Improving the Signal-to-Noise Ratio of Single-Channel Earthquake Data With an Attention-Based UNet3+ and Phase-Difference Mask

Data denoising is a critical yet challenging task for many seismic applications, especially for weak and small signals with a low signal-to-noise ratio (SNR). We propose a novel denoising algorithm for the single-channel passive seismic data based on an attention-based UNet3+ network, which can efficiently learn a sparse representation of the data in the time-frequency domain and adaptively recover seismic signals contaminated by various noises. By incorporating the attention mechanism and full-scale skip connections, the proposed network can extract full-scale features and recover the weak and small earthquake signals in the presence of strong noise. The proposed method is trained using the Stanford Earthquake Dataset (STEAD) and a synthetic dataset. The phase-difference masks (PDMs) in the time-frequency domain are used as training labels. Results of test runs on more than 15000 constructed waveforms show that the proposed method achieves excellent denoising results compared to other widely used methods and improves the SNR on average by about 5 dB over a benchmark deep learning method, DeepDenoiser. Applications to field data show that our approach has good generalization and can effectively improve the detectability of seismic signals. Therefore, the proposed algorithm should have many potential applications such as event location, seismic imaging, microseismic monitoring, etc.