Seismic Denoising by Deep Learning From Natural Repeating Earthquakes

Effective seismic waveform denoising is crucial for advancing our understanding of Earth's subsurface structures and dynamics. Recently, deep learning-based denoising methods have emerged and shown remarkable performance in improving the signal-to-noise ratio (SNR) of seismic waveforms. However, sole improvement of SNR could be misleading, since it is challenging to validate denoising performance in the signal window of real data. Here, we present a new P-wave denoising network (PD-Net), which is trained and validated using naturally similar waveforms from global repeating earthquakes. Our approach employs a novel loss function designed to enhance the similarity of repetitive waveform pairs, enabling effective noise removal while preserving signal fidelity. The testing results demonstrate the effectiveness of our method compared to previous studies, as indicated by the significant improvement of waveform similarity. Our study establishes a new paradigm for evaluating denoising performance and provides insights into training denoising networks for various types of seismic waves.