One of the longest records of landfast ice thickness and duration comes from several coastal stations throughout Canada that date back to the late 1940s, depending on the location. The data set is available online at the Canadian Ice Service (CIS) website (http://www.ec.gc.ca/glaces-ice/, last access: 30 October 2018; see Archive followed by ice thickness data). The thickness measurements are usually performed weekly from freeze-up to break-up, as long as it is safe to walk on the ice. For these reasons, the landfast ice duration at these stations, measured as the number of weeks with measurements, is always biased on the shorter side, possibly by a few weeks. From these station records, we selected the four sites in the CAA that had continuous records up to 2015: Alert, Eureka, Resolute and Cambridge Bay. From these weekly records available from 1960 to 2015, we extracted the landfast ice duration and springtime landfast ice thickness. A thorough analysis of these quantities as derived from these records was initially presented by Brown and Cote (1992) from 1957 to 1989 and recently updated to 2014 by Howell et al. (2016).

For additional ice thickness information we use ice thickness surveys in landfast regions of the CAA carried out by means of airborne electromagnetic induction (AEM) sounding in 2011 and 2015 previously described in Haas and Howell (2015). These surveys were averaged on a 25 km EASE 2.0 grid and are shown in Fig. S1 in the Supplement. We also use weekly ice thicknesses retrieved from CryoSat-2/SMOS in netCDF format for the years 2010–2016, obtained from data.scienceportal.de and remapped using a nearest-neighbour remapping to a 25 km EASE 2.0 grid. The resulting winter maximum sea ice thicknesses are shown in Fig. S2 in the Supplement.

In order to spatially map landfast ice we use the National Ice Center (NIC) ice charts products from the NSIDC (data set ID G02172) and ice charts from the Canadian Ice Service Digital Archive (CISDA). The NIC ice charts are available from 1972 to 2007 but we restrict the time period to 1980–2007 to be consistent with CISDA. Indeed, the CISDA provide ice information before 1980 but landfast ice was not explicitly classified. We refer readers to Tivy et al. (2011) (CISDA) and Yu et al. (2014) (NIC) for in-depth descriptions of ice chart data and their validity as a climate record. Following Galley et al. (2010), who also used the CIS ice chart data to map landfast ice, we consider grid cells with sea ice concentration of 10/10ths to be landfast. We defined pan-Arctic landfast extent using the NIC ice charts (given their larger spatial domain) as the regions that are covered by landfast ice for at least 1 month in the climatology. Both the NIC and CISDA ice charts were converted from shape files to a 0.25∘ latitude–longitude grid and then converted using a nearest-neighbour remapping to a 25 km Equal-Area Scalable Earth (EASE) 2.0 grid. We compute the number of months (equivalent to “percent of the year” in Galley et al., 2012) during which each grid cell was landfast for each time period from September to August.

Climate simulations and reanalyses do not provide a variable that explicitly characterizes landfast ice conditions. This makes it challenging to verify how it emulates landfast ice conditions compared to observations. To circumvent this limitation, we use daily sea ice thickness (hereafter, sit), sea ice concentration (hereafter, sic) and sea ice velocities (hereafter, usi and vsi) to synthetically characterize landfast sea ice conditions using the following procedure:

On the original model grid, we set the land mask to its nearest neighbour and remap using a nearest-neighbour remapping usi, vsi and sit to the sic native grid. Finally, we use a nearest-neighbour remapping to put all variables on a EASE 2.0 grid.

At each grid cell we then compute the number of months in each year with slow, packed ice. Using slow, packed ice is representative because we are interested in one specific aspect of landfast ice: the fact that its growth is primarily driven by thermodynamics and not by the import/export of sea ice. This procedure is used with the Pan-Arctic Ice-Ocean Modeling and Assimilation System (PIOMAS) (Zhang and Rothrock, 2003), a subset of the highest-resolution models (see Table 3, Storto et al., 2011; Forget et al., 2015; Haines et al., 2014; Zuo et al., 2015; Masina et al., 2015) from the ORA-IP (Balsameda et al., 2015; Chevallier et al., 2017). Finally, we use the CESM-LE and CMIP5 models to analyze climatological landfast ice extent and thicknesses. Some ORA-IP models (ORAP5.0, UR025.4) do not provide daily output. For these models, monthly data were first interpolated to daily frequency and from then on the analysis was performed using the procedure described above. It should be noted that sea ice velocities are not provided by all models and only for a few simulations, constraining the scope of the intercomparison presented here (see available models in Table 1). The data for this study were retrieved from the ESGF using the cdb_query tool (https://github.com/laliberte/cdb_query, last access: 9 April 2017). Finally, the 1980–2005 historical experiment followed by the 2006–2015 Representative Concentration Pathway 8.5 (RCP85) experiment (Taylor et al., 2012) are used with daily sea ice velocities, thickness and concentration.

In the summer, the sea ice concentration drops below 100 % for some models but it stills remains relatively high throughout the melt season. In these models (e.g. NorESM1 and ACCESS1.0), the reduction in summer ice concentration is not associated with increased sea ice speed (i.e. close to 0 correlation between the two variables over a year), unlike in the PIOMAS product, where a strong anti-correlation between the two variables can be measured. This suggests that these models may indeed have an ice concentration below 100 % during the summer but the import/export of sea ice remains quite limited because the packed ice never becomes mobile enough in narrow channels, particularly within the CAA. As a result, one must allow for some flexibility in the definition of packed ice in modelled products and a number below 100 % needs to be chosen as a cut-off. Here, we have chosen 85 % because (i) it represents landfast ice that ice grows according to thermodynamics and not because of export/import and (ii) it is widely accepted that in historical observational products a 15 % uncertainty in sea ice concentration is to be expected. Since we are using historical observation products in our comparison, we argue that the same 15 % uncertainty should be used when assessing model behaviour. We acknowledge that, by using an 85 % ice concentration to define packed ice, the lead fraction could be large at the boundary of the slow, packed ice, due to the proximity of mobile ice. In these regions, the argument presented above might break down. In this work, we will primarily focus on archipelagoes and marginals seas where this is not an issue. It is, however, important to keep in mind that, for applications that focus on those boundary regions, this criterion might be too lenient.

The models listed above do not represent the grounding of pressure ridges. Hence, they are not expected to perform well in regions where grounding is known to be an important mechanism for the formation and stabilization of a landfast ice cover. Observations show that grounding is important in the Laptev Sea (Haas et al., 2005; Selyuzhenok et al., 2018), in the Beaufort Sea (Mahoney et al., 2007) and in the Chukchi Sea (Mahoney et al., 2014). Nevertheless, these models can simulate landfast ice in some regions because the models dynamics take into account the aforementioned mechanical interactions. For most of these sea ice models, ice interactions are represented by a viscous–plastic rheology with an elliptical yield curve (Hibler, 1979).

Recently, a basal stress parameterization representing the effect of grounding was developed (Lemieux et al., 2015). This parameterization calculates, based on simulated ice conditions, the largest ridge(s) at each grid point. When these subgrid-scale ridge(s) are able to reach the sea floor, a basal (or seabed) stress term is added to the sea ice momentum equation. This grounding scheme clearly improves the simulation of landfast ice in regions such as the Alaskan coast, the Laptev Sea and the East Siberian Sea. However, in deeper regions such as the Kara Sea, Lemieux et al. (2015) pointed out that their model systematically underestimates the area of landfast ice. As the grounding scheme is less active in these deeper regions, Lemieux et al. (2016) modified the viscous–plastic rheology to promote ice arching.

Following the work of Lemieux et al. (2016), we conducted simulations that combined the grounding scheme and a modified viscous–plastic rheology. We used the optimal parameters k1=8 and k2=15 Nm-3 for the grounding scheme (Lemieux et al., 2015). Given a certain mean thickness in a grid cell and a concentration, the grounding scheme determines whether the parameterized ridges reach the sea floor or not (which depends on k1) and defines the maximum seabed stress that can be sustained by the grounded ridges (which is proportional to k2). As opposed to the standard elliptical yield curve, the ellipse aspect ratio is set to 1.5 (instead of 2) and a small amount of isotropic tensile strength is used (kt = 0.05).

For these simulations, we used the ocean model NEMO version 3.1 and the sea ice model CICE version 4.0 with code modifications to include the grounding scheme and to add tensile strength (Lemieux et al., 2016). Our 0.25∘ grid is a subset of the global ORCA mesh. It covers the Arctic Ocean, the North Atlantic and the North Pacific. This ice-ocean prediction system, which includes tides, was developed as part of the CONCEPTS (Canadian Operational Network of Coupled Environmental PredicTion Systems) initiative. We refer to our 0.25∘ model set-up and simulations as CREG025 (CONCEPTS-regional 0.25∘).

Note that, while adding the tides to our ice-ocean prediction systems, we found that unrealistic sea thicknesses developed in late winter in tidally active regions (e.g. Foxe Basin). To mitigate this problem, the Hibler (1979) ice strength parameterization is used as opposed to the default Rothrock (1975) formulation. The ice strength parameter P* was set to 27.5 k Nm-2 for our CREG025 simulation.

The sea ice model was initialized with sea ice thicknesses and concentrations from the GLORYS2V1 ocean reanalyses. The ocean model was initialized by the World Ocean Atlas (WOA13) climatology and forced at open boundaries by GLORYS2V3 (Ferry et al., 2010; Chevallier et al., 2017). A spin-up from October 2001 to September 2004 was performed. Free runs (no assimilation) are then restarted from the fields in September 2004 and conducted up to the end of 2010. The simulation was forced by 33 km Environment Canada atmospheric reforecasts (Smith et al., 2014).

The CAA is almost entirely covered by landfast ice for up to 8 months of the year (i.e. November to July) (Canadian Ice Service, 2011) and is therefore a useful region in which to begin evaluating a model representation of landfast ice duration and thickness. Figure 1 illustrates the relationship between landfast ice thickness and duration within the CAA for the observed data sets (e.g. CryoSat-2, AEM and in situ) in addition to PIOMAS and CREG025. When combining these heterogeneous data sources, a general picture of their representativeness of ice thickness over landfast ice duration emerges. Based on in situ observations landfast ice within the CAA lasts from 4 to ∼9 months and grows to ∼2 m, which is in agreement with previous studies (e.g. Brown and Cote, 1992; Canadian Ice Service, 2011; Howell et al., 2016). For PIOMAS, CREG025 and CryoSat-2 ice thickness standard deviations are close to the variability observed at the in situ locations. However, very thick ice upwards of ∼4 m is encountered at the 95th percentile in both the CryoSat-2 and the PIOMAS data when the landfast ice lasts for more than 9 months. These very stable and thick landfast conditions are the result of large multi-year ice floes, thus increasing the average ice thickness. It has long been known that MYI forms in situ within the CAA, and very thick MYI from the Arctic Ocean is also advected into the region (e.g. Melling, 2002), which is evident from the airborne EM measurements thickness values (Haas and Howell, 2015). This mix of ice types makes it challenging for models to represent ice thickness within the CAA but, overall, they are in reasonable agreement with observations.

The spatial distribution of annual landfast duration from observations (CIS and NIC), PIOMAS and selected ocean reanalysis models is shown in Fig. 2. Both ice charts products (CIS and NIC) show similar landfast ice extents and durations in the CAA (Fig. 2a and b). This landfast ice extent is also very similar in the two ice chart products over their regions of overlap (Fig. 2a and b, magenta curve). In PIOMAS, the duration of slow and packed ice conditions compares relatively well to the overall landfast extent and duration in the ice chart products (Fig. 2c). There is, however, too much of the slow and packed ice in the Beaufort Sea but too little in the Laptev and Kara seas. Most ocean reanalysis products have a suitable representation of slow, packed ice conditions in the CAA, the notable exception being CGLORS and UR025.4 (not shown). In the CGLORS case, the ice component appears to still be in spin-up at the beginning of the integration period because there is an unphysical interannual variability in the first few years of the simulation, and therefore results should not be expected to conform to observations. In the UR025.4 case, winter ice is packed but is too mobile in the Parry Channel and the M'Clintock Channel.

The spatial distribution of annual landfast ice duration in CMIP5 models with higher resolution is illustrated in Fig. 3b–h. These models exhibit a reasonable slow, packed ice extent and duration but it is mostly confined to the CAA (Fig. 3b–h). The exception is the MRI-ESM1 (and applies to the other models from the MRI), which simulates slow, packed ice conditions year-round across the Arctic (not shown). This is likely due to its sea ice being modelled as a simple viscous fluid without a sophisticated rheology. Compared to the NIC analyses, all the CMIP5 models and reanalyses do not have enough months of landfast ice on the Russian coast. GFDL-ESM2G , CESM-LE and PIOMAS are the ones that provide the best landfast ice simulation in the Laptev, Kara and East Siberian seas (Figs. 2c and 3d, f). Another important feature of the import/export of sea ice in coupled models (ACCESS 1.0, CESM-LE, GFDL-ESM2G) seems to be the tendency for many of them to emulate year-round or close to year-round landfast ice in the Parry Channel regions of the CAA (Fig. 3d, f, ACCESS 1.0 not shown). This is peculiar, since this would mean that ice likely takes years to transit through the Parry Channel, allowing thermodynamic forcing to create very thick ice in a region. Note that, in the remaining models, the MIROC5 and MPI-ESM-MR both emulate a landfast ice duration in the Parry Channel that is too short (Fig. 3c, e).

The largest observed negative trends in landfast ice duration of up to 1 month decade-1 is found primarily in the East Siberian Sea but a general negative trend is located across the Arctic (Fig. 4a, b), as also reported by Yu et al. (2014). In the CAA, trends are larger in the NIC ice charts but both the CIS and NIC show relatively weak decline in duration in the Parry Channel and the M'Clintock Channel. These relatively small trends are in stark contrast with the very large trends almost everywhere in the CAA in the PIOMAS simulations. For CGLORS, the model with sea ice still in spin-up, there is a large positive increase in slow, packed ice duration (not shown). Such increases are also seen in the Beaufort Sea in the GLORYS2V3 reanalysis, indicating that towards the end of the reanalysis the Beaufort Sea is covered by slow, packed ice for a few months per year (Fig. 4f). This is in complete disagreement with observations and mandates that extra care should be taken when using this product to analyze the import/export of sea ice in the Beaufort Sea. In summary, reanalysis products appear to have a particularly difficult time reproducing the long-term stability of the slow, packed ice distribution, suggesting that targeted efforts to improve this aspect of their import/export of sea ice are likely necessary.

CMIP5 models sea ice simulations (except the MRI models for the reason explained above), on the other hand, fare relatively well at representing negative trends in landfast ice duration when compared to observations (Fig. 5). Most models tend to show an enhanced disappearance of slow, packed conditions along the Beaufort Sea edge of the CAA and declines that are in general agreement with observation in the Parry Channel. One exception is the CESM-LE where some year-round slow, packed ice conditions do not decline over the 1980–2015 period (Fig. 5d). The models with less slow, packed ice than in observations, MIROC5 and MPI-ESM-MR, show relatively strong declines that, if they continued, would indicate an almost complete disappearance of slow, packed ice by the middle of the 21st century.

We now take a closer regional examination of landfast ice extent in the CAA, Northwest Passage (Parry Channel route) and Laptev seas. These regions are expected to experience increases in shipping activity from the middle to late 21st century according to model simulations (Smith and Stephenson, 2013; Melia et al., 2016). Instead of using an absolute measure of extent, we report extent as a fraction of the ocean surface within the bounds of the NIC 1-month duration landfast ice extent climatology (magenta line in Fig. 2b). This approach is necessary to appropriately compare observations to models that represent the islands and channels of the CAA differently.

Over the 1980–2015 time period, landfast ice extent has declined dramatically for durations longer than 5 months with a marked decline in the extent of landfast ice with a 7 to 8 months duration within the Northwest Passage (Fig. 6). What is, however, striking is how the extent of landfast ice extent with duration of 5 months or less has been mostly constant over the last 35 years (Fig. 6). It is because of this observation that we have not included a trend analysis in Fig. 6. If the trend in landfast area depends so strongly on landfast ice duration, it would probably be misleading to attribute a hard number to the decline in landfast ice. If sea ice-albedo feedback is an important player in recent sea ice decline (e.g. Perovich et al., 2007) then it is not entirely surprising that during the polar night landfast ice conditions re-establish themselves year after year, even in the context of rapid Arctic warming. Finally, it is also worth noting that Fig. 6a indicates that the small amounts of multi-year landfast ice within the CAA have virtually disappeared in recent years (i.e. the 11-month line has been at 0 since 2002), consistent with Alt et al. (2006).

Landfast ice extent in the CAA and Northwest Passage is well represented in the PIOMAS data assimilation product as it compares well with the CIS and NIC ice chart products, although the NIC product does exhibit stronger interannual variability (Fig. 7a, b). In the Laptev Sea, PIOMAS clearly underestimates the area of landfast ice when compared to the NIC ice charts (Fig. 7c). This is likely due to the fact that PIOMAS does not represent the effect of grounding, an important mechanism for the formation and stability of the Laptev Sea landfast ice cover (Selyuzhenok et al., 2017). Despite this area of landfast ice in the Laptev Sea being too small, PIOMAS exhibits a decline of ∼25 % of the landfast extent over the last 35 years, which is consistent with the one from the NIC ice charts.

Comparing current (1980–2015) to projected (2070–2080) landfast ice extent from CMIP5 in these regions reveals considerable changes, which are summarized in Table 1. The seven models with the smallest extent of 1979–2015 CAA slow, packed ice (ACCESS1.0, ACCESS1.3, BCC-CSM1.1(m), GFDL-CM3, MIROC5, MPI-ESM-LR, MPI-ESM-MR) lose most of it by 2070–2080, while the four models with a large extent of 1979–2015 CAA slow, packed ice (CESM-LE, GFDL-ESM2G, GFDL-ESM2M, NorESM1-M) retain most of it by 2070–2080. As mentioned earlier, two models have unrealistic behaviour (MIR-ESM, MRI-CGCM3) because their sea ice model is based on a simple perfect fluid.

Looking specifically in the CAA, current conditions (Fig. 8a, black) indicate that the CMIP5 distribution is trimodal: one mode has an extent comparable to observations (at 0.6 to 0.8 of NIC extent), a second mode has a much lower extent (at 0.2–0.6 of NIC extent) and a third mode has an extent that covers most of the area (∼1.0 of NIC extent). In the CAA, this trimodal distribution yields a bimodal distribution in 2070–2080 projections (Fig. 8a, yellow): one mode still has an extent comparable to observations and a second mode has virtually no 5-month landfast ice extent left. In the Northwest Passage, the story is much simpler (Fig. 8b). All considered models are entirely covered with slow, packed ice conditions at least 5 months every year for their historical simulations but in the 2070–2080 projections about half become devoid of it, while the other half retain their historical conditions. This highlights the difficulty of projecting how the import/export of sea ice will react to anthropogenic forcing in the narrow channels of the CAA. Finally, in the Laptev Sea, almost all considered models have little slow, packed ice now and by 2070–2080 (Fig. 8c).

The picture is generally clearer for the CESM-LE. In that model, the CAA and the Northwest Passage has slow, packed ice comparable to observations (Fig. 8d, e). In the projection, the CAA is expected to lose only 0.2 of its slow, packed ice coverage and almost none in the Northwest Passage. In the Laptev Sea, the CESM-LE is only performing marginally better that the CMIP5 multi-model ensemble and the projection shows the complete disappearance of 5-month slow, packed ice by 2070–2080 (Fig. 8f).

When we look at ice thickness, models show a wide range of ice thicknesses over landfast ice during the 1980–2015 period for all regions (Fig. 9a–c). However, for the 2070–2080 period they are essentially in agreement, indicating that in all three regions landfast ice thickness is found to grow between 1 and 2 m over the cold season (Fig. 9a–c). Moreover, the projections indicate about a 0.5 m decrease in landfast ice thickness towards the end of the 21st century. A similar growth range is apparent when just looking at the CESM-LE but there is, however, a larger magnitude of projected thickness decreases towards the end of the 21st century (Fig. 9d–f).

The results we have presented so far have been focused on high-resolution observational data sets, 25 km resolution reanalyses and coarser climate models. From these different data sources we were able to demonstrate the capabilities and limitations at emulating landfast ice conditions of models of the current generation. In the remainder of this section, we will look at our 6-year CREG025 simulations and see the benefits of using landfast ice parameterizations.

As evident in Fig. 10, the CREG025 simulations show quite an accurate representation of landfast ice duration in the Laptev Sea, the East Siberian Sea and along the Alaskan Coast where grounding is crucial for simulating landfast ice (Lemieux et al., 2015). The overestimation of landfast ice north of the CAA is likely a consequence of our imperfect criterion for determining whether the ice is landfast or not (slow-drifting ice for a NIC analyst can be identified as landfast in our study).

Overall, in the CAA, the CREG025 landfast ice duration is in good agreement with the ones of the NIC and CIS (Fig. 2a, b). In both NIC and CIS products, the duration of landfast ice is small in tidally active regions such as the Gulf of Boothia, Prince Regent Inlet, Lancaster Sound and Foxe Basin. In accordance with observations, the CREG025 simulation (which includes explicit tides) exhibits mobile ice in these regions throughout the winter (Fig. 10b). However, CREG025 underestimates the landfast ice duration in Barrow Strait and north of M'Clintock Channel.

We are currently conducting a thorough investigation of the impact of tides (and the mechanisms involved) on simulated landfast ice. This will be the subject of a future publication. Preliminary results suggest that including tides is crucial to properly simulate landfast ice in certain regions of the CAA. We speculate that the fact that many models (e.g. GFDL-ESM2G, CESM-LE, PIOMAS) presented in this paper, overestimate landfast ice in parts of the CAA (e.g. Gulf of Boothia and Prince Regent Inlet) is due to the absence of tides in their simulations.

Looking at time series of 5-month landfast ice extent, the CREG025 simulation follows observations very closely in the CAA and Laptev Sea (Fig. 7a, c). In the Northwest Passage, however, the CREG025 simulation leads to too little landfast ice (again due to the underestimation of landfast ice in Barrow Strait and north of M'Clintock Channel). This could be due to the fact that our CREG025 simulation seems to have ice thinner (and therefore weaker) than observations (see Fig. 1). Overall, however, landfast ice extent in CREG025 is much more in line with observations in all three regions than most Earth system models (shown in Fig. 8).