The most robust way is to mask hues and act on them individually. This allows to refine the mask to better follow feathered edges (hair, actual feathers, etc.) and avoid discontinuities in masks that may lead to chroma noise (the biggest flaw of the current color zones).
My mid-term plan is to add spectral capabilities to the color calibration. Currently, you have red (sort of), green (sort of) and blue (sort of), imagine instead you have 8 channels sampled evenly across visible wavelengths. @hanatos has given me published and unpublished material to achieve that, so once his latest paper gets accepted, I’m cleared to start working on that.