[ ] Add fpzip to the image. Make sure to also have it in the worker image.
[ ] Adapt Readme to make sure to only provide a binder link for the cloud provider which has the LLC4320 data (Google, I think).
[ ] Demo cases
[ ] A completely water-filled box in the open ocean (x, y, t). (This should look fine and give a good impression about what's possible with lossy compression.)
Calculate EKE (eddy kinetic energy = 0.5 * (var[u] + var[v])) for the uncompressed data, show map.
Compress / decompress lossy (different rates) and calculate / compare EKE from lossy decompressed data.
[ ] Repeat EKE from uncompressed / lossy data for a box with islands / coastline. (This will show problems that can be solved with learning / training how to handle masked areas and might show different mitigation ideas.)
Show problems for plot across full box. Errors will be present near coasts.
Select a section / line that crosses the coast at some point and show how the error is reduced with increasing distance from coast.
Mitigation 1: Repeat same section but select sequence of data sent to fpzip that does not include masked data. (For the demo, a hand-picked path works fine. For production across huge datasets, we'll need ML to learn how to pick paths automatically.)
Mitigation 2: Adapt algorithm to change the stencil size near masked data. (Not sure we can easily do this based on fpzip?)
Mitigation 3: Fill in masked data in a way that allows for only sending smooth data to fpzip. Extrapolating into land-covered regions worked fine when I played with this in December. _(Not an ML appraoch, however. ?)
fpzipto the image. Make sure to also have it in the worker image.eddy kinetic energy = 0.5 * (var[u] + var[v])) for the uncompressed data, show map.fpzipthat does not include masked data. (For the demo, a hand-picked path works fine. For production across huge datasets, we'll need ML to learn how to pick paths automatically.)fpzip?)