Author(s): Runxin Xu; Wenbo Dong; Li Zhou; Linlin He; Hongrun Wu; Xiaowei Zhang; Qiyou Wu; Qixin Lan
Linked Author(s):
Keywords: Sea ice segmentation; Scribble-guided Partitioned Gaussian Mixture Model (SPGMM); Arctic shipborne imagery; U-Net; Polar ship
Abstract: Pixel level annotation of Arctic shipborne sea ice imagery is difficult and time consuming because of fragmented ice, melt ponds, fog, and weak ice water boundaries. To address this problem, a Scribble guided Partitioned Gaussian Mixture Model (SPGMM) is investigated for weakly supervised sea ice segmentation. In the proposed method, sparse scribble annotations are used as class guidance, and the image is transformed into the LAB color space and partitioned along the vertical direction for local Gaussian mixture modeling. In this way, dense sea ice masks can be generated while adapting to the spatially varying appearance characteristics of shipborne scenes. The results show that SPGMM can preserve major ice water boundaries and maintain structural consistency in representative Arctic conditions. Quantitative evaluation on 50 randomly selected images, with fully manual annotation used as Ground Truth, gives a Dice of 0.923, a precision of 0.954, and a recall of 0.896. In addition, the average annotation time per image is reduced from 3730.45 s for manual annotation to 289.80 s for SPGMM, corresponding to a time reduction of 92.23% and an efficiency gain of 12.87 times. These results demonstrate that SPGMM provides an efficient and reliable solution for Arctic shipborne sea ice annotation and segmentation.
Year: 2026