Where does the dry mask enter the RPS calculation?
This question came up in the technical town hall in https://renkulab.io/gitlab/aaron.spring/s2s-ai-challenge/-/issues/13, thanks @david.landry for asking!
The verification notebook and scorer do not explicitly use the dry mask.
However, we already masked the dry regions already in the ground truth/observations, which are used as observations
in xs.rps
. Masked grid points are np.NaN
in observations
and therefore lead to xs.rps
np.NaN
, thus not contributing to the final aggregated score.
The categorized observations forecast-like-observations_2020_biweekly_terciled.nc
, which are generated in renku_datasets_biweekly.ipynb
are masked by tp_arid_mask
in make_probabilistic
, see In [30]
in renku_datasets_biweekly.ipynb
.
I did not remember this info in the call and just dived into the code. I did not need to change anything.
Here's a screenshot of the dry mask (commented out in In [26]
in in renku_datasets_biweekly.ipynb
):
See here how this masks/whitens the deserts:
Looks plausible @david.landry?