In precision agriculture, the generation of maps accounting the variability of soil attributes is crucial. Geostatistic has been used successfully to this purpose, since kriging, its interpolation algorithm, is a best linear unbiased estimator. However, not always the data concerning the target variable are available in a way that the stationary assumptions are met. Without stationarity the advantage of kriging fail, and other interpolations methods must be found. Artificial Neural Network (ANN) has recently been used to generate soil pedotransfer functions and mapping. The purpose of this paper is to analyze the efficiency of artificial neural network, based on coordinates of the area, to generate maps of spatial variability of field capacity. The study area belongs to an agro ecological farm located at Seropédica-RJ/BR. The data set was composed of 125 undisturbed soil samples collected in a depth of 0,10 m. In each sample, water retention at 10 kPa, was determined by Richards's extractor, and in order to analyze the spatial dependence, in each sample point, coordinates were determined. The methods of interpolation were compared through the parameters from cross validation and jack-knife. Analyzing table and figure 1, its possible to observe that the parameters of the spherical and linear models are very similar, however, considering the lower values of reduced mean, linear model was considered the best to interpolate field capacity. ANN presented the worst performance, probably because the coordinates information alone is not sufficient to do estimation of water retention. Besides, ANN presented higher estimated mean than ordinary kriging with spherical and linear model, and tended to super estimate field capacity. Despite the lower performance, ANN could be used to interpolate field capacity. Figure 1 – Model's of semivariograms fitted to the experimental data Table-1 Parameters from jack-knife and cross validation