Prediction of Pan-Arctic Sea Ice Using Attention-Based LSTM Neural Networks

• Published in: Frontiers in Marine Science
• Main Authors: Jianfen Wei, Renlong Hang, Jing-Jia Luo
• Format: Article in Journal/Newspaper
• Language: English
• Published: Frontiers Media S.A. 2022
• Subjects: Arctic; sea ice; forecast; deep learning; attention-based LSTM; Science; Q; General. Including nature conservation; geographical distribution; QH1-199.5; Climate change; Sea ice
• Online Access: https://doi.org/10.3389/fmars.2022.860403; https://doaj.org/article/8a58ffe94b8d4660b40fc185ded43d7a

Within the rapidly changing Arctic region, accurate sea ice forecasts are of crucial importance for navigation activities, such as the planning of shipping routes. Numerical climate models have been widely used to generate Arctic sea ice forecasts at different time scales, but they are highly dependent on the initial conditions and are computationally expensive. Recently, with the increasing availability of geoscience data and the advances in deep learning algorithms, the use of artificial intelligence (AI)-based sea ice prediction methods has gained significant attention. In this study, we propose a supervised deep learning approach, namely attention-based long short-term memory networks (LSTMs), to forecast pan-Arctic sea ice at monthly time scales. Our method makes use of historical sea ice concentration (SIC) observations during 1979–2020, from passive microwave brightness temperatures. Based on the persistence of SIC anomalies, which is known as one of the dominant sources of sea ice predictability, our approach exploits the temporal relationships of sea ice conditions across different time windows of the training period. We demonstrate that the attention-based LSTM is able to learn the variations of the Arctic sea ice and can skillfully forecast pan-Arctic SIC on monthly time scale. By designing the loss function and utilizing the attention mechanism, our approach generally improves the accuracy of sea ice forecasts compared to traditional LSTM networks. Moreover, it outperforms forecasts with the climatology and persistence based empirical models, as well as two dynamical models from the Copernicus Climate Change Service (C3S) datastore. This approach shows great promise in enhancing forecasts of Arctic sea ice using AI methods.