brt tutorial

[eeholmes]

• get argo data argopy.ipynb
  • get base code working
  • get tutorial to describe the argo data and argopy package done
  • get the chla and bbp700 data
  • merge the data into one parquet file
  • put the data into GitHub
  • DONE with this step
• do the rrs matchup
  • get base code working
  • merge into one parquet file with the argo data
  • put the data into Github
  • DONE with this step
• ? do sst matchup
• fit brt
• predict
• finish the matchup_advanced tutorial section that relates to this

wrinkles

• pca for rrs instead of bands directly

[eeholmes]

Keep the explanatory variables minimal but physically meaningful:

Primary teaching model:

Rrs hyperspectral (log-transformed; maybe PCA reduced)

SST

lat, lon

(optional: sin/cos DOY if paired with lat/lon)

Do NOT include Rrs-derived products as predictors.
Add optional predictors only if you want a “realistic science” extension:

PAR

wind speed

surface pressure

atmospheric correction metadata

These show how ML can blend optics + physical context, but they aren’t needed in the “core teaching” version.