Fully-automated pipeline to extract ice floes from WV03 panchromatic imagery with a U-Net. Trained on a small set of hand-labelled Antarctic pack-ice and background images and a much larger set of weakly-labelled pack-ice images obtained with a watershed segmentation algorithm. Best results are obtained using test-time-augmentation. Model predictions are largely robust to context, adding more flexibility in applications when compared with threshold-based methods typically employed in sea ice segmentation.
- Leverages fine-tuning from synthetic data to greatly improve out-of-sample performance.
- > 0.85 F1 score in a non-trivial, hand-annotated test set.
- Over 30% improvement when compared with threshold-based methods.
- Best model weights (incoming) are easily loaded with the PyTorch Segmentation Models package.
- Over 850 random-search experiments ran for hyperparameter tuning with the Bridges2 supercomputer.
- Training script.
- Model evaluation script.
- Prediction script (incoming).
- Saved model weights (incoming).
- Dataset classes.
- Several implementations of Semantic Segmentation loss functions.