Characteristics of StationDerived Convective Cold Pools Over Equatorial Africa

Due to their potential role in organizing tropical mesoscale convective systems, a better understanding of cold pool (CP) dynamics in such regions is critical, particularly over land where the diurnal cycle further concentrates convective activity. Numerical models help disentangle the processes involved but often lack observational benchmarks. To close this gap, we analyze nearly 43 years of five‐minute resolution near‐surface timeseries, recorded from 12 automatic weather stations across equatorial Africa during 2019– 2023. We identify 4,218 CPs based on criteria for temperature and wind. The identified CP gust fronts, which exhibit respective median temperature and specific humidity decreases of 5.3 K and 2.8 g kg − 1 , closely correlate with satellite‐observed brightness temperature discontinuities. Despite weak diurnal variation in precipitation, observed CP occurrence shows a pronounced diurnal cycle with an afternoon peak — a feature we attribute to low‐level moisture conditions. Our findings can serve as observational benchmark to improve simulations of CP organization.


Introduction
Convective cold pools (CPs) are caused by the evaporation of rainfall beneath deep convective clouds (Zuidema et al., 2017).The resultant denser air volume spreads out laterally along the surface and can cause a so-called "gust front" (GF) along its edges (Charba, 1974).The GF features strong horizontal and vertical winds along with moisture and temperature anomalies which together can give rise to additional deep convective events, for example, under collisions (Feng et al., 2015;Purdom, 1976).CPs are thus important agents in mediating interactions between deep convective cells and thus the self-organization of thunderstorm systems (Haerter et al., 2019;Jensen et al., 2021;Simpson, 1980;Tompkins, 2001b).
Indeed, the importance of collecting information on CPs and precipitation in deep tropical regions has been pointed out (Adams et al., 2015) but systematic, climatological studies on CPs in such regions are still rare or lacking.This may partially be due to difficult environmental conditions which pose challenging demands on equipment and maintenance (Parker et al., 2008).Also the availability of funds may hinder systematic long-term campaigns in some regions.A notable exception is the trans-African hydro-meteorological observatory (TAHMO) which offers a promising network of station measurements in many sub-Saharan African countries (van de Giesen et al., 2014).Using a range of stations from the TAHMO network, we here present a climatology of CP measurements for equatorial Africa and compare findings to previous work in other geographic regions.

Station Data
We utilize data from twelve ATMOS41 automatic weather stations (AWS) (Figure 1), operated by TAHMO.The stations are situated in Cameroon, the Democratic Republic of Congo (DR Congo), Nigeria, and Uganda.To investigate the influence of regional climatic differences on CPs, we group stations according to their respective deployment countries in our analysis.
The AWS provide data at a five-minute temporal resolution.All stations are installed at an approximate height of two m above the surface.We employ the station records of precipitation, atmospheric pressure, air temperature, relative humidity, and wind gust speed.ATMOS41 determines the latter by measuring instantaneous wind speed every 10 seconds and outputting the maximum instantaneous wind speed value within any given five-minute interval.If an instantaneous wind speed is larger than eight times the running average of the previous 10 instantaneous measurements, the measurement is rejected.While this method may prevent spurious spikes in the wind record under normal conditions, it can cause missing wind data in cases of large and sudden wind changes.Given the frequent occurrence of such strong wind variations associated with CP GFs, approximately 22% of all identified CPs have an incomplete wind record.
We analyze the data recorded by these AWSs from 1 January 2019, to 30 September 2023.As not all AWSs were operational throughout the entire period, we limit our analysis for each station to days with complete air temperature records.Additionally, we require the air temperature to be recorded for a minimum of 10 consecutive minutes from the previous day and for the subsequent 120 min on the following day.The resulting number of analysis days per station is indicated in Figure 1.In total, we analyzed 15,602 days and thus nearly 43 years of station data.
Based on the station-measured variables, we additionally compute both mixing ratio, r and saturated mixing ratio, r sat (see Text S1 in Supporting Information S1), and derive the specific humidity, q ≡ r/(1 + r) and the specific humidity deficit, q D ≡ q sat q = r sat /(1 + r sat ) q.

Satellite Data
Apart from the station data, we utilize infrared brightness temperature measurements which we derive from satellite-measured effective radiances.The radiances are extracted from Meteosat Second Generation (MSG) 0°p roducts provided by the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT).
The data has a baseline repeat cycle of 15 min and a spatial resolution of 3 km in the sub-satellite point.To convert the radiances to brightness temperatures we employ equation 5.3 and the corresponding regression coefficients of EUMETSAT (2012).

Cold Pool Detection Algorithm
(i) Temperature criterion A potential CP event is detected at a given time t if three conditions apply: (a), similar to Kirsch et al. ( 2021) we require a substantial temperature decrease ΔT ≤ 2 K, within the 20 min window from t 5 min to t + 15 min.Additionally, we require (b), the decrease of ΔT to be monotonic and (c), T(t) T(t 5 min) ≤ 0.5 K.
While the chosen conditions minimize the risk of high false alarm rates due to typical diurnal temperature changes, some weak or dissipating CP GFs might be missed.Choosing a ΔT of 1.5 K as used by Kruse et al. (2022) leads to approximately 28% more identifications, marginally weakening CP-associated median temperature and specific humidity decreases by 11%-12%, without impacting overall results.

(ii) Wind criterion
To confirm detected potential CP events, we adapt the wind criterion introduced by Kruse et al. (2022).For this purpose, we compute the wind gust speed anomaly for each time t as where u g is the wind gust speed (see Section 2.1) and u g its centered 2-hr running mean, that is, the mean value of the 25 wind gust speeds recorded during the corresponding 2-hr window.
For a potential CP event at time t we identify the maximum wind gust speed anomaly, Δu max g , between t 20 min and t + 40 min.We retain it as potential CP event if with n w = 3, the centered 24-hr running mean of the wind gust speed anomaly, Δu g , and the corresponding 24-hr running standard deviation, σ Δu g .While higher n w mainly increase the number of missed CPs, smaller n w values give rise to more false positive detections.For instance, using n w = 2 would yield 12% more identifications, with 2.5 times as many events that cannot be attributed to convective events in satellite data and, therefore, are rejected in step (iv).
As the potential CP onset is defined based on the temperature criterion, we also search for associated wind gusts in a 20 min time window before this onset.We choose 20 min rather than the 10 min used by Kruse et al. (2022) since our temperature criterion involves a minimum decrease of 0.5 K within 5 min to define the onset of potential CPs and might thus delay the onset in comparison to Kruse et al. (2022).The 40 min time window after potential CP onset allows significant wind offsets while ensuring a temporal relation between ΔT and Δu max g .Halving both time windows would reduce CP identifications by about 7%, with negligible impact on overall results.
In case of missing wind gust speed anomalies between t 20 min and t + 40 min, we identify the maximum wind gust speed, u max g within this time window rather than Δu max g and retain the event as potential CP if with n miss = 2, the centered 2-hr running mean, u g , and the corresponding standard deviation, σ u g .Approximately 25% of the detected CPs is identified based on Equation 3 rather than Equation 2. By evaluating u g and σ u g 80 min before potential CP onset, we keep again a 20 min offset between the onset and the 2-hr time window of the reference values.Since portions of the wind record are missing, we choose n miss conservatively.Yet, neither the number of identifications nor the results are sensitive to the selection of n miss .
If no wind gust speed data has been recorded between t 20 min and t + 40 min, or if the reference values u g (t 80 min) and u g (t 80 min) could not be computed due to missing data, we consider the event as "no CP." Differing from Kruse et al. (2022), we evaluate the wind criterion based on wind gust speed rather than wind speed.Since we work with station data with a temporal resolution of 5 min in contrast to 1 min in (Kruse et al., 2022), we find wind gust speed a better indicator for CP GFs than wind speed.
(iii) Duplicate detection check Often, a CP fulfills the defined criteria (i) and (ii) not only at time t, but also at subsequent time steps.Depending on the evolution of temperature and wind gust speed behind the CP GF, time steps in which the criteria are met can even be separated from each other by time steps in which the criteria are not met.To avoid duplicate detection of a given CP, we drop detected events if at least one other event was detected within 20 min before that particular event.Given the variety of environmental conditions under which we observe CPs at our station locations, we find this definition to be more permissive than the absolute 60 min time window after detected temperature decreases, within which Kirsch et al. ( 2021) considers any detected decrease as part of the same event.
(iv) Space-borne verification In some cases, especially during the Nigerian dry season, neither condition (i) nor (ii) may suffice to completely prevent false positive detections caused by strong diurnal temperature changes.Consequently, we validate each detected event using satellite-measured 10.8 μm brightness temperatures, BT 10.8 .We consider an event at time t as CP if the minimum BT 10.8 recorded between t 3 hr and t within a 2°radius around the station is less than 240 K.
The parameters of the verification ensure the presence of temporally related convective events in the vicinity of the potential CP event.Applying this verification to all detected events, 71 false positive detections are identified and dropped.Reducing the radius around the station from 2°to 1.5°would increase the number of rejected events to 85.

Determination of Cold Pool Anomalies
We analyze the effects of a detected CP with respect to different station-measured meteorological variables by considering a time window relative to CP onset, t 0 , from t 0 40 min to t 0 + 120 min.Within this time window, we evaluate the CP associated anomalies y′(t) ≡ y(t) y ref for a meteorological variable y based on an unperturbed reference state y ref , which we define as the temporal mean of instantaneous measurements in a time interval before CP onset.Since the onset is defined based on the temperature drop and thus could be different for other variables, we choose the time interval from t 0 40 min to t 0 20 min to keep a sufficient margin of 20 min to the CP onset while preserving the required temporal proximity.To minimize any distortion of the reference state through the diurnal cycle, we deviate from this definition only for temperature anomalies and follow the approach of the refined temperature drop from Kruse et al. (2022) instead, that is, we consider the maximum temperature of the two measurements in the 10 min time window preceding the CP onset as unperturbed reference temperature.
Due to the coarser temporal resolution, we extend the time window in which we analyze the anomalies before CP onset to 60 min for satellite-measured 10.8 μm brightness temperatures, BT 10.8 , and define the reference brightness temperature, BT ref 10.8 , as the mean of the three observations in the time interval from t 0 60 min to t 0 30 min.As there might not be a brightness temperature observation at the station-derived CP onset, t 0 , we define the closest satellite time step as t0 and measure the CP time relative to it.The brightness temperature anomalies, BT′ 10.8 , are then computed analogously to those for station-measured variables.Moreover, to further investigate the space-borne CP signature, we additionally determine the temporal change of BT′ 10.8 , as ΔBT′ 10.8 (t) = BT′ 10.8 (t) BT′ 10.8 (t 15 min).
The overall impact of the described parameter choices on the observed anomalies is relatively weak.For instance, doubling the time interval for computing the unperturbed reference state from 20 to 40 min, while maintaining the 20 min offset to CP onset, would result in a 1.4% enhancement in the median specific humidity drop.Similarly, calculating the unperturbed reference temperature based on the mean value of the four measurements before CP onset, instead of using the maximum of the previous two, would lead to a 1.3% reduction in the median temperature drop.

Seasonal and Diurnal Cycle of Observed Cold Pools
First, we derive the seasonal and diurnal cycles of CPs in the different sub-regions (Figure 1) and relate them with precipitation, convection depth and moisture conditions (Figure 2).With about 0.3-0.6CPs per day in the high seasons (Figure 2a, Table S2 in Supporting Information S1), equatorial Africa boasts particular CP  (Nov-Mar, dashed).g, Analogous to (e) but for specific humidity deficit, q D .h, Analogous to (f) but for mean q D .abundance compared to previous climatologies in other continental regions with about 0.1-0.3CPs per day (Kirsch et al., 2021;Kruse et al., 2022;Redl et al., 2015).In every sub-region, the number of CPs peaks twice during the course of the year with a first maximum between March and May and a second maximum between September and October.The bi-modality in the annual cycle of CPs largely corresponds to the latitudinal migration of the Inter-Tropical Convergence Zone (ITCZ) as reflected in the precipitation seasonal cycles (Figure 2c).However, we note that precipitation may not explain all the features of the annual cycle of CPs and the differences between sub-regions.For instance, Nigeria presents a single precipitation peak in September whereas CP occurrence peaks in both May and September.There, the strong CP activity during May may be related to the combination of deeper convection (Figure 2e) fed by high equivalent potential temperatures (θ E ; Figure S1 in Supporting Information S1) and of higher low-level specific humidity deficit (q D ; Figure 2g) boosting rain evaporation.We also note that Uganda receives the least precipitation among sub-regions while experiencing most frequent CPs during the year.We attribute the larger number of CPs in Uganda to generally drier conditions at low levels (Figure 2g).
The diurnal cycle of CPs strongly peaks between 15 LT and 18 LT in all regions except Nigeria, where the peak is reached between 18 LT and 21 LT (Figure 2b).The high CP activity during the afternoon can be directly related to the afternoon peak in (deep) convection, highlighted by maxima in precipitation (Figure 2d) and lower brightness temperatures (Figure 2f).Consistently with earlier studies (Andrews et al., 2023;Camberlin et al., 2018;Zhang et al., 2016), precipitation shows secondary nocturnal peaks in Uganda and Congo, and remains high during the night in Nigeria, whereas the proportion of CPs displays local minima during these hours.This mismatch between precipitation and CPs is likely to be related to both the decline of convection during the night (leading to weaker rainfall intensities and downdrafts) and to moister conditions at the surface (reducing rainfall evaporative cooling; Figure 2h) which both inhibit CP formation (Zuidema et al., 2017).

Observed Cold Pool Characteristics
We further characterize equatorial African CPs by relating their temperature and moisture anomalies (defined in Section 3.2).On average, such CPs are accompanied by 5 K drops in temperature -compared to 3 K in Germany; Kirsch et al. ( 2021) -occurring within approximately 30 min, with little variability among subregions (Figure 3a, Table S2 in Supporting Information S1).In distinction to Kirsch et al. ( 2021), the CPs we analyze here generally show decreases in specific humidity after CP passage (Figure 3d).Interestingly, we find that the magnitude of this decrease is smaller for the elevated (Table S1 in Supporting Information S1) stations of Uganda ( 1 g kg 1 ) characterized by less deep convection (Figures 2e and 2f) and drier low-level environments (Figures 2g and 2h) compared to the other sub-regions (about 3 g kg 1 ).Less deep convection is likely to be associated with lower (relative to the ground) convective downdrafts origins (Zuidema et al., 2017), importing less upper-level dry air to the surface.This may combine with enhanced rain evaporation due to drier environments, thus explaining the more modest reduction in specific humidity over Uganda.When considering the 25% driest (moistest) low-level pre-CP environments, we further evidence the large impact of moisture conditions, and thus of rain evaporation, on CP temperature and moisture anomalies in all sub-regions (Figures 2b, 2c, 2e and 2f): CP anomalies typically are 3 K cooler and 2 g kg 1 moister in the driest pre-CP conditions than in the moistest pre-CP conditions.We note weak maxima in specific humidity occurring few minutes after CP onset -so-called moisture rings (Langhans & Romps, 2015;Schlemmer & Hohenegger, 2016;Tompkins, 2001a) -over Cameroon, Uganda and Congo for the driest pre-CP environments.Finally, the temporal evolution of q D (Figures 3g-3i) reveals that, in the driest environments, rain evaporation may not be sufficient to saturate the low-level air -similar to Germany; Kirsch et al. (2021).
Addressing CP cloud characteristics, we find that 96% (Uganda) to 100% (Congo) of CP GFs are accompanied by shallow or deep convective clouds (Figure S2a in Supporting Information S1).More specifically, CPs are generally accompanied by a strong decrease in BT 10.8 (Figure 4a), for example, reaching 36 K in Congo.The BT 10.8 minimum is typically achieved 30-45 min after CP onset.While this minimum is delayed w.r.t. the CP onset, we find a minimum of the time derivative of BT′ 10.8 to be synchronized with CP onset in all sub-regions (Figure 4b).This observation suggests that CPs in equatorial Africa, and potentially other regions, might be detectable from space-borne satellite data.

Summary and Discussion
The present study provides multi-year statistics of cold pool characteristics in equatorial Africa, based on fiveminute near-surface weather data.Using detection methods similar to those in previous studies focused on mid-latitude continental regions, key findings include that temperature drops upon gust front passage often exceed 5 K and specific humidities typically decrease by more than 3 g kg 1 .Weak moisture rings can only be identified in some of the sub-regions for the driest pre-cold pool environments -in agreement with Kruse et al. (2022) for data in the Netherlands where moisture rings were generally not detected.
Seasonally, the rate of cold pool occurrence roughly follows precipitation statistics.Diurnally, in contrast, cold pool occurrence sharply peaks during the drier late afternoon times, departing from the weak precipitation diurnal cycle.
We suggest this may be due to the nocturnal boundary layer often being close to saturation, thus diminishing cold pool occurrence there.This finding may have important implications for thunderstorm organization through cold pool activity: the limited time window where cold pools actually occur during the day means that self-organization may be limited to relatively short periods of the day.One could speculate that it is the lack of cold pool activity that limits the duration of mesoscale convective systems, often less than 12 hr, rather than the precipitation itselfwhich is more spread out over the day.Yet, further investigations are needed to substantiate this claim, for example, by using additional radiosonde data to explore the vertical coordinate Future studies should also analyze if deep convection is more scattered during nocturnal periods when fewer cold pools occur.Comparisons with oceanic cold pools and their organizational effects, which tend to be weaker (Zuidema et al., 2017), would be insightful.
Our cold pool detection algorithm can be adapted to other regions, provided that there are in-situ weather stations measuring surface wind and temperature with at least 5-min temporal resolution.However, in-situ weather stations meeting this requirement are still limited in the tropics, whereas cold pools are abundant.Encouragingly, our findings may have implications for satellite-based cold pool detection: we show that gust front passage clearly correlates with discontinuities in satellite-derived brightness temperature.We generally observe a significant decrease in brightness temperatures around the time of the gust front passage, with maximum decrease rates at the station-derived cold pool onset.Our findings thus suggest that cold pools in equatorial Africa, and potentially other regions, could be directly detectable from geostationary satellite data on a continental scale.While significant decreases in brightness temperature alone are not a sufficient criterion for identifying cold pool gust fronts, the concurrence with spatiotemporal patterns like radially spreading cloud arcs and rapidly expanding deep convection (Text S2, Figure S3 in Supporting Information S1) could aid neural networks, such as those developed by Hoeller, Fiévet, et al. (2024), in limiting the number of false positive detections.Even in cases where not all parts of a cold pool gust front exhibit brightness temperature drops (Figure S2b in Supporting Information S1), the presence of such spatiotemporal patterns may enable the neural networks to accurately track the gust front.
Both the code for the cold pool gust front identification and the processed data sets are licensed under Creative Commons Attribution 4.0 International and were used in version 2.0 (Hoeller, Haerter, & Da Silva, 2024).The raw data of the automatic weather stations was provided by TAHMO and is not publicly available.For more

equatorial
Africa based on temperature and wind criteria • The occurrence and intensity of the observed cold pools are related to lowlevel moisture conditions and the depth of convection • The identified cold pool gust fronts closely correlate with satelliteobserved brightness temperature discontinuities Supporting Information: Supporting Information may be found in the online version of this article.

Figure 1 .
Figure 1.Weather station data employed.Map of equatorial Africa showing station locations (filled black circles).Numbers within colored circles and the size of the circle area represent available days of station data with a complete record of air temperature.Stations are grouped into four regions denoted by the colored circles: Cameroon (blue), Democratic Republic of Congo (yellow), Nigeria (green), Uganda (red).

Figure 2 .
Figure 2. Observed seasonal and diurnal cycles.Line colors indicate distinct sub-regions (see legend).a, Mean number of daily cold pool (CP) events for each month.Lines interpolate linearly between markers to facilitate the interpretation; colors indicate different regions.The number of CP events is normalized based on the number of analyzed days per month and region.b, Proportion of CP events at different times of the day.Each marker represents the proportion of CP events observed within a given 3-hr time interval, starting with the interval [0,3) for the marker at 0 LT.c, Analogous to (a) but for precipitation.d, Analogous to (b) but for precipitation.e, Mean 10.8 μm brightness temperature, BT 10.8 , of deep clouds (BT 10.8 ≤ 240 (k) for each month.f, Mean BT 10.8 of deep clouds at different times of the day.The two lines for Nigeria represent rainy (Apr-Oct, dotted) and dry months (Nov-Mar, dashed).g, Analogous to (e) but for specific humidity deficit, q D .h, Analogous to (f) but for mean q D .

Figure 3 .
Figure 3. Station-derived cold pool (CP) properties relative to CP onset, t 0 .a, Mean temperature anomalies, T′, for different regions; shading indicates the 95% confidence interval.b, Analogous to (a) but for the 25% driest pre-CP environments of each region w.r.t. the reference specific humidity deficit, q ref D , prior to t 0 .c, Analogous to (b) but for the 25% moistest pre-CP environments.d-f, Analogous to (a)-(c) but for mean specific humidity anomalies, q′.g-i, Analogous to (a)-(c) but for mean specific humidity deficits, q D .Note that only timeseries of CPs, where t 0 is more than 120 min apart from other CP onsets, are included in the analysis.

Figure 4 .
Figure 4. Space-borne signatures of cold pools (CPs) relative to CP onset, t0 a, Mean 10.8 μm brightness temperature anomalies, BT′ 10.8 , of station-derived CPs for different regions; shading indicates the 95% confidence interval.b, Analogous to (a) but for the corresponding derivative ΔBT′ 10.8 .