An Explainable Deep Learning Model for Daily Sea Ice Concentration Forecast

As Arctic sea ice rapidly declines, accurate daily sea ice concentration (SIC) forecasts become crucial for Arctic research and operations. Emerging deep learning (DL) forecast models have the advantage of consuming fewer computational resources compared with numerical forecast systems. However, DL forecast models face challenges of lacking sea ice domain knowledge, low forecast accuracy in the marginal ice zone (MIZ), and poor explainability. This study proposes an explainable DL model, SIFNet, for daily Arctic SIC forecasts, which considers the domain knowledge of teleconnections and lagged correlations by integrating the convolutional block attention module (CBAM) and temporal feature convolution module (TFCM). To improve SIFNet's accuracy in the MIZ, an MIZ-weighted loss function is used for fine-tuning, leading to a notable 1.36% reduction in the mean absolute error (MAE) of MIZ forecasts (MIZ MAE). In the January 2020-December 2023 test scenario, forecasting SIC for the future 49 days using the past 21 days' data, SIFNet demonstrates forecast accuracies with the MAE of 4.69%, binary accuracy (BACC) of 95.16%, structural similarity index (SSIM) of 95.13%, and MIZ MAE of 18.22%. Compared with the other two DL models and nine numerical forecast systems, SIFNet's forecasts achieve higher accuracies. Using layerwise relevance propagation (LRP) technique, the LRP- \vert z\vert rule is constructed to improve SIFNet's explainability. During Arctic minimum sea ice extent (SIE) periods, SIFNet forecasts confirm upward and downward surface solar radiation as predictors for minimum SIE. Similarly, over the past 30 years, precipitation phase is considered a potentially significant factor influencing SIC forecasts.