Daskify 'rasters' module: roadmap

[remi-braun]

There is still numerous functions that needs to be daskified, without clear implementation in already used libraries.

• rasters.sieve: No clear implementation found anywhere. Maybe directly call xr.apply_ufunc like here?
• rasters.merge_gtiff: No clear implementation found anywhere. Existing implementation in rioxarray but rasterio-based and no sign of Dask in it.
• rasters.unpackbits: No clear implementation found anywhere. Maybe directly call xr.apply_ufunc because np.unpackbits don't exist in dask?
• rasters.read_bit_array: Should be straightforward once rasters.unpackbits is daskified
• rasters.vectorize: use Geoutils
• rasters.hillshade and rasters.slope: Use xarray-spatial (or xdem, but seems less mature)?
• rasters.collocate: Use xarray-regrid in case rewriting ref coordinates in place isn't sufficient and reprojecting is too much?

[remi-braun]

Use xarray-spatial in f4b1ff1

[remi-braun]

It would be better to use geoutils + xdem rather than xarray-spatial as soon as:

• we switch to Python 3.10 as minimum dependency
• this PR is merged in order to have a xarray accessor for geoutils rather than converting our objects to Raster objects
  Add Xarray accessor mirroring Raster class GlacioHack/geoutils#446
• this PR is merged in order to have a xarray accessor for xdem rather than converting our objects to DEM objects
  Add dem Xarray accessor mirroring DEM class GlacioHack/xdem#656

These two libraries allow to daskify a wider range of functions and drop the datashader dependency needed for xarray-spatial.

In geoutils / xdem, I found:

• ✔️ daskified reprojection + collocate (with Match-reference: a reference raster can be passed to match resolution, bounds and CRS during reprojection. !)
• ❌ daskified sieve
• ✔️ daskified merge
• ❌ daskified unpackbits
• ✔️ daskified vectorize
• ✔️ daskified hillshade, slope, aspect, ... in xdem