A Markov random field (MRF) model is employed to learn the statistical properties of an equivalent current situated above a rough surface, where this equivalent current represents near-field electromagnetic scattered fields. The MRF parameters are learned by considering numerical scattering from a relatively small rough surface, which can be solved efficiently via existing numerical methods. The MRF parameters may then be used to synthesize multiple realizations of the equivalent current associated with scattering from a much larger rough surface. These equivalent currents may be used to efficiently calculate the far-field scattered-field statistics for large surfaces. The advantages of this approach are manifested in the fact that once the MRF parameters are learned, a forward model is no longer needed to generate random realizations of the equivalent currents, and the MRF model can be employed to generate equivalent surfaces of size beyond the capability of existing numerical models. To improve the accuracy of the MRF model, a narrowband filter is used to preprocess the computed current distributions prior to MRF model training. One-dimensional rough surfaces are considered here to demonstrate the idea, and the method of ordered multiple interactions (MOMI) is employed to perform the numerical forward modeling. Promising MRF-generated results are demonstrated, with comparison to direct MOMI solutions. © 2008 IEEE.