Estimation of primaries by sparse inversion with scattering-based multiple predictions for data with large gaps