Applying regridder on dataset drops coordinate attributes

[yt87]

I am not sure it is a bug or an undocumented feature. Consider the following code:

import numpy as np
import xarray as xr
import xesmf as xe

ds = xr.tutorial.open_dataset(
    "air_temperature"
)
ds['lat'].attrs
ds_out = xr.Dataset(
    {
        "lat": (["lat"], np.arange(16, 75, 1.0), {"units": "degrees_north"}),
        "lon": (["lon"], np.arange(200, 330, 1.5), {"units": "degrees_east"}),
    }
)
regridder = xe.Regridder(ds, ds_out, "conservative")
ds_regridded = regridder(ds)
ds['lat'].attrs
ds_regridded['lat'].attrs

The output is:

{'standard_name': 'latitude',
 'long_name': 'Latitude',
 'units': 'degrees_north',
 'axis': 'Y'}

{}

{'units': 'degrees_north'}

Actually, all coordinate attributes are gone. Attributes for data variables, in this case air, are preserved. Interestingly, regridder assigns attributes to lat and lon for the output (but not to time), copied from ds_out

[yt87]

A quick workaround is to pass ds.copy() to regridder

[aulemahal]

@huard This looks like the same issue you were raising on slack!

The copy of attributes from ds_out is a feature indeed. As xESMF only handles the spatial dimensions, the regridding can only assign attributes to lat and lon in your case.

But the loss of attributes on ds is a problem. We need to investigate.

[yt87]

I think it would make sense to preserve the non-spatial coordinate attributes, such as time in the above example. Regridding does change their meaning.

[aulemahal]

The issue I see hee, is that the regridding does not know which coordinates of ds_out will appear on the regridded output, except for the spatial dims. Also, in you specific case, time was simply absent from ds_out. I think it makes more sense to preserve the attributes from the input (ds) exactly, instead of trying to guess how to "update" them following the regridding.

But if you have an idea on the logic behind this hypothetical update, this could be implemented!

[yt87]

Sorry, I missed the keep_attrs argument to BaseRegridder.__call__. Interestingly, if set to True, it also preserves attributes in the argument, that is, my problem disappears. I still think there is a bug, the input dataset should not be modified in either case.

[yt87]

This seems to be xarray issue:

import xarray as xr

def magnitude(a, b):
    func = lambda x, y: np.sqrt(x**2 + y**2)
    return xr.apply_ufunc(func, a, b, keep_attrs=False)

array = xr.DataArray([1, 2, 3], coords=[("x", [0.1, 0.2, 0.3], {'axis': 'X'})])
array
magnitude(array, -array)
array['x'].attrs

The last line prints an empty dictionary. The documentation states https://github.com/pydata/xarray/blob/main/xarray/core/computation.py#L836-L1210

keep_attrs : {"drop", "identical", "no_conflicts", "drop_conflicts", "override"} or bool, optional
   - 'drop' or False: empty attrs on returned xarray object.

Returned, not passed, so likely this is not an intentional behaviour.

[charlesgauthier-udm]

I opened an issue on xarray regarding this bug.

[huard]

Should we close this here, then ?

[yt87]

I have nothing against closing it here.