IceTFT v1.0.0: interpretable long-term prediction of Arctic sea ice extent with deep learning

• Published in: Geoscientific Model Development
• Main Authors: B. Mu, X. Luo, S. Yuan, X. Liang
• Format: Article in Journal/Newspaper
• Language: English
• Published: Copernicus Publications 2023
• Subjects: Geology; QE1-996.5; Arctic; Climate change; Global warming; Sea ice
• Online Access: https://doi.org/10.5194/gmd-16-4677-2023; https://doaj.org/article/054dcd58ba1349a4946166f117698908

Due to global warming, the Arctic sea ice extent (SIE) is rapidly decreasing each year. According to the Intergovernmental Panel on Climate Change (IPCC) climate model projections, the summer Arctic will be nearly sea-ice-free in the 2050s of the 21st century, which will have a great impact on global climate change. As a result, accurate predictions of Arctic sea ice are of significant interest. In most current studies, the majority of deep-learning-based SIE prediction models focus on one-step prediction, and they not only have short lead times but also limited prediction skill. Moreover, these models often lack interpretability. In this study, we construct the Ice temporal fusion transformer (IceTFT) model, which mainly consists of the variable selection network (VSN), the long short-term memory (LSTM) encoder, and a multi-headed attention mechanism. We select 11 predictors for the IceTFT model, including SIE, atmospheric variables, and oceanic variables, according to the physical mechanisms affecting sea ice development. The IceTFT model can provide 12-month SIE directly, according to the inputs of the last 12 months. We evaluate the IceTFT model from the hindcasting experiments for 2019–2021 and prediction for 2022. For the hindcasting of 2019–2021, the average monthly prediction errors are less than 0.21 ×10 6 km 2 , and the September prediction errors are less than 0.1 ×10 6 km 2 , which is superior to the models from Sea Ice Outlook (SIO). For the prediction of September 2022, we submitted the prediction to the SIO in June 2022, and IceTFT still has higher prediction skill. Furthermore, the VSN in IceTFT can automatically adjust the weights of predictors and filter spuriously correlated variables. Based on this, we analyze the sensitivity of the selected predictors for the prediction of SIE. This confirms that the IceTFT model has a physical interpretability.