Abstract
The rapid advancement of Distributed Acoustic Sensing (DAS) technology has opened up extensive prospects within the field of seismic exploration. However, unforeseeable noise present in actual DAS seismic records has led to the submergence of valuable information beneath intense noise, significantly disrupting reflective signals and diminishing the signal-to-noise ratio (SNR) of seismic data. Consequently, subsequent processing, such as migration and imaging, and interpretation tasks are hindered. In pursuit of an effective denoising approach for DAS data, this study proposes a Hybrid Multi-Resolution Network (HMR-Net), which concentrates on extracting coarse or intricate features from multi-resolution feature maps, thus delving into profound seismic characteristics across diverse scales and resolutions. The integration of error-resilient up-sampling and down-sampling processes serves to optimize the feature extraction ability and mitigate losses arising from sampling procedures. Furthermore, a highly authentic dataset was compiled by utilizing real DAS noise data along with synthetic records obtained through forward simulations. Through validation against both synthesized records and actual seismic records, the effectiveness of the proposed approach in substantially suppressing noise and enhancing the SNR has been demonstrated.