DualSeg26: Instance segmentation of ice and snow thickness sections from shipborne imagery with task-decoupled feature pyramids

Abstract

Shipborne near-field optical imagery directly captures visible ice and snow thickness sections during polar navigation, but these targets are often thin, elongated, low-contrast, and locally ambiguous. Adjacent ice pieces may expose touching sections of the same class, making semantic segmentation prone to region merging when instance-level separation is required. We therefore formulate visible thickness-section delineation as an instance segmentation problem and propose DualSeg26, a YOLO26-based task-decoupled framework whose core design combines separate detection and segmentation necks with a conservative near-identity cross-stream coupling module adapted from manifold-constrained hyper-connections. A semantic P2 head-only branch provides supporting high-resolution refinement for mask prediction. Experiments on a 511-image shipborne dataset show that the final model achieves 0.498 Box AP50, 0.251 Box AP50:95, and 0.267 Mask AP50:95, improving Mask AP50:95 by 0.043 over YOLO26x-Seg. A three-run reproducibility check also gives the highest average AP among the main ablation variants. The resulting segmentation masks can support downstream functions, such as estimating physical ice thickness from calibrated image geometry.