Satellite Data Assimilation Sharpens Antarctic Sea Ice Forecasts at Subseasonal Scales

Antarctic sea ice has confounded scientists for decades. Unlike its Arctic counterpart, which has declined almost continuously since the satellite era began, the sea ice surrounding Antarctica grew modestly from the late 1970s until around 2015, and then swung dramatically in the opposite direction. In the austral spring of 2016, the ice retreated at an unprecedented rate, producing the lowest summer sea ice extent on record in 2017. Record minima followed in February 2022 and again in February 2023, and the ice cover has remained abnormally low ever since. As commercial shipping, fishing, and tourism accelerate in the Southern Ocean, the need for reliable forecasts of where and when the ice will appear has become urgent. A new study published in The Cryosphere by Yong-Fei Zhang of Princeton University and colleagues at the NOAA Geophysical Fluid Dynamics Laboratory demonstrates that a well-established forecasting technique, when applied to the often-neglected Antarctic, can significantly sharpen predictions of sea ice on subseasonal timescales of up to 45 days.

The technique in question is data assimilation, the mathematical machinery by which numerical models are continuously corrected with real-world observations. The team focused on sea ice concentration, the fraction of ocean surface covered by ice, as measured by passive microwave satellite sensors and processed with the NASA Team algorithm at the National Snow and Ice Data Center. These observations were fed into the sea ice component of the Geophysical Fluid Dynamics Laboratory’s global ice-ocean model using the Data Assimilation Research Testbed, a community software facility. At the heart of the system sits an Ensemble Adjustment Kalman Filter, a statistical method that blends a spread of model simulations with observations, weighting each according to its uncertainty. With a horizontal localization radius of roughly 170 kilometers, the filter directly updated the concentration of each of the model’s five sea ice thickness categories, while the aggregate thicknesses of ice and snow adjusted implicitly in response. The result was a reanalyzed sea ice state, spanning 1992 to 2017, that reflects both the physics of the model and the reality captured from orbit.

To test whether this improved initial state actually translates into better forecasts, the researchers ran two parallel suites of reforecast experiments with the SPEAR_MED prediction system, part of GFDL’s Seamless System for Prediction and EArth System Research. SPEAR couples the Modular Ocean Model version 6, the Sea Ice Simulator version 2, an atmospheric model, and a land model, all sharing a nominal one-degree resolution for ocean and ice. The baseline suite, called SPEAR, was initialized the standard way: ocean conditions came from an ocean data assimilation experiment, while the atmosphere was nudged toward the NCEP Climate Forecast System Reanalysis and sea surface temperature toward NOAA’s Optimal Interpolation product. The second suite, called SICDA, was identical in every respect except that its sea ice initial conditions came from the new assimilation product. Each experiment comprised fifteen ensemble members, initialized on the first day of every month from 1992 through 2017 and integrated for a full year, with the analysis focused on the first 45 days.

The comparison of initial conditions alone was striking. The standard SPEAR initial state overestimated Antarctic sea ice extent in most seasons, with the largest positive biases in winter and spring across the Weddell, Amundsen-Bellingshausen, Indian, and West Pacific sectors, while underestimating summer extent in the West Pacific and Ross sectors. After assimilating sea ice concentration, these biases shrank substantially in nearly all months and sectors. The wintertime and springtime positive biases in the Weddell and Ross sectors, regions notorious for the open-water polynyas that the unperturbed model simulated too actively, were almost entirely eliminated. The assimilated state also tracked the observed year-to-year variability of regional sea ice extent more faithfully, showing higher detrended anomaly correlation coefficients, and it reduced the root-mean-square error of grid-cell-level concentration almost everywhere, with the greatest gains in the peripheral zone where ice grows and retreats seasonally.

An important and somewhat cautionary side effect emerged in the ice thickness fields. Because the underlying model carried a systematic positive bias in concentration, the assimilation tended to reduce concentration, and thickness followed in the same direction. The assimilated initial state was thinner than the standard one nearly everywhere, with exceptions in the interior Weddell and Ross sectors, where the ice actually thickened. Lacking direct observations of Antarctic sea ice thickness, the team benchmarked their results against two independent model-based references, the Global Ice-Ocean Modeling and Assimilation System and the Southern Ocean State Estimate, both of which reproduce the broad patterns of Antarctic ice volume despite their own biases. The comparison revealed that the four datasets disagree considerably on the magnitude and even the timing of the annual maximum in ice volume, underscoring how poorly constrained Southern Hemisphere ice thickness remains.

When the two forecast suites were evaluated against satellite observations, the assimilation paid clear dividends. Measured by the detrended anomaly correlation coefficient of sea ice concentration, SICDA outperformed SPEAR in nearly all initialization months, with the largest gains in austral spring, followed by winter and summer, and only minor differences in autumn. Both dynamical systems initially trailed a statistical benchmark, a damped anomaly persistence forecast that extrapolates the most recent observed anomaly while accounting for its expected decay, but they overtook it within the first week when averaged across the year. Notably, the skill advantage of the assimilated forecasts decayed gradually with lead time rather than vanishing immediately, and the magnitude of improvement was larger in the Antarctic than in parallel Arctic experiments, where the greatest benefit had previously been found in summer rather than spring.

The improvements extended to the sea ice edge, arguably the most operationally relevant forecast target for ships navigating ice-infested waters. The team quantified edge skill with the Spatial Probability Score, normalized against a trend climatology so that values below one indicate skill exceeding climatology. In winter and spring-initialized forecasts, the standard SPEAR system showed essentially no edge skill, with normalized scores at or above one throughout the 45-day window, reflecting large initial errors in edge position. SICDA transformed this picture: its scores were significantly lower than SPEAR’s for the entire 45 days in winter and about 40 days in spring, and it began beating the damped persistence benchmark around day 15 in winter. Averaged across all twelve initialization months, SICDA maintained significantly better edge skill than SPEAR for 35 days in the Antarctic, compared with only 22 days in the Arctic, and it outperformed the benchmark after roughly day 10 annually, a feat the standard system never achieved in the Southern Ocean.

The study is candid about its one notable failure mode. In forecasts initialized in December and January, SICDA initially led SPEAR but lost ground after about 20 days, with the degradation concentrated along the Antarctic coast and in the interior Weddell and Ross sectors. The mechanisms differ between these regions. Along the coast, the assimilated state featured lower concentration and thinner ice, which through the ice-albedo feedback absorbed more solar radiation, warmed the sea surface by up to 0.8 degrees Celsius relative to the standard forecasts, and accelerated bottom melt, compounding the initial deficit. In the interior Weddell and Ross sectors, by contrast, the assimilated state carried more extensive and thicker ice, which slowed an already sluggish model melt rate and allowed positive concentration biases to grow. The authors suggest that adjusting highly coupled fields such as concentration and thickness through assimilation can introduce inconsistencies that only surface at longer lead times, a caution relevant to any data assimilation system.

The broader implications reach beyond the Southern Ocean. Previous assessments of subseasonal prediction systems in the S2S database found that only one system outperformed simple benchmarks for Antarctic ice edge forecasts, and that all systems fared worse in the Antarctic than the Arctic, a gap attributed largely to deficient initial conditions. By showing that assimilating a single, widely available satellite observable can close much of that gap, the study offers a relatively inexpensive path to better polar forecasts, and the authors note that their assimilated system performs comparably to the leading European Centre for Medium-Range Weather Forecasts system in both hemispheres. The work also highlights a persistent observational deficit: without ground truth for ice thickness, it remains impossible to confirm whether the assimilation adjusts volume in the right direction, strengthening the case for expanded thickness measurements from satellites, aircraft, and ships. As Antarctic sea ice enters what may be a new, lower state, the ability to anticipate its fluctuations weeks in advance has shifted from an academic curiosity to a practical necessity, and this study shows that the tools to meet that challenge are already within reach.

Subject of Research: Improving subseasonal Antarctic sea ice forecasts through satellite sea ice concentration data assimilation

Article Title: Subseasonal forecast improvements from sea ice concentration data assimilation in the Antarctic

Article References: Zhang, Y.-F., Bushuk, M., Winton, M., Gregory, W., Hurlin, B., Jia, L., & Lu, F. (2026). Subseasonal forecast improvements from sea ice concentration data assimilation in the Antarctic. The Cryosphere, 20(10), 5867-5882. https://doi.org/10.5194/tc-20-5867-2026

Image Credits: AI Generated

DOI: 10.5194/tc-20-5867-2026

Keywords: Antarctic sea ice, data assimilation, subseasonal forecasting, sea ice concentration, sea ice edge, SPEAR model, GFDL, The Cryosphere, sea ice thickness, Kalman filter, Southern Ocean, polar prediction