Smarter regridding

[lewisjared]

The problem

Currently we decimate the horizontal gridded data in an extremely naive way.

• For lat/lon grids: Take the first 10 points in the lat/lon dimensions
• For i/j grids: First 1 points of the i dimension and 10 points from the midpoint in the j dimension (to try and get something that is valid ocean data.

This is a small subset of the grid and very flakey. We should adopt a smarter regridding approach. There are a bunch of packages that support regridding on climate data. Perhaps: https://xesmf.readthedocs.io/en/latest/index.html?

This will require the migration to a conda based workflow. I'd prefer using pixi for managing conda dependencies

Definition of "done"

• MIgrate to use pixi
• Pick a regridding package
• Regrid to something with a similar number of horizontal cells (100-1000)

Additional context

[nocollier]

The current method of decimation does not yield test data that is useful for land variables (ILAMB). One simple change to your method could be to take a 10x10 patch relative to a given lat/lon. I would pick somewhere over the Amazon forest. Maybe there is a lat/lon for atmosphere/land and one for ocean?

[nocollier]

Or maybe the play is to pick a location on a flat stretch of coast and take -5:+5 cells around it so you definitely get atmos, ocean, and land variables from one location.

[lewisjared]

@nocollier The grid can be different, but it at least needs to overlap right?

[nocollier]

Right, for the moment I have made my tests tolerant of the fact that everything will be Nan, but moving forward it would be great t get some numbers to compare.

[lewisjared]

This is blocking Climate-REF/climate-ref#107 so I'll take a look tomorrow