Common-mode noise separation in distributed acoustic sensing vertical seismic profile data: A self-supervised deep learning approach with enhanced blind network

Distributed fiber-optic acoustic sensing (DAS) has emerged as a transformative technology for seismic exploration, offering adaptability, environmental sustainability and cost-efficiency, especially for vertical seismic profile (VSP) acquisition. However, practical applications face the challenges of diverse random noise, incoherent background noise, and weak target signals. Conventional supervised learning methods for noise separation are based on noise-free labels, which limits their applicability in seismic exploration. To address this gap, a novel label-free self-supervised learning approach, called dual-model self-supervised selective learning (dSSSL) is proposed to separate common-mode noise, specifically coherent noise typical of DAS-VSP data, from the desired signals. Our methodology leverages blind horizontal networks (BHNs) to mitigate the common-mode noise, and overcomes the identity mapping challenges inherent in U-Net with the introduction of the blind spot network (BSN). The effectiveness of our method is validated with synthetic DAS-VSP and field data from Janghang-ri, showing that the signal recovery is superior compared to conventional methods. This approach is promising for improving seismic data processing and interpretation in distributed fiber-optic acoustic sensing applications, particularly for vertical seismic profile acquisition.