Ice Floes

The Sentinel-2 Level-2 surface reflectance products provided by the European Space Agency (ESA) serve as the primary data source for extracting ice floe distribution and motion information. Only clear-sky imagery with minimal cloud cover is selected to ensure data quality. True-colour composites are generated using Band 2 (blue, 458–523 nm), Band 3 (green, 543–578 nm), and Band 4 (red, 650–680 nm), each with a spatial resolution of 10 m. To enable accurate area calculations, the projected coordinate system is converted from WGS 1984 UTM Zone 47N to the North Pole Lambert azimuthal equal-area projection. Ice–water classification is performed using a dynamic threshold determined by the local minimum between the two peaks in the grayscale histogram of each image, resulting in a binary mask where white pixels represent ice and black pixels represent water. Small or spurious ice floes with blurred boundaries are removed through a morphological “erosion–marking–dilation” procedure, yielding independent ice floes for further analysis. For deep learning model training, 226 representative ice floe images acquired between 2017 and 2023 are selected. Paired images are generated via affine and perspective transformations, and pseudo-labels are created using homographic adaptation, following a self-supervised strategy. The SuperPoint model is trained with the Adam optimizer (initial learning rate = 1×10⁻⁴, batch size = 8) for 20,000 iterations, with checkpoints saved every 2,000 iterations. The SuperGlue model is subsequently trained end‑to‑end using the transferred SuperPoint weights, also with Adam (initial learning rate = 1×10⁻⁴, decay factor = 0.1 every 10 epochs) for up to 500 epochs (batch size = 4), employing early stopping if the validation loss does not improve for 50 consecutive epochs. All experiments are conducted on a Windows 10 machine equipped with an NVIDIA RTX 4060 GPU (8 GB memory) using CUDA 11.8 and Python 3.11.5. Keypoint extraction and feature matching are performed with a keypoint threshold of 0.001 and a match threshold of 0.1. To facilitate reproducibility, the complete code and trained models will be made publicly available at the URL provided in the paper.