The CryoSat-2 radar altimetry mission was launched April 2010, providing estimates of ice thickness during the ice growth season. CS2 provides freeboard estimates, or the height of the ice surface above the local sea surface, which when combined with information on snow depth, snow density and ice density can be converted to ice thickness assuming hydrostatic equilibrium (e.g., Laxon et al., 2013). Here we evaluate ice thickness fields provided by three different data providers in order to assess robustness of the observed thickness anomalies. Thickness is retrieved from ice freeboard by processing CS2 Level 1B data, with a footprint of 300 m by 1700 m, and assuming snow density and snow depth from the Warren et al. (1999) climatology (hereafter W99), modified for the distribution of multi-year vs. first-year ice (i.e., snow depth is halved over first-year ice) (see Laxon et al., 2013 and Tilling et al., 2017 for data processing details).

While the three data providers rely on W99 for snow depth and density, each institution processes the radar returns differently. In general, the range to the main scattering horizon of the radar return is obtained using a retracker algorithm. This can be based on a threshold (e.g Laxon et al., 2013; Ricker et al., 2014; Hendricks et al., 2016) or a physical retracker (Kurtz et al., 2014). While the CPOM and AWI products use a leading edge 70 % threshold retracker, Kurtz and Harbeck (2017) rely on a physical model to best fit each CryoSat-2 waveform. This will lead to ice thickness differences based on different thresholds applied: Kurtz et al. (2014) found a 12 cm mean difference between using a 50 % threshold and a waveform fitting method.

We note that several factors contribute to CS2-derived sea ice thickness uncertainties, including the assumption that the radar return is from the snow or ice interface (Willat et al., 2011), snow depth departures from climatology and the use of fixed snow and ice densities. In this study we initialize the CICE model simulations described below with the CPOM sea ice thickness fields. Accuracy of the CPOM product has been evaluated in several studies, suggesting mean biases between thickness observations in 2011 and 2012 of 6.6 cm, when compared with airborne EM data (Laxon et al., 2013; Tilling et al., 2015). For April 2017, the CPOM near-real-time product (Tilling et al., 2016) was used in place of the archived product, with a mean thickness bias of 0.9 cm between these products.

In this study, individual thickness point measurements are binned into 5 CICE thickness categories (1: < 0.6 m, 2: 0.6–1.4 m, 3: 1.4–2.6 m, 4: 2.6–3.6 m, 5: > 3.6 m) on a rectangular 50 km grid for each month. The mean area fraction and mean thickness is derived for each thickness category and these values are interpolated on the tripolar 1∘ CICE grid (∼ 40 km grid resolution). Grid points with less than 100 individual measurements and a mean SIT < 0.5 m are not included. Otherwise, all individual observations are included. For November, this effectively limits the area of the Arctic to the region shown in Fig. 1c. Negative thickness values that are retained in the CS2 processing to prevent statistical positive bias of the thinner ice are added to category 1. The novel approach of initializing the CICE model with the full ITD rather than the mean sea ice thickness provides an additional control on the repartition of the ice among different thickness categories. This in turn allows a more accurate representation of ice growth and ice melt processes (Tsamados et al., 2015) compared to initializing with the mean grid-cell SIT and deriving the fractions for each ice category assuming a parabolic distribution. Ice growth and melt strongly depend on SIT: using a real distribution can have a big impact, especially for thin ice.

CICE is a dynamic thermodynamic sea ice model designed for inclusion within a global climate model. The advantages of using CICE for this study is that we can more readily separate thickness anomalies into their thermodynamic and dynamical contributions, examine inter-annual variability and perform longer simulations. For this study, we performed two different CICE simulations. The first is a multi-year simulation from 1985 to 2017 (referred to as CICE-free). The second is a stand-alone sea ice simulation for the pan-Arctic region starting in mid-November and running until the end of April of the following year for the last seven winter periods from 2010/2011 to 2016/2017. This results in seven 1-year long simulations (referred to as CICE-ini), in which the initial thickness and concentration for each of the five ice categories is updated from the CS2 ITD using the CPOM CS2 November thickness fields. For grid points without CS2 data, and for all other variables (e.g., temperature profiles, snow volume), results from the free CICE simulation with the same configuration, started in 1985, are applied. In this way, CICE simulations cover the pan-Arctic region, but in regions where no CS2 are available, we restart SIT values from the free CICE model run. While this approach would be problematic in a coupled model, in a stand-alone sea ice simulation the model adjustment to the new conditions is smooth and the impact of using the vertical temperature profile from the free simulation only affects sea ice thickness in the order of millimeters.

Snow accumulation can depart strongly from the W99 climatology for individual years. Thus, we make the assumption that the deviation of the mean annual cycle of snow depth over the last 7 years from the W99 climatology is small and assume mean winter ice growth to be determined accurately from CS2, and tuned CICE-ini accordingly to match the observed CS2 mean winter ice growth from the CPOM product in the central Arctic (Fig. 1). The excellent agreement for both CICE-ini and CICE-free with CS2 increases the confidence of our model results. Therefore, our approach allows us to study inter-annual variability from two model configurations with different sources of errors, in addition to the three CS2-based products.

For both CICE simulations, NCEP-2 provides the atmospheric forcing. We use NCEP-2 2 m air temperatures because they have been shown to be more realistic for the Arctic Ocean than those from ERA-Interim (Jakobshavn et al., 2012). The setup is the same as described in Schröder et al. (2014), including a simple ocean-mixed layer model, a prognostic melt pond model (Flocco et al., 2012) and an elastic anisotropic-plastic rheology (Tsamados et al., 2013), with the following improvements: we apply an updated CICE version 5.1.2 with variable atmospheric and oceanic form drag parameterization (Tsamados et al., 2014), we increase the thermal conductivity of fresh ice from 2.03 to 2.63 W m-1 K-1, snow from 0.3 to 0.5 W m-1 K-1 and the emissivity of snow and ice from 0.95 to 0.976. While the default conductivity values are at the lower end of the observed range, the new values are at the upper end and have been applied in previous climate simulations (e.g., Rae et al., 2014).

Below, all CS2-derived sea ice thickness anomalies are computed relative to the CS2 time-period: November anomalies are relative to 2010–2016 and for April they are relative to 2011–2017. Results for November and April are only shown for all grid cells that have a minimum thickness of 50 cm and a minimum of 100 individual measurements for each of the seven years. For the month of November, this corresponds to all colored areas shown in Fig. 1c. For April, this region represents the area in red shown in Fig. 1d. The larger region shown in Fig. 1d also corresponds to the region over which the amount of thermodynamic ice growth and dynamical ice growth between November and April are assessed from the CICE simulations. For comparison with CS2, we present the mean thickness of the ice-covered area. In winter, the sea ice concentration in the model generally ranges between 0.98 and 0.995 % apart from locations close to the ice edge. Further note that area-averaged values for November and April are only given for regions shown in Fig. 1c and d, respectively.

The growing season air temperature anomalies (i.e., mid-November 2016 to mid-April 2017, relative to 1981–2010) were positive throughout the Arctic, leading to large reductions in the number of FDDs, computed as the cumulative daily 2 m NCEP-2 air temperatures below -1.8 ∘C, similar to Ricker et al. (2017a). FDDs computed this way reflect both the number of days with air temperatures below freezing and the magnitude of below freezing air temperatures over the specified period. Spatially, FDD anomalies show widespread reductions over most of the Arctic Ocean, with the largest reductions in the Barents and Kara seas, stretching across the pole towards the Beaufort and Chukchi seas (Fig. 2b). In contrast, during winter 2015/2016, FDDs were most notably anomalous within the Barents and Kara seas (Fig. 2a), in agreement with Ricker et al. (2017a). Overall, as averaged from 70–90∘ N, winter 2016/2017 witnessed the least amount of cumulative FDDs since at least 1979 (Fig. 2c).

While ice forms quickly within the central Arctic once air temperatures drop below freezing, 2016/2017 saw large delays in freeze-up throughout the Arctic. Updating results previously reported in Stroeve et al. (2014), freeze-up was delayed by 20 days for the Arctic as a whole, with regions like the Bering, Beaufort, Chukchi, East Siberian and Kara seas delayed by 3 to 4 weeks (Fig. 2d). Within the Barents Sea, the regionally averaged freeze-up was delayed by 60 days. In recent years, the trend towards later freeze-up has increased, with the Barents and Chukchi seas showing the largest trends in the order of +14 days per decade through 2017, followed by the Kara and East Siberian seas with delays in the order of +10 to +12 days per decade. Within the Beaufort Sea, freeze-up is now happening later by +9 days per decade (Table 1).

Before analyzing how the reduced number of freezing degree days impacted winter ice growth during 2016/2017, it is useful to first intercompare the different CryoSat-2 thickness estimates. We start with a comparison of November thickness from the three CS2 data sets from November 2010 to 2016 (Fig. 3). It is encouraging to find that year-to-year variability in the spatial patterns of positive and negative thickness anomalies are generally consistent between the three products despite differences in waveform processing. The AWI and CPOM data sets are in better agreement with each other than with the NASA product, which is expected as they use a similar retracker. Furthermore, all three data sets show widespread thinner ice in November 2011, and widespread thicker ice in November 2013. This is further supported by analysis of regional mean thickness and anomalies computed over the region shown in Fig. 1c (Table 2). For comparison, we also list results from the CICE-free model simulation. In November 2011, the different CS2 data products are in agreement that the ice was anomalously thin (-32 to -46 cm), the thinnest in the CS2 data record. Similarly, in November 2013, all three CS2 products show overall thicker ice in the order of +23 to +38 cm. The CICE-free simulations also show anomalously thinner and thicker ice during these years, but larger anomalies were simulated in 2012 and 2014.

While the overall pattern of years with anomalously thin or thick ice is broadly similar between the three CS2 products, this is not true in 2016. Both the CPOM and AWI thickness estimates suggest slightly thicker ice than average (+4 and +9 cm, respectively), while the NASA product suggests the ice pack was overall slightly thinner (-1 cm). The CICE-free run is in agreement with the NASA data set for the 2016 anomaly. Turning back to Fig. 3, we find that in 2016 the CPOM data set shows +20 to +60 cm thicker ice north of the Canadian Archipelago (CAA) and Greenland, -20 to -60 cm thinner ice on the Pacific side of the pole and +10 to +30 cm thicker ice north of the Laptev Sea. These spatial patterns of November 2016 SIT anomalies are broadly similar to those from AWI but less so with those from NASA. However, despite similar patterns of positive and negative thickness anomalies, AWI shows between +20 and +30 cm thicker ice over much of the central Arctic Ocean, and even thicker ice (up to +60 cm) north of the CAA and Greenland in November 2016, than the CPOM product. NASA, in contrast, shows larger negative anomalies on the Pacific side of the north pole of up to -70 cm and larger positive anomalies directly north of the CAA between +10 and +20 cm.

Since we use CPOM CS2 thickness fields to initialize our CICE model runs, this comparison is useful in determining whether or not the 2016 November thickness anomalies are robust in other CS2 processing streams and provides a measure of CS2 sea ice thickness uncertainty.

However, since we do not have the AWI and NASA ITDs we cannot quantify the impact of using a different thickness data set on our simulations. However, as a result of the negative winter ice growth feedback (discussed below), differences due to model initialization in November will be attenuated until April.

For a more robust analysis of winter ice growth during the record warm winter of 2016/2017, we now include April thickness estimates from CS2 (CPOM, AWI and NASA), the free CICE simulation and the CICE simulations initialized with CPOM CS2 November SIT in Fig. 4. Corresponding values for all other years are shown in Fig. 5 (CS2) and Fig. 6 (CICE). Table 3 summarizes associated mean April thickness and anomalies since 2011, together with contributions from thermodynamics (ice growth) and dynamics (ice transport and ridging) based on the CICE model simulations. The area for which these estimates are provided corresponds to the area shown in Fig. 1d.

We first note that all 5 estimates have different strengths and weaknesses: while the mean annual cycle of sea ice thickness should be more accurate from CS2 than modeled estimates, robust analysis of winter ice growth from CS2 is in part limited due to the impact of climatological snow depth assumptions, which may differ from one year to the next, and differences in waveform processing between CS2 data providers, which may result in inconsistencies in the magnitude and direction of the observed thickness anomalies. In the free CICE simulation, November sea ice thickness is less certain, due to error accumulation during the model run. In the initialized CICE simulation, both these error sources are reduced but inherent model biases remain. While we discuss some of the regional differences below, we are most confident in the model simulations on the Arctic basin-wide scale, over which CICE has been tuned to agree with CS2 winter ice growth.

Despite these limitations, all five approaches show good agreement in most years regarding the direction of the thickness anomalies (i.e., positive or negative) even if they disagree on absolute magnitude. For example, Arctic Ocean mean thickness anomalies are negative in all three CS2 products for April 2013 (ranging from -3 to -25 cm), whereas in April 2014 and 2015 all approaches give positive mean thickness anomalies, ranging from +5 to +20 cm in 2014 and +11 to +22 cm in 2015 (Table 3). In some years, the CICE-free simulation better matches the observed April thickness anomalies (e.g., 2013, 2015), whereas in other years CICE-ini performs better (e.g., 2012, 2014). In contrast, in 2011 and 2017 we find disagreement among the three CS2 data sets. In April 2011, both the CPOM and NASA product have overall negative thickness anomalies for the Arctic Basin (-4 and -8 cm, respectively), whereas they are positive in the AWI product (+7 cm). In April 2017, both the CPOM and AWI are in close agreement that the ice cover was overall thinner (-13 and -12 cm, respectively), as are the CICE-free and CICE-ini simulations (negative thickness anomalies of -13 cm), whereas NASA shows a weak positive anomaly (+3 cm).

Focusing more on April 2017, the three CS2 products suggest widespread thinner ice in April 2017 north of Ellesmere Island (up to -80 cm thinner) relative to the 2011–2017 mean (Fig. 4 top). Thinner ice is also found within the Chukchi and East Siberian seas (on average -10 to -35 cm thinner) despite a mix of positive and negative anomalies. CICE simulations, in contrast, show more widespread thinning throughout the western Arctic, including the Beaufort Sea and positive thickness anomalies north of Ellesmere Island (Fig. 4 middle and bottom). In the Beaufort Sea, there is general disagreement among the three CS2 products, as well as with the CS2 results and the CICE simulations: regional mean anomaly of -5 cm (CPOM), 0 cm (AWI), +20 cm (NASA), -25 cm (CICE-ini) and -30 cm (CICE-free). North of Ellesmere Island, CICE-ini indicates positive thickness anomalies (up to +50 cm), whereas all 3 CS2 products show negative thickness anomalies (up to -80 cm). In this region, the CICE-free simulation also shows mostly negative thickness anomalies (-20 to -80 cm), with a small positive area (up to +25 cm).

While the discrepancy in this region is puzzling, the bias between the CICE-ini simulations and the CS2 products may in part reflect the use of a snow climatology in the CS2 thickness retrievals. As discussed earlier, a positive sea ice thickness anomaly was found in the November 2016 CS2 thickness retrievals north of CAA and Greenland. Yet this positive thickness anomaly is not preserved through April in both the CPOM and AWI CS2 products. Figure 7 shows CICE simulated snow depth anomalies in November 2016 and April 2017. In November, small positive snow depth anomalies occur throughout the Arctic, especially north of the Queen Elizabeth Islands where the anomaly locally increases to 20 cm. By April, the anomalies cover a broader region and increase in magnitude. A positive April snow depth anomaly of 15 to 20 cm relative to W99 would result in an underestimation of the CS2-retrieved April ice thickness (SIT) by 88 to 115 cm using the following equation: SIT=ρsnowHsnow+ρwaterFc(ρwater-ρice), where Fc is the corrected radar freeboard (Fb) for the reduced propagation of the speed of light through the snow cover (Fc=Fb+0.25Hsnow) (Tilling et al., 2017), and using a snow density (ρsnow) of 320 kg m-3 (Warren et al., 1999), ice density (ρice) of 915 kg m-3 and water density of (ρwater) 1024 kg m-3. CICE-ini, which relies on the CPOM CS2 November thickness, maintains this positive thickness anomaly through April despite reduced thermodynamic ice growth. The CICE-free simulation, in contrast, started with negative thickness anomalies in November within this region, and maintains them through April.

Conversely, thickness is also strongly influenced by dynamics, such as convergence against the CAA and Greenland which leads to thicker ice in this region (Kwok et al., 2015). However, during winter 2017, the Beaufort High largely collapsed (Moore et al., 2018), reducing convergence against the northern CAA and Greenland (Fig. 8). One advantage of using CICE, is that we can more readily diagnose thermodynamic vs. dynamical contributions to the observed thickness anomalies. For the region directly north of Ellesmere Island, both the CICE-ini and CICE-free simulations support reduced sea ice convergence, leading to thinner ice from dynamical contributions. At the same time, this region also exhibited reduced thermodynamic ice growth in both CICE simulations. One would expect thermodynamic ice growth to be reduced in regions of enhanced snow depth and thicker November ice. Positive snow depth anomalies extended from this region through the northern Beaufort Sea, in agreement with extended regions reductions in thermodynamic ice growth in both CICE-free and CICE-ini. At the same time, regions of positive 2016 November thickness anomalies are also associated with regions of reduced CICE thermodynamic ice growth.

Overall, the largest reductions in thermodynamic ice growth during winter 2016/2017 occurred within the Chukchi Sea and north of the CAA, extending through the northern Beaufort Sea (in the order of -40 cm). While snow depth and thickness anomalies influenced thermodynamic ice growth north of the CCA, within the Chukchi Sea the negative ice growth anomalies were a result of late ice formation: ice formed a month later than the 1981–2010 mean within the Chukchi Sea. This seems to have been more important than increases in ice thickness from dynamics. Dynamical thickness changes simulated by CICE show an overall thickening of the ice in winter 2016/2017 within the Chukchi and Bering seas (up to 50 cm). Anomalous ridging in this region is in agreement with observed high amounts of deformation along the shore fast ice zone within the Chukchi Sea, as a result of persistent west winds from December to March (http://arcus.org/sipn/sea-ice-outlook/2017/june, last access: August 2017).

An exception to reduced thermodynamic ice growth occurs directly north of Utqiaġvik, Alaska (formerly Barrow), with positive thermodynamic ice growth anomalies of 30 to 40 cm. This enhanced ice growth was offset by ice divergence, leading to overall thinner ice in the CICE simulations. In situ observations of level first-year ice thickness off the coast of Utqiaġvik ranged between 1.35 and 1.40 m during May (http://arcus.org/sipn/sea-ice-outlook/2017/june, last access: August 2017) and appear to be in better agreement with the CICE simulations, as well as the CPOM and AWI CS2 thickness estimates, while the NASA CS2 product shows positive thickness anomalies in that region. Positive thermodynamic ice growth anomalies are also found for small regions north of Greenland and within Fram Strait, as well as within some scattered coastal regions of the Chukchi, East Siberian, Laptev and Kara seas.

Finally, large dynamical thickening was found within the Kara and northern Barents seas (up to 1.2 m) and to a lesser extent over the southern and western Greenland Sea, Baffin Bay and the Labrador Sea (not shown). The CICE-simulated dynamical thickening in the Barents and Kara seas is more anomalous than seen during previous CS2 years (Fig. 6), and likely reflects the influence of the positive Arctic Oscillation (AO) on ice motion (Fig. 8). The AO was positive from December through March, a pattern which results in offshore ice advection from Siberia and enhanced ice advection through Fram Strait (Rigor et al., 2002). This pattern leads to development of thin ice in newly formed open water areas, increasing thermodynamic ice growth in the Laptev Sea, whereas increased ice advection from thick ice regions north of Greenland towards Fram Strait, combined with changes in internal ice stress as the ice cover has thinned, leads to more deformation. Interestingly, while the CICE model runs confirm overall slightly thinner ice within the Barents Sea in April 2016, consistent with the studies by Ricker et al. (2017a) and Boisvert et al. (2016), the thinning from reduced thermodynamic ice growth was largely offset by thickening from dynamical effects (Figs. 5 and 6).

Overall, for the Arctic Basin as a whole, CICE simulations suggest the overall thinner ice observed in April 2017 is largely result of reduced thermodynamic ice growth (-11 to -13 cm), with dynamics adding +1 to +4 cm (Table 3).

Ice growth after the September minima is a result of turbulent heat flux exchanges between the relatively warm ocean mixed layer and the cold autumn and winter air through the snow-covered sea ice. Progressively, as the ice grows to about 1.5 to 2 m thick, the ocean becomes well insulated from the atmosphere and ice growth is slowed. Thus, it is not surprising that we see less thermodynamic ice growth in regions of relatively thick (> 2.5 m) November ice. A case in point is seen in winter 2013/2014 when thermodynamic ice growth was reduced by 9 to 10 cm, despite an overall colder winter.

Conversely, thinner ice regions generally exhibit more vigorous ice growth. For example, during winter 2012/2013, CICE-free, and to a lesser extent CICE-ini simulated thermodynamic ice growth increased throughout much of the Arctic Ocean in areas where the ice retreated in September 2012 (Fig. 6) and where the November 2012 thickness anomalies were negative (Fig. 3). This process of rapid winter ice growth over thin ice regions represents a negative feedback, allowing for ice to form quickly over large parts of the Arctic Ocean following summers with reduced ice cover and thinner November ice.

Thus, while summer sea ice is rapidly declining, several studies have indicated negative feedbacks over winter continue to dominate (e.g., Notz and Marotzke, 2012; Stroeve and Notz, 2015), allowing for recovery following summers with anomalously low sea ice extent, such as those observed in 2007 and 2012. This is further supported in the CICE-free simulations which show the least amount of winter ice growth for the Arctic Basin in 1989, and peak ice growth following the 2007 and 2012 record minimum sea ice extent (Fig. 9). As a result, mean ice growth from November to April in CICE simulations from 1985 to 2017 shows a positive trend that is weakly correlated to winter air temperatures or FDDs (R=0.49). Conversely, we find a strong inverse correlation (R=-0.82) between November sea ice thickness and winter ice growth. Thus, because thin ice grows faster than thick ice, there is an overall stabilizing effect that suggests as long as air temperatures remain below freezing, even if they are anomalously warm, the ice can recover during winter. This stabilizing feedback over winter means that major departures of the September sea ice extent from the long-term trend caused by summer atmospheric variability generally does not persist for more than a few years (Serreze and Stroeve, 2015).

However, since 2012, overall ice growth has declined as winter air temperatures have increased further. This not surprising in that there was a lot of new ice to form in the open waters left after the 2012 record minima. However, 2016 tied with 2007 for the second lowest Arctic sea ice minimum and overall thermodynamic ice growth was significantly less. The correlation from 1985 to 2012 is smaller than over the full record (R=0.34), suggesting a growing influence of warmer winter air temperatures though the difference in correlation is not statistically significant. While there remains a large amount of inter-annual variability in winter warming events, Graham et al. (2017) suggest a positive trend in not only the maximum temperature of these warming events, but also in their duration. Interestingly, there is a modest correlation between detrended FDDs and the winter maxima sea ice extent (R=0.30); not removing the trend results in a correlation of R=0.83. Thus, recent reductions in overall FDDs may have played a role in the last three years of record low maxima extents.