add_dimension: Define behaviour for spatial dimensions better

[m-mohr]

Clarified that type spatial always adds a spatial dimension for the z axis.

[clausmichele]

Why can't x and y be added? A possible use case scenario could consist in:
load_collection over an AOI -> reduce_dimension over x (or y) -> add_dimension x (or y) with an "aggregated" label representing that area.

[soxofaan]

I agree with @clausmichele , is there a particular reason or background story to add this constraint?

[m-mohr]

There's no way to specify the actual axis for the spatial dimensions, so just adding a spatial dimension without axis is ambiguous. The most reasonable (and presumably most common) way out for me was to specify that it's z.

[soxofaan]

maybe use the name argument as axis identifier in some way?

[m-mohr]

That seems a little to implicit for me, what would a user get that specifies lat, lon, long, height, etc. I think defaulting to z is fine. If we ever encounter a usecase for x/y (has anyone encountered one?) we could just add an axis parameter to add_dimension that defaults to z.

[soxofaan]

what would a user get that specifies lat, lon, long, height, etc.

yeah, but when adding a dimension with name "lat" or "lon" or "long" it also does not make sense to do that with axis "z".

Also what if the cube already has an axis "z"?

What about something like

... If the type is spatial, the axis is the first option that does not exist yet from "x", "y", "z"

or

... If the type is spatial, the axis depends on the dimension name: names "x", "lon", "long" map to axis "x", names "y", "lat" map to "y", and all other options map to "z"

[m-mohr]

This makes sense. I've added:

If the type is spatial, the axis is the first option that does not exist yet from x, y, and z. This refers to the axis of the dimension, not the name of the dimension.

The axis depends on the type (according to the cube:dimensions axis property). This makes it irrelevant whether it's named x, lat, lon, y or whatever.

Anyone available for a review? @soxofaan @clausmichele

[clausmichele]

@m-mohr what if I ask to add a spatial dimension but x,y,z already exist? Return DimensionExists exception?

[m-mohr]

Would it make sense to allow multiple z-dimensions? Not sure :-)

[soxofaan]

Would it make sense to allow multiple z-dimensions?

Stacking multiple "other" or "band" dimensions is something imaginable I think: e.g. one dimension with labels "NDVI" and "EVI", and another dimension with "mean", "min", "max", "stddev", which would be orthogonal to that.

But spatial dimensions are tied to an axis (x, y or z), and I don't think it make sense, mathematically, to have for example multiple x axes

[soxofaan]

Apart from also adding a note on some DimensionExists error, I think this is already fine

[m-mohr]

For x and y, I agree. I'm just unsure about the z dimension. Could we repesent e.g. altitude and pressure at the same time in a dataucbe, which afaik in climate are both related to z?! I'm not really into these climate/atmosphere details and how they handle this.

[soxofaan]

"altitude" sounds like a spatial dimension along a spatial axis with a spatial resolution (meter or kilometer).

But I would think "pressure" is a dimension of type "other" with a non-spatial unit (Pascal or bar or whatever is most common in practice)

[m-mohr]

I think I tend to agree, I just don't want us to block any usecases from Climate or similar domains.

[clausmichele]

I think I tend to agree, I just don't want us to block any usecases from Climate or similar domains.

@pierocampa maybe could help here, from our recent discussion he seemed to be fine using other type of dimension, at the end it doesn't make any difference in the result

[pierocampa]

@pierocampa maybe could help here, from our recent discussion he seemed to be fine using other type of dimension, at the end it doesn't make any difference in the result

I have seen multiple time axis but not multiple vertical ones, like having altitude and pressure levels (which anyway would rather be modeled as OTHER-typed like soxofaan said).
Having multiple horizontal spatial axis I have also not seen, and I tend to think it begins to be too much of a freedom for the user, and could generate more confusion than benefits, just my point of view.

I am not sure about the terminology here (axis Vs dimension) and whether you are constraining them to be 1:1 : although the datacube is usually aligned to the axes (or dimensions?) of the space-time-other domain/fishtank, they should not be that coupled imho, so to allow for eg. tilted grids.

So you have grids/datacube axes on one side you can call whatever, and domain dimensions which shall be typed and whose label should be that of the CRS definition in the EPSG database, for what concerns the spatial axes. Each "label" of a datacube, eg. the coordinates of each grid point, will be N-tuples of the N-dimensional domain/fishtank.

This is how OGC models the grids/coverages, but this might be also getting a bit OT : )

[m-mohr]

@pierocampa Thank you, that's good insight, appreciated.