Under the influence of global warming, the Arctic is transitioning from a state dominated by multi-year thick ice to a "New Arctic" characterized predominantly by first-year thin ice. This younger ice is more fragile and prone to melting, which not only exacerbates the instability of the ice cover but also introduces new challenges for sea-ice prediction. Accurate forecasting of sea ice is of significant value for understanding the climate system and ensuring the safety of Arctic navigation. However, due to the combined influence of atmospheric, oceanic, and other factors, precise prediction remains a key international research focus.

Recently, associate professor Baoqiang Tian from the Institute of Atmospheric Physics, Chinese Academy of Sciences, and Professor Ke Fan from Sun Yat-sen University have developed a new real-time prediction method for September Arctic sea-ice extent, based on the interannual increment and stepwise regression approaches. The findings were published in Atmospheric and Oceanic Science Letters under the title "A novel stepwise regression method for predicting September Pan-Arctic sea-ice extent: comparison with long short-term memory neural networks."

The study shows that this method, which integrates initial sea-ice conditions with thermodynamic and dynamic processes, selects effective predictors through stepwise regression and incorporates the interannual increment approach, demonstrates high predictive performance for September Pan-Arctic sea-ice extent. Compared with LSTM (long short-term memory) neural networks, the new method exhibits smaller prediction errors and greater stability in independent tests from 2014 to 2022. Its prediction accuracy also surpasses that of the forecasts released by the Sea Ice Outlook. Although LSTM performs well during the training phase, its real-world prediction robustness is inferior to the new method—a limitation potentially attributable to the limited availability of sea-ice data, which may lead to overfitting in complex machine learning models.

Overall architecture of September Pan-Arctic SIE prediction models based on the SWR and LSTM methods. Credit: Baoqiang Tian.

Professor Ke Fan, corresponding author of the paper, explains that, "Our prediction method not only considers the independence of predictors to avoid overfitting, but also amplifies the prediction signal through the interannual increment approach, thereby enhancing the model's predictive capability."

This study offers a new perspective for improving seasonal predictions of Arctic sea ice.

Citaton: Baoqiang Tian, Ke Fan, 2025. A novel stepwise regression method for predicting September Pan-Arctic sea-ice extent: Comparison with long short-term memory neural networks, Atmospheric and Oceanic Science Letters, 100727, https://doi.org/10.1016/j.aosl.2025.100727.