A method is described to classify cloud mixtures of cloud
top types, termed
© 2019 California Institute of Technology. Government sponsorship acknowledged.
There is increasing evidence of secular cloud trends at regional and global
scales in both satellite observations (e.g., Norris et al., 2016) and climate
general circulation model (GCM) simulations (e.g., Zelinka et al., 2013).
The poleward migration of the extratropical storm tracks (Barnes and
Polvani, 2013) is coupled to systematic changes in cloud-thermodynamic-phase partitioning in forced
Kahn et al. (2018) showed that Atmospheric Infrared Sounder (AIRS)
observations of ice cloud optical thickness (
Statistical classification methods are commonly used to define weather
states or cloud types (e.g., Rossow et al., 2005; Xu et al., 2005; Sassen
and Wang, 2008; Wang et al., 2016). For instance, joint histograms of cloud
top pressure and optical thickness from the International Satellite Cloud
Climatology Project (ISCCP; Rossow and Schiffer, 1999) are useful for
relating cloud types to dynamical, radiation, and precipitation variability,
as well as in evaluating climate model simulations (e.g., Klein and Jakob, 1999;
Jakob and Tselioudis, 2003; Rossow et al., 2005; Tselioudis et al., 2013).
Weather states are typically mixtures of conventional cloud types as shown
by Rossow et al. (2005) and Oreopoulos et al. (2014). Partly inspired by
this methodology, we introduce the concept of
As cloud scenes will be matched to coincident A-Train observations, we begin by defining cloud scenes with cloud types derived from CloudSat and observed within an AIRS/Advanced Microwave Sounding Unit (AMSU) (Chahine et al., 2006) field of regard (FOR) of roughly 45 km resolution. One AMSU FOR within an AMSU swath is spatially and temporally coincident with a “curtain” of 94 GHz CloudSat radar profiles. The likelihood of observing clouds is resolution-dependent and is approximately 80 %–85 % at the AIRS footprint scale of 15 km (Krijger et al., 2007; Kahn et al., 2008). The clouds in AMSU sounding FORs or AIRS footprints are more often broken or transparent and less often uniform or opaque. Yue et al. (2013) showed that about 43 % of the AMSU FORs are mixtures of CloudSat-identified cloud types, implying that roughly half of cloudy soundings contain mixtures of cloud types.
Our purpose in this work is to quantify the scale dependence of cloud type mixtures that are then used to understand the cloud complexity within AIRS cloud-phase and ice cloud property data sets. The AIRS and CloudSat data and the collocation approach are described in Sect. 2. To quantify cloud type distributions and their dependence on horizontal scales, the cloud scenes are first characterized at the AMSU FOR resolution in Sect. 3.1, are extended to larger and smaller scales in Sect. 3.2, and key results of the scale dependence are placed into context in Sect. 3.3. The cloud scenes are used to partition AIRS cloud property retrievals into cloud types, specifically, cloud-thermodynamic-phase histograms in Sect. 4.2, and mean values of ice cloud microphysical parameters are described in Sect. 4.3. A discussion, summary, and suggestions for future investigation are found in Sect. 5.
The AIRS/AMSU/CloudSat matchup product described in Manipon et al. (2016) is used by Yue et al. (2013) and in this investigation. The matchup process uses a nearest-neighbor approach to geolocate all CloudSat profiles within either an AMSU FOR at 45 km spatial resolution at nadir view or a single AIRS footprint at 15 km spatial resolution at nadir view (Kahn et al., 2008). Approximately 45 to 50 (15–17) CloudSat profiles coincide with a single AMSU FOR (AIRS footprint), in a swath of width 30 FORs (90 footprints). The cloud scenes are first defined at the AMSU FOR scale and are then extended to other spatial scales. We use a 2-year period of data extending from 1 July 2006 until 30 June 2008 which contains about 8 million AMSU FORs (or 24 million AIRS footprints).
The CloudSat 2B-CLDCLASS product is used in this work and the algorithm is described in Sassen and Wang (2005, 2008). As summarized in Sassen and Wang (2008) and previous works, the algorithm uses methods developed from ground-based multiple remote sensors that have been tested against surface observer-based cloud typing reports. The cloud classification occurs in two steps. First, a clustering analysis is performed to group cloud profiles into cloud clusters. Secondly, classification methods are used to classify clouds into different cloud types. The decision trees guiding the classification are complex and are based on 23 variables derived from the clustering analysis of the first stage. Geometric quantities such as cloud base, top, and horizontal extents are present in decision trees (Sassen and Wang, 2005). Plan view and zonal average frequencies of 2B-CLDCLASS cloud types at its native resolution are reported in Sassen and Wang (2008).
There are eight CloudSat-defined classes in the 2B-CLDCLASS files: cumulus
(Cu), stratocumulus (Sc), stratus (St), altocumulus (Ac), altostratus (As),
nimbostratus (Ns), cirrus (Ci), and deep convective (Dc) clouds, with a
ninth classification of clear sky designated no cloud (nc). Since each AMSU
FOR contains roughly 50 CloudSat profiles with 125 vertical levels each,
there are
Two additional simplifications are made here: variations in the count of
each CloudSat cloud type are not considered, and the observation sequence of
successive cloud types is disregarded. These two simplifications are applied
to the AMSU field of view (FOV). We define a
One advantage of using a classification to define cloud mixtures rather than an unsupervised learning technique, such as clustering, is that the size of the set of possible cloud mixtures is well defined and finite (here it is 256). A related and important advantage of classification is that one can use this set of classes (cloud scenes) with any parameter matched to any given scene. Here, the spatial-scale dependence of those cloud scenes is described in Sect. 3.2.
An alternative approach may consider the vertical layering of cloud types or cloud features, some form of weighting based on counts of each cloud type, or possibly the sequence of cloud types, which may result in different radiance measurements observed by the AIRS instrument (the radiance emitted within an AIRS footprint is nonuniform and channel-dependent, as described in Schreier et al., 2010). However, the simplified approach outlined above is broadly consistent with the sensitivity and sampling characteristics of nadir-viewing passive infrared sounders. Therefore, we consider the approach outlined above to be an appropriate compromise that retains the diversity of cloud scenes and makes the necessary data processing tractable by reducing the dimensionality for ease of interpretation.
Lastly, the results of Kahn et al. (2018) suggest larger ice cloud particle sizes occur at convective cloud tops compared to thin cirrus at the same cloud top temperature. Given the key assumption of cloud typing only at cloud top, the 2B-CLDCLASS product is better suited for identifying convective clouds in AIRS apart from stratiform clouds, the latter of which are dominant in 2B-CLDCLASS-LIDAR. If 2B-CLDCLASS-LIDAR was used in place of 2B-CLDCLASS, the statistics would be weighted towards the detection of vast areas of cirrus in thin layers above and in proximity to convective clouds. The Ci classification dominates in 2B-CLDCLASS-LIDAR at cloud top and will blur the signals of underlying cumulus and deep convective cloud types that are capped by thin cirrus.
The AIRS version 6 cloud-thermodynamic-phase and ice cloud properties (Kahn
et al., 2014) are geolocated to the CloudSat ground track and are binned by
cloud scene. The cloud-thermodynamic-phase algorithm includes two liquid
tests and four ice tests of brightness temperature (
Kahn et al. (2014) describe a retrieval algorithm that is based on optimal
estimation (OE) theory and derives ice cloud optical thickness (
Histogram of cloud scenes containing relative counts of occurrence
observed at the AMSU FOR and AIRS FOV resolution (
A cloud scene is assigned to every AMSU FOR along the CloudSat viewing path using the methodology outlined in Sect. 2. Using the 2 years of data, a total of 194 out of 256 possible cloud scenes are observed but only 18 of the cloud scenes account for 90 % of all observed scenes (Fig. 1a). The four most common scenes contain one cloud type with or without clear sky, and the most common mixed cloud scene (Ac, Sc) is ranked as the fifth most common scene overall. Intuitively, the more diverse a scene, the less frequently it should be observed. The scene that ranked last (18th) in Fig. 1a is (Ci, Ac, Sc). The least frequently observed cloud scene with a ranking of 194 contains six cloud types (Ci, As, Ac, St, Cu, Ns) and was observed only once in 2 years. Of the 256 possible types of cloud scenes, the number of unobserved cloud scenes is 62, of which 61 include St. The unobserved cloud scenes include the only possible cloud scene with eight cloud types together and the seven possible cloud scenes with seven cloud types together.
The unobserved scenes in the 2-year period contain a median of five different cloud types. This is consistent with the improbability of particular cloud types occurring in rapid succession over a few tens of kilometers. The only unobserved cloud scene that does not contain St is (Sc, Cu, Ns, Dc) and is consistent with the conclusion by Sassen and Wang (2008) that Dc (1.8 %) and Cu (1.7 %) clouds are the least frequent of the cloud types. While Dc and Ns are typically associated with different climatological regimes (tropical convection versus extratropical storm tracks), occasionally, Dc is embedded within extratropical cyclones and Ns is classified in stratiform regions of mesoscale convective systems (MCSs). Given the prevalence of Sc and Cu in Fig. 1a, it is somewhat surprising that the combination (Sc, Cu, Ns, Dc) is not observed.
The relative ranking of cloud scenes within the AIRS FOV along the CloudSat track is shown in Fig. 1b for the same sets of matched pixels. A total of 10 cloud scenes account for 90 % of all observed cloud scenes (Fig. 1b). This shows that fewer cloud scenes are found at the smaller AIRS FOV compared to the AMSU FOR.
Geographic distribution of cloud scenes (Sc) and (Ac, Sc) in panels
Figure 2 depicts the geographic distribution of Sc at the AMSU FOR scale, the most observed scene after clear sky (nc), and (Ac, Sc) is the most observed mixed cloud scene. The Sc classification is consistent with the prevalence of stratocumulus clouds in subtropical subsidence regions and trade cumulus in the tropics and subtropics (e.g., Yue et al., 2011). The (Ac, Sc) cloud scene is identified most frequently in the extratropical storm tracks and the transition from shallow cumulus to deep tropical convection.
In Sect. 3.1, the relative frequencies of cloud scenes were derived for exact collocated matches of AIRS and AMSU observations to the CloudSat ground track. As the CloudSat ground track can oscillate across several AIRS FOVs over a scan line within a given orbit, the numbers of coincident CloudSat profiles matching to AIRS and AMSU will vary. Below, cloud scenes are derived independently of the specific AIRS and AMSU collocation geometry.
To investigate the scale dependence of the number of cloud scenes, the approach described in Sect. 2 is modified for a range of horizontal extents between 1.1 and 1000 km. The number of observed cloud scenes calculated at each horizontal scale is shown in Fig. 3a for 10 to 1000 km. At the finest scale of 1.1 km, only eight possible observed cloud scenes or clear sky are expected. When the scale increases, as expected, the number of cloud scenes quickly increases with a total of 143 cloud scenes observed at a scale of 11 km. As horizontal scale is further increased, the probability of observing cloud scenes with only one or two cloud types is reduced. After a maximum number of cloud scenes is obtained at 105 km, the number of cloud scenes will decrease with increasing scale (e.g., 163 cloud scenes at 990 km) until a limiting case is reached at the largest scale with only one cloud scene with all observed cloud types. The number of cloud scenes observed at least once at the AMSU FOR horizontal scale (indicated by the red vertical line on Fig. 3a) is approximately 190.
The 90th percentile calculated at all horizontal scales is shown in Fig. 3b. The 90th percentile of the maximum number of cloud scenes is 33 between 303 and 440 km in horizontal scale. The number at the nominal 45 km AMSU footprint scale is 16 cloud scenes, while the average number at the AIRS footprint is 9 cloud scenes. (Note that these are slightly smaller than values of 18 and 10 using the exact AMSU and AIRS geometry, respectively, in Sect. 3.1.) While these results show that fewer cloud type mixtures are observed at a decreasing length of 45 to 15 km, a variety of cloud type mixtures is still encountered. While infrared sounding at 15 km resolution does not eliminate the cloud scene complexity encountered for combined infrared and microwave sounding at 45 km, the vast majority of 15 km footprints contain a smaller subset of possible cloud mixtures. In Sect. 4, we will determine whether individual cloud types or cloud type mixtures have meaningful impacts on AIRS cloud property retrievals. (Impacts on temperature and specific humidity soundings are beyond the scope of this investigation.)
The reasons for the maximum number of observed cloud scenes (210) at a particular horizontal scale (105 km) are not immediately clear. The scale preference depends on the physical characteristics of cloud regimes and the degree to which cloud types are mixed together by region and furthermore depend on cloud length distributions (Guillaume et al., 2018). A simple model is described below that is able to approximate the results of Fig. 3 and offers some insight for the observed maximum frequency of cloud scenes and the spatial scale at which it occurs.
The goal of this section is to derive cloud scene scale statistics that are
independent of any regular grid resolution and explore whether these
statistics can explain some features of the number of scenes as a function
of scale observed in the previous section. In particular, we explore whether
these statistics can explain the maximum observed around 105 km. There is
however an inherent difficulty in defining the boundaries that delimit any
given cloud scene in the absence of a predefined horizontal extent. It is
possible that within a given cloud scene there exists several scenes with
the same cloud types but differing lengths making the scene identification
ambiguous. To circumvent this problem, we define a cloud scene and its
maximum length as follows.
We search for a cloud scene containing a predefined mixture of cloud types.
The spatial extent of this scene is delimited by cloud types (or clear sky)
on both ends that do not belong to the mixture. The maximum length of a cloud scene is the sum of all the horizontal lengths
of all the cloud types in the cloud scene. If within a given cloud scene there exist several cloud scenes with the
same cloud types but smaller lengths than the maximum length, the minimum
length of a cloud scene is defined as the smallest length of all those
lengths.
For example, imagine that we will calculate the maximum length of the
specific cloud scene (Ac, Sc). We then identify a location in the CloudSat
data record with the following illustrative succession of cloud types:
(Ci, Ac, Sc, Ac, Sc, Ac, Ns), with the number of CloudSat profiles associated with
each cloud type of 10, 3, 6, 5, 7, 12, and 15, respectively. The Ci and Ns
obviously do not belong to the (Ac, Sc) cloud scene and therefore delimit the
scene as defined in (1) above. The maximum length of the cloud scene (Ac, Sc)
will be the sum of the number of CloudSat profiles for (Ac, Sc, Ac, Sc, Ac),
which is
In the example above, there are four possible sequences that
could be the minimum length: (Ac, Sc, …, …, …),
(…, Sc, Ac, …, …), (…, …, Ac, Sc, …), or
(…, …, …, Sc, Ac). The corresponding lengths are
Before steps (1) and (2) are used to quantify the maximum and minimum lengths for each of the 247 mixed scenes (256 minus the 8 single cloud scenes and clear sky), the locations of each cloud scene must first be identified in the 2-year data record. Starting at the first CloudSat profile, the presence of each of the 247 mixed cloud scenes is determined using (1). For each occurrence of each mixed cloud scene, (2) and (3) are then applied to determine the maximum and minimum lengths for each individual cloud scene. After processing the maximum and minimum lengths for every mixed cloud scene, simple statistics are calculated.
Distribution of
A total of 200 out of 247 possible mixed scenes were identified. The minimum and maximum length occurrence frequencies of five cloud scenes – (Ac, Sc), (As, Sc, Cu), (Ci, As, Cu, Dc), (As, Ac, Ns, Dc), and (Ci, As, Ac, St, Sc) – selected randomly from the 200 present in the 2-year record are shown in Fig. 4a and c, respectively. Recall that the maximum length is defined from (2), while the minimum length is defined from (3), with an illustrative example previously described for (Ac, Sc). From top to bottom, their respective ranks are 1, 26, 51, 76, and 101. It is striking that each frequency histogram in Fig. 4a and c is not monotonic and displays a frequency maximum between 100 and 1000 km. Consequently, the sum of all (200) observed mixed scenes across length scales will result in a curve with a maximum, and these are shown in Fig. 4b and d. Both curves are very similar to Fig. 3a and have maxima for about 180 observed scenes at 77 and 174 km, respectively. Using the methodology outlined in (1) to (3) to estimate numbers of cloud scenes, the scale dependence of the number of observed scenes shows that the maximum will occur somewhere between 77 and 174 km.
Cloud type vertical cross section defined by the values of the cloud_scenario variables of the 2B-CLDCLASS product. Each color corresponds to a different cloud type (legend on right). Color segments on top of the figure indicate the horizontal extent of a cloud measured at its top.
In order to shed additional light on why a maximum in the occurrence
frequency of each cloud scene histogram is obtained, histograms of cloud
length frequency of
Horizontal cloud chord length frequency histograms for each of the eight CloudSat cloud types and clear sky. The cloud chord length was obtained at the cloud top (see Fig. 5) unlike that obtained in Guillaume et al. (2018).
The length of a mixed scene is the sum of the lengths of each cloud type
within it. There are two aspects that will influence the number of scenes
observed at a given length
To illustrate the effects of these opposing behaviors, we consider the scene
(As, Sc, Cu) length distribution. Since the minimum length of all cloud
distributions in Fig. 6 is one CloudSat profile, there is only one possible
cloud length combination (
We will now establish differences in the AIRS thermodynamic-phase and ice
cloud properties in the presence of complex and simple cloud types using
coincident cloud scenes. In this section, the scenes are determined at the
AIRS FOV resolution (approximately 15 km). We briefly summarize general
categories of cloud scene statistics in Sect. 4.1. The AIRS cloud-thermodynamic-phase tests are discussed separately for single and mixed
cloud scenes in Sect. 4.2. The AIRS ice cloud
Horizontal cloud chord length median and median absolute deviation for each cloud type (km).
Table 2 summarizes five types of scenes at the 15 km AIRS FOV scale: (i) clear sky, (ii) cloudy sky with one cloud type, (iii) partly cloudy sky with one cloud type, (iv), cloudy sky with multiple cloud types, and (v) partly cloudy sky with multiple cloud types. The raw counts and the relative percentages for the 2-year observing period are shown. The dominance of clear sky (30.7 %) at 15 km is apparent and is consistent with an absence of thin cloud features in the 2B-CLDCLASS data set. Cloudy sky scenes with one cloud type (multiple cloud types) amount to 31.3 % (10.2 %) of all observed scenes, while partly cloudy sky scenes with one cloud type (multiple cloud types) amount to 23.5 % (4.3 %) of all observed scenes. A total of 41.5 % of AIRS FOVs are completely cloudy while 27.8 % are partly cloudy according to 2B-CLDCLASS. Below the differences in cloud-thermodynamic-phase detection and ice cloud property retrievals are quantified for the types of scenes summarized in Table 2.
The occurrence frequency histogram of the sum of all thermodynamic-phase tests is shown for cloudy sky with one cloud type in Fig. 7. Homogeneous cloud scenes serve as an ideal point of reference for establishing cloud-phase sensitivity benchmarks. Overall, there is strong differentiation in the cloud thermodynamic phase among cloud scenes with single cloud types. Ice tests dominate Ci, Ns, Dc, and As, while liquid and undetermined tests dominate Ac, Sc, and Cu.
The ice tests dominate the Ci cloud scenes and reaffirm the sensitivity of
AIRS to ice clouds. CloudSat-classified clear scenes contain occasional
occurrences of AIRS-detected thin cirrus (
AIRS cloud_phase_3x3 histograms for cloudy sky with one cloud type (i.e., all CloudSat profiles have the same cloud type and no clear sky). The red, green, and blue bars indicate liquid, undetermined, and ice phase, respectively. Each histogram sums to 1.0 and does not show how many counts relative to another histogram. Relative counts could be inferred from the percentages listed in the second to left column of Table 3.
The As cloud scene histogram in Fig. 7 is overwhelmingly dominated by ice.
The undetermined cases in part may result from supercooled liquid or
mixed-phase clouds that potentially could be distinguished with an improved
phase algorithm that factors in the spectral midinfrared signature of
supercooled liquid (e.g., Rowe et al., 2013). The Ac and As cloud scene
histograms are very different from each other, with a majority of
undetermined and liquid for Ac and a majority of ice for As, consistent with
aircraft observations (Mazin, 2006). The preponderance of undetermined phase
for Ac may indicate frequent supercooled liquid cloud tops (Zhang et al.,
2010). Ham et al. (2013) showed that Ac are typically 2–3 km lower in
altitude than As, and this probably explains some of the difference in liquid
and ice phase, as lower clouds are usually warmer. The Ns cloud scene
histogram is dominated by ice detection with occasional liquid and
undetermined cloud tops. The Ns cloud scene also has significant height
overlap with Ac and As, with most tops for all three types typically located
below 9 km. Ice tests dominate in the Dc cloud scene histogram although a
very small proportion of
AIRS cloud_phase_3x3 histograms for partly cloudy sky with one cloud type. All else equal to Fig. 7.
The occurrence frequencies of cloud phase for partly cloudy sky with one cloud type are shown in Fig. 8. The biggest change is the relative ordering of the ranks among cloud scene types between Figs. 7 and 8. Ac is now more common than As, as horizontal extent and frequency both explain reordering of rankings in Figs. 7 and 8 (Miller et al., 2014; Guillaume et al., 2018). There are more subtle changes in the cloud-phase histograms that are consistent with partly cloudy sky. A weaker spectral signature for partly cloudy scenes results in slightly greater counts of unknown phase and also subtle shifts in liquid- and ice-phase tests in Fig. 8 compared to Fig. 7. In the Ac cloud scene histograms, there is a small but discernible increase in ice tests in Fig. 8 compared to Fig. 7. Horizontally heterogeneous Ac appears to have more frequent ice detection than horizontally homogeneous Ac.
Total counts and relative percentages of five cloud scene categories at the AIRS FOV scale: clear sky, cloudy sky with one cloud type, partly cloudy sky with one cloud type, cloudy sky with multiple cloud types, and partly cloudy sky with multiple cloud types.
The nine most frequent cloudy scenes with multiple cloud types are shown in
Fig. 9. The (Ci, Sc) cloud scene ice-phase histogram resembles a hybrid of
histograms for Ci and Sc with undetermined phase the most frequent. (Ci, Sc)
is a common cloud scene in the low latitudes as trade cumulus (Sc cloud
type) and is frequently found under thin cirrus (Chang and Li, 2005).
Furthermore, the spectral signatures of the two types of clouds frequently
cancel, giving an undetermined phase result in the spectral tests used here
(not shown). The (Ci, As) cloud scene shows a slight reduction in liquid
detections and a slight increase in ice detections compared to As alone.
While the As cloud scene in Fig. 7 is dominated by
AIRS cloud_phase_3x3 histograms for cloudy sky with multiple cloud types for the top nine ranked cloud scenes in order of occurrence frequency.
The nine most frequent partly cloudy scenes with multiple cloud types are shown in Fig. 10. As with the differences between Figs. 7 and 8, the biggest change is the relative ordering of the ranks among cloud scene types between Figs. 9 and 10. Furthermore, there are additional (yet subtle) changes in the phase test histograms for the cloud scenes that are common between Figs. 9 and 10.
In most mixed cloud scenes in both Figs. 9 and 10, the characteristics of the histograms are similar either to single types or have combined characteristics of the multiple cloud types contained within the cloud scene. These results are encouraging and reaffirm the capabilities of thermal-infrared cloud-phase determination (Jin and Nasiri, 2014) and exhibit consistency with cloud types from the CloudSat radar. We note, however, that the AIRS phase determination has some ambiguity in overlapping ice and liquid cloud layers as previously shown by Jin and Nasiri (2014).
Cloud ice properties for cloudy sky with one cloud type (i.e., all
CloudSat profiles have the same cloud type). Proportions and relative errors
are in percent. The effective radius is in micrometers (
AIRS cloud_phase_3x3 histograms for partly cloudy sky with multiple cloud types for the top nine ranked cloud scenes in order of occurrence frequency.
The mean ice cloud property retrievals are summarized in Table 3 for cloudy
sky with one cloud type only for the ice-only portions of the cloud-phase
histograms depicted in Fig. 7. Scenes identified as clear sky exhibit
properties of a small population of thin cirrus detected by AIRS (Fig. 7)
with mean values of
Cloud ice properties for partly cloudy sky with one cloud type. All else the same as Table 3.
The Ci cloud scene has mean values of
The Ns cloud scene in Table 3 contains larger mean values of
The mean Cu value of
The relative variations between the ice cloud retrieval properties for
cloudy sky with one cloud type in Table 3 are consistent with expectations
of infrared sensitivity. CloudSat-observed Ci cloud scenes have smaller
error estimates and higher information content in comparison to Sc,
consistent with Sc scenes containing tenuous cirrus that goes undetected by
2B-CLDCLASS. Larger
The mean ice cloud property retrievals are summarized in Table 4 for partly
cloudy sky with one cloud type with cloud-phase histograms depicted in Fig. 8. The biggest difference between Tables 3 and 4 is the relative frequency
of occurrence with large differences between cloud scenes with or without
clear sky. Another significant change is an overall reduction in AKs and
magnitude of
The As and Ac cloud scenes in Table 4 are very similar to As and Ac cloud
scenes in Table 3 except for slight reductions in
The differences between Tables 3 and 4 are more significant for the
convective ice clouds, however. The Cu cloud scene
Cloud ice properties for cloudy sky with multiple cloud types for the first nine most observed cloud scenes at the AIRS FOV scale. All else the same as Table 4.
Cloud ice properties for partly cloudy sky with multiple cloud types for the first nine most observed cloud scenes at the AIRS FOV scale. All else the same as Table 5.
The ice cloud property retrievals for cloud scenes that contain multiple
cloud types are summarized in Table 5 for cloudy scenes and Table 6 for
partly cloudy scenes. These tables list the nine most frequent cloud scene
types as depicted in Figs. 9 and 10. Four of the nine cloud scenes are
common between Tables 5 and 6. There is a general tendency for reductions of
To summarize Tables 3–6, larger differences in ice cloud property retrievals are found between different cloud types than between cloudy and partly cloudy scenes. However, the differences between cloud scene types are the sharpest for the subset of cloudy scenes with one cloud type (Table 3). The AIRS cloud property retrievals are not greatly impacted by mixtures of cloud types within the AIRS footprint, and ice cloud property differences among cloud scenes are broadly consistent with the expected performance of infrared retrievals among these cloud types.
A method is described to classify cloud mixtures of cloud top types, termed
The cloud scenes are organized into five categories: (i) clear sky, (ii) cloudy sky with one cloud type, (iii) partly cloudy sky with one cloud type,
(iv), cloudy sky with multiple cloud types, and (v) partly cloudy sky with
multiple cloud types. Summarizing AIRS cloud top property retrievals for
cloudy sky with one cloud type, there is strong differentiation in the cloud
thermodynamic phase. Ice phase dominates Ci, Ns, Dc, and As, while liquid
and undetermined phase dominate Ac, Sc, and Cu. The results are similar for
partly cloudy sky with one cloud type with an increase in unknown cloud
phase and
The relative magnitude of differences in
The fidelity of AIRS-retrieved cloud-phase and ice cloud microphysics was tested within scenes with both uniform and nonuniform cloud cover, as well as one or more cloud types within the scene. As with phase, retrieval differences are shown to be larger among cloud types rather than between uniform and mixed cloud scenes.
New methodologies for simultaneous retrievals of cloud microphysical properties and temperature and specific humidity profiles that include clouds in the forward radiative transfer (e.g., De Souza-Machado et al., 2018; Irion et al., 2018) necessitate careful investigation of the effects of cloud mixtures on retrieved cloud properties. The bias and root-mean square error of AIRS temperature and specific humidity soundings depend on cloud type (Yue et al., 2013; Wong et al., 2015). A more rigorous evaluation of scene complexity is necessary for optimizing the retrieval configuration of future sounding algorithms (Irion et al., 2018) and for validating their products.
This investigation shows that careful inspection of footprint-scale AIRS cloud property retrievals is consistent with expectations of infrared sensitivity to different cloud types defined with the 94 GHz CloudSat radar. Other cloud observations, such as MODIS, may be used in a similar analysis to the one described here. MODIS captures the off-nadir portion of the AIRS swath and the fine-scale variability within AIRS footprints. Wang et al. (2016) used the cloud typing in CloudSat to cross validate with cloud typing using MODIS-defined cloud types. This establishes a link between cloud types obtained from CloudSat and MODIS. A rigorous estimation of the pixel-scale relationships between cloud properties obtained from CloudSat, MODIS, and AMSU will help to further advance multisensor and multivariate geophysical retrievals (e.g., Irion et al., 2018).
CloudSat data were obtained through the CloudSat
Data Processing Center (
AG designed and implemented the cloud scene classification scheme as well as the cloud scene and cloud type scale dependence studies. BK designed and AG implemented the study of AIRS thermodynamic-phase and ice cloud properties as a function of cloud scenes and types. GM designed, implemented, and generated the AIRS/AMSU/CloudSat matchup product. GM, BW, and HH designed the data system that generated this product. AG and BK prepared the manuscript with contributions from all coauthors.
The authors declare that they have no conflict of interest.
Part of this research was carried out at the Jet Propulsion Laboratory (JPL), California Institute of Technology, under a contract with the National Aeronautics and Space Administration. The authors thank two reviewers for helpful and insightful comments that led to an improved manuscript. This project was supported by NASA's Making Earth Science Data Records for Use in Research Environments (MEaSUREs) program.
This research has been supported by NASA (grant no. NNH17ZDA001N).
This paper was edited by Alexander Kokhanovsky and reviewed by two anonymous referees.