Pan-Arctic winter sea ice classification using Sentinel-1 dual-polarized SAR images

Sea ice type information is important for climate change research and human activities in polar regions. Synthetic Aperture Radar (SAR) data, particularly from the Sentinel-1 (S1) mission, has become a key source for high-resolution sea ice mapping. With the growing volume of S1 SAR data, deep learning (DL) methods have been explored for sea ice classification. However, the limited availability of fine labels and overlapping backscatter intensity among ice types remain major obstacles to achieving automated fine-scale sea ice classification across the pan-Arctic region. This paper proposes a model, named IceDeepLab, for the automatic mapping of sea ice type across the pan-Arctic region in winter using S1 dual-polarized SAR images, based on the Deep Learning framework DeepLabv3+. The IceDeepLab can identify open water, young ice, first-year ice and old ice, generating pixel-wise classification maps at 400 m resolution. A sea ice type dataset, comprising 174 finely labelled S1 images from November 2019 to March 2020 and covering the pan-Arctic region with diverse sea ice conditions, is established for model training and validation. In particular, radar incidence angle is included as an input to mitigate its effects on HH-polarized SAR data. The model integrates the spatially optimised ConvNeXt backbone, the newly proposed Enhanced Atrous Spatial Pyramid Pooling module, and tailored training and inference schemes. These designs make it more suitable for processing whole SAR images, effectively reducing ambiguity between different sea ice types and eliminating stitching lines, thereby improving overall performance. Experiments on the test dataset show that IceDeepLab achieves an overall accuracy of 91.9%, a mean intersection over union of 82.3%, and a mean F1-score of 90.2%, significantly outperforming the traditional DeepLabv3+ by 16.0%, 10.0%, and 11.1%, respectively. When applied to 4153 S1 images collected from November 2020 to March 2021, IceDeepLab maintains over 90% agreement with operational ice charts in the pan-Arctic region, demonstrating the model's robustness across different years. Ablation experiments, feature visualizations, and statistical analyses further validate the model's effectiveness and offer insights for future improvements in sea ice classification algorithms. Furthermore, evaluations using both the radiometer-derived sea ice concentration product and a year-round ice type dataset are conducted to explore its potential for broader applications in operational sea ice monitoring. • A deep learning based model was proposed to classify Arctic sea ice types by Sentinel-1 SAR dual-polarization data. • Inputting SAR incidence angle data significantly improves the classification Accuracy. • The model additionally presents good performance on detection of sea ice leads.