A new data-processing technique is presented for the separation of initially up-traveling (ghost) energy from initially down-traveling (primary) energy on reflection seismograms. The method combines records from two or more shot depths after prefiltering each record with a different filter. The filters are designed on a least-mean-square-error criterion to extract primary reflections in the presence of ghost reflections and random noise. Filter design is dependent only on the difference in uphole time between shots, and is independent of the details of near-surface layering. The method achieves wide-band separation of primary and ghost energy, which results in 10-15 db greater attenuation of ghost reflections than can be achieved with conventional two- or three-shot stacking (no prefiltering) for ghost elimination.The technique is illustrated in terms of both synthetic and field examples. The deghosted field data are used to study the near-surface reflection response by computing the optimum linear filter to transform the deghosted trace back into the original ghosted trace. The impulse response of this filter embodies the effects of the near-surface on the reflection seismogram, i.e. the cause of the ghosting. Analysis of these filters reveals that the ghosting mechanism in the field test area consists of both a surface- and base-of-weathering layer reflector.