Phase association (Tracking issue)

[yetinam]

SeisBench should eventually include phase association models. Together with the pickers, this would allow to create (preliminary) catalogs directly within SeisBench. This issue tracks discussions, options and progress in this direction.

[jawooll]

Just posting the current idea on this, from internal discussions. The aim would be to integrate high-level interaction through the obspy API in a similar way to how it was done for model.annotate. This will be through some base method associate. The obspy event object hierarchy (e.g. obspy WaveformID, Pick, Event, Catalog, etc.) are used for outputting the stored Catalog. First point to investigate is whether this integration will introduce a significant overhead and we should instead use our own streamlined version of these objects (1).

The second point is, we need to think about what types of models will be performing association (2). As we can have association models which directly use the waveforms themselves to associate events,WaveformModel will need an associate method. But, more generally, you can have association models which operate on just metadata information (e.g. just picks). Do we provide a new general base class for such models? Complete with its own associate base method? Or, we can also stick an abstractmethod associate in WaveformPipeline keeping it general enough to take any input parameter and for each association model we override this in.

It just will take a little more thinking about compared to the picking and detection models, as this was a simpler case where we knew the format of the inputs and the outputs, as they always operate on the waveforms.