As summer-warmed seas nibble away at the ice’s buried edges, the sea ice floating atop the Arctic Ocean will fall to its smallest size this year in the next week or so.
Sea ice levels are unlikely to reach new lows this year, according to scientists. At its lowest point in 2020, ice covered 3.74 million square kilometers of the Arctic, coming dangerously close to an all-time low. Sea ice covers little under 5 million square kilometers of Arctic waters right now, putting it on course to be the tenth-lowest extent of sea ice in the region since satellite records began in 1979. Given that sea ice reached a record low for this time of year in early June, the outcome was unexpected.
The surprise stems in part from the fact that the most advanced statistical and physics-based forecasting systems can accurately anticipate sea ice extent just a few weeks in ahead, while long-range forecasts are less accurate. Researchers write in Nature Communications on August 26 that a new method that employs artificial intelligence to construct sea ice forecasts promises to improve their accuracy — and can complete the analysis rather rapidly.
According to Tom Andersson, a data scientist with BAS’s Artificial Intelligence lab, IceNet, a sea ice forecasting system developed by the British Antarctic Survey, or BAS, is “95 percent accurate in forecasting sea ice two months ahead — higher than the leading physics-based model SEAS5 — while running 2,000 times faster.” IceNet can create a forecast in less than 10 seconds on a laptop, but SEAS5 takes around six hours on a supercomputer. Andersson and his colleagues discovered that the system had a startling ability to forecast anomalous ice events — extraordinary highs or lows — up to four months ahead of time.
Keeping tabs on the effects of climate change requires tracking sea ice. While that is more of a long-term strategy, IceNet’s earlier notice could have more immediate benefits. For example, it might supply data that Indigenous communities need to make economic and environmental decisions, as well as giving scientists the time they need to assess and plan for the hazards of Arctic fires or wildlife-human interactions.
Since 1979, when satellite data began, Arctic sea ice extent has consistently decreased in all seasons (SN: 9/25/19). For decades, scientists have attempted to improve sea ice forecasts, but success has eluded them. “Forecasting sea ice is extremely difficult because sea ice interacts with the atmosphere above and the water below in intricate ways,” explains Andersson.
The Arctic sea ice extent decreased to its second lowest level since satellite tracking began in 1979 in 2020. These observations were used to create this animation, which depicts the shift in sea ice coverage from March 5, when the ice was at its thickest, to September 15, when it was at its thinnest. From 1981 to 2010, the yellow line depicts the average minimum extent. These changes may now be predicted weeks in advance using current forecasting technologies. Several months ahead of time, a new AI-based technique can forecast these changes with nearly 95% accuracy.
To predict how sea ice will change in the future, existing forecast technologies convert physics laws into computer code. However, these models struggle to create accurate long-range projections, in part due to uncertainty in the physical mechanisms that drive sea ice.
Andersson and his colleagues used a technique known as deep learning to load observational sea ice data from 1979 to 2011 as well as climate simulations from 1850 to 2100 into IceNet to teach it how to forecast the state of future sea ice by analyzing historical data.
The team compared IceNet’s outputs to observed sea ice extent from 2012 to 2020, as well as projections generated by SEAS5, a widely cited tool used by the European Centre for Medium-Range Weather Forecasts, to determine the accuracy of its forecasts. IceNet was 2.9 percent more accurate than SEAS5, resulting in an additional 360,000 square kilometers of ocean correctly categorized as “ice” or “no ice.”