Annual forecast challenge and other research highlight the complexity of sea ice prediction

In 2015, NOAA's Climate Program Office (CPO) invited grant proposals from sea ice and climate scientists looking to better understand and predict Arctic sea ice behavior, on timescales ranging from days to decades. This is our second story on some of the resulting research.

Bottoming out

In early October 2019, the National Snow and Ice Data Center, in Boulder, Colorado, announced the average monthly Arctic ice extent for September 2019 was 1.67 million square miles (4.32 million square kilometers). It was effectively tied for the second-lowest September in the continuous satellite record going back to 1979.

Sea ice extent dipped to its second-lowest level (tied with 2007 and 2016) in September 2019. Image by Climate.gov based on the National Snow and Ice Data Center’s Charctic tool.

The final extent was a number Uma Bhatt, a professor of atmospheric sciences at the University of Alaska Fairbanks, had been waiting for since June. As a member of the Sea Ice Prediction Network, Bhatt works with other researchers on a crowdsourcing project to predict Arctic sea ice extent for the month of September, the time of the summer minimum. The first call for forecasts is issued in June.

The effort—partially funded by NOAA, and involving dozens of partners from around the world—is more than a scientific guessing game. A range of activities are dictated by the timing of the fall freeze-up, from the mandatory cessation of oil and gas drilling to the length of the Alaska native hunting and fishing seasons.

Plus, Bhatt explains, “The northern part of Alaska isn’t on a road system, and they have to ship everything. If the marine industry knows what the sea ice will be like, they can plan their barge trips better based on the price of fuel. They can say, ‘Okay, the price of fuel is going to go down, I can afford to wait a month.’ If we make progress in forecasting, that has a huge economic benefit,” she says. “But it's a hard science problem.”

Simple formulas give mixed results

The simplest approaches to predicting sea ice behavior take advantage of big-picture patterns and trends in Arctic-wide sea ice extents. Extrapolating from the long-term trend in sea ice decline, for example, or using the last date sea ice retreats from a region in the spring to estimate when it will re-advance in autumn. A 2016 study led by Julienne Stroeve, affiliated with NSIDC and University College, London, found that the latter method predicts well in some Arctic regions, well in some years in other Arctic regions, and not at all in others.

The smaller seas comprising the Arctic Ocean vary in climatology as well as location. Image by NOAA Climate.gov.

This approach has a chance of working because sea ice has a high albedo: It reflects a majority of incoming sunlight. The earlier the ice retreats in the spring, the longer the ocean has to absorb incoming sunlight, and the longer the fall re-freeze may be delayed. Still, explains Stroeve, “Ice-albedo feedback isn’t the sole driver of the evolution of summer ice cover.” Other factors can complicate the relationship. Some of those complications are well known: storms and other atmospheric influences, ice thickness and how it affects drift. But research continues to turn up new ones.

Stroeve, for example, was part of a 2016 study funded by NOAA’s Climate Program Office that linked the timing of the spring thaw in the Laptev Sea to the retreat of snow cover in the West Siberian Plain, more than 1,600 miles to the east. “As snow retreats, land warms,” explains Stroeve. The land surface heats the overlying air. “Using back trajectory analysis, we can see where that heat ends up, and the link to the West Siberian Plain was strongest in the Laptev Sea.”

Julienne Stroeve was one of hundreds of researchers to participate in Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC), which began in September 2019. The mission embedded the German research icebreaker Polarstern in Arctic sea ice to study the region’s ice, ocean, atmosphere, and ecosystem for the next year. Research often entailed braving fierce storms and, in winter, round-the-clock darkness. Image courtesy Stefan Hendricks, MOSAiC.

Another influence is right in the ocean itself. A 2019 NOAA-funded study found that decaying organic matter flushed into the ocean by Arctic rivers can lead to “ocean yellowing,” which simultaneously heats surface waters and cools deeper waters. The net influence on sea ice is complicated. Surface heating may hasten summer melt and delay fall re-freeze. But when the deep, cold water resurfaces in winter, it can accelerate sea ice formation and thickening.

Crowdsourcing sea ice forecasts

The annual September Arctic sea ice minimum is the most closely watched metric of a warming Arctic. Since 2008, the Sea Ice Prediction Network team has been leveraging that interest to study different forecast strategies. Started as an informal, grassroots effort, the team puts out a call in early summer for individuals or research groups to submit forecasts for the September average ice extent.

Predicting the minimum attracts the efforts of both experts and amateurs. Contributors generally use three approaches—dynamical, statistical, or heuristic—or sometimes a mixture. Dynamical models simulate the ocean, atmosphere, and sea ice with mathematical equations. Statistical approaches forecast future events based on past trends and variability. Heuristic forecasts can be characterized as educated guesses.

Participants can submit September forecasts starting in June, and they can update them each month through August. The predictions are bundled together and made available to the public as the annual Sea Ice Outlook. Then the waiting begins. 

In 2019, the project received 31 forecasts in June, 39 forecasts in July, and 42 forecasts (a record number) in August. Mitch Bushuk, a research scientist at NOAA’s Geophysical Fluid Dynamics Laboratory, submitted a forecast using a dynamical model. He has found that predicting summer and predicting winter sea ice extents pose different challenges.

“The key to predicting sea ice is asking, ‘What direction is the sea ice edge moving?’” Bushuk explains. “In winter, the ice edge is moving southward, and it will keep growing until it bumps into ocean temperatures too warm for it to keep growing. In summer, the edge is moving northward toward the pole, and how far it can melt backward is a function of how thick the sea ice is.”

In both seasons, accurate measurements of initial conditions are crucial. “If we know the subsurface heat content, or temperature, that gives us prediction skill for the following winter,” Bushuk says. For the summer minimum, “the key source of predictability is the initial thickness in spring and early summer.”

His previous research indicates that in some regions, including the Barents, Greenland, Labrador, and Bering Seas, and the Sea of Okhotsk, winter ice extent can potentially be predicted as far as 12 to 36 months in advance. Summer minimums are trickier. During the Northern Hemisphere spring, factors affecting melt—timing of melt onset and inherent chaos in the atmosphere, for example—vary so much that accurate dynamical forecasts for the September minimum can’t be generated before May, conservatively before June 1. Research by Bushuk and others suggests the existence of a melt-season “predictability barrier” that hampers prediction early in the spring. Whether this barrier genuinely exists in nature, or whether it’s a limitation of the current forecasting system is an open question.

Compared to ensemble models initialized at other times of year, models initialized in May showed the earliest divergence and loss of skill in predicting sea ice volume. Image by NOAA Climate.gov based on Bushuk et al. 2019.

What might break through the barrier are reliable estimates of sea ice thickness recorded in late spring and early summer. The European Space Agency’s CryoSat-2 provides sea ice thickness estimates, and Bushuk has relied on thaw data in the past—at other times of year. But melt ponds forming on the sea ice surface hamper data collection, so the CryoSat-2 team halts data release around April 15. So current CryoSat-2 observations can’t help him break the spring predictability barrier. Efforts continue to extend reliable CryoSat-2 thickness data into May and even June. NASA’s recently launched ICESat-2 mission may also break the barrier.

Though pretty to human eyes, melt ponds on sea ice can confound satellite sensors aiming to gauge ice thickness. Photo courtesy of NASA Operation IceBridge.

Picking the winner

Each year, the Sea Ice Prediction Network team releases a post-season report describing how the different approaches fared. Bhatt led the 2018 and 2019 post-season reports, which she describes as collaborative efforts. Asked to pick a winning technique, she replies, “There’s never a simple answer. I would say typically the dynamical models do better, though in some years, the statistical models do pretty well.”

The winning approach also changes with the lead time. For forecasts made in June and July, statistical methods came closest to matching the observed September 2019 extent. For forecasts issued in August, however, dynamical methods performed best. Forecasts made using statistical methods estimated too low. Heuristic methods generally underestimated September extent at all lead times.

The Sea Ice Prediction Network’s 2019 post-season report summarized the predictions received, categorized by model type, and how close they came to predicting the actual minimum extent. The rectangles show the middle 50 percent of the values, with the horizontal lines through the boxes indicating the median predictions. The whiskers indicate the lowest and highest 20 percent of predicted values. The colored dots show outliers—values that are very high or low compared to other predictions. Graph adapted from SIPN 2019 Post-Season Report.

In 2016, Stroeve coauthored a study with Lawrence Hamilton of the University of New Hampshire reviewing the successes and failures of the 400 forecasts that had been submitted as part of the Sea Ice Outlook project between 2008 and 2015. The most obvious finding wasn’t about differences in skill between different forecasting methods; it was about differences in skill between years.

Regardless of the forecast strategy, some years appear to be easy to predict, and others are hard. In easy years, the September extent doesn’t deviate much from the prior year, and many forecasts come reasonably close to the observed extent. In hard years, unusual weather events or anomalies lead to big differences from the prior year. Unusual years stymie all methods.

Still, the study found that forecasts based on statistics and those based on ice-ocean-atmosphere (dynamic) models modestly out-performed other approaches, including dynamic models incorporating only ice and ocean. Because statistical methods predict ice extent based on previous years, they do especially well when the sea ice minimum is close to the long-term average.

More hard years ahead

Participation in the Sea Ice Prediction Network grows yearly, but so do the challenges for prediction. Asked whether climate change is making accurate prediction more difficult, Bhatt responds, “Absolutely.”

The loss of thick, perennial ice contributes to that difficulty. According to the Arctic Report Card: Update for 2019, sea ice older than four years comprised 33 percent of the Arctic Ocean ice pack in March 1985. In March 2019, such old ice comprised just 1.2 percent of the ice pack. First-year ice (formed in the most recent fall and winter) now constitutes 70 percent of the Arctic sea ice cover, compared to 35–50 percent in the 1980s. In contrast to old ice, first-year ice is thin. It’s more prone to melt, and more prone to move with ocean currents and winds. Ocean basins may oscillate between ice-covered and ice-free states from year to year.

In short, year-to-year variability will increase in the 21st century. That’s the conclusion of a 2019 study in The Cryosphere led by John Mioduszewski. Using a suite of climate models, Mioduszewski and his colleagues found that when sea ice area retreats to an average of about 3–4 million square kilometers, year-to-year variability—and by extension, forecasting challenges—peaks. For winter months, this peak won’t come until late this century. For summer ice area, though, we’re already on the threshold. Models indicate the period of maximum unpredictability will be the 2020s–2030s. (Sea ice area is calculated slightly differently than sea ice extent, with extent values tending to be higher. The National Snow and Ice Data Center offers a primer on how the measurements differ.)

Mioduszewski and coauthors compared year-to-year sea ice area variations by month. The blue lines show differences in sea ice area, and the orange lines show running standard deviations. The orange lines show where the interannual areal differences peak. For the months of August–October, the period of maximum variability is predicted to occur over the next couple decades. Image adapted from Mioduszewski et al. 2019.

Marika Holland, senior scientist at the National Center for Atmospheric Research, was a coauthor of the 2019 study. She explains why increasing variability poses risks to Arctic operations, saying, “If you can reliably know this area will be open or sea-ice-covered, you can use the ocean there for hunting or shipping. You have knowledge of what sea ice will be like. The challenge with increases in variability is now you don't know.”

After the 2030s, the study predicts, variability will plummet for the simple reason that sea ice will retreat completely from the Arctic in summer. Steve Vavrus, senior scientist at the Nelson Institute for Environmental Studies at the University of Wisconsin-Madison is another coauthor of the study. When asked to predict just which year the Arctic will lose all its summertime ice, he responds, “We can't say precisely. When it actually will occur for the first time will depend on weather conditions in a particular year, and we have no way of knowing that [decades] in advance.”

Whatever the weather, Vavrus suspects—based on where the last remnants of thick, multiyear ice are now—that ice will linger longest along the northern edge of the Canadian Arctic Archipelago. The location where summer sea ice will last the longest may be the one thing sea ice scientists and climate scientists can predict with high confidence.

References

Bonan, D.B., Bushuk, M., Winton, M. (2019). A spring barrier for regional predictions of summer Arctic sea ice. Geophysical Research Letters, 46, 5937–5947. https://doi.org/10.1029/2019GL082947.

Bushuk, M., Msadek, R., Winton, M., Vecchi, G.A., Gudgel, R., Rosati, A., Yang, X. (2017). Skillful regional prediction of Arctic sea ice on seasonal timescales. Geophysical Research Letters, 44, 4953–4964. https://doi.org/10.1002/2017GL073155.

Bushuk, M., Msadek, R., Winton, M., Vecchi, G., Yang, X., Rosati, A. (2019). Regional Arctic sea–ice prediction: potential versus operational seasonal forecast skill. Climate Dynamics, 52, 2721–2743. https://doi.org/10.1007/s00382-018-4288-y.

Bushuk, M., Yang, X., Winton, M., Msadek, R., Harrison, M., Rosati, A., Gudgel, R. (2019). The value of sustained ocean observations for sea ice predictions in the Barents Sea. Journal of Climate. https://doi.org/10.1175/JCLI-D-19-0179.1.

Crawford, A.D., Horvath, S., Stroeve, J., Balaji, R., Serreze, M.C. (2018). Modulation of sea ice melt onset and retreat in the Laptev Sea by the timing of snow retreat in the West Siberian Plain. Journal of Geophysical Research: Atmospheres, 123, 8691–8707. https://doi.org/10.1029/2018JD028697.

Gnanadesikan, A., Kim, G.E., Pradal, M.A.S. (2019). Impact of colored dissolved materials on the annual cycle of sea surface temperature: Potential implications for extreme ocean temperatures. Geophysical Research Letters, 46, 861–869. https://doi.org/10.1029/2018GL080695.

Hamilton, L.C., Stroeve J. (2016). 400 predictions: the SEARCH Sea Ice Outlook 2008–2015. Polar Geography, 39(4), 274–287. https://doi.org/10.1080/1088937X.2016.1234518.

Mioduszewski, J.R., Vavrus, S., Wang, M., Holland, M., Landrum, L. (2019). Past and future interannual variability in Arctic sea ice in coupled climate models. The Cryosphere, 13, 113–124. https://doi.org/10.5194/tc-13-113-2019.

Perovich, D., Meier, W., Tschudi, M., Farrell, S., Hendricks, S., Gerland, S., Kaleschke, L., Ricker, R., Tian-Kunze, X., Webster, M., Wood, K. (2019). Sea Ice. Arctic Report Card: Update for 2019.

Sea Ice Prediction Network (SIPN2) https://www.arcus.org/sipn. Accessed February 19, 2020.

Stroeve, J.C., Crawford, A.D., Stammerjohn, S., (2016). Using timing of ice retreat to predict timing of fall freeze‐up in the Arctic. Geophysical Research Letters, 43, 6332–6340. https://doi.org/10.1002/2016GL069314.

Stroeve, J.C., Mioduszewski, J.R., Rennermalm, A., Boisvert, L.N., Tedesco, M., and Robinson, D. (2017). Investigating the local-scale influence of sea ice on Greenland surface melt. The Cryosphere, 11, 2363–2381. https://doi.org/10.5194/tc-11-2363-2017.