Daily Prediction of Sea Ice Drift at Kilometer-Scale Using Deep-Learning Method

Arctic sea ice drift (SID), a crucial dynamic characteristic of sea ice, is also an important parameter for sea ice mass change quantification as it directly determines the regional sea ice distribution and influences the annual polar sea ice export. Common methods of obtaining high spatial resolution SID depend on the time-series Synthetic Aperture Radar (SAR) data, necessitating multiple data scenes. Here, we developed a deep learning model incorporating single scene high spatial resolution SAR data to predict daily SID at 1 km resolution, achieving a Mean Absolute Error of 1.14 cm·s-1 compared to IABP buoy data. The proposed method has significant advantages over traditional machine learning methods in establishing complex relationships between multiple variables and SID. Additionally, we found that wind has the greatest impact on both the speed and direction of SID prediction. Latitude and longitude together serve as the second most important predictors of SID, after wind. Furthermore, the SAR data inputs have reduced the Mean Absolute Error for speed and direction by 0.24 cm·s-1 and 1.62 degrees, enhancing the SID predictions' accuracy, particularly in regions with thick ice.