Prediction of the Arctic sea ice concentration using deep-learning approaches

Abstract

Changes in the Arctic sea ice affect atmospheric circulation, ocean currents, and polar ecosystems. There have been unprecedented decreases in the amount of Arctic sea ice, due to global warming and its various adjoint cases. A more accurate short/long-term prediction of the Arctic sea ice system, such as sea ice extent (SIE) and sea ice concentration (SIC), is crucial in the management of the Earth's climate system. However, sea ice prediction is a challenging task under the changing Arctic climate system together with uncertainties in observation systems. Especially, a more accurate prediction of the marginal ice zone (MIZ), which is a more vulnerable region to change (i.e., the boundaries between the sea ice and open seas) is important to develop a robust sea ice prediction model. In this research, a novel Arctic sea ice prediction model is proposed, with multi-satellite data using a deep-learning approach. This thesis consists of three parts: (i) monthly prediction of Arctic SIC using convolutional neural networks (CNN), (ii) seasonal prediction of Arctic SIC using UNET approach, and (iii) weekly prediction of Arctic SIC using cross-attention with positional encoding approach. The first part examines the monthly SIC prediction using SIC in the past, sea surface temperature, 2-meter temperature, albedo, and v-wind based on the CNN model. The CNN model outperformed a random forest model and a persistence model in prediction accuracy. The model demonstrated its reliability even in extreme events, such as drastic sea ice declines in 2007 and 2012. The second part suggests the seasonal (up to 12 months) SIC prediction models based on the UNET approach. The UNET outperformed the three baseline models, the Copernicus Climate Change Service, persistence forecast, and CNN models, exhibiting a low root mean square error (RMSE) of 6.06% on average. The UNET with a 12-month lead time had RMSE<8% for the 2007, 2012, 2019, and 2020 extreme summers. In the last part, this thesis suggests a weekly, seven-to-seven days prediction using cross-attention with positional encoding. The model showed RMSE of 6.40% and MAE of 2.32% in seven-day prediction. As a result, the proposed short to long-term sea ice prediction models can contribute to expanding the current knowledge about the Arctic sea ice forecasting system and Arctic shipping route planning, and provide a unique dataset regarding spatio-temporal changes in the sea ice system.