Recorded seismic data are obtained at a plurality of receivers from a plurality of sources, and a set of sources from the plurality of sources is selected using spatial criteria based on a location of each source such that any two sources in the set of sources are separated by a predefined minimum distance of separation sufficient to reduce cross talk between sources. The set of sources is combined in a non-encoded manner into a composite source, and forward modeling for the composite source is performed to generate a synthetic seismic data set. A composite recorded seismic data set for the set of sources is determined, and the synthetic seismic data set and composite recorded seismic data set are used to determine a residual seismic data set. Backward modeling generates a gradient update used to generate an updated earth model.