Open andyjenkinson opened 6 months ago
Thanks for getting the discussion started.
I'm wondering if we should separate this into two notions:
1) A universally agreed-upon definition of a given boundary, which we would move to the core. 2) A dataset-specific definition of a label for the region, which we keep in the extension. This is nice because it grants flexibility to the person building a dataset.
FWIW we've been using https://github.com/Kitware/kwcoco on another project to represent similar data for training ML models. Perhaps there are some lessons to be drawn from that.
Hi @hannah-rae @jacobsn thinking we could start to consolidate some of the fields already, in particular the 'types'.
The 'creation_method' is the same concept as was intended for the 'determination_method' in core, perhaps you could comment on the allowed values in the enum? The list was based on the outcome of AgGateway WG17: Field Boundaries Use Cases and Definitions. I expected there'd also be more method-specific detail (e.g. the ML model and version) so good you have that in the AI extension but you may want to rename it to match.
I think we could also propose to put the 'boundary_type' into core somehow. For reference, this is what AgGateway WG said on this concept:
Notably, it was seen as not mutually exclusive (i.e. a boundary can have multiple types...) Personally I feel that there is something not captured here which is what is the thing on the ground the geometry represents, which is not quite the same as its purpose - for example does it represent a physical boundary feature, or a single crop.