Non-local self-similarity guided graph attention network for DAS-VSP noise and signal separation

Abstract

Fiber-optic distributed acoustic sensing (DAS) technology represents an emerging seismic exploration technique. Due to multicomponent noise contamination, noise suppression and separation have been paramount tasks in DAS. Non-local self-similarity (NSS) serving as an effective prior in seismic data was initially utilized for separating noise and signal. However, the existing deep learning nonlocal methods are pixel-based, they are prone to biases when confronted with the strong noise characteristics of DAS VSP data. To address the aforementioned issues, this paper proposes and validates the effective prior of nonlocal self-similarity based on patch-wise analysis for DAS VSP data. Inspired by the powerful ability of graph attention networks (GATs) for non-local feature aggregation, we introduce a dynamic attention graph learning model. This model utilizes patch-wise graph convolution to capture the long-range correlations in DAS, significantly mitigating the impact of noise levels on the separation of noise and signal. Furthermore, we design the graph-based channel attention module (CE) to dynamically construct the node and edge sets of the graph, generating adaptive thresholds for each node to select neighbors with similarity higher than the threshold as edges for graph connection. This enables the model to focus on key information in seismic signal. In experiments with synthetic DAS data, our method achieved an improvement in signal-to-noise ratio (SNR) of approximately 26.6 dB. In experiments with real DAS data, the separating performance was also superior to other methods. The proposed network can serve as an effective tool for processing DAS VSP data.