We use CloudSat observations of boreal summer tropical ocean cumuliform clouds to evaluate the behavior of the non-parameterized cumuliform clouds in the Nonhydrostatic Icosahedral Atmospheric Model (NICAM), with a particular emphasis on deep convective clouds (DCCs). The CloudSat cloud mask and radar reflectivity profiles for cumuliform clouds are sorted by large-scale environmental variables taken from the Aqua satellite and NCEP/NCAR reanalysis. The variables are total precipitable water (TPW), sea surface temperature (SST), and 500 hPa vertical velocity (W500), representing the dynamical and thermodynamical environment in which the clouds form. The sorted CloudSat profiles are then compared with NICAM profiles simulated with the Quickbeam CloudSat simulator. We first use the cloud mask to examine the transition between shallow clouds and deep clouds rooted in the planetary boundary layer. We find that NICAM simulates this transition fairly realistically. However, the transition occurs at slightly higher TPW and W500 values than the observations show. This may be indication of NICAM's inability to represent the formation of isolated narrow DCCs in marginally favorable environments. We then use simple metrics of the DCC-only radar reflectivity profiles (cloud top height, cloud top reflectivity gradient, maximum reflectivity) to quantitatively compare the observations with NICAM. The results show that while the observed and simulated results agree generally, there are some disagreements in key respects. There is disagreement on the sensitivity of cloud top height to environmental conditions and on the transition between shallow and deep clouds in environments marginally suitable for deep convection.
 The representation of clouds in atmospheric and climate models has been an important problem throughout the history of modeling. One particular aspect of this problem has been the modeling of cumuliform clouds within general circulation models (GCMs). Because the horizontal resolution of “traditional” GCMs is much larger than the horizontal extent of the largest individual deep convective clouds (DCCs), cumulus parameterizations must be used to represent these types of clouds. Many cumulus parameterizations treat cumuliform clouds as rapid restorative-adjustment elements of the climate system. That is, large-scale processes such as radiative cooling and advection tend to destabilize the troposphere over longer timescales (days to weeks), and convective processes, once “activated”, re-stabilize the troposphere on much shorter timescales (minutes to hours). Because of this, cumulus parameterizations are often formulated so that they provide a realistic large-scale balance to the destabilizing processes. This (among other reasons) raises an important question about the GCMs' ability to realistically simulate certain aspects of the climate system in different climatic situations that exists today. If the model gets the “right” answer in current climatic conditions for the “wrong” reason, how can it be trusted to get the “right” answer in different climatic conditions? It is apparent that cumulus parameterization is one of the major causes of climate simulation uncertainty [Randall et al., 2007].
 One possible solution to this is to simply remove the need for cumulus parameterizations, by increasing the horizontal resolution enough so that cumuliform clouds (at least the largest ones) can be simulated directly by the model's dynamical core. This is the approach used by Global Cloud Resolving Models (GCRMs). The Nonhydrostatic Icosahedral Atmospheric Model (NICAM) is an example of a modern GCRM [Satoh et al., 2008]. NICAM uses a high-resolution icosahedral grid to reduce horizontal grid spacing to a minimum of 3.5 km (though 7 km is commonly used). This is small enough to represent entire cumulonimbus clouds directly by the grid, though smaller-scale turbulent processes must still be parameterized. The high horizontal resolution has a significant cost in both vertical resolution (only 40 levels from 0 km to 40 km) and in the total model time. The longest model runs at 7 km resolution are only for a month, and the model runs for 3.5 km are only a week. So NICAM cannot yet be considered a true climate model, even though it has the potential to be one. But as computational power increases with time, NICAM and other GCRMs will become increasingly useful and important for the atmospheric and climate science community. Even now, with their limited ability, they are useful for examining certain aspects of local-scale/large-scale interactions important to climate, such as the Madden-Julian Oscillation (MJO) [e.g., Miura et al., 2007], so it is well-worth examining their behavior.
 NICAM has already been used for investigating important topics in atmospheric science. Perhaps most well known is the aforementioned work on the MJO. Simulating the MJO realistically has been a major difficulty for many atmospheric and climate models, and at least one likely reason for this is the difficulty of cumulus parameterizations [Maloney and Hartmann, 2001]. NICAM has demonstrated a capability to realistically simulate an MJO event already in progress, including a realistic propagation speed, the development of prominent features such as the westerly wind burst, and a reasonable size range for cumulonimbus clouds and mesoscale convective systems. Miura et al.  demonstrated that NICAM is capable of generating an MJO event spontaneously. These studies provide strong evidence that NICAM's non-reliance on cumulus parameterizations is very useful for the ability to handle an existing MJO realistically. NICAM has also been used to investigate the development and life cycle of a tropical cyclone [Fudeyasu et al., 2010]. The high horizontal resolution allows the investigation of the relationship between the individual convective cells within the tropical cyclone and the evolution of the cyclone-scale circulation. It is interesting to note that this relationship appears similar to the proposed “vortical hot tower” path to tropical cyclogenesis, where vigorous convective cells concentrate environmental vorticity and then transfer it to the cyclone-scale circulation [Montgomery et al., 2006]. Finally, NICAM has been used to investigate aerosol/cloud interactions for cumulus, stratus, and stratocumulus by Suzuki et al. . They used a modified version of NICAM, which included an aerosol transport model, to investigate the realism of aerosol effects in the GCRM. They found that the simulated aerosol effect on the microphysical properties (liquid water path and cloud droplet effective radius) of the aforementioned cloud types is similar to the aerosol effect estimated by satellite observations. While there is some disagreement in the absolute magnitude of the aerosol effect between the observations and NICAM, there is closer agreement between observations and NICAM than observations and traditional GCMs, as described previously.
 The A-Train suite of satellites has been used to investigate cloud properties from the observational side [L'Ecuyer and Jiang, 2010]. There is a large number of ways to analyze the observational data in a way that can be quantifiably compared with climate model output. One method that has been used frequently in recent research is “conditional sampling”: examining the sensitivity of cloud variables to changes in large-scale environmental variables. There had been a few earlier studies utilizing this method to examine the effect of large-scale variables on cloud radiative forcing (CRF). Ramanathan and Collins  used Earth Radiation Budget Experiment (ERBE) and in situ observations to estimate the relationship between sea surface temperature and CRF, and Bony et al.  did a similar study with simulated CRF and 500 hPa vertical motion, with the intent to separate dynamical and thermodynamical influences on CRF. CRF is an important quantity to consider for large-scale and global energy budgets, but this is only one important aspect of cloud behavior among many. The A-Train is particularly useful for doing this kind of analysis, given the large amount of available co-located data from the different sensors. The CloudSat [Stephens et al., 2002; Stephens et al., 2008] and the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIPSO) satellites are particularly important for cloud observational studies, as they can observe clouds in situations where traditional VIS/IR sensors cannot (e.g., multiple cloud layers), and can observe vertical profiles of cloud properties with higher spatial resolution than microwave sensors.
 One of the earlier studies using CloudSat and another A-Train member (Aqua) was that of Su et al. , which organized the tropical oceanic CloudSat vertical profiles of cloud water content by various large-scale environmental variables. This method, “conditional sampling”, was particularly useful for identifying the transition between shallow clouds and DCCs. In addition, their analysis provided some quantitative knowledge of variability in the vertical structure of cloud water content (CWC). This is an important piece of information about cloud properties that is not available from other cloud sensors and is useful for evaluating the cloud properties simulated in GCMs and other models. Su et al.  followed up on the initial study by comparing observations with two traditional GCMs. They found that the GCMs disagreed with each other as much as they did with the observations, particularly on the occurrence of mid-tropospheric clouds. Part of the disagreement was caused by “bogus” atmospheric states, that is, unrealistic combinations of environmental variables that bias the average cloud behavior in the GCMs.
Kubar et al.  used CloudSat and CALIPSO to examine the changes in cloud top height with 2 m air temperature and the vertical gradient of moist static energy. While this study only uses cloud top height as the cloud variable, as opposed to CWC vertical profiles, it should be noted that CloudSat and CALIPSO both have the ability to see through higher cloud layers and detect cloud layers below. This allows for a more reliable analysis of cloud top heights, particularly in convectively active large-scale regions where large cirrus shields can obscure low clouds from traditional cloud sensors. Del Genio et al.  used a similar approach to investigate the relationship between cloud top height and moisture in the MJO environment. They noted that the transition between shallow clouds and deep clouds is best understood in a statistical average sense—individual DCCs can form in a wide range of large-scale environmental conditions, depending on the local meteorology. Su et al.  provided useful insight on this topic, but CWC alone, without filtering for specific cloud types, yields an incomplete representation of all cloud processes. It is necessary to include multiple types of cloud variables for different cloud types in any comprehensive evaluation of simulated clouds in GCMs and other models. Finally, Forsythe et al.  examined changes in vertical cloud occurrence frequency as a function of total precipitable water anomaly (difference between instantaneous value and weekly climatological mean value). While environmental variable anomalies are a very useful measure in examinations of observational data, they are not practical to use for analyzing model output from modern GCRMs because of the relatively short time domain. They might be useful for analysis of GCMs, though.
 Conditional sampling should be a useful way for comparing model data with observational data. Direct comparisons between reflectivity profiles can be difficult because of the problems with simulating realistic attenuation of the radar signal in the reflectivity simulators. So, for example, a higher average maximum reflectivity in the NICAM DCCs versus the observations may not necessarily mean that NICAM DCCs are more vigorous than real life DCCs. It may rather be a problem with the reflectivity simulator's attenuation. This analysis method is one way around this problem—by looking at changes in reflectivity with environmental variables in the model and observations separately, and then comparing the model changes with the observed changes, we can remove the ambiguity caused by the conflation of attenuation errors and actual DCC characteristics in the model.
 CloudSat has been used previously to examine convective behavior in NICAM. Masunaga et al.  used CloudSat and the Tropical Rainfall Measurement Mission (TRMM) to compare the observed vertical reflectivity profile of tropical clouds in the MJO region with simulated reflectivity profiles from NICAM. They found that NICAM reflectivity profiles generally agree with the observed profiles, except in the frozen portion of DCCs where too much snow is produced. Inoue et al.  confirmed the excess snow production and found that the model can compensate for the excessive snow by altering other microphysical variables (e.g., the fall speed of snow). Satoh et al.  continued this study using a local CRM as a testing bed for microphysical parameterization schemes. They found that the realism of simulated vertical reflectivity profiles can be improved by implementing more sophisticated microphysical schemes. It should be noted that both of these studies use a limited sample of CloudSat data (11 overpasses of the active MJO region), so their results may not entirely represent the general characteristics of tropical DCCs. Though the previous work with NICAM convection provides substantial evidence that NICAM handles convection more realistically than traditional GCMs, there still seems to be a lack of attention on the behavior of the individual cumuliform clouds themselves. While individual convective cells were examined briefly in the work on tropical cyclones and the MJO [Fudeyasu et al. 2010], there has yet to be an attempt to evaluate the general behavior of cumuliform clouds under a variety of conditions. Such a study is needed for the same reason it is needed for GCMs—so that NICAM can be trusted to produce realistic results in conditions beyond the present day.
 This research project seeks to investigate the relationship of clouds and environmental variables in NICAM, specifically cumuliform clouds, in order to identify potential sources of error and uncertainty that may not be immediately obvious when examining larger-scale weather/climate features such as the MJO and tropical cyclones. This involves using the data to answer two principle questions: First, do the observed characteristics of DCCs according to this analysis method agree with what has been found by previous observational studies? And second, how well do the characteristics of NICAM DCCs agree with the observed characteristics?
2 Data and Methodology
 In order to quantitatively compare the cloud properties observed by CloudSat and those simulated by NICAM, the work of Su et al.  is used as a template on which this project is built. In that paper, the authors derive a relationship of vertical cloud structure and large-scale environmental variables (LSEVs) by binning CloudSat profiles into groups based on the co-located value of each meteorological variable. The meteorological variables are taken from a range of sources, including the Atmospheric Infrared Sounder, the Advanced Microwave Scanning Radiometer for EOS (AMSR-E), the Tropical Rainfall Measurement Mission observations, and the National Center for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) reanalysis data. The results display how the typical vertical structure for tropical oceanic clouds (displayed as CWC) varies as a function of LSEVs—precipitable water, sea surface temperature, etc. Because of the straightforward nature of the methodology, it should be useful for comparing the observational data with many types of model output, such as NICAM.
 To address the main questions of this paper, the properties of DCCs are divided into two main types: the transition between shallow clouds to DCCs as LSEVs change and the change of radar reflectivity characteristics of DCCs as LSEVs change. The transition between shallow clouds and DCCs has been studied quite extensively with both observation and simulation, as it is important for certain meteorological features like the MJO and various tropical waves [e.g., Straub and Kiladis, 2003, Peters and Bretherton, 2006, Del Genio et al., 2012]. Because of this, there is a relatively sizable pool of knowledge with which to compare the results of this paper. Knowledge of the variability of reflectivity properties in DCCs is less well represented in the literature, as it is a presently ongoing topic of research (Luo et al.  provide one example). So this paper will present information that is for now more difficult to corroborate with previous research (see section 2.2 for more details).
 To begin the study, the results of Su et al.  were replicated with a slightly modified methodology to better fit the goals of this project. The time domain is JJA 2006–2009 during western Pacific MJO events, and the spatial domain is the 15°S–15°N ocean-only region. The meridional domain is smaller than the one used in Su et al. (30°S–30°N) because of the possibility of interference by baroclinic waves within the larger Su et al. domain. Ocean-only is used for now because of the difficulty of the limited coverage of certain data types (e.g., water vapor) over land.
 For the cloud vertical structure information, only CloudSat is used. As part of the A-Train (except for a brief time in 2011, when technical issues temporarily forced CloudSat out of formation), CloudSat has a 705 km, 98° inclination orbit, which cross the equator at 01:30 A.M./P.M. local time. CloudSat follows the Aqua satellite within 2 min, minimizing time lag discrepancies. CloudSat's primary instrument is the Cloud Profiling Radar (CPR), a 94 GHz radar with a 1.1 km wide effective footprint and 480 m vertical resolution, oversampled to create a 240 m effective vertical resolution [Stephens et al., 2002]. The specific CloudSat products used are the 2B-Geoprof Radar Reflectivity and 2B-Geoprof Cloud Mask.
 The three LSEVs used presently are total precipitable water (TPW), sea surface temperature (SST), and 500 hPa vertical velocity averaged over 2° × 2° horizontal areas (W500). TPW and SST data are taken from the AMSR-E instrument onboard the Aqua satellite [Kawanishi et al., 2003], specifically version 5 of the AMSR-E ocean algorithm [Wentz and Meissner, 2000; data available at http://www.ssmi.com]. W500 data are taken from the NCEP/NCAR reanalysis [Kalnay et al., 1996]. TPW is binned by 3 mm, SST is binned by 1K, and W500 is binned by 1 cm s−1. The AMSR-E data must be spatially co-located with CloudSat, but temporal co-location is not necessary because of the short (~1 min) lag between CloudSat and Aqua overpasses. For missing AMSR-E data caused by clouds, a simple linear interpolation is used to fill the missing values. The NCEP reanalysis data must be spatially and temporally co-located with CloudSat, as the data exist only at 6 h resolution. No interpolation methods were applied.
 It should be noted that these LSEVs are not entirely independent of each other. In particular, TPW and SST are highly correlated [Raval and Ramanathan, 1989; Inamdar and Ramanathan, 1994; Stephens, 1990]. This is a direct result of the Clausius-Clapeyron relationship: warmer air temperature allows for higher water vapor content. There is a more subtle connection between TPW/SST and W500—large-scale ascent tends to be associated with convectively active regions and corresponding lower tropospheric convergence, which in turn require relatively high SST and TPW values. Joint PDFs of the three LSEVs from the observed/simulated data used in this paper (not shown) demonstrate these correlations, though the reanalysis W500 is somewhat less strongly correlated with observed TPW/SST than in the simulation. This is likely a result of a less direct coupling between observation/reanalysis than simulation/simulation. In future studies, it might be useful to examine more sophisticated diagnostic variables which are dependent on these LSEVs (e.g., convective available potential energy). But for this paper, we choose to start with the basic LSEVs so that future research has a platform upon which to build.
 Although it is possible to include CALIPSO data to augment the 2B-Geoprof Cloud Mask (increased sensitivity to low clouds and thin cirriform clouds), for now it is not used in order to better facilitate meaningful comparison with NICAM data via the CloudSat Simulator. CALIPSO is commonly used in conjunction with CloudSat because of their complementary capabilities. CloudSat is more useful for observing clouds with higher CWCs and/or deeper vertical extents, while CALIPSO is more useful for examining low-CWC clouds like subvisible cirrus, as well as boundary layer clouds which the CPR has trouble detecting. However, the authors do not have a CALIPSO simulator through which to interpret NICAM output. Furthermore, this study primarily focuses on the properties of DCCs, which are easily observed by CloudSat. Though CloudSat may have difficulty in observing the fringes of convective anvils, we do not examine these specific features of DCCs. For these two reasons, CALIPSO data are not used in this paper.
 The NICAM simulation, set in early July 2006, included a significant MJO event in the western Pacific. The existence and phase of the MJO may be important in influencing the relationship between clouds and LSEVs in the tropics, so it is reasonable to account for the MJO in the observational data in order to avoid discrepancies between simulation and observations related to the MJO. The temporal domain of the observational data is chosen as the time periods during JJA in which the phase of the observed MJO was close to the phase of the MJO simulated in NICAM. The JJA season is used because the NICAM simulation is set in July. The simulated MJO corresponded with an observed MJO identified as a high-amplitude phase 7 event, according to the Real-time Multivariate MJO Index as defined by Wheeler and Hendon . Phase 7 is defined as the time when the deep convective portion of the MJO has moved east of the Maritime Continent into the western Pacific Ocean. Periods in the CloudSat data record that correspond to a high-amplitude (greater than one) phase 7 MJO during JJA are 3 July–16 July 2006, 7 July–13 July 2007, and 8 June–11 June 2009 (there was no significant MJO phase 7 event in JJA of 2008). This time domain will be called “JJAphase7”, and it is time domain used for all observational data shown in the figures for this paper. For comparative purposes, the full JJA record for 2006 (begins on 15 June because of early CloudSat data availability issues), 2007, and 2009 (called “JJA”) was also run through the same procedures as the JJAphase7 dataset, for the purposes of examining differences between the average summer cloud characteristics and MJO-modified characteristics.
 This process was repeated for NICAM output. We use the NICAM run from Suzuki et al. . The time domain is 1–8 July 2006, though this project uses data from only the last 4 days of the simulation. The model output time resolution is 3 h, and the horizontal resolution is 7 km. In order to compare NICAM output with the observational data, the NICAM output is partitioned into a series of meridional cross sections extending from 15°S to 15°N. CloudSat crosses the tropics at an angle close to meridional (~8° from N), so meridional cross sections are reasonably close to CloudSat cross sections. In order to minimize the influence of the diurnal cycle, cross sections are only taken for each time step within 45°-wide zonal domains centered on the longitude corresponding with 01:30 A.M. and 01:30 P.M. local time for that time step (e.g., the 9z time step would have zonal domains centered around 67.5°E and 112.5°W). Fifty cross sections are taken within each zonal domain at each time step. Figure 1 shows the meridional domain of all the data, zonal domain of the data for one NICAM time step, and the orientation of the data cross sections for one time step. The meteorological variables in the cross sections include temperature (skin and atmospheric profile), specific humidity, pressure, TPW, surface precipitation, horizontal wind components, 500 hPa vertical wind, cloud liquid mixing ratio (MR), cloud ice MR, rain water MR, and snow MR.
 It should be noted that the NICAM simulation used by Suzuki et al.  included an aerosol model, which is not part of the “standard” NICAM settings. This addition might cause a slight change in behavior of convective clouds in comparison with other NICAM simulations because of changes in the convective microphysical characteristics. However, because the simulated aerosol effect on convective precipitation was smaller than the observed effect, the aerosol effects on other convective characteristics are likely small.
 The hydrometeor variables are used to generate a cloud mask via the Quickbeam CloudSat Simulator [Haynes et al., 2007], where currently a value of −30 dBZ is used as a minimum threshold for identifying a cloud. The microphysical variables used in the Quickbeam simulation (e.g., drop size distribution, number concentration) were the same as those used in the Suzuki et al. NICAM simulation itself for consistency. The simulated attenuation is not included in the simulated reflectivity, because of the unrealistic behavior of attenuation for large melting hydrometeors. The simulated DCC reflectivity may also demonstrate some unrealistic behavior in regions of heavy precipitation because of problems simulating multiple scattering by large rain drops [Haynes et al., 2009]. Only data over oceans are used, for consistency. Instead of CWC used as the cloud variable as in Su et al. , this project currently uses cloud occurrence and radar reflectivity as the cloud variables. These variables were chosen because of the unreliability of the current 2B-CWC CloudSat product in conditions of heavy precipitation (discussed by Su et al. ). Because this project has a focus on large cumuliform clouds, such a limitation in the data renders it unsuitable.
 This paper seeks to investigate the cloud-LSEV relationship of cumuliform clouds in particular, so it is desirable to reduce the influence of non-cumuliform on the results (e.g., cirrus). One simple way to do this is to include only clouds with a base within or near the planetary boundary layer (PBL). Based on the criteria used in Del Genio et al. , we remove all vertical profiles that do not contain a cloud with a base below 2000 m (we will refer to the clouds rooted in the PBL as PBL-R clouds). Of the remaining profiles, we remove all cloudy bins in each profile that are not connected continuously with the PBL-R clouds. The remaining profiles are then sorted by each LSEV, and cloud occurrence frequency (COF) vertical profiles are calculated. COF is calculated by summing the number of cloud occurrences in each vertical bin for each LSEV bin, then dividing by the total number of PBL-R cloudy profiles for the LSEV bin. This method allows the examination of the sensitivity of cloud vertical growth to the LSEVs without other factors such as changes in cloud cover with LSEV obscuring the relationship. It should be noted that this does not remove PBL-R stratiform clouds from the data. The major stratiform cloud type that may influence these results are the nimbostratus regions of MCSs, and this should be kept in mind when interpreting the deep cloud regions of the various cloud frequency data plots. All the presented COF data have had this procedure applied. The results are shown in Figures 3, 5, and 6.
 To narrow the results further to the properties of DCCs, a cloud identification scheme is necessary that is compatible with both the observational and simulated data. Previous research on the properties of DCCs observed by CloudSat has used the 2B-Cloud Classification product [Wang and Sassen, 2007] in order to identify deep convective clouds and systems. The 2B-CLDCLASS product uses a complex algorithm to determine cloud type from the observed cloud's height, vertical and horizontal extent, reflectivity properties, existence of precipitation, and temperature (the latter is taken from ECMWF data). Such an algorithm would be very difficult to apply to NICAM output without extensive modification, so a highly simplified algorithm designed to detect only DCCs is desirable. The algorithm used in this paper is based on the tentative cloud classification rules presented in Wang and Sassen [2007, Table 2]. A vertical profile (both CloudSat and NICAM) containing a cloud is identified as containing a DCC if (a) the cloud vertical extent as identified by the cloud mask is greater than 6 km, and (b) the reflectivity values for the layer of atmosphere between 15°C and −15°C are greater than −5 dBZ. The vertical range of the minimum reflectivity has been slightly altered from the range given in Wang and Sassen  (originally −20°C to 25°C) in order to improve agreement between DCCs identified by this algorithm and DCCs identified by 2B-CLDCLASS.
 This algorithm consistently agrees 2B-CldClass when identifying the “obvious” DCCs—that is, clouds that are several kilometers tall and have high reflectivity values (greater than 5 dBZ) extending to within 2 or 3 km of the cloud top. However, this algorithm has difficulty distinguishing between DCCs and other deep clouds which 2B-CldClass identifies as nimbostratus. Fortunately, in the deep tropics, these types of clouds are rare, as even the “stratiform” regions of MCSs usually contain high-enough reflectivity values over a large-enough vertical extent to be classified as DCCs by 2B-CldClass. The most common disagreement between this algorithm and 2B-CldClass arises at the edges of large DCC clusters where reflectivity values can be less than −5 dBZ throughout much of the cloud. This occurs because 2B-CldClass assigns cloud type to entire cloud clusters, and this algorithm assigns the DCC type only to individual vertical profiles. Nevertheless, the most important thing for this paper is that the DCC identification process is consistent for both observational and simulation data. The vertical reflectivity profiles are sorted by each LSEV and averaged as with the COF diagrams; these are referred to as deep convective cloud average reflectivity (DCCAR) diagrams. The results are shown in Figures 8-10.
 In order to better quantify the relationship between reflectivity and meteorological variables, it is useful to create indices from the reflectivity data: cloud top height (CTH), maximum reflectivity (MAXREFL), and the distance between the −5 dBZ echo top height (ETH) and CTH (we call this difference N5ETHD). N5ETHD is used in this study to represent the upper DCC vertical reflectivity gradient. An increasing N5ETHD means decreasing vertical reflectivity gradient. While previous studies [e.g., Luo et al., 2011] use 0 dBZ and 10 dBZ ETHs for examining CloudSat data, this study uses the −5 dBZ ETH because of the microphysics-related disagreement between CloudSat and NICAM reflectivity in the upper troposphere discussed previously. This disagreement occurs at 0 dBZ, and less so at −5 dBZ (see Figure 6 and section 3.2 for discussion about this disagreement), so −5 dBZ is a more useful value to use in a comparative study. CTH in DCCs is a complex function of thermodynamic instability, the level of neutral buoyancy, and the entrainment of environmental air. In general, it will tend to increase as thermodynamic instability increases, but this is not a rigid relationship [Luo et al., 2011]. MAXREFL is a function of droplet size and concentration, which in turn is thought to be affected by the updraft intensity. N5ETHD also is thought to be related to the updraft intensity, as stronger updrafts are capable of lofting larger cloud hydrometeors higher into the upper troposphere, bringing them closer to the cloud top.
 It should be noted that the exact relationship between DCC reflectivity profiles and cloud-scale vertical velocity is a topic of ongoing research. While there are reasonable qualitative arguments for expecting certain relationships between reflectivity and vertical velocity, the science has not yet advanced to the point that we can directly cite peer-reviewed research to corroborate this speculation with quantitative analysis. The most relevant research presently is that which has been done with precipitation radars. Zipser and Lutz  found that both the maximum DCC reflectivity and cloud top reflectivity gradient from precipitation radars tend to increase as convective vertical velocity increases (though these results may also be related to other changes in DCC behavior related to land/ocean contrasts). The increased cloud top reflectivity gradient occurs because the high reflectivity values within the DCC occur over a greater depth as a result of large hydrometeors being lofted higher into the upper cloud. Liu et al.  found using TRMM data that higher echo top heights for significant reflectivities (e.g., 20 dBZ), and thus a large cloud top reflectivity gradient, are likely associated with more vigorous convection. In fact, this may be a more reliable indicator of vigorous convection than cloud top brightness temperature, a more traditional measure of convective intensity. But because of the limited knowledge about CloudSat reflectivity relationships, the reflectivity results will be presented here without an attempt to associate them with physical processes in the atmosphere. A comparison of reflectivity indices versus maximum updraft velocity in NICAM (not shown) shows no trend in updraft intensity with increased CTH, an increasing trend of updraft intensity with increasing MAXREFL, and a decreasing trend of updraft intensity with increasing N5ETHD (stronger updrafts tend to occur with more packed cloud tops).
 From the changes of these three reflectivity variables with LSEVs, we calculate linear trend lines to estimate the average changes across each LSEV. Because of the noisiness of the data on the tails of the sample PDFs, only the central 95% of the data (between the vertical white lines in Figures 6-9) are used in the trend line calculations. A “significant” result refers to a trend line slope value within the sufficient data range (between the white dashed lines) that is statistically significant according to the 90% confidence one-tailed Student's t-test. The 90% confidence limit is used (as opposed to higher limits) because it provides a roughly even split between the number of significant and non-significant results, so that a meaningful comparison between observations and simulation is possible (e.g., to avoid the trivial “finding” that all relationships are insignificant because confidence limit is too high). The trend line values and significance test results are summarized in Table 1.
Table 1. Trend Slope Values and Significance Test Results for the LSEVs by Dataseta
aListed are the statistics by LSEV (TPW, SST, W500, rows) and by dataset (JJAphase7, JJA, NICAM, columns). Each statistic includes a trend slope value and a yes/no indicator for whether the trend slop value passes the 90% one-tailed t-test.
CTH (m mm−1)
N5ETHD (m mm−1)
MAXREFL (dBZ mm−1)
CTH (m K−1)
N5ETHD (m K−1)
MAXREFL (dBZ K−1)
CTH (m m−1 s)
N5ETHD (m m−1 s)
MAXREFL (dBZ m−1 s)
 We also conducted a sensitivity experiment for the definition of DCC by increasing the minimum DCC height to 10 km. A height of 6 km is a rather low altitude for DCC top height in the tropics and would include both growing DCCs and a small number of cumulus congestus. The results of the 10 km analysis are very similar to the 6 km analysis, with the exception of MAXREFL for TPW, so the results would not be discussed directly other than this one exception.
3 Cloud Occurrence Frequency
 Before examining the results, it would be useful to discuss what we might expect to find based on previous research and the general principles of atmospheric physics. It is reasonable to expect deep clouds to become more frequent as TPW increases, because DCCs require relatively high moisture to develop (with the exact value depending on both local climatic and meteorological conditions). Developing DCCs and cumulus congestus also moisten the environment, conditioning the atmosphere for deeper convection [Thayer-Calder and Randall, 2009]. PBL moisture is an important component of CAPE, which in turn is necessary for overcoming the atmosphere's resistance to large vertical motions. Also of particular interest is the idea of mid-tropospheric humidity acting as a control on vertical cloud growth [e.g., Jensen and Del Genio, 2006]. Holloway and Neelin  found that most of the variability of TPW (i.e., column water vapor) is directly related to water vapor variability above 850 hPa (PBL water vapor is strongly tied with surface interactions and so has smaller variability). If this is true, then larger values of TPW should often correspond with increased mid-tropospheric humidity and thus increased vertical cloud growth.
 We would also expect deep clouds to become more frequent as SST increases [Zhang, 1993]. SST corresponds closely with PBL temperature, which is another important control on CAPE. There is also an association of SST and SST gradients with low-level moisture convergence [Lindzen and Nigam, 1987]. This, as previously described with regard to TPW, is conducive for deep convection. Previous studies have found a close association with convective activity with SST [e.g., Bony et al., 1997], though there is some question to the precise relationship in the vicinity of tropical warm pools [Liu and Moncrieff, 2008].
 Finally, we would expect deep clouds to become more frequent as W500 increases, particularly when it becomes positive. As described in Bony et al. , it can be difficult in practice to separate the effects of large-scale vertical velocity and thermodynamics in controlling large-scale convective activity; it is a “chicken and egg” problem. However, a basic argument for the relationship between DCCs and W500 is that subsidence warming (lifting cooling) increases (decreases) vertical stability and thus suppresses (enhances) deep convection. Also, because of mass continuity, middle tropospheric ascending (descending) motion is associated with convergence (divergence) in the lower troposphere. This convergence/divergence may not necessarily be located within the PBL itself, so we would not expect any correlation between W500 and PBL convergence to be one.
 It would be useful to know the average vertical COF for observations and NICAM before applying the conditional sampling technique. Figure 2 shows the vertical COFs for the tropical oceans for four sky conditions: all-sky, cloudy, PBL-R cloudy, and all sky with all clouds but PBL-R clouds removed. For all-sky conditions, NICAM produces too many shallow clouds and too few mid-tropospheric clouds relative to observations. For cloudy conditions, NICAM produces shallow clouds almost constantly, while CloudSat observes shallow clouds only half of the time. NICAM also produces too many upper tropospheric clouds and too few mid-tropospheric clouds. For PBL-R-only cloudy conditions, NICAM produces too many shallow PBL-R clouds and too few deeper PBL-R clouds. This suggests that PBL-R clouds do not grow vertically as tall and/or as frequently in NICAM as they are observed to do. Therefore, we should expect to see a deficit in vertical development of PBL-R clouds when conditional sampling is applied.
 Figure 3 shows the COF plots and sample PDFs for TPW. The PDFs of both all-sky and PBL-R cloudy conditions are shown in order to illustrate the effect of TPW on cloud cover (the ratio of the PBL-R cloud PDF to the all-sky PDF). The observed and NICAM sample PDF curves are shaped quite differently, with NICAM having a roughly Gaussian shape with a maximum value at 36 mm and the observed PDF being weighted more heavily at high TPW values with a maximum at 57 mm. For both observed and simulated data, the shape of the sample PDF does not change much between the all-sky and PBL-R cloudy conditions, but the PBL-R cloud cover increases noticeably for TPW values above 51 mm.
 The difference in the shape of the observed and simulated sample PDFs is not easily explained with the results we have presented here. However, we can rule out the possibility of NICAM being unable to simulate PBL-R clouds realistically. Figure 3c clearly shows that the disagreement in PDF shape for PBL-R cloudy conditions is caused primarily by a more fundamental disagreement between NICAM and observations of the distribution of TPW in all-sky conditions, irrelevant to clouds. It should be noted that Su et al.  showed a TPW all-sky sample PDF that was also skewed towards higher TPW values, so it should not be assumed that the skewness shown here is simply a methodological artifact. Further analysis of why NICAM behaves so unrealistically is beyond the scope of the paper, but it would be a topic worth investigating.
 The observations show primarily lower tropospheric clouds for low TPW values and a progressive transition to deeper clouds starting at approximately 45–48 mm. This corresponds with results from Su et al. , where the CWC signature for deep clouds begins just below 50 mm, and with Del Genio et al., 2012, where the typical cloud top height begins increasing at 48 mm. This also agrees with tropical precipitation studies [e.g., Holloway and Neelin, 2009], where the non-drizzle rain signature (which requires vertical cloud growth) begins around 50 mm TPW. The results make reasonable physical sense: as stated previously, we would expect DCCs to become more frequent as TPW increases.
 The NICAM data show a similar progression of COF with increasing TPW. However, the simulated clouds begin transitioning to deep clouds at a higher TPW value of 51–54 mm. A possible explanation for this may be the limited ability of NICAM to simulate deep convection with its horizontal resolution. In the real atmosphere, it is possible for isolated narrow DCCs to form and grow in marginally suitable large-scale conditions, which would inhibit larger DCCs and large-scale convective initiation. But NICAM has a 7 km horizontal resolution, which is too large to realistically simulate these narrow DCCs. Though NICAM can resolve a 7 km or 14 km wide cloud, it cannot resolve the smaller-scale convective features (e.g., localized updrafts and downdrafts) that are important for building/maintaining observed DCCs. So convective initiation of any sort in NICAM might not occur until the environmental conditions become more than just marginally suitable. There are odd “spike”-like features on the ends of some of the plots, e.g., below 15 mm of the TPW panel. These features are likely numerical artifacts caused by the extremely low number of observations in the outlying regions. There are few observations below 15 mm TPW, so the frequency plot in this region is likely unreliable.
 It is difficult to explain the deficit in deeper PBL-R clouds in NICAM through simple causality, because there is a “chicken and egg” problem associated with clouds and their environment. In this case, on one hand, increased environmental moisture helps clouds develop vertically. On the other hand, cloud growth tends to moisten the surrounding environment, at least until clouds grow large enough to precipitate and induce subsidence drying in the environment. One possibility is that there is a dry bias in the middle troposphere in NICAM. As mentioned previously, middle tropospheric moisture is an important control in the development of cumulus congestus and DCCs. If the middle troposphere is too dry, deeper cumulus clouds will have more difficulty developing. And if deeper cumulus clouds have difficulty developing, they will be less effective in moistening the middle troposphere.
 To test for this possibility, we have applied the conditional sampling technique used for the vertical COF profiles to vertical specific humidity profiles. For the “observed” specific humidity profiles, we use the CloudSat ECMWF-Aux product [Partain, 2004], which is data from the European Center for Medium-Range Weather Forecasts (ECMWF) global atmospheric model, co-located to CloudSat data in space and time (a description of the ECMWF model is available at http://www.ecmwf.int/products/forecasts/guide/The_ECMWF_global_atmospheric_model.html). We are particularly interested in comparing the “observed” and simulated specific humidity profiles in atmospheric conditions marginally suitable for vertically developing tropical oceanic PBL-R clouds. “Marginally suitable” is defined as the environmental conditions in which the observations show the transition between shallow and deep clouds occurring; in this case, Figure 3 shows this region to be between 45 mm and 60 mm of TPW. The results, calculated as NICAM humidity minus ECMWF humidity divided by the number of ECMWF samples, are shown in Figure 4. We find that for tropical ocean atmospheric profiles where the TPW value is between 45 mm and 60 mm, NICAM has a dry bias of 5–15% between 2–5 km altitude. We hypothesize that this dry bias is related to the deficit of deeper PBL-R clouds in NICAM discussed previously. Preliminary researches by other NICAM researchers have found a similar dry bias, and they speculate that this is related to misrepresentation of shallow/congestus clouds due to sub-grid convection [A. Noda and M. Satoh, personal communication, 2012]. In the future, it may be possible to remove this dry bias through the use of a sub-grid convective parameterization for shallow and cumulus congestus clouds.
 Figure 5 shows the COF plots and number of samples for SST. The observed and NICAM sample PDFs have roughly the same shape and mean, with the observations having slightly thicker tails than NICAM. It should be noted that NICAM SST is taken from observations, and is not a prognostic variable, so the similarity between the observed and simulated PDFs is not surprising. PBL-R cloudy cover does not appear to change significantly with increasing SST for either observations or simulation. One might expect for PBL-R clouds to become more frequent as SST increases because of increased PBL instability, but these results do not support that idea.
 The observations show the shallow clouds transitioning to deep clouds beginning at 298K, with an abrupt increase in the maximum cloud height (i.e., transition to DCCs) occurring at 300K. This is consistent with Su et al. , which shows the deep cloud CWC signal beginning at 300K. Kubar et al.  found a similar pattern of a gradual increase in CTH with increasing SST followed by an abrupt increase in maximum CTH. However, they found the abrupt transition at 298K, not 300K. It should be noted that their domain was a cross section of the eastern tropical Pacific, not the global tropical ocean. However, this does not fully explain why this paper does not find a stronger DCC signature at 298K. It is possible that there is a disagreement in SST measurements between the AMSR-E product used here and the ECMWF-YOTC product used by Kubar et al. .
 NICAM data show a pattern of slow increase in CTH with SST followed by the abrupt transition at 301K, slightly warmer than observations. However, the COF values for middle and upper tropospheric clouds does not increase above 25%, unlike the observed COF values. This indicates that DCC occurrence is not as strongly coupled with SST in NICAM as in the real atmosphere.
 As with the TPW case, NICAM shows a mid-tropospheric dry bias for regions with SST values marginal for convection (Figure 4). This is not surprising, given the high correlation between SST and TPW.
 Figure 6 shows the COF plots and number of samples for W500. The reanalysis and simulated PDFs have a roughly Gaussian shape, though with the observed PDF being noisier. Interestingly, the observed PDF has a maximum W500 probability at 0.002 m s−1, while NICAM has a maximum value at −0.005 m s−1. It appears that the NICAM deep tropics possesses stronger and more common subsidence than the reanalysis show. Su et al.  showed a reanalysis preference for subsidence, as well. This apparent discrepancy in the reanalysis results is not caused by the presence of a strong phase 7 of the MJO, which is associated with enhanced large-scale ascent over the western Pacific, as a similar sample PDF is found for the full JJA dataset (not shown). However, the reanalysis sample PDF for all profiles (cloudy and clear sky, not shown) shows a preference for subsidence similar to NICAM and Su et al. . This suggests that PBL-R clouds have at least a weak association with mid-tropospheric ascending motion in the real atmosphere, which is not simulated appropriately in NICAM. For both observed and simulated results, the PBL-R cloud cover increases slightly with increasing W500 for values above 0.01 m s−1.
 It appears that DCCs exist in both ascending and descending large-scale regimes for the observations/reanalysis (note the prevalence of the bluer colors for negative w). However, DCCs become more frequent with increasing w beginning at the transition between descending and ascending motion (w = 0). DCCs become more frequent with increasing w until they reach mature DCC height at 0.015 m s−1. This corresponds to the results in Su et al. , which shows the deep cloud signal becoming prominent at neutral vertical velocity. Bony et al.  found from ERBE data that the transition from a shallow cloud radiative regime to a deep cloud radiative regime begins at about 20 hPa/day, roughly −0.004 m s−1. This is reasonably close to the w = 0 value found by this paper, and note that Bony et al. did not exclude cirriform clouds from their calculations.
 NICAM shows no deep clouds forming in the descending regime—all deep clouds exist in the ascending regime. As with TPW, a possible explanation for this is that the 7 km horizontal resolution does not allow for isolated narrow vigorous DCCs to develop in environmental regimes unfavorable for large-scale convection. The simulated DCCs may be too large and weak to overcome the large-scale subsidence and associated stability. Furthermore, the frequency of deeper clouds does not increase as quickly with increasing w as the observations/reanalysis show. As with SST, this is an indication that DCC frequency is not as strongly associated with increases in w (above w = 0) as in the reanalysis.
 It is possible that the observational results have been contaminated with errors in the NCEP/NCAR reanalysis data. Much of the spatial domain for this study covers the remote ocean, where observational wind data are sparse, and thus, the reanalysis relies mostly on model estimates. However, in order to fully explain the weak DCC-W500 relationship mentioned above, the NCEP/NCAR reanalysis would have to have a routinely inaccurate estimate of W500 by at least 0.01 m s−1. This possibility cannot be ruled out by the results presented here, but investigating this possibility further would be beyond the scope of this paper.
 As with TPW and SST, NICAM shows a mid-tropospheric dry bias for regions with W500 values marginal for convection (Figure 4). As mentioned previously, W500 is significantly correlated with SST, though not as highly as SST with TPW, so this result is unsurprising.
4 DCC Average Reflectivity
 Before examining the DCC radar reflectivity versus meteorological variable plots, it is useful to examine the average vertical reflectivity profile of DCCs. Figure 7 shows the contoured frequency by altitude diagrams (CFADs) of reflectivity for all observed and simulated DCCs. The observed DCC CFAD shows the characteristic “arc”-shaped reflectivity profile seen in previous studies [e.g., Masunaga et al., 2008], with a reflectivity maximum of 10–15 dBZ at 4–5 km. The large reflectivity below 0.5 km is contamination from the surface return. The observed local minimum in reflectivity at ~5 km is a “dark band” caused by the melting of larger frozen hydrometeors, which enhances attenuation of the radar signal [Sassen et al., 2007]. The increased reflectivity just below the dim band is the bright band from the recently melted rain drops, though the signal is much weaker for the CPR than it is for precipitation radar. In comparison, the simulated CFAD does not show a strong arc-shaped profile below 5 km. This is because of the non-use of the attenuation by large hydrometeors in QuickBeam [Haynes et al., 2007].
 The most notable contrast between the simulated and observed CFADs is the abrupt reduction in reflectivity values starting at ~8 km. This is primarily a problem caused by the Grabowski microphysics scheme used in the NICAM simulation. The Grabowski scheme has only two microphysical variables—cloud water and precipitation—and the water phase (cloud ice/water, snow/rain) is dependent only on temperature. Importantly, there is no graupel included. Observed DCCs have a significant amount of graupel in the freezing portion of the cloud. Because the Grabowski scheme does not produce graupel, the simulated DCCs instead produce unrealistically large amount of snow in the freezing portion of the cloud [Satoh et al., 2010]. Snow has significantly different reflectivity characteristics from graupel, and this causes the abrupt reduction in simulated reflectivity at ~8 km. The missing dark band is caused by the absence of attenuation in Quickbeam.
 Figure 8 shows the DCCAR plots and number of samples for TPW. Table 1 lists the trend slope values and significance test results for TPW and the other LSEVs. The observed and NICAM sample PDFs both have a near-Gaussian shape, with similar means and variances. NICAM has a slightly higher mode (63 mm) and variance compared with the observations (60 mm). Unlike the difference in the PBL-R cloudy PDFs, the difference between the DCC PDFs cannot be explained simply by the difference in all-sky PDFs. If we use the reasoning discussed in section 3.1, we would expect the NICAM DCC PDF to have a lower mean and mode compared with observations because the NICAM all-sky PDF has a lower mean and mode than the observed PDF. This suggests that the disagreement must be related to the simulated DCC behavior in some manner.
 The most prominent feature for observed TPW is the near-monotonic increase in the cloud top height with increasing TPW value. This corresponds with a significant trend line slope of 89 m mm−1. This result is consistent with the full JJA dataset. The observed N5ETHD shows a non-monotonic decrease, in which it increases from 51 mm to 60 mm and the decreases to 69 mm. This results in a decreasing significant trend line slope of -83 m mm−1. This result is also consistent with the full JJA dataset. The observed MAXREFL has no significant trend line slope with the value decreasing from 51 mm to 57 mm and increasing to 69 mm. It should be noted that in the full JJA dataset, there is a non-monotonic increase in MAXREFL with TPW, with a significant trend of 0.11 dBZ mm−1.
 The NICAM CTH does not change with TPW, which is curiously very different from the observations. In fact, it appears that DCC CTH is insensitive to all LSEVs. It may be the case that in NICAM, CTH is not as variable as it is in the real atmosphere. If this is true, this may cause problems for using NICAM to investigate other aspects of tropical convection, such as convective penetration of the tropical tropopause layer, which is important for troposphere/stratosphere interaction [Luo et al., 2008]. The NICAM N5ETHD non-monotonically decreases with TPW, with a decrease from 51 mm to 69 mm and an increase to 72 mm. The resulting significant trend of −66 m mm−1 is consistent with the observations. The NICAM MAXREFL non-monotonically decreases with TPW, with a decrease from 51 mm to 66 mm and an increase to 72 mm. This results in a significant decreasing trend of −0.087 dBZ mm−1, which again is different from the observations (particularly the full JJA dataset). The NICAM MAXREFL is noticeably larger than the CloudSat MAXREFL. This is related to the lack of attenuation in the Quickbeam simulator, as described in the previous section.
 The one major difference in the results of the sensitivity experiment is the observed trend of MAXREFL with TPW. With a 10 km minimum height for DCCs, the observed results for JJAphase7 give a significant increasing trend of 0.184 dBZ mm−1. The NICAM analysis yields the same result as the 6 km minimum height analysis.
 Figure 9 shows the DCCAR plots and number of samples for SST. Like with TPW, the SST sample PDFs have a Gaussian distribution with similar means and variances. The NICAM PDF has a slightly lower mode (302K) than the observations (303K) and a similar variance. SST does not appear to have as quite an obvious relationship with reflectivity as TPW. While the plots show an obvious increase in observed CTH and MAXREFL with SST, and an increase in CTH for NICAM, there is sufficient “noise” in the data that the t-test values fall just short of the critical values to pass the 90% significance test. While the observations and NICAM may technically agree according to the statistics tests, it may be useful to examine the differences between trend behaviors. The results do not change much with the full JJA record. The only difference between the JJAphase7 and JJA datasets is the observed CTH, which has a significant trend slope value of 91 m K−1 for JJA.
 Figure 10 shows the DCCAR plots and number of samples for W500. The sample PDFs again have a roughly Gaussian distribution shape with similar means and variances, though the curves are noisier because of the relatively higher number of W500 variable bins—and thus a smaller bin interval—than for TPW and SST. Observed CTH has a general increase with increasing W500, though the increase is very much non-monotonic. This leads to a significant trend slope value of 30,000 m m−1 s (or 300 m cm−1 s). The observed N5ETHD increases only slightly with W500 and has an insignificant trend slope value of 6500 m m−1 s (65 m cm−1 s). This is an unexpected result. As discussed previously, increasing W500 is generally associated with environments more suitable for stronger convective updrafts (e.g., larger CAPE, lower subsidence warming). If N5ETHD is associated with convective intensity, then N5ETHD should change with changing W500. So the lack of a significant trend is surprising. The observed MAXREFL is highly noisy, and so the calculated trend slope value of 16,000 dBZ m−1 s is very unrealistic and is almost certainly unrepresentative of the data. It should be noted that the full JJA record also has a significant increase in N5ETHD, with a trend slope value of 12,000 m m−1 s (120 m cm−1 s).
 The NICAM CTH shows an insignificant increase with a trend line value of 7600 m m−1 s (76 m cm−1 s). The NICAM N5ETHD shows a non-monotonic decrease with W500, giving a significant trend slope value of 23000 m m−1 s (230 m cm−1 s). Unlike the observed/reanalysis results, this result is closer to what we might expect. The NICAM MAXREFL shows a decrease with W500, with a significant trend slope value of −29 dBZ m−1 s (−0.29 dBZ cm−1 s). It should be noted that the NICAM MAXREFL is much less noisy than the observed MAXREFL, despite having a much smaller sample size.
 This study builds upon previous research of using conditional sampling of clouds and LSEVs to evaluate the behavior of atmospheric models. NICAM has an advantage over traditional GCMs of directly simulated DCCs instead of using parameterization, so we focused our analysis on the cloud-LSEV relationship of DCCs in the observations and NICAM. Because no single cloud variable can capture the full relationship between cloud and environment, we used two cloud variables in the analysis. First, we reduced the full dataset of vertical profiles to only the vertical profiles containing PBL-R clouds. Non-PBL-R clouds in these profiles were removed. The purpose of this was to examine the transition between shallow clouds and DCCs as related to LSEVs. Main conclusions from this part of the study are as follows:
 The observations presented in this study generally agree with previous similar studies, though there are some minor disagreements over the exact values at which the shallow cloud/DCC transition occurs.
 The NICAM results generally agree with the observations in simulating the general pattern of the transition from shallow clouds to DCCs. There are some small but interesting differences to note, however. Most of the disagreement occurs in situations marginally favorable for deep convection; NICAM appears to resist convection slightly more strongly than the real atmosphere.
 NICAM and observations both show the increase in COF depth as TPW increases, though NICAM has the transition occurring at slightly higher TPW values than the observations.
 NICAM and observations both show the increase in cloud depth as SST increases, though NICAM has the transition occurring at slightly higher SST values than the observations, and the COF for DCCs is not as strongly associated with SST as it is for observations.
 NICAM and observations both show the increase in cloud depth as W500 increases, though NICAM underestimates the COF of DCCs in large-scale descent conditions.
 Next, the cloudy profiles were further reduced to those that likely contain a DCC. Because the vertical profile of radar reflectivity is related to the characteristics of the DCC, reflectivity was used as the cloudy variable. Main conclusions for this portion of the study are as follows:
 The most obvious difference between the observations and NICAM is the insensitivity of the NICAM CTH to changes in the LSEVs.
 The most notable relationship is found in TPW, where both observation and NICAM respond significantly to increases in TPW. The only insignificance responses are NICAM CTH and observed MAXREFL. The full JJA record has a significant MAXREFL trend.
 Both observations and NICAM show only a small response to SST, which is not strong enough to be significant. The only significant response is the observed full JJA CTH.
 Observations and NICAM have different responses to W500, with the observations having a significant trend with CTH and NICAM having significant trends with MAXREFL and N5ETHD. The full JJA also shows a significant trend for N5ETHD.
 It is difficult to explain these results simply, because there are many processes that effect both convective behavior and radar reflectivity. However, the results suggest a few possibilities that may be worth looking at in the future. CTH is not simply a function of and convective instability and the level of neutral buoyancy but also the convective entrainment rate. If the simulated convective entrainment rates are unrealistic, then the simulated CTH may also be unrealistic. Observing convective entrainment rates has historically been a very difficult task, but recently, Luo et al.  described a method using A-Train data to estimate the entrainment rates of observed DCCs. It would be interesting to compare observed entrainment rates with those calculated from NICAM. It is also possible that the coarse model vertical resolution of NICAM at the typical DCC CTH, about 1 km, prevents NICAM from correctly simulating the cloud/environment interactions that govern cloud top behavior (discussed by Luo et al. ). It would be useful to examine the effect of model vertical resolution at the DCC CTH on the behavior of DCC cloud tops using a local CRM. On the observation/reanalysis side, the lack of significant relationship between W500 and N5ETHD is difficult to explain through simple physical reasons, as mentioned previously. It would be hard to explain this through errors in the observation/reanalysis integration, because it would require an error in the NCEP/NCAR reanalysis of multiple cm s−1. This question probably merits more investigation.
 The 3.5 day time domain in this study leaves some room for improvement in future investigation using larger time domains. A study using a multi-month time domain over a season would help clarify the effect of the short time domain on our results. For example, does the apparent insensitivity of NICAM CTH to TPW depend on the time domain? There are now multiple NICAM datasets available that would allow such investigations (examples).
 As more climate models with cloud-resolving capabilities come into use, the type of analysis used in this study should be considered to evaluate how realistically the simulated clouds behave. It would be very interesting and useful to apply this methodology to other simulations with NICAM. In particular, it would be interesting to examine NICAM runs with 3.5 km horizontal resolution and compare the effect on DCCs of changing the resolution. This would add information to previous work of this nature [e.g., Inoue et al., 2008]. It would also be useful to use this methodology for other GCRMs, as well as models with cloud-resolving elements, which are also becoming more frequently used in a variety of climate-related research.
 Another use would be to extend the spatial domain of the analyses to regions outside of the tropics. Most of the literature involving NICAM focuses on the behavior of clouds and related meteorological features in the tropics. While the tropics has a critical role in regulating Earth's weather and climate, the extra-tropics covers fully half of Earth's surface as well. Using this same analysis method, it should be possible to evaluate the behavior of DCCs in climate models in various extratropical regions. It has been well established that the behavior of DCCs in the extra-tropics—particularly over land—is significantly different than their behavior over the tropical oceans [Lucas et al., 1994]. It may be possible that a cloud-resolving model that behaves realistically over the tropical ocean may not behave realistically elsewhere, with different environmental conditions. One particular place of interest is the continental United States, where the large amount of in situ observations and other measurement methods would offset the limitations of satellite sensors over land. There has already been some research of continental United States DCCs using the A-Train [Luo et al., 2011], and it would be reasonable and useful to continue exploring this avenue.
 This work has been supported by the National Science Foundation Science and Technology Center for Multi-Scale Modeling of Atmospheric Processes, managed by Colorado State University under cooperative agreement. No. ATM-0425247. Part of the research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. In addition, this work has been supported by the CloudSat award #NAS5-99237, and the Department of Atmospheric Science, Colorado State University.