sdba - Broadcasting with two new coordinates of differing dimensions fails

[aulemahal]

• xclim version: master
• Python version: 3.8
• Operating System: Linux

Description

In QuantileDeltaMapping.adjust with grouping (here 'time.month') there is a broadcasting operation that fails when sim and hist have additional dimensions (other then quantiles and time).
In fact, calls of xclim.sdba.utils.broadcast fails when grouping is used, interp is not 'nearest' and both grouped and the coordinate in sel have dimensions that are not trageted by the broadcasting operation.

What I Did

from xclim.sdba.adjustment import QuantileDeltaMapping
QDM = QuantileDeltaMapping(group='time.month', interp='linear'
QDM.train(ref, hist)
scen = QDM.adjust(sim)

Where, ref has one spatial dim + 'time' and hist and sim have the same spatial dim, 'time' and an additional dim, in my case realization.

What I Received

ValueError: axes don't match array

The traceback is quite long, but it all comes down to xr.interp.

[aulemahal]

xarray's interp doesn't understand multidimensionnal interpolation when the new coordinates have differing set of coords, so I tried to broadcast all arrays of the sel dictionary together which makes it work but enlarges the arrays a lot. Hopefully, no data is created so memory should be ok. However, dask arrays chunked along one of the additional dimensions will make broadcast fail. So, that's not a solution.

Kinda stuck on this. A way would be to call a xr.apply_ufunc with a wrapper of scipy.interpolate.interpn, but that's a lot of machinery just because xarray as kind of a bug hasn't implemented the thing yet.

[huard]

Is there an open issue about this ?

[aulemahal]

Not that I can found.

[huard]

I suggest you design a short, self-contained example illustrating the issue and open an issue with it.

[aulemahal]

I didn't realize up to now, but it is worth noting that xarray raises a "NotImplementedError". It's not a real bug.

[aulemahal]

Corresponds to issue pydata/xarray#4058.