Allow custom dtypes and fills in _zonal_stats_rasterize

[martinfleis]

I suggest that instead of forcing a float32 dtype, it's better to check the input raster data type first. If the input data type is float64 or float32, the output should use the original data type. If integer, then the output can use float32.

Also, I think it's better to set fill and dtype as two input parameters for the _zonal_stats_rasterize function so that the user can be aware of that and may change it as needed for memory efficiency. Cheers.

Originally posted by @coolzhao in #85

[martinfleis]

Wanted to spit it from #85 as this is a different issue.

Maybe rather than exposing, we could use the smallest int possible (based on the number of geoms) but then we need to remove the fill value from groupby. And subsetting using where would again do a float array and fill locations of original fill value with nan. There may be some smarter way, like remove the group from groupby but we would need to explore that. So far, I think that hard-coded float32 is not that critical.

[dcherian]

like remove the group from groupby b

You can pass labels to UniqueGrouper (https://docs.xarray.dev/en/stable/user-guide/groupby.html#grouper-objects) to exclude specific values

[martinfleis]

@dcherian but is that doable in our case where we pass a DataArray as groups?

xvec/xvec/zonal.py

Line 65 in ab3f22d

groups = acc._obj.groupby(xr.DataArray(labels, dims=(y_coords, x_coords)))

It seems to me that I can use UniqueGrouper only if grouping along an existing dimension.

[dcherian]

yes we made this choice.

You can do:

obj = acc._obj.copy()
obj["__labels__"] = xr.DataArray(labels, dims=...)
groups = obj.groupby(__labels__=UniqueGrouper(labels=...))