Seismic signal processing often relies on general convolutional neural network (CNN)-based models, which typically focus on features in the time domain while neglecting frequency characteristics. Moreover, down-sampling operations in these models tend to cause the loss of critical high-frequency details. To this end, we propose a feedback information distillation network (FID-N) in the non-subsampled contourlet transform (NSCT) domain to remarkably suppress seismic noise. The method aims to enhance denoising performance by preserving the fine-grained details and frequency characteristics of seismic data. The FID-N mainly consists of a two-path information distillation block used in a recurrent manner to form a feedback mechanism, carrying an output to correct previous states, which fully exploits competitive features from seismic signals and effectively realizes the signal restoration step by step across time. Additionally, the NSCT has an excellent high-frequency response and powerful curve and surface description capabilities. We suggest converting the noise suppression problem into NSCT coefficient prediction, which maintains more detailed high-frequency information and promotes the FID-N to further suppress noise. Extensive experiments on both synthetic and real seismic datasets demonstrated that our method significantly outperformed the SOTA methods, particularly in scenarios with low signal-to-noise ratios and in recovering high-frequency components.

The Lianggaoshan Formation in the Gongshanmiao block of the Sichuan Basin is characterized by narrow channel sand bodies and thin layers, resulting in weak seismic responses, which leads to poor identification of small-scale fault–fracture systems. Initial high-density 3D seismic data exhibit strong shallow surface waves and significant shot-to-shot variations in energy and frequency, necessitating amplitude-preserving noise attenuation and broadband wavelet consistency processing. First, pre-stack multi-domain amplitude-preserving noise attenuation is applied, integrating surface-wave forward modeling, stationary wavelet transform, and matrix singular value decomposition to suppress complex noise. Next, robust deconvolution constrained by a target wavelet improves broadband consistency. Subsequently, anisotropic depth-domain velocity modeling and imaging under rugged topography are conducted using a well-constrained TTI initial velocity model and full-azimuth angle-domain grid tomography. Compared with conventional data, the processed high-density data significantly enhance bandwidth, structural imaging, and thin-layer resolution. Imaging continuity of small faults (6–10 m throw) is markedly improved, and the channel characterization accuracy of the Liang-2 Member increases from 180 m to 60 m. This workflow delivers high-SNR, high-resolution, and high-fidelity results, providing a reliable basis for thin-sandbody prediction and reservoir evaluation.

Denoising is central to geophysical data processing because the quality of a seismic record largely determines how reliably subsurface structures can be interpreted and resources evaluated. Field data, however, are typically corrupted by random noise, and clean labels are seldom available in practical surveys. In this paper, we address this gap with an improved self-supervised Trace2Trace framework for seismic denoising. Our method builds input-target pairs from the odd and even traces of a seismic record and, by exploiting the spatial correlation between neighboring traces, learns a denoising mapping purely from noisy observations, so no clean reference is needed. Conventional Trace2Trace tends to over-smooth precisely because neighboring traces are so similar, and to counter this we add a neighborhood structure regularization mechanism: a structural constraint term incorporated into the loss function keeps local discontinuities and edges intact, sharpening the representation of discontinuous geology, while a dedicated structure difference matrix strengthens the capture of faults and other critical discontinuities. Based on an idealized noise-free network assumption, we further develop a dual-path regularized optimization strategy that suppresses residual noise without eroding useful signal. Across both synthetic and real seismic datasets, the proposed method achieves improved denoising performance compared with existing baselines. On the synthetic dataset, the proposed method achieves an SNR of 17.91 dB, improving the SNR by 0.98 dB over the best-performing baseline, while preserving reflection events, local amplitude variations, and fault boundaries more faithfully. Its performance is further examined on real seismic data.