Model Ensemble With Dropout for Uncertainty Estimation in Sea Ice Segmentation Using Sentinel-1 SAR

Despite the growing use of deep learning in sea ice mapping with synthetic aperture radar (SAR) imagery, the study of model uncertainty and segmentation results remains limited. Deep learning models often produce overconfident predictions, posing challenges for domain experts to assess prediction quality and identify potential errors. This is especially important in sea ice mapping, where misclassifying ice and water can increase risk to marine navigation. In this article, we incorporate and compare dropout and model ensemble within a convolutional neural network (CNN) segmentation architecture to highlight regions with prediction uncertainty and investigate the effects of loss function choice on the results. We evaluate model generalization and uncertainty characterization by training models on a primary dataset and testing on a secondary benchmark dataset. We obtain test F1 results higher than 0.97 for the primary dataset. Although the F1 performance for the secondary dataset is reduced to 0.93, the generated sea ice maps are reasonable across several Sentinel-1 scenes, and our proposed strategy helps in identification of misclassified and uncertain regions for human quality control. Our models seem to be robust against banding noise in Sentinel-1 SAR, and the prediction uncertainty frequently highlights ice regions misclassified as water, indicating its potential for real-world applications. Our study advances the field of machine-learning-based sea ice mapping and highlights the importance of uncertainty estimation and cross-dataset evaluation for model development and deployment. Our approach can be adopted for other remote sensing applications as well.