This paper investigates the safety risks associated with sea ice along the Arctic Route by using polar sea ice images obtained by Gaofen-3 (GF3) Synthetic Aperture Radar (SAR) satellites. We collected three SAR datasets, representative of GF3 satellites' operational modes, and constructed semantic segmentation datasets through meticulous annotation using concurrent optical satellite imagery obtained by Landsat satellites to ensure precision. We propose SI-CTFNet, an innovative sea ice semantic segmentation model that integrates Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for enhanced multi-scale feature extraction. SI-CTFNet features a dual-pathway architecture designed to optimize feature extraction, complemented by the BiAFusion module, which effectively merges local and global features to improve decoding accuracy. Additionally, we introduce the MSDAM module to facilitate a comprehensive multi-scale contextual analysis, addressing the diverse distribution of ice types in the imagery. The model incorporates advanced decoding techniques, including a progressive upsampling (PUP) approach for the CNN-fusion branch and an efficient All-MLP module for the Transformer branch. Performance evaluations across three distinct datasets reveal that SI-CTFNet significantly outperforms existing methods in key metrics and maintains efficacy with supplementary Sentinel1A C-band satellite data. Furthermore, we present a streamlined variant of SI-CTFNet, which achieves a threefold increase in inference speed with minimal reduction in classification accuracy. The ultimate objective of this work is to advance a precise sea ice forecasting and navigation system for polar regions, aimed at automating sea ice classification within a smart polar shipping framework.

As a critical indicator of polar ecosystem dynamics, sea ice monitoring plays a pivotal role in climate change. However, as global warming accelerates the melting of sea ice, the complexity in the Arctic poses growing challenges for achieving high-precision sea ice classification. To address this issue, this study begins with the creation of a multi-source sea ice dataset based on GaoFen-3 fully polarimetric SAR data and Landsat optical imagery. In addition, the study proposes a Global–Local enhanced Deformable Convolution Network (GLDCN), which effectively captures long-range semantic dependencies and fine-grained local features of sea ice. To further enhance feature integration, an Adaptive Channel Attention Module (ACAM) is designed to achieve adaptive weighted fusion of heterogeneous SAR and optical features, substantially improving the model’s discriminative ability in complex conditions. Experimental results show that the proposed method outperforms several mainstream models on multiple evaluation metrics. The multi-source data fusion strategy significantly reduces misclassification among confusable categories, validating the importance of multimodal fusion in sea ice classification.

Accurate large-scale sea ice classification is vital for Arctic maritime activities. However, this task faces a fundamental challenge. Operational surveillance is restricted to wide-swath dual-polarization data, which limits classification precision due to polarimetric information deficiency. Conversely, while the quad-polarization mode offers the comprehensive scattering details required for more accurate classification, its narrow swath width prevents efficient large-scale monitoring. To address this challenge, we propose an enhanced sea ice classification method relying on limited co-region quad-polarization information to enhance dual-polarization data classification accuracy across larger spatiotemporal scales. Specifically, we construct a polarization-guided cross-mode knowledge distillation framework featuring an asymmetric teacher–student architecture. In this design, a hybrid CNN-Transformer teacher extracts robust scattering features from quad-polarization data to guide a lightweight student network operating on dual-polarization inputs. Through this transfer, the student acquires rich feature representations comparable to quad-polarization data, effectively compensating for the missing polarimetric scattering information. Experimental results on GF3-02 data demonstrate that the proposed method significantly outperforms the standalone dual-polarization network baseline, achieving an overall accuracy of 86.15%. This validates the effectiveness of the proposed method in enabling high-precision sea ice classification for large-scale monitoring.