AGL-UNet: Adaptive Global-Local Modulated U-Net for Multitask Sea Ice Mapping.

The increasing demand for Arctic route planning, climate change studies, and the growing volume of satellite sensor data have made automated sea ice mapping an essential task. In this study, we propose a multi-task sea ice mapping framework based on the U-Net architecture, which supports multi-sensor data integration and automatically modulates global and local features. The model consists of ARC blocks for enhanced multi-sensor feature fusion, a GLCM block for non-local and local feature modulation, and an adaptive loss weighting strategy to balance multi-task training. The proposed method is evaluated on the AI4Arctic RTT dataset, which includes multi-sensor inputs and ice chart-derived labels. Compared with the best-performing method in the AutoIce Challenge, the proposed approach achieves a 1.33% improvement in the combined score. In addition, the F1 scores for stage of development (SOD) and floe size (FLOE) increase by 2.85% and 3.44%, respectively. Although the R2 score for SIC shows a slight decrease of 1.25%, this behavior is consistent with the practical trade-offs commonly observed in multi-task optimization. Ablation studies further demonstrate the effectiveness of the proposed blocks and the multi-task adaptive weighting strategy, confirming their potential for handling multi-sensor data and supporting ocean environment monitoring.