Comment by andrewla

To expand on this differently -- if we have a model that produces estimates of the "true value of improved land", then we can validate that model whenever property sells. It would be a slow process but if we have a model and the model parameters capture a significant portion of the variance, then we would expect the model to converge over time.

Nobody does this, of course; it is not usually politically expedient to do so for a number of reasons, not least of which is the predictability of taxes for a given parcel. But at least it is theoretically asymptotically achievable. Not so for unimproved land value because the value you are trying to estimate is not an actual quantity.