CLARA-A3 (The CMSAF Cloud, Albedo and Surface Radiation dataset from AVHRR data – third edition) is a climate data record derived from the polar-orbiting AVHRR instrument, of cloud properties, surface albedo, and surface radiation budget products for the period 1979–2024 45 years;. The major advantage of this dataset is the global coverage and the time-period available for analysis. The CLARA-A3 edition contains retrieved cloud properties such as CTH, CPH and cloud water path (CWP). The CLARA-3 dataset is available for instantaneous, daily, and monthly periods and was released in mid-2023. The data used in the analysis have a spatial resolution of 5 km. The CLARA A3 uses an artificial neural network (ANN), trained on collocated AVHRR-CC data to obtain cloud-top height. In the case of cloud microphysics, traditional retrieval methods look-up tables (LUTs) are created based on .

The AVHRR PATMOS-x provides a climate record of atmospheric cloud properties and brightness temperatures retrieved by merging the information from AVHRR, and the collocated High-resolution Infra-Red Sounder HIRS,. The algorithm was developed by the NOAA and the Cooperative Institute for Meteorological Satellite Studies (CIMSS). The climate data record covers a period of 44 years (1979–2023), with a resolution of 0.1°×0.1° grid globally . The dataset includes CTH, CER, COD, and ice water path (IWP). Multiple algorithms are referenced in the product, in particular the Advanced Baseline Imager Cloud Height Algorithm ACHA;, MODIS CO2 slicing algorithm , and Daytime Cloud Optical and Microphysical Properties DCOMP;. In this study, the ACHA CTH algorithm uses 11 and 13.3 µm observations along with the split-window approach , and data from the AVHRR Extended (CLAVAR-x). When AVHRR is combined with HIRS information, the CO2 CTH can be determined.

Figure  shows the difference in CTH for the top most layer between the active and passive sensors, alongside the SST for the full period and by season. When the data is analysed as a function of latitude band in the ocean region from 62–70° S, both passive sensors generally overestimate the CTHs compared with the active sensor CTHs, except during winter (JJA), when only MODIS overestimates the CTH. This region is typically covered by sea ice during this time of year. The SST also shows variation during corresponding season with a low SST in winter and a high SST in summer. However, there does not seem to be a strong correlation with the differences. In the poles over the land area irrespective of the seasons, AVHRR CMSAF tends to underestimate the CTHs, while MODIS tends to overestimate the CTH. As per , the cloud amount detected by the active sensors and passive sensors varies overall when comparing L3 monthly data. As CALIOP is an active sensor, it is more sensitive to high-thin clouds (CTH>8km) and the vertical distribution of the clouds. The presence of atmospheric temperature inversions may cause the passive sensor to either underestimate or overestimate the CTHs in the boundary layer . Thus, factors, such as the presence of multilayer clouds, differences in sensor sensitivity and resolution, and differences in retrieval algorithms may account for some of the poor agreement.

Table  summarises the comparisons, with mean CTH values in Table a and the statistics for the comparison with CALIOP in Table b by season and overall. It can be observed that the active sensor correlation with the passive sensors for the overall monthly CTHs is low with values ranging from 0.36 for CALIOP-AVHRR CMSAF (C-A) to 0.32 for CALIOP-MODIS (C-M). The CTH difference between the C-A sensors for DJF ranges from -1 to 1 km for the study domain, while for C-M it ranges from 0–4 km. The MBE is 0.46 km for C-A and 1.12 km for C-M, showing a larger offset for MODIS. The RMSE for C-A is 2.35 km; for C-M, the value is 2.60 km; and for A-M, it is 0.98 km, which is consistent with both correlation and MBE trend. Overall, the MBE between the MODIS-AVHRR CMSAF (A-M) passive sensors is larger for the CLARA-A3 dataset (0.98 km), but the correlation is 0.68. Overall, there is a positive mean bias error for the passive sensors AVHRR CMSAF and MODIS CTHs.

The seasonal comparison of the C-M data over the course of a year shows that the austral autumn (SON) has the weakest correlation (0.28) and MBE (1.25 km). Meanwhile, the inverse is true for the austral summer (DJF) with a correlation of 0.36 and MBE of 1.05 km with MBE>1km, positive bias over the mid-latitudes, and smaller bias (MBE±0.5km) over the high-latitudes. Another intriguing observation is that in austral winter (JJA), MODIS overestimates the CTHs in the high-latitude (MBE>1km) and underestimates in the mid-latitudes(MBE>1.5km), with an overall season MBE of 1.20 km. The seasonal comparison of CTHs for C-A shows that the correlation is highest in the austral spring (SON) for C-A (0.34) and lowest in the summer (DJF -0.15). AVHRR CMSAF tends to overestimate the CTHs in the mid-latitudes and underestimate them in the high-latitudes during winter, with an overall MBE of 0.25 km. In all other seasons, AVHRR CMSAF tends to overestimate the CTHs across the region, with the exception of Antarctica.

The passive sensor comparison shows that DJF has the highest correlation at 0.68, whereas JJA has the lowest at 0.45. Additionally, DJF exhibits the lowest mean bias error of 0.09 km for C-A among all seasons and 0.98 km for C-M. This may be due to the seasonality of cloud cover over the SO, as JJA tends to have more high-clouds and passive sensors fail to detect the cloud due to poor sensor sensitivity to optically thin high clouds. The results indicate that the AVHRR CMSAF generally shows a tendency to overestimate the CTHs compared to CALIOP, although in some subregions (e.g. Antarctic land) it underestimates. Conversely, MODIS tends to underestimate the CTHs overall, but can overestimate in specific latitude bands such as 62–70° S.

The comparison of CFC for the active and passive sensor data overall and seasonally is shown in Fig. . The overall 7 year mean CFC values for the SO were 0.82 for CALIOP, 0.86 for MODIS, and 0.85 for AVHRR CMSAF. This is consistent with the findings of the review study conducted by . The cloud fraction for all sensors, both passive and active, has more cloud cover over the mid-latitudes (40–60° S) than over the high-latitudes (>60° S). In general, the cloud cover decreases (<70%) towards very high-latitudes near the coast of Antarctica for all sensors, which is also consistent with findings of . This decrease, especially in winter, is likely due to the presence of sea ice near the Antarctic coast. This inhibits the flux of water vapour (surface evaporation), as presented in studies by and . The high cloudiness for this area can be attributed to the frequent synoptic-scale and mesoscale depressions and intense cyclonic activity around the Antarctic Continent . In general, the passive sensors have higher cloud coverage than CALIOP observations for the higher latitude.

The detailed mean and statistics for the CFC comparison are given in Table . It's intriguing to note that the correlation between the passive sensors (AVHRR CMSAF and MODIS) and the active sensor (CALIOP) is low (0.52 and 0.48) over the whole period. On the other hand, the correlation between the passive sensors is higher (0.87), which suggests that the L3 passive retrievals may have a different systematic bias. also observed a difference in the cloud cover for the passive and active sensors, they attributed this to the spatial resolution issues of the active sensor when compared to the passive sensor. The results of this analysis indicate that the active sensor cloud cover was less than the passive sensors, by approximately 4 % for MODIS and AVHRR CMSAF. Note that we are comparing against the CALIOP passive sensor adjusted data set.

We found that CFC values are slightly lower in JJA (winter), with a mean of 0.83 and 0.82, respectively, compared to DJF (summer) when we compared AVHRR CMSAF and MODIS L3 data. In the case of MODIS, the high latitude CFC for JJA is approximately 60 %, the lowest amongst all the sensors. This can be attributed to the presence of sea ice. For the active sensor, JJA has the lowest CFC (0.81). It's also observed that there is a notable difference in cloud cover over land for different passive sensors. MODIS overestimates the cloud cover over land; on the other hand, AVHRR CMSAF underestimates the cloud cover compared to the active sensors. In the RMSE and MAE seasonal analysis for all sensor comparisons, it was found that JJA has the highest value for both amongst all seasons. However, for the MBE, it's the opposite, with JJA having the lowest value (≈0.01km) for all sensors. The correlation also follows the same trend as MBE errors: low in JJA and high in DJF. For comparisons among the passive sensors, the agreement of pixels identified as cloudy by both sensors is more consistent, and the root mean square error (RMSE) is low for all seasons. All of the comparison's seasonal findings are consistent with the conclusions of .

The results indicate that CALIOP detects fewer clouds than the passive sensors. This could be due to a difference in resolution or in the approach taken to make the L3 data set comparable with passive sensors. The total cloud cover for the SO was found to be around 86 % (MODIS), 85 % (AVHRR CMSAF) for passive sensors and 82 % for CALIOP.

The overall cloud fraction of liquid water clouds (hereafter CFL) and seasonal distributions are illustrated in Fig. . For the entire year, CALIOP, AVHRR CMSAF, and MODIS show the presence of liquid clouds at around 50 %. In terms of the correlation for the CFLs, the passive sensor comparison is better (0.71) than the active sensor comparison (0.37 C-A and 0.44 C-M). observed that over Antarctica, ice clouds are more prevalent and along the coast of Antarctica, mixed-phase clouds are more prevalent. Other studies also observed that mixed-phase clouds are prevalent over the SO as they are usually misclassified more over the SO. From Fig. , we can conclude that AVHRR CMSAF is identifying more liquid clouds over the high-latitudes than other products, while other passive products classify the clouds as ice clouds or mixed-phase clouds. The comparison's mean and statistical analysis are given in Table . The comparison shows that AVHRR CMSAF primarily overestimates the CFL, while MODIS tends to underestimate it when compared to CALIOP. However, around the high-latitudes, towards the coast of Antarctica the AVHRR CMSAF underestimates the CFL when compare to CALIOP. In the case of seasonal comparisons, the correlations for passive sensors against active sensors are low overall in all seasons. For all the sensors, the CFL is consistently higher in DJF and the lowest in JJA. When we look at the MBE, AVHRR CMSAF has a negative bias for all seasons, while MODIS has a positive bias for all seasons compared to CALIOP. AVHRR CMSAF exhibits the smallest bias in JJA and the highest in DJF, while MODIS displays the smallest bias in DJF and the highest in JJA. This suggests a discrepancy in cloud classification during the liquid phase between the sensors in DJF. This disagreement can be attributed to the presence of supercooled liquid in DJF . The C-A seasonal comparison reveals that the AVHRR CMSAF overestimates the liquid cloud cover across the study region in DJF and JJA, with the exception of Antarctica, where it underestimates the amount of liquid cloud. However, in SON and MAM towards the latitudes>55° S, both sensors underestimate the CFL. The presence of sea ice explains this difference in SON and MAM. In summary, the comparison of the L3 data for the CTH shows a bias in the retrieval between passive sensors, underestimating the CTH by around 1 km compared to the active sensor: the mean bias error was 0.96 km (C-A) and 1.12 km (C-M). The study of L3 CFC and CFL revealed that AC overestimates the amount of clouds compared to observations made by MODIS and CALIOP, with the overestimation increasing during the winter months.

To illustrate these differences in retrieval behaviour, we present a case study over the SO on 19 October 2015, which was used to examine how passive and active sensors capture cloud properties under varying cloud structures. Figure  shows the 11 µm brightness temperature swaths for MODIS and CTH swath for MODIS (MYD), AVHRR PATMOS-x (AP) and AVHRR CMSAF (AC) and in black the merged CloudSat-CALIOP track for the 19 October 2015 at 08:37 UTC. Although all three sensors observed a similar high cloud pattern around 50° S: AC retrieves the CTH at around 11 km height, while MYD and AP retrieve the CTH at around 9 km height. Finally, AP has a significant region of cloud with CTHs at around 5 km height located at 40° S and 80° E, while AC and MYD observed low clouds with CTHs at around 2 km height for the same region.

Figure  shows a curtain plot for the CC track and the corresponding CTH, COD, CER, and CPH for each instrument. The top panel shows the CTH for all passive sensors (AC, AP, MYD, and CC), as well as the vertical profile of COD from the CC dataset (2BFL). The CTH for CC at a COD limit of 0.5 is also shown (cyan) to illustrate the potential differences in optical depth sensitivity between the passive and active sensors. The second panel illustrates the COD for the four sensors, along with the multilayer mask derived from CC. The COD was derived using the 3.7 µm channel, as it is common to all passive sensors. The third panel displays the CER, as well as the sea ice mask used in the AP retrieval. The bottom panel shows the cloud phase for each instrument; for CC the phase is the phase detected at the top of the cloud. This example illustrates the highly variable nature of the cloud field across the Southern Ocean. Our discussion focuses on three sub-sections: multilayer clouds (Fig. a[i]); thick upper-level clouds over thick lower-level clouds (Fig. a[ii]); and optically thick boundary layer clouds (Fig. a[iii]).

In the first panel, the CTH ranges from 0–12 km height and shows substantial differences between the CTH of passive and active sensors, with both underestimation and overestimation by passive sensors compared to CC. Figure a[i] shows a multilayer cloud structure with an optically thin high-level cloud over optically thick mid-level clouds. MODIS and AP detect the high-level cloud, however since it is optically thin over an optically thick lower layer cloud, it places the CTH at an effective CTH between the upper and lower layers. For the AC retrieval, the CTH is similar to the CC height. In Fig. a[ii], AC produces higher CTH values than CC, MYD yields lower values, and AP's results are closest to those of CC. This difference could be due to the different numerical weather prediction (NWP) reanalyses used temperature profiles. Indeed, the CTH can vary substantially, particularly at high altitude and depends on the difficulty in identifying the inversion at the tropopause .

Figure a[iii] shows a region of boundary layer clouds around 41.5° S. AC does a good job of retrieving the boundary layer cloud, while AP boundary layer CTH can show large over estimations, and MYD occasionally over and underestimates. Boundary layer clouds can be difficult to estimate, particularly when temperature inversions are present, as they can present multiple solutions when using the thermal characteristics to retrieve CTH . AP misclassifies the phase between 41.5–40° S which may explain the high CTH retrieval and highlights some internal dependency in the retrieval approach.

The second panel (Fig. b) displays the CC multilayer mask and identifies the areas with multilayer clouds well. The COD for each retrieval is also compared in the second panel. The COD of CC (2BFL) exhibits high variability when compared to the COD of the passive sensors, particularly for ML clouds. CC COD is substantially higher than that of the other sensors. The COD has large differences among the passive sensors and active sensor in the presence of a multilayer structure (Fig. a[i] and [ii]); In single-layer areas, the COD of MYD, AP and AC are more in agreement with the active sensor measurements. The results align with the product statement for CC COD which states that the COD is similar to MYD for single layer and varies with multilayer clouds. The third panel (Fig. c) shows the CER comparison and the sea ice flag. While the general pattern of the CER from the passive sensors is in agreement, the values of AP CER are higher than MYD and AC CER retrievals in general, particularly where ice clouds detected are quite different.

The cloud phase is shown in the fourth panel (Fig. d). For CPH, each retrieval algorithm uses a slightly different phase classification scheme. The CC retrievals classify the phase as mixed, liquid or ice phase on the basis of cloud top and base temperatures, radar reflectivity and integrated attenuated backscattering coefficient across all the cloud layers in the vertical column . For the passive sensors (MYD, AP and AC), the CPH is the phase detected at the top of the cloud. In the case study, all sensors share ice and liquid as common classes, with an additional clear sky class for MYD, a supercooled liquid class for AP, and a mixed phase class for CC. Further classification of the phase can be seen in Table . The agreement between the passive sensors overall is good with occasional differences; for example, at 43° S, AC and MYD classify the cloud phase as liquid while CC and AP classify it as ice, with cloud optical depth indicating a thin upper layer.

To summarise, we observe significant variations in the retrieved cloud properties in the presence of multilayer clouds, with passive sensors both overestimating and underestimating the retrieved CTH compared to active sensors. The dependency on the optical depth on the topmost cloud layer, the penetrative property, and the sensitivity of passive sensors are seen in this case study. The case study demonstrates that cloud property retrievals differ based on cloud type, retrieval algorithm, and sensor sensitivity. Multilayer clouds often exhibit the greatest discrepancies.

The cloud mask products from the passive sensors are evaluated against the CloudSat-CALIPSO product (Table ). The true positive rate (TPR) is similar for AVHRR CMSAF (0.94) and AVHRR PATMOS-x (0.94) and smaller for MODIS (0.77). The false positive rate (FPR) vary considerably among the products, 0.22 for AC, 0.07 for MYD, and 0.51 for AP. The Kuiper Skill scores are relatively similar, 0.71 for AC, 0.70 for MYD, except for AP, 0.43. The higher resolution of MODIS may contribute to the higher KSS compared to AVHRR PATMOS-x.

According to , when the global cloud masks of the CLARA A3 and CALIPSO (5 km cloud product) were compared, for the years 2006–2015, the overall hit rate was 0.82 and the KSS score was 0.68. The hit rate for high latitudes in both hemispheres was 0.85, and the KSS score was 0.69. For polar regions, the hit rate was 0.69, and the KSS score was 0.49. The KSS score from our analysis comparing AC cloud mask and CC is 0.71, is similar, but the hit rate is higher (0.94). The differences could be attributed to the active sensor data considered (CC and CALIPSO) i.e. differences in CC cloud masking algorithms and the validation method. The hit rate for MYD and CC comparison is the lowest (0.77); on the other hand, the FPR is also the lowest amongst the sensors (0.07). In conclusion, the comparison of KSS scores and hit rates across different sensors and regions revealed variations that highlight the importance of balancing FPR and hit rate, passive cloud masking algorithms, and validation methods in evaluating cloud detection accuracy.

Figure  shows a 2D histogram plot of passive satellite CTH compared with active (CC) CTH for the year 2015. A summary of the comparison of the average CTH with and without a COD limit>0.5 is given in Table . We chose a COD threshold of 0.5 to exclude optically thin clouds that are often detected by CALIOP but missed or poorly retrieved by passive sensors, consistent with previous studies (Karlsson et al., 2017; Marchant et al., 2020; Yost et al., 2023).

The average CTH of CC is higher without the COD limit, indicating that the high thin cloud is removed when this threshold is applied. In general the average height of passive sensors is higher than the CC with CTH threshold>0.5 and lower than the CTH without threshold. Analysis of the full dataset (2015) reveals mixed agreement between passive and active sensors, with AC (5.37 km) demonstrating better agreement than CC CTH without threshold (5.14 km), and AP (4.24 km) and MYD (4.11 km) demonstrating better agreement with CC CTH with COD>0.5 (3.91 km). For both CC mean CTH (with and without COD>0.5), single-layer clouds (3.41 and 3.84 km), passive sensors exhibit good agreement with each other.

Multilayer cloud identification was applied using a mask derived from CC CTH. The results for the multilayer cloud are unsurprisingly less accurate and inconsistent. The AC mean CTH (7.18 km) shows better agreement with the CC mean CTH without the threshold (8.40 km), indicating that it is more sensitive to the mean CTH of high thin clouds than MYD (5.37 km) and AP (5.35 km), which show better agreement with the CC CTH COD>0.5 threshold (4.81 km).

Figure  shows a 2D histogram comparison of passive satellite CTH with active (CC) CTH for 2015. The CC CTH plotted was CTH with a COD>0.5. Each row shows the results for a single passive satellite: AC is the top row (Fig. a–c), MYD is the second row (Fig. d–f) and AP is the bottom row (Fig. g–i), each column shows the result, from left to right: the full dataset (2015), single-layer clouds, and multilayer clouds.

A number of features become apparent in Fig.  MYD and AP plots display a horizontal band, signifying the passive sensor's underestimation of the CTH, which aligns with the earlier analysis. In general, the CTH of low clouds are overestimated by the passive sensors. The passive sensors underestimate the CTH of high clouds in MYD and AP. An explanation for the overestimation of low CTH is the presence of a sharp temperature inversion that provides two solutions for the same IR temperature from the NWP reanalyse profile, one above and one below the actual CTH (Fig. a[iii], ). For clouds that have underestimated the CTH, there are two possible reasons for this. Firstly, the clouds are thin and the retrieval does not properly account for the extinction of the cloud or a contribution to the TOA radiance from the surface . Secondly, the passive sensor IR channels often penetrate into the clouds, as seen clearly in the case study Fig. a[ii].

The differing performance of the passive data sets is in part due to different algorithm approaches. The AC CTH has been derived by training a neural network with collocations of CALIOP cloud top heights. While the MYD retrievals uses a physical model based on CO2 slicing and IR window approach for the retrieval of CTH. In AP, CTHs are retrieved using physical models, the ACHA algorithm and the CO2 slicing method . As the results of the CO2 slicing method were poor, we have considered only the ACHA method CTHs. The ACHA methods employs a combination of IR channel observations and the split-window approach , to retrieve the CTHs.

The analysis of L2 CTH properties has shown significant differences between products. There are differences not only between retrievals using the same instrument (AP and AC), but also between retrieval algorithms and between different instruments. The agreement between all sensors was reasonable for cloud top heights in the case of single-layer scenes. However, scenes classified as multilayer exhibited a lower level of agreement. The general findings on CTH are in agreement with previous studies . observed the maximum CTH uncertainty for MODIS (L2) for a single-layer unbroken low cloud of 0.540±0.690km. In the MYD collection, CTH bias with CALIOP was reduced to 0.197 km for low-level boundary liquid clouds .

In our comparison of the overall L2 collocated data of active and passive sensors, the mean bias in each sensor for CTH, against CC without threshold, varies with MODIS (MYD) 1.36 km, AVHRR CMSAF (CLARA A3) 0.33 km and AVHRR PATMOS-x (V6) 1.23 km. When it comes to multilayer classified pixels, the passive sensors mean bias substantially underestimates the CTH (MYD – 3.03 km, AC – 1.22 km, and AP – 3.05 km). Given that the total number of multilayer identified scenes represents one-fourth of the total scene, the bias is considerable. Overall, the AC neural network retrievals performed the best, delivering higher sensitivity to thin clouds and performing significantly better for multilayer and high clouds.

This section compares CER from the passive sensors as a function of CPH and the presence of single or multilayer clouds. The passive sensors categorise thermodynamic phase into various classifications according to their retrieval algorithms and the categories used by each dataset are presented in Table . In the results presented, the number of classes was reduced: classifications of water, fog, and supercooled liquid are grouped as liquid; ice, cirrus, and overlap are treated as ice. When this generalised CPH classification is used, more liquid-phase clouds are found in SL cases, while ice dominates in ML cases. A comparison between the different sensors AC, MYD, and AP is illustrated in Fig.  for the year 2015, stratified for SL and ML cases, and summarised in Table .

The liquid CER probability density functions (PDFs) are in broad agreement across the three sensors for all cases. The mean values for liquid clouds are approximately 14 µm for AC, 13 µm for MYD, and 17 µm for AP. The higher CER values from AP suggest potential misclassification of liquid clouds as ice clouds as the proportion of ice clouds in AP is over five times higher than that of liquid clouds. The liquid phase occurs more frequently in SL clouds than ML clouds across all sensors, with similar mean values for both cases (Table ).

In contrast, the CER comparisons for ice-phase clouds reveal notable differences across the sensors. The overall mean for AC is 15.58 µm, while MYD and AP report means of 21.02 and 20.26 µm, respectively. AP also exhibits a distinct bimodal CER distribution with peaks at approximately 10 and 30 µm, particularly evident in SL ice-phase retrievals. ML cases show more similar CER distributions across the sensors. The SL–ML CER difference is relatively small for AC (≈1µm), modest for AP (≈2µm), and larger for MYD (≈5µm). Sensitivity to ice crystal habit selection and the corresponding LUT structure also influences retrievals. MYD uses scattering functions derived from the General Habit Mixture (GHM) developed by and , implemented as band-specific LUTs . The ice habitat selection includes severely roughened aggregates, hollow columns, bullet rosettes, plates, and droxtals. AC uses radiative transfer simulations based on narrower droplet size distributions and an ice crystal model including severely roughened aggregated solid columns, hollow columns, and bullet rosettes . AP uses LUTs from Baum/Yang scattering models with reduced spectral dimensionality , and its ice habit selection includes roughened bullet rosettes, aggregates, and hollow columns . These distinctions in habit selection and LUT configuration influence the consistency and comparability of retrieved cloud microphysical properties across sensors.

Supercooled liquid clouds are consistently retrieved by both AC and AP, particularly in SL regimes. AC reports a substantial supercooled component in full-year statistics, with a mean CER of 14.53 µm, due to its extended phase classification . AP also distinguishes supercooled liquid from warm liquid, frequently identifying it in SL clouds, with a slightly higher mean CER of 14.54 µm. MYD lacks a dedicated supercooled category but includes an “undetermined liquid” retrieval class (ULQ; mean 19.74 µm), likely encompassing supercooled or glaciating clouds not clearly resolved by its phase algorithm . Both AC and AP indicate persistent supercooled liquid presence over the Southern Ocean, consistent with prior observational studies .

The differences observed between ML and SL cases suggest that multilayer clouds influence retrieval outcomes. had similar findings and demonstrated that multilayer conditions, especially non-opaque clouds over liquid, are among the leading sources of phase retrieval uncertainty in passive satellite observations.

In this section, we investigate the distribution of cloud properties as a function of sea ice coverage. The sea ice flag in AVHRR PATMOS-x(AP) retrieval was used to define the presence of sea ice, based on fractional sea ice coverage from NOAA/NCEP Climate Forecast System Reanalysis (CFSR) data . The distributions of the cloud properties CER, CTH, and COD were analysed as a function of sea ice coverage for latitudes>60° S, as shown in Fig. 10. The narrow latitude range was selected to reduce the differences in microphysics caused by different meteorological regimes. The MYD cloud properties are shown in red, AC in black, and AP in yellow. Cloud properties such as COD, CER, and CTH are compared for sea ice detected and non-sea ice detected; the mean and median are summarised in Table .

The mean COD for AC was 34.62 and 14.51; for MYD it was 24.39 and 18.68; and for AP it was 42.55 and 12.77 for sea ice and non-sea ice cases, respectively. The optical depth of the retrieved pixels increased in the presence of sea ice for all sensors. The results disagree with the findings of and , which concluded that clouds were more frequent, and cloud optical depth tends to be higher over ice free regions. While the disagreement could arise from differences in latitude range and the limited temporal scope of our dataset compared to the multi-year analyses and . It is hypothesised that the difference is due to two effects both originating in flaws of the retrieval algorithms: 1 – sea ice (with no associated cloud) is misidentified as cloud; this could be the cause of the spike in observations with greater than 100 COD over sea ice, and 2 – the albedo of sea ice used in the retrieval algorithms introducing a systematic positive COD bias. It is difficult to define an accurate sea ice albedo; a systematic underestimation of the albedo could cause the optical depth of clouds to be overestimated.

In the comparison of CTH and CER between sea ice and non-sea ice cases, the differences for AP are small (0.2 µm for CER and -0.02km for CTH). The MYD and AC products also show small positive differences in CER (≈1µm) over sea ice. For non sea ice cases, the CTHs are 0.09 km (MYD) and 0.6 km (AC) lower than sea ice cases. Although CER and CTH exhibit positive differences, these were much smaller than the differences observed for COD.

The seasonality of the COD, CER, and CTH anomalies over sea ice was also examined. We observed that the disagreement between retrievals over sea ice and non sea ice is higher in JJA, where sea ice cases show (not shown here) lower COD and CTH values. COD values remain below 50 for AC and MYD and below 80 for AP across all seasons, with generally lower values over sea ice in winter (JJA). CER distributions are relatively consistent across seasons, ranging from 0–35 µm for AP, 5–65 µm for MYD, and 5–45 µm for AC. Seasonal variation is more evident in CTH: low clouds (less than 2 km) dominate for MYD and AP in JJA and for all three sensors in SON, while DJF and MAM show a greater presence of mid-level clouds (3–6 km), though low clouds remain prominent for MYD and AP. It can be concluded that in the case of CTH and CER, the seasonality is not an important factor, while for COD, the seasonality and presence of sea ice are important factors in the retrieval of the product.