A climatology of trade-wind cumulus cold pools and their link to mesoscale cloud organization

mesoscale cloud organization Raphaela Vogel1, Heike Konow2, Hauke Schulz3, and Paquita Zuidema4 1LMD/IPSL, Sorbonne Université, CNRS, Paris, France 2Meteorological Institute, Universität Hamburg, Hamburg, Germany 3Max Planck Institute for Meteorology, Hamburg, Germany 4Rosenstiel School of Marine and Atmospheric Science, University of Miami, Miami, FL, USA Correspondence: Raphaela Vogel (raphaela.vogel@lmd.ipsl.fr)

Vertical profiles of hydrometeors (i.e. cloud and rain droplets) at approx. 30 m vertical resolution are derived from two 35.5 GHz (Ka-Band) Doppler cloud radars. Radar returns with an equivalent radar reflectivity lower than −50 dBZ are removed to eliminate signal from sea salt aerosol (Klingebiel et al., 2019). To identify individual 2D cloud entities, a cloud segmentation algorithm is applied (Konow, 2020). Radar reflectivity is converted to a binary mask and morphological closing is applied. The 90 resulting mask is used to segment cloud entities with connected components analysis with 8-connectivity. A minimum cloud size of four pixels is applied, everything smaller than four pixels is discarded as clutter. For the resulting cloud entities, the overall cloud-base height (cbh ID ), overall cloud-top height and the cloud length (i.e. the duration times the wind extrapolated from the surface to cloud base assuming a power law) are determined. To focus on clouds connected to the sub-cloud and trade-wind layer, cloud entities with a cbh ID >4 km are excluded. 95 From the remaining clouds, we derive timeseries of the hydrometeor fraction, the lowest cloud-base height (CBH) and the highest cloud-top height (CT H) for every radar profile. The cloud cover is further split up into contributions from precipitating cloud segments if CBH ≤ 300 m (CC prcp ), from cloudiness near the lifting-condensation level (CC lcl ; 300 m < CBH ≤ 1 km), and from cloudiness aloft (CC aloft ; 1 km < CBH ≤ 4 km). The latter two categories were also used in many previous studies (e.g Nuijens et al., 2014;Vial et al., 2019). A given radar profile can only count to one of the three categories, such that 100 e.g. a 2 km deep cloud with a CBH<300 m will only be counted in the CC prcp category. Note that the above classification into the different CBH categories does not account for the information of the cloud entity and all radar profiles are classified independently. A similar analysis accounting for the cloud entity by classifying cloud cover contributions of different cloud types by their cbh ID is shown in Appendix A.
From the cloud radar we also derive a deep-cloud mask, which is set to 1 if a radar signal between 4.5-8 km is detected. 105 With this deep-cloud mask, periods of active deep convection reaching above the melting level can be omitted, while periods with only cirrus-clouds are retained.

Doppler lidar
The vertical velocity in the sub-cloud layer is measured by two Halo Photonics Streamline Pro Doppler wind lidar systems at 30 m vertical resolution. The Doppler lidars measure vertical velocities of up to ±20 m s −1 with a 1500 nm laser in altitudes 110 from about 50 m to 1 km, depending on the atmospheric conditions and the aerosol loading. The precision is <20 cm s −1 for a signal-to-noise ratio (SNR) of -17 dB. Measurements with a SNR smaller than -18.3 dB are discarded. Data from the first system that was operated in vertically-pointing mode with a temporal resolution of 1.3 s is used from March 2016 to October 2019. A second system is operated in horizontally-scanning mode since February 2019 and has a temporal resolution of 3 s, with 2 out of 7 profiles measured in vertically-pointing mode. Vertical data from this second lidar is used from November 2019 115 to March 2021.
We derive both the average vertical velocity in the sub-cloud layer (SCL) as the mean over 15 range gates from 75-495 m (w SCL ), and the vertical velocity near the sub-cloud layer top at 450 m as the mean over the four range gates from 405-495 m (w 450 ).
The neural network based on the Retinanet algorithm (Lin et al., 2017) has been initially trained on and applied to visible images in Rasp et al. (2020), and later retrained and applied to infrared images by Schulz et al. (2021). The use of infrared images also allows study of the diurnal cycle of the mesoscale organization (Vial et al., 2021). The classifications of the neural network are rectangles of various sizes that belong to either the Sugar, Gravel, Flowers or Fish pattern. We select every classified rectangle that overlaps with the BCO location. Periods without a classification are labelled as 'No'. For conditioning 130 on cold pools, the 30-min data is downscaled to 1-min by using a given pattern for the 15 min before to after the classification time. If a given pattern is present for more than 75% of the duration of a cold pool, the pattern is attributed to this cold pool.
At any given time, multiple rectangles of different sizes of the same and of different patterns can occur. Multiple rectangles of the same pattern are combined and counted only once, while multiple rectangles of different patterns are counted separately. This leads to timesteps being classified e.g. as both Gravel and Flowers. Excluding situations with multiple patterns 135 only marginally influences the results, but reduces the sample size considerably (as previously noted in Vial et al., 2021).
Ambiguities in the classification can be physical-for example due to regime transitions or similarities between patterns-or related to ambiguities introduced to the neural network by disagreement in the human classifications. The occurrence of multiple patterns can be reduced if a stricter threshold is used for the agreement score representing the confidence of the neural network prediction (here set to 0.4 as in Schulz et al., 2021;Vial et al., 2021), but this again reduces the sample size.

Cold-pool detection algorithm
We detect cold pools by identifying abrupt drops in the surface temperature timeseries following Vogel (2017). We first filter the 1-min averaged temperature timeseries with an 11-minute running average. We then classify all temperature drops δT = T fil (t) − T fil (t − 1) < −0.05 K (per minute) in the filtered timeseries as a cold-pool candidate (see Figure 1 for an illustration).
For every candidate cold pool, we detect the time of the cold-pool front onset (t max ), the time of the minimum temperature 145 (t min ), and the end of the cold pool (t end ) as follows: 1. t max : the onset of the cold-pool front t max is defined as the last instance of δT > 0 K within 20 min before the initial abrupt temperature drop. If the temperature is falling continuously in this period, t max is chosen as the time of the maximum temperature (that is, 20 min before the abrupt temperature drop). We refer to the smoothed temperature at t max as T max . 2. t min : the time of the minimum filtered temperature T min marks the end of the cold-pool front and is identified as the minimum of contiguous temperature minima. Subsequent candidate cold pools with δT < −0.05 K occurring within 20 min of the previous minimum are combined if the temperature does not rise by more than 0.5 K above the previous minimum in between.
3. t end : the end of a cold pool is defined either as the minimum of (a) the time when the filtered temperature first exceeds 155 its minimum by ∆T /e, where ∆T = T max − T min , or (b) the onset of the next cold pool. If using condition (a) or (b) leads to any temperature between t min and t end to be smaller than T min − 0.1 K, then t end is defined as (c) the time when the filtered temperature first decreases again after increasing for some time following t min . Cold pools with t end defined by (a) are referred to as recovered.
The period between t max and t min is referred to as the cold-pool front, and the period between t min and t end as the cold-pool  Our cold-pool detection algorithm is similar to the one used by de Szoeke et al. (2017), but with the important modification that we only identify cold pools for situations with abrupt temperature drops. With our algorithm we thus both filter out turbulent fluctuations and advective or diurnal patterns of temperature variability. The threshold of δT < −0.05 K is subjectively chosen based on visual impression and represents distinct variations in temperature. For an 11-min averaging window, a δT of 165 −0.05 K corresponds to about 2% of the data. Figure 2 shows example cold pools for all patterns and illustrates the workings of the algorithm. In next subsection we briefly discuss the strengths and weaknesses of the algorithm based on these examples.

Example cases
Timeseries of example cold-pool days along with corresponding satellite images are shown for every pattern in Figure 2. The example cases highlight how well the detection algorithm works in these diverse situations. Abrupt strong temperature drops 170 are reliably detected, successive fronts sensibly combined into one single cold pool, and even the 6 h long cold pool with frontal character on the Fish day is correctly identified.
The example cases also indicate some challenges of the cold-pool identification. Although they look like cold pools, some temperature drops on the Gravel and Sugar day are not identified as cold pools because they are either not abrupt enough (δT > −0.05 K) or not strong enough (∆T > −0.4 K). The difficulty in defining the end of the cold-pool wake is illustrated 175 in the Fish case: the cold pool starting shortly before 16 LT lasts until well after 18 LT, but the temperature drop near 17 LT causes a premature end of the cold pool, as such a temperature drop could also be caused by the daily cycle in temperature.
The cold-pool end definition could be improved by an additional rain or downdraft requirement, to more robustly distinguish between cold-pool activity and other processes. Because most analyses and diagnostics computed in this study focus entirely on the cold-pool front (see next section), not fully representing the wake of rare long-lasting cold pools is a minor issue and 180 only influences the overall cold-pool fraction and the duration statistics.
As mentioned in Section 2.2, the organization pattern definition is not unambiguous and also among the example days shown in Figure 2 some cold pools pertain to multiple patterns. For the Flowers case, the 2 h at the beginning and end of the period shown are also classified respectively as Gravel and Fish. In the Sugar case, only the period between 9-16 LT is exclusively classified as Sugar, while the periods before and after are also partly classified as Gravel. Most surprisingly, the textbook Gravel 185 day is also entirely classified as Flowers, and also setting a stricter agreement score of 0.5 leaves half of the day co-classified as Flowers. This indicates that distinguishing Gravel from Flowers can be particularly challenging (as also shown in Vial et al., 2021). The Fish day is very confidently classified and no other pattern is detected for the entire day.

Selection criteria and diagnostics
For the subsequent analyses, we apply a number of selection criteria to make the comparison of cold pools more robust. Namely,

190
we only consider cold pools with ∆T < −0.4 K and less than two missing values in the filtered temperature timeseries during the entire cold-pool duration (set all with 9234 cold pools). For the analyses of the cold-pool properties we further apply a criterion of no non-recovered cold pool in the hour prior to the cold-pool onset (set noprev with 8772 cold pools), which selects cold pools moving into an initially undisturbed atmosphere that is not modified by previous convection. For most of the analyses we also focus on the dry winter regime from December-April (set noprevWI with 3889 cold pools), which is characterized by 195 steady easterlies, subsiding large-scale motion in the free troposphere and the predominance of shallow trade-wind convection (Brueck et al., 2015).
As shown in the brackets, all these selection criteria reduce the cold-pool sample size considerably. They represent a tradeoff between assuring a robust and unbiased sample to address our research questions, while not being unnecessarily strict and removing too many cold pools. The selection criteria are thus somewhat subjective and also differ among studies. For example,  Chandra et al. (2018) used the criterion of no rain in the hour prior to the cold-pool onset to select cold pools unmodified by previous convection, whereas we achieve the same goal with the criterion of no non-recovered cold pool in the prior hour, which excludes about 2500 less cold pools in our case. Instead of focusing on the winter regime, we could have also set a criterion based on the cloud-top height to focus on trade cumulus cold pools. However, as this would restrict the analysis to periods when the radar is running, and-as we are relying on single-site measurements-the parent convection might not move 205 over the BCO in its entirety, we would likely exclude too many cold pools with a CT H criterion, without even being sure that periods of deep convection are really excluded. Despite the rather strict criteria applied here, the long timeseries leads to a much larger number of cold pools analysed than in previous studies.
Another potential sampling issue regarding the single-site measurements is that it is not clear at which stage of its lifecycle we sample the cold pool, and where we sample it with respect to its center. Assuming isotropic wind variations around the 210 cold-pool center, which in case of little wind shear is a good approximation (Touzè-Peiffer et al., 2021), the change in wind direction from the mean direction prior to the cold-pool onset could give a hint as to the location relative to the cold-pool center.
However, due to our large sample size potential biases are likely to be small.
If not mentioned differently, the cold-pool diagnostics are computed either as the minimum difference (∆X min ) or maximum difference (∆X max ) of a variable X between its value at t max and the values between t max + 1 and min(t end , t min + 20).

215
Similarly, X mean or X max are the mean or maximum of variable X over the same analysis period (indicated in dark red in Figure 1). For the Doppler lidar vertical velocities, we diagnose w maxSCL (w max450 ) as the maximum w SCL (w 450 ) in the first half of the front (including the last 10 min before t max ), and w minSCL as the minimum w SCL in the second half of the front (including the first 10 min after t min ). Unless otherwise stated, the surface meteorology diagnostics are computed from the 11-min filtered timeseries.

220
Along with most diagnostics and composites we show the standard error (SE), which measures how well the median or mean of a given sample can be estimated. The SE of the median is computed as IQR/ √ n, where IQR represents the inter-quartile range and n the sample size, and the SE of the mean as σ/ √ n, where σ is the standard deviation. As not all instruments were running all the time, some diagnostics are only available for a subset of the cold pools and the sample size is adjusted accordingly when computing the SE.

Cold-pool climatology
In this section we present the climatology of trade cumulus cold pools detected at BCO for the winter seasons of the years 2011-2021. The first subsection presents general statistics, followed by a discussion of the composite temporal structure of the cold pools in Section 3.2. The daily cycle of cold-pool statistics is shown in Section 3.3. While our focus lies on the winter regime, Appendix B also briefly discusses the seasonal cycle of the cold-pool statistics.

General statistics
In total we detect 3889 cold pools that meet the criteria of ∆T < −0.4 K and less than two missing values in T fil in the winter seasons considered. We find that cold pools are very frequent at BCO and on 73% of days at least one cold pool is detected.
The BCO is on average affected by cold pools during 7.8% of the day (i.e. 112 min) and by a cold-pool front during 4.4% of the day, with the medians being about one-third smaller than the means mentioned The mean cold-pool fraction of 8.6%   Table 1 presents statistics of the most important cold-pool properties for the set of winter cold pools with no non-recovered cold pool in the prior hour (noprevWI). It shows that 50% of the cold pools have a temperature drop exceeding 0.9 K across 240 the front (the unfiltered temperature drop is 0.3 K stronger), a ∆q max exceeding 0.2 g kg −1 and a ∆q min below −0.43 g kg −1 , decreases in θ e and θ v exceeding −2.1 K and −0.96 K, respectively, and a ∆U max larger than 1.14 m s −1 . The median rain intensity measured by the MRR is 0.9 mm h −1 . Furthermore, 50% of the cold pools are associated with a maximum cloud-top height exceeding 3 km, and w maxSCL and w minSCL of 0.9 m s −1 and −0.55 m s −1 near the onset and end of the front, respectively.
The average cold-pool duration is 33 min, of which a bit more than half of the time pertains to the front. Multiplying the duration 245 with the surface wind speed yields a median cold-pool length larger than 13.3 km.
The IQR shows that all these medians are associated with substantial variability, especially for the humidity and rain variables. However, focusing on the winter regime generally reduces the IQR of the diagnostics compared to all seasons (not shown), suggesting that this criterion indeed results in a more homogeneous cold-pool sample representative of the tradecumulus regime. The median duration of 33 min and length of about 13.3 km of the cold pools may seem small compared to 250 satellite imagery, in which mesoscale cold-pool arcs can easily span 100 km. Also the largest 2% of cold pools are hardly larger than 40 km. The smaller cold-pool sizes found here are likely due to the algorithm sampling mostly the edge of the cold pools, and due to the challenges of defining the cold-pool end purely based on the surface temperature timeseries (see discussion in Section 2.4). Table 1 also compares the median±IQR of the 25% strongest and weakest cold pools in terms of ∆T . The strongest cold 255 pools last longer, follow each other more quickly (lower ∆t nextcp ), and are associated with deeper clouds, more rain, stronger downdrafts, humidity drops and wind gusts, and larger positive vertical velocities at the beginning of the front compared to weaker cold pools. Similar but slightly smaller differences between stronger and weaker cold pools are found when comparing cold pools associated with the 25% strongest versus weakest downdrafts or the 25% deepest versus shallowest CT H max (not shown). The downdraft strength w minSCL is the diagnostic that correlates best with ∆T (R 2 =0.23), and together with the front 260 duration it explains a lot of the variability in ∆T for the noprevWI set (multiple R 2 =0.49). The 25 and 75% quartiles of w minSCL also distinguish the rain diagnostics best.
That CT H max also distinguishes the cold-pool properties very well indicates that the parent convection triggering the cold pool is sampled well by the single-point measurements. The CT H usually scales with the precipitation amount for trade cumuli (Byers and Hall, 1955;Kubar et al., 2009;Nuijens et al., 2009), so other factors like the environmental humidity do not seem 265 to influence rain evaporation and downdraft strength much further. We also compared the properties of the 25% driest and moistest cold pools in terms of ∆q min (not shown), which does not strongly distinguish other cold-pool properties, not even the RR that was shown to be particularly related to ∆q min in the literature (Barnes and Garstang, 1982). The specific humidity signal is generally also very variable and the response to the cold-pool onset hard to define in one diagnostic, as will be shown next.  (Young et al., 1995;de Szoeke et al., 2017;Zuidema et al., 2017). The temperature of the composite-mean cold pool, after increasing slightly before t max , decreases rapidly in the front and recovers by ∆T /e within 16 min after t min . The temperature remains about 0.5 K below T max in the hour after the frontal passage. The temperature drop in the front of the 25% strongest cold pools is by definition stronger, but with a mean tendency of −0.070 K min −1 also more than twice as abrupt compared to the weakest cold pools. The strongest cold pools also take longer to recover than the weakest.

280
The temporal structure of the specific humidity response is intriguing. The composite-mean humidity starts to increase already 8 min before t max and increases by about 0.2 g kg −1 until t max . In the first quarter of the front, the humidity increases by another 0.2 g kg −1 , before it drops to its minimum of −0.25 g kg −1 at t min , which is hardly lower than the pre-front value.
The humidity recovers much more quickly than the temperature and remains slightly elevated compared to its pre-front value https://doi.org/10.5194/acp-2021-420 Preprint. Discussion started: 1 June 2021 c Author(s) 2021. CC BY 4.0 License.
very interesting that the prefrontal peak shows this little variance interesting that the stronger cold pools "live" in a rather moisture than drier environment.
interesting, that no increase observed before t_max in the hour after. The fast humidity recovery might be due to the trapping of surface moisture fluxes in the shallow mixed layer typically associated with cold pools (Touzè-Peiffer et al., 2021). Another reason might be continued evaporation of precipitation, which would cool and moisten the air in the cold-pool wake and thus speed up the humidity recovery but slow down the temperature recovery.
The specific humidity response of the strongest cold pools only differs significantly from the weakest cold pools at t min , with the humidity drop at t min being about −0.4 g kg −1 and thus about twice stronger than the drop for the weakest cold pools.

290
If the entire set of cold pools including the summer season with deeper convection is used, the strongest cold pools have a significantly weaker positive humidity anomaly at the beginning of the front, and a significantly faster and stronger humidity reduction at t min compared to the weakest cold pools (see Figure B1c-d). As discussed by de Szoeke et al. (2017), the humidity increase just before t max might be mostly due to the increasing saturation specific humidity associated with the increasing temperature before t max (as seen by the relative humidity anomaly in panel d being slightly below zero), and as such likely 295 also related to the way we identify T max .
The temporal structure of the equivalent potential temperature is similar to the humidity structure, but with a stronger drop across the front, and a stronger difference between the weaker and stronger cold pools governed by the temperature drops. The relative humidity signal in the front is mostly governed by the temperature decrease, with RH being 8% larger at t min for the strongest cold pools. The in-front wind speed increase has a maximum in the middle of the front. After the frontal passage, 300 the wind speed decreases slightly below the pre-front level. The strengthening winds in the front and the slackening winds in the wake are again significantly more pronounced for the strongest cold pools, with a maximum of 1.5 m s −1 and a minimum smaller −0.5 m s −1 in the front and wake compared to the value at t max . Figure 3f-g show the composite mean R freq and RR measured by the MRR. Both rain variables increase rapidly after the onset of the cold pool, peak towards the middle or end of the front, and start to decrease shortly before t min . The strongest cold pools have much larger rain rates and rain frequencies 305 during the entire front compared to the weakest cold pools, and the rain frequency of the strongest cold pools also remains strongly elevated until more than an hour after t min .
The last panel of Figure  inside the front, respectively. The strongest cold pools have significantly stronger downdrafts and also updrafts compared to the weakest cold pools (see also Table 1), the latter highlighting the potentially enhanced triggering of new convection by stronger cold pools. For the vertical velocity averaged over the entire sub-cloud layer (w SCL ), the picture is similar, but the peak w maxSCL is slightly smaller for the strongest cold pools and more similar compared to the weaker cold pools (Table 1).
As already shown in Table 1, Figure 3 shows that the strongest cold pools are also the driest and the rainiest, and have the 315 strongest wind and vertical velocity anomalies in the front. The relationships and timings discussed are mostly the same when considering all cold pools meeting the noprev criterion (i.e. also including summer periods), just with larger anomalies and the differences mentioned above for the humidity structure. The mean temporal structure for all variables-except for the specific humidity and partly for the wind speed-is also similar to previous observations of tropical deep convective cold pools during The initial increase in humidity at the edge of the front at BCO might be explained by enhanced surface fluxes due to the strengthening winds (Langhans and Romps, 2015;Torri and Kuang, 2016), or by an accumulation of moisture from evaporation of precipitation of the parent convection, which was pushed to the edge of the front (Tompkins, 2001). Analyses of the various isotope measurements made during the EUREC 4 A field campaign  might help elucidate the origin of 330 these moisture rings. This could also help understand why cloud-resolving models seem to have difficulties in representing the humidity structure in the cold-pool front correctly (Chandra et al., 2018).
The cloud radars at BCO also allow study of how the cloud properties change across the cold-pool passage (Figure 4).
The mean cloud-top height (CT H) increases rapidly by ∼ 500 m after the cold-pool onset and peaks at the end of the front.
CT H remains elevated by ∼ 300 m compared to the pre-front value in the following hour. The 25% strongest cold pools are 335 associated with significantly deeper clouds throughout the entire period shown, especially so at the end of the front, when the CT H is on average higher than 3300 m. The cloud-base height (CBH) starts to decrease already slightly before t max and reaches its minimum near the end of the front at ∼ 500 m. This decrease is due to the more frequent precipitation with very low echo-base heights, and is most pronounced for the strongest cold pools.
The total hydrometeor cover (CC) increases rapidly at the beginning of the cold-pool front, remains about 25% larger 340 compared to the pre-front value inside the front, and then decreases slowly in the wake. The mean CC of the 25% strongest cold pools reaches nearly 100% at the end of the front and is significantly larger than the CC of the weakest cold pools during the entire period shown, especially so in the wake. Figure 4d-f show that the enhanced CC of the strongest cold pools in the prior hour is entirely due to cloud segments with CBH above 1 km (CC aloft ), whereas the enhanced CC in the front and wake of the strongest cold pools is mostly due to precipitating cloud segments with CBH below 300 m. The rapid increase in CC lcl 345 up to its peak at t max strongly contributes to the CC increase at the edge of the front. This peak is also larger for the strongest cold pools, consistent with their larger w 450 at t max . CC lcl and CC aloft are lower at the end of the front for the strongest cold pools, as the lowest CBH is mostly below 300 m and the cloud segments thus count to the CC prcp category (note that a given time can only count to one of the three categories).
In Figure 4d-f the cloud cover is split into contributions from cloud segments with different CBH without accounting for 350 the information of the cloud entity. Based on a similar analysis that accounts for the entity information (see Appendix A), we find that the peak in CC lcl at t max is mainly due to edges of precipitating clouds that have a CBH > 300 m. Assuming that this cloud population represents the clouds evident as mesoscale arcs in satellite imagery, this suggests that the cloudiness at the gust front is mostly characterized by well-developed precipitating clouds. The entity analysis also shows that more than also suggested by the time-height plots of the composite-mean hydrometeor fraction shown in Figure 4g-i. These panels nicely summarize what was discussed in the previous paragraphs, and again highlight the differences between the 25% strongest and weakest cold pools in terms of the cloud response.   They show that cold-pool periods are much cloudier than the average winter trades. Cold-pool periods also have much deeper 360 clouds, which is expected as it needs deeper precipitating clouds to form cold pools. The enhanced CC in the wake of cold pools compared to the long-term mean is nevertheless surprising, as convection might be expected to be suppressed in the coldpool wake. Mesoscale arcs encircling vast decks of deeper cumuli with stratiform layers therefore seem more representative for periods of cold-pool activity than the more classical picture of trade cumulus cold pools as mesoscale arcs enclosing broad clear-sky areas.

365
Despite the various significant differences between the strongest and weakest cold pools highlighted in the previous paragraphs, there is a lot of variability among individual cold pools. The variability is illustrated in Figure 5, which shows the temporal structure of the most important variables for individual cold pools ranked according to their ∆T . Especially the individual differences in humidity and wind in the front and the beginning of the wake can by far exceed the mean differences among the strongest and weakest cold pools shown in Figure 3. Tendencies of more frequent (and intense) rain, deeper clouds

Daily cycle
The long timeseries also allows to study the variability of the cold-pool frequency and characteristics at the daily timescale. During nighttime between about midnight and 04 LT, cold pools are associated with significantly deeper clouds, stronger mean rain rates, stronger downdrafts and updrafts, larger CC, and slightly stronger humidity drops and weaker wind gusts compared 380 to daytime cold pools between about 08-16 LT. There is also a hint of slightly stronger ∆T during nighttime compared to daytime, but neither in the median nor in the 25% quartiles is this daily cycle significant. It is somewhat surprising that we find no pronounced daily cycle in ∆T , although the daily cycle of e.g. w minSCL and CT H max would suggest that ∆T should be stronger at nighttime compared to daytime. There is a climatological background daily cycle in temperature of about 1.2 K due to the daytime solar heating (minimum and maximum temperatures near 5 and 12 LT, respectively), but this should not affect 385 the cloudy cold-pool periods much and would at best contribute to lower ∆T in the morning. Other diagnostics like ∆q max and R int do not show a pronounced daily variability (not shown).
The pronounced daily variability in the cold-pool frequency and most diagnostics is not surprising given the distinct daily  The peaks in the cold-pool frequency at 09 and 14 LT are shifted by a few hours compared to the peak in the surface 395 precipitation between 03-06 LT (Nuijens et al., 2009;Vial et al., 2019). This suggests that cold pools help extend the daily cycle of shallow convection into the early afternoon, which could be due to cold pools reinforcing each other and triggering subsequent cold pools. This hypothesis is supported by the shorter median interval between subsequent cold pools of 121 min between 07-14 LT compared to 182 min between 22-04 LT. Also the daily cycle of cloud cover seems to be slightly extended into the morning, with CC tot decreasing below the daily mean about 4 h later compared to the climatological CC. The textbook Gravel day is characterized by many short and often weak cold pools quickly following each other, interspersed by stronger cold pools. The cold pools are associated with the presence of strongly precipitating deeper clouds (note that the 415 radar did not work prior to 12 LT). The many cold pools present on this day clearly imprint their signature on the satellite image in the form of mesoscale arcs.
The cold pools on the Flowers day are associated with the large cloud system whose stratiform layer reaches the BCO at 10 LT. Three cold pools are directly associated with the large system, with the first one starting at 11 LT showing a very strong ∆T of −3.85 K. The large system has rain rates up to 3.6 mm h −1 and is announced by a weaker cold pool associated with the 420 very thin mesoscale arc visible in the satellite image, which goes along with a strong increase in humidity of 1.3 g kg −1 .
The Fish day features a 6 h long cold pool associated with steady and intense rain (maximum RR of 11.6 mm h −1 ), continued strong downdrafts and very large humidity throughout its entire duration. The temperature fully recovers within about 20 min of the cold-pool end, and 3 h later two subsequent pronounced cold pools follow that are again characterized by continued precipitation and downdrafts. The satellite image shows the fish-bone like cloud band typically associated with the Fish pattern.

425
The occurrence of the Fish pattern is strongly connected to trailing cold fronts of extratropical origins (Aemisegger et al., 2021;Schulz et al., 2021). The more front-like character of the Fish cold pools with steady showers and downdrafts is clearly evident in the example timeseries.  As expected, only 36 cold pools are detected during Sugar periods. Many cold pools are also associated with the No category (341). When we look at the fraction of time a given pattern is subject to a cold pool, the picture changes and the Fish pattern is associated with the largest cold-pool fraction (12.8% of time), followed by Flowers and Gravel (9.9% and 724%, respectively).
Figure 7b also shows the cold-pool fractions using different selection criteria, namely that only one pattern is allowed at a time ('.only'; excluding cold pools that pertain to multiple patterns), that all noprev cold pools from all seasons are used (rather than only from the winter months; 'all'), and that periods of deep convection in all seasons are excluded ('all.nodeep'; i.e. no cold pools with any radar signal between 4.5-8 km). For the Gravel pattern, these different criteria hardly influence the cold-pool fraction, whereas for Flowers and Sugar the different sets of criteria tend to change the cold-pool fraction. For

440
Flowers, the cold-pool fraction in winter reduces to 8.6% if periods with multiple patterns and their cold pools are excluded.
Only 85 cold pools are left for Flowers.only, while the rest are shared with Gravel (86), Fish (80) and a few also with Sugar (7).
While excluding periods of multiple patterns more than halves the cold-pool fraction for Sugar (to 0.8%, mostly due to overlap with the Gravel pattern), considering all seasons nearly doubles the cold-pool fraction of Sugar. Despite these differences, the four patterns remain distinct in their cold-pool fractions independent of the criteria considered. The cold-pool fraction of the

445
No category in winter is with 6.4% also substantial. The No category is particularly sensitive to the inclusion of all seasons, and in summer with more frequent deep convection most cold pools pertain to the No category (not shown). Excluding periods of deep clouds ('all.nodeep') therefore mostly affects the No category, as deep convection is usually absent when patterns are detected.
That Gravel has the largest number of cold pools but only the third largest cold-pool fraction is partly due to Gravel being the   during the cold-pool periods, the climatological differences in CC and CT H among the different patterns remain (see also Schulz et al., 2021;Vial et al., 2021;. Fish has the largest CC, closely followed by Flowers 1 , and then Gravel and Sugar. The CC differences are mostly due to the differing contributions of CC aloft , whereas CC lcl is fairly similar among the patterns both for the cold-pool periods and the climatological mean . Also the temporal structure of 475 most rapidly back to its pre-front value. Fish also tends to have the deepest mean CT H associated with the cold-pool periods, closely followed by Gravel and Flowers. The mean CT H of Gravel cold pools increases more rapidly in the front compared to Flowers and Fish, but also decreases a bit faster in the wake of the cold pools. Again, the cold-pool onset has the strongest CT H imprint for the Sugar pattern, with a mean CT H increase exceeding 1 km between t max and t min . The differences in the cloud properties of the different patterns associated with the cold-pool passages are again summarized As mentioned before, Vial et al. (2021) find the daily cycle of trade cumuli to be strongly linked to the daily cycle in the 490 occurrence frequency of the mesoscale organization patterns. Figure 9a shows strong daily variations of the number of cold pools associated with the different patterns. These variations are strongly connected to the daily cycles in the occurrence frequency of the patterns (Figure 9b and Vial et al., 2021). The maximum number of Gravel cold pools occurs just after midnight, followed by Flowers around 7 LT, and Fish cold pools at 10 LT. The number of Sugar cold pools is very low throughout the day.
495 Figure 9a suggests that the extension of the daily cycle of convection into the early afternoon due to cold pools may largely be explained by the Fish pattern, together with a substantial contribution of the No category to the peak at 14 LT. Despite the strong connection between the daily phasings of Figure 9a-b, especially the Fish pattern also shows a daily cycle of the cold-pool fraction with a peak in the afternoon (Figure 9c), which is broadly in phase with the occurrence frequency. The daily cycle in the cold-pool fraction might be due to cold pools lasting a while once they are formed, which is supported by the 500 much weaker daily cycles of the cold-pool front fraction (dashed lines in Figure 9c). Once present, cold pools often trigger new cold pools, as indicated by the 33% shorter interval between subsequent fronts during daytime compared to nighttime (see discussion in Section 3.3). From the present analyses, it is difficult to disentangle causal relationships between the pattern occurrence, cold pools, and the daily cycle. It is also difficult to pin down the evolution from one pattern to another, and the role of cold pools therein. As the number of cold pools per pattern and hour is quite low (especially in the case of Flowers), 505 more data is needed to draw robust conclusions on this.
The pattern-associated daily phasing of the cold-pool number might give a clue about why ∆T varies little on the daily timescale (Figure 6c), although the daily cycle of most cold-pool properties would suggest that ∆T should be stronger at night compared to day. The daytime Fish pattern has significantly stronger ∆T compared to the nighttime Gravel pattern (Figure 7e), which might compensate for the opposite expectation due to the daily phasing of CT H max and w minSCL .

Conclusions
This paper presents a longterm climatology of trade cumulus cold pools based on more than ten years of in-situ and groundbased remote sensing data from the Barbados Cloud Observatory (BCO; Stevens et al., 2016). Cold pools are detected by abrupt drops in low-pass filtered temperature timeseries and their associated changes in surface meteorology, cloudiness and sub-cloud layer dynamics are extracted. The cold-pool climatology is combined with a neural network classification of the 515 four mesoscale organization patterns Sugar, Gravel, Flowers and Fish  based on GOES-16 ABI infrared images . To focus on trade cumulus cold pools, most analyses are restricted to the set of 3889 cold pools detected in the dry winter regime from December to April that have no non-recovered cold pool in the hour prior to their onset.
We find cold pools to be ubiquitous in the winter trades-they are present about 7.8% of the time and on more than 73% of days at least one cold pool is detected. The average cold-pool passage is characterized by a 0.9 K temperature drop, a 0.2 g kg −1 520 humidity increase at the onset and a −0.4 g kg −1 humidity decrease at the end of the front, wind speed increases of 1.15 m s −1 , and rain intensities of 0.9 mm h −1 . The vertical velocity at the sub-cloud layer top shows a pronounced peak of 1 m s −1 near the cold-pool onset and sub-cloud layer averaged downdrafts of −0.55 m s −1 near the end of the front. Strong signals of coldpool passages are also found for all cloud macrophysical properties analysed: cloud-top height increases, cloud-base height decreases (due to the very frequent precipitation), and cloud cover increases with the cold-pool onset. Cloudiness at the gust 525 front is mostly due to cloud segments near the lifting-condensation level that pertain to larger precipitating cloud entities.
Similarly, cloud segments with bases above 1 km in the cold-pool wake are mostly part of large precipitating clouds, and not from detached stratiform layers.
The strength of the cold-pool signature depends strongly on the intensity of the temperature drops (∆T ). Cold pools with stronger ∆T are associated with deeper clouds, stronger precipitation, downdrafts, and humidity drops, stronger wind gusts 530 and updrafts at the edge of the front, and larger cloud cover compared to cold pools with weaker ∆T . Stronger cold pools also last significantly longer and follow each other more quickly than weaker cold pools. We find that also the minimum vertical velocity averaged over the sub-cloud layer and the maximum cloud-top height distinguish stronger and weaker cold pools very well. Especially the downdraft strength is a very robust indicator of cold-pool strength and together with the cold-pool front duration it explains 50% of the variability in ∆T .

535
The cold-pool frequency and characteristics also show pronounced daily variability. There are significantly less cold pools and a lower cold-pool frequency between 16-22 LT compared to the rest of the day. We find that cold pools extend the daily cycle of convection into the early afternoon, with a peak in both the cold-pool number and fraction at 14 LT. Also most coldpool diagnostics show a pronounced daily cycle, with significantly deeper clouds, stronger mean rain rates, stronger downdrafts and updrafts, larger cloud cover, slightly stronger humidity drops and weaker wind gusts associated with nighttime compared 540 to daytime cold pools. The phase of these daily signatures is consistent with their background climatological daily cycle, but shifted to much larger values. For the vertical velocity minima and maxima, also the amplitude of the daily cycle is much more pronounced during cold-pool periods.
In the wet summer regime, cold-pools are about 30% more frequent relative to the average winter regime. Summer cold pools are also associated with significantly stronger temperature and humidity drops, deeper clouds and stronger downdrafts-545 consistent with the frequent deep convection and stronger precipitation of this season (Brueck et al., 2015). On the other hand, the summer cold pools have weaker updrafts and humidity maxima at the beginning of the front, suggesting that they might be less effective in triggering new convection. While the temporal structure of cold-pool passages for most meteorological variables in both seasons resemble those of previous observations of tropical deep convective cold pools (de Szoeke et al., 2017;Chandra et al., 2018;Zuidema et al., 2017), especially the humidity structure and also the generally larger anomalies 550 render the summer cold pools more similar to the deep convective cold pools from previous studies.
We also analysed if the cold-pool frequency and characteristics depend on the pattern of mesoscale cloud organization. The most pronounced difference among the patterns lies in the occurrence frequency of cold pools, with Fish having the largest cold-pool fraction (12.8% of time), followed by Flowers and Gravel (9.9% and 7.2%, respectively). As expected, the cold-pool fraction of Sugar is negligible (1.6%). Fish cold pools last significantly longer than cold pools from all the other patterns, 555 and they are also associated with the strongest temperature drops and downdrafts. Gravel cold pools are associated with the strongest updrafts at the cold-pool onset and the deepest cloud-top height maxima.
Given the distinct daily cycle in the occurrence frequency of the four patterns found in Vial et al. (2021), it is not surprising that we find strong daily variations of the number of cold pools associated with the different patterns. The maximum number of Gravel cold pools occurs around midnight, followed by Flowers around 7 LT, and Fish cold pools around 10 LT, in line with 560 the daily cycles in the occurrence frequency of the patterns. The Gravel, Flowers and Fish cold pools can thus explain a large fraction of the daily cycle in the cold-pool occurrence, as well as their extension into the early afternoon. Note also that the unclassified cold pools have a non-negligible contribution to the peak at 14 LT. Interestingly, the climatological differences in the cloud cover and cloud-top height among the different patterns are also present during cold-pool periods-the overall cloud cover and cloud-top height for all patterns is just enhanced compared to their respective climatological values.

565
This study paves the way for more in-depth analyses of the cold-pool properties and their relation to the environment in the trades. Especially the complex humidity signals deserve a more detailed investigation, also using data from the recent EUREC 4 A field campaign  and from realistic large-eddy simulations. Together with the vertical velocity statistics, the humidity anomalies can help shed light on the triggering of new convection at the cold-pool front Additional measurements of the mixed-layer depth from radiosondes and the Raman or Doppler lidar could help refine the cold-pool end 570 definition, which is only poorly constrained by the surface temperature data. Such additional data could also provide interesting insight into the cold-pool recovery process. A systematic matching with satellite imagery would also help collocate the clouds sampled at BCO with the broader view of the entire cold pool seen from space.
Overall, we find that the cold-pool periods are about 90% cloudier relative to the average winter trades. The larger cloudiness is mostly due to larger cloud cover from precipitating and stratiform cloud segments. Also the wake of cold pools is 575 characterized by above average cloudiness, indicating that the classical image of trade cumulus cold pools as mesoscale arcs enclosing broad clear-sky areas is rather the exception than the rule. Our study suggests that a better understanding of how trade-cumulus cold pools interact with and shape their environment is important to understand the variability in cloud cover and cloud organization in the trade-wind regime.
Code and data availability. The BCO data used in the analysis and other supplementary information that may be useful to reproduce the

585
Appendix A: Cloud cover contributions from different types of cloud entities The contributions to total cloud cover from clouds at different height levels can either be computed by classifying every radar profile independently based on its CBH (see Figure 4d-f), or-if a cloud segmentation mask is available-by classifying the entire cloud entities according to their cbh ID (i.e. their overall lowest CBH). As both approaches can provide valuable insights, Figure A1 also shows the temporal structure of the cold-pool signatures for the latter classification method. For this, the cloud 590 cover is again split up into contributions from precipitating clouds with cbh ID ≤ 300 m (CC ID.prcp ), LCL clouds (CC ID.lcl ; 300 m < cbh ID ≤ 1 km), and stratiform clouds (CC ID.aloft ; 1 km < cbh ID ≤ 4 km). The difference between CC ID.prcp and CC prcp is that edges or slanted sides of precipitating clouds that have a CBH > 300 m are counted in their entirety to the CC ID.prcp category, while they would be counted in the CC lcl or CC aloft category if the cloud ID was not considered. Due to the potential presence of cloud entities at different heights, the sum of the three height categories (CC ID.tot ) can be larger than one. CC ID.prcp already starts to increase before t max and continues to increase until the middle of the front for all the coldpool sets shown. For the 25% strongest cold pools, the end of the front is entirely covered by precipitating clouds. CC ID.lcl in Figure A1c for all sets is relatively stable at about 17.5% before the cold-pool onset, decreases abruptly after t max to a minimum near t min , and then slowly recovers back to the pre-front value. CC ID.lcl shows the strongest impact when the cloud 600 entities are considered through the cbh ID and thus the strongest difference to the structure of CC lcl (Figure 4e). The absence of a peak in CC ID.lcl near t max indicates that the CC lcl peak there is almost entirely due to edges of precipitating clouds with a CBH > 300 m, and not due to (not-yet or) non-precipitating trade cumuli.
The temporal structure of CC ID.aloft resembles the structure of CC aloft (Figure 4f), yet with substantially lower coverage as most cloud segments with CBH > 1 km are connected to a precipitating core. This shows that nearly half of the CC aloft in 605 the cold-pool wake is part of large precipitating clouds, and not from detached stratiform layers.

Appendix B: Seasonal cycle of cold-pool characteristics
While this study focuses on the cold-pool climatology of the winter regime, it is also interesting to look at the seasonal cycle of the cold-pool characteristics at BCO. Using all cold pools of the noprev category, we find the largest median %-of-day in 610 cold pool in the summer months from July-November, and another peak in January ( Figure B1a). Only 13% of days have no cold pool at all in summer, compared to 27% in winter. The same monthly variability is found for the %-of-day in front, but with 45% lower values due to the shorter duration of the front compared to the entire cold pool (not shown).  ∆q min , CT H max and R int , as well as slightly stronger ∆T and w minSCL , consistent with the relationships discussed in Section 3. However, w max450 is significantly lower by 0.2 m s −1 and ∆q max by 0.1 g kg −1 in summer compared to winter, indicating that cold pools in summer might be less successful in triggering new convection. Furthermore, CC tot of summer cold pools is also significantly smaller compared to winter cold pools by about 10%. The differences in the cold-pool characteristics between the summer and winter regime are not surprising, as the summer regime is referred to as the wet season in Barbados 620 and characterized by frequent deep convection and much larger precipitation (Brueck et al., 2015). When excluding periods of deep convection (defined by the presence of a radar signal between 4.5-8 km), the number of cold pools detected in summer strongly decreases compared to winter, and the median summer cold pool also becomes weaker compared to the median winter cold pool (not shown).
Author contributions. RV initiated the project in exchange with PZ. HK prepared the radar cloud mask and HS the neural network classifi-