To detect changes on daily, monthly, seasonal, and decadal scales, several measurements per day at an adequate spatial resolution must be made over several decades. The polar-orbiting spectrometer suite comprising GOME, SCIAMACHY, and GOME-2 (Table  for their specifications) makes measurements of RλTOA at the same solar zenith angle and at several times per day as a result of the instruments' swath widths. They are a suitable choice, given the length and coverage of their records and their high spectral resolution, for the creation of the RλTOA time series. Description of GOME can be found in , while SCIAMACHY and GOME-2 are, respectively, described in and . The detailed steps to harmonize RλTOA measured by sensors of different technical specifications are given in Appendix .

While the measurement of solar radiation scattered back to the TOA by GOME, SCIAMACHY, or GOME-2 takes place only during daylight, radiation in the thermal infrared (λ≳ 4 µm), required to record the thermal emission from the surface and the atmosphere, is not measured by these sensors. Because of the different sensors' swath widths, the RλTOA measurements in the solar spectral range have a northern latitude boundary (or terminator). This boundary is illustrated by plotting the pan-Arctic annual cycle of RλTOA in Fig. . At the three wavelengths 510, 560, and 760 nm, the seasonality shows that summer months have lower RλTOA and RλTOA is higher otherwise. This darkening of the Arctic can also be seen by comparing the years at the beginning of the recording from 1996 with the most recent ones. However, this behavior occurs only between April and September. These are the months when the individual terminator of the three sensors reaches the latitude 85∘ N, this being the spatial threshold of common spatial coverage we set in the monthly average. As shown in Fig. , the other months (October to March inclusive) show that recent years are brighter (higher RλTOA) than those at the beginning of the time series. This is because the individual terminators move further south (Fig. c) and the coverage is considered insufficient for this to be studied further.

From Fig.  we identify two distinct behaviors of RλTOA. The first is a period of steepest decrease, from April to June, and the second is a plateau of relatively flat RλTOA, between July and September. The changes in surface reflectance between April and May are attributed to snow cover changes and those in June to sea ice changes . Over water, the timing of such transitions increasingly approaches the summer solstice, which is the day of strongest solar insolation, while it moves further away from it over land . It is therefore reasonable to regard this day as a demarcation point between Arctic spring and summer.

In summary, we group April–May–June (AMJ) as Arctic spring and July–August–September (JAS) as Arctic summer. This distinction is explained by the sensors' measurement strategies and by the time-dependent physical processes leading to the transition from high to low Arctic reflectance in June to the minimum sea ice extent in September. We note that the definition of seasons is arbitrary and is determined by the breakpoints of the variable under consideration. In general, seasons can be astronomical, meteorological, or climatological. Ignoring the astronomical definition, the meteorological seasons are not suitable for our purposes because in May and June (respectively, the last month of meteorological spring and the first of summer) multiple scattering between the surface and the atmosphere still prevails, thus coupling both radiatively. The definition of ad hoc Arctic seasons ensures that the computed trends describe only those changes of RλTOA caused by distinct underlying processes, which in turn determine the breakpoints in the time series of RλTOA shown in Fig. .

In our study, the RλTOA data are complemented by a record of cloud properties and broadband fluxes at TOA and the surface (see Sect. ). These are inferred from the afternoon orbit (PM) of AVHRR sensors on board the POES missions. Despite the availability of the morning orbit (AM) AVHRR series, we found that only the AVHRR PM series fulfilled the calibration stability requirements which allow trend assessment to be made. Inspection of the time series of cloud properties and fluxes for the AM series shows that the drifts in local overpass time of the NOAA-12 platform before 2003 led to calibration offsets and that the scan motor errors of the NOAA-15 platform led to data gaps .

One good reason for choosing this AVHRR record is the number of studies using these data in the Arctic. Our choice is driven by the maturity of the AVHRR data set of measurements; by its popularity; and by its successful use by the advanced, most recent, retrieval algorithm exploiting it. This AVHRR data set is in its third reprocessing, and the algorithm used to generate it has 15 years of development starting with ATSR-2 on board ERS-2. While improvements and validation have been documented in traceable documents (https://climate.esa.int/en/projects/cloud/key-documents/, last access: 23 February 2022), the cloud and flux records are presented by and references therein. Recently, compared CFC, COT, LWP, and cloud top height (CTH) of this record with colocated measurements from four high-latitude stations across the Arctic and found no scale biases for the large majority of satellite-derived cloud products, except for the site located at Summit, Greenland (72.59∘ N, 38.42∘ W).

Some features that distinguish this data record from older AVHRR records are as follows: (1) the channels in the solar spectral range have been cross-calibrated with SCIAMACHY channels. SCIAMACHY is recognized for its accurate radiometric and spectral calibration. Because the part of our study dealing with RλTOA is conceived in a way that the record is radiometrically coherent with SCIAMACHY (see Appendix ), this intra-band correction relates reflectance changes at visible wavelengths detected by SCIAMACHY to those by AVHRR, ingested in the cloud retrieval algorithm, which calculates τ and cloud albedo. (2) The cloud mask uses a neural network, trained on CALIOP data to take into account the extent of the underlying bright Arctic surface. (3) CTH has been calibrated using CALIOP profiles to account for the penetration depth of radiation inside a cloud. This is needed because the retrievals of CTH from all infrared thermal channels are influenced by this effect and yield a radiative cloud top height lower than the physical cloud top

In this AVHRR satellite record, the cloud phase can be only liquid or ice. The input signal for AVHRR comes from the reflectances measured at 0.6, 0.8, and 3.7 µm. Given the different complex refractive indices of the water and ice phase across the SWIR wavelengths, the method is effective in separating the two phases. It is worth noting that Arctic cloud tops are predominantly in the liquid phase, whereas the mixed phase occurs in the middle of the clouds. This is the outcome of four airborne measurement campaigns, totaling 18 flights, reported in . Nonetheless, the mixed phase is not identified, despite its all-season occurrence and role in the Arctic climate . This is because the data set is a neural network trained on the CALIOP cloud phase, which does not natively provide information on the mixed phase in clouds.

The application of the cloud algorithm to MODIS measurements, which take place in the same wavelengths as the AVHRR channels, has shown that the retrieval scheme is well aligned with the reference standards of CloudSat and CALIPSO data for CFC, CTH, τ, and the liquid thermodynamic phase. While agreeing on the sorting of cloud tops between water and ice phases, higher variability for IWP values lower than 50 g m-2 is found as compared to that in the reference DARDAR cloud data products , but IWP histograms across the full range do not substantially differ . Version 3 has improved version 2 in terms of precision, accuracy, and stability . Even more relevant to our purpose is the scheme adopted to calculate broadband fluxes with the cloud properties described above.

The broadband fluxes in the solar and IR spectral regions are computed by solving the radiative transfer combining the two-stream approximation by for the bulk bidirectional reflectance, transmission, and source terms within a plane-parallel atmospheric slab and the spectral band model by for gaseous absorption. Six bands in the SW and 12 bands in the LW are calculated sequentially, ingesting local properties of clouds retrieved with a Bayesian technique , which provides estimates of the individual uncertainty at pixel level.

Specifically, effective radius and cloud optical thickness are the primary inputs for flux calculations together with solar zenith angle and ancillary data from MODIS climatologies of visible and near-infrared surface albedo, linearly interpolated to each spectral band center. Local vertical atmospheric profiles from ERA-Interim account for the p–T variations, while a constant aerosol optical depth of 0.05 and concentrations of well-mixed gases are assumed, the latter being linearly interpolated for their time-dependent increase. The combination of the above factors yields an accuracy of ± 3 W m-2 in outgoing LW radiation (OLR) when compared with observations by the broadband radiometer GERB on board the MSG-2 platform . This value is in line with the radiometric accuracy of GERB, which is 1 % for clear-sky fluxes at TOA and with biases of 4–5 W m-2 against CERES observations, when a similar algorithm for the derivation of broadband fluxes is applied to CloudSat, CALIPSO, and MODIS measurements . Specifically for the Arctic and for latitudes north of 66∘ N, winter OLR from AVHRR has been compared to the multi-model CMIP6 average . Based on their data we calculate an all-sky average bias of ± 3.59 W m-2.

The physical boundaries of clouds are additionally required to correctly compute scattering and absorption along the vertical. From the retrieved CTH and effective radius, the bottom cloud layer is calculated assuming a subadiabatic variation of cloud water path, separately for the liquid and ice phases. While this approach is appropriate for the shallow case , the thickness of deeper clouds is computed by combining a variable increase of water content matching within-cloud temperature profiles. The nominal accuracy limit, in this case, is reached at temperatures of less than 217 K (-56 ∘C), which exceeds the yearly climatological range for the Arctic (-25 ∘C February, +2.5 ∘C July; ), and AVHRR-derived cloud bottom height is found to be in good agreement within ± 369 m against ceilometer observations .

Radiative transfer is solved twice. First, all-sky fluxes are calculated with retrieved cloud properties and then the clear-sky fluxes, assuming that the pixel is devoid of clouds. This approach is in contrast to that employed with the MODIS cloud record and the CERES-EBAF radiation measurements at TOA, by which the interpolation of the measured clear-sky pixels serves as gap filling of all-sky pixels for the monthly aggregation of fluxes at BOA . AVHRR-derived fluxes at BOA have been validated by comparison with BSRN stations and the CERES-EBAF product .

Given the standard notation (all: all-sky, clr: clear-sky, +: upwelling, and -: downwelling fluxes), average comparisons with independent data show a good agreement for all downward fluxes and LW+. The average long-term relative bias of AVHRR-derived fluxes against CERES ranges from +2.9 % for SWall+ to -2.7 % for LWclr+. Validation with BSRN measurements in the period 2003–2016 shows that the bias (correlation) for SW+/- is in the range [-6.16, +1.99] W m-2 (0.93/0.99) and [-3.02, +7.60] W m-2 (0.99/0.99) for LW+/-. The average AVHRR-based estimates tend to be biased high at ≈ 20 W m-2 for SW+ < 100 W m-2, while the opposite holds for SW+ > 250 W m-2 with an average underestimation of up to -50 W m-2. In both ranges, the average relative bias amounts to ≈ 20 % . This bias of higher spread can be due to the surface heterogeneity around the validation site, which influences the comparison of SW+ because of the difference in spatial scales between the satellite footprint and the BSRN effective point measurement.

The surface treatment in the satellite record is also a potential source of error because SW+ is equal to SW- times the surface albedo. While the actual sea ice extent is taken from measurements in the microwave , a fixed value of spectral surface albedo is assumed throughout the record. The albedo of snow- and ice-covered surfaces is set to 0.958 at wavelength 630 nm, 0.868 (910 nm), 0.0364 (1.6 µm), and 0.0 (3.74 µm). The albedo is additionally area-weighted for fractional sea ice or snow cover scenes . Consequently, intra-annual variability and long-term changes in surface reflectivity are not accounted for. This would lead to an underestimate of actual surface albedos in those months having fresh snow and ice (spring) and overestimating during months of melting surface upper layers (summer). In the case of underestimation of surface albedo (or sea ice extent), we expect an overestimation of CRF and thus warming by the clouds and vice versa.

We do not expect differences in BOA fluxes as a function of solar zenith angles because the instantaneous fluxes are corrected for the diurnal cycle of solar illumination by adjusting the surface albedo and the atmospheric path lengths. The LW fluxes have been also corrected by using a cosine function derived from measurements of the geostationary SEVIRI sensor. The final aggregation is a good approximation to a true 24 h average , needed to determine the true climatological mean of SW and LW fluxes and thus CRF. Consequently, also the seasonal averages (i.e., AMJ and JAS) are not expected to exhibit variations induced by solar zenith angle and directionality of surface reflection.

Misclassified cloudy scenes especially over dynamically bright surfaces (i.e., marginal and fractional sea ice zones) impact the calculation of broadband fluxes. This has been already noted in the first studies comparing ERBE and AVHRR cloud radiative forcing derived with different scene classification schemes . The conversion of directional radiance, measured at TOA, to irradiance requires the knowledge of the angular light redistribution function of the surface and atmospheric components. If this conversion is not accurately performed, the irradiance and SWclr+/- above reflecting surfaces cannot optimally be calculated. Using the same data as that of our study, a low sensitivity of trends in cloud radiative forcing to the biases in cloud properties over surfaces of changing brightness has been found Appendix D inp. 7499.

Specifically, assessed possible uncertainties in CRF trends by analyzing CFC biases as a function of sea ice concentrations (SICs) for the seasons defined in Sect. . For season AMJ, the bias is systematically flat from SIC 0 % to SIC 100 %. Given that our trend model is based on anomalies and not absolute values (see Appendix ), any additive component of the bias cancels out and the resulting trend is not affected by it. For season JAS, the bias is not flat and a multiplicative bias in CFC can propagate to CRF via SIC changes. However, the SIC bins of Fig. A1 can also be regarded as the SIC variance over one location in time; therefore this effect is relevant only for those locations with a large dynamic in SIC (e.g., the marginal sea ice zone). If the SIC anomalies over one location in the marginal sea ice zone are not equally distributed about zero, irrespective of any trend, but progressively change over time, their distribution is not Gaussian but skewed. This leads to the addition of the time-dependent component in the CRF trend via CFC. Looking at Fig. 8, the SIC anomalies for the marginal sea ice zone of the enlarged Chukchi Sea are normally distributed. Upon regression, any possible residual of a non-normal SIC distribution, reflected in CFC and propagating into CRF, would still be captured by the trend model (see Appendix ) which accounts for the length of the effective independent sample in the record.

The RλTOA time series, measured by GOME, SCIAMACHY, and GOME-2A over the Arctic region (60–85∘ N); anomalies; trends; and significance were harmonized (for more details see Appendix  and ). They are shown for wavelengths 510, 560, and 620 nm in Figs.  and . The RλTOA values retrieved from the sensors MERIS (on Envisat) and GOME-2B (on MetOp-B) confirm that the correction scheme is successful for the spring (AMJ) and summer (JAS) months. The discrepancy between MERIS and SCIAMACHY in the fall and winter months, as long as sunlight is available, can be tracked to the different swath widths of the respective sensors. MERIS has a swath of 1150 km, whereas SCIAMACHY has a swath of 1000 km. This implies that with the onset of the polar night at high latitudes, the western part of the scan of both sensors (which are polar orbiters in descending node) will include increasingly dark Arctic areas, the MERIS scan being more northward leaning. Therefore, any averages of MERIS measurements will include more dark scenes than those in an average calculated from SCIAMACHY measurements. For this reason, the MERIS reflectances in the fall and winter months are generally lower than those by SCIAMACHY.

A consistent and consolidated data set results from the measurements of the three instruments. Seasonality is the dominant feature of Fig. . Maximum RλTOA occurs in early AMJ when the polar day results in the Arctic being fully illuminated and the ice extent is close to its maximum. Analogously, minimum RλTOA occurs from August to September when the days are shortening and sea ice coverage is at its minimum. The observed seasonal cycle of RλTOA agrees with that calculated by models as do the observations of sea ice extent over the Arctic . This provides evidence to confirm that one dominant parameter in RλTOA variability is surface reflectance .

Figure  shows that the standard deviation of RλTOA for GOME is smaller than the other sensors. GOME has a considerably coarser pixel size than the follow-on sensors (see Table ). This leads to different mean RλTOA and standard deviations because the integration time of acquiring onboard electronics for a coarser pixel is longer than for a finer pixel. This averages out sub-pixel heterogeneity differently. We account for this effect by assessing RλTOA trends not from mean values but from anomalies (see Appendix ) instead. The anomalies are customarily normalized with the standard deviation as a common technique for the analysis of records which might be heterogenous in scale, without changing the underlying sample distribution because standardization of anomalies is a linear transformation .

A negligibly small and statistically insignificant downward trend of RλTOA for the three wavelengths in the solar range is seen in the anomalies of Fig. . The anomaly of RλTOA is the difference between the value of RλTOA and the climatological average value of RλTOA at the given time of the year t (see Appendix ). In a warming Arctic, a substantial decrease in reflectance would have been expected due to sea ice loss.

In , a downward trend of all-sky albedo across the Arctic is reported. This is not compensated by an opposite trend in albedo as a result of increased cloudiness, which thus levels the recent pan-Arctic reflectance trend. However, their analysis is limited to oceanic regions (for open and ice-covered regions) and additional uncertainties are caused by the conversion from clear-sky to all-sky albedo at the beginning of their record. As the clear-sky signal is derived from the sea ice record with sensors for which the atmosphere is almost entirely transparent, the all-sky albedo is computed with a post hoc method adding the atmospheric part and is not the outcome of direct satellite measurements.

Moreover, state that no significant relationship between CFC patterns and sea ice loss is observed during summer, but some are identified in fall months . Such changes are not observable in the pan-Arctic RTOA anomalies. Rather, the reduction in reflectance is small and not attributable to a specific season. As a consequence, we need to ask whether the loss of reflectance associated with sea ice reduction is compensated by increasing CFC or brighter clouds, at the pan-Arctic and regional scale as well, and which processes lead to the small pan-Arctic RTOA trends.

To answer these questions in the following, we show the Arctic sea ice concentration (SIC) in 1996 and 2017 for AMJ and JAS in Fig.  and the RλTOA trends for the wavelengths 510, 560, and 620 nm in Fig. . The mean seasonal sea ice extent (SIE) at 15 % and 75 % SIC is, respectively, coded in orange and red contours. While SIE is usually identified by a SIC threshold of 15 %, a value of 75 % better represents the geographical contours identified using RλTOA. Moreover, identify the 75 % SIE threshold as the demarcation between two distinct regimes of accuracy in broadband fluxes, which depends on the misclassification of satellite-derived CFC above bright surfaces.

Similarly, Fig.  shows trends for the analyzed wavelengths for the 12 Arctic regions that are defined using the geographical subdivision proposed by and (see Fig. ). Trends for AMJ are shown in green, and the JAS trends for selected spectral bands are shown in blue. The red symbols show the absolute averages of the RλTOA values at the beginning of the record for the respective seasons.

There are marked regional differences (Fig. ). Those that are statistically significant (at 95 % confidence level) are shown with red crosses. For AMJ a significant negative trend over the Barents Sea is balanced by a positive RλTOA trend at all three wavelength bands over Greenland, the Canadian Archipelago, and the western Arctic seas, such that the pan-Arctic trend remains almost unchanged. In JAS, the negative trend shifts towards areas of the Kara, Laptev, and Chukchi seas. These are Arctic areas having open oceans and are experiencing significant sea ice loss during the period of study (Fig. ).

In general, the trends are negative and statistically significant in both seasons where sea ice retreats, such as in AMJ for the Barents Sea and the perennial sea ice zone around the North Pole. For the remaining areas that cannot be directly explained by the difference in sea ice extent, we assume patchy residual sea ice concentrations below 50 % closer to Eurasia and the occurrence of melt ponds on the sea ice pack. In both cases, open ocean areas and freshwater lower the albedo of the scene sensed by the satellites, as can be seen comparing the 15 % and 75 % SIC contours in Fig. . The areas that do not show statistical significance are generally above the perennial sea ice during AMJ. These months are characterized by a small standard deviation and by a non-existent SIC trend (not shown).

The spectral dependence of the trends in Fig.  differs as a function of the sign. The negative trends are spectrally neutral in both magnitude and statistical significance. On the contrary, the areas of positive trends like the belt from the Canadian Archipelago and Beaufort and Chukchi seas in AMJ and, to a smaller extent, Greenland in both seasons show an increase in values and spatial statistical significance from 510 to 620 nm.

While we cannot completely rule out the broadband influence of ozone trends (see Appendix ) on reflectances, the spectral patterns are coherent with an increase in some cloud properties conducive to snowfall and a brighter surface. Despite its proximity to the Canadian Archipelago, Baffin Bay has changes in RλTOA trends that would more closely match the eastern Arctic seas region. Over Hudson Bay, the RλTOA trends show unusual patterns. They are largely positive in JAS and relatively strongly negative in AMJ.

Although not of the same magnitude, almost all regions show a reflectance change at 760 nm. This wavelength is the only channel with a very strong gaseous absorption and is not in the broadband continuum like all other channels. The 760 nm wavelength bears more information on light scattering aloft than at the surface because of the strong columnar absorption of atmospheric oxygen largely extinguishing photons before they impinge on the ground. Oxygen absorption is modulated primarily by CTH and, to a lesser extent, by CFC and optical properties such as cloud albedo (CA) and τ. In this context, where a positive trend value of RλTOA at 760 nm is observed, greater than the other channels, we deduce a clear change in the occurrence of clouds or one of their physical or scattering properties. This is the case for Greenland during AMJ and JAS; for the Canadian Archipelago and the Barents, Chukchi, and East Siberian seas only in AMJ; and for the Barents Sea, Hudson Bay, the Atlantic corridor, and the Siberian continent only in JAS. Knowing that RTOA is influenced by scattering and absorption in the atmosphere and that the atmospheric RTOA can be additionally partitioned into the cloud, aerosol, and gas contributions, this prompted us to examine changes in those cloud properties which directly influence the spectral RTOA trends.

The globally validated and consolidated cloud record has first been analyzed across the Arctic (60–85∘ N). The top panel of Fig.  shows the time series of CFC and CTH. Both parameters show small, statistically insignificant, trends over the last 20 years. CFC has slightly increased by about 0.001 (+0.14 %) per decade, while cloud tops are lower by ≈ 6 m (-0.14 %) per decade. This finding excludes an explanation that reflectance loss at visible wavelengths, due to shrinking sea ice extent, is offset by more CFC or that the loss of CFC reveals more underlying bright surfaces. However, the bottom plot of Fig.  shows that the temporal trend over 2 decades for τ of liquid clouds has the opposite sign of that of ice clouds. τ of liquid clouds increases, statistically significantly, by about 0.4 (+2.85  %) per decade while the ice cloud τ decreases by 0.65 (-6.15  %) per decade in the same period. Altogether, the total τ of clouds has not changed, meaning that clouds have experienced a net shift from the ice to the liquid phase without changing their total opacity. The mean values for 1996 and trends of the above cloud properties are given in Table .

Similarly to the approach we used for the RλTOA trends regionally and qualitatively, we map cloud parameters in the bottom panel of Fig. , adding also the albedo of clouds at λ = 600 nm. CFC trends are regionally partitioned and are seen to increase in the range of 5 %–20 %, where the greatest sea ice losses are observed. This occurs during AMJ and less extensively in JAS. Examples of this behavior are found in the Barents, the Kara, and the Laptev seas. On the contrary, large areas of a statistically significant decrease in the range of 2.5 %–10 % are homogeneously observed across land masses circling the inner polar belt. This includes Greenland and the Atlantic corridor, confirming past results . More pronounced trends of the different cloud parameters, irrespective of their sign, occur in AMJ rather than in JAS. Hudson Bay is one of the few regions experiencing a seasonal trend reversal. The AMJ period is characterized by less cloudiness (-5 %), whereas the JAS period exhibits an increase of the order of almost 10 % over the last 2 decades. The resemblance to the trend reversal of all RTOA channels (Fig. ) indicates that CFC changes primarily modulate RλTOA over Hudson Bay. This is inferred from the absence of a change in trend sign for those cloud parameters that influence the reflectance in the solar spectrum, such as τ of liquid water and ice clouds in Fig. .

CTH decreases, especially where statistically significant trends are observed, during AMJ across almost all sectors of permanent and marginal sea ice (Beaufort, Chukchi, East Siberian, Laptev, and Kara seas) and over Baffin Bay. In the last 2 decades, CTH in these regions has decreased by 10 % on average. In JAS, however, CTH increases significantly from the Fram Strait, throughout the Barents and Laptev seas poleward, and in western Siberia. This is coupled with a slightly negative trend for Greenland and the surrounding waters, the southern Baffin Bay (the Davis Strait), the Beaufort Sea, and the East Siberian Sea.

Total τ is split into liquid and solid cloud phases. The geographic distribution of the trends in Fig.  provides insight into which areas are responsible for the positive pan-Arctic trend in τ of liquid clouds (τ-liquid) and the negative trend for ice clouds (τ-ice). τ-liquid increases across the whole Arctic in AMJ except over the Atlantic sector and the southern part of Baffin Bay. The positive trend is maintained over northern Greenland and the Canadian Archipelago, around the North Pole, and on part of the Eurasian continent also during JAS. A positive trend of τ-liquid is correlated with a trend of opposite sign for τ-ice: this holds for all regions of permanent and marginal sea ice, the Canadian Archipelago, and Hudson Bay. Greenland, Baffin Bay, and the Atlantic sector show a different behavior: there is a 34 % increase in τ-liquid during AMJ and a 22 % increase in JAS. Notwithstanding the increase over certain areas (e.g., north Greenland), mean τ-ice over the Arctic regions remains nearly unchanged in different seasons. The liquid phase of clouds does not increase across the Fram Strait, whereas the ice phase decreases by roughly 20 % in both AMJ and JAS periods. Finally, the Atlantic sector (Greenland and the Norwegian seas) shows decreases in the τ for both the liquid and solid cloud phases during AMJ and JAS.

The polar plots of seasonal trends in cloud albedo (CA) in Fig.  show that the magnitude of the positive trends in JAS is larger than those of AMJ, but the spatial extent of the CA trend values are similar in both seasons. To a certain extent, the CA trends are geographically correlated with those of CFC and τ-liquid. Individual regions are grouped similarly to the RTOA polar plots: comparable distribution of CA are found over the most eastern and most western Arctic seas (Beaufort and Chukchi, East Siberian, Laptev, and Kara seas). Positive trends are almost invariably distributed over water masses, the Canadian Archipelago, and the northern part of Greenland, irrespective of the season. In contrast, clouds become less reflective at lower latitudes, southern Greenland, and the Atlantic sector. Over the Siberian land masses, this is not observed, and CA changes in the region are attributed to a competition between changes in CFC and τ-liquid. The loss of albedo due to cloud dissipation is compensated by the increment in albedo through increased τ-liquid.

Figure  shows the trends and the standard error (i.e., 2σ standard deviation) of five cloud properties (CFC, CTH, τ of liquid and ice phase, CA) together with the trend of liquid (LWP) and ice water path (IWP), from the same cloud record . Changes in RλTOA depend in the first place on changes in cloudiness and τ (irrespective of the phase), which in turn is a function of LWP, droplet or crystal effective radius (reff), and air density ρ (i.e., (τ=3/2×LWP/ρreff)). The sign of LWP and IWP trends confirms the τ trends. We infer that τ-liquid has increased as a result of the positive change of LWP.

We compute the net radiative forcing derived only from clouds at the surface, or at the bottom of atmosphere (BOA), CRFBOA, from the differences between the downward and upward fluxes of SW and LW for all-sky and clear-sky conditions as follows: 2CRFBOA=SWdn-SWup+LWdn-LWupall-skyBOA-SWdn-SWup+LWdn-LWupclear-skyBOA. The multi-year mean and trends of SWBOA, LWBOA, and total CRFBOA for AMJ and JAS are plotted in Fig. . The pan-Arctic and regional values are reported in Table  for AMJ and in Table  for JAS. Although not the focus of the current study because of the observational limitations of RλTOA and the retrievals of optical cloud properties during the polar night, an annual perspective on mean CRF can be found in Fig.  and on CRF trends in Fig. , at both the surface and TOA.

The climatological annual pan-Arctic total CRF (see Fig. ) is positive at the surface with the sole exception of the Greenland Sea. Minimum values are found over Baffin Bay and the Barents Sea. Over the Arctic Ocean, the total CRF is positive and amounts to ∼ 7.0 W m-2, which is lower than the 10 W m-2 reported by KE-13 hereafter, while over land masses clouds warm the surface by ∼ 11 W m-2. Our results are directly comparable to those of KE-13 because the algorithm computing the broadband fluxes is based on the same radiative transfer , and the CRF is inferred from the difference between the all-sky and clear-sky atmospheric state, as in Eq. ().

Among the differences that may explain the bias in CRF between our results and those in KE-13, we consider differences in spatial coverage of the Arctic and the spectral albedo of ice- and snow-covered surfaces. KE-13 defines the Arctic as the region between 70 and 82∘ N, while in this study the Arctic is defined between 60 and 85∘ N. The spectral surface albedo in this AVHRR record is 6 % higher for wavelengths in the visible and NIR (0.958 at 630 nm and 0.868 at 910 nm vs. 0.9/0.85 for the dry/melt months in KE-13), while it is lower for wavelengths in the SWIR (0.036 at 1.6 µm and 0.0 at 3.7 µm vs. 0.15/0.05 and 0.05/0.05 for the dry/melt months in KE-13). This means that the Arctic albedo in our record is indicative of dry and bright surfaces at shorter wavelengths but appropriate for melt and darker surfaces towards the infrared. This would lead to an overall underestimation of the (negative) CRF in the SW when compared to KE-13.

Having defined the Arctic as all those areas north of 60∘ N encompassing also low-latitude areas of the relatively dark surface, at a pan-Arctic scale clouds exert a negative SW radiative forcing in both AMJ and JAS, which is larger than the LW component by ∼ 12 and ∼ 18 W m-2 in the same seasons (Tables  and ). However, the seasonal climatological mean CRF is highly variable across the Arctic and is regionally partitioned: clouds' total radiative forcing at the surface is positive over bright areas as a result of LW effects being larger than SW effects. For instance, this holds for the Beaufort, East Siberian, and Laptev seas and over Greenland, where total CRF becomes positive in both seasons and which corresponds to those Arctic areas over which the SW and LW CRF are also the smallest.

The combined effect of the brighter surface and comparatively low τ (irrespective of the phase) over Greenland (τ-liquid ∼ 8 in AMJ and ∼ 7 in JAS) increases SW reflectivity and damps upwelling LW. Minimum values in mean total CRF are seen over Baffin Bay, the Atlantic corridor, and the Barents Sea in both AMJ and JAS. For the same seasons, darker surfaces of the Atlantic corridor and Baffin Bay imply the presence of open water masses, which have higher temperatures and, therefore, emit LW more effectively. However, SW offsets LW and total CRF turns negative owing to larger average values of τ-liquid over the Greenland Sea (τ-liquid ∼ 15 in AMJ and ∼ 16 in JAS) or the Baffin Bay (∼ 15 in AMJ and ∼ 13 in JAS) (see Table ).

At low surface albedos, typically less than 0.2, SW CRF outweighs LW CRF for the great majority of clouds, irrespective of their water content, τ-liquid, and sun illumination. Typical values of solar zenith > 65∘ correspond to latitudes north of 75∘ N, encompassing the Arctic Ocean in both AMJ and JAS. Resorting to Fig. 7, we obtain a lowest LWP threshold of ∼ 20 g m-2 at surface albedo 0.5 and ∼ 250 g m-2 at albedo 0.8. This means that, with increasing surface albedo, SW radiative effects may offset those by LW only at specific values of LWP and sun illumination angles, thus making CRF more sensitive to changes in cloud τ-liquid.

Consequently, the majority of clouds warm the Arctic surface, and our results are qualitatively consistent with current knowledge . The maximum cloud warming at the surface occurs over Greenland and to a lesser extent above sea-ice-covered regions in AMJ (East Siberian, Beaufort, and Laptev seas) and JAS (East Siberian and Beaufort seas). Otherwise, the other Arctic regions show a negative total CRF, from a minimum over the Greenland and Barents seas in AMJ to a less negative CRF over those regions influenced by the climate of the low latitudes (Baffin Bay, Greenland and Barents seas). Hudson Bay and the Kara Sea in JAS, respectively, show a total negative CRF of similar magnitude.

From the CRF trends of the last 2 decades (Fig. ), clouds over the perennial sea ice zone increasingly cool TOA (see Fig. ) and the surface (bottom of atmosphere, BOA) alike, while being neutral to positive over the Atlantic corridor and land masses at low latitudes. In AMJ months, maximal cooling trends at TOA (BOA) are for Kara and Laptev and extend along the Arctic Circle up to the northern section of the Baffin Bay through the Chukchi Sea, albeit dropping in magnitude to -0.9 (-0.8) W m-2 per decade. During AMJ, clouds have increasingly cooled the Siberian land masses and the marginal sea ice zones at an average rate, with the Barents Sea undergoing the strongest CRF drop by 2.5 W m-2 per decade.

Otherwise, the CRF trend at TOA and BOA during JAS varies from slightly positive over land masses, such as Eurasia, and over open waters in the Atlantic sector, the southernmost portion of Baffin Bay, and the Bering Strait. Cooling trends due to clouds are identified over Greenland for both seasons, having a rate of -0.5 W m-2 per decade. The influence of changes in surface albedo is manifested in these results. Where surface albedo remains almost constant (land masses, Greenland, and the Atlantic corridor) then CRF trends are of lesser magnitude. Instead, where the surface experiences more substantial changes, both seasonally and over the long term, trends in CRF are amplified, due to a greater influence of SW over LW.

None of the trends in CRF in Fig.  are statistically significant at 95 % confidence over the 20-year time frame of this data set. Thus, we estimate the time of emergence (ToE) in years for a trend to become statistically significant (see Appendix , Fig. ). The seasonal ToE and regional ToE are reported in Tables  and .

The cloud water path (CWP) is defined as the weighted sum of the two phases, whose relative occurrence is 0.54/0.46 % in AMJ and 0.63/0.37 % in JAS for the liquid/ice clouds, respectively. The seasonal correlation between CWP and its liquid/ice component is, respectively, 0.79/0.75 in AMJ and 0.57/0.84 in JAS, showing that the loss in ice water content is the main driver for the loss of total water condensate in clouds, more in summer than in spring. While highly variable at the pan-Arctic scale, the total change in CWP amounts to -0.51 ± 11.01 % in AMJ and -3.66 ± 7.29 % in JAS.

Notably, the majority of water path changes exceeding natural variability are those of LWP or IWP decrease over areas of sea ice loss and only partly of LWP increase over land masses, the Canadian Archipelago, some spots of Greenland, and the Beaufort Sea in JAS. Additionally, from Fig.  it can be seen that only those CWP trends in both seasons are statistically significant where the LWP and IWP trends are statistically significant too. This holds for the Fram Strait, the northernmost area of the Canadian Archipelago, the Bering Strait, and the coastal area of the Siberian continent. Only in AMJ do more statistically significant patterns of CWP trend emerge, these comprising areas from the Laptev, from the Kara, and throughout the northernmost part of the Barents seas.

In light of the results presented so far regarding the optical thickness and separation of the two cloud phases, it is reasonable to assume that this trend will continue in the future, allowing more patterns of statistical significance to emerge even where they have not been detected with 20 years of data.

Atmospheric moisture fluxes are increasing as a result of more open waters and transport . Marked regionality and seasonality of RλTOA, cloud properties, and CRF across the Arctic are identified in four macro-regions, consistently exhibiting similar behavior: Greenland, the permanent and marginal sea ice areas, the Atlantic sector, and the land masses at lower latitudes.

Greenland has a unique behavior: RλTOA trends at all wavelengths are positive, irrespective of the season (Fig. ). The AMJ RλTOA trends, up to 5 %, are even larger than those for JAS. This result is particularly surprising, given the insignificant CFC trend at the pan-Arctic scale and the local negative CFC trend in both seasons (Figs. , ). Thus, these factors do not contribute to an increase in the overall reflectance. Therefore, we conclude that the increase in RλTOA is due to the enhanced exposure of reflective surface in the southern part of Greenland, while a similar increase in the northern part is due to the simultaneous increase of τ-total (Fig. ) and CWP (Fig. ).

Similar behavior is found in Hudson Bay and the Canadian Archipelago, which show an increase in reflectance, in contrast to a general darkening of the Arctic. The mechanism by which these regions increase RλTOA lies in the link between LWP and CA, through τ-liquid. In fact, τ-liquid changes sustain the correlated RλTOA changes because of the non-linear relationship of CA to τ-liquid via LWP. It follows that a RλTOA loss is overcompensated by more liquid clouds in the northern sector and by increased snowfall in the southern part of the Greenland continent. Cloud LWP has increased by 28 %–30 % over Greenland and by 14 %–16 % over Hudson Bay. The Canadian Archipelago also displays positive τ-liquid trends of 30 %, 14 %, and 22 %, respectively. Notably, the seasonal behavior of τ-liquid, increasing over Greenland, is not associated with CFC loss and a positive CRF change in the last 20 years. In contrast, cloud dissipation, increased by anticyclonic activity and concurrent temperature inversion strengths, is responsible for enhanced insolation at the ground and its concurrent melting effects . In addition to cloud loss (Figs.  and and ), extensive ice melt in Greenland is also known to be enhanced by low-altitude liquid water clouds that have sufficient opacity to enhance downward LW flux but are also optically thin enough to allow a significant amount of SW flux to pass through. This results in the surface being warmed . Such clouds occur in the LWP region between 10 and 60 g m-2.

Figure  shows that the increase in τ-liquid of clouds and LWP over Greenland in spring and summer is among the largest in the entire Arctic (ΔLWP > 20 %–40 %). In both seasons, the cloud fraction decreases, and τ-liquid (as well as the LWP) increases spatially on average. Both effects impact upon the downward SW flux at BOA, but in the opposite direction, resulting in a small net positive change in SW CRF. For decreasing CFC over Greenland and in presence of an increase in near-surface temperatures, we expect a decreasing downward LW flux which might not be compensated by the LW enhancement by more liquid water in the clouds (Fig. , middle panel).

The changes in cloud properties and RλTOA over Hudson Bay are exceptional. A 9 % increase in τ-liquid and minimal CRF changes are correlated to the greatest R560TOA increase in the record in JAS. This area shows one of the largest CFC increases during summer months (Fig. ), also corroborated by similar significant changes in AMJ and JAS observed in the reanalysis data . The total CRF is -30.7 W m-2, while it is 2.7 W m-2 during AMJ. CRF trends point to a cloud cooling of Hudson Bay at a rate of -2.9 (AMJ) and -1.3 (JAS) W m-2 over the last 2 decades.

Cloud forcing at the surface depends on cloud property changes. The behavior is summarized in the seasonal and regional charts of Fig. , in which mean value and trend of SW, LW, and total CRF are shown as a function of τ-liquid of clouds, LWP, and CFC changes. The relationships between total CRF, τ, and LWP are more important in modulating radiation in JAS than in AMJ. This is the case when the underlying surface still has an albedo high enough to modulate CRF, as in the spring months over regions with sea ice. With a decreasing surface albedo, as in the summer months, SW CRF cooling dominates over LW CRF warming. As a consequence, Arctic regionality emerges from the clustering of the regions, especially in AMJ and to a lesser extent in JAS. We conclude that in the last 2 decades the net radiative effect of clouds on the surface is decreasing.

Those regions characterized by a darkening surface undergo a relative increase in SW reflection by more liquid clouds, leading to an increased cooling by clouds (ΔCRF < 0). This takes place over the Barents Sea, a region characterized by early sea ice loss in AMJ, and over the perennial sea ice zone (Beaufort, Laptev, and East Siberian seas), where a CRF decrease at a rate of -1 to 2 W m-2 is associated with greater cloudiness in AMJ and increasing τ-liquid in JAS.

From Fig.  we note that any positive τ-liquid trend corresponds to positive LWP changes for both seasons. Although not surprising, the AMJ changes in CRF do not correlate with either LWP or τ. In the JAS months, however, larger cloud optical densities and LWPs are matched by a decrease in CRF at the surface. This is the effect of darkening the surface that lowers the LWP value necessary for the CRFSW to dominate CRFLW. Excluding the Barents Sea, the variability of ΔCRF during AMJ is narrower (-4.2 to +0.9 W m-2) than during JAS (-6 to +0.4 W m-2). This is evidence of the importance of radiance from the underlying surface, which is larger in AMJ than in JAS. Overall, the radiative effect of CFC and τ is expected to be similar, provided that their changes in time agree in sign. Because CFC and τ change in opposite directions, the decreases in LW CRF and increases in SW CRF suggest a dominant influence of CFC rather than by water content in the clouds over Greenland. This CFC influence is still modulated, but not offset, by the changes in τ and CWP.

One exception is the East Siberian Sea in JAS where τ-liquid of clouds grows despite a lower content of liquid water. Notwithstanding the unexplained contribution of reff, we note that in JAS the East Siberian Sea has experienced a decrease in cloud altitude (see Fig. ), which is a well-behaved parameter in the AVHRR record over most of the Arctic . Assuming that the cloud bases are unchanged, any change in CTH can influence the relationship τ=3/2×LWP/(ρreff) through changes in ρ.

Figure  shows also that CFC changes (i.e., outlined vs. filled circles) modulate mainly the LW portion of cloud radiation in both seasons. The seasonal coefficients of determination r2 of SW CRF by CFC trends are comparable to those by τ-liquid trends. However, for the LW CRF, r2 by CFC is higher than that by τ-liquid (CFC: AMJ 0.98 for both above ocean and all areas; JAS 0.87 above the ocean and 0.94 above all areas. τ-liquid: AMJ 0.39/0.02 above ocean/all areas; JAS 0.65/0.19 above ocean/all areas). This is the case when clouds become optically denser and hence more reflective.

Quantitatively, with values of ΔCRFTotal= -1.4 W m-2 and ΔCF=3.03 %, we obtain the total long-term sensitivity ΔCRFTotal/ΔCF= -0.48 W m-2 %-1 over the Beaufort Sea in AMJ. The sensitivities of the SW and LW parts of CRF amount to -0.56 and +0.84 W m-2 %-1. Although averaged over one multi-year season only, our estimation is in line with measurements reported at the same location during the SHEBA campaign. The SHEBA sensitivity of ∂CRFLW/∂CF=0.65 W m-2 %-1 was seen to offset the SW for most of the year (with ∂CRFSW/∂CF∈[0,1] W m-2 %-1), thereby warming the surface, while cloud cooling took place only in midsummer months with the highest sun illumination and lowest surface albedo in late summer .

Accordingly, we report a net total (SW + LW) sensitivity of -0.13 W m-2 %-1 in JAS, meaning that the SW cooling takes over LW warming during the Arctic JAS in the record. The warming effect from increased CFC in AMJ over these regions is directly linked not only to the retreat of sea ice, the onset of which is in late May , but also to the enhanced convergence of atmospheric water content originating from open Arctic oceans during years with anomalously low sea ice extent. Provided that the ocean cannot be an appreciable source of water vapor in the Arctic boundary layer, attribute an increased downwelling LW flux to the increased atmospheric opacity as a result of the convergence of moisture, in the form of clouds and/or water vapor . Our results imply that this mechanism is evident in the year-to-year variability of exceptional sea ice lows and is also a long-term component at decadal timescales, during which atmosphere–ocean coupling effects are predominant.

From a modeling standpoint, we can validate past results for which the increases in cloud τ-liquid and LWP are projected to extend well beyond the middle of the present century. Constraining the cloud microphysics and thermodynamic phase will be crucial to project future Greenland melting and assess the sign and strengths of total cloud feedbacks . Given the actual and future Arctic temperatures, ice in the clouds will be increasingly depleted. Hence, τ-liquid and LWP will increasingly determine net cloud feedbacks .

When the cloud ice phase turns to liquid water, negative feedback is expected due to the offsetting of LW by SW. This is especially true in those months characterized by low surface albedo, under a stronger interaction with atmospheric radiation by liquid cloud droplets rather than ice crystals. For the rest of the year when the surface albedo is high and sun illumination is low or absent, the cloud feedback is expected to be more positive, which is a warming effect. If climate models do not correctly capture this behavior, i.e., they do not incorporate more supercooled liquid and mixed-phase clouds , unrealistically large amounts of ice result, effectively contributing to the uncertainty in determining the sign of the net cloud feedback.

We consider that this is one reason which may explain in part the discrepancy between the atmospheric components (CAM) of the Community Earth System Model Fig. 2. While show that prescribing in the CESM1-CAM5 weaker scavenging of supercooled liquid droplets by ice crystals in spring months leads to an increase in available atmospheric liquid water and a concurrent increase in downwelling LW flux at the surface, we note that the CAM5 positive cloud feedback at Arctic latitudes becomes negative in CESM2-CAM6 as a result of improved modeling of the cloud phase. Coherently, CAM6 projects increased rainfall rates within a warmer Arctic in JAS at the expense of snow precipitation , as the outcome of poleward moisture streams and more liquid Arctic clouds.

Nevertheless, an improved representation of supercooled liquid clouds in CAM6 models does not necessarily result in better accuracy in describing cloud feedback. Although there is consensus that clouds, twice as bright in CAM6 than in CAM5, increasingly reduce the amount of SW energy accumulated at the surface through optical thickness and phase feedbacks , thereby slowing the Arctic sea ice albedo feedback by 5 years over oceans and 2 years over land , there are indications that clouds might accelerate the albedo feedback in some CMIP6 models . This holds in summer months when the atmospheric contribution to Arctic TOA albedo, dominated by cloud reflectance, is higher than that of the surface. While suboptimal prescribed co-variability of clouds with the underlying sea ice is not ruled out, indicate that future efforts should focus on understanding the parameterization of the cloud microphysics, especially for those models that show a decrease in atmospheric reflectance.

Advances in observational techniques and process-level research are needed to assess unambiguously the relative roles of temperature and atmospheric particulate matter in determining cloud thermodynamic changes. In the absence of a systematic, pan-Arctic, aerosol indirect effect due to decreasing trends of ice-nucleating particles or cloud condensation nuclei (INPs or CCN), higher condensation rates (i.e., positive LWP trends) of small-sized cloud droplets can only nucleate and grow by a combination of changes in Arctic boundary layer depth within a saturated air volume. Different temperature regimes influence cloud albedo by changing the τ–reff–LWP relationship and favor droplet growth over condensation rates and vice versa .

To this end, the role of reff remains the unexplained factor in the relationship between τ and the water path. The reff size spectrum is modulated by the amount of water vapor and available particulate. While model and satellite data show a general moistening of the Arctic , local on-ground evidence of a recent decrease in total aerosol burden is growing. However, INP or CCN cannot be directly inferred from changes in column-integrated extinction of total aerosol load, assuming a CCN decrease is in contradiction with the reff reduction via the Twomey effect. Alternatively, we speculate that the change in size spectrum or aerosol type might lead to optimal INP/CCN size and hygroscopicity , although the total aerosol amount has decreased. This could be the case when anthropogenic aerosols decrease because of emission policy but natural aerosols increase due to more frequent boreal forest fires, increased sea spray, and marine biogenetic activity as a result of more open waters .

Satellite-derived single reff values, such as those in the record analyzed in this work, are only representative of the droplet/crystal population at a level of ≈ 1τ from the cloud top . We recommend that the available and relevant spectral observations are exploited to generate a pan-Arctic picture of in-cloud reff(z) profiles, which would optimally complement surveys based on spaceborne active techniques . reff(z) profiles, together with aerosol speciation at high latitudes and cloud bases , are essential in two ways. First, they constrain INP/CCN activation, supersaturation, and cloud particle number concentrations . Second, cloud fields will be more accurately separated according to their phase (liquid, ice, and mixed phase) and layering (low, mid, high level and multi-layered). We consider our results as upper bounds, and more vertical resolution will improve our understanding of the evolution of clouds in the Arctic.

Finally, a better estimation of the cloud-free surface albedo would enable us to pinpoint the broadband radiative interactions between the surface and the clouds. Recent results suggest that the SW effects of clouds at the surface almost double even in the presence of sea ice and snow. As a result, the total cloud radiative forcing shifts from warming to neutral values already at the beginning of the melt season in mid June . This would imply that the results presented here underestimate the cooling effect of clouds.