SegIceNet: Activation Information Guided PointFlow for Sea Ice Segmentation

Optical remote sensing is the major means of monitoring sea ice, which is beneficial for waterway planning, disaster prevention, and environmental research. Currently, a large number of sea ice segmentation methods have been developed. However, they often ignore the boundaries between sea ice and seawater, leading to unavoidable mis-segmentation. To alleviate this issue, this letter proposes a novel semantic segmentation network applied to sea ice remote sensing images, referred to as SegIceNet. First, we design a partial class activation map (PCAM) to select the feature points whose classes are easily confused in feature layers. By constructing the affinity among these points, the network explores potential semantic information and propagates these feature details among adjacent decoder layers. Second, this letter introduces a point-set optimization strategy, which aims to focus on samples of the aforementioned special points during the training process. Experimental results on a real sea ice dataset show that our method is better than other state-of-the-art semantic segmentation methods. Specifically, even in the imbalanced sea ice dataset, the proposed method achieved 94.88% mean intersection over union (mIoU).