EM ice thickness measurements utilize the contrast of electrical conductivity between sea water and sea ice to determine the distance of the instrument to the ice–water interface . In 2001 during the RV Polarstern cruise (ARK-XVII/2), only ground-based EM (GEM) data were obtained using an instrument (Geonics EM31Mk2) pulled on a sledge across the ice . With GEM measurements, the distance to the ice–water interface corresponds to the ice plus snow thickness (hereafter referred to as EM ice thickness). After 2001, measurements were made with an airborne EM (AEM) system towed 12 to 20 m above the ice surface. Here, the distance to the uppermost snow surface is determined with a laser altimeter. The ice plus snow thickness is then calculated as the difference between the laser and EM derived distance . In 2004, AEM measurements were conducted with a helicopter operated from RV Polarstern (cruise ARK-XX/2) along triangular flight tracks with a side length of 40 to 80 km . In 2010, 2011 and 2012 AEM surveys were conducted with the Polar 5 aircraft during the TIFAX (Thick Ice Feeding Arctic Export) campaigns operating from the Danish Station Nord in Nord-East Greenland . These airplane surveys allow the acquisition of hundreds of kilometers of data along straight flight lines. An overview of the flight tracks surveyed during the individual field campaigns is given in Fig. .

The accuracy of the EM measurements is of the order of ±0.1 m over level sea ice . However, the maximum thickness of pressure ridges can be underestimated by as much as 50 %. The underestimation of peak pressure ridge thickness is a result of footprint smoothing, an effect that is mass-conserving for mean thickness values on the kilometer scale . The thickness PDFs for all profiles presented in this paper were calculated from histograms with a bin width of 0.1 m. The most frequently occurring ice thickness, the mode of the distribution, represents level ice thickness and is the result of winter accretion and summer ablation. Because ridge thicknesses are in general underestimated in AEM data, the mode is most representative of the ice thickness PDF. The fraction of dynamically deformed ice is represented by the length and the shape of the tail of the thickness distribution. In this study, the fraction of ice thicker than 3 m is used to give a relative estimate of the amount of deformed ice. The mean thickness is used to quantify the overall decline in sea ice thickness. Note that before calculating mean and modal thickness from the PDFs, thin ice (less than 0.15 m) and open water were excluded from the analysis. For the investigation of across and along Fram Strait thickness gradients, PDFs, mean and mode were calculated over a 25 km distance for meridional profiles (along Fram Strait) and zonal profiles (across Fram Strait). The distance is equivalent to the spacing between ULS observations of at 79∘ N.

Since per definition EM ice thickness measurements include the snow layer, interannual changes in ice thickness may not be solely related to changes in ice thickness, but also to changes in snow cover. During the presented EM surveys no snow measurements were made, except during the RV Polarstern cruises in 2001 and 2004, where a mean snow or weathered ice thickness of 0.07–0.1 m was observed. Therefore, and due to the general snow climatology of Arctic sea ice where snow completely melts in June and July leaving the ice surface bare in August and September , we assume a 0.1 m thick layer of weathered ice or snow to contribute to the total ice thickness. This assumption is also supported by snow or weathered layer observations in Fram Strait during the months of August and September by . Variations may be due to episodic, short-lasting events of new snow accumulation which typically melt within a few days during July and August. Below we assume the unknown interannual variability of snow thickness to be equivalent to the averaged snow thickness uncertainty on multi-year ice for July and August (± 5.0 cm) provided by .

The examination of interannual changes in the sea ice cover over a certain area requires continuous and overlapping measurements. Despite shortcomings due to logistical and meteorological challenges of air- and shipborne campaigns in the Arctic, we consider our data set to be sufficiently homogeneous with respect to its temporal and spatial coverage. Nevertheless, to ensure a maximum degree of consistency and to limit bias due to warm Atlantic Water , only flights obtained between 82 and 85∘ N and 13∘ W and 20∘ E were selected (compare the red shaded area in Fig. ). A summary of the survey flights obtained during individual campaigns is presented in Table  together with survey dates and length of EM-profiles. In addition, the modal and mean ice thickness, as well as fraction of ice ≥3 m and the open-water fraction are given.

The interpretation of EM thickness measurements in a larger spatial context requires information about the age, drift history and source areas of the surveyed ice. Below we describe the data set that was used to determine age and drift trajectories. In addition, we present the approach to quantify ice area fluxes through Fram Strait.

To investigate pathways and source areas of the surveyed ice, we used the location of the survey lines as starting points for the backtracking algorithm. Figure  shows the trajectories of ice surveyed in the area of interest between 2001 and 2012. The analysis shows that the largest fraction of the ice originated in the Laptev Sea. It took approximately 2–3 years of drift with the Transpolar Drift until the ice was exported through Fram Strait. In contrast, the ice surveyed in 2010 west of the 0∘ meridian mostly originated from the Beaufort Gyre.

Figure  compares the age of oldest ice (source: ) covered by EM measurements with the average age of oldest sea ice exiting through the meridional and zonal flux gates in summer (July–September). The time series shows a consistent decrease of the age of old Fram Strait sea ice at a rate of 0.6 years per decade. This result is significant at the 95 % confidence level. The average age of the surveyed old ice between 2001 and 2012 is 2.56 years. The youngest ice was observed in 2012 (2.1 years), and the oldest ice was observed in 2004 (3.3 years). The surveyed ice had a slightly higher mean ice age than all ice of Fram Strait combined. However, the differences are within the standard deviation (SE) and therefore in reasonable agreement. The good agreement indicates that EM observations are representative for a larger Fram Strait area.

Figure  summarizes EM thickness data obtained between 2001 and 2012. Owing to the rather limited number of campaigns and the snapshot character of the surveys a trend analysis of the time series may be of limited value. Nevertheless, given the overlapping study regions and seasons and the large lengths of surveys, the EM data provide evidence of a changing Fram Strait sea ice cover that stands out of the interannual variability and bias that may arise from year to year varying snow cover (± 5.0 cm, ). According to Fig.  the modal ice thickness has decreased over the past 11 years, with a distinct reduction in ice thickness after 2004, when the mode dropped by 36 % from 2.2 m (2004) to 1.4 m (2012). Similar observations were made in 2007 at the North Pole by . The interannual variability in modal thickness can be explained to some degree by different age compositions. For instance, the higher modal thickness in 2004 is likely the consequence of predominantly older ice (compare Figs.  and ). However, there is no evidence of a change in age composition of surveyed ice towards younger ice that could explain the overall decline in ice thickness. In fact, the length of pathways and age of surveyed ice in 2010 and 2012 does not differ much from 2001, but the modal thickness is significantly lower. Therefore, we assume that the decline in modal thickness observed in Fram Strait rather reflects the thinning of second-year and multiyear ice in the Laptev Sea (source area) and Transpolar Drift than decreasing age. The decrease in modal thickness is accompanied by a decrease in ridged ice (fraction of ice thicker than 3 m). Note that in 2001 and 2004, the fraction of deformed ice is twice as high as in 2010, 2011 or 2012. Similar to the modal ice thickness, some of the interannual variability may be related to a varying age composition, but the overall decline is independent of ice age. Hence, the reduction of the deformed ice fraction points to a reduction in the deformation history in source areas and along pathways, mainly in the Laptev Sea and along the Transpolar Drift, which is in agreement with findings of . Following the authors, the decrease can be associated with changes in wind stress or a loss in perennial ice (decrease in ice age), since younger ice likely contains less consolidated pressure ridges. Another important factor that could explain the observed decrease in deformation is ocean heat, since melt rate is thickness dependent and an increase in ocean heat affects ridges much more than surrounding level ice. The decrease in ice extent , and the speed-up of ice drift along the pathways with the associated increase in lead fraction leads to an increased heat uptake which could in turn result in enhanced melt of deformed ice. The shrinking tail of the ice thickness distribution as well as the decrease in modal ice thickness is also reflected in the mean thickness. Figure  shows that during the past 11 years the mean thickness dropped by 16 % from 2.58 m in 2001 to 2.17 m in 2012. A slight increase in mean thickness takes place after 2010. The increase is related to an increase in the fraction of deformed ice between 2010 and 2012. This is in agreement with who estimate that the contribution of thick, deformed ice towards the mean ice thickness is decreasing from about 70 % in 2001 to about 50 % in 2011.

The comparison of AEM- and GEM-based observations may introduce an additional uncertainty and must be limited to a comparable range of the thickness distribution. Although GEM data were obtained on a daily basis at representative locations along the ship track, the ground-based thickness surveys of 2001 are limited to large floes and predominantly level ice thick enough to walk on. In addition, the footprint of ground-based measurements is smaller than the footprint of airborne surveys which reduces footprint smoothing of pressure ridges. However, thickness distributions obtained by both methods in the same region have very similar shapes and modes (e.g., , their Fig. 3), warranting their combination for this study. To further ensure compatibility with the AEM thicknesses, the GEM data have been regridded to the sampling interval of the airborne data and ice thinner than 0.15 m and open water has been excluded from the analysis of the AEM measurements (see Sect. ). For our study we assume that mean thicknesses obtained with both method are comparable as well. We base this assumption on the high number of available GEM surveys and the general exclusion of thin ice thicknesses from the AEM data, which will be vastly underrepresented in the GEM data.

In Fig.  we compare our thickness measurements with thickness estimates made by and . By means of moored upward looking sonar (ULS) positioned between 79∘ N, 7∘ W and 79∘ N, 3∘ W, reconstructed a time series of sea ice thickness over 21 years (1990–2011). To enable a comparison with our observations, ULS thickness estimates in Fig.  are averaged August measurements, except for 2012 where averaged July measurements are used. Ice thickness measurements taken from were obtained with a GEM during cruises with RV Lance (Norwegian Polar Institute). Measurements cover the width of Fram Strait along approximately 79∘ N in September between 2003 and 2012. For details about data processing and handling we refer to and . A decrease in both, modal and mean thickness with a distinct reduction after 2004 is visible in all three data sets. According to the ULS observations, the mean and modal thickness in August is decreasing by 0.65 m and 0.41 mdecade-1 between 1990 and 2012. GEM observations indicate an even more pronounced thinning of Fram Strait ice cover. A direct comparison of our observations with ULS- and GEM-based data is however difficult. In contrast to the AEM data, the ULS measurements consist of monthly averaged records obtained at single points located approximately 300 km further south. Nevertheless, despite the different locations the agreement between ULS and AEM data for August 2010 and 2011 and July 2012 is high. This indicates that a few but long AEM profiles provide representative information on ice thickness distribution even in areas of highly variable ice age and thickness composition such as Fram Strait. For the last 3 years, the agreement between AEM data and GEM measurements obtained by is high, too. Nevertheless, taking into account that GEM measurements by were obtained approximately 1 month later (September), one would expect the GEM thickness measurements to be lower than ULS and AEM data. According to , the positive offset is likely related to the absence of thin ice classes in the observations and preferential sampling of the survey sites.

The thinning due to atmospheric and oceanographic processes on southward-moving sea ice was investigated during two ice thickness surveys performed in the downstream direction. Figure a shows AEM profiles that were made on 4 August 2011 and 21 July 2012. The first profile started at 81∘ N, 0∘ E and covers a distance of 290 km (south to 79∘ N, 10∘ W).

According to aerial photos taken during the flight, the ice cover was rather homogeneous. Likewise, there is no gradient in ice concentration along the profile or changes in the frequency of open-water occurrence. The high spatial variability in mean thickness makes an identification of a thickness gradient impossible. However, the modal thickness shows a continuous decrease of 0.19 mdegree-1 latitude. The decrease in modal thickness is likely associated with oceanographic and atmospheric processes acting on the pack ice while drifting south: differences in net short- and longwave radiation and the presence of warm Atlantic water may lead to enhanced surface and bottom melt that could explain the observed gradient. A thinning of 0.38 m implies a heat flux of 16 Wm-2. In principle, one could quantify the contribution of warm Atlantic water to sea ice melt via the determination of the atmospheric radiative fluxes and the transit time of sea ice between 81∘ N, 0∘ E and 79∘ N, 10∘ W. For example, using the backtracking approach as described in Sect. , we estimated the transit time of sea ice to be around 80 days. However, an exact determination of the transit time using low-resolution drift products is difficult, since their reliability is significantly reduced in the southern Fram Strait. In addition, atmospheric reanalysis products show large differences in radiative fluxes: the NCEP-based difference in net radiation between 81∘ N, 0∘ E and 79∘ N,10∘ W amounts to 12 Wm-2 for a transit time of 80 days. The difference in radiative fluxes based on ERA-Interim data is much smaller (2.5 Wm-2). Hence, an exact determination of oceanographic and atmospheric processes contributing to the observed gradient remains elusive without additional observations.

In 2012, a second 170 km long flight in the upstream direction was performed. Measurements were a continuation of the transect made in 2011 and started at 80.5∘ N. The ice cover was again rather homogeneous with a few leads. According to Fig.  ice was formed in the western Laptev Sea and transported via the Transpolar Drift towards Fram Strait. The absence of a gradient in modal thickness indicates that enhanced bottom or surface melt due to atmospheric or oceanographic processes is limited to areas south of ≈80∘ N.

The ice thickness gradient across Fram Strait was investigated during two flights in 2010 (22 August) and 2012 (21 July). The long operating distance of Polar 5 enabled us to obtain the first continuous profiles over closed ice pack north of 81∘ N. The across-strait ice thickness profile is presented in Fig. b. Both transects show a negative trend in modal (0.02 m and 0.04 mdegree-1 longitude) and mean (0.03 m and 0.11 mdegree-1 longitude) ice thickness from West to East. The gradient in mean thickness is thereby more pronounced than the gradient in modal thickness. For sea ice at this latitude or higher, one can assume the impact of oceanographic and atmospheric processes on the ice cover to be smaller. This assumption is supported by the absence of a gradient in modal ice thickness for sea ice upstream of 80.5∘ N discussed above. Hence, we assume the observed gradient to be mainly associated with differences in age and deformation of ice provided by the Transpolar Drift system. A comparison to Fig.  reveals that the ice that enters Fram Strait west of the prime meridian is indeed older and therefore most likely thicker than ice that enters through the eastern section. Note that the good agreement between the length of pathways and observed thickness gives us confidence in the performance of the tracking approach.

Earlier quantifications of across-strait gradients were made by and approximately 300 km further south at 79∘ N. Their estimates are based on interpolations between single-point upward-looking sonar measurements and on merged EM profiles obtained during different days. For this position, the authors reported a decline in across-strait modal thickness of -0.1 to -0.3 mdegree-1 longitude and -0.23 mdegree-1 longitude .

To quantify whether coupled sea ice ocean models are capable of reproducing Fram Strait sea ice volume fluxes correctly, validation data are required. Using satellite data, the volume flux in Fram Strait can be described as the product of southward-directed sea ice motion, concentration and mean thickness. Information on ice drift and concentration is available on a year-round basis. However, the availability of satellite-based thickness data from ICESat or CryoSat-2 are restricted to winter months, which is why ice volume flux estimates for summer periods are scarce. In the following, we will therefore use the presented EM measurements together with satellite-based area flux estimates to calculate volume outflows for the periods when thickness surveys were made.

Because of its year-round availability, ice area flux out of Fram Strait is calculated using NSIDC motion estimates together with NSIDC ice concentration information. Figure  shows the monthly ice area export across Fram Strait from 1980 to 2012 (black line) together with the associated uncertainty estimates (gray). Note that the area flux is the sum of meridional and zonal components, with a positive sign referring to ice export, and a negative sign indicating ice import into the Arctic (see Sect. ). The average monthly ice area flux amounts to 46×103 km2 with a standard deviation of 38×103 km2 and an uncertainty of ± 18 × 103 km2. The monthly ice export shows a pronounced seasonal cycle with lowest fluxes in July and August and highest export rates between December and March. During summer, flux rates are significantly lower and can become even negative, such that ice is being imported from southern Fram Strait. The pronounced seasonal cycle and much of the interannual variability of ice area export are associated with changes in SLP gradients across the gate, because gradients are generally lower during summer months and higher during winter. In addition, sea ice concentration in Fram Strait is lower during summer months, which leads to reduced export rates between July and September. The reduced drift speed and ice concentration during summer (May–September) also results in higher uncertainties (± 19 × 103 km2). Likewise, during winter months, when average drift speed and ice concentration are higher, the uncertainty of area flux estimates is lower (October–April, ± 14 × 103 km2). Overall we find a positive trend in monthly Fram Strait area flux of 10.9×103 km2decade-1. The trend is significant at the 99 % confidence level. Following the increase in ice export is the consequence of a positive trend in the local pressure gradient, related to intensification of cyclones over the Nordic Seas. According to that study the sea ice area export has increased by about 25 % since the 1960s. The increase in ice export occurred mostly during winter and is directly connected to higher southward ice drift velocities, due to stronger geostrophic winds. In contrast, other studies did not observe any significant trend in ice area export for the past decades.

The area export for July, August and September accounts for only 6.2 % of the annual fluxes. The months with the lowest net contributions are July and August (1.4 %). Ice area export rates for both months are shown together with associated uncertainty estimates in Fig.  (black and red lines). The net sea ice export during July and August is positive, but estimates show considerable interannual variability with the highest rates occurring in August of 1994, 2006 and 2010 and lowest in 1981 and 1998. The average August ice area flux amounts to 4.3×103 km2 with a standard deviation (SE) of 19×103 km2 and an uncertainty of ±21 × 103 km2. The average July ice flux is a bit lower (3.8×103 km2 ± 13 × 103 km2) with higher uncertainties (±24 × 103 km2). Note that there is a positive trend in the August and July ice export of 5.5×103 and 2.3×103 km2decade-1, respectively. The trend is however not statistically significant.

The associated volume fluxes for the years where GEM/AEM measurements are available (Fig. ) is calculated as the product of area flux and mean GEM/AEM thickness minus the snow thickness assumptions we made for August. The uncertainty of volume fluxes is the product of area flux uncertainties and mean ice thickness plus the snow thickness uncertainty (±5.0 cm). For 2012, where AEM measurements were made 1 month earlier, area transport rates for July (black line) were used. Given the low area export in July and August, the volume transport is low, too. For the investigated months, the average volume export amounts to 16.78 km3 with highest rates in August 2010 (61.25 km3) and lowest in August 2001 (-15.35 km3). The average uncertainty amounts to ±44.48 km3.

The reliability of volume flux depends as well on the accuracy of sea ice motion information in summer as on the available ice and snow thickness information. An additional bias of ± 4 % may arise from uncertainties in sea ice concentration during summer months. This error is however difficult to account for, since we expect it to vary with the presence of thin ice and melt ponds, atmospheric water vapor and water-surface roughening by wind . Assuming that the sea ice thickness PDFs are accurate, and uncertainties related to interannual variability in snow cover are small (±5 cm), the biggest error in volume flux estimates arises from sea ice motion information. Due to the lack of sea ice motion observations obtained from drifting buoys in Fram Strait during summer months, we cannot evaluate the uncertainty of satellite-based sea ice motion information directly. Nevertheless, recent studies of indicate that NSIDC ice motion information suffer from a general underestimation of drift during summer months and a generally reduced accuracy in the narrow Fram Strait. In particular, low drift velocities and ice concentration result in errors of up to 2.0 cms-1. In this study we apply the uncertainty estimates provided by for different drift speeds and ice concentration to evaluate reliability of our flux calculations. As an additional quality control we compare our results with area flux estimates from and (Fig. , red and gray line). Area flux calculations of are based on ice concentration data and manually derived ice motion information from ENVISAT SAR images. SAR WideSwath image pairs were captured 3 days apart with uninterrupted year-round coverage from February 2004 to December 2011. Estimates were made across 79∘ N, 15∘ W and 79∘ N, 5∘ E. Note that at 79∘ N Fram Strait is relatively narrow and therefore only a limited number of images are needed to cover the entire passage. According to the authors, the monthly mean export uncertainties amount to 5 %. used the pressure difference (NCEP/NCAR reanalysis data) between 79∘ N, 25∘ W and 79∘ N, 5∘ E together with SAR-based flux estimates of to estimate the linear regression between geostrophic winds, sea ice drift speed and ice area export. The linear relationship was then used to reconstruct ice area export based on pressure differences for the period between 1957 and 2010. A direct comparison of our area flux estimates with the findings of and is difficult, because area flux estimates are based on different methods and were made at different latitude gates. However, a comparison of our findings with area export estimates of others reveals that the trend in NSIDC export rates is much higher (37.6 %decade-1 for the period 1980–2012) than the trend found by (22.2 %decade-1 for the period 2004–2011, compare Fig. ) or the trend reported by (4.7 %decade-1 for the period 1980–2010), and in contrast to and , who did not observe any significant trend in ice area export for the past decades. A discussion on causes of differences in observed trends is beyond the scope of this paper. Nevertheless, despite large differences in observed trends the concordance between our findings and estimates of and gives us confidence in our results. The agreement in seasonal variability indicates that there is a relative consistency between area fluxes: the correlation coefficient (r) between NSIDC-based estimates and computations of is 0.79 and between NSIDC area flux and SAR-based estimates 0.80. The agreement between SAR-, and SLP-based export rates are of the same order (r = 0.82). A comparison of absolute fluxes for the periods where NSIDC, NCEP/NCAR reanalysis data and SAR-based estimates are available (2004–2010) shows that our estimates are approximately 18 and 20 % lower than estimates of and . If the intercomparison is limited to summer months only (July–September), NSIDC-based export rates are within the range of SAR-based estimates (27×103 km2 vs. 28×103 km2), whereas computations based on pressure differences are higher (53×103 km2).