pan-Arctic super-resolution prediction and explainability mat

This paper proposes an explainable multi-scale stacked spatiotemporal Transformer model, MSs-STFormer, for improving Arctic daily subseasonal SIC SR forecasting during the melting season. The model integrates four tailored modules, MF-F, HFE, ST-TF, and MS-FF, to address main challenges in climate system modeling. Additionally, we employ two post-hoc XAI methods, Gradient SHAP and LIME, to analyze the contributions of environmental factors. MSs-STFormer comprehensively addresses three critical challenges: capturing long-term SIC forecasting trends, overcoming spatial resolution limitations during the melting season, and evaluating model reliability. Experiments confirm its superior performance over traditional methods, offering both scientific insights and operational utility for polar sea ice forecasting.