Sea ice classification from Sentinel-1 data

Using Swin Transformer (SwinT), a new sea ice classification algorithm based on dual-polarized Sentinel-1 Synthetic Aperture Radar (SAR) extra wide swath (EW) data is developed. Combining HH and HV polarized images, RGB false-color composites are made to enhance the texture features and to reduce the impact of the incidence angle and thermal noise on the image. The SwinT is trained with ice charts published by National Ice Center (NIC) for four classes: open water, young ice, first-year ice and old ice. The overall validation accuracy of the SwinT-based sea ice classifier is up to 0.92, which is higher than the CNN-based classifier. The highest accuracy occurs both in old ice and first-year ice, with a value of 0.97. It is difficult to successfully distinguish open water from young ice, whose accuracies are 0.84 and 0.79, respectively. It is believed that an efficient noise removal algorithm is beneficial for high accuracy in sea ice classification.