Select on non-dimension coordinate

[SBouchard01]

The problem

I have a coordinate that has 3 sub-coordinates I want to apply selections to. I don't want to use MultiIndex objects, as I'm using stack and unstack later on, and I don't want those coordinates to unstack and add NaN all over my array. Here is an example of what the DataArray would look like :

v = xr.DataArray(
    values, # of shape (time, space)
    dims=('time', 'space'),
    coords = {
        'time': times,
        'space': x*y*z,
        'x': (('space',), x),
        'y': (('space',), y),
        'z': (('space',), z),
    },
)

The constraint is that values is of shape (time, space) so I can't have x, y and z as dimension coordinates.
I want to select slices of v with conditions on x, y and z, but v.sel(x=slice(0.1, 0.3), y=slice(2.5, 2.7), z=slice(4.2, 4.4)) does not work as x, y and z are not dimensions coordinates:

KeyError: "no index found for coordinate 'x'"

Note

This is a simple example, so I'm using time and space, in reality those would be indices related to each value in the last dimension of the array, hence the impossibility to increase the array dimensions

What I have tried

• MultiIndex objects do not work as expected, and as I said earlier I don't want those to unstack later

TypeError: float() argument must be a string or a number, not 'slice'

• The following is a workaround, but I would like to avoid the hassle of scanning & selecting for non-dimensions coordinates in my dict before calling .sel(dict):

v = v.set_xindex('x')
v = v.set_xindex('y')
v = v.set_xindex('z')

v.sel(x=slice(0.1, 0.3)).sel(y=slice(2.5, 2.7)) # this is not ideal, if it was all in one .sel, that would be perfect :/

Tip

Notice that I call .sel on each coordinate separately, otherwise I get the following error:

ValueError: Xarray does not support label-based selection with more than one index over the following dimension(s):
    'space': 3 indexes involved
Suggestion: use a multi-index for each of those dimension(s).

I'm interested to know if there is a workaround that could allow me to use .sel on v without having a MultiIndex object in the coordinates

[tayfuryldz]

What you are running into is a current limitation of .sel(), not just a missing index setup.

After:

v = v.set_xindex("x")
v = v.set_xindex("y")
v = v.set_xindex("z")

each of those indexes can independently translate its label query into a positional indexer for the same underlying space dimension. A single .sel(x=..., y=..., z=...) would therefore produce three positional indexers for space. Current xarray explicitly rejects that merge; merge_sel_results() counts dimensions returned by each index query and raises exactly the "more than one index ... 'space': 3 indexes involved" error when a dimension occurs more than once.

If the semantics you want are "keep points satisfying all three coordinate ranges", boolean indexing expresses that directly and doesn't require a MultiIndex:

mask = (
    (v.x >= 0.1) & (v.x <= 0.3)
    & (v.y >= 2.5) & (v.y <= 2.7)
    & (v.z >= 4.2) & (v.z <= 4.4)
)

result = v.where(mask, drop=True)

or, if you specifically want positional selection:

result = v.isel(space=mask)

(the exact choice between where(..., drop=True) and isel depends on whether preserving the masking semantics matters for your real object).

Your chained .sel(x=...).sel(y=...).sel(z=...) works because each call resolves only one index at a time against the progressively reduced object. There currently isn't an equivalent single .sel() call that intersects several independent indexes which all target the same dimension.

I checked current xarray's indexing implementation before answering; this restriction is explicit in merge_sel_results(), so adding three set_xindex() calls alone cannot make the combined form work.