Deep Learning Benchmarks for short-term Arctic Sea Ice Forecasting

Abstract. The rapid decline of Arctic sea ice has increased spatiotemporal variability in the core marginal seas along the Northern Sea Route (NSR), thereby hindering reliable short-term forecasting. While deep learning models offer computationally efficient alternatives to physics-based numerical models, previous studies have often relied on global average errors and have provided limited assessment of boundary-preservation performance. In addition, few studies have systematically compared spatial feature extraction strategies within non-recurrent spatiotemporal architectures for sea ice forecasting. To address this gap, this study benchmarks CNN backbone and Transformer backbone families for 10-day sea ice forecasting. The evaluation focuses on five dynamic marginal seas along the NSR. Metrics included Integrated Ice Edge Error (IIEE), Mean Boundary Error (MBE), Intersection over Union (IoU), and Anomaly Correlation Coefficient (ACC). Results indicate that CNN-based models, including PoolFormer, generally show more favorable performance in boundary preservation and prediction stability than Transformer-based models. Regionally, the Barents and Kara Seas were more difficult to predict, whereas the Laptev and East Siberian Seas were relatively more predictable. The Chukchi Sea exhibited particularly high uncertainty during rapid summer ice retreat. Across all models, boundary-based performance degraded significantly as lead time increased, with MBE exceeding 30 km from T+5 onward. These results suggest that, under the current univariate setting, boundary-preservation performance remains relatively stable at short lead times, but degrades thereafter. Overall, CNN-based non-recurrent architectures show relative strengths over the NSR test regions, although future work incorporating multivariate inputs is needed to extend reliable lead times.