If weights are built for a 3D data source and then later a subset of n levels is selected from that Dataset or Dataarray, regridding will still use the first n levels from the weights (which is wrong).
We would need that the Dataset 'remembers' the positions of the original levels.
To address this I suggest that if the Dataset contains a special coordinate 'idx_3d' (depending on the vertical dimension) that will be used as a 'memory' of the original levels.
To explain better: a Dataset initially could have idx_3d which is simply the range of existing levels (i.e. [0, 1, 2 ... 75]) as a function of the vertical dimension (say 'level'), while maybe level could be [1000 900 800 700 600 500 ...] . Then, if later we select specific levels from that dataset (e.g. with .sel(level=[800, 700, 600]), also idx_3d is subselected and will contain [2, 3, 4] . If later we apply smmregrid to this resulting array, the correct weights of levels [2, 3, 4] will be used.
The iplementation would be such that this happens only if idx_3d is among the coordinates of the Dataset/Dataarray.
jhardenberg
commented
10 months ago
If weights are built for a 3D data source and then later a subset of
nlevels is selected from that Dataset or Dataarray, regridding will still use the firstnlevels from the weights (which is wrong). We would need that the Dataset 'remembers' the positions of the original levels. To address this I suggest that if the Dataset contains a special coordinate 'idx_3d' (depending on the vertical dimension) that will be used as a 'memory' of the original levels. To explain better: a Dataset initially could have idx_3d which is simply the range of existing levels (i.e. [0, 1, 2 ... 75]) as a function of the vertical dimension (say 'level'), while maybe level could be[1000 900 800 700 600 500 ...]. Then, if later we select specific levels from that dataset (e.g. with.sel(level=[800, 700, 600]), alsoidx_3dis subselected and will contain[2, 3, 4]. If later we apply smmregrid to this resulting array, the correct weights of levels[2, 3, 4]will be used. The iplementation would be such that this happens only ifidx_3dis among the coordinates of the Dataset/Dataarray.