v0 Embeddings

[brunosan]

Aiming for a set of geographically anchored files that fully cover Earth land and coast, where each one contains a vector with a mathematical representation of the semantics on that location, subject to be fine tuned with higher resolution, or finer set of semantics.

For v0 we can restrict the embeddings to a certain RoI, a certain time window, a certain vector length.

[brunosan]

Updated to re-scope.

[yellowcap]

We will focus on three type of embedding initially

• Instrument (bands). Model should not expect all bands.
• Time with time of day, year, month.
• Location with lat/lon plus extent or resolution. A promising option is Spherical harmonics.

Refs #22 and #12

[brunosan]

Per our call, DevSeed team will create and "embeddings factory" script to run ideally as CI on release versions.

The locations and times to generate will be defined {by Ode's team}, but for v0 ideally not too large.

It also seems that creating embeddings for (some of) the training dataset is a small incremental effort since that data is already prepared.

[brunosan]

v0 documentation #118 includes how to create embeddings for any location and time.