xarray groupby does not drop dimension when squeeze

[ecimoli]

I have MxNxT xarray with the following dimensions and coordinates:

<xarray.DataArray (gridcell: 238, window: 9, traits: 3)>
dask.array<where, shape=(238, 9, 3), dtype=float32, chunksize=(14, 9, 3), chunktype=numpy.ndarray>
Coordinates:
    band         (traits) int32 dask.array<chunksize=(3,), meta=np.ndarray>
    spatial_ref  int32 0
  * traits       (traits) <U7 'trait_1' 'trait_2' 'trait_3'
  * gridcell     (gridcell) MultiIndex
  - latitude     (gridcell) float64 6.054e+06 6.054e+06 ... 6.054e+06 6.054e+06
  - longitude    (gridcell) float64 6.042e+05 6.043e+05 ... 6.043e+05 6.043e+05
  * window       (window) MultiIndex
  - x_window     (window) int64 0 0 0 1 1 1 2 2 2
  - y_window     (window) int64 0 1 2 0 1 2 0 1 2
Attributes:
    scale_factor:  1.0
    add_offset:    0.0

I am trying to apply groupby the "gridcell" dimension so I can perform an operation on each of those groups using xr.map, or apply_ufunc function.

kernel_group = kernels_xr.groupby('gridcell', squeeze=True)

When I do that, the resulting group matrix is 1xNxT, but I only want groups to be NxT, isn't that what squeeze is supposed to do? I don't see any changes whatsoever. Any other suggestions are very much appreciated as well.

Your help would be and immense relief!

Thanks

[mathause]

I am trying to apply groupby the "traits" dimension so I can perform an operation on each of those groups using xr.map function.

kernel_group = kernels_xr.groupby('gridcell', squeeze=True)

Shouldn't the call be kernel_group = kernels_xr.groupby('traits', squeeze=True)?

You are currently grouping over gridcell - can you check if your gridcell actually form groups with more than one member? If not there is nothing to group by...

import numpy as np
import xarray as xr

air = xr.tutorial.open_dataset("air_temperature")
a = air.stack(gridcell=("lat", "lon"))
len(np.unique(a.gridcell.values)) == len(a.gridcell.values)

[ecimoli]

Apologizes mathause, my typo, i want to group by gridcell, and is indeed forming the groups that i desired. Figured out that re assign the coordinates to a non-multindex and float64 vector, it works and drops the specified dimension, but i lose the multindex stacked information.
Any idea? Thanks again for the time and support in replying

[mathause]

Here is a small example. Kind of works

import numpy as np
import xarray as xr

data = np.arange(8).reshape(2, 4)
da = xr.DataArray(data, dims=("y", "x"), coords=dict(y=[1, 2], x=[1, 1, 1, 1]))

da = da.stack(gridcell=("x", "y"))

da_g = da.groupby("gridcell").mean()

It keeps the multiindex but renames the levels (this could be considered a bug).

<xarray.DataArray (gridcell: 2)>
array([1.5, 5.5])
Coordinates:
  * gridcell          (gridcell) MultiIndex
  - gridcell_level_0  (gridcell) int64 1 1
  - gridcell_level_1  (gridcell) int64 1 2

This can be fixed like so:

mi = da_g.indexes["gridcell"]
da_g.assign_coords(gridcell=mi.set_names(["x", "y"]))

Do you experience something different? If so a small example would be helpful. (There is a refactor of the way indexes are treated underway - this may change how this behaves).

[ecimoli]

Thanks mate! That did work out as a solution! So much appreciated.

Xarray is magic, hope it keeps improving!

Emiliano