The 532 nm calibration coefficients in V4 are systematically lower than the V3 values by 3 % to 12 %, depending on latitude, season, and lighting conditions (Kar et al., 2018; Getzewich et al., 2018). These lower calibration coefficients increase the magnitude of the V4 532 nm attenuated backscatter coefficients, thereby facilitating the detection of optically thinner layers than were previously detected in V3. One notable side effect of this improvement is the increased occurrence of weakly scattering features located along the edges of ice clouds. These features are detected at 20 or 80 km horizontal averaging resolutions, and, as illustrated in Fig. 6, occur at the horizontal edges and along the lower boundaries of more robust cirrus layers that are detected at 5 km resolution.

Because these features are, by definition, always found adjacent to cirrus clouds, they are referred to as “cirrus fringes”, which are a new feature in the V4 dataset. As described earlier, 1 month of volcanic aerosol data were added to the 1-year training dataset. These additional training data helped better constrain the characteristic PDF parameters (especially 〈β532′〉0 and χ0′) at high altitudes, and the new V4 aerosol PDFs are now more sensitive to lofted depolarizing aerosols at high altitudes. As illustrated in detail in Sect. 5, this increased sensitivity helps better differentiate lofted aerosol layers from clouds. At the same time, however, there is a cost to be paid for this increase in sensitivity, as the V4 CAD algorithm preferentially classifies cirrus fringes as depolarizing aerosols, not clouds.

Figure 7 shows joint histograms of χ′ and log⁡10(〈β532′〉) for cirrus fringes and the adjacent cirrus clouds for data acquired between 40∘ N and 40∘ S during January, February, and December 2008. The relationship between the two cluster centroids is similar to the separation seen in Fig. 2; just as the aerosols in Fig. 2 exhibit distinctly lower χ′ and log⁡10(〈β532′〉) values than the water clouds, the fringes in Fig. 7 likewise exhibit distinctly lower χ′ and log⁡10(〈β532′〉) values than the cirrus (this is evidenced by probability contours added in Fig. 7). In short, based solely on χ′ and log⁡10(〈β532′〉), the optical properties of fringes are generally more similar to aerosols than clouds, and thus the possibility exists that at least some of these fringes are dust or smoke. However, by definition these fringes do not appear as discrete layers, but instead as areas that are in direct contact with a more strongly scattering layer that has been previously classified as an ice cloud. They are most often detected below semitransparent cirrus clouds, where the attenuation is large but the signal-to-noise ratio (SNR) remains larger than below attenuating cloud regions (see Fig. 6).

The fact that the color ratios in fringes are smaller than in the adjacent cirrus clouds cannot be explained by differential transmittance effects, as the same attenuation at both wavelengths is usually observed. One possible explanation for the change in optical properties seen in cirrus fringes is a reduction in size of the crystals linked to sublimation. The depolarization of the fringes is usually large (>20 %), and fringes are quite frequently observed at latitudes far from dust sources, as demonstrated in Fig. 6. The pervasive presence of fringes and their immediate spatial proximity to layers previously identified as cirrus strongly suggest that these features are most likely also cirrus, and not aerosol layers.

To rectify these perceived misclassifications by the CAD algorithm, a “cirrus fringe amelioration” algorithm has been developed and added as a V4 CAD post-processor. To be identified as a cirrus fringe, a layer must (a) be initially classified by the CAD algorithm as an aerosol, (b) be detected at a 20 or 80 km horizontal averaging resolution, (c) be in direct contact with one or more layers detected at finer resolution and classified as cirrus, (d) have an attenuated backscatter centroid temperature below 0 ∘C, and (e) have a base altitude higher than 4 km above ground level. Layers that meet all of these criteria are classified as cirrus fringes and given a special CAD score of 106.

Application of these simple criteria shows that fringes are ubiquitous within the V4 dataset. For the data acquired between 60∘ N and 60∘ S during 2014, 22 % of all unique layers detected at 20 and 80 km averaging resolutions with bases between 4 and 16 km are identified as cirrus fringes. Note, however, that the fringe amelioration algorithm is not executed if, within an 80 km horizontal extent, 35 % or more of the features with bases above 4 km are originally classified as aerosol. This restriction prevents the amelioration algorithm from operating in those scenes that are likely to contain legitimate cases of clouds embedded in high-altitude aerosols.

Figure 6 demonstrates the routine impact of the fringe amelioration algorithm on CAD. These observations occurred over the remote South Pacific Ocean in June 2008, when dust and volcanic aerosols are not expected at high altitudes, as confirmed by back-trajectory analyses (not shown). Comparing the CAD in V3 (Fig. 6b) to that of V4 without cirrus fringe amelioration (Fig. 6c) shows an increase in the fraction of aerosols found adjacent to cirrus clouds. After application of the cirrus fringe amelioration algorithm (Fig. 6d), the majority of the cirrus fringes are reclassified as cloud rather than aerosol. However, as demonstrated in Fig. 6d, some likely misclassifications of cirrus fringes remain in V4 due to limitations of the amelioration algorithm. Because their depolarization ratios are relatively large, these misclassified layers are most often classified as dust by the aerosol subtyping algorithm (Kim et al., 2018), which may introduce some bias in elevated dust occurrence (see further analyses in Sect. 5).

When applied to features detected at single-shot resolution, the V4 CAD algorithm can encounter difficulties in correctly assigning confidence levels to the classifications of dense water clouds lying beneath thick smoke layers. This situation often happens over the Atlantic Ocean off the west coast of southern Africa during the biomass burning season every year (June–September) (Remer et al., 2008; Chand et al., 2009; Das et al., 2017). The example shown in Fig. 8 occurred on 6 September 2008 at 1:35:29 UTC to the west of the African continent. Smoke attenuates the signal at 532 nm more strongly than at 1064 nm, leading to attenuated backscatter color ratios for the underlying water clouds that often far exceed the expected color ratio of 1.16±0.05 that would be measured in the absence of overlying smoke (Vaughan et al., 2015). The spectrally dependent attenuation of the smoke is clearly evident from the attenuated backscatter color ratio measurements in Fig. 8b that increase dramatically with increasing penetration into the denser parts of the smoke layer (the color changes from orange to red to purple to gray between latitudes ∼6 and ∼13∘ S).

Figure 9 shows the joint distribution of the overlying integrated attenuated backscatter and attenuated color ratio for the water cloud stratus deck below the extensive smoke plume from ∼18 to 0.5∘ S. As can be seen, the attenuated color ratios can reach very high values (2 to 6 times the typical value) for these water clouds because of differential attenuation of the signal at the two CALIOP wavelengths. Because extinction coefficients are not retrieved for layers detected at CALIOP's single-shot resolution (Young and Vaughan, 2009), the attenuated backscatter coefficients in the water cloud cannot be corrected for the overlying signal attenuation from the smoke. Consequently, the water cloud color ratios are abnormally high and entirely inconsistent with the characterization of clouds with no overlying smoke layers. While these features are still classified as clouds, they are assigned very low CAD scores (<20) and thus effectively transformed from high-confidence cloud layers identified in V3 (for which, by default, CAD = 100 for all layers detected at single-shot resolution) into low- and no-confidence cloud layers in V4.

The CAD algorithm is designed so that any layers that fall in the overlap region of aerosol and cloud PDFs or which have incorrect or unphysical parameters due to artifacts introduced in the measurement and/or data processing, as in this case, are assigned low CAD values (Liu et al., 2009). However, the stratus deck over the South Atlantic has been widely observed and studied, and researchers worldwide are highly confident that these layers are unquestionably clouds (e.g., Sakaeda et al., 2011; Schrage and Fink, 2012). Because their identification as low- or no-confidence clouds by the V4 CAD algorithm would thus underestimate the true confidence in classifying these layers, a second CAD post-processor algorithm was designed to rectify the situation.

The classification errors arise primarily from the unphysical cloud color ratios that result from the differential signal attenuation at the two wavelengths as the laser pulses pass through the smoke layers. Therefore, following the assignment of initial CAD scores by the CAD algorithm, a special-purpose algorithm applies additional analyses to clouds with the following attributes: 0≤ CAD score ≤20, 1.4≤χ′≤10, χ′ relative uncertainty <500 %, overlying layer-integrated attenuated backscatter at 532 nm (γ532′) between 0.01 and 0.05 sr-1, and mid-layer temperature >0 ∘C. This set of parameters was chosen to select only those layers that might be classified as high-confidence water clouds were it not for their suspiciously high color ratios. For these layers, χ′ is temporarily reset to 1.10 and the CAD score is recalculated. Only χ′ is changed; all other parameters remain the same as in the original calculation. If the feature is still classified as a cloud in this second assessment by the CAD algorithm (a typical result, but not guaranteed), the CAD score is reset to the newly calculated value, and χ′ is restored to its original value. Otherwise, the original CAD score remains in effect. Implementing this procedure effectively eliminated the majority of this kind of anomalous classification and was particularly effective in the smoke source region in southern Africa as well as the transport region off the coast.

As described in the opening paragraph of Sect. 3, nominal values for the CAD scores reported in the V4 CALIPSO data products range between -100 and 100. However, under special circumstances, the CALIOP scene classification algorithms will assign CAD scores that lie outside this range. These special CAD scores are enumerated in Table 1.

The Taklimakan, located in the Tarim Basin in northwest China at about 40∘ N, is one of the world's major deserts and most prolific dust sources (Prospero et al., 2002). The dust activity over the Tarim Basin area is persistent almost all year long, reaching a maximum in the spring (Z. Liu et al., 2008b; D. Liu et al., 2008). The Tarim Basin is surrounded by high mountains, with the Tian Shan mountains in the north and the Kunlun Mountains in the south and southwest. These mountains create circulations in the basin that are favorable for dust to remain suspended aloft for long periods of time (Tsunematsu et al., 2005). Taklimakan dust can often reach altitudes high into the troposphere and subsequently be transported long distances by westerlies (Huang et al., 2008; Uno et al., 2009).

One of the issues with the V3 CAD was that dense dust layers over the Taklimakan area were often misclassified as cloud when they were lofted to relatively high altitudes and/or transported far to the north (>∼40∘ N), where the occurrence of ice clouds becomes more significant compared to dust (Jin et al., 2014). In these cases, classification skill has been improved in the V4 CAD algorithm by reducing the latitude bands of the PDFs from 10∘ in V3 to 5∘ and optimizing the height-dependent characteristic scattering parameters used in the PDFs.

Dense dust layers detected at single-shot resolution were classified as cloud by default in V3. In V4, the CAD algorithm is now also applied to single-shot layers. This extended application of the CAD algorithm has significantly reduced the misclassification of single-shot dust layers. An example of dense dust over Taklimakan, observed by CALIOP around 20:15:32 UTC on 4 May 2008, is shown in Fig. 11a–c. Multilayered dust (yellow–red–grayish areas) appears between 38.24 and 45.4∘ N and extends from the surface to ∼12 km, with the densest layer (red–grayish area) at 2–4 km. An attenuating water cloud is embedded in the dust at ∼4 km north of 43.6∘ N. While the V3 CAD algorithm correctly identified the water cloud, a large portion of the dust north of ∼40∘ N was misclassified as high-confidence cloud (mostly as ice cloud). As can be seen in Fig. 11c, the V4 CAD correctly classifies these layers as aerosols with good confidence levels (CAD scores between -100 and -50).

Another example of dense dust over Taklimakan, which occurred at 07:08:39 UTC on 7 August 2008, is shown in Fig. 11d–f. In this example, part of the heavy dust portion at ∼4–5 km is misclassified as cloud in V3 and remains misclassified as cloud in V4, but with much lower CAD scores. Very dense dust layers located above ∼4 km and north of ∼40∘ N will sometimes be misclassified as cloud because north of ∼40∘ N a significant amount of ice cloud exists at altitudes of 4 km and below. The V3 misclassification of high-altitude and high-latitude dense dust is not completely corrected in V4, as Fig. 11 illustrates, but the frequency of these dense dust cases is very low.

When a dust layer is misclassified as cloud, the high layer depolarization ratio can cause it to be further misclassified as an ice cloud by the cloud-phase algorithm (Hu et al., 2009; Avery et al., 2018), and hence the CALIPSO data products can sometimes report ice clouds that have temperatures warmer than 0 ∘C (Liu et al., 2009). These so-called “hot cirrus” (e.g., Liu et al., 2009) are highly likely to be misclassifications of dust. To more quantitatively evaluate the changes of the V4 CAD compared to the V3 CAD, Fig. 12 shows seasonal variations in (a) V3 aerosol fraction, (b) V4 aerosol fraction, (c) V3 clouds that changed to V4 aerosol, (d) V3 hot cirrus fraction, (e) V4 hot cirrus fraction, and (f) V3 hot cirrus clouds that changed to V4 aerosol for a selected geographic region (35 to 45∘ N, 75 to 90∘ E) that contains the entire Tarim Basin where dust is the dominant aerosol type (Wang et al., 2008). Hot cirrus occurs most frequently between 3 and 5 km, where dense dust can be lofted and clouds start to occur more frequently. The column mean of V3 hot cirrus fraction reaches a maximum of ∼8 % in August 2008 (Fig. 12d), and close to 80 % of the V3 hot cirrus changed to V4 aerosol (Fig. 12d and f). The hot cirrus fraction in V4 is significantly reduced, reaching a monthly maximum value of less than 3 % during the summer (Fig. 12e). This indicates that V4 has been improved significantly in this particular geographic region, especially at those altitudes where dense dust is most frequently misclassified as cloud in V3. There should also be a certain fraction of misclassified dust with temperatures colder than 0 ∘C that was classified as ice cloud by the cloud-phase algorithm in V3 and still remains misclassified as ice in V4. However, it is very difficult to quantify this type of misclassified dust.

Overall, there are more aerosols in V4 than in V3 because many of the layers misclassified as clouds in V3 are correctly classified as aerosols in V4. These changes occur mainly at relatively high altitudes, as seen in Fig. 12c, and reach a maximum of ∼20 % at altitudes of 4–5 km during the spring. For the column average from 0 to 7 km over this geographic region, the fractional change from cloud to aerosol varies from 5 % in the late fall to 16 % in the summer. The change from V3 ice contributes the most and accounts for 47.1 % of the total V3 cloud to V4 aerosol change, followed by the change from V3 “hot ice” (29.6 %) and V3 no-confidence cloud (19.2 %). Only a very small fraction of V3 water clouds changed their type, accounting for 4.1 % of the total change.

The V4 CAD algorithm also makes significant improvements in classifying lofted layers of Asian dust and polluted dust that are transported to the Arctic each spring. The Arctic regions have long been known to be impacted by aerosols from midlatitude sources, with the primary evidence being the springtime haze in the lower troposphere (Garrett and Verzella, 2008). Pollution and dust from Asian sources can reach the Arctic in 3 to 5 days, carried by midlatitude cyclones (Di Pierro et al., 2011, 2013; Z. Huang et al., 2015). In the earlier CALIOP data product releases, dust layers over the Arctic could be misclassified by the CAD algorithm as ice clouds (Di Pierro et al., 2011). An example measured at 18:28:54 UTC on 1 March 2008 is shown in Fig. 13. The attenuated backscatter data in Fig. 13a show numerous faint layers visible at altitudes of 5 to 10 km between 50 and 80∘ N. These layers show enhanced depolarization, albeit with depolarization ratios smaller than the typical values for ice clouds. Back trajectories analyzed using the HYSPLIT model from a representative point (74∘ N, 138∘ E) along the CALIPSO transect (Fig. 13d) indicate that most of these air parcels were originated and lifted up from the surface close to the Taklimakan Desert within the previous 5 days. These layers are thus likely dust transported from the lower latitudes, as demonstrated by Di Pierro et al. (2011, 2013). Many of these layers were misclassified as ice clouds in V3, as shown in Fig. 13b, and are now correctly classified in V4 as aerosols with high CAD scores, as shown in Fig. 13c.

In polar winter, ice crystals often form in clear skies when temperatures become very cold, due to slow isobaric cooling of moist air advected from lower latitudes, especially over the Antarctic plateau (Lachlan-Cope, 2010). Crystal concentrations tend to be low and were often misclassified as aerosol in V3, although in V3 aerosol in polar regions above ice and snow surfaces could only be classified as clean or polluted continental. In V4, the misclassification of tenuous ice crystals as aerosol still occurs, as pointed out by di Biagio et al. (2018). Note that in V4, the restrictions on aerosol type in polar regions have been removed and these tenuous ice crystals tend to be classified in V4 as mineral dust when misclassifications occur, even in regions where detectable intrusions of mineral dust from midlatitude source regions are rare.

Smoke plumes are often found in the upper troposphere and are occasionally injected above the tropopause by pyrocumulonimbus convection triggered by fires (Fromm et al., 2010; de Laat et al., 2012; Khaykin et al., 2018). A prominent example is the Black Saturday plume from the bushfires of Australia on 7 February 2009 (de Laat et al., 2012; Glatthor et al., 2013). Figure 14 shows smoke plumes at high altitudes ∼20 to ∼40∘ S on 10 February 2009. The low depolarization ratio (below 6 %, Fig. 14b) and increasing color ratio from top to base (not shown here) suggest that these layers are smoke, in contrast to the cloud layers with large depolarization and relatively uniform color ratio between ∼2 and ∼20∘ S. Whereas most of these layers were classified as clouds by the V3 CAD, the V4 CAD now correctly identifies these high-altitude smoke layers as aerosols. Note that the plume at lower altitudes between ∼20 and ∼28∘ S has high confidence (large CAD scores) in V4 and the plume at high altitudes between ∼30 and ∼40∘ S has low or no confidence (small CAD values). This is an indication that the probability for aerosols to be present at higher altitudes is smaller than at lower altitudes.

Table 3 shows scene classification confusion matrices for the year of 2008, calculated from individual range bins within the troposphere obtained from the 5 km profile products. These comparisons use only those range bins that are classified as either cloud or aerosol in both V3 and V4. Range bins reporting new features detected in V4 only and range bins that report features detected in V3 that were not detected in V4 are excluded. The first two rows represent the V3 clouds and aerosols and the first two columns indicate the V4 clouds and aerosols. The two diagonal elements represent the percentage of range bins that remain unchanged; the off-diagonal elements show the percentage of range bins for which the feature type changed. The third column of the first two rows is the percentage of the V3 cloud or aerosol relative to the total number of V3 features detected. Similarly, the first two columns of the third row give the percentages of the V4 cloud or aerosol relative to the total number of V4 features detected. The diagonal element of the third row and third column is the percentage of the total number of range bins that remains unchanged.

We see from this table that, for both daytime and nighttime, less than 5 % of V3 clouds are changed to aerosols in V4. While 4.6 % of V3 daytime aerosols are reclassified as clouds in V4, this change is more than 2 times larger (11.1 %) at night. Our preliminary analyses suggest that more cirrus fringes are detected at night than during the day, presumably because of the better nighttime SNR. Overall, the net cloud and aerosol fractions remained the same during the night, whereas the V4 net aerosol faction increased by ∼2 % during day.

Because the lower 532 nm calibration coefficients in V4 increase the magnitude of the attenuated backscatter coefficients, the V4 feature finder detection totals increased by 12.6 % and 6.2 %, respectively, at night and during the day (refer to Table A1 in Appendix A). At the same time, about 3.6 % of cloud and aerosol features reported in V3 during both night and day were not detected in V4, resulting in a net increase of 9.0 % in the V4 nighttime data and 2.5 % in the daytime data. The new features that are excluded in Table 3 are classified about equally as clouds and aerosols in V4.

To investigate possible spatial patterns in the CAD classification changes, we used the same 5 km profile products to locate where and when the classification changes occurred. Figure 15 presents geographic distributions of V4 cloud and aerosol fractions and the corresponding fractional change of V3 aerosol to V4 cloud or V3 cloud to V4 aerosol relative to the total V3 cloud and aerosol. The distribution patterns of the changes essentially follow the patterns of the cloud and aerosol distributions. More changes of V3 cloud to V4 aerosol are seen in the dust and smoke regions (Fig. 15b and f). As shown in Fig. 16a and b, the fractional change of V3 aerosol to V4 cloud decreases with increasing altitude because the aerosol occurrence is relatively low at higher altitudes. However, the fractional change relative to the V3 aerosol is very large at high altitudes (not shown). The significant changes from V3 clouds to V4 aerosols that are seen at 5–10 km appear to correspond to Asian dust activity over the sources and transport to the Arctic (during March–May; see Fig. 16c and d) or smoke plumes in the central and southern Africa (during August–October; see Fig. 16g and h).

Figure 17 presents joint altitude–latitude distributions of V3 and V4 aerosols and clouds in the left and middle columns, respectively. In Fig. 17b there appears to be a mode in the V4 daytime aerosol distribution in the tropical upper troposphere that shows a correlation to the tropical cloud distribution in Fig. 17e. This is mainly a residual of ice fringe candidates that were not changed to ice after applying the fringe amelioration algorithm. Shown in the right column in Fig. 17 are the changes of V3 aerosol to V4 cloud or V3 cloud to V4 aerosol relative to the total number of clouds and aerosols in each grid. Although most of the V3 aerosols at altitudes above ∼10 km have been changed to clouds in V4, more high-altitude V3 clouds were converted to aerosols in V4. In general, this behavior is expected, as the V4 CAD PDFs were deliberately designed to be more sensitive to the presence of lofted high-altitude aerosols. However, when misclassifications occur, the residuals are most often classified as dust by the aerosol subtyping algorithm (Omar et al., 2009; Kim et al., 2018).

There does not appear to be a clear high-altitude tropical aerosol mode in the nighttime V4 aerosol distribution in Fig. 17h, providing evidence that the mode seen in the daytime data is the result of classification errors. Our analysis shows that there are about 3 times more high-altitude layers detected at 5 km resolution and subsequently classified as dust during the daytime than at night. This can partly explain the day and night difference in the V4 aerosol distributions seen in Fig. 17 because layers detected at 5 km resolution are not processed by the fringe amelioration algorithm.

The CAD algorithm was originally designed for cloud and aerosol layers in the troposphere and in previous versions did not include data from the higher altitudes in its training set. As such, layers detected in the stratosphere were not characterized in versions prior to V4 and were instead simply called stratospheric features. The training set used to develop the V4 CAD PDFs included stratospheric data from all of 2008 and from two volcanic eruptions in June 2011 (see Sect. 3). These PDFs were subsequently applied to all layers detected in the stratosphere in the V4 operational algorithm, thus enabling the classification of volcanic and other stratospheric aerosol and cloud layers. Figure 18 shows two examples of V4 CAD performance when classifying stratospheric volcanic layers.

The extensive and clearly visible layer of enhanced backscatter between ∼50 and ∼80∘ N in Fig. 18a is a volcanic layer injected by the Kasatochi eruption (55∘ N) in August 2008 (Vernier et al., 2013). The persistence of this layer during the 3 months after the eruption underscores the need to characterize these layers properly. δv for this extended layer was relatively low, suggesting the predominance of sulfate particles (Fig. 18b). As can be seen in Fig. 18c, the V4 CAD algorithm correctly classified almost all of the volcanic layer as aerosol with high confidence (CAD scores approaching -100). Figure 18e shows another plume resulting from the eruption of Puyehue-Cordón Caulle in Chile in June 2011 at an altitude of 10–12 km. In contrast to the Kasatochi volcano, the silicate ash content in this plume was very high (Vernier et al., 2013), as can be seen in the relatively high δv in Fig. 18f. Once again, the V4 CAD algorithm classifies most of the layers as aerosols. However, the CAD scores are not very high (Fig. 18g). This largely reflects the fact that the probability for a relatively dense depolarizing aerosol to be present at high altitudes is low, and the cloud and aerosol PDF overlap region here is large compared with low altitudes. A significant fraction of the Puyehue-Cordón plume is classified as cloud, possibly due to high color ratios and depolarization ratios that fall within the PDF overlap region with ice clouds. Because volcanic eruptions can release large amounts of water vapor, which can then condense into ice cloud particles (Guo et al., 2004), it is not always possible to determine with absolute certainty whether the putative cloud layers in the Cordón plume are misclassified aerosols or actually legitimate clouds.

While the V4 CAD can distinguish aerosols and clouds for stratospheric layers, uncertainties tend to increase as the altitude increases. This increasing uncertainty derives from the fact that the very low aerosol occurrence frequency at high altitudes does not provide a statistically significant sample size to constrain the PDFs, and thus the high-altitude PDFs were created by extrapolation from measurements at lower altitudes. Further, because the SNR of stratospheric layers is typically quite low, there is a widening in the distribution of color ratio and attenuated backscatter for stratospheric features compared to features at lower altitudes, leading to generally lower stratospheric CAD scores. This can be seen in Fig. 19, which shows the CAD scores of both aerosol and cloud layers with bases within 2 km above the tropopause and above 4 km above the tropopause for July–October 2008 between ∼50 and ∼82∘ N. These layers are mostly from the Kasatochi volcano. Within 2 km above the tropopause, the CAD algorithm classifies both aerosols and clouds with good confidence. However, as we go higher up in the stratosphere, the general lack of data, as well as decreasing SNR for weaker features, makes CAD increasingly difficult. Furthermore, the fraction of feature-finder false positives may become significant, especially for daytime measurements over bright surfaces or optically thick stratus cloud decks. These false positives generally have very small CAD scores, which quite rightly reflect a lack of classification confidence. As a result, at very high altitudes, most of the layers classified as clouds exhibit very low or no confidence (CAD score <20) (similar to aerosols, as seen in Fig. 22) and the CAD algorithm generally seems to provide somewhat more confidence in the aerosol classification than the clouds. This is consistent with the general dearth of cloud occurrence at stratospheric altitudes.

PSCs are ubiquitous in both polar regions in local winter and have important consequences for polar ozone loss processes (e.g., Lowe and MacKenzie, 2008). Along with other stratospheric layers, the V4 CAD algorithm is now applied to PSCs as well. The CALIPSO project produces the Level 2 Polar Stratospheric Cloud product that uses the spatial and optical properties of these clouds to classify them according to type (Pitts et al., 2009). By definition a PSC is a cloud. In the CALIPSO PSC product, PSCs are classified by composition as a supercooled ternary solution (STS) of HNO3, H2SO4, and H2O, Mix 1, Mix 2, or ice particles. Mix 1 and Mix 2 are PSC classes denoting lower and higher nitric acid trihydrate (NAT) number density and volume, respectively. Of these, STS may be thought of as closest to being a liquid aerosol particle. In this section, we assess the CAD classification of PSC layers by comparing V4 results with the classifications from the CALIPSO PSC-specific data product. Figure 20 shows a comparison of extensive PSC layers observed over Antarctica on 15 August 2008. Figure 20a shows results from the PSC product, with specific colors assigned to the different PSC classifications. Figure 20b shows the corresponding V4 CAD browse image for the same scene. Generally, there is a good correlation between the spatial distributions of STS in the left panels (red) and layers classified by V4 CAD as aerosols (red).

Figure 21 compares the spatial occurrence of stratospheric aerosol identified by the V4 CAD (left panels) and STS from the PSC product (right panels) for the months of June and January 2008 during the Antarctic and Arctic PSC seasons, respectively. There is a good correspondence between the locations of the peak concentration in latitude and altitude in both hemispheres. Despite the differences in spatial occurrence and the general uncertainty in applying cloud–aerosol terminology to PSCs, this level of correspondence is quite encouraging for the CAD performance.

Figure 22 shows the CAD scores assigned to polar stratospheric aerosols for January and June 2008, corresponding to the cases in Fig. 21. Above about 15 km the CAD scores are all very low (<20) for these aerosol layers, similar to the clouds in Fig. 19. This is not unexpected since these particles are in the process of becoming PSCs. However, noise in the 1064 nm data may also contribute to the classification uncertainties.