Evaluating the Consistency and Continuity of Pixel-Scale Cloud

1Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 5 2 Earth System Science Interdisciplinary Center, University of Maryland, 5825 University Research Court, Suite 4001, 6 College Park, MD 20740. 7 3Science and Technology Corporation, 10015 Old Columbia Road, Columbia, MD 21046 8 4Science Applications International Corporation, 12010 Sunset Hills Road, Reston, VA 20190 9 5NASA Goddard Space Flight Center, Greenbelt, MD. 10 11 12

The Aqua, SNPP, and JPSS satellites carry a combination of hyperspectral infrared sounders 15 (AIRS, CrIS) and high-spatial-resolution narrowband imagers (MODIS, VIIRS). They provide an 16 opportunity to acquire high-quality long-term cloud data records and are a key component of the 17 existing Program of Record of cloud observations. By matching observations from sounders and 18 imagers across different platforms at pixel scale, this study evaluates the self-consistency and 19 continuity of cloud retrievals from Aqua and SNPP by multiple algorithms, including the AIRS 20 Version-7 retrieval algorithm and the Community Long-term Infrared Microwave Combined 21 Atmospheric Product System (CLIMCAPS) Version-2 for sounders, and the Standard Aqua-22 MODIS Collection-6.1 and the NASA MODIS-VIIRS continuity cloud products for imagers. 23 Metrics describing detailed statistical distributions at sounder field of view (FOV) and the joint 24 histograms of cloud properties are evaluated. These products are found highly consistent despite 25 their retrieval from different sensors using different algorithms. Differences between the two 26 sounder cloud products are mainly due to cloud clearing and treatment of clouds in scenes with 27 unsuccessful atmospheric profile retrievals. The sounder subpixel cloud heterogeneity evaluated 28 using the standard deviation of imager retrievals at sounder FOV shows good agreement between 29 the standard and continuity products from different satellites. However, impact of algorithm and 30 instrument differences between MODIS and VIIRS is revealed in cloud top pressure retrievals and 31 in the imager cloud distribution skewness. Our study presents a unique aspect to examine NASA's 32 progress toward building a continuous cloud data record with sufficient quality to investigate 33 clouds' role in global environmental change. by the two satellites while pixel-scale cloud assessment is carried out by comparing cloud 107 parameters determined by hyperspectral IR sounders and high spatial resolution imagers at the 108 minimum spatial scale of individual instrument fields of view. Using this approach, products from 109 both the heritage NASA standard retrieval algorithms and the newly-developed continuity cloud 110 algorithms are analyzed (Table 1). This is essential for retrieval algorithm development and cross-111 validation of multiple sensors and products on Aqua and SNPP, and also important for data 112 continuity extending to future JPSS satellites. and Full Spectral Resolution (FSR), which differ in the spectral resolution of the shortwave and 126 mid-IR CrIS observations transmitted from SNPP (Monarrez et al. 2020). The spectral resolution 127 differences cause subtle differences between the CLIMCAPS FSR and NSR retrievals, especially 128 in the upper tropospheric humidity and trace gases (Wang et al. 2021). 129 In both the AIRS V7 and CLIMCAPS algorithms for AIRS and CrIS, the radiatively effective 130 cloud amount (effective cloud fraction, ECF) and cloud top pressure (CTP) are retrieved by 131 matching the calculated cloudy radiances with the observed radiances for a set of channels that are 132 sensitive to clouds. Then the cloud top temperature (CTT) is derived as the atmospheric 133 temperature matching the retrieved CTP. In this process, best estimates of surface and atmospheric 134 parameters are used to calculate the cloudy radiances, either from the a priori state or from the 135 physical retrieval after the cloud clearing step (Susskind et al. 2003, Susskind et al. 2006 and Barnet 2019). The cloud clearing approach (Chahine 1974) is applied in both the AIRS Science 137 Team algorithms and CLIMCAPS. It predicts a single cloud cleared radiance at one AMSU or 138 ATMS field of regard (FOR) using a priori temperature, water vapor, and surface information and 139 a linear combination of IR radiances from nine AIRS or CrIS FOVs that are co-registered with one 140 AMSU or ATMS FOR (Susskind et al. 2003). The cloud cleared radiances are subsequently used 141 to retrieve surface and atmospheric parameters. Flowcharts of the retrieval steps and differences 142 in these two sounder retrieval systems are given in Thrastarson et al. (2021). 143 The ECF is the product of cloud areal fraction and the IR cloud emissivity, the latter of which 144 is assumed to be spectrally flat in the retrieval of ECF (Susskind et al. 2003). Previous studies 145 show that the AIRS ECF is consistent with the cloud properties such as the cloud frequency and 146 cloud optical depth measured by CloudSat and MODIS (Yue et al. 2011, Kahn et al. 2014). The 147 AIRS and CrIS retrievals of ECF and cloud top properties (CTT and CTP) are reported for up to 148 two cloud layers in each IR sounder FOV (~13.5 km spatial resolution at nadir). 149 There are distinct differences between the AIRS V7 and CLIMCAPS V2 algorithms regarding 150 cloud retrievals, summarized here. The first major difference is how cloud clearing is iterated in 151 the retrieval flow. The second major algorithm difference is quality control (QC) procedures when 152 https://doi.org/10.5194/amt-2021-391 Preprint. Discussion started: 15 December 2021 c Author(s) 2021. CC BY 4.0 License. 1) the physical retrieval of atmosphere and surface is not successful, and 2) the final-stage cloud 153 clearing is not successful (Susskind et al. 2014). The third major difference is the choice of the 154 prior states for the two algorithms. The AIRS Science Team algorithms, including both V6 and 155 V7, iterate cloud clearing multiple times, and cloud parameters are determined after the last 156 iteration of cloud clearing using the retrieved surface and atmospheric conditions (Fetzer et al. 157 2020). In contrast, CLIMCAPS V2 performs a single cloud clearing pass and cloud properties are 158 retrieved using the surface and atmospheric parameters from successful retrievals of surface and 159 atmospheric properties Barnet 2019, Thrastarson et al. 2021). The QC procedure used 160 in the two sounder cloud retrievals are also different. AIRS V7 produces case-by-case QC 161 indicators for each retrieved variable; while CLIMCAPS V2 derives one QC value based on the 162 cloud clearing and retrieval status of temperature and water vapor, and the same QC value is 163 assigned to all retrieved variables for the given FOV, including the cloud parameters. Particularly, 164 in AIRS V7 cloud retrieval process, the final stage of cloud clearing and cloud retrievals uses the 165 surface and atmospheric variable retrievals, except for cases over ocean when the retrieved surface 166 temperature differs from the first guess by more than 5 K. For these cases, the surface temperature 167 and surface emissivity from the a priori are used instead, and cloud properties retrieved under this 168 condition are flagged as valid with QC=1, indicating successful cloud retrievals but potentially 169 higher uncertainty than QC=0. This surface test effectively filters out cases when the cloud top is 170 misidentified as surface and causes extremely small ECF values for overcast cloudy conditions 171 over ocean. For ~1% of cases the final cloud retrieval step does not complete successfully, and a 172 QC=2 flag is assigned to cloud parameters to indicate invalid retrievals. As a result, the AIRS V7 173 cloud retrievals produce a much higher percentage of cases with successful cloud retrievals (cloud 174 variable QC=0 or QC=1) than its temperature and water vapor profile products. For CLIMCAPS 175  (Seemann et al. 2008), which is based on the monthly climatology of MODIS land surface 186 emissivity product (MOD11) in 2008 (Thrastarson et al. 2021). The CLIMCAPS system (Smith 187 and Barnet 2020, Smith et al. 2021), instead, uses concurrent fields from the Version 2 Modern-188 Era Retrospective analysis for Research and Application (MERRA-2, Gelaro et al. 2017) as the a 189 priori and implements the Combined ASTER (Advanced Spaceborne Thermal Emission and 190 Reflection Radiometer) and MODIS Emissivity database for land surface (Hook 2019). Over 191 ocean, both systems use the Masuda IR sea surface emissivity model (Masuda et al., 1988) as 192 modified by Wu and Smith (1997). Since the a priori temperature, water vapor, and surface 193 properties are used in the cloud clearing step, differences in the a priori contribute to the 194 differences between the retrieval products, including cloud properties (Yue andLambrigtsen 2020, 195 Yue et al. 2021). Cloud clearing plays an important role in both retrieval systems, and physical 196 retrievals of surface and atmospheric parameters are obtained from the cloud cleared radiances, 197 which, in turn, impact the determination of cloud properties. 198 In addition to these major differences, the two sounder retrieval systems differ in the prior 199 estimates used for ECF and CTP. CLIMCAPS starts the cloud retrieval with background estimates 200 of 0.5 and 0.25 ECF at 350 hPa and 800 hPa CTP for the upper and lower cloud layers, respectively. 201 AIRS V7 uses 1/6 ECF at 350 hPa for the upper layer, and 1/3 ECF at 850 hPa (or 100 hPa above 202 surface in elevated terrain) for the lower cloud layer. However, since the final cloud retrievals of 203 both systems are shown to diverge significantly from their prior (Yue andLambrigtsen 2020, Yue 204 et al. 2021), it is unlikely that different cloud prior estimates are a main contributor to the sounder 205 cloud retrieval product differences. 206 Although their spectral resolution is coarser than that of AIRS and CrIS, instruments like 207 MODIS and VIIRS provide high spatial-resolution cloud properties through information in 208 multiple narrowband channels covering the visible and IR spectral regions. However, significant 209 differences exist between the two imagers. MODIS measures the reflectance or radiance in 36 210 spectral bands, while VIIRS has an analogous subset of these bands (20 channels) plus a day/night 211 visible channel (Oudrari et al. 2015). The lack of near-IR and IR water vapor and CO 2 absorption 212 channels in VIIRS has important implications on the available information content for clouds with 213 respect to MODIS. This impacts the determination of clouds, especially the detection of multi-214 layer clouds and clear sky in polar night conditions, and the determination of cloud thermodynamic 215 phase. It also impacts the retrieval of cloud-top properties, especially for high thin clouds. 216 Moreover, the difference of spectral location of the VIIRS 2.25 μm channel compared to the 217 analogous 2.13 μm MODIS channel has implications on the retrievals of cloud particle size, optical 218 depth, and thermodynamic phase (Platnick et al. 2020). On the other hand, VIIRS provides a higher 219 spatial resolution of 750 m at nadir in cloud property retrievals, compared to the 1-km resolution 220 in the Collection 6.1 MYD06 and cloud mask products. In addition, VIIRS has an onboard detector 221 aggregation scheme that limits the across-swath pixel growth. VIIRS edge of scan pixel size is 222 roughly 1.625 km x 1.625 km versus roughly 2km x 4.9 km for MODIS (Platnick et al. 2021). The 223 MYD06 products have been shown to provide stable and well characterized cloud data records 224 since 2002 (e.g. Yue et al. 2017). Given these instrument differences between MODIS and VIIRS, 225 and a need to develop a continuous data record extending beyond the MODIS era, the MODIS-226 VIIRS CLDMSK cloud mask (Frey et al. 2020) and CLDPROP cloud-top and optical property 227 (Platnick et al. 2021) continuity algorithms were developed. By applying common algorithms to a 228 subset of channels available on both instruments, the continuity algorithms accommodate the 229 detailed channel differences between the two instruments while maximizing the information 230 content on cloud parameters. 231 The continuity CLDPROP products have direct heritage with the Collection 6.1 MODIS 232 atmosphere cloud retrievals (MYD06), with cloud-top property datasets provided by the CLouds 233 from AVHRR (the Advanced Very High Resolution Radiometer) -Extended (CLAVR-x) 234 processing system (Heidinger et al. 2012(Heidinger et al. , 2014. CLAVR-x produces cloud phase reported as 235 Cloud_Phase_Cloud_Top_Properties in the MODIS-VIIRS continuity cloud products. It replaces 236 the MODIS CO 2 slicing solution for cloud top pressure retrievals for cold clouds with an IR-237 window channel optimal estimation approach coupled with a Cloud-Aerosol Lidar and Infrared of two sounders (Manning and Aumann 2015). In order to ensure a close match between the 253 instruments, the following criteria are used to identify candidate SNOs: 254 • FOV centers between Aqua-AIRS and SNPP-CrIS are within 8 km; 255 • Observations are made within 10 minutes; 256 • Both instruments observe within 3.3° of nadir, which corresponds with +/-1 FOR 257 of AMSU for AIRS or ATMS for CrIS. 258 259

Pixel-scale collocations of imagers and sounders: 260
Utilizing the multi-sensor capability at the pixel scale requires accurate and computationally 261 efficient collocation of sounder and imager measurements. Various collocation methods exist 262 (Schreier et al. 2010, Nagle and Holz 2009, Yue et al. 2013. In this study, the method developed 263 by Wang et al. (2016) is applied by matching the instantaneous multi-sensor observations directly 264 based on line-of-sight (LOS) pointing vectors, defined as the vector from the satellite position to 265 the Earth surface pixel location. The details of this method and its accuracy are discussed at length 266 in Wang et al. (2016). 267 In this study, the same collocation method is applied to both Aqua and SNPP to match the finer 268 resolution imager pixels (MODIS and VIIRS) within a given sounder FOV (AIRS and CrIS). The 269 LOS vectors are calculated using the geolocation datasets for different sensors, which contain 270 latitude, longitude, satellite range, satellite azimuth and zenith angles. Collocation is performed 271 using the criterion that the angular difference between the LOS vectors for sounder and imager 272 should be less than half of the sounder FOV size angle. The CrIS FOV is treated as a 0.963° circle 273 which corresponds to ~41% of the peak response and collects ~98% of total radiation falling on 274 the detector (Wang et al. 2013). AIRS has a FOV half-power width of 1.1° (Fishbein et al. 2001). 275 However, 0.963° is used for both AIRS and CrIS in the collocation. After obtaining collocation 276 indices, the L2 cloud properties from both the imagers and sounders are populated accordingly. 277 The high spatial resolution information from MODIS and VIIRS is retained using higher statistical 278 moments and frequency distributions of cloud properties retrieved by imagers within collocated 279 sounder FOV. These statistical metrics include the mean, standard deviation, skewness and 280 kurtosis of MODIS and VIIRS cloud properties, the occurrence frequency of cloud types and cloud 281 phase reported by the cloud mask and cloud thermodynamic phase variables, and joint histograms 282 on the COD and CTP two-dimensional space following the convention of the International Satellite 283 Cloud Climatology Project (ISCCP, Rossow and Schiffer 1999). In addition to summarizing fine 284 imager spatial information over a coarser resolution sounder instrument, these statistical metrics 285 physically describe a variety of cloud processes at both regional and global scales for a range of 286 cloud types in different climate regimes, which are particularly relevant to sub-grid cloud 287 parameterization in numerical models (e.g. Zhu and Zuidema 2009, Kawai and Teixeira 2010and 288 2012, Kahn et al. 2017). The ISCCP-type of joint histograms have been widely used to dissect the 289 uncertainty of the cloud radiative forcing (e.g. Pincus et al. 2012) and climate feedback (e.g. 290 Zelinka et al. 2012, Yue et al. 2016 and 2019) by cloud regimes (e.g. Oreopoulos et al. 2016). 291 By combining the SNOs and the sounder-imager collocated datasets, a multi-sensor multi-292 satellite investigation is conducted to evaluate, at pixel scale, the self-consistency of cloud 293 properties, to benchmark data continuity from the US polar-orbiting operational environmental 294 satellites. 295 296

Results 297
Both Aqua and SNPP are in the 1:30 PM local equatorial crossing time sun-synchronous polar 298 orbits, but at different altitudes. This altitude difference gives a ~2.667 day repeating pattern for 299 AIRS and SNPP-CrIS observations at the same location. Accordingly, the number of SNOs 300 between these two IR sensors varies with time and a large fraction are located at the high latitudes. 301 In this study, seven focus days in January 2016 are selected for their large numbers of SNO pairs 302 and the full operation for all four instruments. Table 2 lists the focus days and gives the number of 303 observations obtained on each day. Figure 1 shows the latitudinal distribution of the focus day 304 SNOs (black bars, y-axis on the left, Table 2). A significant number of observations (>2,500) are 305 available at all latitudes, including the midlatitudes and tropics where SNOs are harder to obtain. 306 it is difficult to directly compare the mean cloud properties retrieved by imagers and sounders, 312 AIRS V7 produces similar general patterns of latitudinal variation of cloud frequency with the 313 imager products, which shows peaks of cloud occurrence in the tropics and midlatitude storm 314 tracks, and troughs in the subtropics. However, CLIMCAPS V2 cloud retrievals do not show these 315 variations, and its mean ECF values are much lower than AIRS V7 at all latitudes. A higher 316 percentage of cloud frequency in the low latitude regions is reported by AIRS V7 than by imagers, 317 consistent with previous findings showing higher sensitivity of hyperspectral IR sounders to 318 optically thin clouds (Kahn et al. 2014, Yue et al. 2016). An increase of COD with latitude at mid 319 to high latitude regions is detected by imagers, compared to a nearly flat or even decreasing mean 320 ECF retrieved by the sounders. These differences will be further assessed in the following 321 discussions. 322 323

Clouds retrieved by hyperspectral IR sounders 324
In Fig. 1, overlapped with the SNO count histograms are the occurrence frequency of 325 sounder FOVs (colored lines, y-axis on the right) for four composites that satisfy the following 326 four conditions, respectively: ECF > 0.01(general cloudy condition), ECF ≤ 0.01 (clear or very 327 thin clouds), ECF > 0.8 (overcast or very thick clouds), and cases with successful CTP retrievals 328 (QC for CTP is 0 or 1). These ECF values are selected based on the relationships between clouds 329 and the IR sounder spectral information, as well as the retrieval uncertainty. The fraction of the 330 highest quality atmospheric state retrievals below clouds, obtained from IR spectral information, 331 decreases with higher ECF (Fetzer et al. 2006). The combination of IR and MW radiances can 332 facilitate the retrieval of vertically resolved temperature and humidity profiles up to ECF of 333 0.7~0.8 (Yue et al. 2011, Yue and Lambrigtsen 2020, Yue et al. 2021. The ECF of 0.01 is often 334 used as the threshold of cloud detection by IR sounders (e.g. Kahn et al. 2014). Moreover, it has 335 been shown that AIRS V7 cloud retrievals present higher uncertainty on thin, broken clouds and 336 cloud edges when ECF < 0.01 (Yue and Lambrigtsen 2020). 337 For each composite, the occurrence frequency is calculated as the percentage of AIRS or 338 CrIS FOVs with successful cloud retrievals that satisfy the composite condition relative to the 339 total number of FOVs in each latitudinal bin. The QC flags for each cloud parameter are reported 340 in the L2 products and used to determine whether the algorithm reports a successful cloud 341 retrieval (when QC = 0 or 1). Different colors are used to indicate retrieval algorithms for the 342 two sounders. Since AIRS V7 and CLIMCAPS retrieve cloud properties up to two cloud layers 343 over each IR sounder FOV, an effective CTP is calculated as the weighted mean CTP by the 344 ECF reported at each cloud layer. 345 These results show large differences between the AIRS V7 clouds with those from CLIMCAPS. 346 AIRS V7 produces a much larger number of cloudy observations (solid pink line in Fig. 1) and a 347 higher yield for CTP retrievals (dash dotted line, Fig. 1), except in the Antarctic region. The 348 magnitude of this difference reaches up to 30% over the Southern Hemisphere and the tropics. 349 Furthermore, AIRS V7 produces much more overcast or very thick clouds (dash lines, Fig. 1) but 350 fewer clear or very thin cloudy cases (dotted lines, Fig. 1) than CLIMCAPS, which is consistent 351 with smaller mean ECF and lower cloud frequency in the tropics and midlatitude storm track 352 regions by CLIMCAPS V2 in Fig. 2. As discussed previously, this is related to the differences 353 between the two algorithms for AIRS in cloud clearing and cloud retrieval QC, as well as the use 354 of different a priori. These differences are further evaluated in the following sections using the 355 imager observations. 356 Despite the differences of sensors, satellites, and spectral resolutions, the three CLIMCAPS 357 Version 2 retrievals evaluated in this study present similar latitudinal distributions of the cloud 358 property distribution and cloud detection. As seen from Fig. 1 with AIRS V7 (y-axes) for both ECF and CTP. The generally good agreement among the 373 algorithms and sensors, especially for CTP, is encouraging, which shows the robustness of these 374 products and consistency of information for clouds in hyperspectral IR sounders. However, 375 CLIMCAPS reports a large number of cases with ECFs between 0 and 0.1, for which AIRS V7 376 reports ECFs ranging from 0 (clear sky) and 1 (completely cloudy). This issue is further illustrated 377 in Fig. 4. For cases where CLIMCAPS-Aqua V2 retrieved ECF is less than 0.1, AIRS V7 (the 378 magenta line) shows two peaks in the ECF occurrence frequency. The first peak is located at V7 379 ECF < 0.1, indicating the two algorithms agree with each other in cloud amount detection. The 380 larger second peak shows that more than 25% of cases with CLIMCAPS ECF < 0.1 have AIRS 381 V7 ECF values of 0.8~0.9. As a result, the correlation coefficient (r) between ECF retrievals from 382 AIRS V7 and CLIMCAPS V2 is only 0.27, which increases to 0.79 when neglecting ECF < 0.1 383

observations. 384
A tighter agreement between CLIMCAPS V2 and AIRS V7 is seen for CTP retrievals as shown 385 by points densely located along the identity line in Fig. 3. The correlation coefficients between 386 CLIMCAPS-Aqua and AIRS V7 CTP are 0.69 for all cases and 0.92 for ECF > 0.1, respectively. 387 High cloud cases (AIRS V7 CTP < 440hPa) show a much higher CTP correlation (r = 0.87) than 388 for low clouds (AIRS V7 CTP > 600 hPa, r = 0.43). When both algorithms identify low clouds in 389 the FOV, CLIMCAPS reports a slightly lower cloud top (larger CTP) than AIRS V7, with a median 390 value difference of 12 hPa; whereas for high clouds, CLIMCAPS V2 reports a higher cloud top 391 with its median CTP 13 hPa smaller than the one by AIRS V7. 392 In the next section, these differences among the various sounder cloud retrieval products are 393 further evaluated using the cloud parameters determined by collocated MODIS and VIIRS data. As shown in Fig. 4, more than 50% of these cases are optically thick clouds with large cloud 406 amount (ECF > 0.7) reported by AIRS V7 and COD values ranging from 2 to 10 by MODIS and 407 VIIRS. Secondly, the comparisons between CLIMCAPS and imager cloud products do not have 408 the cluster corresponding to cases with both high ECF and large COD values, as in the comparison 409 between AIRS V7 and imagers. As discussed previously, this is related to misidentification of 410 cloudy cases as clear or thin cloud conditions by CLIMCAPS. However, another main cause is 411 that CLIMCAPS cloud retrievals have the same QC flags as the physical atmospheric state 412 retrievals; as a result, cases with large cloud amount are filtered out. In general, AIRS V7 products 413 exhibit better agreement with MODIS and VIIRS in detecting cloud amount and occurrence. 414 CLIMCAPS V2 cloud retrievals could be further improved with better cloud clearing flow and 415 more careful treatment when retrieving clouds with unsuccessful atmosphere physical retrievals. 416 The sounder and imager CTP retrievals are also compared in the bottom rows of Fig. 5 and 6. 417 Despite instrument and algorithm differences, when both sounder and imager detect high clouds 418 (CTP < 440 hPa, including ECF < 0.1 cases), CTP retrievals agree with each other well. The 419 correlation coefficients with MYD06 CTP are 0.77, 0.52, and 0.62 for AIRS V7, CLIMCAPS-420 Aqua, and CLIMCAPS-SNPP-FSR, respectively. When imagers detect low clouds (CTP > 680 421 hPa), IR sounders determine the majority of cases as low clouds but with a tail toward CTP values 422 corresponding to high and mid-level clouds (middle row). The disagreement mainly occurs when 423 sounder retrieved ECF is less than 0.1 as shown by the magenta contour lines. These are cases 424 when larger uncertainty in infrared cloud retrieval exists, as discussed previously. After removing 425 these cases, the sounder-imager discrepancy in the low cloud conditions is reduced greatly (bottom 426 row), especially for AIRS V7. These differences are consistent with the known limitation of 427 imagers such as MODIS, which tend to miss high and thin cloud layers (Holz et al. 2008) when 428 compared with AIRS (Kahn et al. 2014). However, the analysis presented here cannot completely 429 rule out the impact of uncertainty in the IR sounder cloud retrievals. When both hyperspectral 430 sounders and narrowband imagers detect low clouds, sounders tend to retrieve smaller CTP than 431 imager. For AIRS V7, the median difference in this condition is -65, -77, and -80 hPa with MYD06, 432 CLDPROP_MODIS, and CLDPROP_VIIRS products, respectively. good agreement between MODIS and VIIRS, and between the MYD06 and continuity products is 438 seen. All correlation coefficients are greater than 0.8. For the three cloud parameters, correlation 439 is always the highest between products derived from the same instrument (MYD06 and 440 CLDPROP_MODIS), and the lowest between MYD06 and CLDPROP_VIIRS (but still reaching 441 0.81, 0.88, and 0.81 for COD, CTP, and Re, respectively) when both instrument and algorithm are 442 different. From the same instrument MODIS but different algorithms, the correlation is lowest for 443 CTP retrievals (r = 0.89) compared to COD (r = 0.97) and Re (r = 0.97). This is because MYD06 444 and the continuity cloud algorithm uses different methods and spectral channels to determine CTP. 445 However, a relationship near one-to-one is seen, indicating the consistency between the 446 operational and continuity cloud products from MODIS, at least for the cloud properties averaged 447 at the sounder resolution (~13.5km). Correlations between MODIS and VIIRS cloud products are 448 lower than those from MODIS alone (with different algorithms), even when both products are 449 derived from the same continuity algorithm. The degradation of agreement is larger for COD and 450 Re than for CTP (Fig. 6). This reflects the effect of spectral channel and spatial resolution 451 differences between MODIS and VIIRS, as well as the related adjustments made to the continuity 452 algorithms, such as the liquid phase LUT for cloud microphysical retrievals. Another possible 453 factor is the collocation error existing in the SNOs, but this is ruled out since results with more 454 conservative collocation criteria remain largely the same (not shown). 455 To further analyze the differences between the imager cloud products and the subpixel cloud 456 heterogeneity over the sounder FOVs, the standard deviation and skewness of the imager cloud 457 property distributions over the sounder FOVs are shown in Fig. 8 and 9, respectively. Correlations 458 are weaker in these higher statistical moments, yet for standard deviation they remain larger than 459 0.6. Similar to comparisons for mean values, tight one-to-one relationships are seen for standard 460 deviation at the sounder FOV scale between the two MODIS cloud products. Similar to mean value 461 comparisons, the CTP standard deviation has the lowest correlation coefficient (r = 0.63) compared 462 to the ones for COD (r = 0.96) and Re (r = 0.87). However, skewness only shows significant 463 correlations for COD (r = 0.78) and Re (r = 0.70) between the two MODIS datasets, but poor 464 correlations (r < 0.3) for CTP. The impact from the differences in CTP algorithms thus shows up 465 more strongly on the higher statistical moments. When evaluating data from different sensors, no 466 correlation is seen for skewness of any of the cloud parameters even with the same retrieval 467 algorithms (Fig. 9, middle and right columns), different from the comparisons using mean value and imager clouds is also found for mid-level and low cloud clusters over ocean (Fig. 11) and for 487 high and mid-level clouds over land (Fig. 12). Over frozen surfaces (Fig. 13), the sounder clouds 488 show optically thin and high clouds, especially in CLIMCAPS V2; a large percentage of mid-level 489 clouds with medium to large ECF values are seen in AIRS V7, more consistent with the cloud 490 histograms from imager observations. However, MODIS and VIIRS cloud detection and retrievals 491 suffer a higher uncertainty over frozen surfaces (Chan and Comiso, 2013), and the small 492 atmospheric thermal contrast with frozen surfaces presents additional challenges for hyperspectral 493 IR sounder retrievals (Yue and Lambrigtsen 2020). Therefore, more accurate cloud measurements 494 from in-situ or active space-borne instruments are needed to further quantify the quality of these 495 imager and sounder cloud retrieval products in snow-and ice-covered regions. circulation (Su et al. 2017), especially in the tropics. Therefore, the differences of the cloud 504 frequency histograms from various imager retrieval products in the tropics are further analyzed 505 here. In Fig. 14, the MODIS continuity product (depicted in Fig. 10) is used as the common base 506 to evaluate the differences caused by algorithms and sensors: 1) between current NASA standard 507 MODIS retrievals and the MODIS continuity algorithms, and 2) between the MODIS and VIIRS 508 continuity cloud data records. The magnitude of joint frequency histogram differences is within 509 ±5% using the focus day observations. MYD06 shows more clouds with CTP < 180 hPa but fewer 510 low clouds with CTP > 800 hPa than the continuity product, consistent with findings in Platnick 511 et al. (2021). VIIRS continuity cloud retrievals produce higher frequencies of clouds with COD 512 between 9.4 and 60, but fewer high clouds with COD < 9.4. Whether and how these differences 513 will impact the long-term trend and short-term variability of clouds as seen by the imagers warrants 514 further study. 515

Cloud thermodynamic phase 516
Both MYD06 and continuity cloud products provide cloud thermodynamic phases (Table 1), 517 given by the optical property retrieval (Cloud_Phase_Optical_Properties, in both MYD06 and 518 continuity products) and the CLAVR-x processing system (Cloud_Phase_Cloud_Top_Properties, 519 continuity products only). The Cloud_Phase_Cloud_Top_Properties variable reports flags 520 determining pixels to be cloud free, water cloud, ice cloud, mixed phase cloud, or undetermined 521 phase. The Cloud_Phase_Optical_Propertes flags indicate cloud mask not determined for pixel, 522 clear sky, liquid water cloud, ice cloud, or undetermined phase, the last of which includes mixed 523 phase clouds (Marchant et al. 2016). AIRS thermodynamic cloud phase, which is available in the 524 AIRS V6 and V7 Level 2 Support product, is based on a set of brightness temperature difference 525 and threshold tests using the channels in 960, 1231, 930, and 1227 cm -1 Kahn 2008, 526 Kahn et al. 2014). These tests are applied to AIRS FOVs where ECF > 0.01, and classify the AIRS 527 FOV as containing liquid, ice, or unknown cloud phases. Detailed comparisons of AIRS cloud 528 phases with CALIPSO indicate good agreement with CALIPSO on ice phase detection, and 529 conservative liquid phase determination Nasiri 2014, Peterson et al. 2020). These studies 530 also show that the unknown class of AIRS cloud phase corresponds to scenes containing both ice 531 and liquid particles, and low-level liquid clouds, especially in the trade-wind cumulus cloud regime. 532 occurrence. For clear sky detection, the cloud-mask clear frequencies from all the imager products 540 are similar except over the frozen surfaces, where VIIRS cloud mask shows 10% higher frequency 541 than MODIS. For IR sounders, AIRS V7 produces significantly lower clear-sky frequency than 542 CLIMCAPS and imager cloud products over non-frozen surfaces. Over frozen surfaces, more 543 frequent clear conditions are reported by AIRS V7 than CLIMCAPS, although AIRS V7 is more 544 consistent with the clear frequency from MODIS and VIIRS data. 545 The frequencies of liquid or ice phase clouds are highly consistent between two cloud phase 546 variables in various imager cloud products, except for ice phase determination over frozen surfaces. 547 This is supported by the low uncertainty range of ice and liquid phase for these four conditions as 548 shown in Table 3. Here uncertainty is roughly characterized by the standard deviation of estimates 549 from different products and variables. The Cloud_Phase_Cloud_Top_Properties reports higher 550 percentage of liquid phase than Cloud_Phase_Optical_Propertes. In particular, the VIIRS cloud 551 top cloud phase product always reports the highest frequency of liquid clouds. From both cloud 552 phase variables, MODIS reports more ice and fewer liquid clouds than VIIRS. When looking at 553 Cloud_Phase_Optical_Propertes for MODIS, ice (liquid) cloud frequency is higher (lower) in 554 MYD06 than in the CLDPROP_MODIS products. The undetermined phase by the 555 Cloud_Phase_Optical_Propertes includes both mixed and uncertain phases (Baum et al. 2012). 556 Except in tropics, MYD06 has the higher frequency of undetermined cases than the continuity 557 cloud products, and this is most prominent over the frozen surfaces with MYD06 reporting ~2.8%. 558 AIRS cloud phase retrievals report a higher frequency of ice clouds than imagers under all 559 conditions, especially in the tropics (Fig. 15) and over land (Fig. 17). However, a much lower 560 frequency of liquid clouds is retrieved by AIRS, which is consistent with a more conservative 561 liquid phase determination approach applied by AIRS cloud phase algorithm (Kahn et al. 2014). 562 The unknown phase of AIRS ranges from ~15% over the frozen surfaces to ~45% over ocean and 563 in the tropics, which corresponds with broken and thin low clouds and scenes with both ice and 564 liquid cloud particles (Jin and Nasiri 2014). 565 566

Summary 567
In this study, the pixel-scale collocation between the hyperspectral infrared (IR) sounders 568 (AIRS and CrIS) and high spatial resolution imagers (MODIS and VIIRS) is performed on the 569 pairs of Simultaneous Nadir Observations (SNOs) between Aqua-AIRS and SNPP-CrIS. Using 570 this approach, the cloud parameters retrieved by various algorithms for IR sounders and imagers 571 from different platforms are evaluated at the pixel level. Quantifying uncertainty in the cloud 572 observational data records is important for constraining the high uncertainty of clouds in weather 573 and climate research. This is also crucial in improving the retrieval of atmospheric, surface, and 574 radiation properties since satellite observations are highly subject to uncertainties and limitations 575 associated with cloud conditions in the instrument field of view (FOV) (e.g. Yue et al. 2013, Wong 576 et al. 2015, Tian et al, 2020. Moreover, narrowband imagers and hyperspectral sounders provide analyses presented here will help to assess the capability of the POR, thus to identify potential 580 gaps existed in the POR for cloud properties. 581 Both the NASA standard and continuity retrieval algorithms for sounders and imagers are 582 investigated here in order to quantify the differences among the retrieval products, and to examine 583 the consistency and continuity of the data products from multiple sensors across different satellites. 584 This is essential to the goal of building a continuous record of satellite data using the Terra, Aqua, 585 SNPP, and JPSS series satellites, with sufficient quality to detect and quantify global 586 environmental change. 587 Multiple cloud parameters are analyzed (Table 1). Comparisons are made by investigating the 588 mean cloud parameters, and higher statistical moments of cloud property distributions measured 589 by MODIS and VIIRS over the corresponding AIRS and CrIS FOV. Cloud types indicated by the 590 joint histograms of cloud properties and cloud thermodynamic phases are included. Through these 591 comparisons, good agreement is found between the sounder and imager retrieved cloud products, 592 yet with distinct differences likely arising from algorithm and sensor differences. For IR sounders, 593 cloud top pressure (CTP) retrieved by AIRS Version 7 (V7) and CLIMCAPS (-Aqua and -SNPP) 594 Version 2 (V2) agree, as shown by correlation coefficients of 0.69 for all cases and 0.92 for cases 595 with effective cloud fraction (ECF) greater than 0.1, respectively. Compared to AIRS V7, 596 CLIMCAPS tends to produce a lower cloud top (CTP 12 hPa larger) for low clouds, but higher 597 cloud top (CTP 13 hPa smaller) for high clouds. However, CLIMCAPS V2 significantly 598 overestimates the frequency of clear and optically thin cloud (ECF < 0.1), relative to AIRS V7 and 599 imager products from both MODIS and VIIRS. This is due to the algorithmic differences between 600 CLIMCAPS V2 and AIRS V7 cloud retrieval algorithms. These differences include whether 601 iteration of cloud clearing is performed, the surface/atmospheric states used in the cloud retrieval, 602 the quality control procedures used, and different a-priori states used by AIRS V7 and CLIMCAPS. 603 How these differences affect the downstream atmospheric and surface retrievals in the two 604 algorithms, and the attribution of impacts from each factor, is beyond the scope of this study and 605 warrants further investigation. 606 High consistency is seen among different imager cloud products, especially in the mean and 607 standard deviation of cloud properties from the MODIS atmosphere cloud property retrieval 608 (MYD06) and the MODIS-VIIRS continuity cloud products (CLDPROP). The magnitude of the 609 correlation coefficients closely reflects the impact of algorithm differences and instrument spectral 610 and resolution differences, with highest correlations obtained between two MODIS products (same 611 sensor but different algorithms) and lowest between MYD06 and CLDPROP_VIIRS (different 612 sensors, different algorithms). The correlation coefficients are always higher for cloud optical 613 depth (COD) and particle effective radius (Re) than for CTP. For mean cloud properties, they are 614 as large as 0.97 between MYD06 and CLDPROP_MODIS, and 0.89 for CTP. For standard 615 deviations within the sounder FOV, the correlations are smaller than those for mean cloud 616 properties, ranging from 0.77 to 0.96 for COD, 0.66 to 0.97 for Re, but only 0.60 to 0.63 for CTP. 617 This is likely due to the fact that completely different CTP retrieval methods are used in the 618 MODIS operational and continuity cloud algorithms to accommodate the lack of near-IR and IR 619 water vapor and CO 2 absorption channels in VIIRS. Such algorithm and instrument impacts are 620 more apparent in the higher moment statistics of cloud properties such as skewness. The 621 correlations of COD and Re skewness between MYD06 and CLDPROP_MODIS drop to 0.78 and 622 0.70, respectively. They are further reduced to below 0.4 when comparing MODIS and VIIRS 623 cloud products. For CTP skewness, the correlation coefficients are less than 0.3. 624 Two different cloud thermodynamic phase retrievals are available from imager observations, 625 which are obtained by the optical property retrieval (Cloud_Phase_Optical_Properties, in both 626 MYD06 and MODIS-VIIRS continuity products) and the CLAVR-x processing system 627 (Cloud_Phase_Cloud_Top_Properties, continuity products only). The frequencies of liquid or ice 628 phase clouds are very consistent between two cloud phase variables in different imager cloud 629 products, with uncertainty usually generally less than 4%. The largest uncertainty is reported for 630 ice phase determination over snow and ice covered surfaces. MODIS retrievals report more ice 631 and fewer liquid clouds than VIIRS, consistent with findings by Platnick et al. (2020). Comparing 632 the two different cloud phase retrievals, the Cloud_Phase_Cloud_Top_Properties reports higher 633 percentages of liquid phase than Cloud_Phase_Optical_Properties, and the 634 Cloud_Phase_Optical_Properties in MYD06 detects higher (lower) frequencies of ice (liquid) 635 clouds than that in the CLDPROP_MODIS products. 636 The general consistency of cloud observations among different sensors aboard Aqua and SNPP 637 from various algorithms is encouraging, especially for achieving a continuous multi-decadal 638 climate data record of clouds that can extend beyond the A-Train era and well into the 2030s with 639 the JPSS series. The quantification of algorithm differences has important implications for future 640 retrieval algorithm developments, and will further improve the capability and accuracy of such 641 climate data records. The authors declare that they have no conflict of interest 668 669

Acknowledgements: 670
The research was carried out at the Jet Propulsion Laboratory, California Institute of 671  surfaces and regions are shown respectively for tropics, ocean, land, frozen surfaces, and global. 932 For each condition, five estimates of cloud phase frequencies are available based on two types of 933 imager-derived cloud thermodynamic phase: Cloud_Phase_Optical_Properties determined by the 934 optical property retrieval (provided in both MYD06 and the two continuity products), and 935 Cloud_Phase_Cloud_Top_Properties obtained through the CLAVR-x processing system applied 936 in the continuity cloud algorithm (provided in the CLDPROP-MODIS and -VIIRS cloud 937 products (dash dotted lines, QC for CTP is 0 or 1). Data from the seven focus days are used (see Table 2 showing the mean ECF from the corresponding sounder retrievals. The bottom row is similar to 990 the middle row except that the cases with sounder ECF < 0.1 are removed from the comparison. 991 Different sounder retrieval algorithms are included. From left to right, data from AIRS Version 992 7, CLIMCAPS-Aqua (C-A), and CLIMCAPS-SNPP FSR (C-S-F) are used. The data points 993 located in regions poleward of 60° are excluded. Cases are included only when both retrievals in 994 comparison (x-and y-axes of the plot) report valid retrievals. The cloud properties from MODIS 995 pixels collocated within the same sounder FOV are averaged before comparing with the IR 996 sounder data. Linear correlation coefficients between the variables on x-and y-axes for different 997 conditions are given in each plot.