Multiple-driven near-offset seismic data reconstruction based on deep learning

Near-offset data gaps are a systematic feature of marine streamer acquisition whose accurate reconstruction is critical for multiple attenuation, velocity analysis, and shallow subsurface imaging. Existing methods — whether based on prediction-error filters, transform-domain sparsity, or supervised deep learning — rely exclusively on the observed primary wavefield and therefore struggle under large contiguous gaps, either accumulating prediction errors progressively or lacking the receptive field coverage to bridge extended missing regions, with no explicit physical constraint to prevent geometrically inconsistent reconstructions. We propose a framework integrating the utilization of multiples with a deep neural network for near-offset reconstruction. By exploiting the multiple-order reduction property of the focal transform, surface-related multiples are converted into a physically interpretable focal-domain representation that involves subsurface structural information within the gap, serving as a physically grounded network input and obtained via a self-contained paired supervision strategy constructed directly from down-sampled focal-domain observations. Experiments on synthetic datasets and field data demonstrate superior structural continuity, amplitude fidelity, and robustness relative to conventional multiple-based PEF and supervised deep learning baselines.