从 xarray 数据集中的某些变量中删除维度

Question

I have an xarray Dataset where some variables have more dimensions than necessary (eg, a 3D dataset where the "latitude" and "longitude" variables also vary along time). How do I remove the extra dimensions?

For example, in the dataset below, 'bar' is a 2D variable along the x and y axes, with constant values along the x axis. How do I remove the x dimension from 'bar' but not 'foo'?

>>> ds = xr.Dataset({'foo': (('x', 'y'), np.random.randn(2, 3))},
                    {'x': [1, 2], 'y': [1, 2, 3],
                     'bar': (('x', 'y'), [[4, 5, 6], [4, 5, 6]])})
>>> ds
<xarray.Dataset>
Dimensions:  (x: 2, y: 3)
Coordinates:
  * x        (x) int64 1 2
  * y        (y) int64 1 2 3
    bar      (x, y) int64 4 5 6 4 5 6
Data variables:
    foo      (x, y) float64 -0.9595 0.6704 -1.047 0.9948 0.8241 1.643

Answer 1

The most direct way to remove the extra dimension (using indexing) results in a slightly confusing error message:

>>> ds['bar'] = ds['bar'].sel(x=1)
ValueError: dimension 'x' already exists as a scalar variable

The problem is that when you do indexing in xarray, it keeps around indexed coordinates as scalar coordinates:

>>> ds['bar'].sel(x=1)
<xarray.DataArray 'bar' (y: 3)>
array([4, 5, 6])
Coordinates:
    x        int64 1
  * y        (y) int64 1 2 3
    bar      (y) int64 4 5 6

This is often useful, but in this case the scalar coordinate 'x' on the indexed array conflicts with the non-scalar coordinate (and dimension) 'x' when you try to set it on the original dataset. Hence xarray errors instead of overriding the variable.

To get around this, you need to drop the scalar 'x' after indexing. In the current version of xarray, you can do this with drop :

>>> ds['bar'] = ds['bar'].sel(x=1).drop('x')
>>> ds
<xarray.Dataset>
Dimensions:  (x: 2, y: 3)
Coordinates:
  * x        (x) int64 1 2
  * y        (y) int64 1 2 3
    bar      (y) int64 4 5 6
Data variables:
    foo      (x, y) float64 -0.9595 0.6704 -1.047 0.9948 0.8241 1.643

In future versions of xarray (v0.9 and later), you will be able to drop coordinates when indexing by writing drop=True , eg, ds['bar'].sel(x=1, drop=True) .