No that won't work. You need to mask the weights where the data is NaN. An untested and not very efficient way may be:

```python def coarsen_weighted_mean(da, weights, dims, skipna=None, boundary="exact"):

weighted_sum = (da * weights).coarsen(dims, boundary=boundary).sum(skipna=skipna)

masked_weights = weights.where(da.notnull())
sum_of_weights = masked_weights.coarsen(dims, boundary=boundary).sum()
valid_weights = sum_of_weights != 0
sum_of_weights = sum_of_weights.where(valid_weights)

return weighted_sum / sum_of_weights

```

An example (without NaNs though):

```python import xarray as xr import numpy as np

air = xr.tutorial.open_dataset("air_temperature").air weights = np.cos(np.deg2rad(air.lat))

we need to rename them from "lat"

weights.name = "weights"

c_w_m = coarsen_weighted_mean(air, weights, dict(lat=2), boundary="trim")

to compare do it for one slice

alt = air.isel(lat=slice(0, 2)).weighted(weights).mean("lat")

compare if it is the same

xr.testing.assert_allclose(c_w_m.isel(lat=0, drop=True), alt) ```

That's currently not possible. What you can try is the following:

(ds * coslat_weights).coarsen(lat=2, lon=2).sum() / coslat_weights.coarsen(lat=2, lon=2).sum()

but this only works if you don't have any NaNs.

@ahuang11 in #4210:

I want to do something similar as xesmf's weighted regridding, but without the need to install esmpy which has a lot of dependencies.

Are variations of the following possible? ds.weighted(coslat_weights).coarsen(lat=2, lon=2).mean()

ds.coarsen(lat=2, lon=2).weighted(coslat_weights).mean()

I think da.groupby(...).weighted(weights).mean() would be pretty sensible.

We could probably make both da.groupby(...).weighted(weights) and da.weighted(weights).groupby(...) return exactly equivalent objects, if desired.