Observation noise uncertainty

[adelinehillier]

At some point we should think more deeply about how to set the observation noise variance. Currently observation_map_variance_across_time is being used even when very few time steps are being saved. In such cases, a scaled identity matrix is a decent option. Otherwise, we might try coming up with a more scientific uncertainty estimate, e.g. derived by perturbing the IC or BCs.

[glwagner]

Or perhaps regenerating the LESbrary at a coarser or finer resolution and estimating uncertainty from the differences?