A Dual-Branch Framework Combining Amplitude and Phase Information for SAR Sea Ice Classification

Sea ice detection and classification are critical for navigation safety and climate change research. With its high-resolution, all-weather, and day-and-night imaging capabilities, Synthetic Aperture Radar (SAR) has become a vital data source for these tasks. However, most existing deep learning studies rely on amplitude data, failing to fully exploit complex-valued information. This study evaluates whether locally filtered inter-channel phase information from Sentinel-1 Single Look Complex (SLC) data provides complementary information to conventional HH and HV amplitudes. Using the AI4 Arctic Sea Ice Challenge Dataset (ASID Challenge), we construct a complex-valued dataset and evaluate different input combinations with four representative semantic segmentation networks. We examine the classification results across six different categories in the ASID Challenge and, in conjunction with real navigation scenarios, map these six categories into three classes: Open Water, Early-stage Ice, and Developed Ice. For FCN under direct three-class training, phase augmentation improved the mean test Macro-F1 and mIoU by 4.4% and 6.1%. The results show that although its benefit is not consistent across all six-class classification experiments, phase information provides more reliable improvements when sea ice types are grouped according to broader development stages. These findings suggest that inter-channel phase contains complementary information beyond amplitude and may improve SAR sea ice classification under practical application scenarios.