than 0.6 (Yip et al., 2021). Even though these two previous studies found underestimation of ice phase in terms of mass concentration and occurrence frequency in climate models, several science questions are not fully addressed. For example, it is unclear if the climate models underestimate ice phase at cloud top layer as well as throughout the entire cloud layer. It is also unclear if the dependency of cloud structure and cloud phase with respect to latitudes is captured by climate model simulations. These questions unfortunately cannot be addressed by in-situ aircraft observations which are limited to one-dimensional (1-D) sampling, or by stationary ground-based observations which are limited to one location. Compared with aircraft-based or ground-based observations, ship-based remote sensing observations show a unique advantage by sampling the vertical structure of clouds over a wide range of geographical locations. Even though ground-based remote sensing measurements such as lidars suffer from beam attenuation during heavy rain periods or deep convective systems, they can provide higher-resolution, continuous measurements of low-level stratiform clouds which tend to dominate the SO with relatively lower uncertainty compared with spaceborne remote sensing observations. Thus, using long-term ship-based measurements over the SO will help improve our current understanding of SO cloud properties such as cloud phase, vertical structure, and variations with latitudes, and further improve the current GCM cloud parameterizations.

In this paper, we analyze ship-based measurements over a duration of 5 months taken during the Measurements of Aerosols, Radiation and Clouds over the Southern Oceans (MARCUS) field campaign in late 2017 and early 2018. We identify cloud phase using on-board instrumentation and compare the statistical distributions of observed cloud properties with simulations from the Department of Energy (DOE) E3SM version 1 (hereafter referred as E3SM). We examine different spatial averaging methods needed to conduct a scale-aware comparison between high-resolution observations and coarse-resolution model simulations. We also compare the scale-aware observations with a remote sensing simulator which can replicate the observed variables from the E3SM results. We then investigate the latitudinal variation in cloud phase vertical stratification as the ship travels between the midlatitudes and polar regions. We discuss the implications of the findings and possible reasons for discrepancies between the observations and model results. Such an analysis will help improve our understanding of SO cloud phase and improve cloud representations in GCMs. The remainder of this paper is organized as follows: In Section 2, the remote sensing data set and the variables used for the analysis are introduced along with the simulation setup of E3SM and its instrument simulator. In Section 3, a hydrometeor classification algorithm is developed and discussed. In Section 4, high-resolution case studies are presented, and spatial averaging methods are discussed. Using the observations from the entire campaign, cloud phase distributions are further investigated with respect to temperature and latitude. Comparisons between the observations and model simulations are also shown in this section. Finally, in Section 5, we summarize the key findings from the analysis and discuss possible explanations and implications.

2 Observations and Model Simulations

2.1 Remote Sensing Observations

The MARCUS field campaign took place over the SO between Australia and Antarctica from 21 October 2017 to 23 March 2018. The observations were made using the United States DOE Atmospheric Radiation Measurement (ARM) Mobile Facility 2 (AMF2), which was installed on-board the Australian icebreaker research vessel Aurora Australis (AA). The facility collected continuous data as the ship completed a total of 4 voyages that started from Hobart, Tasmania and stopped at multiple Antarctic bases as well as Macquarie Island (Figure 1). Of the nearly 5-month long campaign, there were 126 days with reliable observations after removing the periods flagged as incorrect or suspect from the ARM data repository. The data spanned between 43°S and 70°S covering a wide range of high-latitude SO cloud measurements (McFarquhar et al., 2021).

Figure 1

Aurora Australis geo-location data during MARCUS field campaign show that the vessel made four voyages (V1, V2, V3, and V4) that started from Hobart Australia and sailed to multiple Antarctic stations as well as Macquarie Island. V1 occurred between 29 October and 3 December 2017. V2 was between 13 December 2017 and 10 January 2018. V3 was between 16 January and 4 March 2018 and V4 was between 9 and 25 March 2018.

The on-board instrumentation included a variety of instruments, of which we used the following: the vertically pointing Marine W-band 95 GHz ARM Cloud Radar was used to obtain radar reflectivity (dBZ), mean doppler velocity (m s⁻¹), and doppler spectrum width (m s⁻¹) (Kollias et al., 2007). The high-frequency radar observations were obtained at 5 s time and 30 m vertical resolution. The micropulse lidar (MPL) operating at 532 nm wavelength was used to obtain the total attenuated backscatter photon count (μs⁻¹) and the linear depolarization ratio (LDR) obtained at 3 s and 30 m resolution, respectively. A two channel microwave radiometer (MWR) was used to obtain the liquid water path (LWP) at 30 s. The Balloon-borne Sounding System was launched four times daily, which provided atmospheric thermodynamic profiles including pressure, temperature, moisture, wind speed and direction. The ARM INTERPSONDE value-added product (VAP) further provides continuous thermodynamic profiles at 1 min temporal and 20 m vertical resolution (Jensen et al., 2022). The ARM Vaisala Laser Ceilometer (CEIL) which operates at 910 nm with a vertical resolution of 10 m was used to obtain estimates of cloud base heights along with the ARM VAP ARSCL which also provides cloud top heights using radar measurements. A smoothing and filtering algorithm was applied to the observational data set in order to remove noise and decrease data granularity. Each radar and lidar data sets were first smoothed using a 3 × 3 moving average convolution method and then filtered to ensure that at least 5 points within the 3 × 3 sliding neighborhood contained non-garbage numeric values. Thereafter, each variable with time and height dimensions was averaged to 1 min in time and 60 m vertical resolution to generate a data grid for further analysis.

2.2 E3SM version 1 model

In this study, we use the atmospheric component of the DOE's E3SM version 1 model (Rasch et al., 2019). In this model, the cloud microphysics in stratiform clouds is parameterized using the two-moment MG2 scheme (Gettelman & Morrison, 2015). The heterogeneous ice nucleation on dust and black carbon in mixed-phase clouds follows the classical nucleation theory (Hoose et al., 2010; Wang et al., 2014). The aerosols are predicted by the four-mode version of the Modal Aerosol Module (Liu et al., 2012, 2016). In addition, the model uses a third-order turbulence closure parameterization named Cloud Layers Unified By Binormals (Bogenschutz et al., 2013; Golaz et al., 2002; Larson & Golaz, 2005) to unify the treatments of shallow convection, planetary boundary layer turbulence, and cloud macrophysics whereas the deep convection parameterization uses Zhang & McFarlane (1995). The E3SM model was integrated from 1 April 2017 to 30 March 2018 at 1° horizontal resolution and 72 vertical layers with a time step of 30 min. The results during the MARCUS campaign are used for analysis, while the first approximate 6 months are treated as model spin-up. To provide an apples-to-apples comparison, the modeled cloud properties were sampled along the MARCUS ship track every 10 min. Moreover, temperature and horizontal wind components were nudged to the Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2) (Gelaro et al., 2017) meteorology with a relaxation timescale of 6 hr utilizing the linear function nudging strategy (Sun et al., 2019).

2.3 Earth Model Column Collaboratory Simulator

The Earth Model Column Collaboratory (EMC²) is an open-source remote sensing instrument simulator and sub-column generator which generates lidar and radar variables from large-scale model output such as the E3SM (Silber et al., 2022). EMC² was run using the microphysics approach which is primarily suited for stratiform clouds with lidar-based hydrometeor diameter calculations consistent with the MG2 microphysics scheme (Gettelman et al., 2015) implemented in E3SM. The first three radar moments, namely radar reflectivity (Z_e), mean Doppler velocity (V_D), and Doppler spectral width (W_D) are also calculated. However, at present, W_D calculations in EMC² do not include spectral broadening terms other than the microphysical broadening, resulting in considerably lower W_D compared with observations_. The simulator generates 20 sub-columns within the E3SM time and vertical resolution. This provides a finer-scale output grid thereby allowing direct comparison of cloud phase between coarse-scale GCM output and fine-scale observations. Thus, the EMC² simulator results were used to obtain a sub-grid sampling approach to the model output and provide a scale-aware comparison with the spatially averaged observations.

3 Methods of Hydrometeor Classification in High-Resolution Observations

An algorithm that automatically identifies seven hydrometeor types (i.e., aerosols, liquid, ice, mixed, rain, snow, and drizzle) as well as clear-sky conditions using high-resolution observations is developed based on the method of Shupe (2007). This classification algorithm includes six main steps using a combination of radar, lidar, LWP, and temperature profiles (Figure 2). The algorithm is applied to each pixel of high-resolution observations in the MARCUS campaign.

Figure 2

Schematic of the cloud phase classification method used to identify cloud phase during the MARCUS campaign. It shows the modification to step 1 of the Shupe (2007) hydrometeor classification algorithm. V1–V4 indicate the Aurora Australis voyage numbers, LDR is the linear depolarization ratio (unitless) and β_p is the MPL particulate backscatter coefficient in m⁻¹ sr⁻¹.

3.1 MPL Hydrometeor Classification

In step 1 of the classification, the algorithm contrasts lidar backscatter intensity and LDR to separate cloud liquid from ice for each pixel. Lidar backscatter is more sensitive to high particle number concentrations usually associated with cloud liquid (Sassen, 1984), whereas LDR is more sensitive to non-spheroidal particles usually associated with cloud ice. Several preparation procedures need to be conducted before using the lidar data for hydrometeor classification. First, the MPL instrument output does not directly measure the backscatter intensity (β_p, also known as the particulate backscatter coefficient) required for liquid and ice phase classification. Such a variable needs to be first calculated (see Appendix A) from the normalized relative backscatter (NRB) and then compared with the MPL LDR. Second, thresholds to separate ice, liquid, and aerosols need to be examined and selected for individual voyages. During voyages, V1 and V2, MPL LDR values suffered from a faulty glass window which resulted in biases in LDR. Thus, we performed an analysis similar to Mace, Protat, Humphries, et al. (2021) on the data set used for this work to identify the optimal thresholds for hydrometeor classification for each voyage (Figure 3). Bivariate distributions of β_p and LDR are examined to determine thresholds for hydrometeor classification. In order to find a clear cutoff between cloud ice and liquid, case studies are chosen when the existence of mixed phase was minimal, since mixed phase often occupies a large region in the β_p—LDR space and may blur the distinction between liquid and ice (Lewis et al., 2020). Comparing two case studies (Figures 3c and 3d), the original threshold of LDR = 0.25 used in Shupe (2007) shows a reasonable separation among liquid, ice and aerosols for 18 November, during V1, while the threshold of LDR = 0.2 works better for 02 March 2018 during V3. For the entire MARCUS campaign, LDR = 0.25 is used for V1 and V2, and LDR = 0.2 is used for V3 and V4 for the classification method. Such a difference was possibly caused by the re-calibration of the instrument housing between different voyages. The LDR threshold for V3 and V4 also agrees with the threshold used by Mace, Protat, Humphries, et al. (2021) for classifying ice during MARCUS. The LDR threshold for V1 and V2 used in this work differs from that used in Mace, Protat, Humphries, et al. (2021), which is 0.3, possibly due to different subsets of data being selected for various analyses. Instrument uncertainty for calibrated MPL LDR has been shown to be less than 0.015 previously (Welton et al., 2018), which is smaller than both LDR thresholds (i.e., 0.2 and 0.25) used in this work for hydrometeor classification. The aerosol category may be re-classified in the following steps if radar reflectivity shows a measurable value (i.e., Z_e > −60 dBZ). For the analysis of cloud thermodynamic phases, aerosols are considered as part of the clear-sky conditions.

Figure 3

Bivariate histograms of micropulse lidar (MPL) linear depolarization ratio (LDR) in relation to the obtained backscatter coefficient (β_p). An offset of 0.1 in LDR induced by instrument errors is seen in (a) voyages 1 and 2 (V1 and V2) in MARCUS campaign. Such an offset is not seen in (b) voyages 3 and 4 (V3 and V4). Similar offset of 0.1 in LDR is seen in hydrometeor population on (c) 18 November 2017 from V1, but not seen in (d) 2 March 2018 from V3. The color bar indicates the number of samples for each bin and the dashed lines show the thresholds used to identify cloud hydrometeors.

3.2 W-Band Radar Hydrometeor Classification

In steps 2–4 of the hydrometeor classification algorithm, radar reflectivity (Z_e) and mean doppler velocity (V_D) are used to complement the lidar measurements. Since the lidar beam is often attenuated before reaching cloud top for optically thick clouds, radar Z_e, V_D, and Doppler spectral width (W_D) along with temperature data from ARM INTERPSONDE become the main variables used to classify cloud phase after lidar attenuation. As for quality control of the radar data, Kollias et al. (2019) showed that ARM millimeter radars suffered from a systematic under-calibration when compared with CloudSat radars by 4–8 dBZ and the same issue also affected the MARCUS data set. This offset is much larger than the instrument uncertainty when compared with satellite measurements (Protat et al., 2011) and thus, the MARCUS radar reflectivity was offset by adding 7 dBZ to account for this under-calibration. First, pixels classified as liquid phase by the lidar are further classified into mixed phase (T < 0°C) and drizzle (T > 0°C) by checking their reflectivity (Z_e > −17 dBZ), Doppler velocity (V_D > 1 m s⁻¹) and spectral width (W_D > 0.4 m s⁻¹). Larger W_D values typically indicate the presence of mixed phase regions and are used to separate them from liquid or ice regions (Shupe et al., 2004). Similarly, strong precipitation such as rain and snow are separated from lidar liquid and ice classification, respectively, for the largest radar reflectivity (Z_e > 5 dBZ) in step 3. Rain is further separated from drizzle using V_D > 2.5 m s⁻¹ following Beard and Pruppacher (1969), which showed that the terminal velocity of typical raindrops (d > 1 mm) is larger than 2.5 m s⁻¹. Figure 4 shows a sample case study for both lidar only (Figure 4a) and lidar + radar (Figure 4b) phase classification for 3 November. Radar observations are able to complement the lidar phase classification and re-classify a number of lidar-classified pixels into mixed phase, rain, and drizzle. Radar observations are also able to identify cloudy regions above lidar attenuation level and classify cloud phase using Shupe (2007) step 4 shown in Figures 4c–4e. Throughout the entire campaign, lidar-detected cloud pixel count is 37.1% lower than radar for low-level stratiform clouds. The bivariate histograms of radar observations also show that the thresholds used in steps 2–4 of Shupe (2007) work reasonably well for the MARCUS data and thus we follow the same steps 2–4 as shown in that previous work.

Figure 4

Cloud phase classification on 3 November 2017 by using (a) lidar only and (b) lidar + radar. The dotted lines show temperature contours in °C. (c)–(e) Bivariate histograms of mean doppler velocity (V_D) in relation to radar reflectivity (Z_e) for different ranges of temperature and mean Doppler spectral width (W_D). The color bar indicates the number of samples for each bin and the dashed lines show the thresholds used to identify cloud hydrometeors, which are consistent with those used in Shupe (2007). V_D with the downward direction is defined as positive.

While V_D values are used to identify cloud phase in the aforementioned steps, it is important to note that during MARCUS, the stabilizing platform on which the W-band radar was installed, suffered some malfunctioning. This platform was supposed to ensure that the radar always pointed toward the zenith by accounting for ship pitch and roll. However, a lag between the ship navigation system and the platform resulted in noises in V_D values, which seem to have a Gaussian-like behavior. To quantify the impacts on cloud phase distributions, we compared case studies of V_D measurements with higher and lower accuracies, such as days from voyage 1 leg 2 and voyage 2 leg 2, respectively (Figure S1 in Supporting Information S1). The cloud phase occurrence frequencies do not show significant differences between these cases, indicating that the impacts of V_D errors are likely averaged out. In addition, since V_D is one of seven variables used for hydrometeor classification and the comparisons between observations and simulations are based on statistical distributions of cloud phases over the entire campaign, it is unlikely that the Gaussian noises in V_D would lead to significant impacts on the comparison results.

Further classification in step 5 includes temperature constraints such as converting any cloud ice identified at temperatures above 0°C to liquid and any liquid below −40°C to ice. Step 6 includes LWP constraints, that is, ensuring existence of cloud liquid when LWP >25 g m⁻² since the uncertainty in MWR LWP measurements is ∼25 g m⁻² (Turner et al., 2007), and reclassifying liquid to ice when LWP = 0 g m⁻². The original step 7 in Shupe (2007) method homogenizes the cloud phase by using a 7 × 7-pixel smoothing algorithm to generate a spatially averaged data set. However, such smoothing algorithm may not be most suitable for comparison with coarse-scale model simulations. In the next section, we examine several methods to optimize scale-aware, definition-aware comparisons with climate model simulations.

4 Results

4.1 Case Studies of Hydrometeor Classification Using High-Resolution Observations

Time-series plots of high-resolution observations (1 min in time and 60 m vertical resolution) as well as corresponding hydrometeor classification from two case studies during MARCUS are shown in Figures 5 and 6. Time series for 3 November 2017 (Figure 5) shows a low-level precipitating stratocumulus topped layer with the melting level passing through just below 1 km in altitude. Mean doppler velocity V_D with a downward direction is defined as positive. Figure 5b shows increasing V_D values just below the 0°C line, indicating the presence of drizzle and rain since liquid precipitation tends to fall with higher velocities. Calculated lidar backscatter coefficient (β_p) values (Figure 5c) show large gradients near locations where cloud base and top levels have been detected by ARSCL, which further increases confidence in the β_p calculations. Pockets of very high LDR (Figure 5d), usually associated with ice, can be seen near cloud top between 1500 and 2100 UTC. The hydrometeor classification algorithm is able to identify these pockets as ice as indicated by the blue patches near regions with high LDR values (Figure 5e). The algorithm also detects drizzle and rain near the surface, as shown by dark green patches, associated with high doppler velocities and LWP values larger than 200 g m⁻² (Figure 5f). Similar remote sensing signals of drizzle were previously seen from observations of closed-cell stratocumulus clouds (Cadeddu et al., 2020). Lower LWP values near 0600, 0900 and between 1500 and 0000 UTC associated with non-precipitating stratiform clouds are also well identified in the hydrometeor classification. We also noticed a small cloud nearly 3 km in altitude between 0900 and 1200 UTC in panels a, b, and e, which was missed by the lidar due to beam attenuation but detected by the radar measurements. The hydrometeor classification of multi-level clouds is often challenging due to lidar attenuation and thus, in this study, we focus our attention on single-layer stratiform clouds and ignore any such secondary higher altitude layers not detected by the lidar. Low latitude observations such as 3 November often showed cloud top ice and liquid near cloud base due to the warmer ocean waters at lower latitudes.

Figure 5

Time-series plot with altitude on the Y axis for 3 November 2017 case study. From top to bottom, each row shows the corresponding radar reflectivity (panel a), mean Doppler velocity (b), lidar backscatter coefficient (c), linear depolarization ratio (LDR), and classified cloud phase (d) and (e). The top x-axis of all panels shows the ship latitude, the dashed lines on (b) and (e) show temperature contours. The black solid and dashed lines on (c) show estimated cloud top and base heights using the ARSCL data.

Figure 6

Similar to Figure 5, except for 18 November 2017 case study.

Time series for 18 November 2017 (Figure 6) consisted of higher latitude observations (∼68°S) compared with November 03 (∼48°S) and shows a liquid-topped precipitating ice cloud. Mean doppler velocity (Figure 6b) does not show a sudden increase as shown in Figure 5, since the temperatures are much lower than in the previous case study and the melting level does not exist within the cloud layer. β_p shows higher values near cloud top (Figure 6c), usually associated with a liquid layer due to higher droplet number concentrations generating a larger backscatter compared with ice. At lower altitudes, the precipitation is primarily ice as indicated by higher values of LDR (Figure 6d) and lower β_p. The phase classification method (Figure 6e) is able to identify a thin cloud top liquid layer with mean thickness of 200 m between 1500 and 2100 UTC, while a thicker layer primarily with ice precipitation is seen beneath the cloud top. The existence of the cloud top SLW layer is confirmed using mean LWP measurements from the MWR (Figure 6f). The LWP measurements show a marked increase in LWP after 0600 UTC, coincident with continued presence of cloud liquid until 0000 UTC as seen in the hydrometeor classification. Extremely low LWP values are seen when no liquid layer is detected. Small gray patches near the surface indicate possible aerosols or noise due to MPL window issues. The SLW layer at cloud top seen here is common at higher latitudes and agrees well with a number of recent SO observational studies (Alexander et al., 2021; Huang et al., 2012; Mace, Protat, & Benson, 2021; Yip et al., 2021).

4.2 Two Methods for Cloud Phase Identification at Coarser Scales

To compare the high-resolution remote sensing observations with coarse-resolution E3SM results, we need to spatially average the observations to match the E3SM grid size. After applying the hydrometeor classification, as shown in Figure 2, eight categories are derived from high-resolution observations, that is, cloud liquid droplets, ice particles, mixed phase, drizzle, rain, snow, aerosols, and clear sky. We further reduce the eight categories to four categories (Figure 7), including a single liquid phase (by combining liquid droplets, drizzle, and rain), a single ice phase (by combining ice particles and snow), mixed phase, and clear-sky conditions (including both aerosols and clear sky). Note that the meaning of mixed phase in hydrometeor classification is different from the mixed phase in cloud phase classification, since the former is a mixture of liquid droplets and ice particles at high-resolution pixel level, while the latter is a mixture of ice phase and liquid phase at coarse resolution containing multiple pixels. With the simplified cloud phase classification, spatial averaging is further applied using two methods (named as M1 and M2). Eventually, coarser-resolution observations are compared with model simulations for the simplified cloud phase categories—liquid, ice, mixed phase as well as clear-sky conditions correspondingly.

Figure 7

A step-by-step schematic of the consolidation and spatial averaging of the observed high-resolution hydrometeor classes into four coarse-resolution cloud phase categories. Similar cloud phase categories are also obtained from E3SM hydrometeor mass output and sub-column averaged EMC² hydrometeor classes. Highlighted red boxes show the steps where spatial averaging occurs.

Method 1: For averaging high-resolution data into coarser spatial scales in horizontal and vertical, we use the following equation to calculate the spatial fraction of a certain phase based on the number of pixels. Note that the eight categories of hydrometeor classification have been reduced to four categories before applying this calculation (Figure 7):

(1)

Here, p represents ice, liquid, and mixed phase. If any phase frequency ≥ 0.9, then the pixel is classified as phase p. If 0.1 < < 0.9 or if ≤ 0.1 and ≥ 0.5, then it is mixed phase, else the pixel is defined as liquid phase. The threshold of 0.1 < < 0.9 is frequently used to define mixed-phase clouds in observations (D’Alessandro et al., 2019; Korolev et al., 1998, 2006; Yang et al., 2021) as well as in cloud-resolving models and other GCM simulations (Fan et al., 2011; Zhao, Liu, Phillips, et al., 2021). Clear sky regions are accounted for by checking if more than 50% of the total number of pixels in the averaging grid box were clear sky, then the center pixel is also classified as clear sky. This method can be considered as a modified version of step 7 in the Shupe (2007) algorithm, suited for comparison with coarse grid model results.

Method 2: Another method to obtain spatially averaged cloud phase is to average the instrument signals such as radar reflectivity, lidar backscatter to the E3SM resolution, and then use these coarser-scale signals to classify and compare the cloud phase (Figure 7). The resultant cloud phases from M2 by averaging remote sensing signals directly are more comparable to satellite remote sensing data than M1. Since both methods have their advantages and disadvantages, we shall explore both methods and compare their results with the E3SM results.

Figures 8 and 9 show a comparison between the two methods mentioned above while also exploring the effects of horizontal and vertical spatial averaging on cloud phase using the 3 and 18 November 2017 case studies as examples. The figures show a step-by-step change in cloud phase classification from high resolution to coarse resolution following Equation 1. For both M1 and M2 methods, cloud phases are first averaged vertically from 60 m to E3SM vertical resolution while maintaining 1-min horizontal resolution, then further averaged from 1-min to 10-min horizontal resolution. Comparing the effects of vertical and horizontal averaging on cloud phase distributions, horizontal spatial averaging seems to have a larger effect on 3 November (Figure 8), while vertical averaging has a larger effect on 18 November (Figure 9). This is likely due to the E3SM vertical grid size being relatively fine (100 m) at 1 km altitude in the former case, while the model grid becomes coarser (500 m) at 4 km altitude in the latter case. This result indicates that the effects of horizontal and vertical averaging depend on the vertical location of the cloud layer and the model grid size at that location.

Figure 8

A comparison of cloud phase distributions on 3 November 2017, using two spatial averaging methods—(a)–(c) M1, (e)–(g) M2, and two simulations (d) E3SM and (h) EMC² simulator. Observations of high-resolution hydrometeor classification are first consolidated into four cloud phase categories, then averaged to three resolutions: (a, e) 1-min horizontal and 60-m vertical resolution, (b, f) 1-min horizontal and E3SM vertical resolution, and (c, g) 10-min horizontal and E3SM vertical resolution.

Figure 9

Similar to Figure 8, except for 18 November 2017.

Comparing M1 and M2 results, M1 shows more mixed phase regions (Figures 8b and 8c) compared with M2 (Figures 8f and 8g). This is because using M1, high-resolution liquid and ice pixels in close proximity likely become mixed phase pixels at coarse resolution if neither liquid nor ice pixels exceed 90% of the population. On the other hand, when averaging instrument signals such as lidar backscatter using M2, the higher backscatter values from cloud liquid dominate the averaged signals, resulting in more liquid pixels and fewer ice or mixed phase pixels. This result agrees with a similar phenomenon seen in satellite-based cloud phase classification, which often shows fewer ice and mixed-phase samples (Mace, Protat, & Benson, 2021). Exceptions of the overestimation of liquid phase in M2 may also occur, especially when the liquid layer is very thin, for example, between 1500 and 1800 UTC on 18 November (Figure 9a). Such thin liquid-containing layer is partially missed by M2 (Figures 9f and 9g), since the high backscatter values of this narrow layer are diluted by the surrounding low backscatter values during spatial averaging. M1, on the other hand, identifies this cloud top layer as mixed phase (Figures 9b and 9c) since the thin layer of liquid pixels at high resolution still sufficiently reduces the fraction of ice pixels to less than 0.9.

Another difference between M1 and M2 is the effect on clear-sky regions. M1 retains some clear-sky regions in Figure 8c between 0300 and 0900 UTC, while these regions are missing in M2 results, replaced by the neighboring hydrometeor signals. This leads to an artificial increase in the number of cloudy pixels. This can be further seen in panel g, where the deeper ice region near 0300 UTC now shows a thicker ice column compared with small open-cell ice regions seen in the high-resolution observations (Figures 8a and 8b). This phenomenon of increased cloud cover due to averaging of remote sensing signals is consistent with previous studies of satellite remote sensing measurements (Cesana et al., 2016; Chepfer et al., 2013). Those studies showed that different spatial averaging methods of Level 1 satellite data lead to various overestimations of cloud cover, which further causes differences in climatologies of cloud properties and cloud radiative flux estimates derived from different satellite products. Thus, based on our findings, M1 provides a better algorithm for conducting scale-aware, definition-aware comparisons with climate model simulations compared with M2 due to its ability to maintain clear-sky regions and also better identify mixed-phase regions.

4.3 Cloud Phase Identification Using E3SM

E3SM cloud phase was defined using the ice mass fraction obtained by taking the ratio of the IWC to the total water content (Korolev et al., 1998) for each grid point.

(2)

If the ice mass fraction () > 0.9, then the grid point is classified as ice phase; 0.1 ≤  ≤ 0.9 is mixed phase, and < 0.1 is liquid phase. This method has been commonly used to define observed cloud phase (D’Alessandro et al., 2019; McFarquhar et al., 2007, 2013; Yang et al., 2021) as well as cloud phase in cloud-resolving models and other GCM simulations (Fan et al., 2011; Zhao, Liu, Phillips, et al., 2021). Along with cloud water/ice content, precipitation in the form of snow and rain is also added to the IWC and LWC, respectively. Cloud phase is calculated only for “cloudy air” defined as regions with cloud fraction >10⁻⁴. This threshold is often not enough to remove extremely small values of IWC and LWC which can add noise to the cloud phase calculations. Thus, an additional in-cloud threshold of either IWC or LWC has to be greater than 10⁻³ g m⁻³ is applied to obtain cloud phase (M. Zhang et al., 2020). Regions that do not satisfy the latter threshold are defined as clear sky.

Figures 8 and 9 show the E3SM cloud phase for case studies of 3 and 18 November 2017, respectively. On 3 November (Figure 8d), the simulated cloud base height agrees well with the observed high-resolution cloud base height (Figure 8a), while the simulated cloud top height is higher than observed by about 500 m between 0300 and 1500 UTC.

For 18 November (Figure 9d), E3SM shows a much deeper cloud with a cloud top altitude of 8–10 km while the observed cloud top height is at 4–5 km. The E3SM simulation is able to capture a mixed-phase cloud layer as seen in the observations, showing a better representation of vertical stratification of cloud phases compared with Figure 8. Yet this simulated mixed-phase layer is located near cloud base instead of at cloud top as seen in observations, and the simulated layer is also thicker (i.e., 1–2 km in depth) compared with the observed layer (500 m) even when observations are averaged to the same model vertical grid scale.

For the spatial heterogeneities of cloud phase distributions, the simulated cloud phase (Figure 8d) shows horizontal variability with time instead of the vertical variability seen in the spatially averaged observations (Figures 8c and 8g). Since the model temperature is nudged toward the MERRA2 data, this horizontal variability is less likely due to model temperature biases. This is corroborated by the fact that simulated temperature contours are very similar to the observed ones. Thus, the simulated cloud phase difference may arise from biases in relative humidity as shown by the previous work of Yip et al. (2021) as well as other microphysical parameterizations such as the Wegener-Bergeron-Findeisen (WBF) process or ice-nucleating particle (INP) concentrations.

4.4 Cloud Phase Identification Using EMC² Simulator

The fine-scale resolution cloud phase identified by EMC² is averaged to the E3SM scale by calculating the individual phase frequency following Equation 1 and using the frequency thresholds described for M1. The cloud phase results for 3 November 2017 in Figure 8 indicate that EMC² results show a similar horizontal stratification of cloud phase (Figure 8h) as the E3SM (Figure 8d). This is understandable since the EMC² cloud hydrometeor allocations follow the E3SM output. The reduction in cloud top height between 0600 and 1500 UTC compared with E3SM is likely due to the sub-grid variability introduced by the sub-columns which can lead to more frequent lidar beam attenuation near cloud top. This reduction in cloud depth results in a better agreement between EMC² and the observations (Figures 8c and 8g) with respect to cloud top height. However, between 0000 and 0300 UTC and between 1800 and 0000 UTC, EMC² shows significantly more pixels of ice and fewer pixels of mixed phase compared with E3SM. A similar phenomenon was seen on 18 November 2017 (Figure 9), where the mixed phase layer in the E3SM results between 4 and 6 km in altitude (Figure 9d) is replaced by ice phase in EMC² results (Figure 9h). This likely occurs due to the lack of spectral broadening terms in EMC² which are expected to increase mixed phase frequency if they were included (see Figure 3d). Occasionally, mixed phase is shown in EMC² results such as between 0900 and 1500 UTC in Figure 8h, due to spatial averaging of liquid and ice pixels from the higher-resolution sub-columns. Overall, fewer occurrences of mixed phase are seen in EMC² results compared with E3SM in both case studies, resulting in increases in liquid and/or ice phase occurrences in EMC² results.

4.5 Statistical Distributions of Cloud Phase With Respect to Temperature

A larger sample size is needed to examine the statistical distributions of cloud phase in relation to temperature and latitude, as well as for statistical comparisons with model simulations. While the entire campaign includes multiple cloud types such as stratiform, convective, and multi-layer clouds, all the following statistical analyses are restricted to single-layer stratiform clouds below 5 km to avoid errors in observations such as lidar beam attenuation which occurs primarily for deeper and multi-layer clouds. From the entire campaign data set, 48 days satisfy the criterion of low-level stratiform cloud type and are used to compare with the E3SM and EMC² results.

Figure 10 shows the comparison of cloud phase occurrence and in-cloud frequency between the high-resolution observations, spatial averaged observations using M1 and M2, E3SM and EMC² simulation results as a function of temperature. The high-resolution observations are at 1-min horizontal and 60-m vertical resolution, while the spatial averaged observations are at 10-min horizontal and model vertical resolution for both M1 and M2. Cloud phase occurrence frequency is calculated as a ratio of a particular cloud phase occurrence to the total cloud phase (ice + liquid + mixed) occurrence for each 5°C temperature bin. Any grid box identified as either ice, liquid, or mixed phase is considered in-cloud, whereas all other pixels are considered as clear sky. The in-cloud frequency is then calculated as the ratio of the total in-cloud occurrence to the total number of samples including both in-cloud and clear-sky conditions in each temperature bin. All bins show a number of in-cloud samples higher than 100 (Figure S2 in Supporting Information S1).

Figure 10

Cloud phase occurrence frequency is shown as a function of temperature for the entire MARCUS campaign. The temperature range is divided into 5°C bins and the latitudes are divided into 10° bins. Columns from left to right show the high-resolution observations (1-min horizontal and 60-m vertical resolution), spatial averaging Method 1 (M1) and Method 2 (M2) at 10-min horizontal and model vertical resolution, E3SM and EMC² simulator results (both at E3SM grid resolution). The top row shows the occurrence frequency for the entire latitudinal domain while the remaining rows show the occurrence frequency for each latitude bin.

Comparing observations at higher and lower resolutions, M1 shows a higher mixed phase frequency (up to 0.4) and correspondingly lower liquid and ice phase frequencies (Figure 10b) compared with the high-resolution observations (Figure 10a). Such increase of mixed-phase frequency in coarser-scale observations is also seen in the previous study using aircraft observations (e.g., Yang et al., 2021), since the coarser grids are more likely to contain both liquid and ice than smaller grids. On the contrary, M2 (Figure 10c) shows an increase in liquid frequency, a decrease in ice frequency, and a relatively similar mixed phase frequency compared with the high-resolution observations, due to direct averaging of remote sensing signals in M2 (e.g., dominating high backscatter values as discussed in Section 4.2). These differences between M1 and M2 are consistently seen at various latitudes.

The ice phase frequency in E3SM is lower than all the observations (i.e., high-resolution, M1, and M2) between −40°C and 0°C (Figure 10d). This is consistent with a recent study that showed that E3SM underestimates ice phase at temperatures below −20°C when compared with in-situ airborne measurements of cloud phase over the SO (Yang et al., 2021). Such underestimation of ice phase in E3SM is consistently shown for various latitudinal ranges, which is even larger in the higher latitudes. Correspondingly, liquid and mixed-phase frequencies are overestimated in E3SM at −20°C to 0°C and −35°C to −20°C, respectively. EMC² results (Figure 10e) show the highest ice phase frequency and the lowest liquid phase frequency at all temperatures similar to case studies in Figures 8 and 9. The phase crossover point which is defined as the temperature where the liquid and ice phase frequencies intersect each other is found to be close to 0°C for the high-resolution observations, M2 and EMC² results. The E3SM crossover point is found near −10°C for the entire latitudinal range. Overall, E3SM shows more comparable results to M1 than M2 for ice and mixed-phase frequencies. Yet E3SM shows a more similar liquid phase frequency to M2 due to the high biases of liquid phase in both M2 and E3SM. EMC² also shows more comparable results to M1 in terms of liquid and ice phase frequencies but shows lower mixed-phase frequency due to the increased sub-grid variability and lack of spectral broadening terms mentioned in the previous sub-section.

An interesting feature is observed at the higher latitudes (Figure 10p), that is, liquid phase frequency slightly increases at temperatures between −10°C and −30°C compared with lower latitudes (Figures 10f and 10k). This is likely due to the presence of SLW layers commonly observed near cloud tops at higher latitudes (Alexander et al., 2021; Huang et al., 2012; Mace, Protat, & Benson, 2021; Yip et al., 2021) which occur at lower temperatures compared with liquid layers found near surface. This phenomenon is represented as higher mixed-phase frequencies in M1 (Figure 10q) and higher liquid-phase frequencies in M2 (Figure 10r) due to reasons discussed in Figures 8 and 9. The E3SM results at the higher latitudes (Figure 10s) show higher liquid and mixed-phase frequencies at −20°C to 0°C and −35°C to −20°C, respectively, compared with lower latitude simulation results. EMC² results (Figure 10t) also show slightly increased frequencies of liquid and mixed phase at the higher latitudes. It is difficult to assess whether such latitudinal variations would still be seen in the simulations if the underestimation of ice phase frequency in E3SM were to be corrected. The E3SM in-cloud frequency shows a sudden increase below −20°C in the 60°–70°S latitudinal bin which is not seen in the observations. This is consistent with a recent study which showed that the Community Atmosphere Model version 6 (CAM6) climate model (i.e., a model with similar cloud microphysics parameterizations to E3SM version 1) tends to overestimate cloud fraction and cloud height at higher latitudes (Yip et al., 2021).

4.6 Impacts of Spatial Averaging Scales

A common challenge of comparing observations and simulations is the disparity of their spatial resolution. A wide range of spatial scales is examined using M1, aiming to assess the possible non-linearity in the effects of spatial scales on cloud phase frequency distribution (Figure 11). We chose M1 since it is shown to be a better spatial averaging method compared with M2 (Section 4.2). High-resolution observations at 1-min horizontal resolution (i.e., 1.2 km) are averaged by 10, 30, 60, and 90 min, which correspond to 12, 36, 72, and 108 km in horizontal scales, respectively. This calculation is based on a mean horizontal wind speed of 20 m s⁻² at 900 hPa in the MARCUS domain. All the observations use the same vertical resolution as the E3SM model grid. Significant changes in cloud phase frequency (i.e., decreasing ice and liquid frequencies and increasing mixed-phase frequencies) are seen when observations are averaged from 1-min to 10-min resolution. The scaling effects become much smaller when further averaging observations from 10-min to 90-min resolution. This result indicates that 10-min (i.e., 12-km) horizontal resolution is likely a critical threshold and future studies are recommended to average observations to horizontal scales no less than this value when evaluating 1° (i.e., ∼100 km horizontal resolution) model simulations. Among the three cloud phases, mixed-phase frequencies show the highest sensitivity to spatial scales of observations, since the decreases in liquid and ice frequencies both contribute to increasing mixed-phase frequencies.

Figure 11

A comparison of occurrence frequencies with respect to temperature for (a, e) ice, (b, f) liquid, (c, g) mixed phase, and (d, h) all in-cloud categories. Panels (a)–(d) show a comparison between E3SM and M1 cloud phase results at various scales to evaluate the effect of spatial averaging on cloud phase. Panels (e)–(h) show a comparison between high-resolution observations, M1, M2, E3SM, E3SM with 100 times enhanced Hallett-Mossop term (HM100) and EMC² cloud phase statistics.

E3SM ice phase frequency (Figure 10a) is consistently lower than the observations at various scales for the entire temperature range, while the liquid phase frequency in E3SM is consistently higher than all the observations. These differences between E3SM and observations are much larger than the variations seen in observations once observations exceed 12 km horizontal resolution. This result indicates that the model-observation discrepancies seen in Figure 10 are unlikely to be solely explained by the differences in spatial scales. This corresponds to a similar reduction in liquid (Figure 10b) and a corresponding increase in mixed-phase distribution (Figure 10c). This suggests that the underestimated ice and overestimated liquid frequencies by E3SM compared with observations are not due to limiting the spatial averaging of the observations to match the E3SM 10-min resolution. Any further averaging of the observations will only further reduce the liquid frequencies and not explain the higher liquid contributions in the E3SM results. The E3SM mixed (Figure 10c) and in-cloud (Figure 10d) frequencies also show higher values below −20°C compared with the observations similar to Figure 10, further confirming overestimated liquid and cloud fractions at lower temperatures as seen in Figure 9. Interestingly, EMC² shows more similar frequencies of all three cloud phases compared with the high-resolution observations (1-min and 1.2 km resolution), even though the 20 sub-columns within 1° model grid box are more likely to represent 5 km than 1.2 km horizontal resolution. For in-cloud frequencies, both E3SM and EMC² overestimate cloud cover at higher latitudes compared with all observations, while the observations regardless of horizontal scales and averaging methods (M1 or M2) show very similar results.

4.7 Impacts of the Hallett-Mossop Parameterization

The underestimation of ice phase in E3SM (as shown in Figures 10 and 11) may be affected by the model parameterization of ice microphysical processes. Previously, Zhao and Liu (2021) showed that improving model parameterizations of SIP processes can substantially increase the global ice water path and decrease the global LWP by 20% in the CAM6 model. Even though these newly developed parameterizations of SIP are still not available in the E3SM model, the current MG2 scheme in E3SM considers SIP from rime splintering, also known as the Hallett-Mossop (HM) process, at temperature between −8°C and −3°C. The effect of the HM process is being examined by artificially enhancing the HM term by a factor of 100 (Figure 11). The HM100 simulation shows no significant differences compared with the default E3SM simulations in terms of cloud phase frequency (i.e., liquid, ice, or mixed) and in-cloud frequency. The underestimation of ice phase is still seen in this sensitivity test. This may be due to the HM process being a relatively small term in the E3SM, or due to other SIP processes such as ice-ice collision fragmentation and droplet shattering during freezing being overlooked in the current E3SM parameterization. Whether improvements in ice microphysical properties can be achieved by using newly developed parameterizations such as those shown in Zhao and Liu (2021) still warrants further investigation.

4.8 Latitudinal Variation in Cloud Type and Cloud Phase Vertical Heterogeneity

Cloud properties across various latitudes can be affected by a combination of various factors, such as changes in temperature vertical profiles, ocean surface temperatures, and aerosol distributions. Latitudinal variations in cloud phase and vertical structure are examined and compared between observations and simulations in Figure 12. This analysis examines different types of cloud vertical structure at various latitudes as well as the spatial heterogeneity of the thermodynamic phase inside cloud columns. The analysis shown in Figure 12 differs from that shown in Figure 10, which does not analyze a column of cloud layer as an entire piece but rather re-groups different horizontal layers of clouds into various temperature bins. For the observations, the top and bottom boundaries of a cloud column are defined by the ARSCL obtained cloud top and base heights. If the cloud phase for the entire vertical column is the same, it is counted as a single-phase cloud column, such as ice, liquid, or mixed phase for the entire cloud column. If more than one cloud phase exists within a single vertical column (which can be either double- or multi-phase), it is classified as a stratified phase column. Panels (a–e) examine the dominant types of cloud columns at various latitudes, separated into single-, double-, or multi-phase cloud columns. Panels (f–j) examine the dominant thermodynamic phase for these columns, separated into those that are solely liquid, ice, and mixed phase or vertically stratified with more than one phase. Panels (k–o) show the dominant thermodynamic phase for cloud top layers. The cloud vertical columns are normalized by cloud thickness such that the top 20% of pixels from each cloud column are identified as cloud top phase. All occurrence frequencies are calculated as the number of a specific cloud type divided by all cloudy samples in that latitudinal bin. The 5° latitudinal bin also allows sufficient samples from the model simulations in each bin (Figure S3 in Supporting Information S1). The resolutions of various observation and simulation data sets follow those used in Figure 10.

Figure 12

(a–e) Occurrences frequencies of different types of cloud columns in relation to latitudes, separated into single-phase (pink), double-phase (cyan), and multi-phase (purple). (f–j) Occurrence frequencies of different types of cloud columns, separated into those that are solely liquid (red), ice (blue), and mixed (green) in the entire column or vertically stratified with more than one phase (with white hatches). The vertically stratified cloud columns are further colored in background by the liquid, mixed, and ice proportions for each latitudinal band. Cloud columns that are solely liquid, ice, or mixed phase would belong in single-phase type in (a–e). The stratified cloud column would belong in either double-phase or multi-phase types. (k–o) Occurrence frequency of the thermodynamic phase of cloud top layer at various latitudes. Rows 1–5 show high-resolution measurements, Method 1 (M1), Method 2 (M2), E3SM, and EMC² simulator results, respectively. For all sub-panels, the occurrence frequency is calculated as a specific category of cloud divided by all cloudy samples in a 5° latitudinal bin.

The results show that single-phase cloud columns have similar frequencies ranging from 0.3 to 0.9 for the three observational data sets, with a minimum frequency of around 55°–60°S. This is consistent with the fact that cloud columns that are entirely composed of liquid phase and ice phase dominate at lower and higher latitudes, respectively, and therefore, the double and multi-phase cloud columns are more frequently observed in between. The fact that high and coarse-resolution observations show similar frequencies of single-phase cloud columns suggests that the spatial scaling effect cannot solely explain the differences between observations and simulations for the phase spatial heterogeneity of cloud columns. A similar latitudinal dependence is seen in cloud top layer phase frequency, which shows dominant liquid and ice phases in lower and higher latitudes, respectively. The stratified columns further display the proportions of ice, liquid, and mixed phase within themselves for each latitude bin and show similar latitudinal dependence of the dominant phase. Similar transition from liquid-dominant to ice-dominant cloud columns with increasing latitudes is also seen in M1 and M2 spatially averaged observations. M1 (Figure 12g) shows a higher percentage of mixed phase columns as well as stratified mixed phase due to the spatial averaging of high-resolution liquid and ice pixels. M2 (Figure 12h) shows higher liquid phase frequency compared with both M1 and high-resolution observations due to signal averaging as discussed in Figure 8.

Similar analysis using E3SM results (Figure 12i) shows considerably lower ice column frequencies compared with high-resolution as well as averaged observations between 65°S and 40°S. Such underestimation of ice phase in E3SM seems to be caused by the underestimation of homogeneously distributed ice phase in the entire cloud column, rather than the ice phase seen as part of the stratified cloud columns. In fact, the frequency of stratified cloud columns in E3SM is overestimated by 0.2–0.4 at most latitudinal bins (i.e., 40°–55°S and 65°–70°S). This result indicates that the underestimation of ice phase in E3SM is less likely caused by the lack of heterogeneities in the simulated cloud vertical structure or the difficulty to represent thin layers of ice phase embedded among other phases in the E3SM simulation. This is consistent with the results shown in the spatial averaging effect using observations (Section 4.6), that is, the differences in spatial scales between observations and simulations are unlikely able to account for all the model biases seen in cloud phase distributions. Comparing EMC² with the observations, the overall transition from liquid dominance to ice dominance with increasing latitudes is still seen but with more variations. Differing from the consistent underestimation of ice phase seen in E3SM, the EMC² simulation shows that total frequencies of ice phase (i.e., including both homogeneous and heterogeneous columns) are overestimated at all temperatures, consistent with Figure 10.

While representing vertical cloud phase distribution accurately is important for GCMs, cloud top phase plays a strong role in determining local radiative budget (Gregory & Morris, 1996). Cloud top phase as a function of latitude is also examined in Figure 12. This analysis provides a different perspective from the satellite-based cloud top phase statistical analysis, especially for the horizontal scales less than 5 km that are traditionally challenging for spaceborne cloud phase retrievals (Morrison et al., 2011). Observations at various scales (Figures 12k–12m) show similar transition from liquid-topped to ice-topped cloud layers as latitudes increase, which is also similar to the phase transition seen in the entire vertical columns (Figures 12f–12h). M1 and M2 continue to show higher mixed phase and higher liquid phase frequencies at cloud top, respectively. Interestingly, E3SM cloud top phase does not show a significant underestimation of ice phase as seen in the phase distributions of entire cloud columns (Figure 12i) or individual samples (Figures 10 and 11). By contrasting the E3SM phase distributions between the entire columns (Figure 12i) and cloud top layers (Figure 12n), ice phase seems to be the main cloud top phase of stratified layers. On the other hand, the observations show mixed phase being the dominant cloud top layer in M1 (Figure 12l vs. Figure 12g) and liquid phase being the dominant cloud top layer in M2 (Figure 12m vs. Figure 12h). This result indicates that the underestimation of ice phase in E3SM mostly occurs below cloud top. In addition, both E3SM and EMC² overestimate cloud top ice frequency at high latitudes compared with coarse-scale observations. This may be caused by higher cloud top height or different vertical stratification of cloud phase in the simulations (i.e., more ice at cloud top) compared with the observations (i.e., more liquid and mixed phase at cloud top).

5 Discussions and Conclusions

The atmospheric community recognizes that observing and accurately simulating SO mixed-phase clouds is vital to our understanding of global changes in the Earth's climate (Atlas et al., 2020; Gettelman et al., 2020). By analyzing the ARM MARCUS field campaign data over the SO, we were able to develop and analyze a 5-month long observational data set of cloud phase based on ship-borne remote sensing measurements. Such a data set is valuable for evaluating GCM results since the MARCUS campaign explored a large spatial domain including multiple latitudes while providing high-resolution vertical profiles of cloud phase for each time-step (Figure 1). This ship-based data set also complements the satellite remote sensing observations over this remote region, since satellite-based cloud phase identification can be affected by cloud top beam attenuation, low altitude clutter, and missed ice detection (Mace, Protat, & Benson, 2021; Marchand et al., 2008).

Several main goals of this work include (a) developing an algorithm of classifying hydrometeor types using high-resolution observations obtained from the MARCUS campaign payload, (b) defining cloud phases at coarse resolution that can be compared with climate model simulations, and (c) quantifying the variations of cloud phase distributions with respect to temperature and latitude. All the statistical analyses of cloud properties are restricted to low-level stratiform clouds only, since this type of clouds is generally measured with higher accuracy by ground and ship-based remote sensing measurements as suggested by previous work such as Lamer et al. (2020).

Cloud phase classification has generally been more difficult using the MPL which does not provide the attenuated backscatter coefficient like other lidars (e.g., HSRL) do. In this paper, the MPL backscatter coefficient was calculated using the theoretical Rayleigh backscatter coefficients during clear sky periods in the MARCUS campaign and calibrated against the HSRL coefficient using the ARM AWARE data set (Figure A1). By using a combination of lidar, radar, MWR, and sounding measurements, an algorithm was developed to identify hydrometeor types using high-resolution observations by modifying the Shupe (2007) method (Figure 2). An additional procedure is developed to derive cloud phases based on coarse-scale observations and simulations. This paper demonstrates the ability to generate cloud phase measurements using the MPL instrument, which is a common type of lidar measurements for a number of field campaigns. It also demonstrates methods of applying data quality control and quantifying impacts due to noises that cannot be easily filtered out based on a ship-borne measurement platform.

Spatial averaging of high-resolution observations of cloud properties is necessary for a scale-aware comparison with coarse-resolution model results. Here, we compared two spatial averaging methods (M1 and M2) and evaluated their effects on cloud phase classifications based on two case studies (Figures 8 and 9) and statistical distributions of the entire campaign (Figures 10-12). We observed that the spatial averaging of high-resolution hydrometeors (M1) preserves a number of the clear-sky regions, whereas spatial averaging of instrument signals (M2) converts clear-sky regions into cloudy ones. This agrees well with previous studies which showed that the horizontal averaging of regions of alternating cloudy and clear air using satellite remote sensing data results in an overestimation of cloud cover (Cesana et al., 2016). Similarly, low vertical resolution has also been found to mix together regions of cloudy and non-cloudy air resulting in larger than normal cloud cover (Chepfer et al., 2013). Such biases in cloud cover can play a significant role in radiative fluxes over closed versus open cell marine stratocumulus cloud layers commonly found over the SO (Haynes et al., 2011). Using signal averages, M2 cloud phase results in cloud liquid dominating the backscatter average, leading to lower mixed phase frequency compared with M1 (Figure 8). M1 is also able to better identify thin cloud top liquid SLW layers compared with M2 (Figure 9). Thus, for the purpose of evaluating statistical distributions of cloud phases, M1 is recommended for future model evaluation studies.

E3SM and EMC² simulations are able to capture the observed transition from liquid-dominant to ice-dominant cloud phase as latitudes increase or as temperatures decrease. Both simulations also show relatively more similar results to M1 averaged observations compared with M2 averaged or high-resolution observations. The main model biases seen in E3SM include the underestimation of ice phase and the overestimation of liquid phase at temperatures between −40°C and 0°C for all cloudy samples. Even though previous studies have reported similar biases in E3SM using high-latitudinal observations (Yang et al., 2021; M. Zhang et al., 2020, 2022), a unique contribution of this work is to investigate the latitudinal dependence of such model bias. The underestimation of ice phase and overestimation of liquid phase are found to be more severe at higher latitudes than lower latitudes, indicating a possible larger radiative bias for underestimating the amount of absorbed shortwave radiation at higher southern latitudes. In addition, by examining the cloud type and vertical structure, it is unlikely that the underestimation of ice phase frequency in E3SM is only due to the coarse grid box scales, since E3SM is able to represent the vertical heterogeneities in cloud columns (i.e., single-, double-, or multi-phase). Interestingly, for cloud top phase, E3SM overestimates ice cloud top frequency at high latitudes, possibly due to simulations having higher cloud top height or different vertical stratification (i.e., ice on top) compared with observations. Either of these two causes may lead to subsequent radiative biases in the model, for example, overestimation of cloud top height can result in excess reflected shortwave and insufficient outgoing longwave radiation to space (Y. Zhang et al., 2019).

The impacts of spatial scales have also been investigated (Figure 11). Occurrence frequencies of cloud phase show larger difference between 1-min and 10-min averages and smaller differences between 10-min and 90-min averages, indicating a non-linear effect of horizontal spatial scales. Future studies are recommended to average high-resolution observations to no less than 12 km for comparisons with 1° climate simulations. Aside from a scale-aware comparison, it is more complex to achieve a definition-aware comparison. Two types of definitions of cloud phases are used in this work, that is, a spatial fraction approach for observations and sub-columns of EMC² simulator and a mass fraction approach for E3SM data. This limitation is difficult to reconcile in this work since accurate LWC and IWC products are currently not available for MARCUS, and separating the simulated cloud fraction into liquid and ice would involve further assumptions. Nevertheless, a recent study used in-situ aircraft observations to examine the relationship between macrophysical properties (ice spatial ratio) and microphysical properties (ice mass ratio) in mixed-phase clouds over the SO (Maciel & Diao, 2022). They found that these two properties are positively correlated and the cloud phase occurrence frequencies at various temperatures show very similar results through two types of cloud phase definitions, that is, using spatial fraction or mass fraction. Future work with possible overlapping airborne in-situ and remote-sensing observations is recommended to quantify the impacts of two types of cloud phase definitions.

When diagnosing the possible reasons for the underestimation of ice phase frequency in E3SM, such bias persists even when enhancing the HM process term by a factor of 100 (Figure 11). This result indicates that other factors may be needed to compensate for the lower amount of ice in the simulation, such as new SIP parameterizations of rime splintering and ice-ice collision (Zhao & Liu, 2021), modifications of WBF process efficiency (Bogenschutz et al., 2020), representation of dust concentrations over SO (Wu et al., 2020), and/or the inclusion of marine organic aerosols as INPs (Zhao, Liu, Burrows, et al., 2021). Super-micron sea spray aerosols which can be activated as INPs at relatively warm temperatures (−10°C) may also explain the underestimation of ice at warmer temperatures (Mitts et al., 2021).

A significant amount of ice has been seen at higher temperatures between −10°C and 0°C based on various scales of observations in the MARCUS domain (Figures 10 and 11). Interestingly, ice phase occurrence frequency in this temperature bin increases from 0.4 to 0.6 as latitudes increase from 43°S to 70°S. Recent studies showed low INP concentrations at high latitudes combined with their low effectiveness of initiating ice nucleation at temperatures higher than −10°C (McCluskey et al., 2019; Vergara-Temprado et al., 2018). Thus, it is commonly speculated that high concentrations of ice particles at higher latitudes and temperatures higher than −10°C are likely generated from SIP processes. The case study of 3 November 2017 (Figure 5) showed a typical structure of ice-topped cloud layers with liquid/drizzle-dominated cloud base regions. The cloud top temperature is higher than −5°C where INP concentrations were previously reported to be the lowest (Kanji et al., 2017), and the liquid/drizzle near cloud base is very likely caused by melting level being located at 1 km above the warm ocean surface. Other studies using satellite observations (Mace, Protat, & Benson, 2021) and airborne observations (Finlon et al., 2020; Zaremba et al., 2021) have also shown possible existence of SIP processes, leading to the high amount of ice over the SO at similar temperature conditions. Several studies mentioned that the warm ocean surfaces over the SO potentially result in stronger updrafts and more turbulence motions, which may be conducive to SIP processes such as rime splintering (Field et al., 2016; Hallett & Mossop, 1974; Korolev & Leisner, 2020; Luke et al., 2021). The fact that this study shows even higher ice phase frequency between −10°C and 0°C at the higher latitudes (60°–70°S) compared with lower latitudes (40°–50°S) indicates that the SIP processes may play a more significant role in single-layer stratiform clouds at the higher latitudes. The physical processes that lead to such an increasing role of SIP processes in the higher latitudes over cold ocean surfaces still warrant more investigation.

Finally, this paper primarily focused on low-level stratiform clouds; however, further analysis and comparison of convective and frontal systems as well as cold sector cloud properties are needed to grasp the complex cloud physics that exists over the SO. We also restricted our analysis to cloud phase and left the analysis of other cloud microphysical properties such as IWC/LWC, effective radii, and hydrometeor number concentrations for a future study. Similarly, understanding the effect of anthropogenic and biological aerosols on these quantities is also required for a complete understanding of SO climatology. Further studies of other important components of the combined cloud systems will help GCMs to better simulate future climate changes over the SO and on the global scale.

Acknowledgments

ND, MD, YS, and XL acknowledge the funding support from the Department of Energy (DOE) Atmospheric System Research (ASR) Grant DE-SC0021211. MD and ND also acknowledge support from the DOE Research Development and Partnership Pilot (RDPP) Grant DE-SC0023155. M. Diao also acknowledges funding from the DOE Lawrence Livermore National Laboratory Faculty Mini-Sabbatical Fellowship in 2021, and the support of the NCAR EOL faculty collaborator fund in 2022. Data were obtained from the Atmospheric Radiation Measurement (ARM) Program sponsored by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, Climate and Environmental Sciences Division. The authors would like to thank Damao Zhang (PNNL) for guidance in using the ARM data repository and data availability. The authors thank Drs. Shaocheng Xie, Meng Zhang, and Hsi-Yen Ma for the interpretation of E3SM results. This research used resources of the National Energy Research Scientific Computing Center; a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231 using NERSC award BER-ERCAP0019880.

Appendix A: MPL Backscatter Coefficient Calculation

The MPL particulate backscatter coefficient (β_p) was obtained from the available MPL backscatter photon count using the methodology described in Silber et al. (2018). The attenuated NRB described by Welton et al. (2002) consists of both particulate and molecular components as,

(A1)

Here, P(r) is the NRB at range r in photoelectrons/μs after dead-time, overlap, after-pulse, background, and range corrections (Campbell et al., 2002); C is the calibration constant; β_m and β_p are the molecular and particulate backscatter cross sections; and T_m and T_p are the molecular and particulate transmittance, respectively. This equation can be reduced to just the molecular backscatter components by selecting average 1-hr clear sky profiles of P(r) as,

(A2)

Penndorf (1957) generated theoretical equations for vertical profiles of β_m and T_m as a function of temperature and pressure. Using observed temperature and pressure vertical profiles from sounding measurements during MARCUS, β_m(r) and T_m(r) values were obtained for the MARCUS campaign. C(r) was then calculated using Equation A2. Following Fernald et al. (1972), β_p is then calculated as,

(A3)

Here, S is the lidar ratio defined as the ratio of the particulate extinction cross section (α_p) to β_p. When the signals have been corrected for overlap, energy output, dead time, and background subtraction, the MPL backscatter uncertainty was shown to be less than 1% in Welton and Campbell (2002). These corrections have been applied to the MPL data during the MARCUS campaign as well (Sivaraman et al., 2017). However, the uncertainty in estimated MPL β_p is more difficult to quantify. Thus, to validate the method and quantify the error, MPL β_p data are first calculated from the ARM West Antarctic Radiation Experiment (AWARE) campaign, just before the MPL was shipped to Hobart for the MARCUS campaign. The MPL β_p data are compared with co-located High Spectral Resolution Lidar (HSRL) β_p derived from different S values. The HSRL operates at the same wavelength (532 nm) as the MPL and automatically separates the particulate and the molecular backscatter (Eloranta, 2005). Figure A1 shows a comparison between HSRL β_p and β_p obtained using raw MPL NRB for a 1-hr case study on 3 January 2016 at McMurdo Station during the AWARE campaign using Equations A1–A3. MPL β_p obtained using S = 5 sr (Figures A1b and A1f) shows the closest similarity to the HSRL β_p (Figure A1a). Panels (c–e) show that higher S values (10–30) are unable to reproduce lower β_p (<1e⁻⁶ m⁻¹ sr⁻¹) values accurately between 1700 and 1720 UTC and between 0.75 and 1.5 km altitude compared with the HSRL.

Figure A1

Calculating micropulse lidar (MPL) β_p using different lidar ratio (S) values and comparing with HSRL β_p at McMurdo Station, Antarctica on 3 January 2016 during the AWARE campaign. (a) HSRL β_p is compared with (b)–(e) β_p derived from MPL using different S values. (f) The average difference between HSRL β_p and MPL β_p for different S values for the 1-hr time period. Note the magnitude of values in panel (f) is on the order of 10⁻⁶, as highlighted on the upper left corner of the panel.

Our method to derive β_p is also consistent with Silber et al. (2018) which used a single value of S = 5 sr for the same MPL instrument in AWARE campaign. Silber et al. (2018) used a constant value of S = 5 sr and showed that even though it primarily characterizes rain and horizontally oriented ice, it is a more stable value for Equation A3 since higher values of S often lead to false extinction levels. Since cloud hydrometeors such as ice and liquid primarily show β_p values greater than 10⁻⁶ m⁻¹ sr⁻¹ in our classification method, the β_p uncertainties indicated by this comparison are unlikely to have significant impacts on the cloud phase classification results.

Open Research

Data Availability Statement

All MARCUS observational data sets can be accessed from the ARM data repository at https://www.arm.gov/research/campaigns/amf2017marcus. The E3SMv1 and the EMC² simulation results are stored in open archive at Mendeley Data (https://data.mendeley.com/) using https://doi.org/10.17632/hwmnnmmjcd.1 (Shi et al., 2023).