Prior-guided frequency-domain adaptation for pixel-level water-ice classification in remote sensing imagery

Reliable water-ice classification in polar and cold-region remote sensing imagery is essential for environmental monitoring, climate modelling, and hazard assessment. Conventional approaches, whether based on spectral indices or deep learning, often struggle with spectral ambiguity between water and ice, and their performance degrades under domain shifts caused by variations in geography, season, and sensor characteristics. To address these challenges, we propose a parameter-efficient, prior-guided segmentation framework that combines a frozen pretrained backbone with lightweight adapters. By integrating a frequency-domain adaptation module, our method selectively modulates amplitude to incorporate domain-specific information with the preserving phase-consistent structural cues. And a stage-wise prior-guided fusion mechanism is also designed to leverage physical priors to resolve ambiguities. Based on the Sentinel-2 imagery, our scheme is evaluated within four Antarctic regions. The quantitative results show that, our method can achieve an average pixel accuracy (PA) of 97.34% and mean Intersection-over-Union (mIoU) of 86.24%, outperforming the strongest baseline (UNet) by 0.97% in PA and 5.12% in mIoU. Moreover, ablation experiments validate the complementarity of the adapter, frequency-domain adaptation, and prior-guided fusion modules, while portability tests demonstrate the robustness of the proposed framework. Therefore, the proposed method provides an effective solution for high-precision pixel-level water-ice classification in remote sensing.