Enhancing Seasonal Arctic Sea Ice Prediction for Late Summer by Transfer Learning of PIOMAS and CryoSat-2-Observed Sea Ice Thickness Data

Sea ice thickness (SIT) is essential for predicting Arctic sea ice concentration (SIC) on seasonal timescales. This study develops a transfer learning approach that combines long-term PIOMAS reanalysis (1989-2010) with satellite-based CryoSat-2 SIT (2011-2019) in a deep learning model, SICNetDB, to enhance summer SIC predictions. The model is pre-trained on PIOMAS SIT and then fine-tuned with CryoSat-2-observed SIT to leverage their advantages. Results indicate that fusing PIOMAS and CryoSat-2 SIT improves September SIC prediction skill by 0.21-0.45 in the anomaly correlation coefficient (ACC) compared to models that rely solely on SIC, and by 0.14-0.26 compared to models using only PIOMAS SIT. In years with less ice, such as 2012 and 2019, transferring CryoSat-2 SIT data results in a 5-unit reduction in mean absolute error (MAE) and an improvement in binary accuracy of up to 11%. These enhancements are especially noticeable at 7-8 biweekly lead times, demonstrating that transfer learning effectively fuses PIOMAS and CryoSat-2 SIT datasets to optimize the spring prediction barrier.