Economists have been advocating for a land tax rather than a reg- ular property tax. There are, however, several challenges to value land for tax purposes. Indeed, data on vacant land transactions are scarce, land and structure are conventionally traded in a bundle and it is hard to capture all factors that determine the value of land. We propose to use a new Bayesian spatial model and apply the model to the uni- verse of vacant and improved land sales from Belgium in 2018. Our results indicate that vacant land prices are substantially more difficult to predict than house prices. However, the predictive performance of the spatial model improves considerably in comparison to a regular linear hedonic approach. Models that combine data from vacant and improved land are unable to improve the predictive accuracy.