Reanalyses Track Arctic Sea-Ice Decline While Coupled Regional Models Miss It

The Barents and Kara Seas have lost more sea ice and warmed at the surface faster than almost anywhere else in the Arctic, and a new evaluation shows that the tools scientists rely on to reconstruct those changes split sharply into two camps. Satellite measurements and three ocean reanalyses all record a statistically significant decline in sea ice area between 1991 and 2020, while two coupled regional climate models capture little or none of it. The study, published in Ocean Science, compared the EUMETSAT OSI SAF satellite record with two regional ocean reanalyses, TOPAZ 4B and RARE 1.15.2, the global reanalysis ORAS5, and the coupled regional models HIRHAM–NAOSIM 2.2 and RASM across the two shallow marginal seas north of Norway and Siberia.

The evaluation focused on spring, from March to May, and autumn, from August to October, because these seasons bracket the maximum and minimum of the regional sea ice cycle. In the Barents Sea, the satellite observations indicate a mean spring sea ice area of about 412,000 square kilometres and an autumn mean of about 29,000 square kilometres, with statistically significant decadal trends in both seasons. All three reanalyses reproduced significant sea ice area trends in both seasons. The coupled models did not. HIRHAM–NAOSIM showed only a near-significant decline in the Barents Sea, with trend p values between 0.05 and 0.1, and RASM displayed no clear long-term trend at all, only interannual variations. In the Kara Sea, neither model produced a statistically significant sea ice area trend in either season.

The reasons for the divergence between reanalyses and models lie in how the products are built. Ocean reanalyses combine a numerical ocean–sea ice model with observations through data assimilation, and they are driven by atmospheric reanalysis at the surface. TOPAZ assimilates satellite sea ice concentration observations directly, along with sea surface temperature, in situ temperature and salinity profiles, and, after 2010, winter sea ice thickness from CS2SMOS and sea surface salinity from SMOS. ORAS5 likewise assimilates satellite-based sea ice concentration information. RARE, by contrast, does not assimilate satellite sea ice concentration; it constrains sea ice indirectly through a surface heat flux relaxation based on sea ice thickness estimates from PIOMAS. The coupled models, HIRHAM–NAOSIM and RASM, evolve their own atmospheres inside the model domain, responding only to quasi-observed conditions at their lateral boundaries, so small differences in the initial state can lead the simulation along a different path than the real climate followed.

That structural difference shows up clearly in the numbers. ORAS5 and TOPAZ agreed most closely with the satellite record for sea ice area, sea ice concentration, and the position of the ice edge, and the authors attribute this partly to the assimilated observations that constrain their historical states. RARE reproduced the direction of the observed decline, and in autumn it even produced the largest statistically significant decadal sea ice area trend of all the reanalyses, but it overestimated the mean sea ice area, especially in autumn, and its spring trend was smaller than those of the other reanalyses and the satellites. Among the coupled models, HIRHAM–NAOSIM overestimated both spring and autumn mean sea ice area in the Barents Sea significantly, though its autumn mean was only about half that of RARE, while RASM had a smaller mean sea ice area bias but no trend. The authors note that the coupled models represent a possible pathway of the climate system rather than a direct reproduction of observed conditions, which is why exact agreement with observations cannot be expected.

The sea ice edge told a similar story through a different metric. The study measured the Integrated Ice Edge Error, the area where a product and the satellite record disagree on whether the ice concentration is above or below 15 percent. ORAS5 and TOPAZ had much smaller errors than the other datasets, and their overestimation-minus-underestimation trend decreased over time, indicating improving ice edge accuracy. HIRHAM–NAOSIM had the highest mean ice edge error and the highest mean overestimation, with the largest errors in December in the Barents Sea and October in the Kara Sea. RASM’s ice edge barely changed over the study period, so the model underestimated the edge at the start and significantly overestimated it at the end, producing an overestimation-minus-underestimation trend more than five times larger than RARE’s. The comparison demonstrated that a realistic total sea ice area does not guarantee a realistic ice edge: RARE reproduced observed sea ice area trends better than RASM, yet HIRHAM–NAOSIM, with a smaller sea ice area bias than RARE, carried a larger ice edge error.

Sea ice concentration maps sharpened the picture of where the models fall short. In spring, the satellite record showed significant negative concentration trends across most of the Barents Sea ice cover, with the largest decline, exceeding 0.2 per decade, northwest of Novaya Zemlya near the retreating ice edge. ORAS5 matched the satellite data most closely, with no large significant differences, and TOPAZ followed the same spatial trend pattern while overestimating concentration in the northern Barents Sea and the Kara Sea, which narrowed its marginal ice zone. RARE overestimated concentration over most of the ice-covered area, with differences above 0.4 south of Svalbard and west of Novaya Zemlya, and captured the observed decline only south of 75 degrees north in the Barents Sea. RASM produced the largest concentration differences of all, exceeding 0.6 south of Svalbard and along the west coast of Novaya Zemlya, and showed no widespread negative trends. HIRHAM–NAOSIM reproduced part of the observed spring decline, but only in small parts of the marginal ice zone in the Barents Sea.

In autumn the gap widened. The satellite record showed significant negative concentration trends across the entire ice-covered area, and the 2011–2020 mean ice edge had retreated to run north of the Barents Sea and along the northern boundary of the Kara Sea, compared with an edge reaching from Svalbard to Novaya Zemlya and the Yamal Peninsula in the 1991–2000 decade. TOPAZ and ORAS5 again showed no significant differences from the satellite data and nearly identical decadal trends. RARE overestimated the autumn ice edge and concentration more than any other dataset, retaining ice in areas where the satellites and the other reanalyses had become ice-free, which left more ice available to decline and pushed its maximum trend farther south. HIRHAM–NAOSIM showed a significant negative trend only in a small region east of Svalbard, and RASM had no meaningful autumn trend anywhere. The authors conclude that the seasonal retreat is too weak in both models, particularly in the Kara Sea.

The mixed-layer temperature results mirrored the sea ice findings. The reanalyses showed statistically significant warming trends in the Barents Sea in both spring and autumn, stronger in autumn, and in the Kara Sea the autumn warming trend approached 1 degree Celsius per decade, exceeding that rate over large areas. HIRHAM–NAOSIM reproduced the Barents Sea warming but showed no significant Kara Sea trend, where its sea ice was overestimated, and RASM produced no widespread significant trends in either region or season. The authors suggest that the colder mixed layers in the coupled models, particularly in autumn, are consistent with their tendency to maintain too much sea ice, since extensive ice cover shields the ocean from atmospheric forcing and limits seasonal warming. RARE showed anomalous behaviour in the Barents Sea, with spring warming concentrated where Atlantic water enters and a significant autumn cooling west of Novaya Zemlya, alongside a significant autumn deepening of the mixed layer in the same location, a contrast the authors link to how vertical mixing and seasonal heat storage are represented in that product.

Correlations with five climate indices added a further layer of evidence. For the satellite record and the reanalyses, the Atlantic Multidecadal Oscillation and the Northern Hemisphere ocean heat content showed the strongest relationships with sea ice and mixed-layer temperature anomalies, with the Atlantic Multidecadal Oscillation correlation about double that of the North Atlantic Oscillation. The reanalyses represented these relationships more consistently than the coupled models. In HIRHAM–NAOSIM, only the Atlantic Multidecadal Oscillation and ocean heat content correlations differed significantly from the observations, while in RASM the Atlantic Multidecadal Oscillation, Pacific Decadal Oscillation, and ocean heat content correlations all differed significantly, and in the Kara Sea both models’ correlations diverged broadly from the observed patterns. The authors interpret this as evidence that the influence of North Atlantic variability is not fully represented in the coupled models, consistent with earlier work showing that Gulf Stream warming contributes to Barents–Kara sea ice decline and that underestimating this warming can lead to underestimating sea ice loss.

The practical guidance that follows is straightforward. For studies requiring a realistic representation of the recent observed climate, ORAS5 and TOPAZ are the most suitable products, though users should remain aware of the influence of data assimilation on sea ice concentration and sea surface temperature. RARE captures the observed trend but overestimates sea ice and shows anomalous mixed-layer behaviour in the Barents Sea, warranting further research into whether the ocean model, the assimilation scheme, or the forcing is responsible. Among the models, the choice depends on the region and purpose: HIRHAM–NAOSIM is recommended over RASM for studies of change in the Barents Sea, while for the Kara Sea, where both models miss the trend, RASM’s better mean sea ice state may make it the preferable option. The stakes extend beyond these two seas, which cover roughly 11 percent of the Arctic Ocean yet accounted for 30 to 40 percent of recent Arctic winter sea ice loss, so improving their representation in coupled models matters for understanding Arctic and global climate change.

Subject of Research: Evaluation of how well ocean reanalyses and coupled regional models reproduce observed sea ice decline and mixed-layer warming in the Barents and Kara Seas from 1991 to 2020.

Article Title: Evaluation of sea ice variability in three ocean reanalyses and two regional models in the Barents and Kara Seas in 1991–2020

Article References: Äijälä, C., Gutiérrez-Loza, L., Dorn, W., Lauvset, S. K., Maslowski, W., & Uotila, P. (2026). Evaluation of sea ice variability in three ocean reanalyses and two regional models in the Barents and Kara Seas in 1991–2020. Ocean Science, 22(6), 3537-3566. https://doi.org/10.5194/os-22-3537-2026

Image Credits: AI Generated

DOI: 10.5194/os-22-3537-2026

Keywords: Arctic sea ice, Barents Sea, Kara Sea, ocean reanalysis, data assimilation, coupled regional models, sea ice concentration, mixed-layer temperature, climate indices, Atlantification, satellite observations, Ocean Science