Assessments of Data-Driven Deep Learning Models on One-Month Predictions of Pan-Arctic Sea Ice Thickness

Abstract

In recent years, deep learning methods have gradually been applied to prediction tasks related to Arctic sea ice concentration, but relatively little research has been conducted for larger spatial and temporal scales, mainly due to the limited time coverage of observations and reanalysis data. Meanwhile, deep learning predictions of sea ice thickness (SIT) have yet to receive ample attention. In this study, two data-driven deep learning (DL) models are built based on the ConvLSTM and fully convolutional U-net (FC-Unet) algorithms and trained using CMIP6 historical simulations for transfer learning and fine-tuned using reanalysis/observations. These models enable monthly predictions of Arctic SIT without considering the complex physical processes involved. Through comprehensive assessments of prediction skills by season and region, the results suggest that using a broader set of CMIP6 data for transfer learning, as well as incorporating multiple climate variables as predictors, contribute to better prediction results, although both DL models can effectively predict the spatiotemporal features of SIT anomalies. Regarding the predicted SIT anomalies of the FC-Unet model, the spatial correlations with reanalysis reach an average level of 89% over all months, while the temporal anomaly correlation coefficients are close to unity in most cases. The models also demonstrate robust performances in predicting SIT and SIE during extreme events. The effectiveness and reliability of the proposed deep transfer learning models in predicting Arctic SIT can facilitate more accurate pan-Arctic predictions, aiding climate change research and real-time business applications.