Unsupervised DAS noise attenuation via double INR networks

Ji Li, Daniel O. Trad

Distributed acoustic sensing (DAS) has gained increased attention in geophysics due to its high sensing density, cost-effectiveness, and environmental benefits. However, the data acquired through this technique often contain various types of strong noise, making interpretation more challenging compared to traditional geophone data. These noises include high-frequency, high-amplitude erratic noise and coherent horizontal noise. To address these issues, we propose an unsupervised denoising framework called Double INR, which leverages two sub-networks of implicit neural representations (INR). One sub-network employs the l₁ norm to attenuate erratic noise, while the other uses the l₂ norm with a total variation (TV) regularization term to model the horizontal noise. The horizontal noisemodelled by the second sub-network is then subtracted from the denoised output of thefirst sub-network. We evaluate our method on both synthetic and real DAS data, demonstratingits effectiveness in noise reduction for DAS applications.