Return multiple indexes from multi-coordinate index `isel`

[benbovy]

Is your feature request related to a problem?

Let's consider the following DataArray with a rasterix RasterIndex associated to the x/y coordinates:

>>> da
<xarray.DataArray 'band_data' (y: 332, x: 316)> Size: 420kB
[104912 values with dtype=float32]
Coordinates:
  * x            (x) float64 3kB -3.938e+06 -3.912e+06 ... 3.912e+06 3.938e+06
  * y            (y) float64 3kB 4.338e+06 4.312e+06 ... -3.912e+06 -3.938e+06
Indexes:
  ┌ x        RasterIndex
  └ y

Data selection on x and y at arbitrary array positions will drop the raster index since it is built around affine coordinate transformation:

>>> da2 = da.isel(x=[10, 31], y=[0, 4])
>>> da2
<xarray.DataArray 'band_data' (y: 2, x: 2)> Size: 16B
[4 values with dtype=float32]
Coordinates:
    x            (x) float64 16B -3.688e+06 -3.162e+06
    y            (y) float64 16B 4.338e+06 4.238e+06
>>> da2.xindexes
Indexes:
    *empty*

It would be convenient if instead default (pandas) indexes were created for those coordinates in the result:

>>> da2
<xarray.DataArray 'band_data' (y: 2, x: 2)> Size: 16B
[4 values with dtype=float32]
Coordinates:
  * x            (x) float64 16B -3.688e+06 -3.162e+06
  * y            (y) float64 16B 4.338e+06 4.238e+06
>>> da2.xindexes
Indexes:
    x        PandasIndex
    y        PandasIndex

This is not currently possible, though, since xarray.Index.isel() returns either one Index object or None.

Describe the solution you'd like

Change Xarray's Index API from:

class Index:

    def isel(...) -> Index | None:

to

class Index:

    def isel(...) -> Index | dict[Hashable, Index | None] | None:

So that it is possible to return different index objects (or drop the index) for each coordinate variable associated with the index.

Describe alternatives you've considered

In the example above, one alternative would be to encapsulate PandasIndex instances within rasterix's RasterIndex. This would make the implementation of the latter much more complicated, though.

Additional context

https://github.com/dcherian/rasterix/issues/50

https://github.com/dcherian/rasterix/pull/43

[dcherian]

Index | dict[Hashable, Index | None] | None:

This type seems pretty magical. Perhaps we just ignore this until there is a real strong need for it.

[benbovy]

Simply have dict[Hashable, Index] as return type would feel less magical ; it is also even more consistent with the Dataset._indexes flat dictionary.