Deep Learning Denoising for Synchronous Surface Three-Component Seismic Data Based on Huainan Deep Underground Observation

Surface-acquired seismic data have long been plagued by near-surface noise contamination, significantly limiting the utility of field records. Recent progress in ultra-low-background observation from deep underground laboratories has provided a new reference benchmark for high-precision multiphysical signal analysis. Leveraging low-background seismic data collected from the Huainan Deep Underground Laboratory, this study proposes a deep-learning-based noise cleansing method that integrates long short-term memory networks with an auto-encoder architecture. Employing a self-supervised learning mechanism, the method extracts essential features of clean seismic signals from deep underground data then leverages the wavefield homology of the collocated synchronous surface and underground observations to achieve adaptive denoising of strongly contaminated surface data. Experimental results show that for ambient noise without valid seismic events, the model effectively suppresses high-frequency noise above 1 Hz, achieving a maximum high-frequency suppression ratio of 0.986, while largely preserving the integrity of the low-frequency background field. For records containing valid seismic events, the model successfully reconstructs masked signals from heavily polluted data, with a maximum signal-to-noise ratio (SNR) improvement of 8.21 dB. Furthermore, comparisons between single-component (1C) and three-component joint (3C) vector processing reveal that the 1C model performs better in characteristic frequency retention and low-frequency signal fidelity, whereas the 3C vector processing approach yields superior high-frequency noise attenuation and overall SNR enhancement.