Assuming that works for the determining of the standardized, publicized assessment method, I think it stops working as well when people look at the published assessment method and have huge incentives to find ways to boost the assessed value of improvements, thus reducing the value assigned to the land when they appeal based on the notion that the arms-length purchase for the combination was $X and the publicized value of the improvements calculates out to $Y, therefore leaving only $Z for the value of the unimproved land.

Actually, now that I think about it more, the output of the model would have to be a direct estimate of $Z (rather than an estimate of $Y, which is surely easier), making some of those concerns moot when the most recent sale was long ago (but not for new transactions).

I still can't figure out what the appeals process when the model gets it wrong (or is perceived to be wrong) is going to be based on. It seems like we're likely to end up right back at "human judgment", though perhaps that of an actual judge trying to interpret the parcel's value taking into account the model's output and the arguments of the plaintiff land owner instead of a city assessor trying to argue comparable sales with a land owner, which is a process most everyone can wrap their head around and have a pretty good feeling how they'd fare if they took it all the way to court.