Version 4 CALIPSO IIR ice and liquid water cloud microphysical properties, Part II: results over oceans

Following the release of the Version 4 Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) data products from Cloud15 Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) mission, a new version 4 (V4) of the CALIPSO Imaging Infrared Radiometer (IIR) Level 2 data products has been developed. The IIR Level 2 data products include cloud effective emissivities and cloud microphysical properties such as effective diameter (De) and ice or liquid water path estimates. This paper (Part II) shows retrievals over ocean and describes the improvements made with respect to version (V3) as a result of the significant changes implemented in the V4 algorithms, which are presented in a companion paper (Part I). The analysis of the three-channel 20 IIR observations (08.65 μm, 10.6 μm, and 12.05 μm) is informed by the scene classification provided in the V4 CALIOP 5-km cloud layer and aerosol layer products. Thanks to the reduction of inter-channel effective emissivity biases in semi-transparent (ST) clouds when the oceanic background radiance is derived from model computations, the number of unbiased emissivity retrievals is increased by a factor 3 in V4. In V3, these biases caused inconsistencies between the effective diameters retrieved from the 12/10 and 12/08 pairs of channels at emissivities smaller than 0.5. In V4, microphysical retrievals in ST ice clouds are 25 possible in more than 80 % of the pixels down to effective emissivities of 0.05 (or visible optical depth ~ 0.1). For the month of January 2008 chosen to illustrate the results, median ice De and ice water path (IWP) are, respectively, 38 μm and 3 g⋅m in ST clouds, with random uncertainty estimates of 50 %. The relationship between the V4 IIR 12/10 and 12/08 microphysical indices is in better agreement with the “severely roughened single column” ice crystal model than with the “severely roughened 8-element aggregate” model for 80 % of the pixels in the coldest clouds (< 210 K) and 60 % in the warmest clouds (> 230 K). Retrievals in 30 opaque ice clouds are improved in V4, especially at night and for 12/10 pair of channels, owing to corrections of the V3 radiative temperature estimates derived from CALIOP geometric altitudes. Median ice De and IWP are 58 μm and 97 g⋅m at night in opaque clouds, with again random uncertainty estimates of 50 %. Comparisons of ice retrievals with Aqua/Moderate Resolution Imaging Spectroradiometer (MODIS) in the tropics show a better agreement of IIR De with MODIS visible/3.7 μm than with MODIS visible/2.1 μm in the coldest ST clouds and the opposite for opaque clouds. In prevailingly supercooled liquid water clouds 35 with centroid altitudes above 4 km, retrieved median De and liquid water path are 13 μm and 3.4 g.m in ST clouds, with estimated random uncertainties of 45 % and 35 % respectively. In opaque liquid clouds, these values are 18 μm and 31 g.m at night, with estimated uncertainties of 50 %. IIR De in opaque liquid clouds is smaller than MODIS visible/2.1 and visible/3.7 by 8 μm and 3 μm, respectively. https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c © Author(s) 2020. CC BY 4.0 License.

Due to its sensitivity to small particles, the split-window technique is an attractive option for retrievals of liquid droplets sizes (Rathke and Fisher, 2000), and microphysical retrievals in liquid water clouds are now included in the V4 IIR products. All other things being equal, the performance of the split-window technique increases with the radiative contrast between the cloud and the surface. Consequently, retrieval uncertainties are larger for liquid water clouds, which typically form relatively close to the Earth's 80 surface, and hence these retrievals were not included in V3. Liquid water clouds such as marine stratocumulus clouds, which are an important component of the Earth system, have optical depths typically larger than 10, well beyond the range of applicability of the technique. However, infrared observations have the potential to provide new insight into the microphysical properties of thin liquid water clouds (Turner et al., 2007;Marke et al., 2016) and of supercooled mid-level liquid water clouds.
The IIR analyses start with the retrieval of cloud effective emissivities in each channel, which are then converted to effective 85 absorption optical depths as a,k = -ln(1 -eff,k), where εeff,08, εeff,10, and εeff,12 are the effective emissivities retrieved in IIR channels 08.65 (k = 08), 10.6 (k = 10), and 12.05 (k = 12), respectively. Effective emissivity is mostly a measure of cloud absorption, and the term "effective" refers to the contribution from scattering, which is the most significant at 08.65 µm. The first IIR microphysical index, eff12/10 = a,12/ a,10, is the ratio of the effective absorption optical depths at 12.05 and 10.6 µm and the second one, eff12/08 = a,12/ a,08, is the ratio of the effective absorption optical depths at 12.05 and 08.65 µm. Two main pieces of information are 90 needed to retrieve these quantities: the cloud Top Of Atmosphere (TOA) blackbody radiance, which requires a good estimate of the cloud radiative temperature, and the TOA background radiance that would be observed if no cloud were present. The former drives the accuracy at large emissivities and the latter the accuracy at small emissivities.
The first step into any retrieval approach is the detection of a cloud and the determination of its thermodynamic phase and radiative temperature. The ability to ascertain cloud amounts and characteristics varies with the observing capabilities of different passive 95 sensors (Stubenrauch et al., 2013). Even though IIR has only three medium resolution channels, its crucial advantage is the quasiperfect co-location with CALIOP observations. Indeed, as emphasized by Cooper et al. (2003), cloud boundaries measured by active instruments provide an invaluable piece of information for obtaining accurate estimates of cloud radiative temperatures. The IIR algorithm relies on CALIOP's highly sensitive layer detection to characterize the atmospheric column seen by each IIR pixel.
CALIOP provides geometrical altitudes, which are converted into radiative temperatures. The radiative temperature, Tr, of a multi-100 layer cloud system is estimated as the thermodynamic temperature, Tc, at the centroid altitude of the CALIOP attenuated backscatter at 532 nm. In the V4 algorithm, this estimate is further corrected when single or multi-layer ice cloud systems are observed (Part I). The thermodynamic temperature is derived from interpolated temperatures profiles of the Global Modeling and Assimilation Office (GMAO) Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) model (Gelaro et al., 2017). 105 The second retrieval step is the determination of the TOA background radiance, which often requires simulations using ancillary meteorological profiles and surface data. These simulations are generally more accurate over oceans than over land because the surface emissivities in the various channels are better known and less variable over oceans, and the skin temperature data are usually more accurate. In this paper, we therefore focus on retrievals over oceans. In the IIR algorithm, the TOA background radiance is preferentially determined using observations in neighboring pixels in those cases when clear sky conditions, as 110 determined by CALIOP, can be found. Otherwise, it is computed using the FASRAD radiative transfer model (Garnier et al., 2012;Dubuisson et al., 2005). In V3, IIR microphysical retrievals over oceans were possible down to εeff,12 ~ 0.05 (or optical depth ~ 0.1) when the background radiance could be measured in neighboring pixels (G13). When the background radiance had to be computed by FASRAD, which represents about 75 % of the cases, inter-channel biases in the model simulations caused discernable flaws in the microphysical retrievals. The inter-channel biases in the FASRAD simulations have been significantly reduced in V4, 115 as discussed in Part I. https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. This paper aims at demonstrating the improved accuracy of the V4 effective emissivities and of the subsequent microphysical indices that result from the changes implemented in the V4 algorithm (Part I), and at illustrating the changes in the retrieved microphysical properties. Our assessment is carried out after carefully selecting the relevant cloudy scenes, following the rationale presented in Sect. 2. Retrievals in ice clouds are presented in Sect. 3, which includes step by step comparisons between V3 and 120 V4, examples of V4 retrievals, and comparisons with MODIS retrievals. Section 4 is dedicated to retrievals in liquid water clouds that were added in V4, and Sect. 5 concludes the presentation.

Cloudy scenes selection
The analysis of the IIR observations is informed by the scene classification provided by the V4 CALIOP cloud and aerosol 5-km layer products. This scene classification is established for layers detected by the CALIOP algorithm at 5-km and 20-km horizontal 125 averaging intervals (Vaughan et al., 2009). An example is shown in Fig. 1, which was extracted from nighttime granule 2008-01-30T09-15-45ZN on January 30 th , 2008. 130 number of cloud layers in the cloud-system; cases with Earth surface as a reference are denoted with black lines (thin: semi-transparent (ST) layers; thick: 1 opaque layer) and in red are the cases with the lowest opaque cloud as a reference; (c): CALIOP "Was Cleared Flag" at 1-km IIR pixel resolution; (d): Ice Water Flag of the cloud system (e): temperatures at cloud top and cloud base (black) and radiative temperature used by the IIR algorithm (red); (f): effective emissivity of the cloud-system at 12.05 µm. See text for details. https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. Figure 1a shows the Level 1 CALIOP attenuated backscatter averaged at 5-km horizontal resolution with the top and base altitudes of the cloud-system shown in black. Cloudy scenes can include one or several layers (Fig. 1b). When the lowest of at least two layers is opaque to CALIOP, this opaque layer is used as a reference assuming it behaves as a blackbody source and the algorithm retrieves the properties of the overlying semi-transparent (ST) layers. An example is found between latitudes -36.45° and -36.7°, highlighted in red in Fig. 1b, where the algorithm retrieves the properties of two ST layers overlying the opaque cloud located at 140 about 8 km altitude. South of -36.7° and down to -37.2°, the portion of this cloud which is used as an opaque reference between -36.45°and -36.7° is included in a single opaque cloud of top altitude equal to 11.5 km, which extends down to the southernmost latitudes. North of -36.45° and up to -34.45°, the atmospheric column includes 1 to 3 semi-transparent clouds. Finally, no cloud layers are seen north of -34.45° where the scenes contain only low ST non-depolarizing aerosol layers (not shown). The atmospheric column might also contain clouds having top altitudes less than 4 km that are detected at single-shot resolution and 145 then cleared before searching for the more tenuous layers typically reported in the 5-km products (Vaughan et al., 2009). These single shot detections are not included in Fig. 1b. The number of these single shot cleared clouds seen within each IIR pixel is shown in Fig. 1c. We showed in Part I ( Fig. 5 in Part I) that the presence of these cleared clouds modifies the background radiance compared to the radiance due to the ocean surface and ultimately biases the effective emissivity retrievals. Because these biases cannot be quantified a priori, scenes that contain single shot cleared clouds should be treated with caution. The Ice Water Flag 150 shown in Fig. 1d characterizes the ice/water phase of the cloud layers included in the cloud system. These layers are classified either as ice, liquid water, or "unknown" by the V4 CALIOP Ice/Water phase algorithm (Avery et al., 2020). Most of the ice clouds are composed of randomly oriented ice (ROI) crystals. Clouds containing significant fractions of horizontally oriented ice (HOI) crystals are also detected, mainly before the end of November 2007, when the platform tilt angle was changed from its initial 0.3° orientation to a view angle of 3° (Avery et al., 2020). In Fig. 1 we find cloud systems composed of ROI only (flag = 1), liquid 155 water (WAT) only (flag = 2), ice and WAT (flag = 4), and some systems that include at least one layer of unknown phase (flag = 9). IIR effective emissivities are reported for all single or multi-layer scenes, regardless of the phase. In V4, the phase information is used to adjust the radiative temperature ( Fig. 1e) estimates in cases containing ice clouds (Part I). For illustration purposes, the V4 retrieved effective emissivities at 12.05 µm are shown in Fig. 1f. In this example, emissivity values in the opaque cloud are mostly around 1, the lowest value being 0.91 at -39.5° where the CALIOP image suggests the presence of a faint signal below the 160 cloud. Effective emissivities in ST clouds vary between 0 and 0.9. This example shows that a cloudy scene can include a variety of conditions for the IIR retrievals. Because the goal here is to present the cloud microphysical properties as retrieved with the IIR V4 algorithm and improvements with respect to V3, we chose to limit the analyses to scenes that contain only ROIs, only HOIs, or only WAT clouds with background radiances from the ocean surface.
Furthermore, in order to facilitate the interpretation of the results, we require that the CALIOP cloud-aerosol discrimination 165 algorithm (Liu et al., 2019) assign high confidence to the cloud classifications and likewise that the ice/water phase algorithm determined the phase classifications with high confidence. Finally, scenes containing single shot cleared clouds are discarded. Table 1 reports the fraction of scenes that fall into these categories. The statistics are for IIR pixels between 60° S and 60° N in January and July 2008. The ROI scenes represent 13 % to 16 % of all the IIR pixels. The HOI scenes represent less than 0.1 % of all the IIR pixels, and we found that they represent less than 1 % at the beginning of the mission when the platform tilt angle was 170 0.3°. Thus, in the rest of the paper, ice clouds will refer to scenes containing only ROI layers. The WAT scenes represent 14 to 19 % of all the IIR pixels.
Clear sky conditions are defined as cloud free scenes with Was Cleared Flag at 1 km resolution equal to zero, with no aerosol layers or only low (< 7 km) semi-transparent "not dusty" layers. Dusty layers are those identified as dust, polluted dust, or dusty https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. marine (Kim et al., 2018) and are discarded because they may have a signature in the IIR channels (Chen et al., 2010). For 175 comparison with the previous categories, the clear sky conditions represent 20 % of the cases for daytime data and 15 % for nighttime data. It is noted that 6 to 10 % of the pixels are rejected as "clear sky" in V4 due to the presence of single shot cleared clouds. These pixels would have been accepted by the V3 algorithm: they represent 25 % and 35 % of the V3 clear sky conditions for daytime and nighttime data, respectively.
180 Table 1: Total number of IIR pixels, fraction of IIR pixels with only high confidence ROI, WAT, and HOI layers in the column and no single shot cleared clouds for retrievals with background radiance from ocean surface between 60° S and 60° N, and fraction of clear sky pixels. layers. This is quantified in Table 2 for the months of January and July 2008. For these months, 45 to 53 % of the selected ROIs are opaque to CALIOP while opaque clouds represent 67 to 90 % of the WATs. Daytime fractions of opaque clouds are larger than nighttime ones, which is likely due daytime surface detection issues. Scenes with only ST layers are spread into three main categories: only one layer, two vertically overlapping layers, and multi-layer configurations with two non-overlapping layers or more than two layers. For both ROI and WAT clouds, the vast majority of the ST scenes have only one layer in the column, which 190 is explained by the fact that we required all the layers to be characterized with high confidence. Thus, the study will be carried out for single-layer cases for simplicity.

Retrievals in ice clouds
The accuracy of the effective emissivity in each IIR channel and of the subsequent microphysical indices is a prerequisite for successful retrievals of cloud microphysical properties. In section 3.1, we use internal quality criteria to demonstrate the improvements in the V4 effective emissivities in ice clouds that result from the revised computed background radiances over 200 oceans and from the revised radiative temperature estimates (Part I). After examining the changes in eff,12 (at 12.05 µm), interchannel effective emissivity differences, Δεeff12-k = εeff,12 -εeff,k, are assessed, keeping in mind that they should tend towards zero on average when εeff,12 tends towards 0 and towards 1 (G13; Part I). Changes in the visible cloud optical depth, τvis, inferred from the summation of absorption optical depths at 12.05 µm and 10.6 µm (a,12 + a,10, Part I) are shown in Sect. 3.2.
The subsequent improvements in the microphysical indices and in the performance of the microphysical algorithm are discussed 205 in section 3.3, where we also illustrate changes in the effective diameters (De) reported in V3 and V4. We recall that De is defined as De = (3/2) × (V/A), where V is the total volume of the size distribution and A is the corresponding projected area (Foot, 1988;Mitchell et al., 2002). The V4 algorithm uses two ice crystal models from the "TAMUice2016" data base (Bi and Yang, 2017;Yang et al., 2013), namely the severely roughened solid column (SCO) and severely roughened 8-element column aggregate (CO8) models, and the model used for the retrievals is selected according to the relationship between βeff12/10 and βeff12/08. IIR retrieved 210 De is the mean of the De12/10 and De12/08 effective diameters when these two values can be retrieved from the respective eff12/k; . Both De12/10 and De12/08 are reported in the product for users interested in specific analyses. The V4 look-up tables (LUTs) that relate microphysical index and effective diameter are computed using the FASDOM (Dubuisson et al., 2008) model and bulk single scattering properties derived using an idealized gamma particle size distribution.
As illustrated in Part I, the microphysical indices are very sensitive to De smaller than 50 µm and the sensitivity decreases 215 progressively up to De = 120 µm which is considered the sensitivity limit of our retrievals in ice clouds.

Effective emissivity: V4 vs. V3.
Because of numerous changes in the CALIOP V4 algorithms, the cloud layers reported in the V3 and V4 CALIOP data products 220 are not identical, so that direct comparisons of the V3 and V4 IIR data products could be misleading. In order to isolate the changes due to the IIR algorithm, the V3 emissivities (hereafter V3_comp) for clouds reported in CALIOP V4 were recomputed using the V3 computed background radiances reported in the V3 product and the V3-like blackbody temperatures derived directly from the centroid temperatures, Tc, which are available in the V4 product along with the V4 blackbody temperatures. The exercise was carried out for V4 scenes over oceans that contain one single cloud layer classified as high confidence ROI with no cleared cloud, 225 as discussed previously in Sect. 2. Illustrations are shown for the month of January 2008 between 60°S and 60° N.

Effective emissivity in channel 12.05
The nighttime (blue) and daytime (red) distributions of εeff,12 are shown in Fig. 2 Fig. 2c and 2d are of the order of 0.015 at εeff,12 < 0.6 and increase up to 0.03 at the largest emissivities, where 230 the uncertainty in εeff,12 is prevailingly due to the uncertainty in the radiative temperature taken equal to ± 2 K. (Part I). Because of retrieval errors, eff,12 can be found outside the range of physically possible values (i.e., 0 to 1). For ST clouds (Fig. 2a), the V3 and the V4 histograms differ mostly at εeff,12 < 0.05, where the changes in the background radiances have the largest impact. In this https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. example, the fraction of ST clouds with negative εeff,12 values is reduced from 12 % in V3 to 3.5 % in V4. For opaque clouds (Fig.   2b), the larger V4 εeff,12 values are due to the radiative temperature corrections introduced in the V4 algorithm (these corrections 235 have no to little impact for ST clouds). For the range of εeff,12 values found in opaque clouds, the corrections are prevailingly a function of the "apparent" cloud thickness, which is larger and closer to the true geometric thickness at night (Part I). Nighttime and daytime εeff,12 distributions peak at larger εeff,12 in V4 (εeff,12 = 0.99 and 0.97, respectively) than in V3 (εeff,12 = 0.94). Overcorrections combined with uncertainties cause an increase of the fraction samples with εeff,12 > 1, from 3 % in V3 to 12 % in V4 at night, and from 1.2 to 3.3 % for daytime data. At night, 98 % of the opaque clouds have V4 εeff,12 > 0.8, or cloud optical depth > 240 3.2. This lower range of optical depths is consistent with V4 CALIOP optical depth retrievals, even though it is recognized that direct comparisons with V4 CALIOP optical depths in opaque clouds are difficult (Young et al., 2018). Nighttime εeff,12 distributions for ST and opaque clouds are essentially mutually exclusive, with a εeff,12 threshold around 0.7. In contrast, these distributions overlap between 0.4 and 0.7 for daytime data. The tail down to εeff,12 = 0.4 (τvis ~ 1) for daytime opaque clouds data is explained by a greater difficulty for the CALIOP algorithm to detect faint surface echoes during the day due to large solar 245 background noise, so that some clouds of moderate emissivity may be misclassified as opaque by CALIOP. Effective emissivities close to 1 are found in clouds where the CALIOP integrated attenuated backscatter (IAB) is larger than 0.04 sr -1 , which is in the upper range of values typically observed in opaque ice clouds (Young et al., 2018). Platt et al. (2011) showed that these large IABs, which are often coupled with small apparent geometric thicknesses, are observed when the CALIPSO overpass is close to the center of a mesoscale convective system. Using cloud retrievals based on AIRS thermal infrared data, Protopapadaki et al. (2017) 250 demonstrated that emissivities close to 1 in the tropics are most often indicative of convection cores reaching the upper troposphere, which confirms our observations based on CALIPSO. https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License.

Inter-channel effective emissivity differences
We recall that effective emissivity retrievals preferably use background radiances observed in neighboring clear sky pixels and otherwise use radiances computed by FASRAD. In order to evaluate V4 computed background radiances, we first examined 260 Δεeff12-k at εeff,12 ~ 0 in ST clouds by separating retrievals that used observed radiances (V4_obs) and those that used computed radiances (V4_comp). The results are reported in Table 3, where Δεeff12-k at εeff,12 ≈ 0 is also reported for V3_comp for reference.
As in V3 (G13), V4 inter-channel biases are minimum when the background radiance can be determined from observations (V4_obs), which represents 30 % of the retrievals in ST clouds for this dataset. When the background radiance is computed (V4_comp, 70 % of the cases), median Δεeff12-k is similar for both channel pairs and smaller than 0.0025 in absolute value. This 265 indicates residual inter-channel biases smaller than 0.1 K in V4 according to the simulations shown in Fig. 1c of Part I, which is consistent with the residual inter-channel differences seen in clear sky conditions (Part I). Because these biases are very small, retrievals using computed and observed radiances are consistent in V4, hereafter the two methods will be referred to collectively as "V4" for clarity. The Δεeff12-k differences were unambiguously too low in V3_comp, especially for the 08-12 pair, so that reliable retrievals were possible only when observed radiances were available (G13). Including retrievals using computed radiances 270 in V4 increases the number of retrievals in ST clouds by a factor 3.3. Table 3: Inter-channel effective emissivity differences at εeff,12 ~ 0 for retrievals in single-layered ST ice clouds over oceans between 60°S and 60°N in January 2008. The variations with εeff,12 of the Δεeff12-k inter-channel effective emissivity differences for the 12-10 and 12-08 pairs are shown in Figs. 3a and 3b, respectively. The curves are median values, and the shaded gray areas are between the V4 nighttime 25 th and 75 th percentiles. The first observation is that median Δεeff12-k are larger in V4 (solid lines) than in V3_comp (dashed lines) at any emissivity. When εeff,12 tends towards 1, Δεeff12-k is minimum at εeff,12 corresponding to the peak of the distributions shown in Fig.   2, which suggests that the peaks should be closer to εeff,12 = 1. This shows that V4 is improved compared to V3, more convincingly 280 for nighttime data, but also that the radiative temperature corrections are likely not sufficient. Consistent with the simulations shown in Fig. 1 of Part I, Δεeff12-k are increased from V3 to V4 at large emissivities, because the radiative temperatures are increased, and the changes are more important in the 12/08 pair than in the 12/10 pair. https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License.

Visible cloud optical depth: V4 vs. V3
The V3-V4 changes in the visible cloud optical depths inferred from εeff,12 and εeff,10 are shown in Figs. 4a and 4b for nighttime and daytime data, respectively. The changes in τvis are smaller than 0.02 on average and not significant for τvis smaller than 2 (or εeff,12 < ~ 0.6), that is for most of the ST clouds. For τvis > 2, V4 τvis is increasingly larger than V3 τvis, owing to the warmer radiative temperature estimates in V4. Consistent with previous observations regarding εeff,12, the τvis increase from V3 to V4 is larger at 295 night ( Fig. 4a) than during the day (Fig. 4b).   The changes in the eff12/10 and eff12/08 microphysical indices resulting from the changes in Δεeff12-10 and Δεeff12-08 (Fig. 3) 310 are illustrated in Figs. 5a and 5b. The sharp variations of the V4 median microphysical indices (solid lines) at εeff,12 < 0.03 and εeff,12 > 0.96 are due to the increasing truncation of the distributions, because both βeff12/k indices can be computed only when 0 < εeff,k < 1 in the three channels. Over-plotted in Fig. 5 are the median V4 random absolute uncertainty estimates, which are minimum and around 0.02 for intermediate emissivity values (G13). The noticeable large dispersion of the βeff12/k values at εeff,12 < 0.1 is largely explained by the random uncertainties. The median βeff12/k values are overall larger in V4 than in V3_comp, with larger 315 changes for the 12/08 pair than for the 12/10 pair. The consequences for the De retrievals are twofold. First, the fraction of βeff12/k values that are larger than the low sensitivity limit (close to 1) is increased in V4, which means that the fraction of samples for which microphysical retrievals can be attempted is augmented. Secondly, the larger V4 βeff12/k yield smaller De12/k. These two main changes are detailed and quantified in the following sub-sections.

Fraction of samples in sensitivity range 320
Figures 6a and 6b show fractions of samples for which βeff12/10 and βeff12/08 are larger than their respective theoretical lower ranges, which were derived for De = 120 µm using the V4 SCO LUT, and in practice are close to 1. For both βeff12/10 and βeff12/08, V4 retrievals are possible more than 80 % of the time for εeff,12 between 0.05 and 0.80 (or about 0.1 -3.2 in terms of τvis). In contrast, the εeff,12 80 % range in V3_comp was only 0.15 -0.7 for the 12/10 pair and only 0.25 -0.7 for the 12/08 pair. As εeff,12 increases from 0.8 to 0.95 (τvis ~ 6), which corresponds to clouds that are opaque to CALIOP (see Fig. 2), the βeff12/k indices 325 decrease and approach the sensitivity limit, and the fraction of possible retrievals in opaque clouds decreases. This fraction is notably increased in V4, and is larger at night than for daytime data, reflecting the impact of the cloud radiative temperature corrections introduced in V4. As in V3, this fraction remains lower for the 12/08 pair. One hypothesis is that cloud heterogeneities in dense clouds could induce a larger low bias in the 12/08 pair than in the 12/10 pair (Fauchez et al., 2015). The V4 nighttime retrieval rate is larger than 70 % up to εeff,12 = 0.95 for the 12/10 pair and up to εeff,12 = 0.9 (τvis ~ 4.6) for the 12/08 pair. 330 https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. Figure 6: Fraction of (a) βeff 12/10 and (b) βeff 12/08 values above the effective diameter retrieval sensitivity limit vs. effective emissivity at 12.05 µm in single-layered ice clouds over oceans between 60° S and 60° N in January 2008 in V4 (solid lines) and in V3_comp (dashed lines) during night (blue) and day (red).

Changes in effective diameters
Because the changes in the microphysical indices are larger for the 12/08 pair than for the 12/10 pair, we now assess the changes in the respective diameters, De12/08 and De12/10. For meaningful comparisons, the exercise is carried out only for clouds for which both βeff12/10 and βeff12/08 are found above the lower sensitivity limit, both in V3 and in V4. The changes in De12/10 and 340 in De12/08 are illustrated in Figs. 7a and 7b, respectively. The solid lines represent median De12/k derived from V4 βeff12/k and the V4 SCO LUT. The dashed lines represent median De12/k derived from V3_comp βeff12/k and the same V4 SCO LUT, so that the differences between the solid and the dashed lines are due only to the different microphysical indices. As a result of changes of different amplitude for De12/10 and De12/08, the consistency between these two diameters is drastically improved in V4 at εeff,12 smaller than 0.5. Similar conclusions would be drawn using the V4 CO8 model. 345 For a complete analysis of the differences between the V3_comp and V4 diameters, the dotted dashed lines show De12/k derived using V3_comp and the V3 solid column LUT (Part I), so that the differences between the dotted dashed lines and the dashed lines are due only to the different LUTs. The changes resulting from the LUTs and from the microphysical indices have an opposite effect, regardless of the specific V3 and V4 LUTs chosen for the analysis. As a result, De12/10 is overall not changed significantly in V4 (solid lines) compared to V3_comp (dotted dashed lines). In contrast, De12/08 is smaller in V4 by up to 15 µm at εeff,12 < 0.2, 350 because the improved (and increased) βeff12/08 has the largest impact, and conversely V4 De12/08 is larger by up to 10 µm at εeff,12 between 0.2 and 0.9. https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License.

V4 microphysical retrievals
We showed in Sect. 3.3 that the fraction of samples with possible microphysical retrievals is significantly increased in V4 (Fig. 6), 360 and that the consistency between the De12/10 and De12/08 diameters is drastically improved (Fig. 7). The significant disagreement between De12/10 and De12/08 in V3_comp was due to biases of different amplitude in βeff12/10 and βeff12/08, and could not be explained by the possible use of an inappropriate ice crystal model. Both in V3 and in V4, De is retrieved using the ice crystal model found in best agreement with IIR in terms of relationship between βeff12/10 and βeff12/08. Because the accuracy of IIR βeff12/k is improved in V4, the residual discrepancies with respect to the ice crystal models are expected to be a genuine piece of 365 information about ice crystal shape. This requires both eff12/k to be found within the sensitivity range, which hereafter will be called "confident" retrievals. Because the population of clouds meeting this requirement is larger in V4 than in V3 and covers a larger range of optical depths, the results in this section will be shown for V4 only.
Theoretically, confident retrievals should be found when De is smaller than 120 µm and eff12/k should tend to the upper sensitivity limit for De > 120 µm. In practice, uncertainties in βeff12/k can trigger non-confident retrievals even if De is truly smaller than the 370 sensitivity limit, and this is more likely to occur when De is close to this limit. Requiring both βeff12/k to be in the expected range of values is meant to reinforce the confidence in the retrievals, but doing so implies no systematic bias between both pairs of channels. This is not exactly true for opaque clouds with εeff,12 > ~ 0.8 (Fig. 6), and consequently the fraction of confident retrievals in opaque clouds is often constrained by the 12/08 pair. Furthermore, the fraction of confident retrievals at large emissivities is larger at night. 375

Effective diameter and ice water path
The histograms of confident De and ice water path retrievals (IWP) are shown in Figs. 8a and 8b, respectively, for ST and opaque clouds, and statistics are reported in Table 4. The IWP histograms are computed in logarithmic scale between 0.01 and 1000 g.m -2 , with log10(IWP) bins equal to 0.1. The random uncertainty in De, noted ΔDe, is computed based on the LUT selected for the retrieval and the estimated random uncertainty in the eff12/k indices. Median ΔDe/De values reported in Table 4 are between 34 380 % and 49 %. The uncertainty in IWP is in large part driven by the uncertainty in De. https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License.  Table 4: Statistics associated to V4 effective diameter (De) and ice water path (IWP) retrievals in single-layered ice clouds between 60° S and 60° N over oceans in January 2008 (see Fig. 8 clouds with no detectable precipitation as retrieved using the combined CloudSat-CALIPSO 2C-ICE product. IWP (Fig. 8b) is found between 0.03 and 100 g⋅m -2 in ST clouds, with the slightly larger daytime values being explained by the cloud selection and the larger occurrence of medium emissivities in the daytime dataset (Fig. 2). The medium values are around 3 g⋅m -2 , with peaks in the distributions at 3 g⋅m -2 and 8 g⋅m -2 for nighttime and daytime data, respectively, and the median relative uncertainty is 50 %. 395 As noted by Berry and Mace (2014), the CloudSat radar is typically insensitive to these thin layers, so that microphysical retrievals in combined CALIPSO-CloudSat products such as 2C-ICE rely on parameterization of the radar reflectivity (Deng et al., 2015) rather than on actual observations. Combining CALIOP and IIR observations appears to be a suitable alternative approach to characterize these thin layers.
The estimated cloud radiative temperature (Tr) is at an equivalent altitude located between the CALIOP cloud base and cloud top 400 (Part I). While in case of ST clouds, IIR De is a layer average diameter, IIR De in opaque clouds is mostly representative of the portion of the cloud seen by CALIOP before the signal is totally attenuated. Median De in opaque clouds is around 60 µm and the distributions peak at 50 µm. The different nighttime and daytime De and IWP distributions in opaque clouds are explained by the https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. different ranges of optical depth and the different amplitudes of the radiative temperature correction (Figs. 2 and 4). In opaque clouds, the retrieved IWP lies between 10 and only 300 g.m -2 . The upper limit is due to the fact that De cannot be larger than 120 405 µm and because cloud optical depths inferred from IIR effective emissivities saturate and are typically smaller than 15 (Fig. 4).

Ice crystal model selection
Recall that De is retrieved using the crystal model (SCO or CO8) that agrees the best with IIR in terms of the relationship between eff12/10 and eff 12/08. As seen in Fig. 9, the SCO crystal model is selected in 80 % of the ST clouds of Tr < 205 K. This fraction steadily decreases down to 60 % as Tr increases up to 230 K (Fig. 9b) and remains stable above 230 K. This result is qualitatively 410 consistent with previous findings using V3 , and, as was discussed in this paper, both the IIR model selection and the mean CALIOP integrated particulate depolarization ratio (in black in Fig. 9b) indicate changes of crystal habit with temperature. The difference between mean De12/10 and mean De12/08 in black and grey in Fig. 9c is a measure of the residual mismatch between IIR observations and the selected model. We see two temperature regimes, that is, below and above 225 K, with a better agreement between IIR and the LUTs at the warmer temperatures. This suggests that the V4 models are better suited for 415 warmer clouds and that they do not perfectly reproduce the infrared spectral signatures of colder clouds composed of small crystals.
It is acknowledged that the highly variable ice particle shapes found in ice clouds (Lawson et al., 2019 and references therein) are likely not fully reproduced through the two models chosen for the V4 algorithm. It is further noted that the Clouds and the Earth's Radiant Energy System (CERES) science team is planning to use a two-habit model for retrievals in the visible/near infrared spectral domain (Liu et al., 2014;Loeb et al., 2018). This model would be a mixture of two habits (single column and an ensemble 420 of aggregates) whose mixing ratio would vary with ice crystal maximum dimension, with single columns prevailing for the smaller dimensions. Interestingly, our findings appear to be consistent with this approach.
The increase of De with temperature ( Fig. 9c) is in general agreement with numerous previous findings (e.g. Hong and Liu, 2015).
In this example, mean De increases from 17 µm at 185 K to 53 µm at 245 K. The decrease between Tr = 250 K and 260 K is possibly due to an increasing fraction of small liquid droplets in these prevailingly ice layers, which would be consistent with the 425 fact that CALIOP integrated particulate depolarization ratio decreases from 0.37 to 0.30 (Fig. 9b).

Retrievals using parameterizations from in situ formulation
The IIR algorithm takes advantage of the relationship between βeff12/10 and βeff12/08 to identify the ice crystal model that best matches the observations and thereby provide information about both ice crystal shape and effective diameter. Another approach would be to use only βeff12/10 and prescribed LUTs. This approach was adopted by Mitchell et al. (2018), who derived four sets 435 of LUTs using extensive in situ measurements rather than pure modeling. In Part I, we compared these four sets of ßeff12/10 -De relationships with the relationships derived from the V4 SCO and CO8 models. The four sets of De derived from ßeff12/10 using this independent approach are reported in the IIR product for the user's convenience. Figure 10 compares De computed by the analytic function derived by Mitchell et al. (2018) with De12/10 from the CO8 and the SCO models. Relationships derived from the SPARTICUS (blue) and the TC4 (red) field campaign were computed in two ways: by setting the first bin of the measured 440 particle size distribution (PSD) (D < 15 μm) to 0 (i.e. N(D)1 = 0, dashed lines) and without modifying the distribution (i.e. N(D)1 unmodified, solid lines). As discussed in Part I, the differences between the six sets of retrievals illustrate the possible impacts of the LUTs and of the PSDs. Because the presence of small particles in the unmodified PSD causes βeff12/10 to increase faster than De, assuming N(D)1 = 0 yields smaller values of De for a given βeff12/10 than when N(D)1 is not modified. Even though this was not the original intent, comparing median De with or without setting N(D)1 to 0 also illustrates the impact of possible vertical 445 inhomogeneities of De within the cloud layer . Nevertheless, the overall impact of vertical variations on βeff12/10 also depends on the in-cloud IIR weighting function, which is related to the cloud extinction profile (Part I). , the mean De calculated from the SPARTICUS unmodified ßeff12/10 -De relationship (applied at mid-latitudes) and the TC4 N(D)1 = 0 ßeff12/10 -De relationship (applied in the tropics) was compared against the in situ climatology of mean volume radius, Rv, reported in Krämer et al. (2020) after converting De to Rv. The retrieved Rv tended to be no more than ~ 20% smaller 450 than the in situ Rv for temperatures between 208 and 233 K.

Comparisons with MODIS
https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. Figure 11 compares IIR confident retrievals and co-located Aqua/MODIS Collection 6 daytime retrievals from the visible/2.1 μm and visible/3.7 μm pairs of channels (Platnick et al., 2017, and references therein) in single-layered clouds classified as high confidence ROI by CALIOP and as ice clouds by MODIS. MODIS τvis and De at 1-km resolution are from the MYD06 product 460 and co-location with CALIPSO is from the AERIS/ICARE CALTRACK product. Analyses are over oceans between 30° S and 30° N in January 2008 separately for CALIPSO ST and opaque clouds. Figures 11a and 11b  K where De is < 40 µm and IIR τvis is < 0.5, and they progressively depart from each other as Tr increases and MODIS 3.7 increases and approaches MODIS 2.1. MODIS τvis is larger than IIR by 0.3 to 0.2. This small but systematic bias is not seen when comparing CALIOP and IIR (not shown). The MODIS 2.1 De -Tr relationships are similar for ST and opaque clouds, which is not the case for MODIS 3.7 and IIR. For opaque clouds, IIR De. is larger than in ST clouds and is in good agreement with MODIS 2.1 at Tr < 475 225 K. MODIS 3.7 De exhibits a similar increase with temperature as seen with the two other data sets, but it is shifted by -10 μm.
At Tr > 225 K, MODIS De 2.1 continues to increase up to 100 μm at 255 K, whereas IIR remains stable around 60 μm and MODIS 3.7 increases slowly to approach the same plateau as IIR around 60 μm. As seen in Fig. 11f, both MODIS and IIR indicate moderate optical depths in these opaque clouds where comparisons are possible, with median values ranging between 2.5 and 6 at Tr < 250 K, IIR being smaller than MODIS by about 0.4. 480 Kahn et al. (2015) found that MODIS 2.1 De is typically larger than AIRS De by 10-20 μm, and that MODIS 3.7 is in better agreement with AIRS on average. These results, which were for clouds of optical depth between 0.5 and 2 over oceans, are consistent with our findings for ST clouds. The MODIS and IIR techniques exhibit different non-linear sensitivities to particle size, so that vertical inhomogeneities of the effective diameter can yield three different retrieved De . This could explain than IIR De is found in better agreement with MODIS 3.7 in ST clouds while MODIS 2.1 is clearly larger (Zhang et al., 485 2010). For clouds of moderate optical depth as found in our population of opaque clouds, MODIS 3.7 is very sensitive to cloud top while MODIS 2.1 senses deeper into the cloud Platnick, 2000), and the smaller MODIS 3.7 De as observed in Fig. 11e suggests that the effective diameter is smaller at cloud top than deeper into the cloud. IIR De might be larger than MODIS 3.7 and in better agreement with MODIS 2.1 for opaque clouds at Tr < 220 K because the IIR weighting function is deeper into the cloud than at 3.7 µm, which is agreement with simulations by Zhang et al. (2010). In conclusion, distinct sensitivity to 490 possible cloud vertical and horizontal (Fauchez et al., 2018) inhomogeneity likely contributes to the observed differences.

VRetrievals in liquid water clouds 500
The only difference between effective emissivity V4 retrievals in liquid and ice clouds is that Tr is taken as the temperature at the CALIOP centroid altitude (Tc) in case of liquid water clouds, whereas this initial temperature estimate is further corrected in case https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. of ice clouds. It is recalled that De of liquid droplets are retrieved using the water LUTs (Part I) and that liquid water path is derived from De and εeff,12 (Eq. 10 in Part I).
Following a similar approach as for ice clouds, the results are shown for scenes over oceans between 60° S and 60° N that contain 505 one single cloud layer classified as high confidence water by the CALIOP phase algorithm. Because liquid water clouds are statistically warmer than ice clouds, the radiative contrast is typically smaller than for ice clouds. Because uncertainties are inversely proportional to this radiative contrast (Part I), they increase very rapidly when the radiative temperature contrast, that is the difference between the clear air TOA background brightness temperature and the TOA blackbody brightness temperature, is smaller than 10 K. In order to prevent very large uncertainties associated with very small radiative contrast, the results are presented 510 for clouds in the free troposphere with centroid altitude above 4 km. For this cloud population, the radiative temperature contrast is larger than 10 K, and it increases on average from 15 K at 4 km to 50 K at 10 km where the highest water clouds are found (not shown). Most of these sampled liquid clouds are composed of supercooled water droplets. 520 Figures 12a and 12b show the distributions of V4 εeff,12 in ST and opaque liquid water clouds, respectively, for the month of January 2008 between 60° S and 60° N over ocean, for clouds with centroid altitude > 4 km. Figures 12c and 12d show the respective median random uncertainties., which are about twice as large as the uncertainties in ice clouds (Figs. 2c and 2d) because of the smaller radiative contrast. Only 17 % of these clouds are ST (Figs. 12a and 12c). Unlike in ST ice clouds, the distributions peak at εeff,12 ~ 0.2, and non-physical negative emissivity values are found in only 2 % of the pixels. The εeff,12 distributions in opaque 525 clouds peak at 1.02 at night and at 0.99 for daytime data, with an estimated uncertainty of ± 0.06. The spread around these peaks is larger than for ice clouds, which is explained by the larger uncertainties and specifically to a larger sensitivity to a wrong estimate https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. of Tr. Thus, the nighttime and daytime fractions of samples with εeff,12 > 1, for which no microphysical retrievals are possible, are 45 % and 27 %, respectively. The daytime distributions in opaque clouds exhibit a tail down to εeff,12 ~ 0.4, while at night, the lowest εeff,12 is ~ 0.65, which is very similar to what was observed for opaque ice clouds (Fig. 2b). This similarity suggests that 530 emissivity retrievals in ice and liquid water clouds are consistent, notwithstanding the unavoidable larger uncertainties in the latter ones.

Inter-channel effective emissivity differences
The variations with εeff,12 of the V4 Δεeff12-k inter-channel effective emissivity differences for the 12-10 and 12-08 pairs are shown in Figs. 13a and 13b, respectively. The nighttime (blue) and daytime (red) curves are median values, and the shaded gray areas are 535 between the V4 nighttime 25 th and 75 th percentiles. As for ice clouds, both Δεeff12-k tend nicely to 0 at εeff,12 ~ 0, owing to the improved computed background radiances demonstrated previously, which has a beneficial effect on retrievals in any ST layer.
Both Δεeff12-k have a second minimum at εeff,12 ~ 1, as expected, and this minimum is found slightly larger than 0. Both Δεeff12-k and therefore both βeff12/k are notably larger than for ice clouds (see Fig. 3), reflecting the presence of smaller particles in the liquid water distributions (Giraud et al., 2001;Mitchell and d'Entremont, 2012). As shown by Avery et al. (2020), the IIR 540 microphysical indices are unambiguously larger in clouds classified as liquid water by the CALIOP phase algorithm than in clouds classified as ice. As previously, retrievals are deemed confident when both eff12/k are found within the sensitivity range, which corresponds to De = 60 μm for liquid clouds. The fraction of confident retrievals is found similar in liquid water clouds of centroid altitude > 4 km and in ice clouds. Following the same presentation as for ice clouds, the histograms of confident De and liquid water path retrievals (LWP) are shown in Figs. 14a and 14b, respectively, for ST and opaque clouds, and statistics are reported in Table 5. 555 https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License.

560
Both in ST and in opaque clouds, the nighttime and daytime De histograms are similar. In ST clouds, median De is 13 µm and median liquid water path is 3.4 g.m -2 with a median random uncertainty of 1.2 g.m -2 . In opaque clouds, median De is 18 µm and median liquid water path is 25-31 g.m -2 with a median random uncertainty of 10-15 g.m -2 . The maximum retrieved LWP is about 100 g.m -2 , consistent with the infrared saturation range of 40-60 g.m -2 reported by Marke et al. (2016) who combined microwave and infrared ground-based observations to improve LWP and De retrievals in "thin" clouds that they defined as LWP < 100 g.m -565 2 .The authors report De between 10 and 14 μm in "thin" clouds of top altitude < ~ 1 km, which agrees well with the peaks of our distributions. Table 5: Statistics associated to V4 effective diameter (De) and liquid water path (LWP) retrievals in single-layered liquid water clouds of centroid altitude > 4 km between 60° S and 60° N over oceans in January 2008 (see Fig. 14 Fig. 15 as a function of Tr, highlighting that most of these liquid clouds of centroid altitude > 4 km are supercooled, with Tr ranging between 235 and 280 K (Fig. 15a). Mean IIR De (Fig. 15b, red) increases steadily from 11 μm at 242 K to 18 μm at 270 K, while mean CALIOP particulate depolarization ratio (Fig. 15c) is constant and 575 https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. around 0.1. At Tr > 270 K, De continues to increase up to 20 µm, while CALIOP integrated particulate depolarization ratio decreases. As Tr decreases from 242 K to 235 K and the number of samples drops quickly, De increases up to 24 μm, and CALIOP depolarization ratio increases up to 0.15, indicating a progressive transition to ice phase. As seen in Fig. 15b, De12/10 and De12/08 are in fair agreement. The mean De12/10-De12/08 difference increases from -2 µm at 275 K to + 3 µm at 245 K. This slight temperature-dependent discrepancy between the IIR observations and the water LUT could be explained by the fact that the 580 complex refractive index is temperature dependent, as reported by Zasetsky et al. (2005) and Wagner et al. (2005), the complex refractive index of supercooled water being intermediate between warm water and ice (Rowe et al., 2013). Further investigations will be carried out to establish whether the residual discrepancy between De12/10 and De12/08 would be reduced by using a new set of temperature-dependent indices, following the approach in Rowe et al. (2013). Nevertheless, these simple observations give confidence in the new V4 IIR De retrievals in ST liquid clouds. 585

Comparisons with MODIS
IIR confident retrievals in liquid water clouds were compared with MODIS Collection 6 retrievals from the visible/2.1 µm and visible/3.7 µm pairs of channels for clouds also classified as liquid water by MODIS. The results are shown in Fig. 16, following the same presentation as in Fig. 11 for ice clouds. Again, cloud centroid altitude is chosen to be higher than 4 km, and, as previously for ice clouds, the comparisons shown in Figs. 16 c-f are limited to those pixels for which the IIR, MODIS 2.1 and MODIS 3.7 595 retrievals (orange curves in Figs. 16a and 16b) were all successful. As seen in Fig. 16a, Tr spans between 235 K and 280 K for these sampled liquid clouds. In ST clouds, the three sets of median De (Fig. 16c) have different variations with temperature at Tr < 270 K: IIR De increases with Tr from 10 to 20 µm whereas both MODIS 2.1 and 3.7 are larger than about 20 µm. In addition, MODIS τvis overestimates IIR τvis by about 50 % (Fig. 16d). In contrast, the three sets of De exhibit similar variations with Tr in opaque clouds (Fig. 16e). IIR De (red) is systematically smaller than MODIS 2.1 (green), by 8 μm on average. This is fairly 600 consistent with findings by Di Noia et al. (2019) who compared MODIS 2.1 with new retrievals from POLDER-3 measurements, and found that MODIS 2.1 effective radius was larger by about 3 μm (De larger by 6 μm) for high oceanic clouds having pressures lower than 600 hPa. MODIS 3.7 retrievals (blue) are weighted closer to the top of the cloud than the corresponding MODIS 2.1 retrievals (Platnick, 2000), and are larger than IIR De estimates by only 3 m. This is encouraging, despite of the seemingly temperature-dependent discrepancy between MODIS and IIR τvis (Fig. 16f), where median IIR τvis (red) saturates around τvis = 5 605 while median MODIS increases up to 15 at 240 K. https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License.

Conclusions and perspectives 615
This paper describes the impacts of the various changes implemented in the V4 IIR Level 2 algorithm on the effective emissivities and microphysical retrievals in ice clouds. We chose to illustrate and discuss the changes for one month's worth of data over ocean using a step by step approach so that data users can understand the differences and improvements that they should expect when https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License.
using the recently released V4 IIR Level 2 data products. Retrievals in liquid water clouds, which were added in V4, are also presented. The IIR retrievals rely heavily on the scene classification reported for exactly co-located CALIOP observations. The 620 results are presented for single-layer cases having the ocean surface as a reference and for which the CALIOP cloud classification and ice-water phase identification are determined with high confidence.
We show that in tenuous ST clouds, emissivity retrievals derived from both observed and computed background radiances are fully consistent in V4, whereas the inter-channel biases that were observed in V3 when the background radiance had to be computed introduced significant biases into the V3 microphysical retrievals. Our assessment is based on internal control criteria; i.e., the 625 analysis of retrieved inter-channel effective emissivity differences at εeff,12 ~ 0. Because the background radiance has to be computed for approximately 70 % of the retrievals in ST clouds, the number of unbiased emissivity retrievals is increased by a factor 3 in V4. In V4, the lowest effective emissivity for which microphysical retrievals are possible in more than 80 % of the pixels is reduced to ~0.05 (or τvis ~ 0.1). In contrast, this lowest emissivity limit in V3 was as high as 0.25 in those cases of computed background radiances and was driven by the large biases in the 12/08 pair. Furthermore, when microphysical retrievals 630 were possible in V3, the different 12-10 and 12-08 inter-channel biases induced large differences between the De12/10 and De12/08 diameters retrieved from the respective microphysical indices.
Perhaps one unique feature of the IIR algorithm is that the ice crystal model is selected according to the relationship between the IIR βeff12/10 and βeff 12/08 microphysical indices. In V4, the "TAMUice 2016" SCO (severely roughened single column) model is selected in 80 % of the cases in ST clouds at Tr < 210 K, and this fraction decreases at larger temperatures. The "TAMUice 635 2016" CO8 (severely roughened 8-element column aggregate) model is selected in 40 % of the cases when clouds have radiative temperatures larger than 230 K. In ice clouds, De12/10 is on average smaller than De12/08, with larger discrepancies below 230 K than above. Employing a technique similar to the IIR algorithm, Heidinger et al. (2015) also noticed differences between effective diameters retrieved from the Aqua/MODIS 32/31 and 31/29 pairs of channels when using the "TAMUice 2013" CO8 model , which was chosen for the MODIS Collection 6 data products for its consistency between visible and thermal infrared 640 optical depth retrievals (Holz et al., 2016). We could not find a perfect agreement between De12/10 and De12/08 in liquid water clouds supposedly composed of spherical droplets. In the range of temperature between 240 K and 260 K, where both ice and liquid water clouds are found, De12/10 is larger than De12/08 in liquid water clouds while it is smaller in ice clouds, suggesting that these mismatches are not due to undetected residual biases in the IIR microphysical indices but instead to our LUTs. As noted earlier, the residual mismatch in liquid water clouds could be explained by inaccuracies in the refractive indices, which are taken 645 constant whereas temperature-dependent indices have been reported (Zasetsky et al., 2005;Wagner et al., 2005). Likewise, the "TAMUice 2016" single-scattering properties are derived using refractive indices at 266 K (Warren and Brandt, 2008), but Iwabuchi and Yang (2011) reported that the temperature dependence of these properties in the thermal infrared is small but not negligible. While in V3 mismatches between IIR retrievals and the LUTs were largely due to inter-channel biases in the IIR retrievals, the improved accuracy in V4 opens the possibility for more detailed comparisons with the theory or modeling. 650 Retrievals in opaque ice clouds are improved in V4, especially at night and for 12/10 pair of channels, owing to corrections of the radiative temperature estimates. Refining the relationship between lidar geometric altitudes and infrared radiative temperature based on theoretical considerations (Part I) is deemed important per se, and quasi-perfectly co-located IIR and CALIOP observations offer a unique opportunity to test our theoretical approach. To make further progress in this topic and assess the V4 radiative temperature estimates in opaque clouds, the next step will be to use CloudSat extinction profiles from the lower parts of 655 the clouds not seen by CALIOP.
Daytime comparisons with Aqua/MODIS Collection 6 data products are presented for co-located pixels where V4 IIR, MODIS 2.1 and MODIS 3.7 all have successful retrievals. This comparison demonstrated that IIR is best suited for retrievals in tenuous https://doi.org/10.5194/amt-2020-388 Preprint. Discussion started: 9 November 2020 c Author(s) 2020. CC BY 4.0 License. clouds of emissivity < 0.2 while MODIS is more efficient for denser clouds of emissivity > 0.8. IIR De is in better agreement with MODIS 3.7 than with MODIS 2.1 in tropical ST ice clouds at Tr < 200 K. In contrast, IIR De is in agreement with MODIS 2.1 in 660 tropical opaque ice clouds at Tr < 205 K and in fair agreement with MODIS 3.7 at warmer temperatures. For opaque liquid water clouds having centroid altitudes greater than 4 km, so chosen to ensure sufficient radiative temperature contrast for the IIR retrievals, IIR De is systematically smaller than MODIS 2.1 by 8 µm and smaller than MODIS 3.7 by 3 µm. The IIR technique appears to be perfectly suited for retrievals in ST supercooled liquid water clouds.

Author contribution
AG and JP defined the content and methodology of the paper and wrote the original draft. AG performed the data analysis and prepared the figures. NP was in charge of software development. MV provided assistance for the use of the CALIOP data. PD provided the FASRAD and FASDOM radiative transfer models and bulk scattering properties. PY provided the ice crystal models 680 from the "TAMUice 2016" database. DM provided the analytical functions derived from in situ measurements. All authors contributed to the review and editing of this paper.

Competing interests
Author Jacques Pelon is a co-guest editor for the "CALIPSO Version 4 Algorithms and Data Products" special issue in Atmospheric Measurements Techniques but will not participate in any aspects of the editorial review of this manuscript. All other authors declare 685 that they have no conflicts of interest.