Semi-Supervised Multi-Source Sea Ice Classification in Small-Data Regime

Sea ice type classification is essential for climate change research and maritime safety. Traditionally, this process relies on manual ice charting, which is time-consuming, expensive, and requires expert knowledge, making it difficult to scale up for current demands. Automating sea ice type classification is essential to keep pace with rapidly changing sea ice conditions. However, two main challenges limit the development of effective automated classifiers. First, while ice charts provide valuable labeled data, they only offer large-area (polygon) annotations rather than pixel-level labels, leading to a lack of precise training data. Second, although there are additional datasets with useful sea ice information, effectively combining these different data sources remains difficult. To tackle the first challenge, we employed co-training and label propagation, two semi-supervised learning methods, to learn from a small amount of labeled data and a large pool of unlabeled data, thereby improving the accuracy of sea ice classifiers despite limited labeled data. To address the second challenge, we leveraged co-training’s built-in ability to integrate multiple data sources during the training process for the small labeled data. Additionally, we further enhanced data integration by using an ensemble of these co-trained models after training. Our approach demonstrates significant improvements over traditional supervised methods, showcasing the potential of semi-supervised learning methods in addressing two major challenges in developing automated sea ice classification solutions. Our study shows that semi-supervised learning improved F1 scores by 17% for SAR data and 33% for AMSR2 with limited labels, compared to supervised methods, while ensembling further boosted accuracy by 33%.