Hi @maxim.lamare !
Here we meet once more! Thanks for taking the time to answer.
I understand better the effect of angles on the output thanks to your point #1.
As for #2, indeed, separated VV/VH between ascending and descending is not a problem for machine learning applications. While we use these techniques, it’s not always the case. For instance, we could calculate indexes off of VV and VH values, such as the degree of polarization, which is linked to the “fullness” of a canopy. It can be approached as VV/(VV + VH) and can be used on our end to help with agronomic studies, but here we would have a difference of ~10% in mid-February between descending and ascending, making it harder to interpret and use (as humans at least ahaha).
However, we understand that it’s better to keep them unmerged and to use ascending and descending data on their own due to the scattering effects you mentionned, which are actually also usable and valuable data. It seem unusual to try to merge them so I guess we just haven’t figure things out enough yet 
Thanks for your patience and your ressources!
Léa