Statistical modeling of spatial attributes aims to estimate dependence parameters, used to interpolate values at unsampled locations within the study area. Thus, the modeling process is conducted with some statistical criteria that ensure reliable predictions representing the actual local variability. This study evaluates different formulations of the Gaussian geostatistical model to reconstruct the surface of phosphorus over the area from the available measurements on 48 experimental plots located in Xanxerê/SC with emphasis on the methodological framework. Choices of covariates in the model and for data transformation define four modeling options to be assessed. The Matèrn correlation function was used, evaluated at values 0.5; 1.5 and 2.5 for smoothness parameter. Models were compared by the maximized logarithm of the likelihood function and also by cross validation. The model with transformed response variable, including coordinates of the area as covariates and the value of 0.5 for the smoothness parameter was selected. The cross validation measures did not add relevant information to the likelihood, and the analysis highlights care must be taken with globally or locally atypical data, as well as the need of objective choice based on different candidate models which ought to be the focus of geostatistical modeling to ensure results compatible with reality.
geoestatistical modeling; correlation function; estimation
