CALIPSO Quality Statements
Lidar Level 2 Cloud and Aerosol Layer Products
Version Releases: 3.01, 3.02

The CALIOP surface detection routine uses a digital elevation map (DEM), GTOPO30 as the starting point in its search for the lidar surface echo, and thus the reliability of the lidar surface elevations depends to some extent on the accuracy of the information recorded in GTOPO30. The GTOPO30 data is very reliable over oceans, but can be considerably less so in rugged terrain, such as in the Andes mountains of Peru, and over the polar regions. Note too that due to aberrations in the signal caused by a non-ideal transient response in the 532 nm detectors, the geometric thickness associated with the lidar surface elevation (i.e., surface top - surface base) can be extremely misleading. This non-ideal transient response must be carefully considered whenever the (apparent) subsurface portions of the lidar signals analyzed

Bits 1, 2, and 3 indicate the horizontal resolution at which the surface was detected:

• 0 = not detected
• 1 = detected at 1/3-km averaging
• 2 = detected at 1-km averaging
• 3 = detected at 5-km averaging
• 4 = detected at 20-km averaging
• 5 = detected at 80-km averaging

Bits 4 and 5 are not used and are set to zero. Taken together, bits 6, 7, and 8 report the 5-km detection frequency:

• 0 = detection frequency = 0%
• 1 = detection frequency = 20%
• 2 = detection frequency = 40%
• 3 = detection frequency = 60%
• 4 = detection frequency = 80%
• 5 = detection frequency = 100%

The CALIOP layer detection algorithm used for the Version 1 and Version 2 data releases is described in detail in Vaughan et al., 2009 and in the CALIPSO Feature Detection ATBD (PDF). For Version 3, an additional refinement has been incorporated into the base determination procedure. Under certain conditions, described here, the initial estimate of base altitude for those layers identified as aerosols will be extended to a new, lower altitude located 90 m above the local surface. The Layer Base Extended flag identifies those layers for which the base altitude has been altered by this procedure.

The uncertainties associated with detection of cloud and aerosol layers in backscatter lidar data are examined in detail in Section 5 of the CALIPSO Feature Detection ATBD (PDF). The ATBD contains quantitative assessments of feature finder performance derived using simulated data sets, for which all layer boundaries were known exactly. In the real world of layer detection, we do not have access to this underlying truth. Therefore in this document we provide the following set of "rules of thumb" that users can apply to the data products to obtain a qualitative understanding of the layer boundaries reported, and of the optical properties associated with these layers.

1. Strongly scattering features are easier to detect than weakly scattering features. The scattering intensity of each layer is reported in the 532 nm and 1064 nm attenuated backscatter statistics and by the integrated attenuated backscatter at 532 nm and 1064 nm.

2. Detection of layers during the nighttime portion of the orbits is more reliable than during the daytime portion of the orbits. Due to solar background signals, the noise levels in the daytime measurements are much larger than those at night, and this additional noise can obscure faint features, and can lead to boundary detection errors even in more strongly scattering layers.

3. Features become increasingly difficult to detect with increasing optical depth above feature top. Put another way, detection of the lower layers in a multi-layer scene is made more difficult by the signal losses that occur as the laser light passes through the upper layers. (In a sense, this is a restatement of (a), since the backscatter intensity of secondary features is reduced from what it otherwise might be by the signal attenuation caused by the overlying features.) The Overlying Integrated Attenuated Backscatter and the Layer Integrated Attenuated Backscatter QA factor serve as proxies for the optical depth above each feature, and thus provide qualitative assessments of the confidence that users should assign to the reported layer properties.

4. In general, our confidence in the location of the top of a layer is somewhat greater than our confidence in the location of the base of the same layer. For transmissive features, one reason for this is that the backscatter signal is attenuated by traversing the feature, thus degrading the potential contrast between feature and "non-feature" at the base. Additionally, in strongly scattering layers, multiple scattering effects and signal perturbations introduced by the non-ideal transient response of the 532 nm detectors can also make base determination less certain.

5. The Opacity Flag is used to indicate features that completely attenuate the backscatter signal. For these features, the base altitude reported must be considered as an "apparent" base rather than a true base.

6. In those cases where the layer base has been extended to 90 m above the local surface, the assumption is that extended region contains aerosol that lies below the detection limits of the standard algorithm. The resulting increase in aerosol optical depths indicates that this procedure is appropriate far more often than not.

7. Stratospheric features reported during daylight -- especially those reported above 20 km between 60N and 60S -- are often noise artifacts and should be treated with suspicion.

Interpretation of the number of layers found parameter is straightforward for the 1-km and 1/3-km layer products: individual layers are always separated by regions of "clear air", and layer boundaries never overlap in the vertical dimension. However, this simplicity of interpretation does not always carry over into the 5-km cloud and aerosol layer products. CALIPSO uses a nested multi-grid feature finding algorithm (see the layer detection ATBD), and thus the search for layer boundaries is conducted at multiple horizontal averaging resolutions. While the 1-km and 1/3-km layer products report only those features detected at, respectively, averaging resolutions of 1-km and 1/3-km, the 5-km products report layers detected at multiple averaging resolutions (5-km, 20-km, and 80-km in the version 3 products). Because the reporting resolution (5-km) is not always identical to the detection resolution, layers may appear to overlap in the vertical dimension.

Figure 2 shows a wholly fictitious but heuristically useful schematic of layer detection results for a data segment extending 80-km horizontally and 465-m vertically. Yellow/orange/brown colors indicate an aerosol layer detected at horizontal averaging resolutions of, respectively, 80, 20 or 5 km. Shades of blue likewise represent clouds detected at 80, 20, and 5 km resolutions. The white regions are (presumably) clear air, where no features were found. The labeled rows at the bottom indicate the 'number of layers found' that will be reported in the cloud and aerosol layer products for each 5-km column.

Figure 2: interpreting the number of layers parameters

In column 16, the layer labeled F5 (top altitude = 0.285 km, base altitude = 0.165 km) appears to vertically overlap F6 (top altitude = 0.255 km, base altitude = 0.135 km), which in turn appears overlap F7 (top altitude = 0.225 km). However, F5 was detected at an averaging resolution of 5-km, and hence the backscatter data that comprises F5 is removed from consideration before construction the 20-km horizontally averaged profile in which F6 was detected. Similarly, the backscatter data from both F5 and F6 were removed from consideration before constructing the 80-km averaged profile in which F7 was detected. Layers detected at higher spatial resolutions are thus seen to overwrite, rather than overlap apparently collocated layers detected at coarser spatial resolutions. More details can be found in the Feature Detection and Layer Properties ATBD (PDF) and in Vaughan et al., 2009.

For the uppermost layer in any column, the quality of the estimate for γ′₅₃₂ is determined by the accuracy of the top and base identification, the reliability of the 532 nm channel calibrations, and by the signal-to-noise ratio (SNR) of the backscatter data within the layer. For layers beneath the uppermost, the quality of our estimate for γ′₅₃₂ also depends on either obtaining an independent estimate of the two-way transmittance, T², for all overlying layers, or by estimating this quantity directly from the lidar backscatter data. In those situations where an extended region of clear air exists between successive layers, and where the uppermost layer has no more than a moderate optical depth of -- say -- between 0.4 and 2.0, T² can be estimated directly from the attenuated backscatter data (albeit with some uncertainty due to noise and the possibility of aerosol contamination of the clear air regions). Otherwise, the only way to estimate T² is to compute a full extinction retrieval for the profile being examined. In this case, additional error can be introduced into the estimate of γ′₅₃₂ by uncertainties in the approximation of the lidar ratio(s) for the overlying layer(s). Furthermore, the effects of errors caused by misestimating T² can increase sharply as the optical thickness above a layer increases. For the 5 km layer products, the CALIOP processing scheme always attempts to correct estimates of γ′₅₃₂ for the attenuation imparted by previously identified overlying features. As a consequence, we will occasionally report unrealistically large values for γ′₅₃₂ in the 5 km layer products. However, because extinction solutions are only derived for data averaged to a 5-km (or greater) resolution, the γ′₅₃₂ values reported in the 1 km and 1/3 km layer products are not corrected for the signal attenuation effects imparted by overlying layers.

The values reported for γ′₅₃₂ should always be positive, and for the results derived directly from the layer detection algorithm (i.e., in the 1 km and 1/3 km layer products) this is indeed always true. However, in the 5 km products there are certain rare and pathological cases where a secondary layer could only be detected after averaging to 20 km or even 80 km horizontally, and where the overlying layers were detected at 5 km and have vastly different optical depths. In these cases, integrating the reaveraged data within the secondary layer will occasionally yield a negative γ′₅₃₂. Such layers can be identified by a special CAD score of 105. All measured and derived optical properties for these layers are unreliable, and should be ignored. In evaluating the reliability of the spatial properties of these layers, users should carefully consider the layer IAB QA factor.

As is the case for γ′₅₃₂, in the uppermost layer within any column, the quality of the estimate for γ′₁₀₆₄ is determined by the accuracy of the top and base identification, the reliability of the 1064 nm calibration constant, and by the signal-to-noise ratio (SNR) of the backscatter data within the layer. However, unlike the measurements at 532 nm, reliable estimates of T² cannot be derived from an analysis of the 1064 nm backscatter signal in the (assumed to be) clear air regions, and thus in the 5 km products, the T² corrections for the attenuation from overlying layers are always obtained from an extinction solution that uses prescribed values of the lidar ratios for all overlying layers. As is the case at 532 nm, no T² corrections are applied to the γ′₁₀₆₄ values reported in the 1 km and 1/3 km layer products. Furthermore, because the CALIOP layer detection algorithm typically examines only the 532 nm backscatter signals, negative (i.e., non-physical) values may occasionally be reported for γ′₁₀₆₄ in all resolutions of the layer products. Unlike the layers for which γ′₅₃₂ is negative, layers with negative γ′₁₀₆₄ are not indicated by a special CAD score. Negative values of γ′₁₀₆₄ occur most often for very weakly scattering layers (e.g., subvisible cirrus and faint aerosols) and in those layers for which the backscatter signal has been highly attenuated by other, overlying layers.

The quality of the estimate for δ_v is determined by the accuracy of the top and base identification, the reliability of the polarization gain ratio calibration, and by the signal-to-noise ratio (SNR) of the backscatter data within the layer. In general, the CALIOP δ_v estimates are highly reliable. Histograms of δ_v compiled for midlatitude cirrus in the northern hemisphere compare very well with previously reported distributions, e.g., Sassen & Benson, 2001 (PDF).

In regions with acceptable SNR, the accuracy with which the range resolved depolarization ratios can be determined will depend almost entirely on the accuracy of the polarization gain ratio calibration.

The quality of the estimate for χ′_layer is determined by the accuracy of the top and base identification, the reliability of the 532 nm calibration constant and the 1064 nm calibration constant, and by the signal-to-noise ratio (SNR) of the backscatter data within the layer. For the 5 km layer products, the attenuated backscatter coefficients used in the calculation of χ′_layer are corrected for the estimated overlying two-way transmittance. No such correction is attempted for the 1 km and 1/3 km values, as no extinction solution is computed at these resolutions.

Users can have high confidence in the calculation of all of the values in the attenuated total color ratio statistics fields. However, as with the 532 nm depolarization ratio statistics, the meaning of the various numbers can be somewhat misleading. Like the depolarization ratios, the attenuated total color ratios are produced by dividing one noisy number (the 1064 nm attenuated backscatter coefficient) by a second noisy number (the 532 nm attenuated backscatter coefficient). Depending on the noise in any pair of samples, the resulting values can range from large negative values to extremely large positive values. When computing layer means, standard deviations, and centroids, these outliers can dominate the calculation, and thus return entirely unrealistic estimates.

Because all features reported in 1/3-km and 1-km layer products are detected at a single horizontal averaging resolution (i.e., either at 1/3-km or 1-km), the opacity flag is not reported. When using these products, opacity, in the sense described above, can be assessed as follows. If the surface was detected (i.e., the lidar surface altitude field does not contain fill values) then there are no opaque layers in the column. If the surface was not detected, then the lowest layer in the column is considered to be opaque.

• For the vast majority of cases, CALIOP cannot provide a direct measurement of layer optical depth. In these cases, estimates of optical depth are derived using extinction-to-backscatter ratios (i.e., lidar ratios) that are specified based on an assessment of layer type and subtype. Uncertainties in the value of the lidar ratio, which can arise both from natural variability and from occasional misclassification of layer type, propagate non-linearly into subsequent estimates of layer optical depth.

• Retrievals of optical depth from space-based lidar measurements must account for contributions from multiple scattering that are generally considered negligible in ground-based and aircraft based measurements. The theoretical basis for CALIPSO's treatment of multiple scattering is provided in the extinction retrieval ATBD (PDF) and in Winker, 2003 (PDF)

• Similar to the layer detection problem, estimates of layer optical depth become increasingly fraught with error in multiple layer scenes, as errors incurred in overlying layers are propagated into the solutions derived for underlying features.

• IMPORTANT NOTICE: before proceeding, all users of the CALIOP optical depth data should read and thoroughly understand the information provided in the Profile Products Data Quality Summary. This summary contains an expanded description of the extinction retrieval process from which the layer optical depths are derived, and provides essential guidance in the appropriate use of all CALIOP extinction-related data products.

Despite these caveats, users should not be unduly pessimistic about the quality and usability of the CALIPSO optical depth estimates. Figure 3 (below) shows a preliminary comparison of CALIPSO version 3 aerosol optical depths with the optical depths derived from MODIS for all daytime measurements acquired during January 2007. The comparison is generally good, with MODIS appearing to slightly over-estimate values at the lower end of the optical depth range.

Figure 3: Comparison of CALIPSO aerosol optical depths to those derived from MODIS (Preliminary - January 2007, daytime data only) final lidar ratio = initial lidar ratio only)

Calculation of the layer optical depth uncertainty is an iterative process. On some occasions when the SNR is poor, or an inappropriate lidar ratio is being used, the iteration will attempt to converge asymptotically to positive infinity. Whenever this situation is detected, the iteration is terminated, and the layer optical depth uncertainty is assigned a fixed value of 99.99. Any time an uncertainty of 99.99 is reported, the extinction calculation should be considered to have failed. The associated optical depths cannot be considered reliable, and should therefore be excluded from all science studies.

Note: optical depth uncertainties are reported as absolute errors, not relative errors.

The aerosol lidar ratios used for the CALIOP analyses represent well established mean values that are characteristic of the natural variability exhibited for each aerosol species (e.g., see Omar et al., 2005 (PDF); Cattrall et. al., 2005; and Figure 4 below). The clear implication of this natural variability is that even for those cases where the aerosol type is correctly identified, the initial lidar ratio represents an imperfect estimate of the layer-effective lidar ratio of any specific aerosol layer. These same caveats apply equally to the mean values used for the initial cloud lidar ratios. For all layer types, cloud-aerosol discrimination errors can exacerbate the error associated with the specification of the initial lidar ratio. Uncertainty can also be introduced by the cloud ice-water phase classification and the aerosol subtype identification procedures. However, the CALIOP extinction algorithm incorporates some error-correcting mechanisms that in many cases will adjust the initial estimate of lidar ratio so that a more suitable value is ultimately used in the retrieval. Details of the lidar ratio adjustment scheme are provided in the extinction retrieval ATBD (PDF). Algorithm architectural information and generalized error analyses for CALIPSO's cloud-aerosol discrimination algorithms, cloud ice-water phase algorithms, and aerosol subtyping algorithms can be found in the CALIPSO Scene Classification ATBD (PDF).

Figure 4: Distributions for AERONET-derived lidar ratios, computed for aerosol types described in Omar et al., 2005 (PDF);

The version 3.0 aerosol subtyping scheme is different from the previous versions. Two of the aerosol models, dust and polluted dust, discussed in Omar et al. (2009) and used in prior versions have been updated in light of the latest advances in the science. Recent measurements of size distributions of dust aerosol during NAMMA and T-Matrix calculations of the phase functions allowed a more realistic estimate of the dust lidar ratio at 1064 nm which, thus far, was based on a single measurement during SAFARI 2000. The dust phase function (squares in Figure 4a) is determined by T-Matrix calculations using NAMMA measurements of size distributions and refractive indices and the smoke phase functions (circles in Figure 4a) are determined from Mie calculations using size distributions and refractive indices of the biomass burning cluster of the AERONET measurements. The dust lidar ratios determined partly using NAMMA measurements are 40 sr and 55 sr at 532 nm and 1064 nm, respectively. The 55 sr at 1064 nm is a significant departure from the 30 sr used to calculate 1064 nm extinction coefficients in previous versions.

Since the polluted dust model is built from a smoke fine component and a dust coarse component, the above adjustment in the dust model is also reflected in the polluted dust model. A composite phase function (Figure 5b) of dust coarse mode and smoke fine mode from the individual phase functions is shown in Figure 5a. The resulting polluted dust lidar ratios are 55 sr at 532 nm and 48 sr at 1064 nm. The former is a departure from the old values of 65 sr at 532 nm. This is because in the original model most of the polluted dust aerosol was comprised of smoke, while this new model, partly based on NAMMA observations, apportions a significant surface area to the coarse mode dominated by dust. The old model used Mie calculations to generate polluted dust phase functions and lidar ratios while the new model uses Mie model calculations for the fine mode (smoke) and T-Matrix calculations for the coarse mode (dust) to generate the phase functions and lidar ratios.

Figure 5: (a) Smoke and Dust phase functions determined by Mie and T-Matrix calculations respectively, and (b) composite dust and smoke phase functions representative of the polluted dust aerosol model

For weakly scattering features, the lidar ratio is most often left unchanged by the extinction solver, as a physical solution is usually obtained on the first iteration. In these cases, the uncertainties in the final lidar ratio are the same as the uncertainties in the initial lidar ratio. The exception to this statement would be if either the cloud-aerosol discrimination or the layer sub-typing procedures have misclassified the layer. However, for weak layers, the relative error in the lidar ratio is (approximately) linearly related to the resulting error in the derived optical depth estimate.

Retrievals of opaque and strongly scattering layers are very sensitive to the initial lidar ratio selection. Too large a value will cause the retrieval algorithm to, in effect, extinguish all available signal before reaching the measured base of the feature. When that point is reached the retrieval becomes numerically unstable and the calculated extinction coefficients will asymptote toward positive infinity. In these cases, a successful solution can only be obtained by reducing the lidar ratio. The CALIOP extinction routine does this automatically, and will repeat the process until a stable solution is achieved. When this happens, the final lidar ratio reported in the data products is the first one for which a physically meaningful (albeit not necessarily correct) solution was obtained for the entire measured depth of the layer.

The optical depths and extinction profiles derived in those cases where the layer lidar ratio must be reduced are generally not accurate. The current lidar ratio reduction scheme terminates after identifying (a very close estimate of) the largest lidar ratio for which a physically meaningful solution can be generated for the backscatter measured in the layer. However, the optical depths and extinction profiles reported in these situations can only be considered as upper bounds; the true values are somewhat, or perhaps even significantly, lower. Because the associated optical depth uncertainties cannot be reasonably estimated, these data should be excluded from statistical analyses of layer optical properties, and even the most sophisticated users are advised to treat these cases with extreme caution.

When an independent estimate of layer optical depth is available from a measured layer two-way transmittance, the CALIOP extinction algorithm will retrieve the optimal estimate of the layer-effective lidar ratio, irrespective of layer type, and use this retrieved lidar ratio in the extinction retrieval. These so-called 'constrained' retrievals are more accurate than unconstrained retrievals. For constrained retrievals, the uncertainty in the final lidar ratio can be well estimated using equation 7.4 from the CALIPSO Scene Classification ATBD (PDF). An extinction QC value of 1 indicates a successful constrained retrieval.

Ice clouds: for the CALIOP viewing geometry, simulations show multiple scattering effects are nearly independent of range and, as parameterized in the CALIOP retrieval algorithm (Winker et al. 2009), are nearly independent of extinction. In Version 2, ice clouds were assigned a range-independent multiple scattering factor of η₅₃₂ = 0.6. Validation comparisons indicate this is an appropriate value and the same value is used in Version 3.

Water clouds: in Version 2 the multiple scattering factor for water clouds was set to unity, resulting in large errors in retrievals of extinction and optical depth. In Version 3, a value of η₅₃₂ = 0.6 is used. Based on Monte Carlo simulations of multiple scattering, this value appears to be appropriate for semitransparent water clouds (τ < 1). (It is purely coincidental this is the same value used for ice clouds.) For denser water clouds (τ > 1) the multiply-scattered component of the signal becomes much larger than the single-scattered component, η₅₃₂ becomes dependent on both cloud extinction and range into the cloud, and the retrieval becomes very sensitive to errors in the multiple scattering factor used. In these cases the multiple scattering cannot be properly accounted for in the current retrieval algorithm and retrieval results are unreliable.

Aerosols: simulations of multiple scattering effects on retrievals of aerosol layer optical depth indicate the effects are small in most cases. There is uncertainty in these estimates, however, due to poor knowledge of aerosol scattering phase functions. Validation comparisons conducted to date do not indicate significant multiple scattering effects on aerosol extinction profile retrievals. Multiple scattering effects may become significant in dense aerosol layers (σ > 1 /km), but in these cases retrieval errors are usually dominated by uncertainties in the lidar ratio or failure to fully penetrate the layer. In Version 3, as in Version 2, multiple scattering factors for both wavelengths are set to unity.

While the volume depolarization ratio is a direct measurement, the layer integrated 532 nm particulate depolarization ratio, δ_p, is a post-extinction quantity, calculated from ratio of the layer integrated perpendicular and parallel polarization components of particulate backscatter coefficient within the layer, using

Here β_⊥,P and β_||,P are the perpendicular and parallel components of particulate backscatter coefficient at 532 nm, respectively.

The quality of the estimate for δ_p is determined not only by the SNR of the backscatter measurements in parallel and perpendicular channels, but also the accuracy of the range-resolved two-way transmittance estimates within the layer. The two-way transmittances due to molecules and ozone can be well characterized via the model data obtained from the GMAO. The two-way transmittances due particulates, however, are only as accurate as the CALIOP extinction retrieval. Opaque cirrus cloud layers can be particularly prone to errors in the particulate depolarization ratio, as very large attenuation corrections are applied to the weak signals at the base of the layers, and on those occasions where one channel or the other becomes totally attenuated, this situation can generate very large, negative particulate depolarization ratio estimates. For layers that are not opaque, δ_p is generally reliable. However, in weakly scattering layers, the quality of the daytime estimate can be degraded by a factor of 2-4 due to the larger background noise compared with the nighttime estimate.

Based on a one month test data set (January 2007), the median particulate depolarization ratio uncertainties in the aerosol layer products is typically ~0.04 and ~0.16 for nighttime and daytime measurements, respectively.

Note: both in the layer products and the profile products, particulate depolarization ratio uncertainties are reported as absolute errors, not relative errors.

Much like the integrated attenuated total color ratio, the quality of χ_p is governed by the accuracy to which layer top and base altitudes are determined and by the signal-to-noise ratios of the backscatter data within the layer. Additionally, since all of the β_P,λ(z) values are derived by the Hybrid Extinction Retrieval Algorithm (HERA), the quality of χ_p also depends on the success of the HERA profile solver in deriving accurate solutions for β_P,λ(z). As such, the quality of χ_p can be partially assessed via the extinction QC flags which report the final state of the HERA solution attempt. In general, solutions where the final lidar ratio is unchanged (extinction QC = 0) or the extinction solution is constrained (extinction QC = 1) yield physically plausible solutions more often. Conversely, solutions tend to be more uncertain in those cases where the lidar ratio for either wavelength must be reduced.

Note: Uncertainties for layer-integrated particulate backscatter color ratios are reported as absolute errors, not relative errors.

Users can have high confidence in the calculation of all the values in the attenuated particulate color ratio statistics fields. However, particulate color ratios are produced by dividing one noisy number (the 1064 nm mean particulate backscatter coefficient) by a second noisy number (the 532 nm mean particulate backscatter coefficient), resulting in values that can range from large negative values to extremely large positive values, depending on the noise in any pair of samples. When computing layer means, standard deviations, and centroids, these outliers can dominate the calculation, and thus return entirely unrealistic estimates. Therefore, the integrated particulate ratio characterizes the particulate color ratio of a layer more reliably than does the mean value of the individual particulate color ratios within a layer.

In the current release (version 3), the CAD algorithm uses newly developed five-dimensional (5D) probability density functions (PDFs), rather than the three-dimensional (3D) PDFs used in previous versions. In addition to the parameters used in the earlier 3D version of the algorithm (layer mean attenuated backscatter at 532 nm, layer-integrated attenuated backscatter color ratio, and altitude), the new 5D PDFs also include feature latitude and the layer-integrated volume depolarization ratio. Detailed descriptions of the CAD algorithm can be found in Sections 4 and 5 of the CALIPSO Scene Classification ATBD (PDF). Enhancements made to incorporate the 5D PDFs used in version 3 release are described in Liu et al., 2010 (PDF). For further information on the CAD algorithm architecture and the three-dimensional (3D) PDFs used in versions 1 and 2 of the data products, please see Liu et al., 2004 (PDF) and/or Liu et al., 2009.

The standard CAD scores reported in the CALIPSO layer products range between -100 and 100. The sign of the CAD score indicates the feature type: positive values signify clouds, whereas negative values signify aerosols. The absolute value of the CAD score provides a confidence level for the classification. The larger the magnitude of the CAD score, the higher our confidence that the classification is correct. An absolute value of 100 therefore indicates complete confidence. Absolute values less that 100 indicate some ambiguity in the classification; that is, the scattering properties of the feature are represented to some degree in both the cloud PDF and in the aerosol PDF. In this case, a definitive classification cannot be made; that is, although we can provide a "best guess" classification, this guess could be wrong, with a probability of error related to the absolute value of the CAD score. A value of 0 indicates that a feature has an equal likelihood of being a cloud and an aerosol. Users are encouraged to refer to the CAD score when the cloud and aerosol classification results are used and interpreted.

Beginning with the version 2.01 release, several "special" CAD score values have been added. These are listed in the table below. Each of these new values represents a classification result that is based on additional information beyond that normally considered in the standard CAD algorithm.

The bit assignments are additive, so that (for example) an extinction QC value of 18 represents an unconstrained retrieval (bit 1 is NOT set) for which the lidar ratio was reduced to prevent divergence (+2; bit 2 is set), and for which the feature finder has indicated that the layer is opaque (+16; bit 5 is set). For the version 2.01 release, bits 10-15 are not used. Complete information about the conditions under which each extinction QC bit is toggled can be found in the CALIPSO Extinction Retrieval ATBD (PDF)

Correct interpretation of the feature subtype bits depends on the status of the feature type; e.g., the interpretation is different for clouds and aerosols. For aerosols, the feature subtype is one of eight types: desert dust, biomass burning, background, polluted continental, marine, polluted dust, other, and 'not determined'. Desert dust is mostly mineral soil. Biomass burning is an aged smoke aerosol consisting primarily of soot and organic carbon (OC), clean continental (also referred to as background or rural aerosol) is a lightly loaded aerosol consisting of sulfates (SO₄^2-), nitrates (NO₃^-), OC, and Ammonium (NH₄⁺), polluted continental is background aerosol with a substantial fraction of urban pollution, marine is a hygroscopic aerosol that consists primarily of sea-salt (NaCl), and polluted dust is a mixture of desert dust and smoke or urban pollution. Extensive test data generated prior to the version 3 release revealed a negligibly minute number of layers with spurious lidar ratios and aerosol type designations. These trace layers are currently labeled 'not determined'. The 'other' designation is a place-holder for another, yet to be determined, aerosol type. While this set does not cover all possible aerosol mixing scenarios, it accounts for a majority of mesoscale aerosol layers. In essence the algorithm trades off complex transient multi-component mixtures for relatively stable layers with large horizontal extent (10-1000 km).

In Version 1 and 2 data, the base altitudes of optically thick aerosol layers were sometimes biased high due to lidar signal attenuation or signal perturbations, causing aerosol optical depth (AOD) underestimates. In Version 3, to compensate for this, layer base altitudes of aerosol layers meeting the following criteria (hereafter termed, 'extended layers') are lowered to three range bins (90 m) from the surface as reported by the lidar surface elevation. The criteria for aerosol layer base extension are:

1. The layer must be the lowest layer in the 5 km resolution column
2. The surface must be detected below the layer
3. The difference between the original base altitude and Lidar Surface Elevation must be less than the science-tunable parameter, 'maximum gap distance', currently = 2.5 km
4. The mean attenuated backscatter at 532 nm beneath the original base altitude must be positive

These criteria attempt to ensure that only boundary-layer aerosol layers with useable signals beneath their original bases are extended. The possibility of introducing surface contamination in layer optical properties is reduced by assigning the extended base to an altitude 90 m above the local surface height. Hence, surface contamination in extended layers is minimal and confined to regions with rugged terrain. However, this also means that profiles within layers with extended bases always stop 90 meters above the surface. In four months of global test data, the base altitudes of 8.6% of all layers originally classified as aerosol were extended, with average base altitudes lowered by 0.54 km.

Layer descriptors are re-computed after base extension and each extended layer is re-analyzed by the scene classification algorithms to assign feature type, subtype, and CAD Score. Consequently, the feature type or subtype of these layers may change since their optical properties have changed. Hence, layer base extended descriptors are populated with the feature classification flags of these layers prior to base extension so their previous type and subtype can be discerned. In four months of test data, 17% of extended aerosol layer subtypes changed due to base extension. Additionally, 12% of all extended aerosol layers were reclassified as cloud layers, accounting for 0.3% of all cloud layers. This typically occurs in scenes with low SNR, or when surface contamination is suspected. Extended aerosol layers which are reclassified as cloud layers tend to have very low CAD scores before and after base extension (65% have |CAD score| less than 10), so their type was never certain to begin with. These layers cannot be used with confidence. Conversely, 85% of extended layers have |CAD score| > 90 when the type does not change (extended aerosol layer remains an aerosol layer). Users are advised to consult CAD scores to assess confidence in all feature types.

Impact on Version 3 Aerosol Optical Depths

A focused study of two test days (2007-01-01 and 2007-08-27) found that the median optical depth of extended aerosol layers increased by 22%, resulting in a 1% AOD increase in all aerosol layers globally. The figure below shows the change in optical depth for extended aerosol layers (unconstrained retrievals) where the type or subtype did not change for the two test days. By design, optical depth values have increased due to the aerosol base extension algorithm.

Figure 6: Histogram of aerosol optical depth at 532 nm with and without base extension for all extended aerosol layers in the two day test.

γ′_above provides a qualitative assessment of the confidence that users should assign to each layer reported. As noted earlier (see the discussion for layer base and top heights), layer detection, and the assessment of the associated layer descriptors, becomes increasingly uncertain as the overlying optical depth increases. This uncertainty cannot be easily quantified, because backscatter lidars such as CALIOP cannot measure optical depth directly, and must instead derive optical depth estimates in subsequent data processing. However, γ′_above can easily be obtained directly from the calibrated backscatter signal, and hence can provide a qualitative proxy for the optical depth above each layer detected.