Expose sample on the rioxarray accessor

[martinfleis]

I was recently comparing the performance of point sampling from a raster in xvec (xarray-contrib/xvec#81) and learned that when using xarray's sel, the sampling is about 40x slower than if I use rasterio's sample method (see the notebook in the linked issue).

I can possibly use sample from the rasterio object available within the rio._manager as list(dtm_da.rio._manager.acquire().sample(list(zip(x, y)))) but that is using a private API of rioxarray.

Would it be in scope to expose sample directly on the rio accessor? We could then consume it in xvec when dealing with rioxarray-backed DataArrays.

[snowman2]

rio._manager is something that is easily lost after performing operations on the DataArray|Dataset and is only available when the DataArray|Dataset is opened with rioxarray.open_rasterio or when using engine="rasterio". Unfortunately, due to these limitations, exposing sample on the rio accessor would likely cause confusion as it would not always work as expected.

I would be interested to know if this method is any faster.

If it is, that may enable us to add rio.sample:

def sample(self, *args, **kwargs):
    try:
       return self._manager.acquire().sample(*args, **kwargs)
    except AttributeError:
        pass
    with MemoryFile() as memfile:
        self._xds.rio.to_raster(memfile.name)
        with memfile.open() as dst:
             return dst.sample(*args, **kwargs)

[martinfleis]

Using the DTM from the notebook above, the trip via MemoryFile takes 36s, about 4x more than using the built-in sel method of xarray, so there's no point in doing that.

It seems that exposing rasterio's sample is not a feasible option. Thanks anyway!

[snowman2]

These observations are also useful to be aware of: xarray-contrib/xvec#81 (comment)

Thanks for the proposal!