xarray.Dataset.where() method force-changes dtype of DataArrays to float

Question

Problem description

I have a dataset with ints in them, and I'd like to select a subdataset by some criteria but I would like to preserve the integer datatype. It seems to me that Xarray force-changes the integer data to float datatype.

Example setup

Code

import numpy
import xarray

nums = numpy.random.randint(0, 100, 13)
names = numpy.random.choice(["babadook", "samara", "jason"], 13)
data_vars = {"num": xarray.DataArray(nums), "name": xarray.DataArray(names)}
dataset = xarray.Dataset(data_vars)
print(dataset)

Output

<xarray.Dataset>
Dimensions:  (dim_0: 13)
Coordinates:
  * dim_0    (dim_0) int64 0 1 2 3 4 5 6 7 8 9 10 11 12
Data variables:
    num      (dim_0) int64 93 99 49 35 92 14 41 57 28 59 74 1 15
    name     (dim_0) <U8 'babadook' 'samara' 'samara' 'samara' 'jason' ...
In [16]:

Example problem

Code

subdataset = dataset.where(dataset.num < 50, drop=True)
print(subdataset)

Output

<xarray.Dataset>
Dimensions:  (dim_0: 7)
Coordinates:
  * dim_0    (dim_0) int64 2 3 5 6 8 11 12
Data variables:
    num      (dim_0) float64 49.0 35.0 14.0 41.0 28.0 1.0 15.0
    name     (dim_0) <U32 'samara' 'samara' 'jason' 'babadook' 'jason' ...

Answer 1

That's because with numpy (which xarray uses under-the-hood) ints don't have a way of representing NaNs. So with most where results, the type needs to be coerced to floats.

If drop=True and every value that is masked is dropped, that's not actually a constraint - you could have the new array retain its dtype, because there's no need for NaN values. That's not in xarray at the moment, but could be an additional feature.