Blue marble can only aggregate tree color over relatively large regions, but we still need the original vegetation that can make up that color. For example, there's a desert and savanna. Pure sand and pure grass has a relatively stable color over some continuous regions. However, the aggregated color of the transition between the two is created by patches of grass dispersed in the sand (or vice-versa), with various ratios depending whether they are closer to the desert or to the savanna. The algorithm basically computes the probability of land types to exists in given region, and resolves the final values by picking the best set of basic types. It's not trivial: it should account for seasonal changes (using 12-month color inputs), it must pick the possible land types reflecting their probability of occurrence at that particular place, depending on parameters like temperature and precipitation. For this it needs reliable land type and vegetation definitions.
Output of this mapper are probabilities of land types that can occur. This is fed to the fractal terrain mixer, which uses the probabilities to create a naturally looking (with fractal-transitions) blend of the terrain.
Obviously that's a lot of things to make it work, but I believe the results will be worth it.
Regarding your question, it would probably help to match the dominant vegetation and land type with a few world locations where it covers a large area.
