A methodology for the identification of random field models for the description of seismic ground motions recorded over extended areas is presented. The approach (i) embeds seismic ground motion parameter estimation in a probabilistic framework, and (ii) identifies simultaneously point and spatial models for the seismic ground motion random field from either smoothed or unsmoothed non-parametric spectral estimates. It is shown that parametric coherency estimates identified by means of this approach are stable and insensitive to the amount of smoothing performed on the empirical data. The approach and its findings have a significant impact on the evaluation of parametric ground motion models for the seismic response evaluation of lifelines. Copyright © 2003 John Wiley & Sons, Ltd.