Extracting sea-ice in the arctic by integrating microwave scattering mechanisms with superpixel-based synergistic constraints

High-resolution, fine-scale sea-ice extraction is essential for assessing navigation risk in Arctic shipping routes. Synthetic aperture radar (SAR) provides all-weather observations and is therefore a key data source for polar sea-ice classification. However, conventional pixel-level SAR classification methods are vulnerable to speckle noise and often produce fragmented misclassifications, and fine-scale ice-water boundary delineation remains challenging when reliable labelled samples are limited and SAR backscatter responses overlap under complex ice conditions. Here, we propose Microwave-scattering Feature and Superpixel Synergy for Ice Classification (MFSS-IC), a fine-scale sea-ice classification model that integrates microwave scattering mechanisms with superpixel-based synergistic constraints. The model first uses simple non-iterative clustering to reconstruct SAR images into radiometrically homogeneous superpixel objects, and then integrates dual-polarization backscatter intensity, global statistical priors derived from Otsu thresholding, and morphological structural constraints, enabling adaptive suppression of speckle noise and fine delineation of ice-water boundaries at the object scale. Reconstruction of sea-ice distributions from the 2015–2022 Sentinel-1 images over the Arctic Northwest Passage, together with cross-validation, showed that: (1) under complex ice conditions, MFSS-IC achieved an overall accuracy (OA) of 98.90%; its superpixel-based synergistic constraints suppressed local high-frequency noise, maintained the spatial connectivity of sea-ice classification results, and effectively reduced the salt-and-pepper artifacts common in conventional methods; (2) under the same regional evaluation, MFSS-IC achieved the highest Kappa coefficient (0.978) and user’s accuracy (UA; 99.20%) among U-Net, DeepLabv3+, and MFSS-IC; its OA exceeded those of U-Net and DeepLabv3+ by 0.50 and 0.23 percentage points, respectively; and (3) MFSS-IC showed strong spatiotemporal generalization, captured nonlinear interannual oscillations in sea-ice distribution driven by climate forcing, and accurately extracted submesoscale fragmented sea-ice in bottleneck regions of the Northwest Passage. The proposed model improves the spatial representation of submesoscale fragmented sea-ice and provides a more reliable basis for quantitative assessment of navigation risk in polar shipping routes.