3.1 Algorithm

The ATSR-2/AATSR cloud products were produced using the Community Cloud retrieval for Climate (CC4CL) algorithm, developed during the ESA CCI programme. The algorithm has been described in detail in Stengel et al. (2017); Poulsen et al. (2012); McGarragh et al. (2017); Sus et al. (2018). For completeness, the basic concepts are summarised here. The CC4CL algorithm consists of three main components: (1) cloud detection, (2) cloud typing and (3) the retrieval of cloud properties based on an optimal estimation (OE) technique. The cloud mask and cloud phase are both determined using artificial neural network (ANN) algorithms. Each ANN is trained using CALIPSO data co-located with AVHRR and then transferred to the ATSR series of instruments through the application of spectral band adjustments described in Sus et al. (2018).

The optimal estimation retrieval within CC4CL, known as the Optimal Retrieval of Aerosol and Cloud (ORAC), is a non-linear statistical inversion method based on Bayes’ theorem (Rodgers, 2000). A state vector containing all variables to be retrieved is optimised to obtain the best fit between observed top-of-atmosphere (TOA) radiances and those simulated by a forward model. The inversion can accommodate a priori information and its associated uncertainty (though, in this application, only surface temperature is constrained, based on ERA-Interim reanalyses). The method provides a rigorous characterisation of the retrieval uncertainties, including propagation of measurement noise, the uncertainty in parameters assumed by the model and the uncertainty in the forward model itself. The retrieval also provides information about the quality of the fitting, such as the number of iterations it took to minimise the retrieval to an acceptable level and cost. Similar to a χ² statistic, cost is a combination of the squared deviations between the measurements and forward model as well as the retrieved state vector and a priori state vector, each weighted by their associated covariance matrix. The cost provides an indication of how well the measurements fit the model. A cost of less than 5 is taken to mean the model was a good fit to measurements, though the exact threshold used does not greatly affect the conclusions. A higher cost indicates the model is failing to capture the observed conditions, such as multiple layers of cloud.

The radiation products are created using BUGSRad (Stephens et al., 1991) in a similar manner to Fu and Liou (1992). BUGSRad is a radiative transfer algorithm based on the two-stream approximation and correlated-k distribution methods of atmospheric radiative transfer. It is applied to a single-column atmosphere for which the cloud and aerosol layers are assumed to be plane-parallel. Cloud and aerosol properties retrieved using CC4CL, together with co-located visible and near-infrared surface albedo from MODIS (Schaaf et al., 2002), are ingested into BUGSRad to compute both shortwave and longwave radiative fluxes for the top- and bottom-of-atmosphere results. Total solar irradiance is drawn from the Solar and Heliospheric Observatory (SOHO; Domingo et al., 1995). The algorithm uses 18 bands that span the electromagnetic spectrum to compute the broadband flux. In total, 6 bands are used for shortwave and 12 bands are used for longwave radiative flux calculations. To account for the low sampling frequency of the polar orbiting satellite and the dependence of the shortwave fluxes on viewing geometry, an angular dependence correction is applied to the shortwave radiation properties to make the L3C monthly products represent 24 h averages. Further details are outlined in Stengel et al. (2019).

Since V2 was produced, a number of developments have been made regarding the algorithm, resulting in considerable improvement to the ATSR-2/AATSR records as summarised below. Figures 1 and 2 show global maps of yearly average cloud properties from 2008 for V3 Level-3C data compared to that from V2.

• The cloud mask was retrained using a larger data set including 1 km CALIPSO data. This has reduced the number of clouds falsely detected over polar regions (sea, sea ice and land), reduced cloud coverage in the topics and increased the number of clouds detected in stratocumulus cloud banks.

• The cloud phase selection in V2 used a threshold scheme developed by Heidinger and Pavolonis (2009). This has been replaced with an ANN approach for V3 (Stengel et al., 2019). The change significantly increased the number of clouds in the liquid phase and had a follow-on effect on other variables such as the LWP (increase) and correspondingly decreased the IWP and cloud albedo, particularly in polar regions and the northern and southern storm track regions. This change also affects the retrieval of cloud top height, with an overall reduction in the height of the clouds. This is particularly evident in the tropics and the stratocumulus cloud banks, accompanied by an increase in LWP. It also impacted the CER, as liquid clouds have a smaller effective radius than ice clouds.

• The surface reflectance model was revised to correct a bug in the application of large solar zenith angles over bright polar surfaces. This resulted, in a significant decrease in the COT and CER, to much more realistic values. Changes outside the polar regions were minimal.

• The look-up tables (LUTs) are now based on Baum et al. (2014) ice optical properties instead of Baran et al. (2004). This resulted in significantly smaller ice CER and COT, with a corresponding reduction in IWP to more realistic values.

• In V2, maintaining consistency with the earlier sections of the AVHRR record required using lower-resolution (and less accurate) auxiliary data sets for ice and snow, such as European Centre for Medium Range Forecasting (ECMWF) reanalysis, and inferior land sea masks. This resulted in poor results over mountainous and snow- or ice-covered regions. These auxiliary data sets were also not consistent with those used by the AATSR ORAC Aerosol_cci. In V3, we implemented the higher resolution National Snow and Ice data centre (NISE) masks (Brodzik and Stewart, 2016), improving retrievals over snow and ice covered surfaces.

• In V2 there was a discontinuity between the ATSR-2 and AATSR cloud retrievals, particularly in cloud fraction, COT and CER. This discontinuity was caused by a number of factors:

  • The use of the 3.7 µm channel to generate the data set, which differs in dynamic range between ATSR-2 and AATSR.

  • Differences between the two instruments in the availability of shortwave channels across the swath during the day.

• In order to create a record which minimised the inconsistency between ATSR-2 and AATSR (and the aerosol record), in V3 cloud properties were retrieved using the 1.6 µm channel rather than the 3.7 µm channel and only retrieved for the narrow swath mode of ATSR-2 when all channels are present.

The key strengths of the Cloud_cci data sets have been retained in V3.

• The spectral consistency of derived parameters, which is achieved by an OE approach based on a physically consistent cloud model simultaneously fitting satellite observations from the visible to the mid-infrared.

• Uncertainty characterisation, which is inferred at pixel level from OE theory, that is physically consistent (1) with the uncertainties of the input data (e.g. measurements, a priori) and (2) among the retrieved variables. These pixel-level uncertainties are further propagated into the monthly products.

• Comprehensive assessment and documentation of the retrieval schemes and the derived cloud property data sets, including the exploitation of applicability for evaluation of climate models and reanalyses.

Figure 1 Examples from 2008 of Level-3C (yearly average) Cloud_cci AATSR V3 (left column), V2 (middle column) and the difference between V3 and V2 (right column). From the top the rows represent the following variables: cloud fraction (CFC), liquid cloud faction (CPH), cloud optical thickness (COT) and cloud effective radius (CER).

Figure 2As in Fig. 1 but for cloud top height (CTH), liquid water path (LWP), ice water path (IWP) and cloud albedo (CLA).

4.1 Cloud fraction

The AATSR Level-2 cloud fraction products are retrieved at 1 km resolution, and thus the retrievals were co-located with the CALIPSO 1 km cloud products. These are less sensitive to thin clouds than the 5 km products (NASA, 2019), which were used in the evaluation of the Cloud_cci AVHRR products (Stengel et al., 2019). The validation was repeated using the 5 km products (not shown) and the changes were negligible. The Hanssen–Kuiper skill score (KSS), an often used skill score (Hanssen and Kuipers, 1965) is defined as $\text{KSS}=\text{TPR}-\text{FPR}$, where TPR is fraction of pixels correctly identified as cloud and FPR is the fraction of pixels wrongly identified as cloud. The results are consistent with the results found for the AVHRR Cloud_cci product in the same region.

The results of the comparison are shown in Table 3. The cloud detection has clearly improved for V3, with the most significant improvement being the identification of clear pixels. This result is consistent with the reduction in cloud fraction observed over the poles in the global maps. The cloud detection improved for both clear and cloudy pixels. The identification of clear pixels increased significantly from a probability of detection (POD_clear), i.e. the fraction of pixels identified correctly as clear, of 61 % to 76 %.

Table 3 Comparison of co-located AATSR cloud mask and CALIPSO 1 km layer product for V2 (left) and V3 (right). The comparison metrics shown are hit rate, the percentage of pixels identified correctly as either cloudy or clear, the Probability Of Detection (POD) for cloudy and clear pixels separately, the Hanssen-Kuiper skill score (KSS; defined as $\text{KSS}=\text{TPR}-\text{FPR}$), and the total number (cloudy and clear) of co-locations used in the analysis.

Download Print Version | Download XLSX

4.2 Cloud top height

The cloud top height product was validated using the CALIPSO 1 km product. In previous studies (Sus et al., 2018) it was shown that the CTH retrieval is more accurate when the cloud is opaque or single layer. Here, the opacity flag from the CALIPSO 1 km layer product is used to verify this finding. The opacity flag indicates features that completely attenuate the lidar beam (NASA, 2019). Results are summarised in Table 4 and are presented separately for all cloud observations, only opaque clouds, the cloud top height corrected for penetration depth and for all clouds retrievals with a cost less than 5 (as would be expected for single-layer clouds). The cost is an output of the optimal estimation retrieval scheme and is the result of the squared deviations between the measurements and the forward model (which in this scenario is a single layer of cloud) and the retrieved state vector and the a priori state vector, weighted by an associated covariance matrix. Essentially, it is an indicator of whether the observed measurements were a good fit to the forward model. A cost of less than 5 indicates the measurements fit the model well. A higher cost would indicate we are viewing cloud from multiple layers, for example. From V2 to V3, the correlation with the lidar measurements for all the co-located pixels was unchanged, but the bias decreased from 1.3 to 0.9 km. When only opaque clouds are considered, the correlation increases considerably but with a slight increase in bias. The corrected cloud top heights are produced by approximating the observed brightness temperature as emitted from one optical depth into the cloud and assuming the cloud is vertically homogeneous with a constant lapse rate. This product achieves a similar correlation to the all-cloud results but further reduces the bias from 0.94 to 0.85 km for V3. For clouds with a cost of less than 5, the correlation decreased and the bias also reduced. High costs are indicative of multi-layer cloud, such as thin cirrus over liquid cloud. These are typically retrieved as some weighted average of the two layers, returning a nonphysical value (Poulsen et al., 2012). Overall, V3 is a superior cloud top height (temperature and pressure) product. The comparison was also performed with the CALIPSO 5 km layer (not shown), and the sensitivity to optical depth is investigated. The results showed negligible variation with optical depth threshold and with the 5 km product.

Table 4 Comparison of AATSR cloud top height with co-located CALIPSO measurements for 5 d in 2008. On the left V2 is shown, and V3 is shown on the right . The results are shown for all observations, only opaque clouds (as defined by CALIPSO), the corrected cloud top height product and retrievals with a cost of less than 5. All values are in kilometres.

Download Print Version | Download XLSX

4.3 Liquid water path

The liquid water path of the Cloud_cci data sets is evaluated against the Multisensor Advanced Climatology of Liquid Water Path (MAC-LWP) data set (Elsaesser et al., 2017) over ocean. The MAC-LWP climatology is based on retrievals from multiple microwave radiometer instruments, including the SSM/I series (Special Sensor Microwave Imager), the Tropical Rainfall Measurement Mission Microwave Imager (TMI) and the Advanced Microwave Scanning Radiometer for EOS (AMSR-E). Microwave measurements of LWP are typically more accurate than visible imagers because microwave instruments are able to penetrate deep convective clouds and ice over water clouds while also measuring the LWP at lower altitudes, which is not possible for passive imagers. Their disadvantage is the large footprint, up to 0.25^∘.

This evaluation focuses on regions where liquid clouds are dominant (i.e. fewer than 5 % ice clouds), specifically three stratocumulus regions: the oceanic area west of Africa at 10–20^∘ S, 0–10^∘ E (SAF); the area west of South America at 16–26^∘ S, 76–86^∘ W (SAM); and the area west of California at 20–30^∘ N, 120–130^∘ W (NAM), similar to the analyses in the Cloud_cci Product Evaluation Intercomparison Report (PVIR, 2018). The MAC-LWP data set was co-located with the Cloud_cci data set on a monthly basis. No correction was made for the diurnal cycle as it is assumed to be small in the selected regions. A time series of the comparison is shown in Fig. 3 and summarised in Table 5. There was a significant improvement from V2 to V3, particularly in the consistency between the ATSR-2 and AATSR instruments. The V2 data set showed a large offset between ATSR-2 and AATSR, which has almost disappeared in V3. The correlation with MAC-LWP exhibited in V3 is extremely good, over 0.8 for all regions, which is a significant improvement over V2, particularly for the region off the Californian coast. The associated bias and standard deviation are also very low, typically less than 5 % of the total liquid water path.

Figure 3Comparison of ATSR-2/AATSR Cloud_cci LWP time series (coloured) with MAC-LWP (black) for three regions: west of California (NAM, a, b), west of Africa (SAF, c, d), and west of South America (SAM, e, f). Version 3 (a, c, e) has reduced the discontinuity between sensors seen in version 2 (b, d, f).

Download

Table 5 Multi-annual (2000–2012) liquid water path validation results for ATSR-2/AATSR when compared with MAC-LWP monthly data for three regions of predominantly stratocumulus cloud. The results for V2 (left) and V3 (right) are compared for correlation, bias and standard deviation.

Download Print Version | Download XLSX

Figure 4 Examples of Level-3C (yearly average for 2008) Cloud_cci AATSR V3 (left column), CERES (middle column) and the difference between CERES and AATSR (right column): global maps of fluxes, from top to bottom, ${\text{LWF}}_{\mathrm{TOA}}^{\mathrm{up}}$, ${\text{LWF}}_{\mathrm{TOA}}^{\mathrm{up}}$ clear, ${\text{SWF}}_{\mathrm{TOA}}^{\mathrm{up}}$ and ${\text{SWF}}_{\mathrm{TOA}}^{\mathrm{up}}$ clear.

Figure 5 Examples of Level-3C (yearly average for 2008) Cloud_cci AATSR V3 (left column) and CERES (middle column) and the difference between CERES and AATSR (right column): global maps of forcing, from top to bottom, ${\text{LWF}}_{\mathrm{BOA}}^{\mathrm{down}}$, ${\text{LWF}}_{\mathrm{BOA}}^{\mathrm{down}}$ clear, ${\text{SWF}}_{\mathrm{BOA}}^{\mathrm{down}}$ and ${\text{SWF}}_{\mathrm{BOA}}^{\mathrm{down}}$ clear.

4.4 Comparison of radiative fluxes

Examples of the Cloud_cci AATSR flux products are shown in Fig. 4. In the left column the multi-month mean for 2008 from AATSR is shown. The AATSR products are compared with the Clouds and Earth Radiation Energy System (CERES) Energy Balanced and Filled (EBAF) Top-of-Atmosphere (TOA) and Bottom-of-Atmosphere (BOA) fluxes edition 4.1 (Loeb et al., 2018) shown in the middle column, and the right column shows the difference between the two. The two data sets show very similar global patterns. The highest TOA longwave cloudy fluxes are observed over warm land, typically desert regions and ocean regions with low cloud coverage. The highest TOA longwave upwelling clear-sky fluxes are found in the tropics and mid-latitudes. The highest TOA cloudy shortwave fluxes are located over the bright polar regions, deserts and regions of bright cloud, such as the storm tracks and stratocumulus cloud decks. The TOA clear-sky shortwave fluxes are low over the oceans and higher over bright land surfaces. The global mean comparison of TOA fluxes is summarised in Table 6. The means were compared for both ±60^∘ latitude and the whole globe when both CERES and AATSR are present. For the TOA fluxes, the LW fluxes show the largest negative bias over land and largest positive biases over sea, particularly in clear conditions. The CERES team have evaluated the accuracy of the TOA products in Loeb et al. (2018) and state that their all-sky shortwave and longwave monthly uncertainty (which is comprised of both random and systematic error sources and is specified for the global region) is 2.5(3) W m⁻² for Aqua and Terra (Terra only) period, while the clear-sky shortwave and longwave uncertainty is 5(6) and 4.5(5) W m⁻², respectively, for the Aqua and Terra (Terra only) monthly products. The period compared here covers the AQUA period. The agreement between AATSR TOA-derived fluxes and CERES is within this uncertainty. Averaged globally, the AATSR all-sky longwave fluxes are slightly lower (colder) over the sea (particularly in the tropics; see the red areas in the difference map) and slightly higher (warmer) over land. Comparing the all-sky and clear-sky differences suggests that the longwave radiative fluxes associated with clouds (particularly regions with high-altitude ice clouds) are higher in AATSR than in CERES. TOA all-sky shortwave flux shows a similar pattern to the TOA all-sky longwave bias, although it is reversed in sign. Both shortwave and longwave clear-sky TOA fluxes are systematically lower than CERES, indicating a potential underestimation of the surface reflectance and surface temperature in the AATSR product. The AATSR sea surface reflectance model uses a Cox and Munk (1954) formulation. For the longwave channels, a key source of differences could be the sensitivity to the diurnal correction applied to the AATSR data in order to make a like-for-like comparison with CERES. These differences will be investigated for improvement in future versions.

Table 6 Multi-annual (2003–2012) zonally averaged broadband shortwave and longwave fluxes (SWF, LWF) at the top-of-atmosphere (TOA) inferred from the Cloud_cci AATSR V3 data set. Two latitude ranges, −60 to 60^∘ (top) and −90 to 90^∘ (bottom), are presented. The values are compared with the equivalent values from the Clouds and Earth Radiation Energy System (CERES) Energy Balanced and Filled (EBAF) fluxes (all values are given in W m⁻²). The differences and relative differences are also reported.

Download Print Version | Download XLSX

Table 7As in Table 6 but for the bottom-of-atmosphere values (BOA).

Download Print Version | Download XLSX

The BOA longwave downwelling fluxes (all-sky and clear) have a minimum in the cold polar regions and a maximum in the tropics. The corresponding shortwave fluxes are lowest in the southern and northern storm tracks and peak in the tropics. The BOA all-sky shortwave downwelling flux shows the largest regional differences. The BOA shortwave downwelling clear-sky fluxes show AATSR to be higher in regions of high aerosol loading. The downward longwave fluxes are also higher for AATSR. The BOA LW fluxes show the largest disagreement with CERES. The global mean BOA comparisons are summarised in Table 7. The means were compared for ±60^∘ latitude and the whole globe. The CERES data quality statement indicates NASA (2020) uncertainties of 9 and 8 W m⁻² for longwave BOA downwelling all-sky and clear sky, respectively, and 14 and 6 W m⁻² for the shortwave BOA downwelling all-sky and clear sky, respectively. This is consistent with the AATSR shortwave values. The BOA clear and all-sky longwave differences with CERES are very slightly outside this range. We hypothesise that this difference that the AATSR cloud base height is systematically biased. While the AATSR fluxes also use satellite aerosol measurements in the clear-sky calculations, the impact on the shortwave flux is less pronounced than in the CERES product, which used MODIS aerosol products. This difference and difference between longwave downwelling fluxes will be investigated for future improvements to the retrieval.