Deep Learning-Based Blind Denoising for Distributed Acoustic Sensing Seismic Data With Self-Supervised and Transfer Learning

Abstract A distributed acoustic sensing (DAS) technology, extensively utilized in the seabed geological exploration, ocean current analysis, and marine seismic monitoring, faces challenges due to the presence of various noise types in sensing signals, which complicates signal recognition and analysis. This paper introduces a novel deep learning-based strategy for the blind denoising of the DAS seismic data. We propose an advanced neural network training approach that leverages self-supervised learning and transfer learning methodologies, enhanced by the integration of a physical model of seismic transmission. This innovation leads to superior signal correlation across adjacent DAS channels. Our method outperforms other denoising algorithms under identical conditions, achieving the highest windowed signal-to-noise ratio of approximately 26 dB and the fastest processing time of 0.32 seconds, establishing itself as a significant and promising technique for field applications involving DAS data.