Evaluation of the Aqua MODIS Collection 6 . 1 multilayer cloud detection algorithm through 1 comparisons with CloudSat CPR and CALIPSO CALIOP products 2 3

Abstract. Since multilayer cloud scenes are common in the atmosphere and can be an important source of uncertainty in passive satellite sensor cloud retrievals, the MODIS MOD06/MYD06 standard cloud optical property products include a multilayer cloud detection algorithm to assist with data quality assessment. This paper presents an evaluation of the Aqua MODIS MYD06 Collection 6.1 (C6.1) multilayer cloud detection algorithm through comparisons with active CPR and CALIOP products that have the ability to provide cloud vertical distributions and directly classify multilayer cloud scenes and layer properties. To compare active sensor products with an imager such as MODIS, it is first necessary to define multilayer clouds in the context of their radiative impact on cloud retrievals. Three main parameters have thus been considered in this evaluation: (1) the maximum separation distance between two cloud layers, (2) the thermodynamic phase of those layers, and (3) the upper layer cloud optical thickness. The impact of including the Pavolonis-Heidinger multilayer cloud detection algorithm, introduced in Collection 6, to assist with multilayer cloud detection has also been assessed. For the year 2008, the MYD06 C6.1 multilayer cloud detection algorithm identifies roughly 20 percent of all cloudy pixels as multilayer (decreasing to about 13 percent if the Pavolonis-Heidinger algorithm output is not used). Evaluation against the merged CPR and CALIOP 2B-CLDCLASS-lidar product shows that the MODIS multilayer detection results are quite sensitive to how multilayer clouds are defined in the radar/lidar product, and that the algorithm performs better when the optical thickness of the upper cloud layer is greater than about 1.2 with a minimum layer separation distance of 1 km. Finally, we find that filtering the MYD06 cloud optical properties retrievals using the multilayer cloud flag improves aggregated statistics, particularly for ice cloud effective radius.


multilayer (decreasing to about 13 percent if the Pavolonis-Heidinger algorithm output is not 25 used). Evaluation against the merged CPR and CALIOP 2B-CLDCLASS-lidar product shows that 26 the MODIS multilayer detection results are quite sensitive to how multilayer clouds are defined in 27 the radar/lidar product, and that the algorithm performs better when the optical thickness of the 28 upper cloud layer is greater than about 1.2 with a minimum layer separation distance of 1km.

29
Finally, we find that filtering the MYD06 cloud optical properties retrievals using the multilayer 30 cloud flag improves aggregated statistics, particularly for ice cloud effective radius.  (Desmons et al, 2017), in addition to spectral signature differences between monolayer and multilayer cloud scenes determined from forward 50 radiative transfer models (Pavolonis and Heidinger, 2004;Heidinger and Pavolonis, 2005; Nasiri 51 and Baum, 2004;Jin and Rossow, 1997). Several studies have also been dedicated to the 52 inference of cloud optical properties for multilayer cloud scenes, e.g., Watts et al. (2011), 53 Sourdeval et al. (2014) and Chang and Li (2005). Those studies use a two-layer cloud overlapping 54 model approximation coupled with, e.g., optimal estimation, to derive the cloud optical properties 55 associated with the two cloud layers, and thus inherently require robust multilayer cloud detection.

57
Evaluating the performance of multilayer cloud detection algorithms requires appropriate 58 truth datasets and an understanding of the intent of the algorithm itself. For instance, the 59 MOD06/MYD06 multilayer cloud detection algorithm was initially evaluated using forward

108
The algorithm is based primarily on four tests that are collectively used to classify a cloudy 109 pixel as monolayer or multilayer:  2. An above-cloud precipitable water (PW) difference test (ΔPW), using the relative difference 114 between above-cloud PW derived from the CO2-slicing cloud-top pressure result and that 115 derived from the 0.94µm channel with respect to the total PW (TPW) derived from ancillary 116 atmospheric profiles; a relative difference larger than 8% yields a positive multilayer cloud 117 result. 118 3. A second above-cloud PW difference test (ΔPW900mb), similar to the ΔPW test above but 119 assuming the cloud is located at 900mb when deriving above-cloud PW from the 0.94µm 120 channel; again, a relative difference of 8% yields a positive multilayer cloud result.

195
While it is evident in Figure 3 that MYD06 misses a relatively large percentage of multilayer 196 clouds that the radar/lidar merged product detects (7.79% or 11.40% when the PH04 test is 197 included or excluded, respectively), the active sensors are much more capable at detecting multilayer cloud scenes than MODIS. More importantly, as we will see in the next section, in many 199 cases these missed multilayer scenes do not adversely impact the optical property retrieval

205
To better understand the multilayer cloud scenes, we focus on multilayer cloud scenes with 206 only two cloud layers (which represent about 77% of the multilayer cloud population in our co-207 located dataset). Figure 4 shows the probability that MYD06 correctly identifies a multilayer cloud,

215
On the other hand, if the PH04 test is not used (Figure 4b), one can see that the probability of 216 correctly detecting a multilayer cloud scene increases with both d and t. Regardless of the 217 inclusion of the PH04 test, however, the results shown here indicate that it is probable that MYD06 218 will detect a multilayer cloud if the separation distance d is greater than 1km and the upper layer

286
namely those associated with three particle absorptive bands at 2.1, 1.6 and 3.7µm. One can see 287 the differences between the monolayer cloud (blue) and multilayer cloud (red) populations, and 288 that the ice cloud effective radius populations exhibit the largest differences. In particular, the ice 289 cloud effective radius distributions for the multilayer cloud population have a secondary mode at 290 effective radius around 10-15µm. This secondary mode can be explained by a large fraction of 291 cases in the co-located dataset having ice overlapping liquid clouds (see Figure 6, left column).

292
Since liquid droplets are less absorptive than ice crystals in these spectral channels for a given 293 size, identifying these scenes as ice phase can yield smaller ice cloud effective radius retrievals.

294
Indeed, if we remove from the multilayer population those cloudy pixels classified by MYD06 as 295 https://doi.org/10.5194/amt-2019-448 Preprint. Discussion started: 23 January 2020 c Author(s) 2020. CC BY 4.0 License. multilayer, as shown in Figure 9 for cases where MYD06 cloud optical thickness exceeds 4, one 296 can see that the secondary peaks in the ice effective radius distributions for multilayer clouds 297 (red) have disappeared. Therefore, though the MYD06 multilayer cloud detection is not able to 298 detect all multilayer clouds, it can be used to filter cloud effective radius retrievals that are 299 radiatively impacted by multilayer cloud scenes. Even if the PH04 algorithm is ignored in the 300 MYD06 multilayer cloud detection algorithm (Figure 10), the multilayer detection results remain 301 useful for removing most of the differences between the two populations, though some portion of 302 the small ice cloud effective radii remain.

304
If the MODIS cloud optical thickness is lower than 4, the multilayer cloud detection algorithm 305 is not applied since forward modeling indicated that there is not enough information to discriminate 306 monolayer and multilayer clouds (Wind et al. 2010). Figure 11 shows, however, that some 307 noticeable differences remain in the MODIS cloud effective radius distributions for monolayer and 308 multilayer clouds as determined by the 2B-CLDCLASS-lidar products. It is then not possible to 309 directly screen out the cloud effective radius strongly biased by the presence of multilayer cloud 310 scenes as we showed previously.        (2004)