Although accounting for the spatial - temporal relationships in farmland valuation has gained attention in the literature recently, misspecification and incorrectly imposed assumptions on spatial weighting matrix can often produce misleading estimates and inference compared to maintaining ignorance of spatial dependence structure among spatially observed farmland values. In this study I assemble a panel data set using Pennsylvania county level farmland values reported in the U.S. Census of Agriculture between 1982 and 2007, and estimate the spatial weighting matrix among farmland values semiparametrically. A spatial lag panel data model with the consistently estimated spatial weighting matrix is then estimated via maximum likelihood estimation (MLE). The results show that the proposed approach can substantially improve the goodness of fit of the spatial hedonic model of farmland values therefore the reliability of obtained price elasticity estimates.