Landfast ice thickness and corresponding snow depth measurement have been made regularly at many coastal stations throughout Canada since about 1950. These data are quality controlled and archived at the Canadian Ice Service (CIS) and represent one of the few available sources of continuous ice thickness measurements in the Arctic. In general, thickness measurements are taken once per week, starting after freeze-up when the ice is safe to walk on and continuing until breakup or when the ice becomes unsafe. Complete details of this data set are provided by Brown and Cote (1992). The data set is available on the CIS web site (http://www.ec.gc.ca/glaces-ice/, see Archive followed by Ice Thickness Data). Four sites in the CAA were selected for study: Alert, Eureka, Resolute and Cambridge Bay (Fig. 1). Although there are other sites in the database, these sites are the only ones than span the same 55-year period between 1960 and 2014. The record at Mould Bay, used in BC92, terminated in the early 1990s. Together these sites cover ∼ 20∘ in latitude (Fig. 1) adjacent to an area of thick Arctic sea ice that experienced the highest thinning in recent years (Kwok and Rothrock, 2009; Laxon et al., 2013). Values of maximum or end-of-winter ice thickness and corresponding snow depth during the ice growth season were extracted from the weekly ice and snow thickness data at the selected sites (see Supplement). As this study is concerned with annual variability in maximum ice thickness, the main period of interest extends from September to late May.

The other source of observed data used in this study was Environment Canada's monthly mean air temperature records at Alert, Eureka, Resolute and Cambridge Bay for which a complete description is provided by Vincent et al. (2012).

The representation of CAA landfast sea ice thickness within the Coupled Model Intercomparison project phase 5 (CMIP5) is analysed using the 1850–2005 Historical experiment followed by the 2006–2099 Representative Concentration Pathway 8.5 (RCP85) experiment (Taylor et al., 2012) (Table 1). Monthly sea ice thickness (variable sit), sea ice concentration (variable sic), 2 m temperature (variable tas) and snow depth (variable snd) were used. The CMIP5 data were retrieved from the British Atmospheric Data Centre database and accessed through the Centre for Environmental Data Analysis (www.ceda.ac.uk). Ensemble r6i1p1 and r7i1p1 from model EC-EARTH were removed because of corrupted data. We obtain the multi-model mean of trends and their statistical significance at each grid point by creating the distribution of trends through a Monte Carlo simulation. We use a t distribution for the interannual variability and build a noise model to account for internal variability as in Swart et al. (2015) and Laliberté et al. (2016). We obtain the multi-model mean of Pearson correlations and their statistical significance by first performing a Fisher transform and then applying the same method as for the trends. The inverse Fisher transform is applied after obtaining the multi-model mean and its significance. See the appendix for a complete description of the method.

We also investigate ice thickness values from a selection of the highest-resolution models (Storto et al., 2011; Forget et al., 2015; Haines et al., 2014; Zuo et al., 2015; Masina et al., 2015), from the Ocean Reanalysis Intercomparison (ORA-IP) (Balmaseda et al., 2015; Chevallier et al., 2016) (Table 2) and from the Pan-Arctic Ice-Ocean Modeling and Assimilation System (PIOMAS) (Zhang and Rothrock, 2003). Supporting 2 m temperature data were obtained from ERA-Interim (Dee et al., 2011).

The average behaviour of landfast ice at the four sites over the 50+-year record is summarized in Table 3. Ice growth, approximately linear through most of the season, slows after March (Fig. 2). Ice thickness reaches a maximum of ∼ 2–2.3 m by late May at all sites. Values are consistent with those reported by BC92 and with recent observations of Melling et al. (2015) and Haas and Howell (2015). The standard deviations are nearly uniform (at ∼ 0.2 m) across all sites, giving a relatively low coefficient of variation (CV; a measure of relative dispersion defined as the ratio of the standard deviation to the mean) of ∼ 0.1. The thickest ice is found in Eureka with a 1957–2014 mean of 2.27 m, which is likely due to climatologically lower air temperatures in the fall and winter (Table 3).

Snow depth also appears to grow linearly through the season, peaking in May, but unlike ice thickness the monthly variability is high (CV ∼ 0.4) (Fig. 3). Mean October to May snow depths at Resolute, Eureka and Alert range from ∼ 18 to 23 cm compared to only ∼ 8 cm at Cambridge Bay (Table 3). The rapid buildup of the snow cover due to storms in the fall and early winter, which is evident over the Arctic Ocean multi-year ice cover (Warren et al., 1999; Webster et al., 2014), is not seen in these snow depth records within the CAA. The linear behaviour in snow depth is likely maintained by continuous wind-driven redistribution and densification throughout the ice growth season (BC92; Woo and Heron, 1989).

The time series of maximum ice thickness at Cambridge Bay, Resolute, Eureka and Alert are illustrated in Fig. 4 and summarized in Table 1. Statistically significant (95 % or greater confidence level) negative maximum ice thickness trends are present at Cambridge Bay (-4.31 ±1.4 cmdecade-1), Eureka (-4.65 ±1.7 cmdecade-1) and Alert (-4.44 ±1.6 cmdecade-1) (Table 1). A slight negative trend is present at Resolute but not statistically significant at the 95 % confidence level (Table 1). Over the 50+-year record, the ice thinned by ∼ 0.24–0.26 m at Cambridge Bay, Eureka and Alert with essentially negligible change at Resolute. These trends in the CAA are similar to trends on the Siberian coast (-3.3 cmdecade-1) (Polyakov et al., 2010) but lower in magnitude compared to the Barents Sea (-11 cmdecade-1) (Gerland et al., 2008).

For the shorter record (late 1950s–1989, ∼ 30 years) investigated by BC92 there was a negative trend at Alert (-7.1 cmdecade-1), no evidence of a trend at Eureka and a positive trend at Resolute (10 cmdecade-1), but only the positive trend at Resolute was statistically significant at the 95 % or greater confidence level. Our results from the present 50+-year record suggest that the negative trend at Alert is robust and the trend at Eureka is now negative and significant. The trend at Resolute is now slightly negative, but it is not statistically significant.

Typically, ice thickness reaches its maximum in late May with trends toward earlier dates of maximum ice thickness present at all sites (significant at Resolute, Eureka and Alert; Table 3). The significant trends are between -2.0 ±0.1 daysdecade-1 at Eureka and -6.2 ±1.5 daysdecade-1 at Resolute. At Resolute, the date of maximum ice thickness is now on average more than a month earlier than the early 1960s, although this is not reflected in the trend in ice thickness. Freeze onset at these sites also increases at ∼ 3–6 daysdecade-1 (Howell et al., 2009) and demonstrates a shortened growth season at Resolute, Eureka and Alert. Together, the trends of ice thickness and their recorded dates suggest a systematic thinning of landfast ice at Cambridge Bay, Eureka and Alert.

The variability of landfast thickness at these Arctic sites was previously found to be largely driven by interannual variations in snow depth and air temperature (BC92; Flato and Brown, 1996). With the 50+-year record at the four sites, we can examine the corresponding linkages to snow depth and temperature.

For snow depth, the only trend that is statistically significant at the 95 % confidence is Cambridge Bay at -0.8 ±0.4 cmdecade-1 (Table 3, Fig. 5). In contrast, BC92 found a significant positive trend at Alert (4 cmdecade-1), a trend of low significance in Eureka and a negative and significant trend at Resolute (-3.3 cmdecade-1). Looking at the detrended correlations (r) between snow depth and ice thickness reveals the strongest correlation at Resolute (r= -0.71) followed by Eureka (r= -0.66), Alert (r= -0.47) and Cambridge Bay (r= -0.31). Figure 6 provides evidence from extreme years of the role of deeper snow inhibiting ice growth compared to thinner snow, but the positive trends in snow thickness are not significant at Resolute, Eureka and Alert. This may in part be due to the single point-wise snow depth and ice thickness measurements made at each point in time, which fail to capture spatial heterogeneity in the snow depth–ice thickness relationship.

With respect to observed temperature, we find significant warming trends in the spring and fall at all sites over the 50+-year record (Table 3; Fig. 7). Significant warming is also present at all sites in the summer except Resolute and at all sites during the winter except Eureka (Table 3). Warming is highest during the fall, at ∼ 0.6 ∘Cdecade-1 at all sites (Table 3). The detrended correlation between temperature (winter, spring, summer and fall) and maximum ice thickness is weak at all sites. For example, the strongest detrended correlation between maximum ice thickness and temperature (winter and spring) is found at Cambridge Bay during the winter and spring but is only ∼ 0.4.

Also of interest is that the observed temperature trends over this period differ considerably from the earlier period investigated in BC92, in which they reported cooling at all the sites, with a significant cooling trend at Eureka. It was noted that the general cooling over their record coincided with the 1946–1986 cooling trend over much of the eastern Arctic and northwestern Atlantic reported by Jones et al. (1987). This cooling trend halted during the 1980s and the warming, seen in the current and longer record, has resumed (Jones et al., 1999). Arctic land areas have experienced an overall warming of about ∼ 2 ∘C since the mid-1960s, with area-wide positive temperature anomalies that show systematic changes since the end of the 20th century, which continued through 2014 (Jeffries and Richter-Menge, 2015). Recently, warming in Canadian Arctic regions was found to be greater than the pan-Arctic trend by up to 0.2 ∘Cdecade-1 (Tivy et al., 2011).

In order to compare seasonal cycles and trends in landfast ice thickness and snow depth between models and observations, we limit our comparison to models with a reasonable representation of the CAA, i.e. those with an open Parry Channel (i.e. bcc-csm-1-1, bcc-csm-1-1m, CNRM-CM5, ACCESS1-0, ACCESS1-3, FIO-ESM, EC-EARTH, inmcm4, MIROC5, MPI-ESM-LR, MPI-ESM-MR, MRI-CGCM3, CCSM4, NorESM1-M, NorESM1-ME, GFDL-CM3, GFDL-ESM2G, GFL-ESM2M, CESM1-BCG, CESM1-CAM5, CESM-WACCM). In these models, sufficient spatial resolution allows us to find sample points that are almost collocated to in situ observation locations. The sample points were determined by finding the closest ocean grid point where the sea ice is packed for a good portion of the year, but not all year. Grid points with this characteristic therefore share the most important feature of the landfast ice at our observations locations: it is not perennial. Mathematically, we sought sample points where the sea ice concentration is on average above 85 % for more than 1 month but less than 11 months over the 1955–2014 period. The Eureka site is, however, particularly challenging for models because it lies deep in a very narrow channel, which is only resolved by the MPI-ESM-MR in the CMIP5. As a result, for most models, the sample point for Eureka is located on the western shore of Ellesmere Island. This is a consequence of using samples because some models either do not resolve some of the channels in the CAA or have too perennial packed ice cover (e.g. CESM1-CAM5), and then the sample points are further from the observational site than would be desired. We chose to use sample points in our comparison to observations instead of using regional averages for two main reasons. The first reason is that using regional averages would have lumped together different ice dynamics regimes that should not necessarily be expected to compare well to point observations on landfast ice. The second reason is that we are of the opinion that the resolution in many of these models is sufficiently high to warrant such a direct comparison and provides a better benchmark than regional averages for landfast ice modelling in the CAA.

The seasonal cycle (1955–2014) of median ice thickness from CMIP5 (black), ORA-IP models CGLORS, ORAP5.0 and GLORYS2V3 (blue), ECCO-v4 (green) and UR025.4 (red) is shown in Fig. 8. ORA-IP models have been split into three groups based, respectively, on their high, medium and low ice thicknesses at Alert. Ice thickness from CMIP5 is comparable to observations (Fig. 2) at Cambridge Bay and Resolute with maximum ice thickness reaching 200 cm. The ORA-IP models are less consistent. ECCO-v4 tends to have thicker sea ice than observations at Cambridge Bay, Resolute and Eureka but thinner at Alert. CGLORS, ORAP5.0 and GLORYS2V3, however, are comparable to observations at Cambridge Bay, Resolute and Eureka but have extremely thick and perennial ice close to Alert.

The seasonal cycle (1955–2014) of median snow depth from CMIP5 is shown in Fig. 9. CMIP5 models indicate a linear increase similar to observations reaching a maximum of ∼ 20 cm in April or May. This is lower than the observed maximum at Resolute, Eureka and Alert but is about twice as much as at Cambridge Bay. While the snow depth reaches 0 during the summer at Eureka and Alert in models, the sea ice thickness does not (Fig. 8), unlike in observations. This likely reflects the fact that the grid cell thickness in sea ice models with thickness classes a represents the average thickness over these classes. In August the thinner ice classes might have melted but thicker ice classes can still be found, resulting in a substantial average ice thickness over the grid cell. The seasonal cycle over packed ice in these models thus gives a reasonable representation of the seasonal cycle over landfast ice in the CAA, especially in the southern region of the CAA. Overall, this comparison shows how recent improvements in sea ice model resolution allows comparisons with observations that required dynamical downscaling techniques in the previous generation of sea ice models (i.e. Dumas et al., 2005; Sou and Flato, 2009).

Despite relatively high spatial resolution, PIOMAS does not resolve seasonal ice thickness along the coasts and within the very narrow channels within the CAA (not shown). As a result, Cambridge Bay and Resolute Bay sites represent the only long-term monitoring sites within the CAA suitable for comparison since PIOMAS. The monthly time series of PIOMAS ice and snow thickness estimates at Cambridge Bay and Resolute is shown in Fig. 10. The seasonal cycle of ice growth at Cambridge Bay and Resolute is representative compared to observations (Fig. 2) but PIOMAS estimates retain more ice in August and September, particularly at Resolute. Ice growth reaches a maximum in April at Cambridge and in May at Resolute which is 1 month earlier compared to observations. Snow depth follows a linear increase similar to observations (Fig. 3), with good agreement at Cambridge Bay, but considerably underestimates snow depth at Resolute (Fig. 10). Schweiger et al. (2011) performed a detailed comparison of PIOMAS ice thickness values against in situ and Ice, Cloud, and land Elevation Satellite (ICESat) ice thickness observations and found strong correlations. They determined a root mean square error (RMSE) of ∼ 0.76 m and noted that PIOMAS generally overestimates thinner ice and underestimates thicker ice. At both sites within the CAA, PIOMAS ice thickness data are in reasonably good agreement with in situ observations with RMSEs of 0.29 cm at Cambridge Bay and 0.68 cm at Resolute (Fig. 11). The systematic overestimate of thinner ice reported by Schweiger et al. (2011) is more apparent at Resolute than Cambridge Bay (Fig. 11). The higher-latitude regions of the CAA where there is an intricate mix of seasonal first-year ice and multi-year ice is a problem for PIOMAS and thus contributes to the larger discrepancy at Resolute compared to Cambridge Bay.

The spatial distribution of maximum sea ice thickness trends from ORA-IP and CMIP5 is illustrated in Fig. 12. The CMIP5 model mean exhibits a fairly uniform trend pattern, consistent with the different in situ observations (Fig. 4) but with overestimated negative thickness trends. Although for individual models this pattern is far from uniform, the general pattern and magnitude of thickness trends tend to be roughly in accordance with temperature trends (not shown). A similar behaviour is observed in the ORA-IP models, with the notable exception of CGLORS, where positive thickness trends are found almost everywhere (Fig. 12a). This is robust and it appears that the model is not completely equilibrated in the CAA and exhibits large month-to-month adjustments. Model ORAP5.0 also is not completely equilibrated in the region for years 1979–1984. During those years, it exhibits large interannual changes in thickness. For this reason, we are only considering years 1985–2013 for this model.

For PIOMAS, the north–south overestimated trend is also present (not shown) as with CMIP5 and ORA-IP. Looking specifically at trends computed from 1979 to 2014 near the observed sites shows the mean maximum ice thickness linear trend from at Cambridge Bay to be -13.4 +3.4 cmdecade-1, which is almost double the observational trend of 6.2±2.4 cmdecade-1. At Resolute, the PIOMAS linear trend is 24.0±4.1 cmdecade-1, which is considerably stronger than the observational trend of -4.9 ±3.51 cmdecade-1.

Even though ORA-IP models have unrealistically large thickness trends, the pattern of interannual correlation (detrended) between winter temperatures and thicknesses is roughly consistent across models (Fig. 13). Some ORA-IP models also experience positive correlations (e.g. CGLORS, ORAP5.0, GLORYS2V3 and UR025.4) that are mostly located north of the CAA or within the CAA in regions where multi-year ice is known to be present. It is possible that warmer temperatures are associated with an increased flux of thicker multi-year ice into the CAA, which is known to occur (e.g. Howell et al., 2013), but the driving processes responsible for these positive correlations require more investigation. In CMIP5 models, no model exhibits positive correlations with temperature that resembles ORA-IP models over the CAA. Although the time series for the ORA-IP models is short and the positive correlations are only statistically significant at a few grid points in CGLORS and UR025.4, this behaviour is sufficiently problematic to recommend that care should be taken when using these ORA-IP models to study the interannual variability in the Canadian Arctic.

In the CMIP5 models, significant winter snow depth trends are more strongly negative in the north than in the south (Fig. 14). This is in disagreement with point observations presented in the previous sections that showed no significant trends snow depth trends at Alert but negative and significant trends at Cambridge Bay. Although only based on limited point in situ observations, this suggests that over the last decades changes in winter precipitation at Alert must have compensated the increased melting driven by increasing temperatures, a compensation that is clearly not captured in CMIP5 models.