This study investigates the regional and seasonal variability of cloud occurrence, located below 3000 m above ground level (a.g.l.), in the Arctic (60–82° N) between 2007 and 2016 using the DARDAR-MASK v2.23 product . In this satellite product, the vertical profiles of the 532 nm attenuated backscatter coefficient measured by the CALIOP lidar onboard CALIPSO are merged with the 94 GHz (W-band) reflectivity profiles from the Cloud Profiling Radar (CPR) onboard CloudSat on the same resolution grid. Since the lidar signal is more sensitive to the cloud optical extinction and the radar reflectivity to the size of large cloud particles (ice crystals), the combination of both measurements is used to detect the hydrometeor type (or phase) within a cloud layer. Therefore, the DARDAR-MASK v2.23 product provides a vertically resolved classification of the cloud hydrometeor type and aerosols (see Fig. S1 in the Supplement) along the A-Train track, with each atmospheric pixel (60 m × 1.7 km) assigned to one of 18 classes . These classes include clear sky, aerosols, stratospheric features, and several categories of hydrometeors.

Warm liquid water pixels are identified by strong lidar attenuation (β532≥2×10-5 m⁻¹ sr⁻¹) at temperatures above 0 °C. Supercooled liquid water pixels are defined by subfreezing temperatures (−40 °C < T < 0 °C), along with strong lidar attenuation, and for a layer thickness lower than 360 m . Mixed-phase pixels are characterized by a strong lidar attenuation caused by supercooled liquid droplets combined with the presence of ice crystals detected by the radar reflectivity signal. All remaining cloudy pixels at below-freezing temperatures are classified as ice-only when the lidar backscatter at 532 nm is weak (β532<2×10-5 m⁻¹ sr⁻¹), indicating no strong lidar attenuation, or when a radar detection signal is measured CloudSat radar mask values > 30 were considered as cloud detection;; these conditions may occur simultaneously and are both indicative of ice-only clouds.

It should be noted that the lidar signal can be rapidly attenuated in clouds with an optical depth greater than 3 to 5, depending on their microphysical composition . If the signal is extinguished or fully attenuated, the underlying cloud pixel phase diagnosis relies only on the radar reflectivity. Backscattering signals from spaceborne cloud radars are known to be affected by surface returns such as ground clutter or false ground detection. This can result in the misclassification of hydrometeors in the lowest atmospheric layers . Previous studies have shown that the vertical extent of the CloudSat CPR blind zone varies greatly depending on the surface type. It is narrow over open ocean and sea ice (approximately 750 m) but can extend up to approximately 1.2 km over land . The DARDAR-MASK product includes dedicated classes to identify potential contamination by clutter (class −1: surface and subsurface), thus reducing classification biases close to the surface. In , the impact of radar clutter was reduced by setting a lower-altitude threshold of 500 m. This threshold was determined through comparisons with a limited set of ground-based observations. However, this fixed threshold does not adequately account for surface-dependent radar contamination, especially over land.

In this study, the lower limit of the analysed cloud layer is no longer fixed. Instead, the altitude of the lowest possible detectable cloud layer is individually determined for each CloudSat profile, based on the height of the radar clutter. This adaptive approach explicitly accounts for variations in surface elevation and terrain complexity, ensuring consistent cloud detection over the ocean, sea ice, and land. Consequently, the cloud occurrence results presented in this study correspond to clouds detected between the radar clutter height and 3000 m (a.g.l.). A dedicated statistical analysis of clutter height distributions categorized by surface type shows median clutter heights of approximately 660 m over open ocean and sea ice, and 830 m over land (see Fig. S2 and Table S1 in the Supplement).

DARDAR-MASK pixel classification issues were also reported by . Comparisons with co-located in situ observations in low-level Arctic clouds showed that the liquid or the ice phases were accurately diagnosed by DARDAR-MASK 90 % of the time in single-phase clouds. However, in situ measurements revealed that almost 50 % of the DARDAR-MASK “ice” pixels within boundary layer mixed-phase clouds were actually a mixture of supercooled water droplets and ice crystals. The main sources of disagreement were attributed to the full attenuation of the lidar signal by the upper cloud layers and the contamination of near-surface pixels by radar ground echo.

To mitigate these limitations, we used the DARDAR classification program to estimate the occurrence of five distinct types of clouds. This program was originally developed for Arctic clouds by and later updated and extended by to study low-level clouds over the Southern Ocean, hence the name DARDAR-SOCP (DARDAR for Southern Ocean Cloud Phase). The algorithm analyzes the vertical structure of each atmospheric column to extract the cloud composition between clutter height and 3000 m (a.g.l.). The atmospheric column sampled is divided into “pixels” with a vertical resolution of 60 m, which are all assigned to a specific DARDAR-MASK class value. A cloud is defined as at least three vertically adjacent cloudy pixels (180 m), in accordance with in situ observations of low-level clouds . A single column may contain several distinct cloud layers, provided that they are separated by at least three clear-sky pixels. Accordingly, low-level clouds are first separated into two broad thermodynamic categories: warm clouds (T > 0 °C) and cold clouds (T ≤ 0 °C). Within the cold-cloud category, three phase-based subcategories are further distinguished: ice clouds, unglaciated supercooled liquid clouds (USLCs), and mixed-phase clouds (MPCs), which are described below:

Warm clouds: consisting exclusively of liquid water droplets at T > 0 °C.

Due to the CloudSat CPR's coarse vertical resolution, very weak ice precipitation, which is commonly observed in stratiform mixed-phase clouds, may remain undetected. Furthermore, without Doppler velocity measurements, the cloud radar cannot reliably distinguish between cloud ice and snow. Consequently, all radar-detected frozen hydrometeors fall into a single ice category within the DARDAR classification framework. Additionally, the lowest observable altitude is determined by the changing height of the radar clutter. Ice precipitation occurring entirely below this level cannot be detected, which could lead to an overestimation of the occurrence of USLCs. To assess this effect, we distinguish between USLCs that overlap the height of the clutter and those that do not. As the majority of identified USLCs (80 %) do not overlap with the clutter, the potential bias associated with undetected low-level ice precipitation is expected to remain limited (see Fig. S3 in the Supplement). Furthermore, although liquid-phase precipitation is not explicitly included in the classification, it is implicitly incorporated into the warm-cloud category. Drizzle can also occur in supercooled liquid containing clouds. Drizzle-sized liquid drops and falling ice crystals are expected to produce similar radar reflectivity signatures. Therefore, ambiguity in phase attribution is inherent when relying on spaceborne active sensors, particularly under conditions of lidar attenuation.

In DARDAR-SOCP, the requirement of at least three consecutive cloudy pixels to define a cloud can bias the detection of mixed-phase clouds, which often consist of successive single-phase layers. As a result, the occurrence of MPCs can be overestimated compared to the occurrence of ice and liquid clouds. We showed that these cloud-type classification issues lead to systematic biases in cloud occurrence of about 15 % for ice, 20 % for liquid, and 10 % for mixed-phase clouds see Supporting Information of. A potential underestimation of liquid phase may also occur as a result of lidar signal attenuation by an overlying liquid water layer located above 3000 m. The cloud occurrence for each type (OCn) is calculated as the ratio of the number of observations of that type (Nn) to the total number of satellite footprints (Nfootprint), expressed as a percentage: 1OCn=NnNfootprint×100

Cloud occurrences are computed for each cloud type on the basis of 1 d satellite data averages, ensuring statistical robustness at a regional scale. All footprints located between 60 and 82° N (14 overpasses per day in the Arctic region) are processed and aggregated into 1° × 1° latitude–longitude boxes. The years 2011 and 2012 are excluded due to an anomaly in the CloudSat battery, which caused data loss. Before 2010, both daytime and nighttime observations were used, while after 2012, only daytime observations were considered in our study because of the same battery-related issues. Relying only on CloudSat's daylight operations mode (2013–2016) led to fewer observations, particularly during winter and at high latitudes . This limitation introduces a seasonal bias in cloud climatologies, with a slight overestimation of winter occurrences (≈5 %), especially over oceans . More details are provided in the Supplement (see Figs. S4, S5, and Sect. S1).

As explained in the previous section, it is well known that CloudSat observations and thus the DARDAR product may be affected by a blind zone near the surface due to radar ground echoes. As previous studies based on ground and airborne observations have highlighted the prevalence of Arctic clouds near the surface, it is important to evaluate whether and to what extent the DARDAR cloud occurrences presented here are underestimated. To this end, an independent evaluation of DARDAR cloud occurrences has been conducted, based on comparisons with ground-based data from the Cloudnet network. Cloudnet observations based on multi-remote sensing instruments are used to produce a “classification product” that includes different cloud types for details. This product is comparable to the DARDAR product used in this study. We use Cloudnet classification products from the Ny-Ålesund (Norway, 78°55^′ N, 11°56^′ E) and the Hyytiälä (Finland, 61°84^′ N, 24°28^′ E) Arctic sites, as they are the only ones in the Arctic that were available for a significant period (several months) when CALIPSO and CloudSat were operational (June to December 2016 at Ny-Ålesund and March to August 2014 at Hyytiälä). To ensure representative occurrences comparable to ground-based observations, DARDAR cloud occurrences have been calculated in a square of 5° of latitude by 5° of longitude (±2.5° in both latitude and longitude), centered over each ground site location.

The results for the Ny-Ålesund site are shown in Fig.  (Results for the Hyytiälä site are provided in Fig. S6 in the Supplement). In this figure, occurrences of total (Fig. a), cold (Fig. b), warm (Fig. c), ice (Fig. d), mixed-phase (Fig. e) and cold liquid-containing (Fig. f) clouds in the altitude range up to 3000 m above ground level are presented. Weekly (blue colors) and monthly (green colors) occurrences have been computed. Ground-based occurrences are displayed in thick straight lines. We note that the last cloud type (CLCC for cold liquid-containing clouds) was only computed for this evaluation. The supercooled liquid class is not included in the Cloudnet classification product. Consequently, the USLCs class could not be evaluated directly. However, to be coherent with the Cloudnet product, cold liquid-containing clouds, including USLCs and MPCs cloud types, have been computed from DARDAR data.

Generally, a good level of consistency is observed when comparing ground-based and space-based occurrences. The weekly and monthly variability are very similar for both observation sites. However, differences are observed in the values of cloud occurrences, particularly those averaged over a week. These differences may be due to the blind zone, which could affect the satellite products. This could explain the higher frequency of cloud occurrences retrieved from ground-based instruments. However, these discrepancies may also be due to the effect of averaging occurrences within a 5 × 5° box. To quantify the differences, the average difference was calculated (see Table S2 in the Supplement). The average differences between ground-based and spaceborne cloud occurrences remain very small at only a few percent and are evenly distributed around zero (ranging from -5.8 % to +5.2 %, as can be seen in Fig. ). This indicates that there is no bias, meaning that occurrences are neither under- or overestimated. The standard deviations show non-negligible values (up to 23 %). Therefore, the main source of uncertainty is probably the spatial representativeness of the data (i.e. the comparison of a 5 × 5° box to a single location) rather than the blind zone. Thus, the general good agreement between the two datasets with no systematic bias leads to the conclusion that the impact of ground clutter on the low-altitude cloud occurrences from DARDAR presented in the paper remains limited.

Thermodynamic data are used to analyze the impact of environmental conditions on cloud occurrence and type. DARDAR-SOCP retrieves thermodynamic variables from ECMWF analyzes, which are integrated into DARDAR-MASK files as part of the ECMWF-AUX collocated products . ECMWF-AUX is an intermediate product in which the 3 h thermodynamic property forecasts of the ECMWF model on a 0.5° × 0.5° grid are interpolated to each CPR radar profile . The set of parameters used in this study is summarized in Table . In addition, three metrics are calculated to characterize the vertical structure of the lower troposphere and the advection of heat and moisture at a regional scale:

Lower-Tropospheric Stability LTS; is defined as the potential temperature difference between the free troposphere and the surface:2LTS=θ700hPa-θsurfLTS provides a bulk measure of static stability across the lower troposphere and has been widely used as a large-scale predictor of low-level cloudiness . Under Arctic conditions, where low-level clouds often reside well below 700 hPa, LTS primarily reflects the thermodynamic contrast between the cloudy boundary layer and the overlying free troposphere, rather than the stability within the cloud layer itself, and low LTS values should not be interpreted as convective instability.

In this study, we also account for the type of surface underlying the cloud: land, open ocean, or sea ice. In particular, sea ice concentration (SIC) is used to characterize sea and ocean surface conditions. SIC data are obtained from the AMSR-E instrument on board AQUA (2002–2011) and AMSR2 on board GCOM-W1 (2012–present). Both are passive microwave radiometers that measure brightness temperatures in several channels (18–89 GHz). SIC is retrieved with the ARTIST Sea Ice (ASI) algorithm , which has been validated against other retrieval methods and in situ data . The typical uncertainty for 100 % sea ice cover is 5.7 % , with higher uncertainties in the marginal ice zone due to mixed ocean/ice pixels and atmospheric contamination.

The selection of this limited set of environmental factors is consistent with previous studies carried out in mid-latitudes and in the Arctic . The environmental parameters are averaged over 1 d and collocated on a 1° × 1° grid to match the temporal and spatial resolution used for cloud-type occurrences. The environmental parameters are partitioned in the same way as cloud-type occurrences to ensure temporal consistency. This timescale is thus appropriate both to obtain representative cloud-type occurrences and to investigate the variability of environmental parameters at the regional scale.

In this study, we implement multiple linear regression (MLR) analyzes to identify the environmental factors that contribute the most to the regional and seasonal variability of cloud-type occurrences. This statistical approach is used to predict the variation of a dependent variable (e.g., cloud occurrence) as a linear combination of several explanatory variables (e.g., environmental parameters). It relies on a least-squares estimator to minimize the error between predicted and observed values . In this study, we apply MLR to relate several environmental predictors to the target variable (OCn) as follows: 5OCn=β0+β1x1+…+βixi+ϵ where β0 represents the intercept, βi the regression coefficients associated with each predictor xi, and ϵ the residual. MLR variables (xi) are normalized to ensure a direct comparison of the relative importance of each parameter . We follow the same MLR methodology as in , ensuring consistency with recent approaches applied to Arctic cloud studies. Before each regression, we evaluated the multicollinearity between predictors and explanatory variables based on the Variance Inflation Factor (VIF, see Fig. S7a in the Supplement). According to , variables with a VIF greater than 10 are excluded from multiple linear regression analysis. Non-significant predictors, either showing a p-value greater than 0.05 or exhibiting no clear influence (e.g., wind; see Fig. S7b in the Supplement) in the simple correlations with cloud-type occurrences (R2<0.1), were excluded to retain only robust explanatory factors for each cloud type. The relation between the cloud occurrence and the environmental parameters is expected to be nonlinear, leading to a relatively low R2, but it remains satisfactory at the first order with R2 greater than 0.3. To assess potential dependencies between predictors, we also computed pairwise correlations using Spearman's rank method. The correlation matrix shown in Fig.  reveals strong links between several environmental variables that are inherently interdependent and are often correlated to varying degrees. This is especially the case for the temperature variables (RSurf_Temp/Temp_850hPa2=0.88; Fig. ) and the humidity variables at different atmospheric levels, as well as for EIS and LTS (REIS/LTS2=0.91; Fig. ).

Accordingly, to minimize the collinearity and redundancy between the different predictors, eight variables are selected for the MLR analyzes: surface temperature (Surf_Temp), surface pressure (Surf_Pres), integrated water vapor (IWV), lower-tropospheric stability (LTS), marine cold air outbreak (MCAO), specific humidity at 700 hPa (SH_700hPa), geopotential height at 850 hPa (GH_850hPa), and sea ice concentration (SIC). For MLR, LTS was preferred over EIS because it exhibits weaker correlations (than EIS, Fig. ) with the other predictors, thereby reducing multicollinearity within the MLR framework. Some thermodynamic parameters cannot be computed over all regions (e.g., LTS over high-elevation areas such as central Greenland); grid cells with missing parameter values are therefore excluded from the MLR analysis. These parameters characterize the basic properties of the lower troposphere thermodynamic structure and the surface conditions. We expect this limited set of parameters to be sufficient to statistically investigate how varying environmental conditions impact the regional distribution of low-level clouds.

During the eight-year study period, the observational dataset used in the MLR analysis consists of approximately 2 450 300 data points distributed across 1° × 1° grid cells. However, some low-latitude grid cells can sometimes contain a limited number of observations. This makes the calculation of cloud occurrence frequencies and associated statistics less robust. To ensure adequate sampling, a minimum sampling threshold for the number of observations per grid cell is defined based on the first quartile (Q1) of the distribution of daily observations in the lowest latitude band (60–61° N), where sampling is weakest. We find a Q1 value equal to 12 observations per day. Therefore, only 1° × 1° grid cells containing at least 12 DARDAR pixels per day are retained in the cloud occurrence and MLR analyses.

The remaining dataset is randomly divided into two subsets. 80 % of the data are used to train the MLR models, and the remaining 20 % serve as an independent test set for validation. The performance of the regression model is quantified using the coefficient of determination (R2). In addition, the statistical significance of each regression coefficient is assessed using a Student's t-test, and the corresponding p-values are reported for all predictors. These metrics provide a measure of the precision and robustness of the regressions.

The cloud–surface coupling state is determined from the thermodynamic structure of individual cloud profiles. Building on the work of , who introduced a simplified version of the coupling algorithm originally proposed by , we analyze the vertical profile of the potential temperature (θ) from the cloud liquid layer base to the surface. The cumulative mean of θ is computed for each individual cloud profile over this layer and compared to the local θ profile. A cloud profile is classified as decoupled from the surface if the difference between the mean θ and the instantaneous θ exceeds 0.5 K at any level. Cloud profiles that do not exceed this threshold are considered to be related to surface-coupled clouds. For ice-only clouds, the procedure is carried out using the base of the ice cloud rather than the base of the liquid layer. The coupling state is therefore diagnosed profile-by-profile and is subsequently aggregated in time and space for the statistical analyses presented in this study, consistent with the methodology applied in previous observational studies e.g.,.

This section examines the distribution of cloud occurrence (OC) in different regions of the Arctic. On average, the analysis of the DARDAR-MASK products shows that clouds are present 65 % of the time throughout the atmospheric column between the clutter height and 12 km. Figure  shows the geographical distribution of the occurrence of low-level clouds in the Arctic for different cloud types: total clouds, cold clouds, warm clouds, ice clouds, USLCs and MPCs. Across the Pan-Arctic (defined as the area between 60 and 82° N, encompassing all surface types), the median annual occurrence of low-level clouds (OCall-ll) at altitudes ranging from the clutter height and 3 km (a.g.l.) is 47.7 % (Fig. a). We show that the cloud distribution is highly inhomogeneous across the different regions of the Arctic. The vast majority of these low-level clouds occur in thermodynamic conditions where cloud temperatures are below 0 °C. These cold low-level clouds are present more than 46 % of the time (OCcold-ll; see Fig. b). Their occurrence progressively decreases toward lower latitudes, with the most pronounced reduction observed over oceanic regions. In contrast, warm liquid clouds are comparatively rare in the Arctic. The median occurrence of these clouds (OCwarm) generally remains below 1.5 % (Fig. c). The average distribution of fully glaciated cloud occurrences (OCice) shows geographical contrasts. OCice median value is 21 % over the Arctic but can reach 30 %–35 % in the continental and mountainous regions of Siberia, Alaska, and Greenland (Fig. d). Figure e shows that clouds composed exclusively of supercooled water droplets (USCLs) are less common than ice clouds, with a median occurrence of around 8 % (OCUSLCs). Finally, mixed-phase clouds, which are characterized by a mixture of water droplets and ice crystals, are observed on average 17 % of the time over the Arctic (Fig. f). The geographical variability of these two types of liquid-containing clouds is also pronounced. They prevail over regions influenced by the North Atlantic Ocean and the Bering Sea (OCMPCs > 30 %). However, it is important to note that the respective contribution of MPCs and USLCs to the liquid-containing cloud population may potentially be subject to uncertainties. When the cloud base of a USLCs is located within the radar clutter zone, its phase could be misdiagnosed. The lowest layers of these near-surface USLCs may consist of ice crystals fully masked by the radar clutter. This suggests that clouds initially categorised as USLCs could actually be MPCs. Removing these cloud cases from our analysis decreases the median USLCs occurrence from 8.1 % to 6.5 % (gray values shown in Fig. e). Consequently, the median occurrence of MPCs would increase from 17.2 % to 18.8 % (Fig. f). These results show that the potential underestimation of the ice phase in cloud layers impacted by the radar clutter remains a moderate but non negligible issue. Most of the USLCs cloud bases are located above the radar clutter (80 %). Therefore, the values of 6.5 % and 18.8 % should be considered as the lower and higher bounds of the median occurrence of USLCs and MPCs, respectively.

Based on the spatial distribution of the cloud-type occurrence, on the main weather patterns (high- and low-pressure systems) and geographical limits (latitudes, oceans, continents), we choose to subdivide the Arctic into 12 distinct subregions (Fig. a). The six subregions in Zone 1 have latitudes ranging from 60 to 70° N, whereas the northernmost regions are located within the high-latitude ring, which extends from 70 to 82° N (Zone 2). Subregions 1B and 2B, as well as 1C and 2C, are merged to increase the statistical significance of the results (1-2B and 1-2C). The purpose of this regionalization is to highlight the specific features of the low-level cloud occurrence in the Arctic. A statistical analysis of total cloud occurrences for each region is presented in Fig. S8 and discussed in Sect. S2 of the Supplement.

The northeast Atlantic sector (region 1A) is the area with the highest occurrence of warm clouds (≈ 5 %–10 %). MPCs are also frequent in this region, as well as over the seas surrounding the Svalbard Archipelago (region 2A), where they occur up to 35 % of the time. These two regions are influenced by the Gulf Stream and low-pressure systems. They are characterized by the highest low-level cloud occurrence (OCall-ll > 60 %). A specific feature of region 2A is the prevalence of MPCs and USCLs over the Greenland, Barents, and Norwegian Seas (OCUSLCs with peaks around 20 %). Low-level ice clouds are more prevalent in mountainous regions of Alaska, Greenland, and western Canada (1E) (OCice > 25 %) than in eastern provinces of Canada (1D) and, to a lesser extent, than in the Labrador Sea and Baffin Bay (1-2C). In these latter regions and parts of eastern Europe (1H), USLCs occur more frequently than in continental areas with relief of western Canada (1E) and central and eastern Russia (1G), for which OCUSLCs is lower than 10 %. The cloud-type cover over the Bering Sea sector (1F) differs from that in neighboring regions. The frequency of occurrence of low-level clouds and of liquid-containing clouds in particular is much higher (OCMPCs≈25 % and OCUSLCs > 10 %) than in surrounding regions subject to high-pressure systems. The northernmost eastern sector of the Arctic (2F) is more cloudy (OCall-ll≈50 %) than the Beaufort and Chukchi Seas (2E) and the northern Nunavut region of Canada (2D). Region 2E experiences a 5 % higher occurrence of MPCs than region 2D. In summary, these first results show that the distribution of low-level ice clouds and liquid-containing clouds is characterized by strong regional contrasts partly influenced by the latitude-driven air temperature and the type of surface (ocean, sea ice, continental relief).

In this section, we investigate the seasonal variations of low-level cloudiness (between the clutter height and 3 km a.g.l.) for seven contrasting subregions of the Arctic. The results are presented in Fig. . It shows that in most regions, the annual cycle of total low-level cloud occurrence exhibits two distinct seasonal maxima. The first peak occurs in late spring, when low-level clouds are present 45 % of the time in the Arctic (Fig. b, dashed black line). Previous studies have highlighted that the first cloud maximum is related to changes in the regional atmospheric circulation resulting from the weakening of the tropospheric polar vortex at the end of winter . The second peak has a larger amplitude, reaching 55 % in autumn. However, the magnitude and timing of this double seasonal peak of high cloudiness exhibit regional variability. Most of the regions north of 70° N (2D, 2E, and 2F) are characterized by well-marked peaks with above average low-level cloud occurrences in May and in October or November. This feature is particularly striking in the eastern Arctic region (2F), where low-level cloud occurrence reaches almost 45 % in May and 68 % in October. Over the Bering Sea (1F), the spring and autumn local maxima are also clearly defined but more likely occur in April and November. This pronounced seasonal variation is less evident in regions influenced by the Atlantic Ocean (1A and 2A). Typically, higher and steadier occurrences are observed throughout the year (from 48 % to 65 % in region 1A). The second peak of cloud occurrence observed in the autumn can be attributed to changes in surface conditions. In the northern regions (2E and 2F), sea ice retreat peaks in September. The open ocean facilitates the vertical transfer of moisture to the atmosphere, which promotes the development of low-level clouds until the return of the ice pack .

The seasonal cycle of warm liquid clouds and ice clouds does not exhibit the two seasonal maxima observed for total low-level clouds. As expected, the occurrence of low-level warm clouds is characterized by a single seasonal maximum, occurring in summer (Fig. c). The pan-Arctic median OCwarm reaches 5 % in July. These clouds are mainly observed in low-latitude Arctic regions (1A, 1D, and 1F), where a maximum frequency of occurrence of 11 % is recorded over the Northeast Atlantic region (1A) in July. Region 1A is also the only sector to experience warm cloud cover from mid-autumn to late spring. The monthly variations of low-level ice clouds are opposite to those of warm clouds. The occurrence of ice clouds is at its maximum in all regions from mid-autumn to late winter (Fig. d). During this period, OCice ranges from 20 % to 30 %. In the southern Arctic (1A, 1D, and 1F), the amount of ice clouds decreases considerably in spring, dropping to a minimum value of below 10 % in summer. However, this decrease is less steep over the northern seas of the western (2D and 2E) and eastern (2F) Arctic, where the persistence of low-level ice clouds is favored by the colder atmospheric conditions. In these regions, OCice is typically greater than 10 % even during the summer months.

At the scale of the Arctic, the temporal evolution of the frequency of occurrence of USLCs (OCUSLCs) and MPCs (OCMPCs) is similar to that of total clouds. Figure e and f show a seasonal cycle featuring two peaks. The late spring peak is just over 10 % for USLCs and over 20 % for MPCs. The second peak occurs in September for USLCs and in October for MPCs, reaching more than 10 % and 25 %, respectively. The secondary seasonal peak is also broader for USLCs than for MPCs. Significant differences are observed when these seasonal patterns are investigated at a regional scale. In autumn, when the second peak occurs, OCMPCs reaches 35 % over the seas north of Siberia (2F). Differences in OCMPCs of almost 15 % are also observed in spring between the Northeastern Atlantic region (1A) and region 1F. Moreover, regions 2F and 2E shift from being amongst the least covered by MPCs in summer (less than 20 % and 15 %, respectively) to becoming the regions with the highest MPCs occurrence in mid-fall (OCMPCs is 30 % to 35 %). In these regions, the opposite trend is observed for USLCs, with relatively high and steady values of OCUSLCs observed in summer, followed by a sharp decrease in autumn. Regions influenced by the transport of heat and moisture from the Atlantic Ocean (1A and 2A) exhibit the highest occurrences of MPCs during winter and spring (up to 25 % in region 1A). In contrast, the western and eastern Arctic regions (1F, 2D, 2E, and 2F) have lower than average MPCs occurrence values (8 %–10 %). These regional differences also exist for USLCs but remain less pronounced. Interestingly, the significant reduction of OCUSLCs observed at the beginning of spring (March) in the marine regions 1A, 2A, and 1F is balanced by an increase in ice clouds and a more moderate one in MPCs. From a broader perspective, the regional variations in the seasonal occurrence of these liquid-containing clouds cannot be explained solely by latitudinal temperature gradients and surface conditions. The frequency of occurrence of these clouds appears to be controlled by several environmental factors. Furthermore, our results show that USLCs and MPCs have a contrasted seasonal distribution at a regional scale, supporting the relevance of studying these clouds separately.