Further modifications to NdArray specification

[jonblower]

I realised that the latest changes to the NdArray specification are not compatible with some of the Playground examples. Given that some users may have based their own CoverageJSON documents on these examples, I thought it was better not to break them.

I think we have to permit "zero dimensional" arrays after all, but perhaps we can clarify that this means that:

• axisNames and shape can be omitted or set to [].
• values must contain exactly one data value. (presumably it can't be empty?)

I think the modification we recently made to the Domain definition is still valid, because I think it's still true that a Domain must have at least one axis.

What do you think, @letmaik?

I also realised that I don't think we define how Domains and NdArrays need to relate to each other. I think the rules are:

• Any domain axes with more than one value must be included as a dimension in each of the NdArrays in the range.
• Any axisName mentioned in the NdArray must correspond with the name of a Domain axis.
• Domain axes with one value MAY be omitted from the corresponding NdArray. This means that, for PointDomains, the corresponding NdArrays can be zero-dimensional.

Do we need to define the above anywhere?

[letmaik]

Ah shoot, you're right, I didn't think about the Point case.

I agree that the Domain changes are still valid. I would just revert the NdArray changes without extra clarifications, to me at least the two bullets about shape and axisNames are clear enough.

Regarding the relation between Domain and NdArray there's this text within Coverage:

A coverage object MUST have a member with the name "ranges" where the value is a range set object. Any member of a range set object has as name any of the names in a "parameters" object in scope and as value either an NdArray or TiledNdArray object or a URL resolving to a CoverageJSON document of such object. [...] The shape and axis names of each NdArray or TiledNdArray object MUST correspond to the domain axes defined by "domain", while single-valued axes MAY be omitted.

This corresponds to your three bullet points, right?

[jonblower]

OK great, I'll revert the NdArray changes. The only thing I wasn't clear about with the original NdArray text was whether a 0D array must have exactly one data value. I might add a sentence to that effect.

This corresponds to your three bullet points, right?

Yes, I think that covers it, thanks

[jonblower]

Addressed in https://github.com/opengeospatial/CoverageJSON/pull/77

[jerstlouis]

@chris-little @jonblower

Ah shoot, you're right, I didn't think about the Point case.

I was thinking about that when that was brought up and was slightly concerned. Glad that 0D is back in :)

[jonblower]

@jerstlouis A Point would still have been possible in the >=1D case. There could be 4 axes, each of which has a single value, leading to a single-valued NdArray.

The problem is that we also allow single-valued dimensions to be omitted from the NdArray (although they must still be present in the Domain). So some older CoverageJSON documents (legally) use a 0D NdArray with a 4D Domain and we didn't want to break backward compatibility.

In general, there is a distinction between the dimensionality of the coordinate space and the dimensionality of the geometry within it. E.g. it's possible to have a 0D geometry (a point) or a 1D geometry (trajectory) in a 4D coordinate space (domain).

[chris-little]

agreed to merge at 2022-05-18