Reproduce and improve on the U-Net as seen in the AI4SeaIce paper

[MagnusOstertag]

AI4SeaIce: Toward Solving Ambiguous SAR Textures in Convolutional Neural Networks for Automatic Sea Ice Concentration Charting the code for the model

Do we see similar issues as reported? What ideas do we have to mitigate them?

• noise phenomenon: visible as long vertical lines and grained particles resembling small sea ice floes

• landfast ice predictions are still troublesome for the model

• [x] visualize well the errors of our model

[MagnusOstertag]

They have provided the code of their model with small adaptations in the challenge repo. We only have to use the deeper and smaller layers as they described in their paper:

Used 'filters' parameters:
   'filters': [16, 32, 32]  # 3-level
   'filters': [16, 32, 32, 32, 32]  # 5-level
   'filters': [16, 32, 32, 32, 32, 32]   # 6-level
   'filters': [16, 32, 32, 32, 32, 32, 32]  # 7-level
   'filters': [16, 32, 32, 32, 32, 32, 32, 32]   # 8-level

The 8-level model performed best. For those and the other parameters used see here.

• [x] try all those models

In the outlook of the paper they wrote:

• [ ] Naturally, further expansion of the number of levels of the U-Net could also be of interest but it requires changing the choice of testing scenes or refraining from downsampling the SAR pixel spacing from 40 to 80 m. However, this would reduce the effective spatial receptive field of the models. It is very unclear to me why they stopped at eight levels when the performance kept improving and why they did improve the receptive field only with depth and not more with larger filters.
• [ ] Solving the landfast sea ice prediction problem could potentially be addressed using larger training patch sizes and additional scenes from, for example, Scoresbysund where this phenomenon is well-known. Adding a class for landfast ice could also encourage the model to address it specifically. Providing additional auxiliary data such as distance to land and season/month of image acquisition could enable the model to better understand regional and seasonal variations. -> make as a feature what is to complex for the model (problem: we officially do not know on what scenes we get scored on)

Other ideas from the paper:

• [ ] sample the more complicated parts of the scenes closer to the land/ the sea ice-open water border/ in specific regions (Scoresbysund, fjords, ...) / summer more - importance weighting, instead of weighting it by only the number of pixel available
• [x] what is the class distribution for train and for test? -> see the distributions.ipynb
• [x] -> reweight the distribution of the training set to match the testset!
• [ ] there might be sth like a human attention layer for the final prediction to be learned: the granularity of the prediction varies
• [x] SIC can also be formulated as a regression problem, and they are still evaluated as regression problem with R2
• [ ] as testing is on full scenes, there might be more error in the parts near the borders
• [ ] think about the homogeneity and level of detail trade-off, maybe even more details in the marginal ice zone, the model does not seem to learn in Fig. 9 that there is no open sea between the old ice! Does this also mean we should start from furthest/closest to land such to inform later predictions by the values of our previous predictions? (could also help the fjord problem)
• [x] use the old layers as initialization, transfer learning