Ensure consistency of SST conditioning in multi-modal inference methods #42
Description
The SST conditioning uses different infilling strategies between the modalities of ERA5 and ICON. For ERA5, we use AMIP style zonal interpolation, while for ICON there is a land-sea mask and land pixels are filled with a constant global value.
For inference, this needs to be reproduced for the image-to-image use case so that interpolation conditioning is used for ERA5 and constant filling for ICON, i.e. the SSTs need to match the flag passed to the model. Channel infilling could be done in the same way, or conditioning could be dropped out here altogether.
For training, in the future this different conditioning could be resolved with a single conditioning technique, e.g. applying ICON's land mask to ERA5 and infilling with the same value.
Above a visualization of the AMIP-style interpolated SSTs