Self-Supervised Attention-Guided Multi-Loss Framework for Wide-Band Random Noise Suppression in an Iranian Oil Field

Noise attenuation is a fundamental step in seismic data processing, as noise can distort both the energy and frequency content of target reflections. This study presents a self-supervised blind-spot, attention-guided multi-loss framework designed for seismic random noise attenuation. The network employs a UNet architecture for feature extraction. It is driven by three components: (i) a masked pixel-wise MAE loss, (ii) a Structural Similarity Index Measure loss and (iii) a coherency-based attention module operating on the residual map to distinguish noise-like regions from coherent signal leakage. Model training was performed in two stages. First, the network was pretrained on synthetic seismic data contaminated with broadband random noise (۰–۸۰ Hz) and validated on blind synthetic test data. Second, the pretrained model was fine-tuned on a field data in Iran. Results demonstrate that the proposed framework not only preserves the true seismic amplitude but also enhances reflection continuity and suppresses broadband random noise in both synthetic and field seismic data. Owing to its amplitude-preserving nature, the denoised outputs can be directly utilized for subsequent post-stack inversion.

Keywords:

Denoising , Deep learning , Self-supervised , Attention , UNet