Missing Seismic Data Reconstruction Using Low‐Frequency Structure‐Guided 3D Diffusion

ABSTRACT In field seismic acquisition, consecutively missing traces, local acquisition gaps and large‐scale sampling incompleteness commonly arise from complex near‐surface conditions, acquisition‐geometry constraints, receiver failures and field obstacles, which can degrade detailed imaging. Compared with scattered dead traces, consecutively missing traces interrupt the lateral coherence of reflection events over a wider spatial range, making event continuation, fault‐boundary continuity and thin‐bed response recovery more difficult. Conventional prediction filtering, transform‐domain interpolation and low‐rank completion methods usually rely on local linearity, sparsity or low‐rank assumptions. Under strongly nonstationary structures and wide missing‐trace bands, these assumptions can lead to oversmoothing, boundary mismatch and high‐frequency loss. Deep‐learning methods have improved the non‐linear representation capacity of reconstruction models, but slice‐wise 2D modelling does not fully exploit 3D spatial correlation, whereas existing 3D methods often suffer from high training cost, long sampling chains and insufficient stable structural constraints. To address these issues, we propose a low‐frequency structure‐guided 3D conditional diffusion method for seismic data reconstruction. The method first extracts stable low‐frequency components from an incomplete 3D seismic volume and completes the long‐wavelength structure within the missing‐trace band by combining 3D structure‐tensor dip estimation, event‐oriented low‐frequency interpolation and reliability‐weighted fusion. The completed low‐frequency volume preserves the dominant reflection geometry, long‐wavelength phase trend and 3D horizon distribution and is used as a structural condition for subsequent full‐band recovery. The reconstruction problem is then formulated as a 3D conditional diffusion process constrained by the low‐frequency structure, in which a 3D U‐Net denoiser progressively restores mid‐ and high‐frequency reflection details during reverse sampling. To adapt the model to the spatial organization of 3D seismic wavefields, the volumetric convolutional feature extractor supports an optional lightweight 3D attention module to strengthen local waveform information and structural associations across the missing‐trace band along the time, inline and crossline directions. The experiments include tests with consecutively missing traces using the SEG C3 synthetic data and a controlled test with artificially removed traces in a prepared field‐data volume. Ablation tests and comparisons with 3D tensor completion, slice‐wise 2D diffusion and 3D U‐Net show that the completed low‐frequency structural condition reduces the structural ambiguity inside missing‐trace bands and enables the diffusion model to recover high‐frequency details while preserving event continuity, phase consistency and amplitude compatibility. By combining low‐frequency structural modelling, 3D wavefield continuity and conditional posterior generation, the framework provides a feasible route for high‐fidelity reconstruction of seismic data with wide missing‐trace bands.