Presentation Information
[ACG68-P04]Development of a sea-ice prediction model with a deep learning
*Jun Ono1, Motomu Oyama2, Hironori Yabuki1 (1.National Institute of Polar Research, 2.Graduate School of Engineering, Kogakuin University)
Sea ice is one of the most important subsystems of the Earth’s climate system. Through the reflection of solar radiation (the albedo effect), air–sea heat exchange, and water and thermohaline circulations, it plays a fundamental role in regulating the global energy, heat, and water budgets. Observed sea-ice extent has decreased in the Arctic since the beginning of the satellite era and in the Antarctic since approximately 2016. In 2025, global sea-ice extent reached a record minimum in February. In addition, the winter Arctic sea-ice extent recorded the lowest annual maximum in March, intensifying both scientific and societal concerns. Sea-ice reduction is a prominent indicator of ongoing global warming; therefore, improvements in sea-ice prediction accuracy are highly significant from a societal perspective. Traditionally, sea-ice prediction has relied mainly on numerical and statistical methods.
Meanwhile, machine-learning and deep-learning models capable of learning complex spatiotemporal dependencies from large datasets have been developed (e.g., Andersson et al., 2021). In this study, as a first step, we performed several experiments to predict the annual minimum Arctic sea-ice extent using a Long Short-Term Memory (LSTM) model. Then we investigated how well the learns learn seasonal cycles and interannual variability. The preliminary results indicate that data after May 1 are critical and that the model can reasonably reproduce interannual variations in the annual minimum Arctic sea-ice extent using data up to July 31, considering the lead time. The predicted range of the 2025 annual minimum Arctic sea-ice extent is estimated to be 4.08–4.60 million km². Future improvements will include extending the model to predict not only the minimum value but also its timing. In addition, we are currently developing a model to predict sea-ice concentration distributions using a convolutional LSTM.
Acknowledgements. This work was a part of the Arctic Challenge for Sustainability 3 (ArCS-3), Program Grant Number JPMXD1720251001.
References
Andersson, T. R., et al., 2021: Seasonal Arctic sea ice forecasting with probabilistic deep learning.
Nat. Commun., https://doi.org/10.1038/s41467-021-25257-4.
Meanwhile, machine-learning and deep-learning models capable of learning complex spatiotemporal dependencies from large datasets have been developed (e.g., Andersson et al., 2021). In this study, as a first step, we performed several experiments to predict the annual minimum Arctic sea-ice extent using a Long Short-Term Memory (LSTM) model. Then we investigated how well the learns learn seasonal cycles and interannual variability. The preliminary results indicate that data after May 1 are critical and that the model can reasonably reproduce interannual variations in the annual minimum Arctic sea-ice extent using data up to July 31, considering the lead time. The predicted range of the 2025 annual minimum Arctic sea-ice extent is estimated to be 4.08–4.60 million km². Future improvements will include extending the model to predict not only the minimum value but also its timing. In addition, we are currently developing a model to predict sea-ice concentration distributions using a convolutional LSTM.
Acknowledgements. This work was a part of the Arctic Challenge for Sustainability 3 (ArCS-3), Program Grant Number JPMXD1720251001.
References
Andersson, T. R., et al., 2021: Seasonal Arctic sea ice forecasting with probabilistic deep learning.
Nat. Commun., https://doi.org/10.1038/s41467-021-25257-4.