A comparison of Arctic Ocean sea ice concentration among the coordinated AOMIP model experiments



[1] Sea ice concentration is a fundamental property of the Arctic ice-ocean-atmosphere system reflecting both dynamics and thermodynamics. Concentration integrates across space and time and is useful for characterizing both observed and numerically simulated systems. Concentration is reasonably well measured by remote sensing, and several high-quality sea ice concentration data sets exist beginning with the satellite era. In this paper we examine the simulated sea ice concentration from nine ice-ocean numerical models that are part of the coordinated experiments of the Arctic Ocean Model Intercomparison Project (AOMIP). Spatial patterns of means and differences between models and observations, and among models, are compared for a multiyear record and for the September sea ice minimum. Interannual variations are assessed on data with monthly climatology removed. As a proxy for the annual cycle of open water for each model, the total areas with concentration less than 10% are compared among models. Mean ice statistics are computed for grid points with greater than 1% and greater than 10% concentrations. The results show that the models have similar characteristics for the winter months when 100% cover is produced, and most models reproduce an observed minimum in sea ice concentration for 1990. The compared observational data sets use the NASA Team algorithm (Goddard Space Flight Center data, the adjusted or Walsh data, and the Hadley Centre data) and the Bootstrap algorithm. Variability in sea ice concentration is less among the four observational records than among models.

1. Introduction

[2] Profound changes are occurring in the Arctic. Many recent changes suggest rapid and dramatic shifts in Arctic sea ice, a key indicator of climate change. The sea ice trend over the past decade shows decline, with record and near-record minimums in sea ice concentration and extent during the past five years [Meier et al., 2005]. Passive microwave satellite data for 24 years [Fowler et al., 2007] suggest that the oldest and thickest sea ice is being preferentially lost from the Arctic [Belchansky et al., 2005]. These changes warrant our attention.

[3] Global climate modeling studies show the Arctic to be one of the most sensitive regions to climate change [Manabe and Stouffer, 1994; Rind et al., 1997; Dickson, 1999; Serreze et al., 2000]. Climate models indicate that the poles are more sensitive to climate change than lower latitudes because of feedback processes driven by sea ice [Moritz et al., 2002].

[4] Recent sea ice studies indicate a 3 to 4% decline per decade in sea ice extent [Parkinson, 2000], while under-ice measurements from submarines indicate a 40% decline in thickness in the central Arctic Ocean prior to 1999 [Rothrock et al., 1999, 2003]. The consistency in the sign of these observations suggests real, unambiguous decline.

[5] Identifying the scales of variability is particularly difficult because of the spatial heterogeneity of the Arctic system. Sea ice anomalies show variability at the 4 to 6 year time scale in the Western Arctic, but decadal scale variability in the Eastern Arctic [Mysak and Manak, 1989], making detection of long-term trends difficult. Ultimately, long-term observations over all seasons are needed to assess model performance that accounts for thermodynamic processes that dominate the annual cycle and dynamic processes that dominate interannual cycles [Walsh et al., 1985].

[6] To improve our understanding of the Arctic ice-ocean system and to improve accuracy of models simulating this system, a series of coordinated model experiments by the Arctic Ocean Model Intercomparison Project (AOMIP) resulted in a suite of nine model simulation runs from 1979 to 2001 using consistent forcing. The details and motivation for these experiments are found in Holloway et al. [2007].

[7] In this paper, the long-term and September mean sea ice concentrations from each model are compared with the nine-model ensemble median and with the observations., Following removal of the monthly climatology, interannual variability is assessed by comparing monthly anomalies.

[8] “Open water” is defined here as having concentrations that are less than 10%. The 10% contour is often used on historical ice-charts to mark the open water boundary. Monthly time series of open water are compared among models and observations. We also tabulate the time-mean ice extent defined by concentrations greater than 1%, and greater than 10%, for models and observations.

2. Models and Data

[9] Monthly mean sea ice concentration data were made available for this analysis by the AOMIP modelers. Table 1 lists key model parameters for all the models discussed here. The AOMIP models are AWI1 and AWI2: Alfred Wegener Institute (2 models with AWI1 having higher resolution); GSFC: Goddard Space Flight Center, ICM: Institute for Computational Mathematics and Mathematical Geophysics, IOS: Institute of Ocean Sciences, LANL: Los Alamos National Lab, NPS: Naval Postgraduate School, RAS: Russian Academy of Science, and UW: University of Washington. All models use an ice-ocean virtual salt flux except GSFC, which follows Mellor and Kantha [1989]. Atmosphere-ice heat and moisture exchanges are parameterized using bulk formulas with exchange coefficients 1.2 × 10−3 and 1.5 × 10−3, respectively, except AWI2, which uses 1.75 × 10−3 for both. While most models also apply a bulk parameterization with a drag coefficient for ice-ocean stress to both the ice and the ocean, as shown in Table 1, the AWI models use the bulk form for computing the bottom stress on the ice but employ the formulation of Hibler and Bryan [1987] for the surface stress acting on the ocean. See Holloway et al. [2007] for additional model details.

Table 1. Model Configuration and Selected Parametersa
  • a

    See Holloway et al. [2007] for additional details.

  • b

    Regional domains generally include the North Atlantic and Arctic oceans.

  • c

    Resolution in degrees refers to the rotated grids used by regional models, except for the ICM model that varies from 35 km at the north pole to 1° south of 65 N. LANL uses a nonuniform, displaced pole grid with the north pole singularity in North America.

  • d

    Multiple-category ice thickness distributions incorporate more than one thickness “category” or bin within each grid cell. Here, “1” refers to one category of ice plus open water. In the AWI and UW models, a single prognostic ice thickness is distributed over 7 categories for the heat flux calculation.

  • e

    VP, viscous plastic; EVP, elastic-viscous-plastic.

  • f

    There is one ice albedo in NPS's “0-layer” ice model, which does not include snow processes.

  • g

    Atmosphere-ice momentum exchange approaches vary greatly among the models. The AOMIP protocol calls for a variable exchange coefficient given by (1.1 + 0.4W) × 10−3, where W is wind speed. ICM uses a variable coefficient that depends on atmospheric stability. IOS applies NCEP wind stress directly.

  • h

    Jordan et al. [1999].

  • i

    See text for additional details.

  • j

    Mellor and Kantha [1989].

Resolutionc0.25°∼0.8°35 km-1°0.5°0.4°18 km40 km
Ice Δt21600 s900 s1080 s10800 s43200 s1800 s7200 s1800 s1440 s
Ice Physical Parameterizations
Salinity, ppt445Profile4profile044
Thickness categoriesd1(7)1(7)1515181(7)
Advectioncorrected upstreamcorrected upstreamcentered differencesupstream + remapmodified Prather, '86incremental remapcentered differencescentered differencescentered differences
Cold snow0.810.810.810.810.800.81 0.800.84
Melting snow0.770.770.770.770.700.77 0.600.75
Cold ice0.700.700.700.700.600.700.73f0.600.75
Melting ice0.680.680.680.680.500.68 0.500.66
Surface Momentum Exchange Coefficients
atm-icegAOMIPAOMIP1.1 × 10−3JAM1999hNCEPAOMIP1.1 × 10−3AOMIPAOMIP
ice-ocni5.5 × 10−35.5 × 10−3MK1989j5.5 × 10−35.5 × 10−35.5 × 10−35.5 × 10−35.5 × 10−35.5 × 10−3

3. Comparisons Between Observations and Models

3.1. Observations

[10] The gridded sea ice concentration data sets analyzed in this paper include three that use the NASA Team Algorithm (the NASA GSFC sea ice concentration data, the NSIDC (Walsh and Chapman, henceforth “Walsh”) data, and the Hadley Centre concentration data), and one that uses the bootstrap algorithm (henceforth “Bootstrap”).

[11] The GSFC sea ice data set uses the Nimbus-7 SMMR and DMSP SSM/I passive microwave measurements to provide a consistent time series of concentration. It is available online (see http://nsidc.org/data/docs/daac/nsidc0051_gsfc_seaice.gd.html). The Walsh data set is based on passive microwave data with adjustments to correct for low summer concentrations. It blends several data sets and uses the historical Walsh data from 1901 to 1930. For the comparisons done here, the dominant input to this data set is monthly climatology based on the 1979–1995 satellite passive microwave data. The Walsh and Bootstrap sea ice concentration data sets are available from NSIDC (see http://nsidc.org/data/docs/noaa/g00799_arctic_southern_sea_ice/index.html). The Hadley Centre global sea ice and SST data set (HadISST) is on a 1° grid from 1870 to the present [Rayner et al., 2003]. The Hadley Centre data is also based on satellite microwave concentration with compensations applied for surface melt effects. The Hadley data incorporates both the Walsh and GSFC data (see http://www.hadobs.org/). In this paper we use the monthly averaged sea ice concentration [Cavalieri et al., 2005] and the monthly averaged model output.

3.2. Models

[12] Each model has a different and unique computational grid making direct comparisons difficult. To alleviate this, the monthly data from each model were linearly interpolated to a polar stereo grid, exactly like the uniform 12.5 km grid used by NSIDC for their products with the result being roughly a “square” of data 469 by 457 points (Figure 1). The NSIDC grid was selected for convenience for comparison of all model data because this is the grid already used to map observed sea ice concentration. In the following discussions, all model results for means, medians, temporal histograms and time series are based on monthly model data interpolated to the uniform grid.

Figure 1.

The uniform 12.5 km polar grid with points labeled following the NSIDC products that are used in this paper. Note cutoff in the Bering Sea. Data from all models were interpolated to this uniform grid with a master land mask applied.

[13] Minor grid point differences among the models near islands and coastlines motivated us to create a master “land-mask” so that only those points that were “active” for all models were used in the following calculations. In all cases, model values that were used to compute statistics were “live”, and any grid points that were constant (usually zero) were added to the land mask. All results discussed in the remainder of this paper are based on analysis of the uniformly interpolated model data with the master land mask applied.

3.3. Comparisons

[14] The 1979 to 1995 mean sea ice concentration was computed for all four observational data sets (Figure 2). The 1995 limit was used because of the 1995 satellite passive microwave data limit in the Walsh data. The means are generally similar, but differences at the highest concentrations are evident north of the Lincoln Sea, north of the Canadian Archipelago, and over the western central Arctic. The Barents Sea, Chukchi Sea, Baffin Bay and the perimeter of the Arctic Ocean all have similar mean sea ice concentrations.

Figure 2.

Mean sea ice concentration for GSFC, Walsh, Hadley Centre, and the Bootstrap data sets for 1979–1995. The yellow contour is 0.9. Note the differences among the highest concentrations that appear north of the Lincoln Sea, north of the Canadian Archipelago, and over the western central Arctic.

[15] Figure 3 (left) shows the record-length grand mean for all models combined. The large range of individual model means motivated us to calculate the median of the ensemble of record-length model means (Figure 3, right). The two maps are generally similar, but with differences evident in the shape of the 0.9 contour line in the central Arctic, and larger differences in the Canadian Archipelago and around the periphery of the Arctic basin. Concentration exceeds 0.9 over most of the Eurasian Basin, Beaufort Gyre and Canadian Archipelago. Along the perimeter of the Arctic Ocean, in the Beaufort Gyre, the East Siberian Sea, the Laptev Sea, the Barents Sea, along East Greenland Current, and much of Baffin Bay the concentration falls to about 0.5–0.6. Much of the GIN Sea, Labrador Sea, and the southern Baffin Bay are below 0.5.

Figure 3.

(left) Grand mean sea ice concentration for all models from 1979–1999. (right) The ensemble median of the record-length model means. Panels are largely similar except for the shape of the yellow 0.9 contour, the marginal seas, and the Canadian Archipelago.

[16] Differences between the 1979–1999 mean GSFC data and the 1979–1999 modeled mean concentration are shown in Figure 4. All the models show reasonable agreement in the central Arctic Ocean but with increasing differences toward the basin periphery and lower latitudes. The models were visually ranked based on their mean sea ice concentration differences from the GSFC mean over the Arctic Ocean. Red (warmer) colors are below the mean and blue (cooler) colors are above the mean. Differences near the periphery can result from differences in modeled extent. Models AWI2, AWI1, and UW (top row) underestimate the concentration in the Arctic Ocean. Models NPS, IOS and ICM slightly underestimate the concentration in the central basin, but with larger differences along the periphery. Model ICM strongly underestimates concentration in the Laptev, Kara and Barents Seas. Models LANL, GSFC, and RAS slightly overestimate the sea ice concentration across the main basin and more so over the periphery. Most models overestimate the sea ice concentration, compared to the observations, east of Greenland.

Figure 4.

Differences for each model between the mean model sea ice concentration and the mean sea ice from GSFC for 1979–1999. The dark line is the 0.001 concentration contour from the GSFC data. Models from left to right and top to bottom are AWI2, AWI1, UW, NPS, IOS, ICM, LANL, GSFC, and RAS. Scale is from −0.4 (red) to +0.4 (blue) with values nearer zero having less color saturation. Saturated colors indicate larger differences from the observations with red below and blue above the observed.

[17] The annual minimum in sea ice concentration in the Arctic Ocean occurs during September. Monthly mean sea ice concentrations were computed for each of the observational data sets, and the September sea ice minimum is shown in Figure 5. The highest concentrations, above 0.9, are quite different among the data sets. For the GSFC data, the highest concentrations occupy the coastal area north of the Canadian Archipelago and Greenland. In the Walsh data the high concentration region extends well into the western Arctic Ocean and curves around the pole, a possible artifact of data filling near the pole. In the Hadley Centre and Bootstrap data, the central and western Arctic Ocean are broadly covered by high sea ice concentration in September, including the pole for the Hadley data, although the Bootstrap data shows a larger area above 0.9, and highest concentrations north of the Canadian Archipelago. There are clear differences among the data sets for the lower concentrations along the periphery of the Arctic Ocean.

Figure 5.

September mean sea ice concentration from the GSFC data, the Walsh, the Hadley Centre data, and the Bootstrap data (1979–1995). The highest concentrations (the yellow line is 0.9) are different among the observations. The Hadley and Bootstrap data are similar.

[18] The model ensemble September mean, and median of the model September means, were computed and are shown in Figure 6. The highest concentrations are generally found north of the Canadian Archipelago and Greenland, with concentration decreasing away from this area. For September, the mean and median are different in terms of the shape of the 0.9 contour across the central basin, the Canadian Archipelago, and the periphery suggesting wide extremes in the model September means. There are notable differences between the observed September mean and the modeled mean and median. The Hadley and Bootstrap observations clearly show higher sea ice concentrations than the models across the central Arctic Ocean. The model mean and median have higher concentrations near the periphery than the observations. This suggests that the models are underestimating concentration where concentrations are higher (in the interior, where there is multiyear ice), and overestimating ice where concentration is lower (in the periphery, where first-year ice often dominates).

Figure 6.

(left) Mean September sea ice concentration from all models. (right) The median from the ensemble of model mean September concentrations. The highest concentrations extend well into the Beaufort Gyre in the figure on the right, and the midrange concentrations penetrate farther to the periphery.

[19] Because the Hadley data incorporates information from the GSFC and Walsh data, and because of the close similarity between the Hadley and Bootstrap methods, we select the Bootstrap data for additional comparison with the models. Differences for each model between the Bootstrap observed sea ice concentration and each model mean for September are shown in Figure 7. We again focus on the Arctic Ocean region to rank the models by inspection from those that are below the observed September mean to those that are above the September mean. Although there are regional variations, models AWI2, AWI1, and UW underestimate the concentration over the Arctic Ocean compared to the data. Models NPS and IOS underestimate the concentration in the central basin with overestimates along the periphery. Model ICM overestimates along the periphery near Alaska and the East Siberian Sea and underestimates in the Laptev Sea. Models GSFC and LANL slightly underestimate concentration in the central Arctic Ocean. Except for RAS, which slightly overestimates across the entire basin, all the AOMIP models underestimate the September concentration in the central Arctic compared to the observed data.

Figure 7.

The difference in sea ice concentration for each of the model September means and the Bootstrap September mean for the 1979–1999 period. The dark line is the 0.001 contour from the Bootstrap data. Models are shown by visual inspection from under to overestimation of sea ice. From left to right and top to bottom: AWI2, AWI1, UW, NPS, IOS, ICM, GSFC, LANL, and RAS. Scale is from −0.4 (red) to +0.4 (blue) with values nearer zero having less color saturation. Red is less than and blue is more than the observations.

4. Comparisons Among Models

[20] We computed each model's 1975–1999 mean, monthly means (climatology), and the deviations for each month from their monthly means (anomalies). Differences between each model's record-length mean and the ensemble median (from Figure 3) were computed and are shown in Figure 8. The models were ranked by inspection of the central Arctic Ocean for results that fell below the median to those that were above it. Models AWI2 and AWI1 are below the median ice concentration across the central Arctic Ocean. Model UW is below the median in the Beaufort Gyre and the East Siberian Sea. Model IOS is slightly above the median in the Amerasian basin and below it in the Eurasian basin. Models LANL, GSFC, and RAS are slightly more than the median north of Greenland and the Canadian Archipelago, and are well above it across the rest of the Arctic basin.

Figure 8.

The difference in sea ice concentration for each model for the record-length mean and the median of the ensemble of model means. From upper left to lower right the models are AWI2, AWI1, UW, NPS, IOS, ICM, LANL, GSFC, and RAS. Red is negative (less ice) and blue positive (more ice) on a difference scale from −0.4 to 0.4. Color saturation indicates increasing difference.

[21] The ensemble median of the nine-model mean September sea ice concentration was subtracted from the September mean concentration for each model (Figure 9). Red shades denote values lower than the median and blue shades denote values higher than the median. Focusing on the Arctic Ocean basin, models AWI2, AWI1, UW, and NPS are generally less than the median September sea ice concentration. Model IOS is generally near the median but is lower in the central Arctic basin. Model ICM is below the median in the Laptev and Kara Seas, and well above it elsewhere. Models LANL, GSFC, and RAS are above the median September concentration over the entire region.

Figure 9.

Differences between each model September mean sea ice concentration and the median of the ensemble of nine model September means. Models were ranked by inspection from below to above the median with red denoting values below and blue denoting values above. Models, in order, are AWI2, AWI1, UW, NPS, IOS, ICM, LANL, GSFC and RAS. Models AWI2, AWI1, UW, and NPS are generally below the median sea ice concentration, and models ICM, LANL, GSFC and RAS are above the median in the Arctic Ocean. Model IOS is closest to the median.

[22] Not shown are the differences between the model mean concentrations for April and the median of the ensemble of model April means. The nine models are more similar in April than in September because most models produce concentrations of 1 in the winter.

5. Interannual Variations

[23] To understand better the interannual behavior of the observed and modeled sea ice, we computed monthly histograms of concentration and contoured them in time. Figure 10 shows contours of counts from histograms of the observations for the GSFC, Walsh, and Bootstrap data sets. The top row shows results (1979–1992) from the monthly mean data. The bottom row is from data with the record length monthly climatology removed. Positive values (to right of midline) show when there is more ice than the monthly mean, and negative values (left of the midline) show when there is less ice than the monthly mean.

Figure 10.

Histograms of sea ice concentration for the observational data sets: GSFC, Walsh, and the Bootstrap data. Top row: contours of histogram counts for the monthly mean data. Bottom: histogram counts with the monthly climatology removed. The GSFC data set has a broader range of interannual variability than the Walsh data set. The Bootstrap data show little variation at the lowest and highest concentrations. All three data sets show the 1990 reduction in concentration (bottom row), and it is more evident in the GSFC Bootstrap data sets.

[24] As expected, all data sets show a clear annual cycle (top row). The GSFC and Walsh data sets show more variability at the low and midrange of concentration than Bootstrap. The Bootstrap data show weak variations at the lowest and highest concentrations. Summer melting from the highest concentrations is strongest in the GSFC data, weaker in the Walsh data, and nearly non-existent in the Bootstrap data. Because of the different grids for these data and the numerical precision in the given concentrations, we also computed histograms with different bin sizes and normalization schemes. The above results still apply (see other examples at http://ak.aoos.org/∼gaffigan/).

[25] The bottom row of Figure 10 displays contours of histograms with the climatology removed. The GSFC and Bootstrap data sets have a wider range of interannual variability than the Walsh data set as indicated by the “width” of the contours away from the midline. The Bootstrap data has stronger seasonal variations above, but weaker variations below, the annual cycle. All three data sets show a 1990 reduction in concentration that is more evident in the GSFC and Bootstrap data than in the Walsh data.

[26] Histograms for all nine models have clear annual cycles that produce winter ice that partially melts back each summer. As examples of models that were below, at, and above the ensemble, results for models AWI2, IOS, and RAS are shown in Figure 11 for the period 1979–1992. Model AWI2 shows strong melting each summer especially from the highest concentrations, with very little summer ice above 0.9 each summer. Model IOS clearly has less summer melting above 0.9, and a seasonal cycle that varies at lower concentrations between 0.1 and 0.2. Model RAS has the least summer melting with longer, less variable, winters as shown by the duration, area and smoothness of the contour lines. This range of model variation is greater than the range shown by the observations in Figure 10.

Figure 11.

Histograms of concentration for models AWI2, IOS, and RAS from 1979 to 1992. Model AWI2 shows strong melt back each summer especially at the higher concentrations. Model IOS shows robust temporal variations with the seasonal cycle visible between 0.1 and 0.2. Model RAS shows little summer melt back and longer winters, based on the duration and smoothness of the contour lines. From left to right, note the differences in both area and duration of summer reductions in concentrations above 0.9.

[27] All models show variability at interannual time scales. Figure 12 shows contours of concentration from monthly histograms with their individual monthly climatology removed for six representative models. They are sorted by inspection of the range of variation about the climatology: NPS, IOS, GSFC, LANL, RAS, and ICM. Model NPS shows seasonal variation about both winter and summer months. Compared to NPS, model IOS has less variation in values below the climatology (compare the left and right sides). Interannual variations are less extreme in model GSFC as shown by contour lines that are closer to the centerline. Model LANL has smooth temporal variations across all seasons. Models RAS and ICM have the least interannual variation.

Figure 12.

Contours of counts from monthly histograms of concentration for the period 1979–1992 after the monthly mean climatology is removed. Models are sorted (NPS, IOS, GSFC, LANL, RAS, and ICM) from more to less variation about the climatology. Model NPS shows seasonal shifts in both winter and summer seasons. Model IOS has weaker variation in values below climatology. Interannual variations are less extreme in model GSFC as shown by contours lines that stay closer to the centerline. Model LANL has strong variations across all seasons. Models RAS and ICM have progressively less interannual variation.

[28] As in the observations, there is a visible reduction in concentration in summer of 1990 in models NPS, LANL, AWI1, UW, IOS, and AWI2. Models GSFC, ICM, and RAS do not show a significant reduction at this time

6. “Open Water”

[29] One diagnostic in the annual cycle of sea ice occurs within the marginal ice zone defined as the region where the sea ice concentration falls below 10%. The 10% contour, from a navigation perspective, often delineates the “ice edge” from “open water” and is sometimes used to define areas for navigation. A time-series was created by calculating the number of grid points from the uniform grid in each model for each month where the sea ice concentration was less than or equal to 10%. This was done for the Arctic Ocean basin, and the Laptev, Barents and Kara marginal seas to the latitude of Svalbard. The GIN sea area was not included. Figure 13 shows an annual cycle in most models. The responses for models GSFC and ICM are different from the other models, with a smaller amplitude annual variation. All other models show an annual cycle in the area of open water.

Figure 13.

(top) Time series of the area with concentration 10% or less for the Arctic Ocean including the Laptev, Barents and Kara Seas. No values south of Svalbard, were included. The models are from top to bottom NPS, LANL, AWI1, UW, IOS, GSFC, ICM, RAS, AWI2, and observational data GSFC and Bootstrap. The maximum value in UW for 1996 normalizes the suite of time series. Models GSFC and ICM have low amplitude variations for the “marginal ice zone” where concentrations are less than 10% compared to the others.

[30] We also calculated three proxies for mean ice extent. Two proxies were computed by summing over all months concentrations greater than 1%, and greater than 10%, for all AOMIP models and the GSFC and Bootstrap observational data sets for the Amerasian and Eurasian Basins (Table 2). A third proxy used the area from the NSIDC pixel-area data set where we multiplied the grid point concentration by the area from the NSIDC area data set, to create mean sea ice for both basins and all nine models and the two observational data sets. These are shown in Table 2 as well. These results show that, not surprisingly, the models and the observations create consistent extents of ice during April. Table 2 shows that, for September, AWI2 generally has less ice than the observations and other models, and RAS has more ice than observations and other models.

Table 2. Sea Ice “Extent” for the Amerasian and Eurasian Basins for the Nine AOMIP Models and the GSFC and Bootstrap Observational Data Setsa
  • a

    Extent is calculated as the number of grid points above 10%, and above 1%, normalized by the total number of grid points. Pixel area values are calculated from concentration multiplied by area and normalized by total area using the NSIDC pixel-area data set.

Amerasian Basin > 10%1.0001.0001.0001.0001.0001.0001.0001.0001.0001.0001.000
Eurasian Basin > 10%1.0001.0001.0000.9941.0001.0001.0001.0001.0001.0001.000
Amerasian Basin > 1%1.0001.0001.0001.0001.0001.0001.0001.0001.0001.0001.000
Eurasian Basin > 1%1.0001.0001.0000.9991.0001.0001.0001.0001.0001.0001.000
Amerasian Basin (pixel area × concentration, normalized by total area)0.9860.9730.9900.9690.9740.9910.9850.9920.9930.9840.975
Eurasian Basin (pixel area times concentration, normalized by total area)0.9820.9580.9900.9100.9490.9820.9760.9920.9930.9710.969
Amerasian Basin > 10%0.8320.7380.9860.9930.9010.9710.8880.9710.7010.8510.865
Eurasian Basin > 10%0.7480.4940.9930.8070.8170.9820.8450.9630.7960.8800.900
Amerasian Basin > 1%0.8820.8120.9950.9970.9270.9840.9320.9910.7770.8910.871
Eurasian Basin > 1%0.8670.6390.9960.9130.8740.9940.9240.9900.8670.9170.905
Amerasian Basin (pixel area × concentration, normalized by total area)0.6100.4580.8510.8810.7440.8830.6500.8940.4790.6660.764
Eurasian Basin (pixel area times concentration, normalized by total area)0.3950.2190.8290.6640.5620.8910.5280.8640.4910.6640.780

7. Discussion

[31] Our goal in this paper is to present the differences among the models and from observations. Documenting differences among models is, we think, a worthwhile scientific goal, and a necessary step toward model improvement. We hope that modelers will use the results from this and similar papers to discuss model differences and determine why the models are different, and explore ways to improve them. A full discussion of the model details that may account for such differences is beyond the scope of this paper. A paper is in preparation discussing the causes for these differences. A few comments, however, at this stage of our analysis are worthwhile.

[32] Not surprisingly, multiple processes are at work in each of these models. For example, the LANL and the UW models tend to have higher albedos than many of the other models, with AWI1 and AWI2 having lower albedos and hence more ice. Likewise in the LANL model, the AOMIP imposed air temperatures were deemed too warm by about 2°C in the summer, melting too much ice, so air temperatures were limited to be at or below 0.5°C if there was more than 10% ice cover. From the present analysis, this may have been too strong a correction. The combination of high albedos and modified temperatures cause the LANL model to have more ice than many of the other AOMIP models.

[33] Another issue that can influence the sea ice concentration is whether or not the model uses an ice thickness distribution. The AWI models have seven thickness categories in each grid cell, ICM and LANL have five, and RAS has eight. Thinner ice that is available in multiple thickness categories is thought to melt more quickly and ridge more easily. Thus, a prescribed thickness distribution has the potential to allow for more open water formation due to ice pack deformation, and may exhibit more variability via the ice-open water albedo feedback.

[34] We also note that by separating the ice pack into a marginal ice zone (as defined by 10%) and the “main” pack, which is generally the multiyear part, we are isolating the behavior of the seasonal advance and retreat of the marginal ice zone. The seasonal variation in ice extent is largely determined by the modeled ocean heat flux, which in turn is determined partially by solar heating and transport from lower latitudes. Such functions differ among the various ocean models, making this diagnostic more useful for distinguishing among the ocean model behavior than for comparing sea ice models.

8. Summary

[35] Differences among the sea ice concentrations computed by the AOMIP models are greater than differences among four observational data sets. The observations show a robust annual cycle in the area with less than 10% that is not reproduced by some models, and some models do not reproduce the loss of sea ice in 1990 seen in other models and in all observational data sets. Regardless of the different model physics and parameters, the results here show that the models have more variability than observed, and that, compared to observations, almost all the models underestimate the September sea ice concentration in the central Arctic Ocean. This underestimation may have important implications for sea ice forecasts. These and other differences warrant further investigation.


[36] This research is supported by the National Science Foundation Office of Polar Programs under cooperative agreements OPP-0002239 and OPP-0327664 with the International Arctic Research Center, University of Alaska Fairbanks. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. E. Hunke is supported through the U.S. Department of Energy's Climate Change Prediction Program.