📝 SummarXiv

Abstract

We introduce SeisDiff-denoNIA, a novel diffusion-based seismic denoising framework that trains directly on field noise, eliminating the reliance on synthetic datasets. Unlike conventional denoising methods that require clean signal labels, our approach leverages field noise extracted prior to first arrivals as training targets, allowing the diffusion model to explicitly learn the true noise distribution. The model demonstrates robust performance on field DAS-VSP data contaminated by different noise types, and significantly outperforms traditional signal-based diffusion models under low SNR conditions in synthetic tests. The results suggest that explicitly modeling noise is not only viable but advantageous for seismic denoising tasks.