Joint Arctic Sea Ice Forecasting Based on Graph-Structured Spatial Modeling and Temporal Transformers

Rapid changes in Arctic sea ice exert significant impacts on regional climate feedbacks and high-latitude maritime activities, increasing the demand for accurate short-term forecasting of key sea ice variables. This study proposes a GraphTransformer-based framework for joint forecasting of sea ice thickness (SIT) and sea ice concentration (SIC), designed to address their strong spatiotemporal coupling under irregular Arctic Ocean geometries. A static spatial graph is constructed over effective Arctic marine grid cells, where neighborhood aggregation is applied at each time step to explicitly encode spatial correlations. A shared-parameter temporal Transformer is subsequently employed to model node-level long-range temporal dependencies and to perform direct multi-step forecasting. The model generates 14-day daily forecasts of SIT and SIC in a single forward pass. Experiments are conducted using multi-source daily data spanning from 1 January 2019 to 15 May 2025, with evaluation restricted to valid marine grid nodes. Results indicate that the proposed GraphTransformer achieves either the best or second-best performance among the compared models in multi-step forecasting accuracy. Ablation experiments further confirm the critical role of graph-based spatial encoding in enhancing spatial coherence and mitigating error propagation.