Probabilistic site characterization usually requires a parametric model such as the Gaussian random field to begin. This paper proposes a nonparametric approach to characterizing soil spatial variability based on the maximum entropy method. The finite-dimensional distribution specifying a random field is decomposed into marginal distributions and vine-structured copulas. Then, the distribution is adaptively developed under the moment constraints, which are classified into two categories. The first category describes the uncertainty of the random quantity at an arbitrary point, while the second category defines the dependence between the random quantities at any two points. The marginal distribution and vine-structured copulas are developed under the two types of moment constraints, respectively, which are formulated as two optimization problems. The Akaike information criterion (AIC) is adopted for model selection, and the hypothesis test based on the probability integral transformation (PIT) is used to examine whether the spatial variability is modeled appropriately. Analysis results of a hypothetical case and several real soil profiles indicate that the nonparametric approach is feasible. The nonparametric models are comparable or superior to the parametric or Gaussian models, and no prior assumption is made on the distribution of the random field.