Four-Channel Deep Learning for Sea Ice Classification: Integrating SAR Backscatter, Gradient Ratio, and Incidence Angle over the Antarctic Peninsula

Distinguishing sea ice stages of development from spaceborne Synthetic Aperture Radar (SAR) remains challenging due to overlapping backscatter signatures, and prior convolutional neural network (CNN)-based approaches have typically relied on only two or three input channels (e.g., HH/HV backscatter alone, or backscatter with incidence angle treated as a correction factor rather than a direct input). This study introduces a novel four-channel CNN framework that, to our knowledge, is the first to fuse HH backscatter, HV backscatter, a derived Gradient Ratio (GR) using the HH and HV, and local incidence angle as four equally weighted input channels for classifying sea ice types—open water, young ice, first-year ice, and old/multi-year ice (MYI)—from Sentinel-1 Extra Wide SAR imagery over the Antarctic Peninsula, a region selected for its representation of all major ice classes including old ice. The CNN comprises three convolutional blocks (64-64-96 filters) with batch normalization and max-pooling, followed by a 512-unit dense layer with dropout, trained using a weighted focal loss to address class imbalance. Training used 338,539 patches (32×32 pixels) extracted from a December 2020 SAR mosaic and National Ice Center ice charts, with hyperparameters tuned via a 20-trial Optuna search and Adam optimization. On a held-out test set of 67,708 patches, the model achieved 91% overall accuracy and a Cohen's Kappa of 0.87, with F1-scores of 0.92-0.94 for young ice, first-year ice, and old ice—outperforming or matching benchmark CNN approaches that use fewer input channels, while providing added interpretability through the GR and incidence angle terms. Open water showed high recall (0.93) but lower precision (0.66) due to spectral similarity with smooth young ice. This work demonstrates that fusing SAR backscatter with a derived GR and local incidence angle as equal CNN input channels—a combination not previously reported in the sea ice classification literature—yields good classification accuracy while offering a generalizable template for incorporating radiometric correction terms directly into deep learning pipelines, rather than as pre-processing corrections. The four-channel CNN incorporating local incidence angle achieves good performance for automated Antarctic sea ice classification, though open-water/young-ice confusion and ice-chart resolution mismatches remain key limitations warranting multi-frequency or temporal data in future work. Notably, GPU memory constraints limited the training set to 338,539 patches rather than the targeted 1 million, suggesting further accuracy gains are achievable with greater computational resources.