A controlling factor in the seasonal and climatological evolution of the sea ice cover is its albedo α. Here we analyze Arctic data from the Advanced Very High Resolution Radiometer (AVHRR) Polar Pathfinder and assess the seasonality and variability of broadband albedo from a 23 year daily record. We produce a histogram of daily albedo over ice covered regions in which the principal albedo transitions are seen; high albedo in late winter and spring, the onset of snowmelt and melt pond formation in the summer, and fall freezeup. The bimodal late summer distribution demonstrates the combination of the poleward progression of the onset of melt with the coexistence of perennial bare ice with melt ponds and open water, which then merge to a broad peak at α ≳ 0.5. We find the interannual variability to be dominated by the low end of the α distribution, highlighting the controlling influence of the ice thickness distribution and large-scale ice edge dynamics. The statistics obtained provide a simple framework for model studies of albedo parameterizations and sensitivities.
 Here we describe an analysis of daily satellite retrievals of the directional - hemispheric apparent albedo from the Advanced Very High Resolution Radiometer (AVHRR) Polar Pathfinder (APP). (The apparent albedo is what would be measured by upward and downward looking radiometers and hence varies with the state of the atmosphere and the solar zenith angle. For simplicity we refer to this as α.) By culling out the albedo retrieved solely from the ice covered region of the Arctic Ocean from 1982 to 2004 a pixel scale view of the surface is created from which we produce a daily histogram. The principal seasonal surface transitions in the ice cover are observed; the high albedo snow covered spring to snowmelt convolved with the poleward progression of melt onset, the appearance of melt ponds, their coarsening and fall freezeup (Figures 1–3). The data reveal the decadal and intra and inter-seasonal variability of the albedo over multiple timescales. It is hoped that their statistics and trends will be of use in the development of models ranging from the very simple to GCMs.
 The APP dataset has been refined and applied for use in a wide range of polar studies and is described in detail on the NSIDC website [Fowler et al., 2000]. AVHRR channels 1 – 5 range from the visible to the thermal infrared (0.58 – 12.5 μm) and measure Top of the Atmosphere (TOA) reflectances (1, 2, 3A) and brightness temperatures (3, 3B, 4, 5). Additional information includes the solar zenith, relative azimuth and satellite elevation angles, cloud and orbit masks and universal time [see Fowler et al., 2000, Figures 3 and 4]. For the analysis reported here we use the so-called Surface Type Mask and α retrievals from 1 January 1982 through 31 December 2004, taken daily at 1400 hours. The albedo retrieval involves four steps developed and tested by Csiszar and Gutman : (a) normalize channels 1 and 2 with respect to solar zenith angle, (b) convert channels 1 and 2 narrow band reflectance to a TOA broadband reflectance, (c) correct the TOA broadband reflectance for anisotropy, and (d) convert the TOA broadband reflectance to a clear sky surface broadband albedo.
 The Northern Hemisphere is gridded in 1805 × 1805 pixels, with each pixel representing a 5 km × 5 km region. Sea ice is distinguished from land and open water using SSM/I brightness temperatures and filtering the Surface Type Mask data with the NASA Team Sea Ice Algorithm which distinguishes between first-year (FYI) and multiyear ice (MYI) concentrations. The approach ascribes a MYI flag to a region containing at least 50 percent of this ice type. We note that uncertainties depend on the season (e.g., melt pond fraction) and region (near the ice edge), along with the surface type categories [Comiso and Kwok, 1996]. In particular, as will become relevant in our interpretation, errors are enhanced in the presence of surface ablation (especially during snowmelt) because a small amount of liquid water alters the emission characteristics of the surface and it is difficult to distinguish between open water at the freezing point and wet ice surfaces. In their ground truthing study of this retrieval scheme Stroeve et al.  note that the satellite albedo and that measured by radiometers at the surface are equivalent and are functions of atmospheric conditions, solar zenith angle and the satellite elevation angle. Finally, the cloud masking scheme was ground truthed by Maslanik et al.  during the most difficult part of the cloud seasonal cycle (April-July) over boxes from 15 to 305 km on a side. The maximum standard deviation of albedo was found to be 0.05 which reduced precipitously as the length scale decreased.
 On physical grounds the albedo data are filtered to remove any values greater that 1 or less than 0.2. We examine each pixel every day for the presence of ice. Then, for a specified time threshold τth, we average α for that pixel. Different thresholds τth specify the minimum time for which a pixel contained ice. A histogram for each day of the year is thereby produced and represents the number of pixels for each average albedo value bin for that day. Different colors on the histogram specify different thresholds. Red indicates that a pixel has contained ice for the whole time period under consideration, i.e., τth = 23 yr (1982 through 2004). Black specifies if a pixel has been ice for at least τth = 22 yr in the entire time period, which helps us to remove any bias introduced by any minimum during a year. White denotes τth = 17 yr; blue τth = 12 yr; green τth = 8 yr; and yellow τth = 1 year in the entire time period. Our approach is motivated by several factors. (a) The knowledge of the great sensitivity of GCMs to albedo parameterizations [Eisenman et al., 2007]. (b) Our continued development of theoretical treatments for large scale transitions in the ice cover that test the robustness of bifurcations in the state of the system to the ice-albedo (and other) feedback [Moon and Wettlaufer, 2011]. (c) Extending these theories to understand the role of noise. Whence, an observational understanding of the temporal variability, from month to month and interannually is essential, and is simply captured using the threshold idea.
3. Results and Discussion
 The results are shown in Figures 1–3 and we discuss first the entire record shown in red (τth = 23 yr). Because of the limited spatial resolution of the retrieval and the uncertainties in ice type detection and albedo estimation, we do not expect that the numerical values averaged over the ice-covered regions as determined here to map directly onto field observations, but our results nonetheless nicely display the five distinct phases of albedo evolution from spring through fall: dry snow, melting snow, pond formation, pond evolution, and fall freezeup (see Figure 3 of Perovich et al. [2007b, and references therein]). The dry snow value seen here (∼0.75) will by definition be lower than a field observation on MYI (∼0.85) because our pixels contain both MYI and FYI, which have varying amounts of snow cover [Kwok and Untersteiner, 2011], and near the ice edge the pixels include open water. For March, April and May α has a single strong peak about a value indicative of snow covered ice, with an asymmetry that shows a sharp decay at the high values and a broader decay at low values capturing the range of ice types and snow covers that are expected over the basin in the spring. By the middle of June the distribution is clearly bimodal, with a high peak at about the melting snow covered MYI value (∼0.70) and melt pond (∼0.50) signatures beginning to come through as the peaks broaden and move to lower values by mid-July. This is due to the poleward progression of the onset of melting that, in mid-June, can have gradients that span 100's of km taking a week to progress northward from 73 to 82 °N [Winebrenner et al., 1994]. Such a large scale progression is clearly resolved by our method and hence the low latitude semi-zonal region of low α contributes to left hand peak in our distribution. They finally merge to a single broad peak at a value indicative of fractional coverage of melt ponds and perennial ice [see Skyllingstad et al., 2009, Figures 11, 13 and 14]. By mid-September, when freezeup is in full swing, the peak shifts slightly to the right centered around a perennial ice value of α ∼ 0.6. Finally insolation diminishes in October.
 There is a systematic trend of increasing variability as τth decreases and this is dominated by the low α part of the distribution. The variability increases substantially from decadal to seasonal, indicating both larger intrinsic fluctuations in the areal fraction of ice types as well as in the ice edge position. Changes in both will have a signature at small values of α. The integrated influence of this observation is a larger interannual variability in solar insolation received by the Arctic Ocean, consistent with the findings of Perovich et al. [2007b].
 As a means of further quantifying the variability, we plot the daily mean albedo as a function of τth in Figure 4 which demonstrates a monotonic increase in the albedo with τth. The seasonal trend stands out, as does the decrease in the threshold dependence during the early evolution of the melt season. The latter observation is consistent with the notion that as snow and ice melt from the surface and the system approaches the bulk freezing point, the intrinsic variability between ice types and ice and water is suppressed. Moreover, as discussed above, because the retrieval scheme is less accurate as the pack begins to ablate from its surface [Comiso and Kwok, 1996], there is less τth dependence during these periods. However, as the open water, melt ponds and mature ice surfaces in the pack evolve toward freezeup, curves with different τth begin to diverge as might be expected [Skyllingstad et al., 2009; Fetterer and Untersteiner, 1998]. Finally, in Table S1 in the auxiliary material we give the monthly variance in the albedo as a function of τth which provides a slightly different view of the seasonal evolution shown in Figure 4.
 The annual cycle of low stratiform cloud fraction ranges from 20% in the winter to 70% in the summer. According to the observational synthesis of Eastman and Warren  the interannual variations of cloud amounts demonstrate significant correlations with surface air temperature, and total sea ice area as well as the Arctic Oscillation. Septembers with low areal extent are generally preceded by a summer with decreased middle and precipitating clouds and the autumns following such Septembers are enhanced in low cloud cover. They conclude that the total cloud cover appears to be greater throughout the year during low-ice years. Hence, as expected, this increasing cloud cover would promote long wave forcing and ice loss, which may also be enhanced by the observed decrease of summer cloud cover. Interestingly, when the cloud fraction is growing from April to mid June (20–70%) viz., the climatological seasonal cycle, we see the least dependence on threshold (Figure 4) which we believe is due to the suppression of the difference in ice type and water associated with the ablation of the surface. Therefore, an observed trend of decreasing cloud cover in summer would be consistent with the weaker threshold dependence we observe for the onset of melt and the increase in the threshold dependence during autumn freezeup, because of the larger fraction of open water. However, any increase in cloud fraction correlated with minima in ice cover would manifest itself as an increase in the low-α part of the distribution with decreasing τth as already seen here. Thus, to ascribe a cause would require independent measurements, on the same timescales, of all cloud types. Interpretation would also benefit from a clear distinction between the albedos of the different ice types which is the topic of a more extensive study.
 We have analyzed daily retrievals of the directional - hemispheric apparent albedo from summer 1982 to summer 2004 using the APP archive and have produced histograms of daily albedo over ice covered regions. This reveals the principal albedo transitions over the ice pack; high albedo associated with dry snow covered ice in late winter and spring, the onset of snowmelt and melt pond formation and evolution in the summer, and fall freezeup. We find a bimodality of the late summer histogram, displaying the combination of the poleward progression of the onset of melt with the coexistence of perennial bare ice with melt ponds which eventually merge to a broad peak at α ≳ 0.5 consistent with field constrained models [e.g., Skyllingstad et al., 2009]. There is substantial decadal to interannual variability that is dominated by the low end of the α distribution. This is consistent with a recent combined field and remote sensing study that found large interannual variability in the solar insolation in the Arctic [Perovich et al., 2007b], and our method allows us to examine variability on multiple timescales up to the length of the record. In the simple approach described here the variability originates in the natural fluctuations associated with the interannual changes in the fractions of ice types and in the position of the ice edge. By implication, a similar study of the more recent past would presumably reveal a broad low albedo distribution associated with the increased fraction of FYI [see Kwok and Untersteiner, 2011, Figure 2]. Because of the thickness dependence of the albedo [see, e.g., Eisenman and Wettlaufer, 2009, equation 4] and the observed increasing fraction of thin ice, the results highlight the role of the ice thickness distribution in controlling the spatially homogenized albedo. This has implications for the manner in which all classes of models, from the very simple to GCMs, treat this important quantity. Such models could easily utilize our distributions or their moments as input and use the τth-dependent variances as a baseline for treating the variability.
 The authors thank I. Eisenman, R. Kwok and N. Untersteiner for feedback on various aspects of this project as well as the three anonymous referees for their thorough comments. They also thank Yale University for support. WM thanks NASA for a graduate fellowship and JSW thanks the Wenner-Gren Foundation and the John Simon Guggenheim Foundation.
 The Editor thanks three anonymous reviewers for their assistance in evaluating this paper.