Support model initialization time + ensemble member + forecast_tau axes

[ahuang11]

Title

[dcherian]

An example use case would help here.

If there is an associated standard_name, some stuff should already work. I guess I don't know what you're asking for :) Please add examples.

[ahuang11]

To start, I would love if it can label and standardize the coordinates in this dataset (although I know the conventions say IRIDL lol, but in general, forecast datasets with lead and initializations).

import xarray as xr
import cf_xarray
ds = xr.open_dataset('http://iridl.ldeo.columbia.edu/SOURCES/.Models/.NMME/.NCEP-CFSv2/.HINDCAST/.MONTHLY/.sst/dods', decode_cf=False)
ds.cf.coordinates

Once that's done, it'd be nice to have something like verification time alignment where it's able to automatically align a verification dataset with a forecast. https://github.com/mktippett/ENSO/blob/master/ForecastVerification.ipynb

cfsv2_ds["o"] = (( "L", "S"), o.sel(T=T))
cfsv2_ds

To me, it sounds like it's within the scope of cf_xarray since I think cf_xarray already handles rolling longitudes across the edges if I am not mistaken? So maybe a new accessor method: verf_ds.cf.align(model_ds)

[ahuang11]

Maybe the title should state "axes" instead of "coordinates" since these dimensions are usually 1D besides verification time
which consists of (Initialization, Lead), or more generally initialization_time + forecast_lead = verification_target, or I guess it can also be the other way around: verification_target - initialization_time = forecast_lead

ini (ini)
tau (tau)
time (ini, tau)

ini (ini)
time (time)
tau (time, ini)

[dcherian]

Can you expand the attributes for the coordinate variables in your screenshot please?

I am sympathetic to some form of this request.

1. We need to come up with nice names and criteria for these dimensions though.
2. OTOH, I imagine that we'd only be matching standard_name, maybe its enough to generalize _get_axis_coord to do just that.

[ahuang11]

[dcherian]

So you want to be able to do ds.cf.mean("realization") and ds.cf.rename(realization="ensemble_member")?

[ahuang11]

Sure plus:
ds.cf.rename_like -> verf_ds.cf.align_like(fcst_ds)
where
verf_ds contains target/valid time
and
fcst_ds contains initialization/tau
which xarray does with
cfsv2_ds["o"] = (( "L", "S"), o.sel(T=T))

[dcherian]

Sure plus: ds.cf.rename_like -> verf_ds.cf.align_like(fcst_ds) where verf_ds contains target/valid time and fcst_ds contains initialization/tau

This is a little too specialized but we should add support for dims other than x,y,z,t,latitude,longitude,vertical,time. With support for standard_name you could be doing o.cf.sel(target_time=fct_ds.cf["initialization_time"]) assuming target_time and initialization_time are the appropriate standard names.

[ahuang11]

Does rolling handle taking the means across the edges of the dataset e.g. sphere wraps around, or is it simply the equivalent of doing ds.rolling.mean(lon=5)? https://cf-xarray.readthedocs.io/en/latest/examples/introduction.html#Rolling-&-coarsen

Maybe I was mistaken.

[dcherian]

No it does not. Right now all its doing is rewriting args, kwargs to let you make use of attributes.

[ahuang11]

Okay then what I requested in my latter comment is out of scope for now~ but eventually that'd be ideal!

I thought since it was able to detect it's a longitude, make full use of it.
pydata/xarray#2007