Comparative Study of the Performance of Seismic Denoising Methods Using Regional Data

Seismic waveform data are generally contaminated by noise from various sources, which interfere with the signals of interest. Thus, the efficiency of the noise suppression approach used early in the processing pipeline affects the quality of the downstream products. In this study, we implemented and applied different seismic denoising methods and their respective variants to data recorded by the regional network of the University of Utah Seismograph Stations. The denoising methods consist of frequency filtering, approaches based on nonlinear thresholding of continuous wavelet transforms (CWTs, e.g., Langston and Mousavi, 2019), and a convolutional neural network (CNN) denoiser (Tibi et al., 2021). Frequency filtering works by retaining signals within a predefined frequency band, while suppressing anything that lies outside that band. The CWT nonlinear soft thresholding involves first calculating a scale-dependent threshold using the characteristics of the pre-event noise, and assumes that the noise is stationary across the waveform. The denoising step subtracts off the threshold value from the CWT coefficients of the seismic waveform that are above the threshold, while setting to zero coefficients that are below the threshold. The CNN denoiser exploits a machine learning model trained using constructed noisy waveforms with known component (signal and noise) characteristics to process the input seismogram. Results involving 4780 constructed waveforms suggest that on average bandpass filter, the CWT, and CNN denoisers improve the signal-to-noise ratio (SNR) by about 5, 10, and 7 dB, respectively. In terms of waveform similarity and amplitude distortion for the recovered waveforms with respect to the ground truth (GT) seismograms, CNN denoising outperforms both frequency filtering and CWT denoising. The performance of all the approaches are depend on the SNR of the input waveforms; however, for frequency filtering the SNR of the processed waveform decreases significantly faster with decreasing SNR for the input seismogram. Also, we find that the average correlation coefficient value is about 0 for the seismograms processed with frequency filtering, which suggests that these waveforms are significantly different from their respective GTs, i.e., significant changes in waveform shape have occurred.