Scholarlink

<p>Arctic sea ice concentration (SIC) prediction on a subseasonal scale plays an important role in polar navigation. To reduce the high uncertainty of daily forecasts, three time series prediction models are combined with empirical orthogonal function (EOF) decomposition to forecast Arctic pentad-mean SIC, where each month is divided into six pentad-means–the first five each span five days, and the last encompasses the remaining days, which may vary in length. The models were trained on SIC data from 1989 to 2018 and tested from 2019 to 2023, with lead times ranging from 1 to 12 pentad-means. Model skill was evaluated based on SIC spatial patterns, sea ice area (SIA), and the sea ice edge in September from 2019 to 2023. The moving-averaged 2-m temperature helps reduce the long short-term memory model’s error in the Beaufort and Chukchi Seas. Based on the models’ scores for each EOF time series, weighted ensemble prediction results were obtained. These results outperform two benchmark models across all lead times. In addition, the ensemble prediction better reproduces the seasonal cycle of the SIA, with relative errors ranging from 1.04% to 3.85%. The predicted September ice edge closely matches observations, with binary accuracy consistently above 90%. Forecast models show the lowest errors in the central Arctic, while relatively higher errors appear in the Barents and Kara Seas.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodel Ensemble Prediction of Pentad-Mean Arctic Sea Ice Concentration

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Shuo Zhao,
  • Jie Su

摘要

Arctic sea ice concentration (SIC) prediction on a subseasonal scale plays an important role in polar navigation. To reduce the high uncertainty of daily forecasts, three time series prediction models are combined with empirical orthogonal function (EOF) decomposition to forecast Arctic pentad-mean SIC, where each month is divided into six pentad-means–the first five each span five days, and the last encompasses the remaining days, which may vary in length. The models were trained on SIC data from 1989 to 2018 and tested from 2019 to 2023, with lead times ranging from 1 to 12 pentad-means. Model skill was evaluated based on SIC spatial patterns, sea ice area (SIA), and the sea ice edge in September from 2019 to 2023. The moving-averaged 2-m temperature helps reduce the long short-term memory model’s error in the Beaufort and Chukchi Seas. Based on the models’ scores for each EOF time series, weighted ensemble prediction results were obtained. These results outperform two benchmark models across all lead times. In addition, the ensemble prediction better reproduces the seasonal cycle of the SIA, with relative errors ranging from 1.04% to 3.85%. The predicted September ice edge closely matches observations, with binary accuracy consistently above 90%. Forecast models show the lowest errors in the central Arctic, while relatively higher errors appear in the Barents and Kara Seas.