ZSN2N-DAS: Zero-Shot Noise2Noise Denoising for Distributed Acoustic Sensing Data

Distributed acoustic sensing (DAS) is a highly appealing technology for seismology, prized for its flexibility, durability, and high sensitivity to vibrations and strain field changes. However, despite these advantages, optical fibers are susceptible to external interference, which introduces incoherent noise during measurements. Recognizing that this noise in raw DAS data is unstructured and has a zero mean, we propose ZSN2N-DAS, a self-supervised learning method that achieves zero-shot Noise2Noise (N2N) denoising for ground-based DAS data. The effectiveness of ZSN2N-DAS is underpinned by a mathematically proven equivalence. Results demonstrate its ability to effectively suppress incoherent noise and significantly enhance the signal-to-noise ratio (SNR) of microseismic events, as validated using datasets from a DAS array on the Rutford Ice Stream in Antarctica. Furthermore, compared to the current leading deep learning model, N2N-DAS, ZSN2N-DAS achieves a 54% reduction in network parameters while simultaneously improving the SNR of seismic signals by approximately 20% and mitigating the issue of signal leakage. Consequently, in arrival time picking, ZSN2N-DAS enhances picking quality and increases the average number of effectively picked arrivals by roughly 28%. The denoised data produced by ZSN2N-DAS can benefit downstream seismological analyses like magnitude estimation, hypocenter location, and focal mechanism inversion.