BAFormer: boundary-aware transformer for sea ice leads extraction from Sentinel-1 SAR images

Sea ice leads are crucial for investigating the mechanisms of polar climate change and ensuring the safety of maritime navigation. Although deep learning techniques have been widely employed for ice leads recognition, they still demonstrate significant limitations. These limitations are particularly due to substantial classification errors in the boundary transition regions between ice and leads, as well as difficulties in effectively identifying narrow leads. To address these challenges, we propose a neural network model named BAFormer, which employs a two-stage decoder composed of a coarse-level decoder and a coarse-to-fine decoder to refine the classification of boundary regions. Furthermore, a boundary evaluation metric is incorporated to enhance the loss function during the training phase. We applied BAFormer to Sentinel-1 images to obtain high-resolution ice lead maps and conducted comparative experiments with ten classical segmentation models. Results demonstrate that the proposed BAFormer achieves state-of-the-art performance, delivering optimal boundary adherence with leads extraction accuracy reaching 99.21% and at least 18.20% improvement in Intersection over Union (IoU) metrics. We further integrate the generated results with Sentinel-2 images and MODIS ice surface temperature (IST) products for combined analysis, and verify that BAFormer maintains stable performance on multi-scale ice lead datasets.