Extension of a Multisensor Satellite Radiance-Based Evaluation for Cloud System Resolving Models

As an alternative approach to previous multisensor satellite evaluations for cloud system resolving models (CSRMs), a technique for precipitation clouds over the ocean of CSRMs is presented using combined infrared and microwave channels. This method quantitatively analyzes precipitation clouds using cloud-top temperatures and ice scatterings from infrared 11 μm and high frequency microwave (89.0 GHz) brightness temperatures (TBs). The TB threshold at low frequencies (18.7 GHz) is used to identify precipitation regions. This method extends a previous approach based on tropical rainfall measuring mission (TRMM) precipitation radar which uses a narrow coverage, by incorporating a wide passive microwave sensor swath and ice cloud sensitivity. The numerical results of the non-hydrostatic icosahedral atmospheric model, NICAM, with two cloud microphysics schemes were evaluated over the tropical open ocean using this method. The scattering intensities in both simulations at 89.0 GHz were different due to the parameterizations of the snow and graupel size distributions. A bimodal snow size distribution improved the TB underestimation at 89.0 GHz. These results exhibited similar structures to the joint histograms of cloud-top temperatures and precipitation-top heights generated using the previous method; the frequencies of overestimated scattering intensities in this study and the frequencies of high precipitation-top heights above 12 km in the previous study. It was observed that the change in the snow size distribution in the cloud microphysics scheme can lead to better agreements of simulated TBs at 89.0 GHz. Furthermore, we investigated the impacts of nonspherical snow assumptions using a satellite simulator. The effect of a nonspherical snow shape in the radiative transfer model caused a smaller change in TBs at 89.0 GHz compared to the difference between the TBs of the two simulations without nonspherical assumptions.


Introduction
Recently, various methods have been proposed to evaluate and improve the cloud microphysics schemes in cloud system resolving models (CSRMs) using satellite data.One method is a radiance-based evaluation using a satellite simulator, which avoids making different settings of the microphysics between retrieval algorithms and CSRMs (Masunaga et al. 2010;Hashino et al. 2013;Matsui et al. 2014).Using the tropical rainfall measuring mission (TRMM) and a satellite simulator (Matsui et al. 2009(Matsui et al. , 2016)), Roh and Satoh (2014) (hereafter RS14) improved cloud properties over the tropical Pacific Ocean simulated by the non-hydrostatic icosahedral atmospheric model (NICAM; Tomita and Satoh 2004;Satoh et al. 2008;Satoh et al. 2014).These properties included precipitation cloud statistics in terms of cloud-top temperature (CTT) and the precipitation-top height (PTH), and the contoured frequency altitude diagrams of radar reflectivities.RS14 improved their single-moment bulk microphysics scheme using the preferred size distributions by multiple sensitivity tests to reproduce realistic cloud statistics, accumulated precipitation, and outgoing longwave radiation.Roh et al. (2017) expanded RS14 to evaluate global simulations with 3.5-km mesh horizontal resolutions and observed improvements in various cloud types over the whole of tropic compared with the TRMM and CloudSat observations.
RS14 and Roh et al. (2017) used the TRMM Triple-Sensor Three-Step Evaluation Framework (T3EF) proposed by Matsui et al. (2009).This method is innovative because multisensor observations are used to evaluate the CSRMs using satellite simulators.However, as TRMM precipitation radar (PR) has a relatively narrow swath (250 km), T3EF is limited to short-term and small domain simulations.
Passive microwave observations in satellites have been used to quantitatively estimate precipitation.The low frequencies (< 20 GHz) in microwave sensors have direct physical relationships with the pathintegrated water content of rain and radiances over the ocean.For high frequencies (> 80 GHz), the radiances are depressed by the scattering of large ice particles such as snow, graupel, and hail.The microwave observations enable us to obtain precipitation information, including ice clouds, in deep convective systems.
Previous studies have used microwave observations to evaluate CSRM precipitation systems (Eito and Aonashi 2009;Matsui et al. 2009;Han et al. 2010).For example, Eito and Aonashi (2009) used the Japan Meteorological Agency non-hydrostatic model (JMA-NHM) and observations from the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) to obtain an agreement between the simulated and observed brightness temperatures (TBs) at 18.7 GHz.However, simulated scattering intensities at high frequencies (36.5 GHz and 89.0 GHz) were stronger than that observed.
In this study, we introduce a method of radiancebased evaluation using microwave and infrared channels and satellite simulators over the ocean.This method is an extension of T3EF and offers two advantages: (1) a wider range of coverage compared with the TRMM PR and (2) ice cloud information.We evaluate the numerical results from NICAM using two cloud microphysics schemes over the tropical open ocean and investigate the impact of snow and graupel size distributions on these results.

Numerical experimental design and observational data
NICAM is a global, non-hydrostatic model that can be applied regionally by transforming the horizontal grid system to focus on a region of interest (stretched NICAM; Tomita 2008a).Simulations using the stretched NICAM were evaluated following the same experimental design in RS14.The analysis domain is the region of the tropical central Pacific Ocean between 10°S to 10°N and 170°E to 170°W.Mesoscale convective systems (MCSs) are dominant and the convective band occurs at approximately 5°N.The minimum horizontal grid spacing is 2.4 km at the central point of the domain and most grid sizes are less than 5 km.We focused on the period from 06:00 UTC on January 1 to 06:00 UTC on January 6, 2007.Two microphysics schemes, the original NICAM Single-Moment Water (NSW6) scheme (hereafter, CON; Tomita 2008b) and the modified NSW6 scheme (hereafter, MODI) following Roh et al. (2017), were used and evaluated in this study.
The microwave data are from AMSR-E on the Aqua satellite.AMSR-E has a 1450-km surface swath and measures microwave emissions at six separate frequencies (i.e.,6.925 GHz,10.65 GHz,18.7 GHz,23.8 GHz,36.5 GHz,and 89.0 GHz).The dual-polarized microwave radiometer operates at each frequency for both the vertical and horizontal polarizations.The polarization differences between the vertically and horizontally polarized TBs at 18.9 GHz (D19; Liu and Curry 1998;Eito and Aonashi 2009) and the polarization corrected 89-GHz TBs (PCT89; Spencer et al. 1989) are derived from AMSR-E L2A (Ashcroft and Wentz 2013), which has a spatial resolution of 39 km and 12 km at 18.7 GHz and 89.0 GHz, respectively.D19 represents the atmospheric emission from liquid water and is close to 0 in strong emission areas (Liu and Curry 1998).PCT89 was calculated to reduce the impact of surface emissivity inhomogeneity (Kidd 1998).
The CTTs were derived from 11-μm infrared channels with a 0.04° resolution and a 30 min interval on the Multifunction Transport Satellite (MTSAT) geostationary satellite.The CTTs were matched with AMSR-E data from similar measurement time.We sampled observation data during the same period as the simulations.
To compare our results with RS14, joint histograms of the CTT and PTH were constructed (Masunaga and Kummerow 2006;Matsui et al. 2009), following RS14.The PTH was identified as the highest altitude of the layer above a 20 dBZ PR reflectivity.The PTH was calculated using 13.8 GHz reflectivities of TRMM 2A25 product, and the CTT are from the TRMM 1B01 product.
We used two satellite simulators to compare the observational and NICAM radiances.The first satellite simulator was developed at the Florida University (hereafter the Liu simulator; Liu 1998;Liu 2008) for the 18.7 GHz and 89.0 GHz microwave TBs of AMSR-E.The second is the Satellite Data Simulator Unit (Masunaga et al. 2010) version 2.1.4for 11-μm TBs and radar reflectivities at 13.8 GHz.The fourstream radiative transfer model was used and a nonspherical database for the ice clouds was implemented in the Liu simulator.A Gaussian beam convolution was applied to the two microwave channels in the Liu simulator, following Masunaga and Kummerow (2005), to reduce the antenna pattern uncertainty of passive microwave sensors.The hydrometeor extinction and scattering properties were calculated based on the Mie calculations in the two simulators.The effective dielectric snow and graupel constant was calculated using the Maxwell-Garnett approach (Maxwell Garnett 1904), generalized by Bohren and Battan (1982).We assumed snow densities ( ρ s ) of ρ s = 100 kg m −3 and ρ s = 0.15 D −1 kg m −3 in CON and MODI, respectively, where D is the diameter of snow in meters.We compare the results of the spherical assumption with those of the nonspherical databases in the Liu simulator (Liu 2008;Nowell et al. 2013).Liu et al. (1995) presented a cloud classification method using the CTT of a geostationary meteorological satellite (GMS)-4 and a microwave index from the scattering and emission channels of the special sensor microwave/imager (SSM/I) over ocean.Matsui et al. (2014) introduced an evaluation method based on joint diagrams using the CTT (MODIS 11-μm TBs) and the PCT89 of AMSR-E to classify cloud types over land, following Liu et al. (1995).The two previous studies used the microwave index and PCT89 to distinguish between non-precipitation and precipitation clouds.In this study, we propose an evaluation method for precipitation clouds using the CTT and PCT89 as the scattering intensities from large ice particles.To improve the results of Matsui et al. (2014), we added D19 to enhance shallow precipitation cloud detection.This method enabled us to investigate high and low precipitation systems using CTT and PCT89 over the ocean.The precipitation areas were identified using the D19 threshold.

Methodology
Figure 1a shows the joint histogram of the CTT and PCT89 from MTSAT and AMSR-E over the analysis domain.There are high frequency mixtures of clear skies and cloudy regions at CTT and PCT89 values greater than 280 K (Fig. 1a). Figure 1a shows all cloud types and does not distinguish between non-precipitation and shallow precipitation clouds.Therefore, we used a D19 threshold of 50 K to focus on the precipitation areas, corresponding to approximately 1 mm hr −1 for a typical tropical profile according to Liu and Curry (1998) (in their Fig. 1a).Figure 1b shows the joint histogram of the CTT and PCT89 for the areas where D19 is lower than 50 K.Using this histogram and CTT and PCT89 thresholds, we classified the precipitation clouds into three categories, "warm", "cold", and "deep" clouds.A CTT threshold of 245 K clearly divided the system into low and high precipitation clouds.We refer to clouds with a CTT higher than 245 K as "warm clouds".We subdivided clouds with a CTT less than 245 K into two categories: "cold clouds" for clouds with a PCT89 greater than 250 K and "deep clouds" for clouds with a PCT89 less than 250 K.
The terminology introduced in this study does not follow conventional definitions in the literature."Warm clouds" are defined as mixtures of shallow and cumulus congestus precipitations.According to Machado et al. (1998), the infrared threshold of clouds associated with deep convection is 245 K. "Cold clouds" comprise stratiform precipitation in MCSs and shallow/congestus, overlapped with cirrus (CTT < 245 K and PCT89 > 250 K). "Deep clouds" (CTT < 245 K and PCT89 < 250 K) are the convective and stratiform precipitations of MCSs related to the snow and graupel-induced depressions of PCT89.

NICAM evaluation
Using the method introduced in the previous section, we evaluated two simulation results using NICAM.Figure 2 shows the relationship between average D19 and rain rate from the simulations.In these simulations, the D19 decreased as the precipitation increased, matching the previous study (Liu and Curry 1998).This phenomenon is due to the fact that the emission of precipitating liquid hydrometeors at 18.7 GHz weakens D19 from the polarized ocean surface.Due to microwave radiation saturation, D19 is limited at precipitation rates greater than 12 mm hr −1 .The precipitation rate for a D19 of 50 K is approximately 1 mm hr −1 .The relationship between the D19 and the rain rate is consistent in tropical settings without ice (Liu and Curry 1998).We found that D19 was not sensitive to the microphysics schemes in CON and MODI despite the fact that the CON D19 reproduced slightly larger average precipitation rates than MODI between 25 K and 60 K.According to Sasaki et al. (2007), the effect of the rain size distribution on the simulated TBs at 19 GHz is small, suggesting that D19 is a reliable surface precipitation index.
Figure 3 shows the joint histograms of the CTT and PCT89 from the CON and MODI results.The MODI (Fig. 3b) results showed high cold cloud frequencies and a minimum PCT89 value near 210 K, which matched the observations (Fig. 1b) more accurately compared with the CON results (Fig. 3a).In contrast, the CON results overestimated the deep cloud frequency and underestimated the cold cloud frequency.
Deep clouds in the CON data had PCT89 values lower than 200 K, which were not observed or detected in MODI.MODI reproduced a smaller fraction of deep clouds (5.7 %) than was observed (10.6 %).Both the CON and MODI overestimated warm clouds compared to the observations.To understand the PCT89 improvements in MODI, we examined its sensitivity to the ice hydrometeor size distribution (Fig. 4).First, the MODI snow size distribution was changed to the distribution used in CON (Fig. 4a).The use of the CON snow size distribution underestimated PCT89 compared with the result of the MODI snow size distribution.The MODI bimodal size distribution was parameterized by temperatures and ice water contents based on the study reported by Field et al. (2005), which led to an increase in PCT89 compared to the CON scheme, assuming a constant intercept parameter with a negative exponential distribution.In addition, only the MODI graupel size distribution was added to CON, which had a small impact compared with that of the snow size distribution.The joint histogram is more widespread than the MODI histogram.As the interceptor parameter of the graupel in MODI was 100 times greater than that of CON, MODI produced smaller graupel sizes than CON.Thus, we concluded that the snow size distribution in MODI was the primary factor influencing the PCT89 improvements.
In Fig. 5, we compare the CTT and PCT89 joint histogram with the joint histogram of CTH and PTH from RS14.To capture precipitation regions, we focused on the areas with surface radar reflectivities above 20 dBZ (note that a TRMM PR of 21.6 dBZ corresponds to approximately 1 mm hr −1 for convective systems) (Fig. 9 in Schumacher and Houze 2000).
We divided the cold and deep clouds using a PTH of 6 km instead of a PCT89 of 250 K.The two joint histograms captured a similar contrast between CON and MODI.For example, CON showed higher frequencies below 220 K compared with those above a 12 km PTH (Figs. 3b, 5b).Deep clouds with high PTH values indicate large graupel and snow at high altitudes, and according to RS14, the snow size distribution affects the PTH distribution.In this study, we found that the snow size distribution primarily affected the PCT89 distribution.

Snow shape dependency
The advantage of using PCT89 is that it is sensitive to ice clouds.Snow has various shapes and affects the scattering properties of high frequency microwave channels such as PCT89.Snow shapes have not been fully accounted for in CSRMs and therefore have created uncertainties (Geer and Baordo 2014).The previous results are based on a soft sphere assumption (SP) according to the Mie calculation with an effective dielectric constant from the Maxwell-Gartnett method.The impacts of the nonspherical snow assumption on the normalized probability distribution of PCT89 were investigated over the analysis domain using the Liu simulator in off-line mode (Fig. 6).We used the pre-computed optical properties of five snow shapes, short column (SN1), thin plate (SN4), six bullet rosettes (SN8), dendrite snow (SN10), and aggregate (SN11).These snow shapes were in accordance to the discrete dipole approximation (DDA, Purcell and Pennypacker 1973) in a database of microwave single-scattering properties for nonspherical ice particles (Liu 2008;Nowell et al. 2013).The normalized PCT89 probability distribution showed that MODI reproduced a more realistic distribution compared with the observations of CON.The CON PCT89 values were underestimated due to the overestimation of ice particle scattering.In CON, simple snow structures such as the short column led to lower PCT89 values than complicated shapes such as six bullet rosettes, dendrite snow, and aggregate.Dendrite snow has a similar distribution to the soft sphere assumption in CON and MODI.Snow shapes affected MODI to a lower degree because the snow size was smaller in MODI than in CON and the nonspherical assumption significantly affected large-sized snow.Thus, the PCT89 distribution pattern depended more strongly on the snow size distribution than on the nonspherical assumptions.

Summary
We propose a CSRM precipitation cloud evaluation method using 11-μm TB from a geostationary satellite and PCT89 and D19 from a passive microwave satellite over the ocean using satellite simulators based on the studies reported by Liu et al. (1995) and Matsui et al. (2014).This method can investigate precipitation clouds using the CTT and PCT89, which represent the cloud-top height and large ice particle scattering intensity, respectively.This method enables quantitative evaluation of simulated shallow and deep precipitation clouds by comparing them to observational data.
Two NICAM simulations using different cloud microphysics schemes were evaluated over the tropical open ocean.We found that the simulation results with MODI were closer to the observations than CON.We also found that the PCT89 improvements in RS14 were primarily due to the change in the snow size distribution rather than changes in graupel size.
The new method presented in this study showed similar results to the TRMM-based analysis in RS14.The passive microwave satellite has limitations in distinguishing shallow and congestus precipitation compared with PTH from TRMM PR.The distribution of the PTH and CTT joint histograms clearly shows three modes: shallow (PTH < 4 km and CTT < 273 K), congestus (4 < PTH < 6 km and CTT > 245 K), and deep clouds (PTH > 6 km and CTT < 245 K).The advantage of using the CTT and PTH joint histogram is that it can distinguish shallow and congestus clouds using PTH and CTT thresholds of 4 km and 245 K, respectively (Fig. 5a).TRMM PR has a finer horizontal resolution (5 km) for detecting precipitation regions compared with the resolution of D19 (approximately 18 km).
However, passive microwave satellites, e.g., AQUA, TRMM, and Global Precipitation Measurement, are abundant and their swaths cover approximately 1450 km compared to the 247 km coverage of TRMM PR.Thus, the present method is more appropriate when collecting data with wider coverage than TRMM PR.
In addition to the wider swath coverage, passive microwave satellite sensors are sensitive to ice cloud properties.We tested nonspherical snow assumptions, with respect to the Mie calculation, using an optical property database from DDA with a satellite simulator.The snow size distribution had a stronger impact on the PCT89 depression than the nonspherical snow assumption.The impact of the nonspherical assumption on the PCT89 distribution depended on the snow size distribution parameterization in the cloud microphysics schemes.
As there are a greater number of satellites carrying passive microwave sensors than those carrying precipitation radar sensors, more sampling data are available from passive microwave sensors.Therefore, by using these data, it is possible to evaluate CSRM simulations with smaller domains and shorter integration times than with active microwave sensor methods.For improvements of CSRMs, quantitative investigations of warm and deep precipitation clouds in different environmental conditions should be needed.And more importantly, ice cloud properties, which are one of the most uncertain parts of CSRMs, should be evaluated and improved.

Fig. 1 .
Fig. 1.PCT89 and CTT joint histogram from AMSR-E and MTSAT over (a) the tropical open ocean and (b) the precipitation regions defined by the 50 K D19 threshold.The color bar is in units of % K −1 K −1 .

Fig. 3 .
Fig. 3. PCT89 and CTT joint histogram over the tropical open ocean with D19 lower than 50 K in (a) CON and (b) MODI.The color bar is in units of % K −1 K −1 .

Fig. 4 .
Fig. 4. (a) Same as Fig. 3b, except for the CON snow size distribution, and (b) same as Fig. 3b except for the CON graupel size distribution.The color bar is in units of % K −1 K −1 .

Fig. 5 .
Fig. 5. PTH and CTT joint histogram over the tropical open ocean from (a) TRMM, (b) CON, and (c) MODI.The color bar is in units of % km −1 K −1 .