The aim of this paper is to present a method for the spatial prediction of agricultural crop yield. This estimation represents an essential issue in the definition of the economic policy of a country. Agricultural data are usually represented by spatially distributed observations and the corresponding statistical analysis is often based on the assumption of stationarity of the estimated parameters. This hypothesis is patently violated when the data are characterized by information relative to predefined but unknown sub-groups of the reference population. It is clear that for spatial data which follow this hypothesis, the main analytic goal is not to estimate the model parameters or to introduce a structural dependence among observations, but to identify the geographical units where the model parameters can be considered as locally stationary. In order to predict agricultural crop yield, we propose an approach based on the Simulated Annealing algorithm. We present an application of the proposed algorithm for the prediction of the yield of durum wheat of Foggia province (Italy) by using AGRIT 2005 data. The objective is to produce a map of potential yield for durum wheat through a regression of three purely geographical covariates, the two coordinates and the elevation, and on an agro- meteorological estimated yield model. The assumption of non-stationarity for agricultural data largely improves the obtained results. The spatial prediction of the yield of our proposed approach highlighted a greater precision in terms of the coefficient of variation if compared with the estimates obtained with Horvitz-Thompson estimator.