One hot encode world cover data

[gabrieltseng]

https://allenai.slack.com/archives/C08BBV886SY/p1744650084967499

Example run: https://wandb.ai/eai-ai2/helios_debug/runs/ugttz4lc?nw=nwusergabrieltseng

This seems to do pretty well (although the biggest gains are on Pastis, which - as we discussed - is noisy).

[gabrieltseng]

This is maybe a small nit but why do we have 0s in the data?

I am guessing this is masked values in the data (i.e. ocean). Would it be sensible to replace it with 80 (permanent water body) which I am assuming is the case for the majority of pixels?

[took a look]

Okay this is indeed the case. This tif file is padded with 0s, which are missing / masked:

ESA_WorldCover_10m_2021_V200_S12E051_Map.tif

Do we want to replace these with ocean? Is it a safe bet that the majority of missing values are ocean pixels?

Edit: it seems this does worse (brown) than treating the missing class as its own class (beige), so that's what we do now.

[pjreddie]

Noting preferences from last weeks meeting: Favyen wants separate modality to preserve backward compatibility, I'm interested in this happening in the data creation phase as opposed to in the loop. I think consensus was to punt on this for now unless your ongoing experiments show significant results since overall it doesn't seem to affect performance substantially. Let me know if I misunderstood though!