Question on encoding categorical ranges

[murithigeo]

Supposing a categorical parameter {....,"categoryEncoding": { "http://example.com/land_cover/categories/grass": 1, "http://example.com/land_cover/categories/forest": [2,3] }. Does the categoryEncoding value represent a range, i.e. 3>=x>=2 or fixed values?
I'm asking because of the complexity of encoding values sampled from an image

[jonblower]

The categoryEncoding is one or more fixed values (integers). It can be an array to allow the same category (e.g. grass) to be encoded by multiple alternative values in the range object. This is sometimes useful, e.g. if the range object contains many different kinds of "forest" but you want to treat them as a single category.

But in most cases, I would imagine that the categoryEncoding is a single value.

Does that make sense?

[murithigeo]

What is the optimal way to represent bounded values in categories ie ndvi values without having to create arrays crammed full of all observed values

[jonblower]

I don't think there's a way to do that without having two separate Parameters, one for the continuous-valued NDVI values, and one for the classification.

I think you are describing a new use case, which could be interesting to accommodate, particularly if this is important for other kinds of observation

[chris-little]

@murithigeo @jonblower Is this a requirement for a generalised lookup table (LUT)? That is, there is a domain with a range of possibly continuous observed values, which are then quantized into discrete buckets, and these bucket labels are the categories? If I had this use case, I would not try to encompass everything into one object, but have two linked objects - the category definitions which would be relatively stable, or even in a controlled register, and the observed categories.
Or am I missing it completely?

[jonblower]

@chris-little I agree with you, that having linked objects would be a neat solution. It's a lot like having a colour palette definition, which specifies the mapping between real-valued data and colours.

[murithigeo]

@murithigeo @jonblower Is this a requirement for a generalised lookup table (LUT)? That is, there is a domain with a range of possibly continuous observed values, which are then quantized into discrete buckets, and these bucket labels are the categories? If I had this use case, I would not ry to encompass everything into one object, but have two linked objects - the category definitions which would be relatively stable, or even in a controlled register, and the observed categories. Or am I missing it completely?

Essentially yes. Land cover classes or similar classifications are often continuous and a single label can represent a gamut of values falling into a single category.

[chris-little]

@jonblower `It's a lot like having a colour palette definition, which specifies the mapping between real-valued data and colours.

Ii is exactly like that, but generalised.