Lightweight sea-ice segmentation for resource-constrained polar navigation applications

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate sea-ice segmentation requires both reliable delineation of complex ice–water structures and controlled computational cost, particularly in scenes containing fragmented floes, dense boundaries, and strong scale variation. This study develops IceAware-UNet, a lightweight segmentation framework that organizes model design around two complementary operating points in the accuracy–complexity space. IceAware-UNet-B retains a larger representation budget to emphasize segmentation accuracy under reduced model complexity, whereas IceAware-UNet-L adopts a more restrictive feature budget to achieve stronger compression while maintaining high segmentation quality. The framework is evaluated through extended architectural comparisons, component ablation, image-level paired statistical analysis, morphology-conditioned assessment, and supplementary evaluation on held-out high-resolution scenes. IceAware-UNet-B achieves an average IoU of 0.9456 and Dice coefficient of 0.9720 with 7.62 M parameters and 10.27 G MACs, placing it in the highest-accuracy region of the extended comparison while using substantially less computational capacity than conventional high-accuracy references. IceAware-UNet-L reduces the model to 1.99 M parameters and 2.49 G MACs while retaining an average IoU of 0.9388 and Dice coefficient of 0.9684, defining the compact end of the framework. Paired bootstrap analysis shows that the accuracy separation from lightweight and multi-scale references remains positive across validation images, while differences among the strongest U-Net-based models are narrower. Ablation results identify channel–spatial recalibration as an effective contributor to the balanced configuration and multi-scale contextual aggregation as the dominant component in the compact configuration, with skip-feature attention showing its clearest contribution under highly fragmented morphology. Morphology-conditioned analysis further shows that fragmentation and dense ice–water interfaces are the principal sources of segmentation degradation, whereas scale disparity alone does not produce a monotonic difficulty trend. These results establish IceAware-UNet as a unified lightweight framework in which segmentation accuracy, computational cost, and morphology-dependent behavior are jointly characterized rather than optimized in isolation.