argmax/argmin?

[soxofaan]

In machine learning (classification) use cases, one typically need an operation like argmax or argmin. It's currently not very obvious/discoverable how to do that in pure openEO processes.

There are a couple of options however, e.g. for argmax:

• based on order : first(order(array, asc=false))
• based on array_find: array_find(array, value=max(array))

Possible options to improve the situation:

• simple: just improve docs, e.g. from max and min docs: link to order/array_find how these can be used to emulate argmax/armgin
• more work: add argmax/argmin processes

[m-mohr]

We could also add a section for common ML practices in the cookbook: https://openeo.org/documentation/1.0/cookbook/

[jdries]

clients can also add convenience processes for argmin/argmax, creating the corresponding graph based on standardized processes.
or we can publish them as UDP's?

[PondiB]

• more work: add argmax/argmin processes

I can add them later to the ML processes I am working on.

[JorisCod]

I would like add that having an argmax process available is advantageous for useability of the openEO API.
ML algorithms commonly output probabilities, and typically converted to classes using argmax. It's also a common mathematical operator, available in many packages. Users will look for this in the list of processes.

Some practical use cases:

1. openEO PG using a machine learning algortihm giving probabilities, converting this into classes
2. openEO PG for on-the-fly visualisation: e.g. Copernicus browser is increasingly relying on openEO PG's, support for this function would enable on-the-fly visualizations. Multiple probability bands can be displayed as single images, either as a) more complex categorical maps with hierarchical legends or b) RGB's which are for visualization only and not meant for data processing.