Using machine learning to distinguish km-scale climate models and observations on a regional scale

Using conceptual idea of Labe & Barnes (2022):

• On CORDEX domain using: CORDEX models, nextGEMS models, observations (e.g. ERA5-Land)
• Train a CNN to distinguish models from each other (multi-class problem)
• Apply classifier to observations to see which model the observations are assigned to.
• Then use the same approach on a smaller domain (Austria): add ÖKS15 for training
• Again apply the classifier to observations to see which model the observations are assigend to (Is ÖKS15 now the "best" on this domain?)

The results would be interesting in two ways:

• If observations are most frequently assigned to ÖKS15: at least for smaller domains, nextGEMS has no added value compared to regional downscaled data
• If observations are most frequently assigned to nextGEMS: nextGEMS has high added value, we may no longer need regional downscaling