Tropical Convection Overshoots the Cold Point Tropopause Nearly as Often Over Warm Oceans as Over Land

Tropical convection that overshoots the cold point tropopause can impact the climate by directly influencing water vapor, temperatures, and thin cirrus in the upper troposphere‐lower stratosphere (UTLS) region. The distribution of cold point overshoots between land and ocean may help determine how the overshoots will affect the UTLS in a changing climate. Using 4 years of satellite and reanalysis data, we test a brightness temperature proxy calibrated by radar/lidar data to identify cold point‐overshooting convection across the global tropics. We find evidence of cold point‐overshooting convection throughout the tropics, though other cirrus above the cold point cover an area 100 times larger than overshooting tops. Cold point‐overshooting convection occurs 30%–40% more often over convectively active land areas than over the warmest oceans. This proxy can be generalized to evaluate the fidelity of cold point overshoots simulated by storm‐resolving models.


Introduction
The tropical upper troposphere-lower stratosphere region (UTLS, ∼12-20 km) plays a critical role in the Earth's climate by influencing the composition of the lower stratosphere.The Brewer-Dobson circulation helps loft air from the upper troposphere into the lower stratosphere, which gets "freeze-dried" as it passes through the coldest temperatures near the tropopause, helping to set stratospheric moisture content (Brewer, 1949;Dessler, 2002;Holton & Gettelman, 2001).This mechanism is robust in observations; variations in stratospheric moisture are highly correlated with variations in the cold point temperature on subseasonal to interannual time scales (e.g., Randel & Jensen, 2013;Randel et al., 2004;Zhou et al., 2001).This is significant for climate change.Many climate models project stratospheric water vapor will increase under 21st century greenhouse warming (Dessler et al., 2016;Tian et al., 2023), producing an additional greenhouse effect which would further increase the surface warming rate (Solomon et al., 2010).
However, this is a challenging modeling problem.In addition to nonlocal dynamical drivers, a complex, multiscale mixture of local physical processes is thought to regulate the cold point and overlying lower stratosphere.Though infrequent, convection that overshoots the cold point can alter the cold point temperature through entrainment (Chae et al., 2011;Gettelman et al., 2002;Kuang & Bretherton, 2004;Randel & Park, 2019) and modify stratospheric air through turbulent mixing of ice and water vapor (e.g., Corti et al., 2008;Dion et al., 2019;Ueyama et al., 2023) and removal of vapor by deposition onto sedimenting ice crystals (Jensen et al., 2007;Khaykin et al., 2022).Observations during the southeast Asian monsoon suggest ice injected by convective overshoots over land strongly affects the regional UTLS composition (Bucci et al., 2020).Cold point-overshooting convection may drive much of the transport between the upper troposphere and lower stratosphere (Pommereau, 2010;Vernier et al., 2011).It can also support the formation of thin UTLS cirrus through anvil detrainment or by generating gravity waves where ice nucleates in the cold perturbations (Chang & L'Ecuyer, 2020;Jensen, Toon, Selkirk, et al., 1996;Jensen et al., 2016;Krämer et al., 2016).
Thin cirrus are common just below the cold point, especially around areas of frequent deep convection (Virts & Houze, 2015), with possible impacts on the UTLS from ice removal and radiative lofting (Fueglistaler et al., 2009;Jensen, Toon, Pfister, et al., 1996, Jensen, Toon, Selkirk, et al., 1996).An outstanding question is whether cold point-overshooting convection or thin cirrus has a greater impact on cold point temperatures and thus lower stratospheric moisture.Even global storm-resolving models, which explicitly simulate deep convection that reaches into the UTLS, have substantial intermodel spread in simulating the convection, cold point temperature, and cirrus that affect the UTLS (Nugent et al., 2022;Turbeville et al., 2022).
An important but imperfectly understood aspect of cold point-overshooting convection is its geographical and seasonal distribution, and how it differs between land and ocean regions.When defined using the 20 dBZ echo top height, the deepest tropical convection (Zipser et al., 2006) occurs mostly over land (e.g., N. Liu & Liu, 2016;N. Liu et al., 2020;Xian & Fu, 2015).But far too little large ice is typically lofted in cold point overshoots to reach such reflectivities.Other observational studies using more sensitive radar thresholds (e.g., −20 to −30 dBZ) to identify high convective cloud tops have found comparable frequencies of overshoots in convectively active land and ocean regions (Iwasaki et al., 2010;Li et al., 2022;Luo et al., 2008;Takahashi & Luo, 2014).Studies that used brightness temperature thresholds have even found more overshoots over oceans (Gettelman et al., 2002;Rossow & Pearl, 2007).The weaker land/ocean contrast in cold point overshoots versus in intense mid-tropospheric updrafts is not fully understood.The upper-tropospheric thermal environment surrounding very deep oceanic convection may be more conducive for relatively weaker updrafts to reach high altitudes (Kelley et al., 2010).Convective updraft velocity measurements in the UTLS are difficult and sparse, and extreme updraft strengths predicted by global storm-resolving models at 14 km altitude vary widely (Nugent et al., 2022).
A recent modeling study by Wu et al. (2023) found that in a warming climate, overshoots will increase much more over tropical oceans.Hong et al. (2005) and Aumann et al. (2018) also found warmer sea surface temperatures are correlated with more overshoots.Thus, documenting and understanding the spatial distribution of present-day cold point overshoots over ocean and land regions is important for predicting how overshooting convection may influence the UTLS in a changing climate.
Past studies using brightness temperature or −20 to −30 dBZ definitions of overshooting convection have been somewhat limited by relatively coarse data (Gettelman et al., 2002;Rossow & Pearl, 2007) or the twice-daily sun-synchronous sampling by the NASA A-Train (Li et al., 2022;Luo et al., 2008;Takahashi & Luo, 2014), which misses overshoots over land during the late afternoon diurnal cycle peak.Other studies used just 1 year of data (Iwasaki et al., 2010) or had only aggregate land-ocean statistics (Luo et al., 2008).
The goal of this paper is to achieve a climatology of the spatial distribution of cold point-overshooting convection across different land and ocean regions using a method that is generalizable to model output.Via calibration with active sensor data, we show more convincingly than prior studies that a combination of passive satellite brightness temperatures and reanalysis with frequent global coverage is now sufficiently accurate to develop the desired climatology.In 4 years of data, we find evidence of cold point-overshooting convection throughout the warmest, moistest parts of the tropics, with comparable frequencies over the Pacific warm pool and tropical land hot spots.Our proxy provides a useful observational benchmark for testing storm-resolving model simulations of tropical convective overshoot.

Data Sets
We use ice water contents (IWC) from DARDAR-CLOUD v3.10 (Delanoë, 2023).The DARDAR product joins retrievals from the CloudSat cloud profiling radar and CALIPSO lidar on the NASA A-Train.The data have a horizontal footprint of 1.4 km and a vertical resolution of 60 m (Cazenave et al., 2019;Delanoë & Hogan, 2008, 2010).We use brightness temperatures from the NCEP/CPC L3 Half Hourly 4 km Global (60S-60N) Merged IR V1 (GPM_MERGIR) data set, which combines infrared brightness temperature data from several geostationary satellites (Janowiak et al., 2017).The GPM_MERGIR data set has a horizontal resolution of 4 km and a temporal resolution of 30 min.

10.1029/2023GL105083
3 of 10 Temperature profiles come from the ECMWF Reanalysis v5 (ERA5) model level data, which is available hourly on a 0.25° × 0.25° grid with a vertical resolution of 300-400 m in the UTLS (Hersbach et al., 2017).The ERA5 reanalysis incorporates both Global Navigation Satellite System Radio Occultation (GNSS-RO) data and radiosonde measurements into the upper-level temperatures (Hersbach et al., 2020).We use reanalysis rather than the GNSS-RO or radiosonde data directly to have a collocated cold point temperature for each DARDAR retrieval.Overshooting convection occurs so infrequently that further restricting to near-coincident DARDAR and GNSS-RO data would insufficiently sample overshoots.
We consider two seasons: December-January-February (DJF) and June-July-August (JJA).Our DARDAR climatology is limited to 2007-2010, after which CloudSat only operated during the daytime.The ERA5 and GPM_ MERGIR data are mapped onto the DARDAR track by selecting the data point closest in space and time to each DARDAR retrieval.However, the mapped T b values may not represent the true brightness temperatures at each pixel since the GPM_MERGIR and DARDAR data could be offset by up to 15 min.

Analysis Regions
We focus our analysis on four regions of approximately equal size for each season (Figures 1a-1d).These locations (three oceanic, one land) were selected to enclose areas of frequent active convection based on the climatological mean precipitation rate.For DJF, the regions are Amazonia (AMZ), the southern Indian Ocean (IOS), and the South Pacific Convergence Zone (SPC).For JJA, the analogous regions are Africa (AFR), the equatorial Indian Ocean (IOE), and the West Pacific (WPC).We also consider the East-Central Pacific (ECP) as a control, since this area rarely experiences deep convection as intense as that over the warmest oceans (e.g., C. Liu & Zipser, 2005;C. Liu et al., 2007), despite having similarly high time-mean precipitation rates.Table S1 in Supporting Information S1 lists the coordinates of each region.Across all regions, there are approximately 6.5M retrievals in DJF and 8.4M in JJA.

Cold Point Tropopause
We define the cold point as the level of the minimum temperature in the hourly 0.25° ERA5 temperature profiles.The 0.25° grid is small enough to capture fine spatiotemporal variations in the cold point but large enough to avoid interpreting individual convective cloud tops as cold point fluctuations.The 2007-2010 time-mean cold points (see Figure S1 in Supporting Information S1) are higher (>17 km) and colder (∼191 K) in DJF than in JJA (altitude <17 km and temperature ∼194 K), consistent with previous studies (e.g., Kim & Son, 2012;Seidel et al., 2001).
Compared to GNSS-RO and radiosonde data, the ERA5 climatological tropical cold point is generally <0.5 K warmer and ∼150 m lower (Tegtmeier et al., 2020), although local variations may be larger.Hoffmann and Spang (2022) calculated an uncertainty of ±120-200 m in the lapse rate tropopause (LRT) height globally for ERA5 reanalysis.The ERA5 cold point, typically at or 0.5-1 km above the LRT (Munchak & Pan, 2014;Pan et al., 2018;Tseng & Fu, 2017), likely has a comparable uncertainty.In Section 3.3 we identify cold point overshoots by binning CloudSat-detectable echoes at intervals of 500 m relative to the local ERA5 cold point height.Echoes in the bin 500 m above the cold point presumably will lie above the true cold point despite these uncertainties.

Stratospheric Cirrus and Brightness Temperatures
To focus on cirrus associated with convective cold point overshoots, we bin DARDAR cloud ice by the difference between the brightness temperature (T b ) and the collocated cold point temperature (T CP ).This method is similar to that of Dauhut and Hohenegger (2022), who identified very deep convection in GSRM output by binning ice by the outgoing longwave radiation (OLR).Since brightness temperatures vary much more than cold point temperatures, large values of T b − T CP correspond to high T b and vice versa.We vertically bin the DARDAR data at cold-point-relative altitude levels.The ERA5 cold point height has a standard deviation of about 600 m within each analysis region, so using fixed altitude levels could incorrectly identify cold point overshoots.Following Pan and Munchak (2011), we instead use levels −500, 0, +500, and +1,000 m relative to the local ERA5-estimated cold point height.We interpret "stratospheric" cirrus detected by DARDAR in the +500 and +1,000 m bins as very likely to lie above the cold point, despite uncertainties in measurement, reanalysis, and collocation.
Figure 1 shows the bin-mean IWC and bin counts at the cold point-relative levels in the four analysis regions shown in the top row.In all regions, cirrus is occasionally detected at least 500 m above the cold point across most of the T b − T CP bins, corresponding to a broad range of brightness temperatures (bottom row).In fact, more stratospheric cirrus is collocated with high T b than low T b , which may seem counterintuitive.However, the coloring for overall count in each T b − T CP bin (below the thick black line) indicates that there are many more bins with high than low brightness temperature.Thus, an atmospheric column with high T b (i.e., relatively thin and/or low-lying cirrus) is much less likely to include cirrus above the cold point than one with low T b .
When T b − T CP is below ∼10 K, there is enhanced cloud ice (middle row; ≥5 × 10 −6 kg/m 3 ) at 500 m above the cold point in all regions except the ECP.At these low brightness temperatures, ice is almost always detected at 500 m above the cold point; that is, the stratospheric cirrus bin counts are almost the same as the total bin counts.Together, these bin means and counts suggest that some of the DARDAR-detected stratospheric cirrus may indicate overshooting tops in both land and warm ocean regions.
There is some seasonal variability in the bin means and counts (see Figure S2 in Supporting Information S1 for JJA), but the overall patterns between analogous regions in DJF and JJA are the same.One difference is that bin counts at the +500 m level are lower in the WPC (Figure S2f in Supporting Information S1) than in the SPC (Figure 1f), meaning cirrus above the cold point is rarer in JJA over the Pacific warm pool.The ECP (Figure S2h in Supporting Information S1) has more ice above the cold point in JJA than in DJF (Figure 1h), but still has much less than in other regions.

Identifying Cold Point-Overshooting Tops
Most of the detections of ice above the cold point in Figure 1 and Figure S2 in Supporting Information S1 are associated with thin cirrus clouds with high T b .For the low T b − T CP bins, the cirrus likely overlies deep convection.An actual overshooting top should not just have dense cloud ice but also larger ice particles lofted in the strong updraft that supports it; these particles should be detectable by the cloud radar.We therefore make binned plots restricted to include only radar-lidar pixels (Figure 2).
We anticipate that the dense cloud ice and low temperatures of overshooting tops should also cause particularly low brightness temperatures, possibly lower than the cold point temperature.The bottom row of Figure 2 shows that indeed, radar-detected ice above the cold point is almost exclusively associated with T b − T CP < 10 K, and at the coldest brightness temperatures, ice is almost always detected by the radar 500 m above the cold point.Thus we interpret radar/lidar-detected cirrus above the cold point as a cold point overshoot.We use the term "other stratospheric cirrus" for the remaining cirrus above the cold point that is detected only by lidar; some of this cirrus may be thin, while some of it may overlie extensive convective anvils or updrafts that do not overshoot the cold point.
Overshooting tops only comprise 1%-2% of all tropical stratospheric cirrus, as inferred by comparing the fraction of DARDAR cold point overshoot pixels containing ice at the +500 m level detected by both radar and lidar (Figure 2 and Figure S3 in Supporting Information S1) versus lidar-only detections.The occurrence frequency of other stratospheric cirrus (0.9%-2.5% in DJF, 0.4%-1.5% in JJA) is approximately 50-100 times larger than the convective overshoot occurrence frequency (0.01%-0.04% in both seasons).This relationship also holds for the ECP in JJA, but not in DJF when ice almost never occurs above the cold point there.

Brightness Temperature Proxy for Overshoots
Using the DARDAR data, we can infer a relationship between brightness temperature and cold point overshoots.If the relationship is strong enough, T b , which is much more broadly available than radar/lidar data, can be confidently used as an overshoot proxy.The reliability of this proxy will depend on the infrared opacity of the cloud extending above the cold point.We expect cold point overshoots to be ice-rich, but whether their T b over the size of a GPM_MERGIR pixel is necessarily less than the T CP is less obvious.
We calibrate the brightness temperature proxy by finding the fraction of DARDAR detections at a given value of T b − T CP that are associated with ice at 500 m above the cold point (top row in Figure 3).The probability of cold point overshoots steadily increases once T b − T CP falls below 10 K.In AFR, AMZ, and SPC, there is about a 50% chance of ice detected at 500 m above the cold point when T b = T CP (solid lines in Figure 3).The probability of radar-detected ice (dashed lines in Figure 3) is slightly lower than for ice from all detections and does not exceed 50% when T b = T CP .The pattern is the same in the IOS and IOE (Figures S4a and S4c in Supporting Information S1).In the ECP, the probability of cold point overshoots never reaches 50%, and the brightness temperature is rarely less than the cold point (Figures S4b and S4d in Supporting Information S1).In the WPC, the probability of cold point overshoots at +500 m never exceeds 30%, but does exceed 50% at slightly lower heights (green lines in Figure 3h).Unlike in other regions, the conditional probability of a convective overshoot over the WPC does not monotonically increase as T b − T CP becomes more negative; more investigation is needed to understand what other factors may enable T b < T CP with a low conditional probability of stratospheric cirrus in this region.Luo et al. (2008) argued that T b < T CP is an unreliable indicator of cold point overshoots because overshoots with (without) radar-detected ice above the cold point may have brightness temperatures warmer (colder) than the cold point.While this may be true for individual cases, our DARDAR results suggest that the T b < T CP proxy is a statistically unbiased and physically plausible threshold for convective overshoots that are often CloudSat-detectable.
The joint histograms of brightness and cold point temperatures in the bottom row of Figure 3 show the probability distributions for the entire GPM_MERGIR data set (i.e., not conditioned on DARDAR retrievals) in each study region/season.These probability distributions are similar to those from Gettelman et al. (2002) for the global tropics.Despite the differently-shaped distributions between land and ocean, the percentage of data points where T b < T CP (below the solid lines) is similar: 0.008/0.007% in AMZ/SPC and 0.024/0.015% in AFR/WPC.In both Indian Ocean regions, 0.004% of data points have T b < T CP , and there are even less in the ECP (0.0002% in DJF, 0.002% in JJA).These patterns suggest there are slightly more cold point overshoots over hot spots of tropical convection over land than over the warmest oceans, and that convective overshoots are rare over convectively active but slightly less warm ocean regions such as the Pacific Intertropical Convergence Zone.
Finally, in Figure 4 we apply this proxy to estimate the spatial distribution of cold point overshoots within our study regions over the entire 4-year GPM_MERGIR climatology, which past studies have not documented.With more data analysis, one could extend this analysis to look at the full global distribution and diurnal cycle of overshoots.Overshoots are 3-4 times more likely in JJA than DJF for analogous regions.Unsurprisingly, they are most common in AFR.Cold point overshoots are 30%-40% more frequent over land regions (AFR, AMZ) than the warmest ocean regions (WPC, SPC).On average, the regional frequencies of cold point overshoots for DJF are 0.011% in the AMZ, 0.008% in the SPC, 0.006% in the IOS, and negligible (0.0004%) in the ECP.In JJA, 7 of 10 they are 0.045% in AFR, 0.032% in the WPC, 0.016% in the IOE, and 0.014% in the ECP.The frequencies of cold point overshoots and convective cirrus are similar everywhere but the SPC.Convective cirrus are much more frequent for the SPC, which may result from enhanced ascent from the Brewer-Dobson Circulation in the Pacific warm pool during boreal winter that can help maintain anvils produced by cold point-overshooting convection.Such anvils could be thick enough to be radar-detectable but not so thick that T b < T CP .
These results corroborate past findings that cold point overshoots somewhat favor tropical land areas over oceans from Luo et al. (2008) and other studies that defined convective overshoot using low radar reflectivity thresholds.Even if we adjust the proxy to T b − T CP < −2 K following the 50% probability of radar-detectable ice in Figure 3, this pattern is robust.

Conclusions
We argue that T b < T CP is a suitable proxy to identify cold point overshoots that uses only high-resolution (4 km) IR brightness temperature from a geostationary satellite and ERA5-inferred cold point temperature.At such low brightness temperatures, we find that there is a high probability of cloud ice occurring above the cold point.Applying this proxy to 4 years of data over convectively active tropical land and ocean regions, we corroborate past findings that cold point-overshooting convection is only about 30%-40% more common over land (AFR and AMZ) than Pacific warm pool regions (WPC and SPC).The cooler East Pacific region (ECP) has very few cold point overshoots, and the Indian Ocean regions (IOE and IOS) fall somewhere in between.Even in the regions where cold point overshoots are the most frequent, thin cirrus above the cold point cover ∼100 times more area than the cold point-overshooting tops.By using more data than available to earlier investigators, we obtain seasonal maps of the overshoot frequency over our study regions that sample the full diurnal cycle and could easily be extended to the global tropics.
Our finding that cold point overshoots are comparably frequent between warm land and ocean areas does not contradict the widely-held view that convection is more intense over land.Oceanic convection has been well-documented to contain weaker and narrower updrafts (e.g., Fierro et al., 2012;LeMone & Zipser, 1980;Zipser & LeMone, 1980).Other classic measures of convective intensity where land dominates, including frequent lightning and tall 40 dBZ echo top heights (Zipser et al., 2006), need not always coincide with overshooting tops above the cold point.We expect that the cold point overshoots we have identified over land are more intense by these measures than those we have identified over oceans, but this does not necessarily imply less frequent cold point overshoots.
In a follow-up paper, we will adapt our T b < T CP proxy to global storm-resolving model (GSRM) output from the DYAMOND intercomparison (Stevens et al., 2019) to compare cold point overshoots between models and 10.1029/2023GL105083 8 of 10 observations.GSRMs explicitly simulate deep convection and have small enough horizontal grid spacing to capture convective overshoots.Understanding how well GSRMs can reproduce observed cold point overshoots in the current climate over both land and ocean will test the reliability of these models for simulating cold point overshoots and their influence on the UTLS in a warming climate.

Figure 1 .
Figure 1.Cold point-relative cloud ice binned by T b − T CP for all DARDAR pixels in DJF 2007-2010.(a-d) Maps of each analysis region; (e-h) bin-mean IWC at 500 m below to 1,000 m above the cold point height; and (i-l) bin counts conditioned on when DARDAR detects ice at each cold point-relative level.The row below the thick black line in (i-l) shows the total brightness temperature bin counts.The dashed lines in (e-l) mark where the brightness temperature equals the cold point.The bin width is 2 K.

Figure 2 .
Figure 2. As in Figure 1 for select regions in DJF and JJA, but restricted to pixels in which the radar and lidar both detect ice.

Figure 3 .
Figure 3. (a-d) Conditional probability of cold point overshoots (ice at 500 m above the cold point) as a function of T b − T CP for all retrievals and radar-detected ice.The horizontal dashed line indicates a 50% chance of cold point overshoots.(e-h) Joint brightness temperature-cold point histograms for all GPM_MERGIR data points.In all panels, the solid lines mark where the brightness temperature equals the cold point.The same regions as in Figure 2 are shown.

Figure 4 .
Figure 4. Frequency of cold point overshoots over time for (a) DJF and (b) JJA.