The spatial variability of structures in regionalized variables are defined with the aid of geostatistical techniques, which facilitate the estimation of values for these variables in unsampled localizations and generate thematic maps to be used in decision making for localized treatments in the area under study. The quality of these maps depends on the trustworthiness of these estimates that can be modified with the choice for the sample design. The objective of this work was to establish an optimal size and shape of the sample designs in order to enhance the efficiency of sampling plans for the prediction of space dependent variables. These designs were obtained with the use of a stochastic search method called Simulated Annealing. This method is based on a sampling grid with a large number of points. Here, it is initially used to consider simulated data sets with distinct spatial dependence structures and is then used to consider real data on soy productivity. The simulated results are used as reference for the achievement of the best sample design with the lowest number of sample points that can efficiently represent the spatial dependence structure of soy productivity in a commercial area harvested by the harvester monitor. The results reported for the simulations and soy productivity data show that the optimization process was efficient in determining sample designs with reduced size, especially when using the Global Accuracy as the measurement to be maximized.