DAS-VSP coupled noise suppression method using weakly supervised structured feature extraction

Abstract Coupled noise suppression is one of the major denoising challenges in vertical seismic profiling (VSP) data acquired using distributed fibre-optic acoustic sensing (DAS) technology. Traditional methods to suppress coupled noise often require complicated parameter tuning, resulting in relatively low efficiency. Purely supervised deep learning methods demand a large amount of clean data as training labels, the acquisition of which is costly. Autoencoder (AE), as a deep learning framework that does not require manually annotated labels, has been widely used for random noise attenuation. In this study, we extend the AE to the suppression of coherent coupled noise in DAS-VSP data and propose a weakly supervised two-step coupled noise suppression workflow. Firstly, time windows dominated by coupled noise are manually selected from the target DAS-VSP data, and the AE is trained separately on these windows to reconstruct and remove the stable, coherent coupled noise component. Secondly, the AE with the same architecture is trained to further reconstruct the entire dataset, aiming to enhance event continuity and attenuate residual noise. We provide a mathematical argument to explain why the AE tends to reconstruct coupled noise preferentially. Both synthetic and field examples demonstrate that the proposed method can effectively suppress coupled noise in DAS-VSP data, outperforming traditional methods. Furthermore, by introducing Gaussian white noise to simulate complex noise contamination, the method is shown to be robust and generalizable to more complex noise environments. Under our experimental settings, the result further shows that for the same network depth, incorporating skip connections weakens the latent space’s preferential representation of coupled noise, leading to overall inferior performance compared with AE without skip connections.