a. Data products

CloudSat, release version 5, data over ocean from 2006 to 2010 and from 60°N to 60°S latitude are used throughout this study. The CloudSat GEOPROF product provides a measurement of the normalized surface cross section Σ₀ and radar reflectivity factor Z as described in Tanelli et al. (2008). The GEOPROF product also includes a cloud mask (Marchand et al. 2008), however, CloudSat is not sensitive enough to detect many shallow clouds. For this reason, a combined radar–lidar hydrometeor mask called GEOPROF lidar has also been developed (Mace and Zhang 2014). This product includes a variable that lists the layer-top heights of up to five distinct cloud layers, which we use as a more precise cloud/clear discriminator than can be provided from the radar only. We also use the 2-m temperature from the ECMWF auxiliary (ECMWF-AUX) product, which is a weather analysis interpolated in space and time to the CloudSat footprint (Partain and Cronk 2017).

Cloud properties from MODIS are also used: specifically, the collection 6 MODIS cloud products from Aqua subset to a 15 pixel (±5 km) swath centered on the CloudSat ground track called MAC06S0 (Savtchenko et al. 2008). These files contain the same data fields as their full-swath parent product MYD06 (Platnick et al. 2017). Specifically, we use the visible cloud optical depth τ_vis and effective radius derived from the 3.7-μm channel r_e. An adiabatic cloud liquid water path (Bennartz 2007) is calculated as W = (5/9)ρ_lτ_visr_e, where ρ_l is the density of liquid water.

The 1° × 1° cloud regimes derived using a clustering algorithm applied to MODIS collection-6 level-3 daily cloud-top-pressure and cloud-optical-depth joint histograms (Oreopoulos et al. 2016) are used to place results in the context of specific cloud regimes (CR). This dataset assigns each 1° × 1° region into 1 of 12 CRs, which have distinct geographical distributions, cloud morphology, radiative effects, and precipitation characteristics (Oreopoulos et al. 2016; Leinonen et al. 2016a). There are 12 unique cloud regimes in the MODIS dataset listed in Table 1 along with a brief description. The reader is referred to the above references for more detailed information on the CRs. We use the CR’s to understand and quantify what (if any) cloud-type or regime dependence exists in the SRT and MODIS W_cld data.

Table 1

Brief description of each of the 12 MODIS cloud regimes. See Oreopoulos et al. (2016) for more detail.

View Table

b. Retrieval method

Uncertainty in temperature results in a small contribution to the total uncertainty because of the relative insensitivity of α to temperature. Figure 1 shows the inverse two-way extinction quantity α plotted as a function of temperature. Local variations in this parameter are between 0.24% and 1.28% K⁻¹ over the range of temperatures shown. Throughout this work, we approximate the cloud temperature using the ECMWF 2-m temperature, the CALIPSO cloud height, and a fixed 7.5 K km⁻¹ lapse rate (Zuidema et al. 2009) to define α. We attempted to use MODIS cloud-top temperatures; however, we found that the cloud-top temperature data were missing for many small cumulus clouds. In contrast, this indirect method for estimating cloud temperature is available for every pixel. Zuidema et al. (2009) find that the temperature lapse rate varies between 6.9 and 7.6 K km⁻¹ in subtropical marine boundary layers. Assuming typical cloud heights vary between 0.5 and 2.5 km, we estimate the potential range of error in W_cld due to incorrect cloud temperature to be 0.08%–2.24%.

View Full Size

The two-way inverse extinction coefficient α(T) for liquid water plotted as a function of temperature at 94.05 GHz.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

We note that the distribution of Σ_0,clr occasionally contains outliers indicating a highly nonnormal distribution, which can significantly affect the sample mean and variance. To evaluate the impact of these cases, we perform a test of normality (Shapiro and Wilk 1965) for each distribution of Σ_0,clr associated with individual retrievals. The Shapiro–Wilk test returns a p value indicating the likelihood that the sample distribution is normally distributed. A p value < 0.05 indicates that the null hypothesis of a normal distribution can be rejected with 95% confidence. We emphasize that this statistical test cannot confirm that the distribution is in fact normally distributed.

c. Data filtering

The retrieval method is only applied to a limited set of pixels following a filtering procedure. Some of these filters are necessary while others have an arbitrary element. The filters applied are explicitly described below:

1. Surface type—Retrievals are limited to ocean surfaces only. The navigation_land_sea_flag variable = 2 in GEOPROF is used to identify ocean pixels.

2. Precipitation—The retrieval is limited to nonprecipitating clouds. The Precip-Column variable precip_flag = 0 is used to identify nonprecipitating clouds. Precip_flag = 0 corresponds to an attenuation-corrected reflectivity in the lowest bin exceeding −15 dBZ (Haynes et al. 2009). This is a conservative criterion that excludes clouds that are suspected of containing drizzle.

3. Cloud temperature—The retrieval is limited to clouds with a cloud-top temperature (CTT) > 273.15 K, as calculated indirectly from the 2-m air temperature, the cloud-top height (CTH) from CALIPSO, and fixed temperature lapse rate of 7.5 K km⁻¹.

4. Cloud height—The retrieval is limited to clouds with a CTH ≤ 5 km indicated in the GEOPROF lidar cloud_layer_heights.

5. Cloud identification—The GEOPROF lidar cloud_layer_heights variable = 0 to indicate clear sky. Note that there is a possibility that a CloudSat footprint contains partial cloud cover that is undetected by the CALIPSO lidar because the lidar footprint does not fill the radar footprint area.

6. Missing data—On extremely rare occasions, Σ₀ is missing, in which case no retrieval is performed.

a. Evaluation of errors using clear sky

This clear-sky experiment provides us with an opportunity to confirm our understanding of the measurement uncertainty model. To ensure that the sampling is commensurate with the retrievals performed on cloudy pixels we require that at least one pixel in the window contain a cloud with CTT ≥ 273.15 K and CTH ≤ 5 km. For these clear-sky retrievals we assume a cloud temperature of 280 K for the purposes of translating uncertainties in PIA to a retrieved W_cld. Results would scale according to Fig. 1 for different cloud temperatures.

Table 2 and Fig. 2 show a summary assessment of the retrieval mean bias and the ability of the analytical uncertainty estimate to reproduce the observed variability. The table shows separate rows for all pixels and for subsets of data in which pixels with p value ≤ 0.05 have been removed. The removed pixels are those for which the null hypothesis that the distribution of Σ_0,clr is normally distributed can be rejected with 95% confidence. A final row shows a subset in which pixels with p value ≤ 0.05 and surface wind speed < 4 m s⁻¹ have been removed. First, the results demonstrate that, as expected, the mean bias is negligibly small. Second, the observed variability and analytical error estimates are significantly reduced when the p value ≤ 0.05 pixels are removed from the population and further reduced when wind speeds < 4 m s⁻¹ are removed. Third, the results show that the analytical uncertainty estimate is systematically too large and indicates that outliers are biasing the variance in the estimate of Σ_0,clr. Furthermore, Fig. 2 clearly shows a sharply peaked, nonnormal distribution of retrieved W_cld and significant overestimate of uncertainty when the p value ≤ 0.05 and low-wind-speed pixels are included. This is an unexpected result for such a large sample size and is the likely result of occasional outliers causing the distributions of Σ₀ to behave pathologically. We are drawing from real-world sample distributions whose exact properties cannot be determined; however, it can easily be shown that drawing from analytical distributions such as the Cauchy distribution, which does not have a defined finite mean and variance, results in Eq. (7) consistently underestimating variance. Removing the low-p-value and low-wind-speed pixels reduces the uncertainty and results in a more normal distribution; however, there remains a slight overestimate of the observed variability by the analytical model. Figure 3 demonstrates that the extent of the uncertainty overestimation is a function of p value. The uncertainty overestimate systematically decreases with increasing p value and significantly increases for p values < 0.05. Even for p value = 1 there is a small residual overestimate of the analytical model because the statistical test cannot confirm the null hypothesis that the distribution is in fact normally distributed. From these results we define an empirical uncertainty adjustment term to account for nonnormality:

${\mathrm{\Delta }}_{\epsilon }=-7.89+1.77\hspace{0.17em}\ln (p\text{ value}),$(8)

which is only valid for p value > 0.05. The adjustments would become significantly larger for the smallest p values; however, we will exclude these pixels from the remainder of the analysis. The final column in Table 2 shows the adjusted uncertainty is indeed in much better agreement with the observed standard deviation than the unadjusted estimates.

View Full Size

(a) The histogram of the derived clear-sky water paths along with the standard deviation (STD) of the observed distribution and the analytical error estimate (Error). (b) The same, filtered for p value > 0.05 and wind speed > 4 m s⁻¹ and also including the adjusted error (Adj Error) from Eq. (8).

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

View Full Size

(left) The observed STD minus the unadjusted analytical error estimate as a function of the Shapiro–Wilk p value and (right) a histogram of the p values. Note that p value = 0 is not included in either panel. For p value = 0 cases, the difference between the observed STD and the analytical error estimate is −105.6 g m⁻², which is substantially larger than for nonzero p values < 0.05.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

Table 2

Summary of clear-sky retrievals of W_cld. The uncertainty estimate is derived from Eqs. (5) and (7), and the adjusted uncertainty estimate includes the p-value correction shown in Eq. (8).

View Table

The value of Σ_0,clr has a strong dependence on wind speed and a much smaller dependence on the sea surface temperature (SST). The dependance on wind speed is particularly strong for wind speeds less than about 4 m s⁻¹ (Tanelli et al. 2008; Haynes et al. 2009), whereas the magnitude of Σ_0,clr is relatively stable at higher wind speeds. There are larger natural variations in the individual values from which we estimate $\widehat{{\text{Σ}}_{0,\text{clr}}}$ in low-wind-speed environments, and as a result we expect errors in estimating $\widehat{{\text{Σ}}_{0,\text{clr}}}$ to be substantially larger in low-wind-speed regimes. Figure 4 demonstrates that this is indeed the case. Both the standard deviation of the clear-sky retrievals and the analytical error estimates have a strong nonlinear increase below wind speeds of 4–5 m s⁻¹. Figure 4 also shows that despite the increasing error the mean bias is still relatively close to 0 albeit with a small negative bias at low wind speeds. We conclude that we are still able to adequately quantify the error in low-wind-speed environments; however, that error is large and for the remainder of the paper we will remove pixels in which wind speed is below 4 m s⁻¹. While these pixels correspond to a small fraction of our global sample, they are concentrated regionally in the western Pacific Ocean warm pool and parts of the intertropical convergence zone, where larger errors would be frequently expected when using the SRT retrieval.

View Full Size

(top) The observed STD of retrieved clear-sky W_cld (blue) and the adjusted error estimate (red) as a function of near surface wind speed for all clear-sky pixels with p value > 0.05. (bottom) The retrieved mean W_cld.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

Figure 5 shows a more detailed evaluation of the retrieval uncertainties in terms of cloud regimes. Figures 5a and 5b show the retrieval mean, standard deviation, error, and adjusted error estimate as a function of cloud fraction and MODIS cloud regime, respectively. Most important, the retrieved mean is generally near 0 g m⁻² regardless of cloud regime or cloud fraction, indicating a relatively bias-free estimate of PIA_cld across a diversity of cloud conditions. Again, we see that the analytical error estimates consistently overestimate the observed variability; however, the adjusted error estimate does a credible job of bringing the uncertainty into agreement with the observed standard deviation across regimes even though the adjustment term has no knowledge of the cloud regime or cloud fraction. Last, the measurement standard deviation shows a clear dependence on the cloud regimes, which is well tracked by the adjusted error estimates. The mean adjusted error estimated is 34 g m⁻² (Table 2), and Fig. 5 shows that after filtering by p value and wind speed, this value is fairly representative across the diversity of cloud regimes.

View Full Size

Results sorted according to (a) the lidar cloud faction or (b) the MODIS CR. For (a) and (b), the top plot shows the observed STD, analytical error estimate, and adjusted error (Adj Error). The bottom plot shows the mean retrieved W_cld. The top plot can be interpreted as random error, and the bottom plot can be interpreted as measurement bias.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

We conclude from the evaluation of clear-sky W_cld that random retrieval errors originating from application of the SRT are well understood. There is negligible bias in estimating $\widehat{{\text{Σ}}_{0,\text{clr}}}$ across a diversity of cloud regimes. The random uncertainty in PIA_cld translate into typical uncertainties in W_cld of 34 g m⁻² after filtering for pixels of the Σ_0,clr distribution with p value > 0.05 and wind speeds > 4 m s⁻¹. This precision is well characterized using an analytical framework after applying an adjustment factor to account for the nonnormal distribution of the observed Σ_0,clr.

b. NUBF

Thus far, these clear-sky retrievals neglect a potentially important source of error, that is the effect of NUBF of the field of view in the cloudy pixels. The SRT retrieval is an inversion of the observed attenuation that is an exponential function of τ_,cld [Eq. (1)]. Because the exponential function is convex everywhere, Jensen’s inequality states that the retrieved optical depth must be less than or equal to the true mean optical depth (${\tau }_{\text{cld},\text{ret}}\le \overline{{\tau }_{\text{cld}}}$). Since τ_,cld maps linearly into W_cld, we see that NUBF always results in an underestimate of the true cloud water path.

We use an LES of the Barbados Oceanographic and Meteorological Experiment (BOMEX; Siebesma et al. 2003) field experiment to evaluate the expected biases due to NUBF on the retrievals. BOMEX was a nonprecipitating shallow marine cumulus field experiment where clouds very similar to those that are the ideal targets of the SRT retrievals were observed. The LES model of Matheou and Chung (2014) is used for the simulation with the initial/boundary conditions and forcings described in Siebesma et al. (2003). The LES is run with horizontal and vertical resolution of 20 m on a domain size of 12.8 × 12.8 km². The vertical profile of the atmosphere above the LES top boundary (located at 3 km) is filled in from Modern-Era Retrospective Analysis for Research and Applications, version 2 (Gelaro et al. 2017), data appropriate for the region and weather regime. We couple a radar simulator to six different snapshots of the BOMEX LES as described in Roy et al. (2021). The simulator assumes a constant unattenuated radar surface cross section, but it does include resolved subpixel variability in both cloud liquid and water vapor. To evaluate the NUBF, the Σ₀ and W_cld fields are averaged over the radar’s two-way beam pattern using an idealized Gaussian CloudSat-like antenna pattern. We then normalize the simulated Σ₀ such that the mean is 0 dB for clear-sky pixels separately at the model and radar resolution. After accounting for CloudSat’s along-track integration by performing a time-average of the antenna pattern, the resulting footprint has size 1.4 km (cross track) by 1.7 km (along track), where the widths are defined using the 6-dB point of the two-way propagation pattern (Tanelli et al. 2008). The LES domain was divided into an integer number of parallel along track segments that are separated by the cross-track beamwidth of 1.4 km to ensure footprint independence. Each along-track pixel is separated by the along-track width of 1.7 km again to maintain statistical independence of each pixel. The end result is a reduction in the number of CloudSat-resolution pixels relative to the native model resolution by an approximate factor of 50². Last, an estimate of PIA_cld was made by subtracting the individual Σ₀ from the mean Σ_0,clr.

Figure 6 (left panel) summarizes the influence of NUBF on the retrieved W_cld for BOMEX. The figure shows the relationship between W_cld and PIA at both the model resolution and the CloudSat resolution. The relationship is linear at the model resolution with the scatter due to water vapor absorption. We fit a line to this relationship with intercept forced through 0 and find a slope of α = 105 g m⁻² dB⁻¹. We then use the CloudSat-resolution simulated PIA_cld to invert this linear fit to derive the W_cld providing simulated retrievals. The mean bias of these synthetic retrievals is −8%, although it is clear in Fig. 6 that the biases can be significantly larger as W_cld increases. We reemphasize that the negative bias results because the relationship between PIA_cld and W_cld is not linear at the CloudSat resolution. To provide a heuristic understanding of the relationship at the radar resolution we approximate the NUBF by a pixel that is filled with a homogenous cloud element with subpixel cloud fraction f_cld and optical depth τ_cld and an area of homogenous clear sky with area 1 − f_cld and τ_cld = 0. To further simplify the model, we assume that there are no subpixel variations in the gaseous attenuation. The observed PIA averaged over the radar pixel in decibel units can then be parameterized in terms of f_cld as

$\overline{{\text{PIA}}_{\text{cld}}}=-10\hspace{0.17em}{\log }_{10}\left(f_{\text{cld}}\hspace{0.17em}\exp \left\{-\left[\frac{\ln (10)\overline{W_{\text{cld}}}}{10\alpha f_{\text{cld}}}\right]\right\}+(1-f_{\text{cld}})\right).$(9)

The right panel of Fig. 6 shows the fractional bias of a partially filled footprint according to Eq. (9) as a function of f_cld and $\overline{W_{\text{cld}}}$. Three key points can be gleaned from Eq. (9) and Fig. 6. First the NUBF bias increases as f_cld decreases. Second, regardless of the cloud fraction the NUBF bias tends to 0 as $\overline{W_{\text{cld}}}$ goes to 0. Third, for a given cloud fraction the bias increases monotonically as the $\overline{W_{\text{cld}}}$ increases. While a more realistic description of NUBF changes the magnitude of these effects, it does not alter these three general conclusions.

View Full Size

(left) The relationship between W_cld and the apparent PIA from the BOMEX LES experiment. Here the y axis shown as PIA is the difference of the individual simulated Σ₀ from the mean of the clear-sky pixels so that it includes the effect of water vapor and cloud liquid water variability. The relationship is shown at both the model native resolution (20 m²) and the CloudSat radar footprint resolution (1700 × 1400 m²). The linear fit at the model resolution is PIA = W_cld/105. (right) The CloudSat-resolution fit, based on Eq. (8) with f_cld = 0.32.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

We find the value of f_cld that minimizes the mean square deviations of residuals of the CloudSat-resolution data, giving f_cld = 0.32. This should be contrasted with the actual cloud fraction in the BOMEX simulations is 0.21. In this sense it is important to interpret f_cld as an effective cloud fraction that encompasses all of the effects of NUBF as opposed to an actual cloud fraction. Nevertheless, the approximation of Eq. (9) provides a possible approach to mitigate NUBF biases in the SRT by using high-resolution imagery from the MODIS bands 1 and 2 that resolve visible and near-infrared radiances at 250 m to estimate radar subfootprint cloud fraction.

We comment that the effects of NUBF can be significantly larger when using the SRT to infer the total water path (including precipitation), in which case optical depths are substantially larger and the α term itself is spatially variable and introduces considerable additional inhomogeneity within the radar beam. Lebsock and Suzuki (2016) showed that when estimating the total water path in shallow convective precipitating cloud field this bias could be as large as −50%.

c. Retrieval yield

The retrieval for W_cld presented here is powerful in that it is extremely simple and has a fairly well-understood error characterization, including minimal sources of mean bias. However, the retrieval can only be performed reliably on nonprecipitating, marine, liquid-phase clouds, for which a clear-sky surface cross section reference can be established. We further limit the retrieval to pixels in which p value > 0.05 and wind speed is > 4 m s⁻¹. From 2006 to 2010 we find 80 257 183 cloudy pixels with a CTT ≥ 273.15 K and CTH ≤ 5 km between 60°N and 60°S over ocean. Of these, 13.8% are flagged as precipitating (including drizzle). We are unable to estimate $\widehat{{\text{Σ}}_{0,\text{clr}}}$ for 25.2% of pixels, and 16.3% have a p value ≤ 0.05. Last, 12.9% have wind speed < 4 m s⁻¹. These conditions combine to limit retrieval yield to 43.1% between 60°N and 60°S latitudes. Note that this is the retrieval yield defined relative to only single layer clouds liquid clouds. There is a notable geographical pattern to the retrieval yield that is also evident in the yield binned by MODIS cloud regimes (Fig. 7). The retrieval yield is found to be maximized for the MODIS CRs 1 and 10–12, which correspond to shallow subtropical shallow cumulus convection (CR10–11), disorganized cloud fields, with low cloud fraction (CR12), and tropical cirrus (CR1). Note that, while the retrieval yield is fairly high for CR1, the actual number of retrievals in this regime is relatively small because single layer liquid clouds are rarely observed because of the preponderance of high cloud. There is a significant reduction in the retrieval yield for the other CRs that can be attributed primarily to the inability to consistently estimate $\widehat{{\text{Σ}}_{0,\text{clr}}}$ relative to CRs 10–12. Precipitation occurrence, which is a significant factor in reducing retrieval yield, is relatively constant across the cloud regimes.

View Full Size

(top) Map of the retrieval yield, (middle) the retrieval yield binned by the MODIS cloud regime, and (bottom) the frequency of various factors decreasing retrieval yield binned by MODIS cloud regime.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

d. Cloud retrievals

The geographical distribution of the retrieved mean W_cld is shown in Fig. 8. We emphasize that this W_cld is conditioned on low cloud occurrence for which a retrieval is performed and is not suitable for comparison with all-sky datasets or model outputs. The geography of the distribution is fairly straightforward, with relatively constant values between 20 and 30 g m⁻² over much of the midlatitude oceans and the eastern margins of the subtropical ocean basins. Each of these areas is characterized by significant coverage of stratocumulus clouds. In contrast, cumulus-dominated geographies such as the western Pacific warm pool and the Indian Ocean have smaller mean W_cld near 10 g m⁻². We note that these low values of W_cld in cumulus regimes are a direct result of the large numbers of small lidar-detected cumulus that go undetected by CloudSat and often do not have MODIS cloud retrievals, which highlights the unique role for the SRT retrieval to fill the gaps in the other retrieval methods.

View Full Size

The map of the conditional W_cld for the SRT.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

Next, we compare the SRT retrievals with MODIS W_cld derived according to the adiabatic assumption and the 3.7-μm r_e. Figure 9 shows good correlation between the retrieved mean SRT W_cld binned by MODIS W_cld. However, there is a significant bias between the two estimates. MODIS has a mean W_cld of 43.1 g m⁻², and the SRT has a mean W_cld of 30.0 g m⁻². Also shown in Fig. 9 is a range of possible biases for the SRT retrieval from two sources: 1) NUBF from the parameterization in Eq. (8) and Fig. 6, which always biases the SRT negatively (low), and 2) systematic water vapor attenuation bias of 10 g m⁻², which may bias the retrieval positively (high) as estimated from Lebsock and Suzuki (2016). Over much of the range of MODIS W_cld, it appears that NUBF could explain some of the MODIS–SRT bias, but with the caveat that this parameterization is based on a single LES case and may not be generalizable.

View Full Size

(left) A scatterplot of the mean SRT binned by MODIS W_cld. The gray error bars show the STD of the SRT retrievals within each MODIS bin. The red error bars show a range of potential bias based on NUBF using the parameterization in Fig. 6 and Eq. (9) and water vapor bias assumed to be 10 g m⁻² from Lebsock and Suzuki (2016). (right) The mean difference between SRT lidar and MODIS W_cld as a function of MODIS W_cld.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

Some of the bias between MODIS and the SRT can be explained by NUBF causing an underestimate in the latter. Applying the NUBF correction (e.g., Fig. 6) to every pixel increases the mean SRT W_cld to 35.3 g m⁻², which is in better agreement with MODIS. However, there is a great deal of uncertainty in the applicability of this parameterization, which is based on a single LES, to global data, as stated earlier. Furthermore, we expect that water vapor attenuation bias in the SRT cannot explain the bias because it should only further decrease the SRT and thus increase the discrepancy.

Of course, MODIS LWP is also subject to bias, particularly due to possible overestimates in r_e that would result in overestimates of W_cld. Recently, unresolved variability has been shown to account for a 1–2 μm overestimate of the MODIS r_e (Zhang et al. 2016; Werner et al. 2018). Three-dimensional radiative effects have been estimated using Multiangle Imaging SpectroRadiometer (MISR) optical depths near cloud-bow scattering angles to suggest a zonal mean bias in r_e derived from the 3.7-μm channel of 2–7 μm (Liang et al. 2015). Unresolved variability and three-dimensional effects also influence the retrieved optical depth. Depending on the scattering geometry and cloud morphonology these biases can be either positive or negative (e.g., Várnai and Marshak 2002); for the subtropical cumulus cloud regimes that are overrepresented in this dataset, the optical depth bias is likely negative because of large solar zenith angles. The combined effect of potential r_e and τ bias on the MODIS W_cld is difficult to assess.

Figure 10 shows the scatterplot of the mean SRT against MODIS categorized by MODIS CRs and Table 3 shows the mean W_cld for each CR. The geographical distributions of MODIS CRs are shown in Oreopoulos et al. (2016), and a brief description of each was provided here in Table 1. In general, there is correlation between the SRT and MODIS retrievals regardless of regime, however, there are clearly regime dependent biases. The only MODIS CRs in which SRT tends to be larger than MODIS is CR2, which contains the strongest convective storms, and has a very small number of SRT retrievals. For all other CRs, MODIS exceeds the SRT. This is particularly true for CRs 8–12, which correspond to stratocumulus, cumulus, and broken cloud regimes. These CRs also dominate the retrieval yield (Fig. 7) and therefore largely drive the overall SRT–MODIS bias. It is plausible that these regimes contain a disproportionate amount of small cumulus that only fill part of the CloudSat footprint and would therefore be especially susceptible to NUBF-induced underestimates for the SRT.

View Full Size

Scatterplot of the mean SRT lidar binned by MODIS W_cld for each of the 12 MODIS cloud regimes. The gray error bars show the STD of the SRT retrievals within each MODIS bin.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide

Table 3

CloudSat W_cld and MODIS W_cld (g m⁻²) averaged over different cloud regime.

View Table

The regime dependent biases in W_cld suggest that there may be geographic patterns of bias as well. Figure 11 shows that there are indeed indications of coherent geographical patterns in the differences between the data. MODIS is uniformly larger than SRT almost everywhere with the exception of the western Pacific warm pool and small regions off the east coast of Central America and North Africa. These are regions where deeper liquid-phase cumulus convection exists than is found in the subtropical trade winds and are home to the CR 2, which is the regime in which the SRT mean exceeds the MODIS mean.

View Full Size

The mean CloudSat SRT − MODIS W_cld. Only common pixels are used in this plot.

Citation: Journal of Applied Meteorology and Climatology 61, 8; 10.1175/JAMC-D-21-0235.1

• Download Figure
• Download figure as PowerPoint slide