Apparently there is a linear relationship between the $NDVI$ and the $K_c$ value, which would make it possible to derive landuse-specific timeseries of the $K_c$ values. However, the $K_c$ refers to the "crop" water requirements under standard conditions. In reality, most of the "natural" vegetation should be under water stress, and derivation of the $K_c$ curves would only be possible for a few years since the land use change is quite fast and would soon lose its validity going back in time. This means that the $K_c$ then should reflect the growth conditions due to precipitation and weather conditions of that particular year and as such should be more related to the actual evapotranspiration than the potential evapotranspiration. This should distort things when estimating the actual evapotranspiration as $$ET_a = ET_{ref} \times K_c \times f(z) $$ where $f(z)$ is a function that reduces the potential evapotranspiration according to the soil moisture, such as described here.

The $K_c$ values obtained for fallow land with this method are considerably below the simulated values from the MABIA method, which results in an $ET$ that is too low, leaving too much water in the system and hence producing too much recharge (around 200mm for fallow).

The question now is: Is this $K_c$ approach transferable at all to non-agricultural landuse classes, if so, how can I obtain these values and how would it be used with a soil moisture model and if not, what other ways are there to get a rough estimate of the potential evapotranspiration in these classes?

I am aware that this will yield very rough approximations at the very best, but it is a starting point.