Support for metadata

[miquelmassot]

Datasets often come with more information than just images. Sometimes there's timestamp information, location, depth, temperature, salinity, etc. Models could make use of that for their training/classification. We would need support for that, in a way that would be scalable and easy to use.

[ots22]

The intake-xarray driver can handle this (e.g. this notebook, catalogue) - metadata are additional data variables in the Dataset. In practice, this seemed to divide opinion a little (some found it a bit confusing, especially if not used to xarray). The other thing we discussed was a numpy + pandas format, which wouldn't lose this metadata.

[ots22]

(related issue: #98)