unstructured grids

I tried regionmask on unstructured grids and it works with a tweak: https://gist.github.com/aaronspring/1b45cf22f80360ddc4ef13af1cadec63

Unstructured grids like ICON just have one spatial dimension like ncells and 2 spatial coordinates clon and clat.

To avoid:

/work/mh0727/m300524/conda-envs/pymistral/lib/python3.8/site-packages/regionmask/core/mask.py in _get_LON_LAT_out_shape(lon, lat, fill)
    467 
    468     if ndim == 1:
--> 469         LON, LAT = np.meshgrid(lon, lat)
    470     elif ndim == 2:
    471         LON, LAT = lon, lat

and use L470 instead, I add a fake dimension dummy:

ds = ds.assign_coords(clon=ds.coords['clon'].expand_dims('dummy'))
ds = ds.assign_coords(clat=ds.coords['clat'].expand_dims('dummy'))

To identify unstructured grids, we could maybe use the metadata CDI_grid_type from the xr.DataArrays variables:

ds[v].attrs
{'standard_name': 'sea_water_potential_temperature',
 'long_name': 'sea water potential temperature',
 'units': 'deg C',
 'code': 2,
 'CDI_grid_type': 'unstructured',
 'number_of_grid_in_reference': 1}

or we just check for the number of dimensions used in lon and lat and add such a fake dimension.

Do you think regionmask could handle such unstructured grids without such tweaks? A solution should also work for other unstructured output from MPAS @mgrover1, FESOM @koldunovn, etc

[mgrover1]

Here are a couple use cases. Both of these can be accessed using the geocat-datafiles repository Here is an MPAS file:

<xarray.Dataset>
Dimensions:           (nCells: 163842, Time: 1)
Dimensions without coordinates: nCells, Time
Data variables:
    lonCell           (nCells) float64 ...
    latCell           (nCells) float64 ...
    surface_pressure  (Time, nCells) float64 ...

And a CAM-SE (unstructured atmospheric climate model output) case

<xarray.Dataset>
Dimensions:  (ncol: 48602)
Dimensions without coordinates: ncol
Data variables:
    T850     (ncol) float32 283.6 283.1 282.5 282.2 ... 265.8 266.3 267.0 267.3
    lon      (ncol) float64 315.0 315.8 317.2 318.0 ... 135.0 133.7 136.3 135.0
    lat      (ncol) float64 -35.26 -35.65 -36.26 -36.62 ... 36.68 36.68 36.05
Attributes: (12/16)
    np:                        4
    ne:                        30
    Conventions:               CF-1.0
    source:                    CAM
    case:                      b1850cam5cn_camse
    title:                     UNSET
    ...                        ...
    initial_file:              /glade/scratch/juliob/cami_minimal_ne30np4_L30.nc
    topography_file:           /fis/cgd/cseg/csm/inputdata/atm/cam/topo/USGS-...
    history:                   Thu Jul 21 11:14:15 2011: ncatted -O -a yrs_av...
    nco_openmp_thread_number:  1
    yrs_averaged:              3-10
    description:               This file is a subset of a CAM-SE file. It was...

You can access these files for testing using the following:

import geocat.datafiles
import xarray as xr

# MPAS data
mpas_file = 'netcdf_files/MPAS.nc'
mpas_ds = xr.open_dataset(geocat.datafiles.get(mpas_file))

# CAM-SE data
camse_file = 'netcdf_files/camse_unstructured_grid.nc'
camse_ds = xr.open_dataset(geocat.datafiles.get(camse_file))

@aaronspring let me know if you have any questions

[mathause]

only do meshgrid when data are 2d and the coordinates are 1d?

I don’t look at the data only the coordinates. @aaronspring can you share a repr of the dataset? I could only call meshgrid if the dimension name is not equal to the coords name. I.e. call meshgrid:

<xarray.Dataset>
Dimensions:  (lat: 25, time: 2920, lon: 53)
Coordinates:
  * lat      (lat) float32 75.0 72.5 70.0 67.5 65.0 ... 25.0 22.5 20.0 17.5 15.0
  * lon      (lon) float32 200.0 202.5 205.0 207.5 ... 322.5 325.0 327.5 330.0
...

don’t call meshgrid:

<xarray.Dataset>
Dimensions:  (idx: 25, time: 2920)
Coordinates:
  * lat      (idx) float32 75.0 72.5 70.0 67.5 65.0 ... 25.0 22.5 20.0 17.5 15.0
  * lon      (idx) float32 200.0 202.5 205.0 207.5 ... 322.5 325.0 327.5 330.0
...

That might work.

At the end of the day unstructured meshes are not that common, so maybe just an additional argument to the function would be enough?

Yes that is probably necessary (although I prefer to not add yet another argument). Unless we have enough trust in the above solution.

---

b.t.w. - the __version__ was wrong when installing with conda (see the gist) - that should be fixed.