Unicorn: U-Net for sea ice forecasting with convolutional neural ordinary differential equations

Area of Science:

• Climate Science
• Artificial Intelligence
• Oceanography

Background:

• Accurate sea ice forecasting is crucial for understanding global climate dynamics.
• The complexity of interacting variables makes precise sea ice prediction challenging.
• Neural networks are increasingly utilized for sea ice forecasting due to their ability to handle multiple inputs.

Purpose of the Study:

• To introduce a novel deep learning architecture, Unicorn, for weekly sea ice forecasting.
• To enhance sea ice forecasting performance by integrating multiple time series images.
• To improve the capture of spatiotemporal dynamics using a bottleneck layer with neural ordinary differential equations.

Main Methods:

• Developed a novel deep architecture named Unicorn.
• Integrated multiple time series images into the model.
• Incorporated a bottleneck layer within the U-Net architecture, functioning as neural ordinary differential equations with convolutional operations.

Main Results:

• Unicorn demonstrated significant improvements over state-of-the-art models for sea ice concentration forecasting, achieving an average 12% Mean Absolute Error (MAE) improvement.
• The model outperformed existing methods in sea ice extent forecasting, with an approximate 18% improvement in classification performance.
• Real data analysis from 1998-2021 validated the model's superior performance.

Conclusions:

• The proposed Unicorn model offers a superior approach to sea ice forecasting.
• Integrating multiple time series images and employing neural ODEs enhances spatiotemporal dynamic capture.
• The model shows significant potential for improving climate modeling and prediction accuracy.