In this case it's more likely a bug in the auxiliary vegetation density data (which is problematic and there are several interpretations of its value), or a manifestation of some of the other known issues (precision in texture sampling) that have to be worked around. Otherwise the problem would show everywhere, as the resolution of the color data is constant. Color data aren't used directly, they go through the fractal mixer that also uses other variables to determine the output, and the transitions are normally fractal-like even if the input is a rough raster.
Fractal mixer currently does not have access to the noise with longer wavelengths, and thus can't disperse rough color samples when they are not also accompanied with usable variations in the terrain topology. However, the existing natural layer isn't supposed to handle artificial, man-made changes like fields or parks, because they do not have a fractal distribution anyway. That's the task for the vector overlay. Those color map texels that map to agricultural areas should be actually blurred to make the distribution more natural-like, removing the aliasing altogether. The vector overlay will then produce the right polygonal shapes for the man-made changes. To fix the issue we need either to fix the biome data (blurring and anti-aliasing the civilization areas), or to add long-wavelength fractal channels able to handle the coarse transitions. We'll probably do both as the forest distribution needs the latter, and there are other uses for it (snow coverage ...).
