Feasibility Study of Denoising Instrumental Noise of Raw SOV-DAS Data with Deep Learning

Summary This study explores the potential of using deep learning to denoise raw Surface Orbital Vibrator-Distributed Acoustic Sensing (SOV-DAS) data in geophysical applications. Distributed Acoustic Sensing (DAS) has gained prominence for its cost-effectiveness and high spatial density in seismic monitoring. However, the risk of noise-signal mixture during data processing remains a challenge. The research begins by detailing the mechanics of SOV-DAS data processing, highlighting the need for deconvolution and the risk of signal-noise mixtures. To address this challenge, a Noise2Noise (N2N) deep learning model is applied for denoising before deconvolution. The study uses data from a CO2 injection experiment conducted at OTIC in 2021. Results show that the N2N approach effectively attenuates high-frequency noise in the SOV-DAS dataset, improving the Signal-to-Noise Ratio (SNR) and data quality. Deconvolution of denoised data enhances the contrast between signals and noise, outperforming stack gathering methods. In conclusion, this research demonstrates the feasibility of denoising instrumental noise in raw SOV-DAS data using deep learning. The denoising process leads to enhanced data quality, increased SNR, and potential efficiency gains in data acquisition and processing. This approach holds promise for improving 4D time-lapse monitoring in geophysical applications, offering valuable insights for researchers and practitioners in the field.