| Summary: | Despite the growing use of deep learning in sea ice mapping with SAR imagery, the study of model uncertainty and segmentation results remains limited. Deep learning models often produce overconfident predictions, a concern in sea ice mapping where misclassification can impact marine navigation safety. We incorporate and compare dropout and model ensemble within a convolutional neural network segmentation architecture to highlight regions with prediction uncertainty, and explore the impact of loss function choice. We evaluate model generalization and uncertainty characterization by training and evaluating models on the AI4Arctic Sea Ice Challenge Dataset (primary). We further explore model uncertainty by testing the trained models on the Extreme Earth version 2 Dataset (secondary). The primary and secondary datasets vary in number of scenes as well as in the available data and preprocessing. 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.
|
|---|