中国科学院大气物理研究所李熙晨的团队报道了泛北极海冰厚度一个月预测的数据驱动深度学习模型的评估。该项研究成果发表在2024年1月5日出版的《大气科学进展》上。

研究基于ConvLSTM和全卷积U-net (FC-Unet)算法，构建了两个数据驱动的深度学习模型，并对CMIP6历史模拟进行迁移学习和再分析/观察以进行微调。这样就可以在不考虑所涉及的复杂物理过程的情况下，对北极SIT进行月度预测。通过对季节和区域预测能力的综合评估，结果表明，使用更广泛的CMIP6数据集进行迁移学习，以及将多个气候变量作为预测因子，有助于获得更好的预测结果。

两种DL模型均能有效预测SIT异常的时空特征。FC-Unet模型预测的SIT异常，特别是与再分析的空间相关性在各月份均达到89%的平均水平，而时间异常相关系数在大多数情况下接近1。此外，该模型在预测极端事件期间的SIT和SIE方面也表现出较好的效果。有目的深度迁移学习模型在预测北极SIT中的有效性和可靠性，可以促进更准确的泛北极预测、气候变化研究和实时商业应用。

据悉，近年来，深度学习方法已逐渐应用于北极海冰浓度的预测任务，但由于观测和再分析数据的时间覆盖范围有限，对更大时间和空间尺度的研究相对较少。与此同时，深度学习预测海冰厚度(SIT)的研究仍然很少受到关注。

附：英文原文

Title: Assessments of data-driven deep learning models of one-month prediction of pan-Arctic sea ice thickness

Author: Chentao Song, Jiang Zhu, Xichen Li

Issue&Volume: 2024-01-05

Abstract: In recent years, deep learning methods have been gradually applied to the prediction tasks of Arctic sea ice concentration, but relatively few works have been done for larger spatial and temporal scales, mainly due to the limited time coverage of observations and reanalysis data. Meanwhile, deep learning predictions of sea ice thickness (SIT) have still received little attention. In this study, based on the ConvLSTM and fully convolutional U-net (FC-Unet) algorithms, two data-driven deep learning models are built and trained using CMIP6 historical simulations for transfer learning and reanalysis/observations for fine-tuning. These enable monthly predictions of Arctic SIT without considering the complex physical processes involved. Through comprehensive assessments of prediction skills by season and region, the results suggest that using a broader set of CMIP6 data for transfer learning, as well as incorporating multiple climate variables as predictors, contribute to better prediction results. And both two DL models can effectively predict the spatiotemporal features of SIT anomalies. The predicted SIT anomalies of FC-Unet model, in particular, the spatial correlations with reanalysis reach an average level of 89% over all months, while temporal anomaly correlation coefficients are close to 1 in most cases. Besides, the models also show robust performance in predicting SIT and SIE during extreme events. The effectiveness and reliability of purposed deep transfer learning models in predicting Arctic SIT canfacilitate more accurate pan-Arctic predictions, climate change research and real-time business applications.

DOI: 10.1007/s00376-023-3259-3

Source: http://www.iapjournals.ac.cn/aas/en/article/doi/10.1007/s00376-023-3259-3viewType=HTML