A prototype-based metric-learning model for sea-ice type mapping from Sentinel-1 SAR and ancillary data

Limited labeled data remains a central challenge for deep-learning-based sea-ice type mapping, while cross-channel relationships remain underexploited, and interaction mechanisms are lacking. This study proposes ESICM, combining Sentinel-1 SAR, AMSR2 brightness temperatures, and ERA5 environmental variables for pixel-level mapping of open water (OW), young ice (YI), first-year ice (FYI), and multi-year ice (MYI). ESICM adopts prototype-based metric learning to fully exploit labels, incorporates a channel multiplication-division module to enhance cross-channel interactions, and introduces two interaction functions, supports-input selection and user dense prompts, to improve classification in hard-to-classify scenes. On the AI4Arctic dataset, trained on 492 AI4Arctic scenes and tested on 20 scenes, ESICM achieved 93.34% overall accuracy, 93.08% F1 score (F1), 88.40% Kappa coefficient, and 77.29% mean intersection over union, exceeding the second-best results by 0.85%–1.55%. Under challenging melt-affected conditions, MYI F1 and intersection over union exceeded the second-best results by 3.13% and 3.93%. Across forward-temporal, cross-region, reduced-training-data, and cross-dataset experiments, ESICM ranked first for 14 and second for two of 16 metrics, indicating relatively strong generalization. In two hard-to-classify scenes, simulated user prompts covering 10% of valid pixels improved all four metrics by more than 4% and 20% over prompt-free inference, respectively, supporting the technical feasibility of interaction. SHAP attributions were broadly consistent with established sea-ice remote-sensing knowledge, enhancing the model’s post hoc interpretability. Overall, ESICM improved sea-ice mapping with limited labeled data and showed application potential under controlled distribution shifts and simulated interaction.