Include all CORDEX domains from py-cordex

[shaussler]

Description

The project py-cordex includes all CORDEX domains: https://py-cordex.readthedocs.io/en/latest/. It should be quite easy to generate pyresample metadata on-the-fly from the py-cordex project to offer all CORDEX domains as predefined projections in pyku.

Concretely, the pyku project permits to regrid data with either pyresample or xesmf by offering standardized projections. For example, one can project to EUR-11 like so:

ds.pyku.project('EUR-11')

The projections are defined using the pyresample format. For example:

EUR-11:
  description: EUR-11 rotated longitude latitude grid in accordance with CORDEX
    domains requirements https://cordex.org/domains/
  projection:
    proj: ob_tran
    o_proj: longlat
    o_lon_p: 0
    o_lat_p: 39.25
    lon_0: 18
    datum: WGS84
    no_defs: null
    type: crs
  shape:
    height: 412
    width: 424
  area_extent:
    lower_left_xy: [-28.43, -23.43]
    upper_right_xy: [18.21, 21.89]

CF-conform projection information are defined as separate metadata:

EUR-11:
    crs_name: rotated_pole
    crs_data:
        grid_mapping_name: rotated_latitude_longitude
        grid_north_pole_latitude: 39.25
        grid_north_pole_longitude: -162.0
        long_name: coordinates of the rotated North Pole
    y_coordinate: 'rlat'
    x_coordinate: 'rlon'
    CORDEX_domain: 'EUR-11'

WKT or Proj4 metadata can be automatically determined with the pyku dependency pyresample, or added to the configuration file.

Strictly speaking, the data from py-cordex can already be used to generate the transformations without change:

https://deutscherwetterdienst.github.io/pyku/tutorials/geographic_projections.html#Custom-projection-definition

Also of note is that pyku already offers possibility to resample unstructured grids. Here however, work will be needed in pyku to simplify the usage. Documentation on how to regrid ICON data can be found here:

https://deutscherwetterdienst.github.io/pyku/tutorials/CMORization.html#Attach-the-georeferencing
https://deutscherwetterdienst.github.io/pyku/tutorials/CMORization.html#Regridding-the-data

[shaussler]

@vmaurerDWD @larsbuntemeyer

[larsbuntemeyer]

@shaussler thanks, i have not much experience with pyresample, but i did some prototyping with GLCC data here. In that example, i also grab grid data from py-cordex and convert it for pyresample area defs. It works much easier with xesmf but pyresample seems to be the only package i know to be reasonably parallelized using dask under the hood.

[larsbuntemeyer]

Also note that the rotated pole grid definitions might be subject to change between CORDEX-CMIP5 and CORDEX-CMIP6. The actual grid definitions can be found in this repo: https://github.com/WCRP-CORDEX/domain-tables

[shaussler]

That exactly. Your prototyping is the core idea that we have built in pyku and how we began. Over the years we extended that core more than a bit. With respect to multiprocessing with pyresample, we use the pyresample numpy functions together with dask.array.map_blocks. So our usage of pyresample in pyku is fully parallelized using dask. I am sure other libraries can be faster, but it never was a bottleneck for us. And xesmf is also built in anyway. But I have personally a preference for pyresample. The installation of xesmf is a little hard and I like that we can just pip install pyresample.

Thanks a lot for the warning about the possible change of the rotated pole grid definition!