Cloud Properties Derived From Airborne Cloud Radar Observations Collected in Three Climatic Regions

Clouds observed by the airborne High‐Performance Instrumented Airborne Platform for Environmental Research (HIAPER) Cloud Radar (HCR) were classified into twelve categories, based on their convective/stratiform nature. Dimensional and convective cloud properties were analyzed in three climatic regions: The subtropical easterlies off the coast of California, the Southern Ocean, and the tropics surrounding Central America. The convective properties of the stratocumulus clouds in the subtropical easterlies are closely related to the degree of boundary layer decoupling. In regions of strong boundary layer coupling, convectivity and updrafts in the clouds are weak and precipitation is light. In regions where the boundary layer is more decoupled, convective properties increase together with cloud top altitudes and cloud depth. Cloud properties of stratocumulus and cumulus clouds over the Southern Ocean show similarities to those observed in the subtropics, but overall they are less convective, indicating a strongly coupled boundary layer. Sea surface temperatures are much lower and the development of clouds is driven by transient synoptic conditions rather than zonal ocean temperature gradients. Clouds observed over the tropical oceans are much more convective in nature. As in the two other regions, they are mostly shallow, but clouds in regions with high sea surface temperatures have high convectivity and reflectivity values and stronger updrafts. Some of them grow to extreme depths (>14 km) and widths (>500 km). They have strong and large updraft regions and are heavily precipitating throughout their life cycle as they transition from the convective to the stratiform stage.


Introduction
Clouds are fundamental features of the weather and climate system.Understanding their properties, evolution, and formation and dissipation processes is therefore essential for their prediction.Accurate modeling of cloud evolution strongly influences the quality of fundamental weather forecast quantities such as temperature or precipitation.For climate models, cloud properties strongly influence the radiation budget, and uncertainties in the modeling of clouds are still associated with large uncertainties in the prediction of future climates (Bretherton, 2015;Zelinka et al., 2020).The feedback of clouds on the climate system depends on the altitude levels at which the clouds occur, with low clouds having a particularly strong effect on the net energy balance of the earth (e.g., Ceppi et al., 2017;Hartmann et al., 1992;Wood, 2012).Convective properties of clouds and storm systems, such as their convective or stratiform nature, determine their latent heating and mass divergence profiles (e.g., Houze, 1997).
To investigate the effect of clouds and their properties on the weather and climate system, cloud classification algorithms have been developed.For example, the International Satellite Cloud Climatology Project (ISCCP, Schiffer & Rossow, 1983;Rossow & Schiffer, 1999, Young et al., 2018) provides cloud type and property information derived from satellite information.Using radar observations, convective systems can be classified Abstract Clouds observed by the airborne High-Performance Instrumented Airborne Platform for Environmental Research (HIAPER) Cloud Radar (HCR) were classified into twelve categories, based on their convective/stratiform nature.Dimensional and convective cloud properties were analyzed in three climatic regions: The subtropical easterlies off the coast of California, the Southern Ocean, and the tropics surrounding Central America.The convective properties of the stratocumulus clouds in the subtropical easterlies are closely related to the degree of boundary layer decoupling.In regions of strong boundary layer coupling, convectivity and updrafts in the clouds are weak and precipitation is light.In regions where the boundary layer is more decoupled, convective properties increase together with cloud top altitudes and cloud depth.Cloud properties of stratocumulus and cumulus clouds over the Southern Ocean show similarities to those observed in the subtropics, but overall they are less convective, indicating a strongly coupled boundary layer.Sea surface temperatures are much lower and the development of clouds is driven by transient synoptic conditions rather than zonal ocean temperature gradients.Clouds observed over the tropical oceans are much more convective in nature.As in the two other regions, they are mostly shallow, but clouds in regions with high sea surface temperatures have high convectivity and reflectivity values and stronger updrafts.Some of them grow to extreme depths (>14 km) and widths (>500 km).They have strong and large updraft regions and are heavily precipitating throughout their life cycle as they transition from the convective to the stratiform stage.
Plain Language Summary Clouds were observed in three climatic regions by a radar mounted under the wing of an aircraft.The clouds were organized in twelve different categories.These categories are based on the nature of the clouds, and whether they developed more vertically or horizontally.Cloud properties, such as their horizontal/vertical extent, the strength of the up and downward air motion within the cloud, or the nature of their precipitation, were calculated for the different cloud categories.Comparing these cloud properties for clouds observed in the different climatic regions shows that the cloud categories are related to the different developmental stages of the clouds.In mid-latitude and subtropical regions, clouds develop more horizontally, and their nature is determined by conditions in the lowest layer of the atmosphere.The development of tropical clouds is related to the temperature of the underlying ocean, where warm ocean surface temperatures allow some clouds to grow into large storm systems.
HCR has been deployed in five field campaigns and this study focuses on the three larger projects (Figure 1a).The Cloud Systems Evolution in the Trades (CSET, Albrecht et al., 2019) study focused on the characterization of the cloud fields over the Pacific in two different regimes (Figure 1b).The stratocumulus regime is associated with the eastern flank of the subtropical anticyclone, while the fair-weather cumulus regime is located within the subtropical easterlies of the anticyclone in the warm tropics.Clouds within these two regimes, and the transition zone in-between, were observed during 16 research flights between California and Hawaii in the summer of 2015.Albrecht et al. (2019) observed a deepening boundary layer from the northeastern stratocumulus fields to the southwestern cumulus regions which was associated with an increase in sea surface temperature (SST).The boundary layer became more decoupled (American Meteorological Society, 2023) toward the southwest and the stratocumulus layer transitioned into precipitating cumulus clusters with increased cloud-top heights and a decrease in the cloud and rain fraction (Bretherton et al., 2019;Schwartz et al., 2019).
Motivated by challenges in modeling Southern Ocean clouds, these were observed during 15 research flights south of Tasmania during the Southern Ocean Clouds, Radiation, Aerosol Transport Experimental Study (SOCRATES, McFarquhar et al., 2021, Figure 1c) in the austral summer of 2018.Remote sensing instruments, such as HCR, were used to sample the vertical distribution and properties of clouds as a function of latitude.It was found that low clouds account for a large portion of model radiative biases (McFarquhar et al., 2021).An additional focus was the observation of supercooled liquid water, which is common in the pristine environment of the Southern Ocean (D'Alessandro et al., 2021;Romatschke & Vivekanandan, 2022;Schima et al., 2022).While many of the studies that are based on SOCRATES observations focus on microphysical properties of clouds (D'Alessandro et al., 2021;Efraim et al., 2020;Saliba et al., 2021;Schima et al., 2022;Twohy et al., 2021;Y. Wang et al., 2021;Zaremba et al., 2021;Zhao et al., 2020), few studies (e.g., Truong et al., 2020) have focused on the physical properties of the observed clouds.
Deep convective clouds in a tropical environment were the focus during 22 research flights in the Organization of Tropical East Pacific Convection (OTREC, Fuchs-Stone et al., 2020) field campaign in the summer of 2019.Flights were conducted over three distinct regions (Figure 1d).The Caribbean region was located east of Central America over the Southwest Caribbean Ocean.The other two regions were located west of Central America, one just off the west coast of Colombia (the Colombian coastal region), and the other one further west over the Eastern Pacific Inter-Tropical Convergence Zone (ITCZ region), extending over eight latitudinal degrees.Thermodynamic and cloud properties studied in these three regions showed distinct characteristics (Fuchs-Stone et al., 2020).The ITCZ region showed bottom-heavy vertical mass flux patterns (i.e., positive mass flux was 10.1029/2023JD039829 3 of 27 observed at low altitudes), with deeper convection at higher latitudes, while the Colombian coastal region showed deep convection with mixed vertical mass flux patterns.Convection over the Caribbean region was generally weaker.
While studies that focus on one of these field campaigns are relatively numerous, the availability of observations from all three regions collected by the same radar presents an opportunity to compare cloud characteristics and properties from these different climatic regions using the same methodology.In this study we first develop methods to identify individual clouds and classify them based on their convective and stratiform nature.We then use the classified clouds to answer the following research questions: What types of clouds occur in different climatic regions?What are their physical properties?How do their characteristics vary within one region?How do their physical properties compare between climatic regions?Section 2 describes the data and methodology used in this study.Locations of clouds and their dimensional properties in the three climatic regions are analyzed in Section 3. Convective properties and updraft regions within clouds are described in Section 4. Section 5 compares precipitation shafts between the different cloud categories.Conclusions are presented in Section 6.

Data and Methodology
In this section we describe the data and methodology used to classify clouds and calculate their properties.In Section 2.1, we describe HCR and its observations from different climatic regions.In Section 2.2, we provide a

HCR Observations in Different Regions
HCR is a 94.4 GHz (W-band) cloud radar which is deployed in an underwing pod on the National Center for Atmospheric Research (NCAR) HIAPER Gulfstream V aircraft (NCAR/EOL HCR Team, 2014; Romatschke et al., 2021;Vivekanandan et al., 2015).It is a polarimetric Doppler radar operating at a peak power of 1.6 kW with a pulse repetition frequency (PRF) of 10 kHz.The beam width is 0.73° and the radar's sensitivity is −37 dBZ at a Signal-to-Noise Ratio (SNR) of −10 dB at 1 km range.The radar data is provided at a temporal resolution of 10 Hz which, at typical aircraft speeds, translates to a ∼20 m horizontal resolution.Hence, combined with the ∼20 m range resolution, each data grid point is roughly square.HCR can be operated in zenith (upward) or nadir (downward) pointing mode, or it can scan the 180° arc (away from the fuselage) between these two pointing directions.In staring mode, the beam is stabilized every 20 ms for changes in roll and pitch angles caused by platform motion, which results in reliable Doppler velocity measurements even during aircraft ascents, descents, and turns.
HCR provides the two fundamental radar moment variables, reflectivity (DBZ) and radial Doppler velocity (VEL), among others.In addition to the moments, the radar data set includes fields such as the ERA5 reanalysis fields of temperature (TEMP), pressure (PRESS), sea surface temperature (SST), and others (European Centre for Medium-Range Weather Forecasts, 2019) interpolated to the radar grid (Romatschke et al., 2021), melting layer detections (Romatschke, 2021), stratiform/convective echo type classifications, and convectivity (Romatschke & Dixon, 2022).Information on how to access HCR data is given in the data availability statement at the end of this paper.
The flight tracks, colored by SST, for all three field campaigns are shown in Figure 1.CSET flights covered an SST range from ∼15 to 30°C (Figure 1b) and the transition zone from the stratocumulus to the cumulus regions is located around the 23°C isotherm (Albrecht et al., 2019).Flight patterns in CSET partly consisted of ferry legs at ∼6-10 km altitude to or from target areas.During the ferry legs HCR was mostly operated in nadir pointing mode.Once in the target area the aircraft descended to below-cloud altitudes, sometimes only 150 m above the sea surface.HCR was operated in zenith pointing mode during the below-cloud legs.These were followed by flight legs at altitudes just above the cloud tops where HCR pointed at nadir again.Below-and above-cloud legs were alternated and complemented with ascents and descents through the cloud layers, the so-called sawtooth patterns.During these maneuvers, HCR was pointed so as to best sample the available clouds, either below or above the aircraft.SOCRATES flights spanned an even larger SST range from just below 20°C near Tasmania to the freezing point south of Antarctica (Figure 1c).The flight patterns were similar to those from CSET with high altitude ferry legs and below-, above-and in-cloud maneuvers in specific target regions.
During OTREC, SSTs in the Caribbean region were uniform and very warm, just below 30°C (Figure 1d).The Colombian coastal region has a relatively small range of SSTs from ∼26.5 to ∼28.5°C.It was specifically chosen for its uniform SSTs (Fuchs-Stone et al., 2020).The ITCZ region spans the largest range of SSTs from just below 30°C down to ∼26°C.During OTREC the aircraft flew only at high altitudes of ∼14 km and HCR was operated in nadir pointing mode, except for take-off and landing.The OTREC flight patterns were ideal for airborne radar observations because the radar was able to sample almost the whole troposphere, which avoids the partial sampling of clouds as was often the case during the in-cloud maneuvers in CSET and SOCRATES.

Echo Classification and Convectivity
The cloud identification and cloud classification algorithms described in Sections 2.3 and 2.4 below heavily rely on the convective or stratiform classifications of the radar echoes.The convective/stratiform echo classification is one of the output fields of the ECCO-V algorithm (Romatschke & Dixon, 2022) and is provided in all HCR data sets.The second quantity provided by ECCO-V is convectivity, which is used in the cloud analysis in Sections 3 to 5. In this section, we reiterate the main concepts behind ECCO-V and the echo classification and convectivity fields it provides.Further details can be found in Romatschke and Dixon (2022).
ECCO-V utilizes the fact that stratiform radar echoes tend to be horizontally homogeneous while convective echoes present as vertically oriented features of high reflectivity and are therefore heterogeneous on the horizontal axis (Houze, 2014).The basic principle of ECCO-V is therefore to derive a measure of the horizontal heterogeneity of reflectivity.This is accomplished by calculating a quantity called reflectivity texture, where DBZ is the radar reflectivity and stdev is the standard deviation.This calculation is carried out for each radar grid point, using a line segment of a finite number of reflectivity values on each side of that grid point, on the horizontal time axis.Note that the standard deviation measures the heterogeneity while the square preserves the strength of the echo.A similar calculation is carried out over the radial velocity field which provides velocity texture.
The reflectivity and velocity texture fields are then scaled, multiplied with each other, and finally transformed with a linear transfer function into a field, which we call convectivity.The scaling and transformation coefficients are chosen in such a way that convectivity becomes a unitless quantity which ranges from 0 (all stratiform) to 1 (all convective).Thresholds on convectivity are then used to create a qualitative convective/stratiform echo classification: Grid points with convectivity values below 0.4 are classified as stratiform, those with values above 0.5 are classified as convective, and those between 0.4 and 0.5 are classified as mixed echo.Again, the scaling and transformation coefficients mentioned previously were chosen specifically to lead to thresholds that would fall close to the half-way point, to separate stratiform from convective echo.
Following the basic convective/mixed/stratiform classification, which is based on thresholds on convectivity, ECCO-V classifies the radar echoes into subcategories using temperature reanalysis data.The current study does not use this subclassification but uses an updated method, which will be described in Section 2.4.
It is important to note that the concept of convectivity was originally developed for long-wavelength radars at S-and C-band (Dixon & Romatschke, 2022).When it is applied to shorter wavelengths such as the W-band of HCR, we need to consider the effects of Mie scattering and attenuation.
Because of Mie scattering effects at W-band, reflectivity values do not significantly exceed 20 dBZ (e.g., Ellis & Vivekanandan, 2011), which results in an artificial upper limit on the reflectivity values for HCR.We need to keep this upper bound on reflectivity in mind when interpreting cloud properties that are based on this quantity.However, this upper bound does not affect the qualitative cloud classification (Section 2.4).
Attenuation artificially lowers the observed reflectivity values as the radar beam is weakened when it passes through the cloud (and to a much lesser extent when it passes through clear air).Attenuation effects accumulate and are most noticeable at the far range gates.Attenuation is particularly strong in heavy precipitation, especially when the precipitation is liquid.As with Mie scattering effects, attenuation lowers reflectivity and convectivity values, but not just at the upper end of the reflectivity range.This needs to be taken into consideration when interpreting these quantities, especially at far range gates.In addition to the quantitative reflectivity and convectivity values, attenuation can also somewhat affect the qualitative classification of radar echoes in cases when the convectivity values become so low that they fall below the 0.5 or 0.4 thresholds that differentiate convective from mixed and stratiform echo.However, as will be described in Section 2.4, the cloud categories that are derived from the qualitative classifications depend on the altitude of the top of convective features.Since thick, liquid clouds, which cause the most attenuation, were mostly sampled when HCR was pointing down, the tops of convective features are mostly observed closer to the radar where attenuation effects are weak.
Both Mie scattering and attenuation effects are strongest in the tropical storms observed during OTREC, which contained heavy liquid and frozen precipitation.They are less problematic for CSET and SOCRATES.Fortunately, HCR was almost exclusively pointing at nadir in OTREC so the effect on the cloud classification is minimal.Effects on reflectivity and convectivity in OTREC are considerable and need to be taken into account when interpreting the quantitative results.

Cloud Identification
To study and compare properties of clouds we first need to define and identify the individual clouds in the HCR observations.We developed a cloud identification algorithm which labels each cloud with a unique number.We call this label field the cloud puzzle.The input for the cloud puzzle algorithm is the FLAG field, which differentiates between cloud echo and other types of echo, and which is included in the HCR data sets (Romatschke et al., 2021).The first step is therefore to remove all echo that is identified as something other than cloud echo in the FLAG field.The basic principle is then to find contiguous echo regions within the remaining cloud echoes and label them as individual clouds.While this identification method seems simple, we need to take some special issues into consideration and properly handle them before identifying contiguous echoes to generate a higher quality cloud puzzle product.
At low altitudes, especially during CSET and SOCRATES, the HCR transmitter was sometimes disabled for short time periods to protect the receiver from potentially damaging high ocean surface echoes.This resulted in short gaps in the data.Therefore, before identifying contiguous echoes, we locate any gaps that are shorter than 10 s and remove them from the data.Taking out these short periods of missing data ensures that cloud echoes on both sides of the gap are not separated into two different cloud entities.The gaps are added back in at the end of the data processing to restore the original dimensions of the data set.An example of the cloud puzzle field is shown in Figure 2a, which shows how the cloud identification algorithm handles the short data gaps (white vertical lines) between 12:46 and 12:52 UTC.
When HCR is rotated from zenith to nadir pointing operations, or vice versa, while the aircraft is within a cloud, the same cloud will be sampled in both operating modes, but the cloud will not be contiguous.We identify such cases and join the cloud pieces together by checking whether echo is present close to the radar both before and after the antenna rotation.An example of such a case is shown in Figure 2a where HCR was first pointing nadir (12:36 to 12:38 UTC), then zenith (12:38 to just after 12:40 UTC), and then again nadir (just after 12:40 UTC), all within the same cloud.This step is particularly important during the in-cloud maneuvers of CSET and SOCRATES.
After these two pre-processing steps are completed we search for contiguous clouds.In the post-processing, we again handle two specific issues.In some cases it is desirable to break up large clouds that logically should be separate as they are only joined by a few data points.Such cases were observed in OTREC when several cumulonimbus clouds were located in such close proximity that they touched (Figure 2c).We test all contiguous cloud echoes with more than 3,000,000 data points to identify those that can be separated into logical smaller entities, 10.1029/2023JD039829 7 of 27 following a procedure similar to the "clumping" method described by Dixon and Romatschke (2022): Within each large cloud we identify large contiguous regions (of over 500,000 pixels) with reflectivities greater than −12 dBZ.Of these elevated reflectivity regions, we only keep for further consideration those that have a vertical extent that is at least 0.7 times the vertical extent of the largest elevated reflectivity region.If more than one elevated reflectivity area meets this criterion, these regions are grown by consecutively adding lower reflectivity pixels.We stop the growing process just before the regions join.All remaining pixels (with lower reflectivity) within that cloud are joined with the closest elevated reflectivity region.The cloud regions that are split apart each get their own label in the cloud puzzle field.This process breaks large clouds into logical sub-entities as shown in Figure 2c where one large cloud (17:29 to 17:53 UTC) is divided into three smaller clouds (dark red, blue, and medium red), and another large cloud (17:52 to 18:10 UTC) is divided into two smaller clouds (green and orange).
In the cloud classification described in Section 2.4 below, the quantitative convective/mixed/stratiform echo types from the ECCO-V algorithm (Section 2.2) will be used.Using these echo types for cloud classification is problematic in cases where new convective cells develop underneath aging stratiform clouds and penetrate into the stratiform regions.In such cases, it is desirable to identify the new convective cell to later classify it separately from the aging stratiform cloud.To identify penetrating convective cells we find clouds with more than 100,000 pixels and with at least half of the pixels classified as convective.In clouds meeting these criteria we loop through contiguous convective regions with 50,000 or more pixels.Newly developing convective cells generally consist of a large convective core, which is surrounded by a thin layer of mixed or stratiform echo.We therefore enlarge the convective region using morphological dilation (Gonzalez et al., 2020) to integrate this mixed/stratiform layer into the convective region.We then identify the pixels bordering the enlarged convective region and check how many of them are empty, that is, outside the cloud.If the fraction of empty pixels compared to the total number of border pixels is greater than 0.7, the enlarged convective region is labeled as a separate cloud entity in the cloud puzzle field.An example of a newly developing convective system penetrating into an existing anvil cloud (and therefore labeled as a separate entity) is shown in Figure 2b in red.
The pixel number and reflectivity thresholds that are used in the just described processing steps were determined experimentally.While they introduce some randomness into the cloud identification process, that randomness is small when we consider the much bigger effect of the flight path of the aircraft on the sampled cloud entities.

Tropospheric Regions
The individual clouds labeled in the cloud puzzle field are classified into twelve categories by a cloud classification algorithm.To facilitate this classification, we separate the troposphere vertically into three regions (which are similar to, but not identical with, the regions used by ECCO-V, Romatschke & Dixon, 2022).The boundary between the low and mid regions is the altitude of the lowest melting level that was detected in HCR observations by the HCR melting layer detection algorithm (Romatschke, 2021).Note that the melting level, that is, the level where the actual melting occurs, is usually observed below the level of the 0°C isotherm at slightly positive temperatures.The boundary between the mid region and the high region is the −25°C isotherm (from ERA5 reanalysis temperatures).
Contrary to the tropospheric levels defined in the International Cloud Atlas by the World Meteorological Organization (WMO, Cohn, 2017), which are based on altitude, we chose to base the tropospheric levels on temperature.The reasoning behind that choice is that it would be difficult to find altitudes that are meaningful for all three geographic regions.For example, the boundary between the mid and the high region separates mid-from deep convective clouds.If we choose a high altitude threshold for that boundary that works well for tropical convection, convection in high-latitude regions would never reach that threshold and we would not observe any deep convection there.If we choose a lower altitude that works for high-latitude clouds, too many convective clouds would be categorized as "deep" in the tropics.We argue that labels such as "deep" or "shallow" need to adjust to the environment.While a convective cloud reaching up to say 6 km may be the norm in the tropics, and therefore not deserving of the label "deep," such a cloud observed over the Southern Ocean would be remarkable and well deserving of the "deep" categorization.
Separating tropospheric regions based on temperature has one disadvantage at high latitudes in that the melting level can intersect with the ground, which eliminates the lower tropospheric region.To avoid that problem, we do not allow the separation boundary between the low and the mid region to fall below 2 km above the ground, and the separation boundary between the mid and the high region to fall below 4 km above ground.This way, we always retain all three regions and the associated cloud classifications described below.The altitudes of 2 and 4 km were chosen following those established by the WMO (Cohn, 2017).

Non-Precipitating Clouds
Once the three regions of the troposphere are established, the cloud classification algorithm steps through each cloud entity identified by the cloud puzzle algorithm (Section 2.3) and classifies them according to the following procedure.It first differentiates between precipitating and non-precipitating clouds.A cloud is classified as precipitating if the lowest pixel is less than 500 m above the underlying topography.Note that this definition focuses on surface and near-surface precipitation only and does not take in-cloud precipitation or virga at higher altitudes into account.The fact that fog would be classified in the precipitating categories does not present a problem in the current study because no fog was observed in any of the three field campaigns.For further sub-classification of the non-precipitating clouds, they are separated into the categories CloudLow, CloudMid, and CloudHigh based on whether the majority of the cloud pixels are located in the low, mid, or high region of the troposphere.We recognize that using the majority of pixels deviates from the WMO cloud classification, which is based on the altitude of the cloud base.However, since it is not possible to reliably identify cloud base in radar data (because it is not easily possible to distinguish cloud echo from precipitation echo) we found our definition to be more robust for the purposes of this research.Clouds in the CloudLow category can be associated with the WMO low level cloud genera of cumulus, stratocumulus, and stratus, while the CloudMid category likely also contains altocumulus and altostratus in addition to cumulus and stratocumulus.The CloudHigh category contains all cirrus-type clouds.

Precipitating Clouds
Precipitating clouds are separated into categories based on their convective or stratiform nature.Clouds with more than 50% convective echo, as determined by ECCO-V, are classified as predominantly convective (Conv).They are further classified as ConvShallow, ConvMid, and ConvDeep depending on whether the cloud top reaches the low, mid, or high region of the troposphere.ConvShallow and ConvMid clouds can represent smaller-scale precipitating cumulus-type convective clouds.They can also represent clouds that have been captured at their early stages of development, and will later grow into mid-or deep convective systems such as cumulonimbus (Cb) or even Mesoscale Convective Systems (MCSs, Houze, 2004).ConvShallow and ConvMid clouds, and in some cases even ConvDeep clouds, may not have reached their full depth at the time of sampling.
Clouds with less than 50% but more than 5% convective echo are classified as ConvStratShallow, ConvStratMid, and ConvStratDeep based on their echo top location.They can represent convective cloud systems (Cbs or MCSs) at the mature stage of their life cycle when significant stratiform echo has already formed but active convective regions are still present.ConvStratShallow and ConvStratMid clouds can also represent layered clouds with embedded convective features, such as lightly precipitating stratocumulus clouds (Houze, 2014;Wood, 2012).
They have likely reached their full depth at the time of sampling.
Clouds with less than 5% convective echo are classified as stratiform (Strat).The stratiform precipitating clouds are classified as StratLow, StratMid, and StratDeep based on whether the cloud top is located in the low, mid, or high region of the troposphere.They are nimbostratus clouds either representing convective systems at their latest, stratiform stage of their life cycle, or layered stratiform precipitating clouds that did not originate from convection (Houze, 1997).Examples of all cloud categories are shown in Figure 3.

Clouds Not Classified
There are several reasons why some clouds were not classified and excluded from the data set.In this study, the focus is on larger clouds and clouds with fewer than 5,000 pixels are not considered.While small clouds are certainly important, we decided to develop the classification and analysis methodology on larger clouds, which promise more robust statistics when cloud properties are calculated (Section 3 below).
Precipitating clouds that consist of more than 80% mixed echo cannot be identified as convective or stratiform and are therefore not classified.
Sometimes ConvStrat clouds have embedded convective cores that do not reach deeply into the cloud, even though the stratiform part of the cloud extends much higher into the atmosphere.Classifying such clouds as 10.1029/2023JD039829 9 of 27 ConvStratDeep would be misleading.ConvStrat clouds that have an embedded convective core that does not reach into the high region of the troposphere, but with more than one third of their pixels located in the high region, are therefore not classified.ConvStrat clouds with embedded convective cores that do not reach the mid-level of the troposphere are not classified if they either have echo in the high region or more than half of their pixels are located in the mid region.Clouds that are not classified for these reasons are rare.
In our analysis of cloud properties, we calculate statistics of variables such as cloud depth, or cloud base and cloud top altitude (Section 3 below).For these types of variables, cases where the aircraft was flying within a cloud are problematic as either cloud top or cloud base cannot be determined based on whether the radar was pointing at nadir or zenith.For each cloud, we therefore calculate the time fraction that the aircraft was flying within the cloud.If the in-cloud fraction is larger than 30%, clouds observed in zenith pointing mode are not classified.For nadir pointing observations the mean temperature at the highest in-cloud radar gates is calculated and if it is higher than −30°C, and the in-cloud fraction is larger than 30%, the cloud is not classified.The black cloud in Figure 3a is such an example.The −30°C threshold was chosen to preserve clouds from flight legs at very high altitudes, for which we are confident that the majority of the cloud was sampled.Such cases frequently occurred in OTREC when the aircraft was flying within an anvil cloud at ∼14 km altitude.Cloud depth and cloud top altitude will be slightly underestimated for these clouds but we do not expect this to have a significant effect on our conclusions.An example is the purple cloud in Figure 3b.It is important to keep in mind that these exclusion criteria apply to a significant portion of clouds during in-cloud maneuvers in CSET and SOCRATES.
The following cloud properties analysis therefore mostly includes clouds observed during the ferry legs of the CSET and SOCRATES flights.
Excluding clouds from the analysis for so many different reasons is unfortunate but unavoidable when using airborne observations from field campaigns.Caution needs to be applied when interpreting the findings presented in this study as they may not be generally representative of the respective region and season.While the use of large samples of spaceborne observations could mitigate the statistical sample problem, spaceborne radars lack the spatial resolution of HCR and cannot capture the detail of the present study.We therefore believe that our findings are a valuable contribution when the appropriate caution is applied for the interpretation of the results.

Number of Clouds per Category
The total numbers of clouds observed in each field campaign and cloud category are listed in Table 1.More than twice as many clouds were observed in OTREC (2,183) than in SOCRATES (952) or CSET (792).In general, the number of clouds in the low and shallow categories are higher than in the mid-and high or deep categories.An exception are the mid-and high non-precipitating clouds in OTREC, where clouds in the CloudMid category are as frequent as those in the CloudLow category and clouds in the CloudHigh category are more frequent than in any other category of all field campaigns.The high number of clouds in the shallow categories is expected, as it is common knowledge that boundary layer clouds are the most frequent cloud type globally (Norris, 1998a(Norris, , 1998b)).Specifically for SOCRATES, Mace and Protat (2018) also found a higher number of low level clouds over the Southern Ocean.In our analysis we only include cloud categories with more than 25 clouds (bold in Table 1).Clouds in CSET only occurred in significant numbers in the shallow precipitating categories and the CloudLow and CloudHigh non-precipitating categories.In SOCRATES clouds were observed in all non-precipitating categories, all shallow precipitating categories and the ConvMid and StratMid precipitating categories.Significant numbers of clouds were observed in all cloud categories except ConvStratMid during OTREC.

Calculation of Cloud Properties
For each cloud, we calculated several cloud properties, which were then averaged over the different cloud categories.Dimensional cloud properties include cloud depth, cloud width (horizontal extent), cloud top and cloud base altitude for the non-precipitating clouds.Cloud top and base are defined as the altitude of the highest and lowest cloud pixel.As already mentioned, cloud base exhibits uncertainty because of the possibility of precipitation underneath the cloud base which does not reach the ground (virga).Minimum, maximum, and mean reflectivity are related to particle size and/or density.Properties relating to the convective nature of the cloud are minimum, maximum, and mean convectivity, and mean and maximum particle velocity in updraft regions.VEL is smoothed with a 5 s median filter on the horizontal axis to remove outliers before the velocity related properties are calculated.The VEL field of each individual cloud is also slightly eroded (using morphological erosion, Gonzalez et al., 2020) to remove outliers at the edges of clouds.It is important to keep in mind that VEL is a measure of particle motion, which is a combination of air motion and particle fall speed.Environmental properties include variables such as minimum, maximum, and average temperature and pressure, melting level and −25°C level altitudes, or SSTs.Note that we use the mean to calculate the mean-type properties over individual clouds but the median when we average over all clouds within one category.
Properties of the updraft regions within clouds have been linked to phenomena such as hail growth (Foote, 1984) or the altitude of the lifting condensation level (Mulholland et al., 2021).D. Wang et al. (2020) found "weak positive correlations between core intensity and core width … and increases in draft intensity with altitude."We identify contiguous regions of upward particle motion that are larger than 0.1 km 2 and analyze their properties such as their horizontal and vertical dimensions, overall area, mean and maximum velocity, and at what altitude percentile they are located within the cloud.
For precipitating clouds, we analyze the properties of the precipitation shafts.We do not attempt to identify the top of the precipitation shaft but rather calculate the properties for the lowest 200 m of the cloud echo.Properties calculated for the precipitation shafts include shaft width (horizontal dimension), mean and maximum reflectivity and velocity as a measure of precipitation strength, and the width fraction of the precipitation shaft compared to the width of the whole cloud.Note that maximum reflectivity will be affected by attenuation (Section 2.2).Not all properties can be calculated for all clouds.For example, the −25°C level will be missing if the radar is pointing nadir and the aircraft is flying below the −25°C level.Therefore, the number of clouds included in the statistical calculations varies for different cloud properties.When calculating the statistics we need to keep in mind that HCR only samples a 2D slice through each cloud at one specific time.Sometimes that slice will be at the edge of a cloud, which can lead to classifications that differ from those that would be determined from the entire 3D entity, assuming that information were available.For example, if the aircraft flies only over the anvil of a cumulonimbus cloud it may not sample the precipitation occurring in a different part of the cloud.The cloud will be classified as CloudHigh even though complete 3D information would lead to a classification in the precipitating categories.Even if a cloud is correctly classified, the properties may not come from the most representative part of the cloud.For example, a stratiform cloud field that is sampled at its edge will have smaller horizontal and vertical dimensions, and lower reflectivity and velocity values, than if the slice were captured in the middle of the cloud.Both cloud width and precipitation shaft width also depend on the flight path of the aircraft relative to the observed clouds.For example, if a cloud street is sampled, cloud width will be different if the aircraft flies perpendicular versus parallel to the cloud street.For this study we have to assume that these effects will somewhat average out when calculating statistics over a meaningful sample size.We therefore only include cloud properties for cloud categories with more than 25 clouds (bold in Table 1) but acknowledge that there still may be some biases.
Another issue to consider is the fact that, for safety reasons, the aircraft cannot penetrate the cores of heavily convective systems.Therefore, especially in OTREC, the deep convective system slices are never from the most extreme deep convective cores but represent the closest distance to the cores that was still safe to sample.Cloud properties from deep convective systems therefore likely underestimate the convective nature of these systems.The occurrence of all stratocumulus stages over most of the CSET area could be partly explained by the observations by Bretherton et al. (2019) who found evidence of decoupling at all longitudes, even though increasing toward the southwest.They also observed more precipitation in the stratocumulus region in the afternoon than in the morning.Increased afternoon precipitation supports our hypothesis of the development of stratocumulus clouds from lightly precipitating StratShallow clouds to ConvStratShallow clouds with increased precipitation, as they develop concurrently with the diurnal heating cycle.The diurnal component is also described by Smalley   2020) observed at least some significantly higher near-surface precipitation rates than expected for stratocumulus clouds, which also points to a decoupling of the boundary layer, even in the stratocumulus region.High non-precipitating clouds are mostly observed in the northeastern and the transition regions (Figure 4e).

Locations and Dimensions of Clouds
ConvShallow clouds in CSET are mostly located in the southwestern region and therefore likely represent cumulus clouds (Figure 4d).Their occurrence over regions with high SSTs is not surprising as previous studies found a connection between cumulus development and high SSTs (Behrangi et al., 2012).
The median cloud base altitude for the CloudLow category in CSET is 0.85 km and cloud base does not vary significantly spatially (Figure 5a).The cloud top altitude of all low and shallow clouds, however, is significantly higher in the southwest than in the northeast (Figure 5b) and is particularly high for the cumulus clouds of the ConvShallow category.The increase in cloud top altitude, and hence cloud depth, is not surprising given the increase in boundary layer depth and has been observed in previous studies (Albrecht et al., 2019;Schwartz et al., 2019).It is noteworthy that even in the southwest, cloud tops do not reach the freezing level of ∼4.6 km and almost all clouds observed in CSET, except for the CloudHigh category, are therefore in the liquid phase only.
Interestingly, the horizontal extent of CSET clouds does not decrease toward the southwest (Figure 5c) as one might expect from decreasing cloudiness and the breakup of the stratocumulus layer reported in previous studies (Albrecht et al., 2019).There are several potential reasons for the lack of decreased width: (a) The stratocumulus field is not contiguous and consists of smaller-scale but frequent clouds.Since the total number of stratocumulus clouds does not decrease significantly in the transition zone (Figure 4), this reason may not seem likely, but since many clouds could not be classified during the sawtooth pattern flight legs, the distributions in Figure 4 may not reflect actual distributions.Bretherton et al. (2019) indeed saw a maximum of occurrence of clouds sampled by HCR between 130 and 135°W.(b) The stratocumulus field is contiguous but HCR only observes parts of the clouds because of sensitivity limitations, which makes it look like the clouds are broken up when in reality they are joined by optically thin cloud parts which are not observable with HCR.This reason is supported by comparisons with data from the Gulfstream V High Spectral Resolution Lidar (GV-HSRL, NCAR/EOL HSRL Team, 2012), which was also deployed on the HIAPER aircraft during CSET (Schwartz et al., 2019), and observations by the radar operators during the flights (Michael Dixon, personal communication).(c) The cloud puzzle algorithm breaks up some extremely large clouds.However, visual inspection of the output of the cloud puzzle algorithm showed almost all break-ups of clouds occurring in OTREC, and not in CSET or SOCRATES clouds.
The cloud base altitude in the CloudHigh category of CSET varies significantly with location and shows two distinct patterns (not shown).Clouds north of 30°N latitude have cloud base altitudes of ∼7-8 km while those south of 30°N latitude are at ∼9-10 km.These altitudes closely correspond to the −25°C level which is at ∼7-9 km in the north but ∼10-11 km in the south (not shown).Cloud top altitude of clouds in the CloudHigh category varies similarly, which results in little spatial variation of cloud depth.The median cloud depth is 1.7 km which is deeper on average than median cloud depth of the CloudLow category (1.0 km).

SOCRATES
In SOCRATES, ConvShallow and ConvMid clouds occur mostly in the northern latitudes with higher SSTs (Figures 1b and 6c and d).The ConvShallow clouds west of 146°E are likely the cumuli clouds described by Scott (2019).ConvStratShallow, StratShallow, and StratMid clouds are found at all latitudes.These distributions show some similarities to those of CSET (Figure 4) where Conv clouds are also confined to the highest SST regions but clouds with significant stratiform echo occur under all conditions.Of course the absolute SST values are vastly different between CSET and SOCRATES (Figure 1  The median cloud base altitude for the CloudLow category is 0.81 km (not shown), almost the same as for CSET.Interestingly, there is no organized variation of cloud base with latitude.Median cloud top altitudes are similar for all low and shallow categories, between 1.4 and 1.9 km.Cloud top altitudes for the precipitating mid-level categories are only somewhat higher and are between 2.3 and 2.5 km.All cloud top altitudes for the low and mid-level categories are right around the altitude of the separation boundary between the low and mid-levels of the troposphere (2-3 km).Separating them into low and mid categories is therefore somewhat artificial as they just happen to top out around the altitude of the separation boundary.Since variations in cloud top altitude within the precipitating shallow, mid, and non-precipitating CloudLow categories are larger than the differences in the median values between clouds of the two levels, we can analyze the spatial variation of cloud top altitude for all these categories together (Figure 7a).While there is some decrease in cloud top altitude toward the south for some of the SOCRATES flights, in general there is little correlation between cloud top altitude and latitude.The lack of a clear dependence on latitude suggests that the influence on cloud top altitude by the specific synoptic conditions of each flight is stronger than a general SST influence.As expected for regions outside the tropics and subtropics, previous studies have shown that clouds in the SOCRATES region are strongly dependent on the passing of highs and lows and the associated frontal features (Naud et al., 2020;Truong et al., 2020).Cloud top altitudes of the mid-level categories generally do not reach the separation boundary between the mid and high regions of the troposphere, which results in the absence of clouds in the deep categories in SOCRATES (Table 1).Cloud top altitudes of the non-precipitating Cloud-Mid category are higher (∼5 km) than those of the precipitating mid-level categories, which indicates that they are indeed a different category of clouds and do not turn into precipitating clouds at a later stage.The independence of the non-precipitating from the precipitating mid-level clouds agrees with findings by Truong et al. (2020) who suggested that the moisture for upper level clouds does not originate from or near the surface.The mid-level precipitating clouds therefore likely originate from low and shallow level clouds, but the non-precipitating clouds form by a different mechanism.
The medians of the cloud width for all SOCRATES cloud categories are similar to each other and vary between 6 and 14 km (Figure 7b).There are outliers of up to ∼40 km and extreme outliers of over 80 km.The extreme outliers occur more frequently in the southern latitudes.

OTREC
Clouds in OTREC show significant spatial variations (Figure 8).Precipitating clouds in the shallow categories are more frequent in the southern part of the ITCZ region (Figures 8d, 8g, and 8j) where SSTs are lower (Figure 1d).The fact that all shallow cloud categories show similar patterns suggests that they represent different stages of development of the same cloud types.Most likely ConvShallow clouds represent the earliest stage of development.They then turn into ConvStratShallow as the clouds mature and form stratiform regions, before they become completely stratiform as StratShallow clouds at their last stage before dissipation.This life cycle implies that the shallow clouds in OTREC are all of convective origin (Houze, 2014) and likely cumulus clouds, unlike the clouds in CSET and SOCRATES where the convective ConvShallow clouds were observed only in the cumulus region with warm SSTs.
Precipitating clouds in the mid-and especially the deep categories are mostly observed in the Caribbean, Columbian coastal, and northern part of the ITCZ regions where SSTs are higher (Figures 8e,8f,8h,8i,8k,8l,and 1d), which agrees with observations of a higher instability index and convective depth in the northern part of the ITCZ region by Fuchs-Stone et al. (2020).Raymond and Fuchs-Stone (2021) found an inversion in the 1-3 km range over OTREC regions with SSTs below 28°C, which decreased in strength with higher SSTs.This inversion likely

10.1029/2023JD039829
15 of 27 plays a role in why clouds over low SST regions stay shallow.They conclude that the "main effect of high sea surface temperatures is to increase the probability of convection with strong entropy divergence and top-heavy mass flux profiles."In regions with higher SSTs and deeper clouds, the lower SST limit agrees with the onset temperature for deep clouds (27°C) reported by Behrangi et al. (2012).The upper SST limit is just below the threshold of 30°C, above which they report a decrease in deep convective activity.Similarly, Li et al. (2014) found an increased occurrence of deep clouds, reaching up to 16 km in height, in regions with SSTs above 27°C on a global scale.It needs to be mentioned that the variation in cloud occurrence in the ITCZ region may also have a diurnal component as flights were carried out from south to north and started roughly around the same time each day (∼12:00 UTC or 06:00 local time).8a-8c) show a tendency to occur more frequently in the same locations as the deep precipitating clouds (Figures 8f, 8i, and 8l).

Non-precipitating clouds (Figures
The median altitude of the cloud base of the CloudLow category is 1.5 km.Cloud base is lower over cold SST regions in the southern part of the ITCZ region (Figure 9a).Cloud top altitudes of the CloudLow and all shallow precipitating categories are also lower over regions with lower SSTs, not just in the ITCZ region but also in the Colombian coastal region (Figure 9c).Cloud tops decrease so significantly over the colder SST regions that total

Convective Properties and Updrafts
Now that we have a general idea of the types of clouds that were observed in the different field campaigns, we analyze the convective properties of the different cloud types (Figure 10).Non-precipitating clouds are mostly stratiform in nature (Figure 10a).For CSET, average convectivity values for all three non-precipitating categories are extremely low for the low stratocumulus clouds (0.02) and high clouds in the northwest (0.02).SOCRATES medians are even lower and less than 0.01 for all non-precipitating clouds.Medians for the non-precipitating clouds of OTREC are slightly higher for the low (0.17) and mid-level categories (0.07) but zero for the high anvil clouds.The slightly higher convectivity values for the low clouds in OTREC, together with their small horizontal  clouds are even lower, mostly below 0.1 for CSET and SOCRATES, but OTREC clouds still have slightly higher median values (0.1-0.2).
The maximum convectivity value of each cloud indicates whether there are embedded convective features within a cloud, which is the case if convectivity reaches the convective threshold of 0.5.Maximum convectivity values (Figure 10b) are higher than average values, as expected.We first evaluate the non-precipitating clouds.In CSET and SOCRATES they are still in the stratiform range but higher for the subtropical CSET than the mid-latitude SOCRATES clouds.Medians of maximum convectivity values for non-precipitating clouds in CSET are similar for the low and high altitude levels.The SOCRATES median maximum convectivity values slightly increase with altitude level (0.00 CloudLow, 0.03 CloudMid, 0.13 CloudHigh).They exhibit the least variability with almost no non-precipitating clouds reaching the convective range.Clouds in the CloudLow category in OTREC are often in the convective range and even the median value is above 0.5 which again points to their potential for growth.Maximum convectivity for non-precipitating clouds in OTREC decreases with increasing altitude level.At the CloudMid level, median maximum convectivity decreases to 0.36 and at the CloudHigh level it is 0.00.
Moving on to the precipitating clouds we notice that maximum convectivity for all precipitating Conv and ConvStrat clouds is 1, except for ConvStratShallow clouds which are slightly below 1 (Figure 10b).Maximum convectivity values for precipitating stratiform clouds increase with increasing altitude level.Medians for the StratShallow category are just below 0.3 for CSET and SOCRATES and around 0.5 for OTREC.At the mid-level, they increase to 0.5 for SOCRATES and 1 for OTREC, and deep stratiform clouds in OTREC all have maximum convectivity values of 1.The high values in deeper stratiform clouds indicate that even though the precipitation has turned to stratiform, there are still some embedded convective features, especially in the large Cb and MCS systems of OTREC.However, these embedded features may be small since only the maximum over the whole cloud is identified.
Medians of mean and maximum reflectivities (Figures 10c and 10d) show similar patterns to convectivity, which is expected since convectivity is partly derived from reflectivity (Romatschke & Dixon, 2022).However, convectivity provides more detailed information on the convective versus stratiform nature, and especially embedded convective features.For example, it is interesting to note that while mean convectivity in OTREC is the highest at the mid-level for precipitating clouds, mean reflectivity reaches its maximum at the deep stage.As described in Section 2.2, reflectivity shows an artificial upper bound because of Mie scattering effects.
The existence and strength of updrafts within clouds are essential for cloud development.We therefore calculate the mean and maximum radial velocity of areas with upward motions (Figures 10e and 10f).As already mentioned, when analyzing radial velocity, it is important to keep in mind that it is a combination of air motion and particle fall speed.Particle fall speed depends on particle shape and size, and it is possible to observe downward radial velocities in regions with upward air motion.Upward radial velocity is only observed in regions of strong updrafts or regions with smaller, lighter, and irregularly shaped particles.We also calculate the area fraction of regions with upward radial velocity, that is, the number of pixels with upward motion divided by the total number of pixels in each cloud (Figure 10g).
Non-precipitating clouds have small medians of mean (<0.25 m s −1 ) and maximum (<2 m s −1 ) upward velocity (Figures 10e and 10f).However, when we look at the area fraction of upward motion, the stratocumulus clouds of the CloudLow category in CSET (median 0.11) and especially SOCRATES (0.2) show higher fractions than almost all other categories.Fractions in the CloudHigh category are smaller for CSET (0.7) and SOCRATES (0.8) but still relatively high compared to other cloud categories.Area fractions of non-precipitating clouds in OTREC show a different behavior.Those of the CloudLow category are insignificant, but those of the CloudHigh category are relatively high (0.18).
For the precipitating categories, Conv clouds have the highest mean and maximum upward velocities.While medians are relatively low for the ConvShallow category (<0.3 m s −1 mean, <2 m s −1 maximum), at the mid-level they increase to 0.4 and 0.7 m s −1 for the mean and 1.8 and 3.9 m s −1 for the maximum in SOCRATES and OTREC, respectively.During SOCRATES, Schima et al. (2022) observed more frozen particles in convective than stratiform clouds and the higher updraft speeds in convective clouds are likely associated with this increased ice production (Y.Wang et al., 2020).Medians for the ConvDeep clouds in OTREC are the highest of any category with 0.9 m s −1 for the means and 7.9 m s −1 for the maxima.The median for the mean upward velocities decreases to 0.5 m s −1 for the ConvStratDeep category, as the systems mature and become more stratiform.Area fractions of updraft regions are generally low for the precipitating Conv and ConvStrat shallow and mid-level categories (medians<0.04, Figure 10g).They are slightly higher for OTREC at the deep stage (∼0.08).Stratiform low and mid-level clouds have higher area fractions of upward motion for CSET and SOCRATES (0.05-0.09) but they are very low for OTREC (0.1-0.2).Only at the deep level are they slightly higher for OTREC stratiform clouds (0.05).
The fact that the convective properties (convectivity, reflectivity, and updraft speed) for CSET and SOCRATES clouds increase with the lifecycle stages proposed earlier (Section 3.1) strengthens our hypothesis that clouds in the CloudLow category develop into StratShallow clouds, and then into ConvStratShallow clouds.The higher values of convective properties in the ConvStratShallow category nicely agrees with the conceptual model of turbulent elements embedded within stratocumulus clouds at their later stages of development (Houze, 2014;Wood, 2012).To analyze the properties of updraft regions in more detail, we focus on those that are larger than 0.1 km 2 in area.We calculate the altitude of each updraft as the mean altitude of its data points.Then we calculate the altitude percentile at which each updraft region occurs within the cloud.The 100th percentile represents cloud top and the 0th percentile represents cloud base (Figure 12a).For this calculation, and all following properties of updraft regions, we treat each updraft region individually, decoupled from its original cloud.Updrafts in non-precipitating clouds observed during SOCRATES are distributed around the vertical middle of the clouds (medians at the 60th percentile for the CloudLow, 48th percentile for the CloudMid, and 50th percentile for the CloudHigh categories).Updrafts in OTREC and CSET are mostly located in the upper half of the non-precipitating clouds, with medians between the 68th and 78th percentiles.Similarly, updraft regions are mostly located in the upper half of precipitating clouds.The fact that altitude percentiles do not vary too much between the different categories means that the absolute altitudes of the updrafts above sea level increase with cloud depth: mid-level clouds have higher absolute updraft altitudes than shallow clouds and deep clouds higher updraft altitudes than mid-level clouds (not shown).Mean updraft altitudes above sea level of OTREC ConvStratDeep and StratDeep clouds are located at ∼12 km altitude.However, even the percentiles of the altitudes of updraft regions do somewhat increase with increasing cloud depth for ConvStrat and Strat clouds, and reach medians higher than the 80th altitude percentile for the deep clouds of these categories (Figure 12a).At least some of that increase in the altitude Another point to consider in this context is that the altitude percentiles of precipitating clouds include the precipitation shaft but those of the non-precipitating clouds do not.The percentiles of non-precipitating clouds are therefore more likely to represent the true percentiles within the clouds (as long as there is no precipitation underneath the cloud that evaporates before it reaches the ground), but percentiles of precipitating clouds are artificially high because of the inclusion of the precipitation shaft.
We calculate dimensional properties of the individual updraft regions.Contrary to the area fraction of regions with upward motions shown in Figure 10g, which were based on the sum of all regions with upward motion within a cloud, we now analyze the area of individual updraft regions.Since we only consider updraft regions that are larger than 0.1 km 2 , we introduce an artificial, but necessary, lower bound to the size distributions.Width and depth of updraft regions are calculated as the horizontal and vertical dimensions of a rectangular box around the region.In case of updraft regions that are tilted, the calculated width will therefore be enlarged.
Median areas of updraft regions for all low, shallow, and mid-level clouds are ∼0.2 km 2 and only at the mid-level the distribution of some categories (SOCRATES CloudMid, OTREC ConvMid) widens to include larger areas, with 75th percentiles reaching 0.4 km 2 (Figure 12b).While medians of high and deep clouds are not much larger (0.2-0.3 km 2 ), the distributions widen significantly, especially for OTREC clouds, where the 75th percentile values reach up to 0.6 km 2 , and extreme values reach up to 1.2 km 2 .Updraft regions are mostly less than 2 km wide and less than 1.5 km deep (Figures 12c and 12d).The ratio of depth divided by width (Figure 12e) shows highest values for the convective categories, which implies that relative updraft depth is important for convective development.The ConvStrat categories have smaller depth/width ratios and they are even smaller for the Strat categories.For the non-precipitating clouds, depth/width ratios are small at the low and mid-level, but higher at the high level.Depth/width ratios increase with stratocumulus development from the CloudLow to the StratShallow, and then the ConvStratShallow stages.They decrease with cumulus development from the Conv to the ConvStrat and then to the Strat stages.
When we analyze the spatial distribution of convective properties, variations in the ConvShallow category are particularly pronounced (Figure 13).In CSET, both mean convectivity and mean upward motion values are higher in the cumulus region in the southwest (Figures 13a and 13c).In OTREC, these values are higher in regions with warm SSTs (Figures 13b, 13d, and 1d).The warm SST regions are also the regions of occurrence of deep convective systems.The spatial overlap of regions with high values of convective properties and deep convective systems in OTREC suggests a connection between convectivity and strength of upward motion with the ability of shallow systems to grow deeper.

Precipitation Shafts
For precipitating clouds, we analyze the properties of the precipitation shafts.For the ConvShallow and ConvStratShallow clouds, the width fraction of the cloud that is precipitating is relatively high, with median fractions between 0.4 and 0.7 (Figure 14a).Median fractions for the StratShallow category are smaller (<0.5) but The total width of the precipitation shafts are small for shallow and mid-level clouds (medians of 5-18 km), but much larger for the deep categories in OTREC (medians of 15-45 km, Figure 14b).The distributions in the deep categories are also much broader with 75th percentiles of up to ∼140 km and extreme values of up to 250 km for the ConvStratDeep category.
Mean reflectivity and mean downward velocity in the precipitation shafts, which can be interpreted as a measure for precipitation rate, follow similar patterns (Figures 14c and 14d).Note that because of the slight erosion of the velocity field (Section 2.5), mean velocity is calculated over a smaller number of precipitation shafts.Starting with the shallow categories we note that mean reflectivity values are high (>-5 dBZ) and mean velocity values are moderate (2-3 m s −1 ) for the ConvShallow category.Both measures somewhat decrease in the ConvStratShallow category.The StratShallow category has the lowest values of all categories, with mean reflectivity values between −20 and −10 dBZ, and mean velocity values between 1 and 2 m s −1 .OTREC values are higher for the ConvStrat-Shallow and StratShallow categories than those of the other two field campaigns, which, again, points to the fact that the CSET and SOCRATES clouds are lightly precipitating stratocumulus clouds while the clouds in OTREC are of convective origin.The increase in precipitation fraction and rate from StratShallow to ConvStratShallow clouds in CSET and SOCRATES agrees with the description of drizzle falling from stratocumulus clouds at an intermediate stage of development, which increases to light rain at a later, more turbulent stage (Houze, 2014).During CSET, Schwartz et al. (2019) observed slightly increasing rain rates from the northeast to the southwest, which agrees with the transition of stratocumulus to cumulus clouds, but they found significant variability of rain rates in all of their longitude bins.Since they did not differentiate between different types of clouds, the variability is expected and the differences in reflectivity and velocity values between our different cloud categories are associated with these variabilities.
At the mid-level, mean reflectivity values for the ConvMid category are similar to those of the ConvShallow category but mean velocity values increase to ∼3 m s −1 for SOCRATES and ∼3.5 m s −1 for OTREC.Similarly, mean reflectivity values for the StratMid category are similar to those of the Strat-Shallow category, but mean velocity values increase, especially for OTREC (3.5-4.5 m s −1 ).Mean reflectivity also does not significantly increase at the deep level, but velocity values are equally high to the OTREC ConvMid category.It is interesting that mean reflectivity does not increase with system depth, but mean velocity does, especially in OTREC.High reflectivity values in radar observations are caused either by high particle concentrations or by fewer, but large particles.The fact that radial velocity increases with system depth may indicate that in the earlier, shallower stages, the high reflectivity values are caused by particle density, while at the deep stages the particles had time to coalesce, and these larger particles cause the same high reflectivity values but larger fall speeds.However, there are three other reasons that could cause an artificial upper limit to reflectivity measurements: (a) HCR cannot measure reflectivity values higher than ∼25 dBZ because of Mie scattering effects.(b) As already mentioned (Section 2.2), the HCR signal is heavily attenuated in precipitation and therefore not capable of sampling deeply into the heaviest precipitation regions.(c) As also mentioned, for safety reasons, the aircraft was not able to fly in or above the heaviest convective regions, which may also limit the sampling of heavy precipitation.This last reason also applies to velocity.It is interesting to note that the mean radial velocity does not decrease as the deep systems become more stratiform.
Looking at spatial variability, we note that for the CSET shallow cloud categories, maximum radial velocity increases from the northeast to the southwest (Figure 15a), which indicates larger rain drops falling from the cumulus than from the stratocumulus clouds.Increased precipitation in the cumulus regime agrees with findings of increased rainwater content in that region (Bretherton et al., 2019).Maximum radial velocity from the shallow OTREC clouds are lower in the low SST regions (Figures 15b and 1d).

Conclusions
Algorithms were developed to identify clouds sampled by HCR in three field campaigns in three climatic regions: CSET in the stratocumulus and the fair-weather cumulus regimes within the subtropical easterlies over the northern Pacific, SOCRATES over the Southern Ocean, and OTREC over the Caribbean ocean, near the Colombian coast, and the East Pacific.The clouds were classified into twelve categories which differentiate between precipitating and non-precipitating clouds of different altitudes, and different developmental stages based on the clouds' convective and stratiform nature.Dimensional and convective cloud properties, and characteristics of updraft regions and precipitation shafts, were calculated and averaged for each cloud category.
When interpreting the results of this study it is important to keep some caveats in mind.HCR is a W-band radar which experiences significant attenuation, especially in liquid precipitation.Attenuation, together with Mie scattering effects, lead to artificially low reflectivity values which needs to be taken into account when interpreting quantitative reflectivity information.It is also important to recognize that the clouds sampled in each region are not necessarily representative of all clouds in this region and season.For example, SOCRATES focused on the sampling of cold sector clouds while during OTREC the aircraft had to avoid the most extreme convective cores for safety reasons.In-cloud maneuvers of the aircraft resulted in incomplete sampling of some clouds which then had to be removed from the analysis.
In CSET, almost all clouds were low or shallow, that is, their cloud tops stayed below the freezing level.A few high level clouds were observed in the northeastern stratocumulus region.Two distinct cloud species dominated in CSET.In the northeast, stratocumulus clouds occur over low SST regions.They develop as non-precipitating clouds in the CloudLow category, then transition into StratShallow, and finally ConvStratShallow precipitating clouds as the boundary layer decouples.While the decoupling is likely stronger in the transition zone to the cumulus region in the southwest, ConvStratShallow clouds are also observed in the northwest, which indicates at least some decoupling, as suggested previously (Bretherton et al., 2019).The early stages of the stratocumulus clouds are very stratiform in nature, but the convective properties, such as convectivity, upward particle motion, and the depth/width ratio of updraft regions increase with the development to StratShallow and ConvStratShallow clouds.Precipitation also increases.
The second cloud species are cumulus clouds that occur in the southwest of the CSET region, in the trade cumulus regime.In our classification they are characterized as ConvShallow clouds with stronger convective signatures and stronger precipitation than the stratocumulus clouds.Cloud top altitudes increase significantly from the northeastern stratocumulus to the southwestern cumulus region as the boundary layer deepens, but cloud base altitude increases only slightly, leading to an increase in overall cloud depth.
Similar to CSET, the dominant cloud species in SOCRATES was also stratocumulus.Cloud tops of stratocumulus clouds in SOCRATES are mostly close to the freezing level, which means that most of them are classified as low and shallow but some extend into the mid-level.The cloud properties of the SOCRATES stratocumulus clouds show similarities to the CSET stratocumulus clouds, as has been noted in previous studies (Mace et al., 2021).They seem to undergo the same life cycle, from CloudLow to StratShallow to ConvStratShallow, but fewer clouds reach the ConvStratShallow stage than in CSET, which points to less decoupling of the boundary layer.The cloud property values are also similar to those of CSET, both qualitatively and quantitatively, which might be surprising, given the vastly different SST ranges of the two regions.However, contrary to CSET, there is almost no latitude dependence as SOCRATES cloud development is driven by transient synoptic conditions.
Convective clouds in the ConvShallow category are sparse in SOCRATES and are associated with warm synoptic conditions.They occasionally grow into ConvMid clouds, but almost never into ConvDeep clouds.As in CSET, they have stronger convective characteristics with higher convectivity and updraft speed values, and they produce heavier precipitation than the stratocumulus clouds.
Non-precipitating mid-level clouds in SOCRATES are distinct from the shallow and low clouds and likely originate from advected higher-level moisture rather than near-surface moisture (Truong et al., 2020).High level clouds occur mostly in the southernmost latitudes, which agrees with previous findings (Truong et al., 2020).
In OTREC, most precipitating clouds are also shallow.In regions with lower SSTs, particularly in the southern part of the ITCZ region, they are capped by an inversion (Raymond & Fuchs-Stone, 2021)  Non-precipitating clouds in OTREC are particularly frequent at the high level, and are likely associated with anvil clouds.Low-and mid-level non-precipitating cumulus clouds are also numerous.
OTREC clouds are generally more convective than those of CSET and SOCRATES and the values of convective properties increase with cloud depth.They have higher convectivity values, higher upward velocity speeds, larger, and particularly deep updraft regions, especially at the deep stage.The convective property values decrease as the clouds become more stratiform at later stages of their development.Precipitation shafts are also larger and precipitation is particularly strong in the OTREC Cbs and MCSs.
An analysis of updraft regions within clouds of all three field campaigns reveals that they are larger in deeper clouds, in both the horizontal and vertical dimensions.However, the depth/width ratio of the updraft regions is larger for convective than stratiform precipitating clouds of all sizes.Non-precipitating clouds have smaller depth/width ratios than precipitating clouds.
Precipitation intensity increases with cloud depth but is lighter in stratiform than in convective clouds, at least at the shallow level.At the deep level, the fraction of the cloud that is precipitating becomes smaller as non-precipitating anvils form.
We hypothesize that our newly developed cloud classification scheme represents clouds at different stages of development.Stratocumulus clouds, observed in CSET and SOCRATES, seem to develop from the CloudLow to the StratShallow and then the ConvStratShallow categories, but cumulus clouds observed in all three field campaigns seem to develop from Conv to ConvStrat to Strat clouds, both at the shallow and the deep levels.Even though these different cloud types may appear in the same cloud classification category, they show distinct differences in their cloud properties, with cumulus clouds being more convective at all stages of development.

Figure 1 .
Figure 1.(a) Location of field campaigns.Flight tracks for (b) CSET, (c) SOCRATES, and (d) OTREC.Colors show ERA5 sea surface temperatures (gray when over land).Gaps represent times when HCR was not transmitting.Coastlines are shown in black.

Figure 2 .
Figure 2. Examples of how the cloud puzzle algorithm handles (a) short data gaps and changes in radar pointing directions, (b) penetration of convective echo tops into stratiform anvils, and (c) the breakup of extremely large clouds into smaller parts.All examples are from the OTREC field campaign.Each color represents an individual cloud identified in the cloud puzzle algorithm.

Figure 3 .
Figure 3. Clouds sampled during OTREC and their classification categories.

Figure 4 .
Figure 4. Number of CSET radar profiles normalized by total number of profiles and flight hour for the (a) CloudLow, (b) StratShallow, (c) ConvStratShallow, (d) ConvShallow, and (e) CloudHigh categories.Only grid cells with more than 10 flight minutes are plotted.Grid cells with more than 10 flight minutes but no cloud profiles are shaded in gray.
and L'Ecuyer (2015), who found that most of the precipitation in stratocumulus regions west of continents falls during the night.Sarkar et al. (

Figure 5 .
Figure 5. CSET.(a) Cloud base altitude of the CloudLow category, (b) cloud top altitude, and (c) cloud width of all shallow and low cloud categories.
). Non-precipitating clouds in the CloudLow and particularly the CloudHigh categories have a preference for the colder SST regions in the south, again showing a similarity to the CSET CloudHigh category.Clouds in the CloudMid category are found at all latitudes.As described by McFarquhar et al. (2021), the objective of most SOCRATES flights was to sample clouds in the cold sector of cyclones.That, together with the difficulty of classifying clouds during sawtooth flight patterns, warrants caution when interpreting spatial cloud distributions.

Figure 7 .
Figure 7. SOCRATES.(a) Cloud top altitude of all shallow and mid-level precipitating categories and the CloudLow non-precipitating category, and (b) cloud width of all cloud categories.

Figure 9 .
Figure 9. OTREC.(a) Cloud base altitude and (b) cloud depth of clouds in the CloudLow category.(c) Cloud top altitudes of all shallow and low cloud categories.(d) Boxplot of cloud widths.Box plot description: On each box, the central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively.The upper (lower) whisker extends to the most extreme data point smaller (larger) than the 75th (25th) percentile plus (minus) 1.5 times the interquartile range.The numbers at the bottom are the sample sizes.Note the separate right-hand y-axis for the deep/high categories in (d).

Figure 10 .
Figure 10.(a) Mean and (b) maximum convectivity.(c) Mean and (d) maximum reflectivity.(e) Mean and (f) maximum upward velocity in areas with upward motion.(g) Area fraction of upward motion.CD, CSD, and SD stand for ConvDeep, ConvStratDeep, and StratDeep.See Figure 9 for box plot description.
median of the maxima is equally high for the ConvStratDeep category (7.8 m s −1 ) as for the ConvDeep category.The decrease in mean upward velocity from the Conv to the ConvStrat stage while the maxima remain the same, is also observed for the ConvStratShallow category, where mean velocity also decreases compared to the ConvShallow category but maximum values stay almost the same.High maximum values of upward velocity at the mature stage of cloud development indicate that at least a few strong updraft regions are still present at this stage.Two reasons could be associated with the decline of the mean values: (a) previously strong updraft regions weaken and (b) weaker updraft regions form in the newly formed stratiform regions, similar to those in the non-precipitating clouds (see above).
To calculate the number of updraft regions per cloud, contiguous updraft regions were identified.On average, between 5 and 35 such updraft regions are observed in non-precipitating clouds (not shown).More updraft regions per cloud are observed in the CloudHigh category than in the lower two categories.Shallow and mid-level precipitating clouds have mostly fewer than 40 updraft regions per cloud.Only ConvStratMid clouds have a higher number of updrafts, just over 100 (not shown).Deep systems in OTREC have much larger numbers of updrafts per cloud.While ConvDeep clouds have an average of 233 updrafts, the number increases to 2,214 in ConvStrat-Deep Clouds and then slightly decreases to 1,456 in StratDeep OTREC clouds.It is interesting to note that StratShallow clouds with high numbers of updraft regions in SOCRATES and OTREC seem to be concentrated in regions with colder SSTs in both field campaigns (Figures 11, 1c and 1d).

Figure 14 .
Figure 14.(a) Width fraction of precipitation shafts, (b) width of precipitation shafts, (c) mean reflectivity of precipitation shafts, and (d) mean velocity of precipitation shafts.CD, CSD, and SD stand for ConvDeep, ConvStratDeep, and StratDeep.See Figure 9 for box plot description.

Figure 15 .
Figure 15.Maximum downward velocity in precipitation shafts from the shallow cloud categories for (a) CSET and (b) OTREC.
and their cloud top altitude decreases with SST toward the south.The clouds likely initiate as ConvShallow clouds, then transition into ConvStratShallow clouds at the mature stage of their life cycle, before they turn into StratShallow clouds in the late stages.In regions with high SSTs, in the Caribbean, Colombian coastal, and northern part of the ITCZ region, some of the ConvShallow clouds grow into ConvMid, and then ConvDeep clouds, as they form Cbs and sometimes MCSs.The ConvDeep clouds then develop stratiform regions and become ConvStratDeep clouds, and eventually transition into mostly stratiform StratDeep clouds.Deep clouds in OTREC can become several hundred kilometers wide and are orders of magnitudes wider than any other cloud type observed in all three field campaigns.OTREC clouds stay either shallow or grow very deep.Mature clouds at the mid-levels are rare.
Note.Categories with more than 25 clouds (bold) are included in our analysis.Number and Percentage of Clouds Observed in Each Category per Field Campaign